Build
Browse filesThis view is limited to 50 files because it contains too many changes.
See raw diff
- build/torch25-cxx11-cu118-x86_64-linux/moe/{_moe_2pimofs7erzvi.abi3.so → _moe_z6j3gzsycn542.abi3.so} +2 -2
- build/torch25-cxx11-cu118-x86_64-linux/moe/_ops.py +3 -3
- build/torch25-cxx11-cu118-x86_64-linux/moe/fp8.py +175 -6
- build/torch25-cxx11-cu118-x86_64-linux/moe/fused_marlin_moe.py +43 -8
- build/torch25-cxx11-cu118-x86_64-linux/moe/fused_moe.py +788 -77
- build/torch25-cxx11-cu118-x86_64-linux/moe/platforms.py +15 -5
- build/torch25-cxx11-cu121-x86_64-linux/moe/{_moe_pqwfgssq5enn2.abi3.so → _moe_tuji4gj3mmhfo.abi3.so} +2 -2
- build/torch25-cxx11-cu121-x86_64-linux/moe/_ops.py +3 -3
- build/torch25-cxx11-cu121-x86_64-linux/moe/fp8.py +175 -6
- build/torch25-cxx11-cu121-x86_64-linux/moe/fused_marlin_moe.py +43 -8
- build/torch25-cxx11-cu121-x86_64-linux/moe/fused_moe.py +788 -77
- build/torch25-cxx11-cu121-x86_64-linux/moe/platforms.py +15 -5
- build/torch25-cxx11-cu124-x86_64-linux/moe/{_moe_lwzoz7knnxf4i.abi3.so → _moe_pss5doo675cd4.abi3.so} +2 -2
- build/torch25-cxx11-cu124-x86_64-linux/moe/_ops.py +3 -3
- build/torch25-cxx11-cu124-x86_64-linux/moe/fp8.py +175 -6
- build/torch25-cxx11-cu124-x86_64-linux/moe/fused_marlin_moe.py +43 -8
- build/torch25-cxx11-cu124-x86_64-linux/moe/fused_moe.py +788 -77
- build/torch25-cxx11-cu124-x86_64-linux/moe/platforms.py +15 -5
- build/torch25-cxx98-cu118-x86_64-linux/moe/{_moe_uhyif3wslpwak.abi3.so → _moe_5uyw6qhdybj5e.abi3.so} +2 -2
- build/torch25-cxx98-cu118-x86_64-linux/moe/_ops.py +3 -3
- build/torch25-cxx98-cu118-x86_64-linux/moe/fp8.py +175 -6
- build/torch25-cxx98-cu118-x86_64-linux/moe/fused_marlin_moe.py +43 -8
- build/torch25-cxx98-cu118-x86_64-linux/moe/fused_moe.py +788 -77
- build/torch25-cxx98-cu118-x86_64-linux/moe/platforms.py +15 -5
- build/torch25-cxx98-cu121-x86_64-linux/moe/_moe_tj3osoay2niyk.abi3.so +3 -0
- build/torch25-cxx98-cu121-x86_64-linux/moe/_moe_xsk7dxl7fy4pk.abi3.so +0 -3
- build/torch25-cxx98-cu121-x86_64-linux/moe/_ops.py +3 -3
- build/torch25-cxx98-cu121-x86_64-linux/moe/fp8.py +175 -6
- build/torch25-cxx98-cu121-x86_64-linux/moe/fused_marlin_moe.py +43 -8
- build/torch25-cxx98-cu121-x86_64-linux/moe/fused_moe.py +788 -77
- build/torch25-cxx98-cu121-x86_64-linux/moe/platforms.py +15 -5
- build/torch25-cxx98-cu124-x86_64-linux/moe/_moe_b25pgchg5o5pa.abi3.so +0 -3
- build/torch25-cxx98-cu124-x86_64-linux/moe/_moe_phlujktdbqekw.abi3.so +3 -0
- build/torch25-cxx98-cu124-x86_64-linux/moe/_ops.py +3 -3
- build/torch25-cxx98-cu124-x86_64-linux/moe/fp8.py +175 -6
- build/torch25-cxx98-cu124-x86_64-linux/moe/fused_marlin_moe.py +43 -8
- build/torch25-cxx98-cu124-x86_64-linux/moe/fused_moe.py +788 -77
- build/torch25-cxx98-cu124-x86_64-linux/moe/platforms.py +15 -5
- build/torch26-cxx11-cu118-x86_64-linux/moe/_moe_ooomuvan6f6yy.abi3.so +0 -3
- build/torch26-cxx11-cu118-x86_64-linux/moe/_moe_zlz7rpd2goyn2.abi3.so +3 -0
- build/torch26-cxx11-cu118-x86_64-linux/moe/_ops.py +3 -3
- build/torch26-cxx11-cu118-x86_64-linux/moe/fp8.py +175 -6
- build/torch26-cxx11-cu118-x86_64-linux/moe/fused_marlin_moe.py +43 -8
- build/torch26-cxx11-cu118-x86_64-linux/moe/fused_moe.py +788 -77
- build/torch26-cxx11-cu118-x86_64-linux/moe/platforms.py +15 -5
- build/torch26-cxx11-cu124-x86_64-linux/moe/_moe_h5rxhm5fum47w.abi3.so +0 -3
- build/torch26-cxx11-cu124-x86_64-linux/moe/_moe_wua27hyvpwmli.abi3.so +3 -0
- build/torch26-cxx11-cu124-x86_64-linux/moe/_ops.py +3 -3
- build/torch26-cxx11-cu124-x86_64-linux/moe/fp8.py +175 -6
- build/torch26-cxx11-cu124-x86_64-linux/moe/fused_marlin_moe.py +43 -8
build/torch25-cxx11-cu118-x86_64-linux/moe/{_moe_2pimofs7erzvi.abi3.so → _moe_z6j3gzsycn542.abi3.so}
RENAMED
@@ -1,3 +1,3 @@
|
|
1 |
version https://git-lfs.github.com/spec/v1
|
2 |
-
oid sha256:
|
3 |
-
size
|
|
|
1 |
version https://git-lfs.github.com/spec/v1
|
2 |
+
oid sha256:9664c7b8a4e935582354443bebc5557041cac1d35b4b483abe73b4559d7c468c
|
3 |
+
size 85827696
|
build/torch25-cxx11-cu118-x86_64-linux/moe/_ops.py
CHANGED
@@ -1,9 +1,9 @@
|
|
1 |
import torch
|
2 |
-
from . import
|
3 |
-
ops = torch.ops.
|
4 |
|
5 |
def add_op_namespace_prefix(op_name: str):
|
6 |
"""
|
7 |
Prefix op by namespace.
|
8 |
"""
|
9 |
-
return f"
|
|
|
1 |
import torch
|
2 |
+
from . import _moe_z6j3gzsycn542
|
3 |
+
ops = torch.ops._moe_z6j3gzsycn542
|
4 |
|
5 |
def add_op_namespace_prefix(op_name: str):
|
6 |
"""
|
7 |
Prefix op by namespace.
|
8 |
"""
|
9 |
+
return f"_moe_z6j3gzsycn542::{op_name}"
|
build/torch25-cxx11-cu118-x86_64-linux/moe/fp8.py
CHANGED
@@ -1,6 +1,11 @@
|
|
|
|
|
|
1 |
import torch
|
|
|
|
|
2 |
|
3 |
-
|
|
|
4 |
|
5 |
|
6 |
def is_hip() -> bool:
|
@@ -49,15 +54,179 @@ def scaled_fp8_quant(
|
|
49 |
if scale is None:
|
50 |
if use_per_token_if_dynamic:
|
51 |
scale = torch.empty((shape[0], 1), device=input.device, dtype=torch.float32)
|
52 |
-
|
53 |
-
output, input, scale, scale_ub
|
54 |
-
)
|
55 |
else:
|
56 |
scale = torch.zeros(1, device=input.device, dtype=torch.float32)
|
57 |
-
|
58 |
else:
|
59 |
# num_token_padding not implemented for this case
|
60 |
assert scale.numel() == 1 or num_token_padding is None
|
61 |
-
|
62 |
|
63 |
return output, scale
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
1 |
+
from typing import Tuple, Optional, Union
|
2 |
+
|
3 |
import torch
|
4 |
+
import triton
|
5 |
+
import triton.language as tl
|
6 |
|
7 |
+
|
8 |
+
from ._ops import ops
|
9 |
|
10 |
|
11 |
def is_hip() -> bool:
|
|
|
54 |
if scale is None:
|
55 |
if use_per_token_if_dynamic:
|
56 |
scale = torch.empty((shape[0], 1), device=input.device, dtype=torch.float32)
|
57 |
+
ops.dynamic_per_token_scaled_fp8_quant(output, input, scale, scale_ub)
|
|
|
|
|
58 |
else:
|
59 |
scale = torch.zeros(1, device=input.device, dtype=torch.float32)
|
60 |
+
ops.dynamic_scaled_fp8_quant(output, input, scale)
|
61 |
else:
|
62 |
# num_token_padding not implemented for this case
|
63 |
assert scale.numel() == 1 or num_token_padding is None
|
64 |
+
ops.static_scaled_fp8_quant(output, input, scale)
|
65 |
|
66 |
return output, scale
|
67 |
+
|
68 |
+
|
69 |
+
@triton.jit
|
70 |
+
def _per_token_group_quant_fp8(
|
71 |
+
# Pointers to inputs and output
|
72 |
+
y_ptr,
|
73 |
+
y_q_ptr,
|
74 |
+
y_s_ptr,
|
75 |
+
group_size,
|
76 |
+
# Avoid to divide zero
|
77 |
+
eps,
|
78 |
+
# Information for float8
|
79 |
+
fp8_min,
|
80 |
+
fp8_max,
|
81 |
+
# Meta-parameters
|
82 |
+
BLOCK: tl.constexpr,
|
83 |
+
):
|
84 |
+
"""A Triton-accelerated function to perform per-token-group
|
85 |
+
quantization on a tensor.
|
86 |
+
This function converts the tensor values into float8 values.
|
87 |
+
"""
|
88 |
+
# Map the program id to the row of X and Y it should compute.
|
89 |
+
g_id = tl.program_id(0)
|
90 |
+
y_ptr += g_id * group_size
|
91 |
+
y_q_ptr += g_id * group_size
|
92 |
+
y_s_ptr += g_id
|
93 |
+
|
94 |
+
cols = tl.arange(0, BLOCK) # N <= BLOCK
|
95 |
+
mask = cols < group_size
|
96 |
+
|
97 |
+
y = tl.load(y_ptr + cols, mask=mask, other=0.0).to(tl.float32)
|
98 |
+
# Quant
|
99 |
+
_absmax = tl.maximum(tl.max(tl.abs(y)), eps)
|
100 |
+
y_s = _absmax / fp8_max
|
101 |
+
y_q = tl.clamp(y / y_s, fp8_min, fp8_max).to(y_q_ptr.dtype.element_ty)
|
102 |
+
|
103 |
+
tl.store(y_q_ptr + cols, y_q, mask=mask)
|
104 |
+
tl.store(y_s_ptr, y_s)
|
105 |
+
|
106 |
+
|
107 |
+
@triton.jit
|
108 |
+
def _per_token_group_quant_fp8_colmajor(
|
109 |
+
# Pointers to inputs and output
|
110 |
+
y_ptr,
|
111 |
+
y_q_ptr,
|
112 |
+
y_s_ptr,
|
113 |
+
group_size,
|
114 |
+
# Num columns of y
|
115 |
+
y_num_columns,
|
116 |
+
# Stride from one column to the next of y_s
|
117 |
+
y_s_col_stride,
|
118 |
+
# Avoid to divide zero
|
119 |
+
eps,
|
120 |
+
# Information for float8
|
121 |
+
fp8_min,
|
122 |
+
fp8_max,
|
123 |
+
# Meta-parameters
|
124 |
+
BLOCK: tl.constexpr,
|
125 |
+
):
|
126 |
+
"""A Triton-accelerated function to perform per-token-group
|
127 |
+
quantization on a tensor.
|
128 |
+
This function converts the tensor values into float8 values.
|
129 |
+
"""
|
130 |
+
# Map the program id to the row of X and Y it should compute.
|
131 |
+
g_id = tl.program_id(0)
|
132 |
+
y_ptr += g_id * group_size
|
133 |
+
y_q_ptr += g_id * group_size
|
134 |
+
|
135 |
+
# Convert g_id the flattened block coordinate to 2D so we can index
|
136 |
+
# into the output y_scales matrix
|
137 |
+
blocks_per_row = y_num_columns // group_size
|
138 |
+
scale_col = g_id % blocks_per_row
|
139 |
+
scale_row = g_id // blocks_per_row
|
140 |
+
y_s_ptr += scale_col * y_s_col_stride + scale_row
|
141 |
+
|
142 |
+
cols = tl.arange(0, BLOCK) # group_size <= BLOCK
|
143 |
+
mask = cols < group_size
|
144 |
+
|
145 |
+
y = tl.load(y_ptr + cols, mask=mask, other=0.0).to(tl.float32)
|
146 |
+
# Quant
|
147 |
+
_absmax = tl.maximum(tl.max(tl.abs(y)), eps)
|
148 |
+
y_s = _absmax / fp8_max
|
149 |
+
y_q = tl.clamp(y / y_s, fp8_min, fp8_max).to(y_q_ptr.dtype.element_ty)
|
150 |
+
|
151 |
+
tl.store(y_q_ptr + cols, y_q, mask=mask)
|
152 |
+
tl.store(y_s_ptr, y_s)
|
153 |
+
|
154 |
+
|
155 |
+
def per_token_group_quant_fp8(
|
156 |
+
x: torch.Tensor,
|
157 |
+
group_size: int,
|
158 |
+
eps: float = 1e-10,
|
159 |
+
dtype: Optional[torch.dtype] = None,
|
160 |
+
column_major_scales: bool = False,
|
161 |
+
) -> Tuple[torch.Tensor, torch.Tensor]:
|
162 |
+
"""Function to perform per-token-group quantization on an input tensor `x`.
|
163 |
+
It converts the tensor values into signed float8 values and returns the
|
164 |
+
quantized tensor along with the scaling factor used for quantization.
|
165 |
+
Args:
|
166 |
+
x: The input tensor with ndim >= 2.
|
167 |
+
group_size: The group size used for quantization.
|
168 |
+
eps: The minimum to avoid dividing zero.
|
169 |
+
dtype: The dype of output tensor. Note that only `torch.float8_e4m3fn`
|
170 |
+
is supported for now.
|
171 |
+
Returns:
|
172 |
+
Tuple[torch.Tensor, torch.Tensor]: The quantized tensor and the
|
173 |
+
scaling factor for quantization.
|
174 |
+
"""
|
175 |
+
if dtype is None:
|
176 |
+
dtype = (
|
177 |
+
torch.float8_e4m3fnuz if current_platform.is_rocm() else torch.float8_e4m3fn
|
178 |
+
)
|
179 |
+
assert x.shape[-1] % group_size == 0, (
|
180 |
+
f"the last dimension of `x` {x.shape[-1]} must be divisible "
|
181 |
+
f"by `group_size` {group_size}"
|
182 |
+
)
|
183 |
+
assert x.is_contiguous(), "`x` must be contiguous"
|
184 |
+
|
185 |
+
finfo = torch.finfo(dtype)
|
186 |
+
fp8_min = finfo.min
|
187 |
+
fp8_max = finfo.max
|
188 |
+
|
189 |
+
x_q = torch.empty_like(x, device=x.device, dtype=dtype)
|
190 |
+
M = x.numel() // group_size
|
191 |
+
N = group_size
|
192 |
+
if column_major_scales:
|
193 |
+
shape = (x.shape[-1] // group_size,) + x.shape[:-1]
|
194 |
+
x_s = torch.empty(shape, device=x.device, dtype=torch.float32).permute(-1, -2)
|
195 |
+
else:
|
196 |
+
shape = x.shape[:-1] + (x.shape[-1] // group_size,)
|
197 |
+
x_s = torch.empty(shape, device=x.device, dtype=torch.float32)
|
198 |
+
|
199 |
+
BLOCK = triton.next_power_of_2(N)
|
200 |
+
# heuristics for number of warps
|
201 |
+
num_warps = min(max(BLOCK // 256, 1), 8)
|
202 |
+
num_stages = 1
|
203 |
+
if column_major_scales:
|
204 |
+
_per_token_group_quant_fp8_colmajor[(M,)](
|
205 |
+
x,
|
206 |
+
x_q,
|
207 |
+
x_s,
|
208 |
+
group_size,
|
209 |
+
x.shape[1],
|
210 |
+
x_s.stride(1),
|
211 |
+
eps,
|
212 |
+
fp8_min=fp8_min,
|
213 |
+
fp8_max=fp8_max,
|
214 |
+
BLOCK=BLOCK,
|
215 |
+
num_warps=num_warps,
|
216 |
+
num_stages=num_stages,
|
217 |
+
)
|
218 |
+
else:
|
219 |
+
_per_token_group_quant_fp8[(M,)](
|
220 |
+
x,
|
221 |
+
x_q,
|
222 |
+
x_s,
|
223 |
+
group_size,
|
224 |
+
eps,
|
225 |
+
fp8_min=fp8_min,
|
226 |
+
fp8_max=fp8_max,
|
227 |
+
BLOCK=BLOCK,
|
228 |
+
num_warps=num_warps,
|
229 |
+
num_stages=num_stages,
|
230 |
+
)
|
231 |
+
|
232 |
+
return x_q, x_s
|
build/torch25-cxx11-cu118-x86_64-linux/moe/fused_marlin_moe.py
CHANGED
@@ -40,7 +40,6 @@ def single_marlin_moe(
|
|
40 |
g_idx: Optional[torch.Tensor] = None,
|
41 |
sort_indices: Optional[torch.Tensor] = None,
|
42 |
w_zeros: Optional[torch.Tensor] = None,
|
43 |
-
override_config: Optional[Dict[str, Any]] = None,
|
44 |
num_bits: int = 8,
|
45 |
is_k_full: bool = True,
|
46 |
) -> torch.Tensor:
|
@@ -61,8 +60,6 @@ def single_marlin_moe(
|
|
61 |
- topk (int): The number of top-k experts to select.
|
62 |
- renormalize (bool): If True, renormalize the top-k weights to sum to 1.
|
63 |
- w_zeros (Optional[torch.Tensor]): Optional zero points to be used for w.
|
64 |
-
- override_config (Optional[Dict[str, Any]]): Optional override
|
65 |
-
for the kernel configuration.
|
66 |
- num_bits (bool): The number of bits in expert weights quantization.
|
67 |
|
68 |
Returns:
|
@@ -90,7 +87,6 @@ def single_marlin_moe(
|
|
90 |
w.shape,
|
91 |
topk_ids.shape[1],
|
92 |
None,
|
93 |
-
override_config=override_config,
|
94 |
is_marlin=True,
|
95 |
)
|
96 |
config = get_config_func(M)
|
@@ -154,6 +150,25 @@ def single_marlin_moe(
|
|
154 |
return torch.sum(intermediate_cache.view(*intermediate_cache.shape), dim=1)
|
155 |
|
156 |
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
157 |
def fused_marlin_moe(
|
158 |
hidden_states: torch.Tensor,
|
159 |
w1: torch.Tensor,
|
@@ -169,7 +184,6 @@ def fused_marlin_moe(
|
|
169 |
sort_indices2: Optional[torch.Tensor] = None,
|
170 |
w1_zeros: Optional[torch.Tensor] = None,
|
171 |
w2_zeros: Optional[torch.Tensor] = None,
|
172 |
-
override_config: Optional[Dict[str, Any]] = None,
|
173 |
num_bits: int = 8,
|
174 |
is_k_full: bool = True,
|
175 |
) -> torch.Tensor:
|
@@ -193,8 +207,6 @@ def fused_marlin_moe(
|
|
193 |
permutation.
|
194 |
- topk_weights (torch.Tensor): Top-k weights.
|
195 |
- topk_ids (torch.Tensor): Indices of topk-k elements.
|
196 |
-
- override_config (Optional[Dict[str, Any]]): Optional override
|
197 |
-
for the kernel configuration.
|
198 |
- w1_zeros (Optional[torch.Tensor]): Optional zero points to be used for w1.
|
199 |
- w2_zeros (Optional[torch.Tensor]): Optional zero points to be used for w2.
|
200 |
- num_bits (bool): The number of bits in expert weights quantization.
|
@@ -248,7 +260,6 @@ def fused_marlin_moe(
|
|
248 |
w2.shape,
|
249 |
topk_ids.shape[1],
|
250 |
None,
|
251 |
-
override_config=override_config,
|
252 |
is_marlin=True,
|
253 |
)
|
254 |
config = get_config_func(M)
|
@@ -350,6 +361,30 @@ def fused_marlin_moe(
|
|
350 |
return torch.sum(intermediate_cache3.view(*intermediate_cache3.shape), dim=1)
|
351 |
|
352 |
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
353 |
if hasattr(ops, "marlin_gemm_moe"):
|
354 |
|
355 |
@register_fake(add_op_namespace_prefix("marlin_gemm_moe"))
|
|
|
40 |
g_idx: Optional[torch.Tensor] = None,
|
41 |
sort_indices: Optional[torch.Tensor] = None,
|
42 |
w_zeros: Optional[torch.Tensor] = None,
|
|
|
43 |
num_bits: int = 8,
|
44 |
is_k_full: bool = True,
|
45 |
) -> torch.Tensor:
|
|
|
60 |
- topk (int): The number of top-k experts to select.
|
61 |
- renormalize (bool): If True, renormalize the top-k weights to sum to 1.
|
62 |
- w_zeros (Optional[torch.Tensor]): Optional zero points to be used for w.
|
|
|
|
|
63 |
- num_bits (bool): The number of bits in expert weights quantization.
|
64 |
|
65 |
Returns:
|
|
|
87 |
w.shape,
|
88 |
topk_ids.shape[1],
|
89 |
None,
|
|
|
90 |
is_marlin=True,
|
91 |
)
|
92 |
config = get_config_func(M)
|
|
|
150 |
return torch.sum(intermediate_cache.view(*intermediate_cache.shape), dim=1)
|
151 |
|
152 |
|
153 |
+
if hasattr(ops, "single_marlin_gemm_moe"):
|
154 |
+
|
155 |
+
@register_fake(add_op_namespace_prefix("single_marlin_gemm_moe"))
|
156 |
+
def single_marlin_moe_fake(
|
157 |
+
hidden_states: torch.Tensor,
|
158 |
+
w: torch.Tensor,
|
159 |
+
scales: torch.Tensor,
|
160 |
+
gating_output: torch.Tensor,
|
161 |
+
topk: int,
|
162 |
+
renormalize: bool,
|
163 |
+
g_idx: Optional[torch.Tensor] = None,
|
164 |
+
sort_indices: Optional[torch.Tensor] = None,
|
165 |
+
w_zeros: Optional[torch.Tensor] = None,
|
166 |
+
num_bits: int = 8,
|
167 |
+
is_k_full: bool = True,
|
168 |
+
) -> torch.Tensor:
|
169 |
+
return torch.empty_like(hidden_states)
|
170 |
+
|
171 |
+
|
172 |
def fused_marlin_moe(
|
173 |
hidden_states: torch.Tensor,
|
174 |
w1: torch.Tensor,
|
|
|
184 |
sort_indices2: Optional[torch.Tensor] = None,
|
185 |
w1_zeros: Optional[torch.Tensor] = None,
|
186 |
w2_zeros: Optional[torch.Tensor] = None,
|
|
|
187 |
num_bits: int = 8,
|
188 |
is_k_full: bool = True,
|
189 |
) -> torch.Tensor:
|
|
|
207 |
permutation.
|
208 |
- topk_weights (torch.Tensor): Top-k weights.
|
209 |
- topk_ids (torch.Tensor): Indices of topk-k elements.
|
|
|
|
|
210 |
- w1_zeros (Optional[torch.Tensor]): Optional zero points to be used for w1.
|
211 |
- w2_zeros (Optional[torch.Tensor]): Optional zero points to be used for w2.
|
212 |
- num_bits (bool): The number of bits in expert weights quantization.
|
|
|
260 |
w2.shape,
|
261 |
topk_ids.shape[1],
|
262 |
None,
|
|
|
263 |
is_marlin=True,
|
264 |
)
|
265 |
config = get_config_func(M)
|
|
|
361 |
return torch.sum(intermediate_cache3.view(*intermediate_cache3.shape), dim=1)
|
362 |
|
363 |
|
364 |
+
if hasattr(ops, "fused_marlin_moe"):
|
365 |
+
|
366 |
+
@register_fake(add_op_namespace_prefix("fused_marlin_moe"))
|
367 |
+
def fused_marlin_moe_fake(
|
368 |
+
hidden_states: torch.Tensor,
|
369 |
+
w1: torch.Tensor,
|
370 |
+
w2: torch.Tensor,
|
371 |
+
w1_scale: torch.Tensor,
|
372 |
+
w2_scale: torch.Tensor,
|
373 |
+
gating_output: torch.Tensor,
|
374 |
+
topk_weights: torch.Tensor,
|
375 |
+
topk_ids: torch.Tensor,
|
376 |
+
g_idx1: Optional[torch.Tensor] = None,
|
377 |
+
g_idx2: Optional[torch.Tensor] = None,
|
378 |
+
sort_indices1: Optional[torch.Tensor] = None,
|
379 |
+
sort_indices2: Optional[torch.Tensor] = None,
|
380 |
+
w1_zeros: Optional[torch.Tensor] = None,
|
381 |
+
w2_zeros: Optional[torch.Tensor] = None,
|
382 |
+
num_bits: int = 8,
|
383 |
+
is_k_full: bool = True,
|
384 |
+
) -> torch.Tensor:
|
385 |
+
return torch.empty_like(hidden_states)
|
386 |
+
|
387 |
+
|
388 |
if hasattr(ops, "marlin_gemm_moe"):
|
389 |
|
390 |
@register_fake(add_op_namespace_prefix("marlin_gemm_moe"))
|
build/torch25-cxx11-cu118-x86_64-linux/moe/fused_moe.py
CHANGED
@@ -1,21 +1,242 @@
|
|
|
|
1 |
"""Fused MoE kernel."""
|
2 |
|
3 |
import functools
|
4 |
import json
|
|
|
5 |
import os
|
6 |
-
from typing import Any, Callable, Dict, Optional, Tuple
|
7 |
|
8 |
import torch
|
9 |
import triton
|
10 |
import triton.language as tl
|
11 |
|
|
|
12 |
from ._ops import ops
|
13 |
-
from .fp8 import scaled_fp8_quant
|
14 |
from .platforms import current_platform
|
15 |
|
|
|
|
|
|
|
16 |
VLLM_FUSED_MOE_CHUNK_SIZE = int(os.getenv("VLLM_FUSED_MOE_CHUNK_SIZE", "32768"))
|
17 |
|
18 |
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
19 |
@triton.jit
|
20 |
def fused_moe_kernel(
|
21 |
# Pointers to matrices
|
@@ -44,8 +265,14 @@ def fused_moe_kernel(
|
|
44 |
stride_bn,
|
45 |
stride_cm,
|
46 |
stride_cn,
|
|
|
|
|
47 |
stride_bse,
|
|
|
48 |
stride_bsn,
|
|
|
|
|
|
|
49 |
# Meta-parameters
|
50 |
BLOCK_SIZE_M: tl.constexpr,
|
51 |
BLOCK_SIZE_N: tl.constexpr,
|
@@ -105,17 +332,17 @@ def fused_moe_kernel(
|
|
105 |
num_tokens_post_padded = tl.load(num_tokens_post_padded_ptr)
|
106 |
if pid_m * BLOCK_SIZE_M >= num_tokens_post_padded:
|
107 |
return
|
108 |
-
offs_token_id = pid_m * BLOCK_SIZE_M + tl.arange(0, BLOCK_SIZE_M)
|
109 |
offs_token = tl.load(sorted_token_ids_ptr + offs_token_id)
|
110 |
token_mask = offs_token < num_valid_tokens
|
111 |
|
112 |
-
offs_bn = (pid_n * BLOCK_SIZE_N + tl.arange(0, BLOCK_SIZE_N)) % N
|
113 |
offs_k = tl.arange(0, BLOCK_SIZE_K)
|
114 |
a_ptrs = a_ptr + (
|
115 |
offs_token[:, None] // top_k * stride_am + offs_k[None, :] * stride_ak
|
116 |
)
|
117 |
|
118 |
-
off_experts = tl.load(expert_ids_ptr + pid_m)
|
119 |
b_ptrs = (
|
120 |
b_ptr
|
121 |
+ off_experts * stride_be
|
@@ -128,8 +355,15 @@ def fused_moe_kernel(
|
|
128 |
b_scale = tl.load(b_scale_ptrs)
|
129 |
|
130 |
if use_fp8_w8a8:
|
131 |
-
|
132 |
-
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
133 |
|
134 |
# -----------------------------------------------------------
|
135 |
# Iterate to compute a block of the C matrix.
|
@@ -151,7 +385,17 @@ def fused_moe_kernel(
|
|
151 |
if use_int8_w8a16:
|
152 |
accumulator = tl.dot(a, b.to(compute_type), acc=accumulator)
|
153 |
elif use_fp8_w8a8:
|
154 |
-
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
155 |
else:
|
156 |
accumulator += tl.dot(a, b)
|
157 |
# Advance the ptrs to the next K block.
|
@@ -164,7 +408,10 @@ def fused_moe_kernel(
|
|
164 |
if use_int8_w8a16:
|
165 |
accumulator = (accumulator * b_scale).to(compute_type)
|
166 |
elif use_fp8_w8a8:
|
167 |
-
|
|
|
|
|
|
|
168 |
else:
|
169 |
accumulator = accumulator.to(compute_type)
|
170 |
# -----------------------------------------------------------
|
@@ -175,6 +422,141 @@ def fused_moe_kernel(
|
|
175 |
tl.store(c_ptrs, accumulator, mask=c_mask)
|
176 |
|
177 |
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
178 |
def moe_align_block_size(
|
179 |
topk_ids: torch.Tensor, block_size: int, num_experts: int
|
180 |
) -> Tuple[torch.Tensor, torch.Tensor, torch.Tensor]:
|
@@ -225,9 +607,34 @@ def moe_align_block_size(
|
|
225 |
(max_num_m_blocks,), dtype=torch.int32, device=topk_ids.device
|
226 |
)
|
227 |
num_tokens_post_pad = torch.empty((1), dtype=torch.int32, device=topk_ids.device)
|
228 |
-
|
229 |
-
|
230 |
-
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
231 |
return sorted_ids, expert_ids, num_tokens_post_pad
|
232 |
|
233 |
|
@@ -237,6 +644,7 @@ def invoke_fused_moe_kernel(
|
|
237 |
C: torch.Tensor,
|
238 |
A_scale: Optional[torch.Tensor],
|
239 |
B_scale: Optional[torch.Tensor],
|
|
|
240 |
topk_weights: torch.Tensor,
|
241 |
topk_ids: torch.Tensor,
|
242 |
sorted_token_ids: torch.Tensor,
|
@@ -248,64 +656,147 @@ def invoke_fused_moe_kernel(
|
|
248 |
compute_type: tl.dtype,
|
249 |
use_fp8_w8a8: bool,
|
250 |
use_int8_w8a16: bool,
|
|
|
|
|
251 |
) -> None:
|
252 |
assert topk_weights.stride(1) == 1
|
253 |
assert sorted_token_ids.stride(0) == 1
|
254 |
|
255 |
if use_fp8_w8a8:
|
256 |
-
A, A_scale = scaled_fp8_quant(A, A_scale)
|
257 |
assert B_scale is not None
|
258 |
-
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
259 |
assert B_scale is not None
|
|
|
260 |
else:
|
261 |
assert A_scale is None
|
262 |
assert B_scale is None
|
263 |
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
264 |
grid = lambda META: (
|
265 |
-
triton.cdiv(
|
266 |
* triton.cdiv(B.shape[1], META["BLOCK_SIZE_N"]),
|
267 |
)
|
268 |
|
269 |
-
|
270 |
-
|
271 |
-
|
272 |
-
|
273 |
-
|
274 |
-
B_scale
|
275 |
-
|
276 |
-
|
277 |
-
|
278 |
-
|
279 |
-
|
280 |
-
|
281 |
-
|
282 |
-
|
283 |
-
|
284 |
-
|
285 |
-
|
286 |
-
|
287 |
-
|
288 |
-
|
289 |
-
|
290 |
-
|
291 |
-
|
292 |
-
|
293 |
-
|
294 |
-
|
295 |
-
|
296 |
-
|
297 |
-
|
298 |
-
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
299 |
|
300 |
|
301 |
-
|
|
|
|
|
|
|
302 |
device_name = current_platform.get_device_name().replace(" ", "_")
|
303 |
dtype_selector = "" if not dtype else f",dtype={dtype}"
|
304 |
-
|
|
|
|
|
|
|
305 |
|
306 |
|
|
|
307 |
@functools.lru_cache
|
308 |
-
def get_moe_configs(
|
|
|
|
|
|
|
|
|
|
|
|
|
309 |
"""
|
310 |
Return optimized configurations for the fused MoE kernel.
|
311 |
|
@@ -317,18 +808,27 @@ def get_moe_configs(E: int, N: int, dtype: Optional[str]) -> Optional[Dict[int,
|
|
317 |
|
318 |
# First look up if an optimized configuration is available in the configs
|
319 |
# directory
|
320 |
-
|
|
|
321 |
|
322 |
config_file_path = os.path.join(
|
323 |
os.path.dirname(os.path.realpath(__file__)), "configs", json_file_name
|
324 |
)
|
325 |
if os.path.exists(config_file_path):
|
326 |
with open(config_file_path) as f:
|
|
|
327 |
# If a configuration has been found, return it
|
328 |
return {int(key): val for key, val in json.load(f).items()}
|
329 |
|
330 |
# If no optimized configuration is available, we will use the default
|
331 |
# configuration
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
332 |
return None
|
333 |
|
334 |
|
@@ -340,21 +840,34 @@ def get_default_config(
|
|
340 |
topk: int,
|
341 |
dtype: Optional[str],
|
342 |
is_marlin: bool,
|
|
|
343 |
) -> Dict[str, int]:
|
344 |
-
|
345 |
-
|
346 |
-
|
347 |
-
"BLOCK_SIZE_K": 32,
|
348 |
-
"GROUP_SIZE_M": 8,
|
349 |
-
}
|
350 |
-
# A heuristic: fused marlin works faster with this config for small M
|
351 |
-
if M <= E or (is_marlin and M <= 32):
|
352 |
config = {
|
353 |
-
"BLOCK_SIZE_M":
|
354 |
-
"BLOCK_SIZE_N":
|
355 |
-
"BLOCK_SIZE_K":
|
356 |
-
"GROUP_SIZE_M":
|
|
|
|
|
357 |
}
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
358 |
return config
|
359 |
|
360 |
|
@@ -364,15 +877,21 @@ def try_get_optimal_moe_config(
|
|
364 |
top_k: int,
|
365 |
dtype: Optional[str],
|
366 |
M: int,
|
367 |
-
override_config: Optional[Dict[str, Any]] = None,
|
368 |
is_marlin: bool = False,
|
|
|
369 |
):
|
|
|
|
|
|
|
|
|
370 |
if override_config:
|
371 |
config = override_config
|
372 |
else:
|
373 |
# First try to load optimal config from the file
|
374 |
E, _, N = w2_shape
|
375 |
-
|
|
|
|
|
376 |
|
377 |
if configs:
|
378 |
# If an optimal configuration map has been found, look up the
|
@@ -380,7 +899,9 @@ def try_get_optimal_moe_config(
|
|
380 |
config = configs[min(configs.keys(), key=lambda x: abs(x - M))]
|
381 |
else:
|
382 |
# Else use the default config
|
383 |
-
config = get_default_config(
|
|
|
|
|
384 |
return config
|
385 |
|
386 |
|
@@ -416,7 +937,8 @@ def fused_topk(
|
|
416 |
return topk_weights, topk_ids
|
417 |
|
418 |
|
419 |
-
# This is used by the Deepseek-V2 model
|
|
|
420 |
def grouped_topk(
|
421 |
hidden_states: torch.Tensor,
|
422 |
gating_output: torch.Tensor,
|
@@ -424,11 +946,25 @@ def grouped_topk(
|
|
424 |
renormalize: bool,
|
425 |
num_expert_group: int = 0,
|
426 |
topk_group: int = 0,
|
|
|
|
|
427 |
):
|
428 |
|
429 |
assert hidden_states.shape[0] == gating_output.shape[0], "Number of tokens mismatch"
|
430 |
|
431 |
-
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
432 |
num_token = scores.shape[0]
|
433 |
group_scores = (
|
434 |
scores.view(num_token, num_expert_group, -1).max(dim=-1).values
|
@@ -444,7 +980,13 @@ def grouped_topk(
|
|
444 |
.reshape(num_token, -1)
|
445 |
) # [n, e]
|
446 |
tmp_scores = scores.masked_fill(~score_mask.bool(), 0.0) # [n, e]
|
447 |
-
|
|
|
|
|
|
|
|
|
|
|
|
|
448 |
|
449 |
if renormalize:
|
450 |
topk_weights = topk_weights / topk_weights.sum(dim=-1, keepdim=True)
|
@@ -454,6 +996,7 @@ def grouped_topk(
|
|
454 |
|
455 |
def get_config_dtype_str(
|
456 |
dtype: torch.dtype,
|
|
|
457 |
use_int8_w8a16: Optional[bool] = False,
|
458 |
use_fp8_w8a8: Optional[bool] = False,
|
459 |
):
|
@@ -461,6 +1004,8 @@ def get_config_dtype_str(
|
|
461 |
return "fp8_w8a8"
|
462 |
elif use_int8_w8a16:
|
463 |
return "int8_w8a16"
|
|
|
|
|
464 |
elif dtype == torch.float:
|
465 |
# avoiding cases where kernel fails when float32 MoE
|
466 |
# use fp16/bfloat16 configs
|
@@ -468,6 +1013,80 @@ def get_config_dtype_str(
|
|
468 |
return None
|
469 |
|
470 |
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
471 |
def fused_experts(
|
472 |
hidden_states: torch.Tensor,
|
473 |
w1: torch.Tensor,
|
@@ -475,16 +1094,80 @@ def fused_experts(
|
|
475 |
topk_weights: torch.Tensor,
|
476 |
topk_ids: torch.Tensor,
|
477 |
inplace: bool = False,
|
478 |
-
override_config: Optional[Dict[str, Any]] = None,
|
479 |
use_fp8_w8a8: bool = False,
|
480 |
use_int8_w8a16: bool = False,
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
481 |
w1_scale: Optional[torch.Tensor] = None,
|
482 |
w2_scale: Optional[torch.Tensor] = None,
|
|
|
|
|
483 |
a1_scale: Optional[torch.Tensor] = None,
|
484 |
a2_scale: Optional[torch.Tensor] = None,
|
|
|
485 |
):
|
486 |
# Check constraints.
|
487 |
-
|
|
|
|
|
|
|
|
|
488 |
assert topk_weights.shape == topk_ids.shape, "topk shape mismatch"
|
489 |
assert hidden_states.is_contiguous(), "Hidden_states must be contiguous"
|
490 |
assert w1.is_contiguous(), "Expert weights1 must be contiguous"
|
@@ -500,6 +1183,7 @@ def fused_experts(
|
|
500 |
config_dtype = get_config_dtype_str(
|
501 |
use_fp8_w8a8=use_fp8_w8a8,
|
502 |
use_int8_w8a16=use_int8_w8a16,
|
|
|
503 |
dtype=hidden_states.dtype,
|
504 |
)
|
505 |
|
@@ -509,7 +1193,7 @@ def fused_experts(
|
|
509 |
w2.shape,
|
510 |
topk_ids.shape[1],
|
511 |
config_dtype,
|
512 |
-
|
513 |
)
|
514 |
|
515 |
config = get_config_func(M)
|
@@ -530,7 +1214,14 @@ def fused_experts(
|
|
530 |
dtype=hidden_states.dtype,
|
531 |
)
|
532 |
|
533 |
-
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
534 |
|
535 |
if inplace:
|
536 |
out_hidden_states = hidden_states
|
@@ -571,6 +1262,7 @@ def fused_experts(
|
|
571 |
intermediate_cache1,
|
572 |
a1_scale,
|
573 |
w1_scale,
|
|
|
574 |
curr_topk_weights,
|
575 |
curr_topk_ids,
|
576 |
sorted_token_ids,
|
@@ -582,6 +1274,8 @@ def fused_experts(
|
|
582 |
compute_type=compute_type,
|
583 |
use_fp8_w8a8=use_fp8_w8a8,
|
584 |
use_int8_w8a16=use_int8_w8a16,
|
|
|
|
|
585 |
)
|
586 |
|
587 |
ops.silu_and_mul(intermediate_cache2, intermediate_cache1.view(-1, N))
|
@@ -592,6 +1286,7 @@ def fused_experts(
|
|
592 |
intermediate_cache3,
|
593 |
a2_scale,
|
594 |
w2_scale,
|
|
|
595 |
curr_topk_weights,
|
596 |
curr_topk_ids,
|
597 |
sorted_token_ids,
|
@@ -603,6 +1298,8 @@ def fused_experts(
|
|
603 |
compute_type=compute_type,
|
604 |
use_fp8_w8a8=use_fp8_w8a8,
|
605 |
use_int8_w8a16=use_int8_w8a16,
|
|
|
|
|
606 |
)
|
607 |
|
608 |
ops.moe_sum(
|
@@ -620,17 +1317,20 @@ def fused_moe(
|
|
620 |
topk: int,
|
621 |
renormalize: bool,
|
622 |
inplace: bool = False,
|
623 |
-
override_config: Optional[Dict[str, Any]] = None,
|
624 |
use_grouped_topk: bool = False,
|
625 |
num_expert_group: Optional[int] = None,
|
626 |
topk_group: Optional[int] = None,
|
627 |
custom_routing_function: Optional[Callable] = None,
|
628 |
use_fp8_w8a8: bool = False,
|
629 |
use_int8_w8a16: bool = False,
|
|
|
630 |
w1_scale: Optional[torch.Tensor] = None,
|
631 |
w2_scale: Optional[torch.Tensor] = None,
|
|
|
|
|
632 |
a1_scale: Optional[torch.Tensor] = None,
|
633 |
a2_scale: Optional[torch.Tensor] = None,
|
|
|
634 |
) -> torch.Tensor:
|
635 |
"""
|
636 |
This function computes a Mixture of Experts (MoE) layer using two sets of
|
@@ -646,20 +1346,28 @@ def fused_moe(
|
|
646 |
- renormalize (bool): If True, renormalize the top-k weights to sum to 1.
|
647 |
- inplace (bool): If True, perform the operation in-place.
|
648 |
Defaults to False.
|
649 |
-
- override_config (Optional[Dict[str, Any]]): Optional override
|
650 |
-
for the kernel configuration.
|
651 |
- num_expert_group: Optional[int]: additional parameter for grouped_topk
|
652 |
- topk_group: Optional[int]: additional parameter for grouped_topk
|
653 |
- use_grouped_topk: If True, use grouped_topk instead of fused_topk
|
654 |
note: Deepseekv2 model uses grouped_topk
|
655 |
- use_fp8_w8a8 (bool): If True, use fp8 arithmetic to compute the inner
|
656 |
products for w1 and w2. Defaults to False.
|
657 |
-
- use_int8_w8a16 (bool): If True, use
|
658 |
-
products for w1 and w2.
|
|
|
|
|
|
|
|
|
659 |
- w1_scale (Optional[torch.Tensor]): Optional scale to be used for
|
660 |
w1.
|
661 |
- w2_scale (Optional[torch.Tensor]): Optional scale to be used for
|
662 |
w2.
|
|
|
|
|
|
|
|
|
|
|
|
|
663 |
|
664 |
Returns:
|
665 |
- torch.Tensor: The output tensor after applying the MoE layer.
|
@@ -693,11 +1401,14 @@ def fused_moe(
|
|
693 |
topk_weights,
|
694 |
topk_ids,
|
695 |
inplace=inplace,
|
696 |
-
override_config=override_config,
|
697 |
use_fp8_w8a8=use_fp8_w8a8,
|
698 |
use_int8_w8a16=use_int8_w8a16,
|
|
|
699 |
w1_scale=w1_scale,
|
700 |
w2_scale=w2_scale,
|
|
|
|
|
701 |
a1_scale=a1_scale,
|
702 |
a2_scale=a2_scale,
|
|
|
703 |
)
|
|
|
1 |
+
# SPDX-License-Identifier: Apache-2.0
|
2 |
"""Fused MoE kernel."""
|
3 |
|
4 |
import functools
|
5 |
import json
|
6 |
+
import logging
|
7 |
import os
|
8 |
+
from typing import Any, Callable, Dict, List, Optional, Tuple
|
9 |
|
10 |
import torch
|
11 |
import triton
|
12 |
import triton.language as tl
|
13 |
|
14 |
+
|
15 |
from ._ops import ops
|
16 |
+
from .fp8 import per_token_group_quant_fp8, scaled_fp8_quant
|
17 |
from .platforms import current_platform
|
18 |
|
19 |
+
logger = logging.getLogger(__name__)
|
20 |
+
|
21 |
+
|
22 |
VLLM_FUSED_MOE_CHUNK_SIZE = int(os.getenv("VLLM_FUSED_MOE_CHUNK_SIZE", "32768"))
|
23 |
|
24 |
|
25 |
+
@triton.jit
|
26 |
+
def fused_moe_kernel_gptq_awq(
|
27 |
+
# Pointers to matrices
|
28 |
+
a_ptr,
|
29 |
+
b_ptr,
|
30 |
+
c_ptr,
|
31 |
+
b_scale_ptr,
|
32 |
+
b_zp_ptr,
|
33 |
+
topk_weights_ptr,
|
34 |
+
sorted_token_ids_ptr,
|
35 |
+
expert_ids_ptr,
|
36 |
+
num_tokens_post_padded_ptr,
|
37 |
+
# Matrix dimensions
|
38 |
+
N: tl.constexpr,
|
39 |
+
K: tl.constexpr,
|
40 |
+
EM,
|
41 |
+
num_valid_tokens,
|
42 |
+
# The stride variables represent how much to increase the ptr by when
|
43 |
+
# moving by 1 element in a particular dimension. E.g. `stride_am` is
|
44 |
+
# how much to increase `a_ptr` by to get the element one row down
|
45 |
+
# (A has M rows).
|
46 |
+
stride_am,
|
47 |
+
stride_ak,
|
48 |
+
stride_be,
|
49 |
+
stride_bk,
|
50 |
+
stride_bn,
|
51 |
+
stride_cm,
|
52 |
+
stride_cn,
|
53 |
+
stride_bse,
|
54 |
+
stride_bsk,
|
55 |
+
stride_bsn,
|
56 |
+
stride_bze,
|
57 |
+
stride_bzk,
|
58 |
+
stride_bzn,
|
59 |
+
block_k_diviable: tl.constexpr,
|
60 |
+
group_size: tl.constexpr,
|
61 |
+
# Meta-parameters
|
62 |
+
BLOCK_SIZE_M: tl.constexpr,
|
63 |
+
BLOCK_SIZE_N: tl.constexpr,
|
64 |
+
BLOCK_SIZE_K: tl.constexpr,
|
65 |
+
GROUP_SIZE_M: tl.constexpr,
|
66 |
+
MUL_ROUTED_WEIGHT: tl.constexpr,
|
67 |
+
top_k: tl.constexpr,
|
68 |
+
compute_type: tl.constexpr,
|
69 |
+
has_zp: tl.constexpr,
|
70 |
+
use_int4_w4a16: tl.constexpr,
|
71 |
+
use_int8_w8a16: tl.constexpr,
|
72 |
+
):
|
73 |
+
"""
|
74 |
+
Implements the fused computation for a Mixture of Experts (MOE) using
|
75 |
+
token and expert matrices.
|
76 |
+
|
77 |
+
Key Parameters:
|
78 |
+
- A: The input tensor representing tokens with shape (*, K), where '*' can
|
79 |
+
be any shape representing batches and K is the feature dimension of
|
80 |
+
each token.
|
81 |
+
- B: The stacked MOE weight tensor with shape (E, N, K), where E is
|
82 |
+
the number of experts, K is the input feature dimension, and N is
|
83 |
+
the output feature dimension.
|
84 |
+
- C: The output cache tensor with shape (M, topk, N), where M is the
|
85 |
+
total number of tokens post padding, topk is the number of times
|
86 |
+
each token is repeated, and N is the output feature dimension.
|
87 |
+
- sorted_token_ids: A tensor containing the sorted indices of tokens,
|
88 |
+
repeated topk times and arranged by the expert index they are
|
89 |
+
assigned to.
|
90 |
+
- expert_ids: A tensor containing the indices of the expert for each
|
91 |
+
block. It determines which expert matrix from B should be used for
|
92 |
+
each block in A.
|
93 |
+
This kernel performs the multiplication of a token by its corresponding
|
94 |
+
expert matrix as determined by `expert_ids`. The sorting of
|
95 |
+
`sorted_token_ids` by expert index and padding ensures divisibility by
|
96 |
+
BLOCK_SIZE_M, which is necessary to maintain consistency in block matrix
|
97 |
+
multiplication across different blocks processed by the same expert.
|
98 |
+
"""
|
99 |
+
# -----------------------------------------------------------
|
100 |
+
# Map program ids `pid` to the block of C it should compute.
|
101 |
+
# This is done in a grouped ordering to promote L2 data reuse.
|
102 |
+
pid = tl.program_id(axis=0)
|
103 |
+
num_pid_m = tl.cdiv(EM, BLOCK_SIZE_M)
|
104 |
+
num_pid_n = tl.cdiv(N, BLOCK_SIZE_N)
|
105 |
+
num_pid_in_group = GROUP_SIZE_M * num_pid_n
|
106 |
+
group_id = pid // num_pid_in_group
|
107 |
+
first_pid_m = group_id * GROUP_SIZE_M
|
108 |
+
group_size_m = min(num_pid_m - first_pid_m, GROUP_SIZE_M)
|
109 |
+
pid_m = first_pid_m + ((pid % num_pid_in_group) % group_size_m)
|
110 |
+
pid_n = (pid % num_pid_in_group) // group_size_m
|
111 |
+
|
112 |
+
# ----------------------------------------------------------
|
113 |
+
# Create pointers for the first blocks of A and B.
|
114 |
+
# We will advance this pointer as we move in the K direction
|
115 |
+
# and accumulate
|
116 |
+
# `a_ptrs` is a block of [BLOCK_SIZE_M, BLOCK_SIZE_K] pointers
|
117 |
+
# `b_ptrs` is a block of [BLOCK_SIZE_K, BLOCK_SIZE_N] pointers
|
118 |
+
num_tokens_post_padded = tl.load(num_tokens_post_padded_ptr)
|
119 |
+
if pid_m * BLOCK_SIZE_M >= num_tokens_post_padded:
|
120 |
+
return
|
121 |
+
offs_token_id = pid_m * BLOCK_SIZE_M + tl.arange(0, BLOCK_SIZE_M).to(tl.int64)
|
122 |
+
offs_token = tl.load(sorted_token_ids_ptr + offs_token_id)
|
123 |
+
token_mask = offs_token < num_valid_tokens
|
124 |
+
|
125 |
+
offs_bn = (pid_n * BLOCK_SIZE_N + tl.arange(0, BLOCK_SIZE_N).to(tl.int64)) % N
|
126 |
+
offs_k = tl.arange(0, BLOCK_SIZE_K)
|
127 |
+
a_ptrs = a_ptr + (
|
128 |
+
offs_token[:, None] // top_k * stride_am + offs_k[None, :] * stride_ak
|
129 |
+
)
|
130 |
+
|
131 |
+
off_experts = tl.load(expert_ids_ptr + pid_m).to(tl.int64)
|
132 |
+
|
133 |
+
if use_int4_w4a16:
|
134 |
+
b_ptrs = (
|
135 |
+
b_ptr
|
136 |
+
+ off_experts * stride_be
|
137 |
+
+ (offs_k[:, None] // 2) * stride_bk
|
138 |
+
+ offs_bn[None, :] * stride_bn
|
139 |
+
)
|
140 |
+
b_shifter = (offs_k[:, None] % 2) * 4
|
141 |
+
elif use_int8_w8a16:
|
142 |
+
b_ptrs = (
|
143 |
+
b_ptr
|
144 |
+
+ off_experts * stride_be
|
145 |
+
+ offs_k[:, None] * stride_bk
|
146 |
+
+ offs_bn[None, :] * stride_bn
|
147 |
+
)
|
148 |
+
|
149 |
+
if not has_zp and use_int4_w4a16:
|
150 |
+
b_zp_num = 8
|
151 |
+
if not has_zp and use_int8_w8a16:
|
152 |
+
b_zp_num = 128
|
153 |
+
elif has_zp and use_int4_w4a16:
|
154 |
+
b_zp_shifter = (offs_bn[None, :] % 2) * 4
|
155 |
+
|
156 |
+
# -----------------------------------------------------------
|
157 |
+
# Iterate to compute a block of the C matrix.
|
158 |
+
# We accumulate into a `[BLOCK_SIZE_M, BLOCK_SIZE_N]` block
|
159 |
+
# of fp32 values for higher accuracy.
|
160 |
+
# `accumulator` will be converted back to fp16 after the loop.
|
161 |
+
accumulator = tl.zeros((BLOCK_SIZE_M, BLOCK_SIZE_N), dtype=tl.float32)
|
162 |
+
for k in range(0, tl.cdiv(K, BLOCK_SIZE_K)):
|
163 |
+
# Load the next block of A and B, generate a mask by checking the
|
164 |
+
# K dimension.
|
165 |
+
|
166 |
+
if not block_k_diviable:
|
167 |
+
k_mask = offs_k[:, None] < K - k * BLOCK_SIZE_K
|
168 |
+
k_other = 0.0
|
169 |
+
else:
|
170 |
+
k_mask = None
|
171 |
+
k_other = None
|
172 |
+
|
173 |
+
a = tl.load(
|
174 |
+
a_ptrs,
|
175 |
+
mask=token_mask[:, None] & (offs_k[None, :] < K - k * BLOCK_SIZE_K),
|
176 |
+
other=0.0,
|
177 |
+
)
|
178 |
+
b = tl.load(b_ptrs)
|
179 |
+
if use_int4_w4a16:
|
180 |
+
b = (b >> b_shifter) & 0xF
|
181 |
+
|
182 |
+
b_scale_ptrs = (
|
183 |
+
b_scale_ptr
|
184 |
+
+ off_experts * stride_bse
|
185 |
+
+ offs_bn[None, :] * stride_bsn
|
186 |
+
+ ((offs_k[:, None] + BLOCK_SIZE_K * k) // group_size) * stride_bsk
|
187 |
+
)
|
188 |
+
b_scale = tl.load(b_scale_ptrs, mask=k_mask, other=k_other)
|
189 |
+
b_scale = b_scale.to(tl.float32)
|
190 |
+
|
191 |
+
if has_zp and use_int4_w4a16:
|
192 |
+
offs_k_true = (offs_k[:, None] + BLOCK_SIZE_K * k) // group_size
|
193 |
+
b_zp_ptrs = (
|
194 |
+
b_zp_ptr
|
195 |
+
+ off_experts * stride_bze
|
196 |
+
+ (offs_bn[None, :] // 2) * stride_bzn
|
197 |
+
+ offs_k_true * stride_bzk
|
198 |
+
)
|
199 |
+
b_zp = tl.load(b_zp_ptrs, mask=k_mask, other=k_other)
|
200 |
+
b_zp = (b_zp >> b_zp_shifter) & 0xF
|
201 |
+
b_zp = b_zp.to(tl.float32)
|
202 |
+
elif has_zp and use_int8_w8a16:
|
203 |
+
offs_k_true = (offs_k[:, None] + BLOCK_SIZE_K * k) // group_size
|
204 |
+
b_zp_ptrs = (
|
205 |
+
b_zp_ptr
|
206 |
+
+ off_experts * stride_bze
|
207 |
+
+ offs_bn[None, :] * stride_bzn
|
208 |
+
+ offs_k_true * stride_bzk
|
209 |
+
)
|
210 |
+
b_zp = tl.load(b_zp_ptrs, mask=k_mask, other=k_other)
|
211 |
+
b_zp = b_zp.to(tl.float32)
|
212 |
+
|
213 |
+
# We accumulate along the K dimension.
|
214 |
+
if has_zp:
|
215 |
+
b = ((b.to(tl.float32) - b_zp) * b_scale).to(compute_type)
|
216 |
+
else:
|
217 |
+
b = ((b.to(tl.float32) - b_zp_num) * b_scale).to(compute_type)
|
218 |
+
accumulator = tl.dot(a, b, acc=accumulator)
|
219 |
+
|
220 |
+
# Advance the ptrs to the next K block.
|
221 |
+
a_ptrs += BLOCK_SIZE_K * stride_ak
|
222 |
+
if use_int4_w4a16:
|
223 |
+
b_ptrs += (BLOCK_SIZE_K // 2) * stride_bk
|
224 |
+
else:
|
225 |
+
b_ptrs += BLOCK_SIZE_K * stride_bk
|
226 |
+
|
227 |
+
if MUL_ROUTED_WEIGHT:
|
228 |
+
moe_weight = tl.load(topk_weights_ptr + offs_token, mask=token_mask, other=0)
|
229 |
+
accumulator = accumulator * moe_weight[:, None]
|
230 |
+
|
231 |
+
accumulator = accumulator.to(compute_type)
|
232 |
+
# -----------------------------------------------------------
|
233 |
+
# Write back the block of the output
|
234 |
+
offs_cn = pid_n * BLOCK_SIZE_N + tl.arange(0, BLOCK_SIZE_N)
|
235 |
+
c_ptrs = c_ptr + stride_cm * offs_token[:, None] + stride_cn * offs_cn[None, :]
|
236 |
+
c_mask = token_mask[:, None] & (offs_cn[None, :] < N)
|
237 |
+
tl.store(c_ptrs, accumulator, mask=c_mask)
|
238 |
+
|
239 |
+
|
240 |
@triton.jit
|
241 |
def fused_moe_kernel(
|
242 |
# Pointers to matrices
|
|
|
265 |
stride_bn,
|
266 |
stride_cm,
|
267 |
stride_cn,
|
268 |
+
stride_asm,
|
269 |
+
stride_ask,
|
270 |
stride_bse,
|
271 |
+
stride_bsk,
|
272 |
stride_bsn,
|
273 |
+
# Block size for block-wise quantization
|
274 |
+
group_n: tl.constexpr,
|
275 |
+
group_k: tl.constexpr,
|
276 |
# Meta-parameters
|
277 |
BLOCK_SIZE_M: tl.constexpr,
|
278 |
BLOCK_SIZE_N: tl.constexpr,
|
|
|
332 |
num_tokens_post_padded = tl.load(num_tokens_post_padded_ptr)
|
333 |
if pid_m * BLOCK_SIZE_M >= num_tokens_post_padded:
|
334 |
return
|
335 |
+
offs_token_id = pid_m * BLOCK_SIZE_M + tl.arange(0, BLOCK_SIZE_M).to(tl.int64)
|
336 |
offs_token = tl.load(sorted_token_ids_ptr + offs_token_id)
|
337 |
token_mask = offs_token < num_valid_tokens
|
338 |
|
339 |
+
offs_bn = (pid_n * BLOCK_SIZE_N + tl.arange(0, BLOCK_SIZE_N).to(tl.int64)) % N
|
340 |
offs_k = tl.arange(0, BLOCK_SIZE_K)
|
341 |
a_ptrs = a_ptr + (
|
342 |
offs_token[:, None] // top_k * stride_am + offs_k[None, :] * stride_ak
|
343 |
)
|
344 |
|
345 |
+
off_experts = tl.load(expert_ids_ptr + pid_m).to(tl.int64)
|
346 |
b_ptrs = (
|
347 |
b_ptr
|
348 |
+ off_experts * stride_be
|
|
|
355 |
b_scale = tl.load(b_scale_ptrs)
|
356 |
|
357 |
if use_fp8_w8a8:
|
358 |
+
if group_k > 0 and group_n > 0:
|
359 |
+
a_scale_ptrs = a_scale_ptr + (offs_token // top_k) * stride_asm
|
360 |
+
offs_bsn = offs_bn // group_n
|
361 |
+
b_scale_ptrs = (
|
362 |
+
b_scale_ptr + off_experts * stride_bse + offs_bsn * stride_bsn
|
363 |
+
)
|
364 |
+
else:
|
365 |
+
a_scale = tl.load(a_scale_ptr)
|
366 |
+
b_scale = tl.load(b_scale_ptr + off_experts)
|
367 |
|
368 |
# -----------------------------------------------------------
|
369 |
# Iterate to compute a block of the C matrix.
|
|
|
385 |
if use_int8_w8a16:
|
386 |
accumulator = tl.dot(a, b.to(compute_type), acc=accumulator)
|
387 |
elif use_fp8_w8a8:
|
388 |
+
if group_k > 0 and group_n > 0:
|
389 |
+
k_start = k * BLOCK_SIZE_K
|
390 |
+
offs_ks = k_start // group_k
|
391 |
+
a_scale = tl.load(
|
392 |
+
a_scale_ptrs + offs_ks * stride_ask, mask=token_mask, other=0.0
|
393 |
+
)
|
394 |
+
b_scale = tl.load(b_scale_ptrs + offs_ks * stride_bsk)
|
395 |
+
|
396 |
+
accumulator += tl.dot(a, b) * a_scale[:, None] * b_scale[None, :]
|
397 |
+
else:
|
398 |
+
accumulator = tl.dot(a, b, acc=accumulator)
|
399 |
else:
|
400 |
accumulator += tl.dot(a, b)
|
401 |
# Advance the ptrs to the next K block.
|
|
|
408 |
if use_int8_w8a16:
|
409 |
accumulator = (accumulator * b_scale).to(compute_type)
|
410 |
elif use_fp8_w8a8:
|
411 |
+
if group_k > 0 and group_n > 0:
|
412 |
+
accumulator = accumulator.to(compute_type)
|
413 |
+
else:
|
414 |
+
accumulator = (accumulator * a_scale * b_scale).to(compute_type)
|
415 |
else:
|
416 |
accumulator = accumulator.to(compute_type)
|
417 |
# -----------------------------------------------------------
|
|
|
422 |
tl.store(c_ptrs, accumulator, mask=c_mask)
|
423 |
|
424 |
|
425 |
+
def ceil_div(a, b):
|
426 |
+
return (a + b - 1) // b
|
427 |
+
|
428 |
+
|
429 |
+
@triton.jit
|
430 |
+
def moe_align_block_size_stage1(
|
431 |
+
topk_ids_ptr,
|
432 |
+
tokens_cnts_ptr,
|
433 |
+
num_experts: tl.constexpr,
|
434 |
+
numel: tl.constexpr,
|
435 |
+
tokens_per_thread: tl.constexpr,
|
436 |
+
):
|
437 |
+
pid = tl.program_id(0)
|
438 |
+
|
439 |
+
start_idx = pid * tokens_per_thread
|
440 |
+
|
441 |
+
off_c = (pid + 1) * num_experts
|
442 |
+
|
443 |
+
for i in range(tokens_per_thread):
|
444 |
+
if start_idx + i < numel:
|
445 |
+
idx = tl.load(topk_ids_ptr + start_idx + i)
|
446 |
+
token_cnt = tl.load(tokens_cnts_ptr + off_c + idx)
|
447 |
+
tl.store(tokens_cnts_ptr + off_c + idx, token_cnt + 1)
|
448 |
+
|
449 |
+
|
450 |
+
@triton.jit
|
451 |
+
def moe_align_block_size_stage2(
|
452 |
+
tokens_cnts_ptr,
|
453 |
+
num_experts: tl.constexpr,
|
454 |
+
):
|
455 |
+
pid = tl.program_id(0)
|
456 |
+
|
457 |
+
last_cnt = 0
|
458 |
+
for i in range(1, num_experts + 1):
|
459 |
+
token_cnt = tl.load(tokens_cnts_ptr + i * num_experts + pid)
|
460 |
+
last_cnt = last_cnt + token_cnt
|
461 |
+
tl.store(tokens_cnts_ptr + i * num_experts + pid, last_cnt)
|
462 |
+
|
463 |
+
|
464 |
+
@triton.jit
|
465 |
+
def moe_align_block_size_stage3(
|
466 |
+
total_tokens_post_pad_ptr,
|
467 |
+
tokens_cnts_ptr,
|
468 |
+
cumsum_ptr,
|
469 |
+
num_experts: tl.constexpr,
|
470 |
+
block_size: tl.constexpr,
|
471 |
+
):
|
472 |
+
last_cumsum = 0
|
473 |
+
off_cnt = num_experts * num_experts
|
474 |
+
for i in range(1, num_experts + 1):
|
475 |
+
token_cnt = tl.load(tokens_cnts_ptr + off_cnt + i - 1)
|
476 |
+
last_cumsum = last_cumsum + tl.cdiv(token_cnt, block_size) * block_size
|
477 |
+
tl.store(cumsum_ptr + i, last_cumsum)
|
478 |
+
tl.store(total_tokens_post_pad_ptr, last_cumsum)
|
479 |
+
|
480 |
+
|
481 |
+
@triton.jit
|
482 |
+
def moe_align_block_size_stage4(
|
483 |
+
topk_ids_ptr,
|
484 |
+
sorted_token_ids_ptr,
|
485 |
+
expert_ids_ptr,
|
486 |
+
tokens_cnts_ptr,
|
487 |
+
cumsum_ptr,
|
488 |
+
num_experts: tl.constexpr,
|
489 |
+
block_size: tl.constexpr,
|
490 |
+
numel: tl.constexpr,
|
491 |
+
tokens_per_thread: tl.constexpr,
|
492 |
+
):
|
493 |
+
pid = tl.program_id(0)
|
494 |
+
start_idx = tl.load(cumsum_ptr + pid)
|
495 |
+
end_idx = tl.load(cumsum_ptr + pid + 1)
|
496 |
+
|
497 |
+
for i in range(start_idx, end_idx, block_size):
|
498 |
+
tl.store(expert_ids_ptr + i // block_size, pid)
|
499 |
+
|
500 |
+
start_idx = pid * tokens_per_thread
|
501 |
+
off_t = pid * num_experts
|
502 |
+
|
503 |
+
for i in range(start_idx, tl.minimum(start_idx + tokens_per_thread, numel)):
|
504 |
+
expert_id = tl.load(topk_ids_ptr + i)
|
505 |
+
token_cnt = tl.load(tokens_cnts_ptr + off_t + expert_id)
|
506 |
+
rank_post_pad = token_cnt + tl.load(cumsum_ptr + expert_id)
|
507 |
+
tl.store(sorted_token_ids_ptr + rank_post_pad, i)
|
508 |
+
tl.store(tokens_cnts_ptr + off_t + expert_id, token_cnt + 1)
|
509 |
+
|
510 |
+
|
511 |
+
# Triton implementation based on:
|
512 |
+
# https://github.com/sgl-project/sglang/commit/ba5112ff691d791a9e38c6c71f59324a5fcb49d0
|
513 |
+
def moe_align_block_size_triton(
|
514 |
+
topk_ids: torch.Tensor,
|
515 |
+
num_experts: int,
|
516 |
+
block_size: int,
|
517 |
+
sorted_token_ids: torch.Tensor,
|
518 |
+
expert_ids: torch.Tensor,
|
519 |
+
num_tokens_post_pad: torch.Tensor,
|
520 |
+
) -> None:
|
521 |
+
numel = topk_ids.numel()
|
522 |
+
grid = (num_experts,)
|
523 |
+
tokens_cnts = torch.zeros(
|
524 |
+
(num_experts + 1, num_experts), dtype=torch.int32, device=topk_ids.device
|
525 |
+
)
|
526 |
+
cumsum = torch.zeros((num_experts + 1,), dtype=torch.int32, device=topk_ids.device)
|
527 |
+
tokens_per_thread = ceil_div(numel, num_experts)
|
528 |
+
|
529 |
+
moe_align_block_size_stage1[grid](
|
530 |
+
topk_ids,
|
531 |
+
tokens_cnts,
|
532 |
+
num_experts,
|
533 |
+
numel,
|
534 |
+
tokens_per_thread,
|
535 |
+
)
|
536 |
+
moe_align_block_size_stage2[grid](
|
537 |
+
tokens_cnts,
|
538 |
+
num_experts,
|
539 |
+
)
|
540 |
+
moe_align_block_size_stage3[(1,)](
|
541 |
+
num_tokens_post_pad,
|
542 |
+
tokens_cnts,
|
543 |
+
cumsum,
|
544 |
+
num_experts,
|
545 |
+
block_size,
|
546 |
+
)
|
547 |
+
moe_align_block_size_stage4[grid](
|
548 |
+
topk_ids,
|
549 |
+
sorted_token_ids,
|
550 |
+
expert_ids,
|
551 |
+
tokens_cnts,
|
552 |
+
cumsum,
|
553 |
+
num_experts,
|
554 |
+
block_size,
|
555 |
+
numel,
|
556 |
+
tokens_per_thread,
|
557 |
+
)
|
558 |
+
|
559 |
+
|
560 |
def moe_align_block_size(
|
561 |
topk_ids: torch.Tensor, block_size: int, num_experts: int
|
562 |
) -> Tuple[torch.Tensor, torch.Tensor, torch.Tensor]:
|
|
|
607 |
(max_num_m_blocks,), dtype=torch.int32, device=topk_ids.device
|
608 |
)
|
609 |
num_tokens_post_pad = torch.empty((1), dtype=torch.int32, device=topk_ids.device)
|
610 |
+
if num_experts >= 224:
|
611 |
+
if VLLM_ENABLE_MOE_ALIGN_BLOCK_SIZE_TRITON:
|
612 |
+
moe_align_block_size_triton(
|
613 |
+
topk_ids,
|
614 |
+
num_experts,
|
615 |
+
block_size,
|
616 |
+
sorted_ids,
|
617 |
+
expert_ids,
|
618 |
+
num_tokens_post_pad,
|
619 |
+
)
|
620 |
+
else:
|
621 |
+
ops.sgl_moe_align_block_size(
|
622 |
+
topk_ids,
|
623 |
+
num_experts,
|
624 |
+
block_size,
|
625 |
+
sorted_ids,
|
626 |
+
expert_ids,
|
627 |
+
num_tokens_post_pad,
|
628 |
+
)
|
629 |
+
else:
|
630 |
+
ops.moe_align_block_size(
|
631 |
+
topk_ids,
|
632 |
+
num_experts,
|
633 |
+
block_size,
|
634 |
+
sorted_ids,
|
635 |
+
expert_ids,
|
636 |
+
num_tokens_post_pad,
|
637 |
+
)
|
638 |
return sorted_ids, expert_ids, num_tokens_post_pad
|
639 |
|
640 |
|
|
|
644 |
C: torch.Tensor,
|
645 |
A_scale: Optional[torch.Tensor],
|
646 |
B_scale: Optional[torch.Tensor],
|
647 |
+
B_zp: Optional[torch.Tensor],
|
648 |
topk_weights: torch.Tensor,
|
649 |
topk_ids: torch.Tensor,
|
650 |
sorted_token_ids: torch.Tensor,
|
|
|
656 |
compute_type: tl.dtype,
|
657 |
use_fp8_w8a8: bool,
|
658 |
use_int8_w8a16: bool,
|
659 |
+
use_int4_w4a16: bool,
|
660 |
+
block_shape: Optional[List[int]] = None,
|
661 |
) -> None:
|
662 |
assert topk_weights.stride(1) == 1
|
663 |
assert sorted_token_ids.stride(0) == 1
|
664 |
|
665 |
if use_fp8_w8a8:
|
|
|
666 |
assert B_scale is not None
|
667 |
+
if block_shape is None:
|
668 |
+
A, A_scale = scaled_fp8_quant(A, A_scale)
|
669 |
+
else:
|
670 |
+
assert len(block_shape) == 2
|
671 |
+
block_n, block_k = block_shape[0], block_shape[1]
|
672 |
+
A, A_scale = per_token_group_quant_fp8(A, block_k)
|
673 |
+
assert triton.cdiv(A.shape[-1], block_k) == A_scale.shape[-1]
|
674 |
+
assert triton.cdiv(B.shape[-2], block_n) == B_scale.shape[-2]
|
675 |
+
assert triton.cdiv(B.shape[-1], block_k) == B_scale.shape[-1]
|
676 |
+
elif use_int8_w8a16 or use_int4_w4a16:
|
677 |
assert B_scale is not None
|
678 |
+
assert block_shape is None or block_shape[0] == 0
|
679 |
else:
|
680 |
assert A_scale is None
|
681 |
assert B_scale is None
|
682 |
|
683 |
+
EM = sorted_token_ids.shape[0]
|
684 |
+
if A.shape[0] < config["BLOCK_SIZE_M"]:
|
685 |
+
# optimize for small batch_size.
|
686 |
+
# We assume that top_ids of each token is unique, so
|
687 |
+
# so num_valid_experts <= batch_size <= BLOCK_SIZE_M,
|
688 |
+
# and we can skip some invalid blocks.
|
689 |
+
EM = min(sorted_token_ids.shape[0], A.shape[0] * top_k * config["BLOCK_SIZE_M"])
|
690 |
grid = lambda META: (
|
691 |
+
triton.cdiv(EM, META["BLOCK_SIZE_M"])
|
692 |
* triton.cdiv(B.shape[1], META["BLOCK_SIZE_N"]),
|
693 |
)
|
694 |
|
695 |
+
if (
|
696 |
+
(use_int8_w8a16 or use_int4_w4a16)
|
697 |
+
and block_shape is not None
|
698 |
+
and block_shape[1] > 0
|
699 |
+
):
|
700 |
+
assert B_scale is not None and B_scale.ndim == 3
|
701 |
+
assert B_zp is None or B_zp.ndim == 3
|
702 |
+
|
703 |
+
fused_moe_kernel_gptq_awq[grid](
|
704 |
+
A,
|
705 |
+
B,
|
706 |
+
C,
|
707 |
+
B_scale,
|
708 |
+
B_zp,
|
709 |
+
topk_weights,
|
710 |
+
sorted_token_ids,
|
711 |
+
expert_ids,
|
712 |
+
num_tokens_post_padded,
|
713 |
+
B.shape[1],
|
714 |
+
A.shape[1],
|
715 |
+
EM,
|
716 |
+
topk_ids.numel(),
|
717 |
+
A.stride(0),
|
718 |
+
A.stride(1),
|
719 |
+
B.stride(0),
|
720 |
+
B.stride(2),
|
721 |
+
B.stride(1),
|
722 |
+
C.stride(1),
|
723 |
+
C.stride(2),
|
724 |
+
B_scale.stride(0),
|
725 |
+
B_scale.stride(2),
|
726 |
+
B_scale.stride(1),
|
727 |
+
B_zp.stride(0) if B_zp is not None else 0,
|
728 |
+
B_zp.stride(2) if B_zp is not None else 0,
|
729 |
+
B_zp.stride(1) if B_zp is not None else 0,
|
730 |
+
block_k_diviable=A.shape[1] % config["BLOCK_SIZE_K"] == 0,
|
731 |
+
group_size=block_shape[1],
|
732 |
+
MUL_ROUTED_WEIGHT=mul_routed_weight,
|
733 |
+
top_k=top_k,
|
734 |
+
compute_type=compute_type,
|
735 |
+
has_zp=B_zp is not None,
|
736 |
+
use_int4_w4a16=use_int4_w4a16,
|
737 |
+
use_int8_w8a16=use_int8_w8a16,
|
738 |
+
**config,
|
739 |
+
)
|
740 |
+
|
741 |
+
else:
|
742 |
+
fused_moe_kernel[grid](
|
743 |
+
A,
|
744 |
+
B,
|
745 |
+
C,
|
746 |
+
A_scale,
|
747 |
+
B_scale,
|
748 |
+
topk_weights,
|
749 |
+
sorted_token_ids,
|
750 |
+
expert_ids,
|
751 |
+
num_tokens_post_padded,
|
752 |
+
B.shape[1],
|
753 |
+
A.shape[1],
|
754 |
+
EM,
|
755 |
+
topk_ids.numel(),
|
756 |
+
A.stride(0),
|
757 |
+
A.stride(1),
|
758 |
+
B.stride(0),
|
759 |
+
B.stride(2),
|
760 |
+
B.stride(1),
|
761 |
+
C.stride(1),
|
762 |
+
C.stride(2),
|
763 |
+
A_scale.stride(0) if A_scale is not None and A_scale.ndim == 2 else 0,
|
764 |
+
A_scale.stride(1) if A_scale is not None and A_scale.ndim == 2 else 0,
|
765 |
+
B_scale.stride(0) if B_scale is not None and B_scale.ndim >= 2 else 0,
|
766 |
+
B_scale.stride(2) if B_scale is not None and B_scale.ndim == 3 else 0,
|
767 |
+
B_scale.stride(1) if B_scale is not None and B_scale.ndim >= 2 else 0,
|
768 |
+
0 if block_shape is None else block_shape[0],
|
769 |
+
0 if block_shape is None else block_shape[1],
|
770 |
+
MUL_ROUTED_WEIGHT=mul_routed_weight,
|
771 |
+
top_k=top_k,
|
772 |
+
compute_type=compute_type,
|
773 |
+
use_fp8_w8a8=use_fp8_w8a8,
|
774 |
+
use_int8_w8a16=use_int8_w8a16,
|
775 |
+
**config,
|
776 |
+
)
|
777 |
|
778 |
|
779 |
+
# Adapted from: https://github.com/sgl-project/sglang/pull/2628
|
780 |
+
def get_config_file_name(
|
781 |
+
E: int, N: int, dtype: Optional[str], block_shape: Optional[List[int]] = None
|
782 |
+
) -> str:
|
783 |
device_name = current_platform.get_device_name().replace(" ", "_")
|
784 |
dtype_selector = "" if not dtype else f",dtype={dtype}"
|
785 |
+
block_shape_selector = (
|
786 |
+
"" if not block_shape or not all(block_shape) else f",block_shape={block_shape}"
|
787 |
+
)
|
788 |
+
return f"E={E},N={N},device_name={device_name}{dtype_selector}{block_shape_selector}.json" # noqa: E501
|
789 |
|
790 |
|
791 |
+
# Adapted from: https://github.com/sgl-project/sglang/pull/2628
|
792 |
@functools.lru_cache
|
793 |
+
def get_moe_configs(
|
794 |
+
E: int,
|
795 |
+
N: int,
|
796 |
+
dtype: Optional[str],
|
797 |
+
block_n: Optional[int] = None,
|
798 |
+
block_k: Optional[int] = None,
|
799 |
+
) -> Optional[Dict[int, Any]]:
|
800 |
"""
|
801 |
Return optimized configurations for the fused MoE kernel.
|
802 |
|
|
|
808 |
|
809 |
# First look up if an optimized configuration is available in the configs
|
810 |
# directory
|
811 |
+
block_shape = [block_n, block_k] if block_n and block_k else None
|
812 |
+
json_file_name = get_config_file_name(E, N, dtype, block_shape)
|
813 |
|
814 |
config_file_path = os.path.join(
|
815 |
os.path.dirname(os.path.realpath(__file__)), "configs", json_file_name
|
816 |
)
|
817 |
if os.path.exists(config_file_path):
|
818 |
with open(config_file_path) as f:
|
819 |
+
logger.info("Using configuration from %s for MoE layer.", config_file_path)
|
820 |
# If a configuration has been found, return it
|
821 |
return {int(key): val for key, val in json.load(f).items()}
|
822 |
|
823 |
# If no optimized configuration is available, we will use the default
|
824 |
# configuration
|
825 |
+
logger.warning(
|
826 |
+
(
|
827 |
+
"Using default MoE config. Performance might be sub-optimal! "
|
828 |
+
"Config file not found at %s"
|
829 |
+
),
|
830 |
+
config_file_path,
|
831 |
+
)
|
832 |
return None
|
833 |
|
834 |
|
|
|
840 |
topk: int,
|
841 |
dtype: Optional[str],
|
842 |
is_marlin: bool,
|
843 |
+
block_shape: Optional[List[int]] = None,
|
844 |
) -> Dict[str, int]:
|
845 |
+
if dtype == "fp8_w8a8" and block_shape is not None:
|
846 |
+
# Block-wise quant: BLOCK_SIZE_N must be divisible by block_shape[0]
|
847 |
+
# BLOCK_SIZE_K must be divisible by block_shape[1]
|
|
|
|
|
|
|
|
|
|
|
848 |
config = {
|
849 |
+
"BLOCK_SIZE_M": 64,
|
850 |
+
"BLOCK_SIZE_N": block_shape[0],
|
851 |
+
"BLOCK_SIZE_K": block_shape[1],
|
852 |
+
"GROUP_SIZE_M": 32,
|
853 |
+
"num_warps": 4,
|
854 |
+
"num_stages": 3,
|
855 |
}
|
856 |
+
else:
|
857 |
+
config = {
|
858 |
+
"BLOCK_SIZE_M": 64,
|
859 |
+
"BLOCK_SIZE_N": 64,
|
860 |
+
"BLOCK_SIZE_K": 32,
|
861 |
+
"GROUP_SIZE_M": 8,
|
862 |
+
}
|
863 |
+
# A heuristic: fused marlin works faster with this config for small M
|
864 |
+
if M <= E or (is_marlin and M <= 32):
|
865 |
+
config = {
|
866 |
+
"BLOCK_SIZE_M": 16,
|
867 |
+
"BLOCK_SIZE_N": 32,
|
868 |
+
"BLOCK_SIZE_K": 64,
|
869 |
+
"GROUP_SIZE_M": 1,
|
870 |
+
}
|
871 |
return config
|
872 |
|
873 |
|
|
|
877 |
top_k: int,
|
878 |
dtype: Optional[str],
|
879 |
M: int,
|
|
|
880 |
is_marlin: bool = False,
|
881 |
+
block_shape: Optional[List[int]] = None,
|
882 |
):
|
883 |
+
# from vllm.model_executor.layers.fused_moe import get_config
|
884 |
+
# TODO: removed when syncing to vLLM, do we need this?
|
885 |
+
# override_config = get_config()
|
886 |
+
override_config = None
|
887 |
if override_config:
|
888 |
config = override_config
|
889 |
else:
|
890 |
# First try to load optimal config from the file
|
891 |
E, _, N = w2_shape
|
892 |
+
block_n = block_shape[0] if block_shape else 0
|
893 |
+
block_k = block_shape[1] if block_shape else 0
|
894 |
+
configs = get_moe_configs(E, N, dtype, block_n, block_k)
|
895 |
|
896 |
if configs:
|
897 |
# If an optimal configuration map has been found, look up the
|
|
|
899 |
config = configs[min(configs.keys(), key=lambda x: abs(x - M))]
|
900 |
else:
|
901 |
# Else use the default config
|
902 |
+
config = get_default_config(
|
903 |
+
M, E, N, w1_shape[2], top_k, dtype, is_marlin, block_shape
|
904 |
+
)
|
905 |
return config
|
906 |
|
907 |
|
|
|
937 |
return topk_weights, topk_ids
|
938 |
|
939 |
|
940 |
+
# This is used by the Deepseek-V2 and Deepseek-V3 model
|
941 |
+
@torch.compile(dynamic=True, backend=current_platform.simple_compile_backend)
|
942 |
def grouped_topk(
|
943 |
hidden_states: torch.Tensor,
|
944 |
gating_output: torch.Tensor,
|
|
|
946 |
renormalize: bool,
|
947 |
num_expert_group: int = 0,
|
948 |
topk_group: int = 0,
|
949 |
+
scoring_func: str = "softmax",
|
950 |
+
e_score_correction_bias: Optional[torch.Tensor] = None,
|
951 |
):
|
952 |
|
953 |
assert hidden_states.shape[0] == gating_output.shape[0], "Number of tokens mismatch"
|
954 |
|
955 |
+
if scoring_func == "softmax":
|
956 |
+
scores = torch.softmax(gating_output, dim=-1)
|
957 |
+
elif scoring_func == "sigmoid":
|
958 |
+
scores = gating_output.sigmoid()
|
959 |
+
else:
|
960 |
+
raise ValueError(f"Unsupported scoring function: {scoring_func}")
|
961 |
+
|
962 |
+
if e_score_correction_bias is not None:
|
963 |
+
# Store original scores before applying correction bias. We use biased
|
964 |
+
# scores for expert selection but original scores for routing weights
|
965 |
+
original_scores = scores
|
966 |
+
scores = scores + e_score_correction_bias.unsqueeze(0)
|
967 |
+
|
968 |
num_token = scores.shape[0]
|
969 |
group_scores = (
|
970 |
scores.view(num_token, num_expert_group, -1).max(dim=-1).values
|
|
|
980 |
.reshape(num_token, -1)
|
981 |
) # [n, e]
|
982 |
tmp_scores = scores.masked_fill(~score_mask.bool(), 0.0) # [n, e]
|
983 |
+
|
984 |
+
if e_score_correction_bias is not None:
|
985 |
+
topk_ids = torch.topk(tmp_scores, k=topk, dim=-1, sorted=False)[1]
|
986 |
+
# Use original unbiased scores for the routing weights
|
987 |
+
topk_weights = original_scores.gather(1, topk_ids)
|
988 |
+
else:
|
989 |
+
topk_weights, topk_ids = torch.topk(tmp_scores, k=topk, dim=-1, sorted=False)
|
990 |
|
991 |
if renormalize:
|
992 |
topk_weights = topk_weights / topk_weights.sum(dim=-1, keepdim=True)
|
|
|
996 |
|
997 |
def get_config_dtype_str(
|
998 |
dtype: torch.dtype,
|
999 |
+
use_int4_w4a16: Optional[bool] = False,
|
1000 |
use_int8_w8a16: Optional[bool] = False,
|
1001 |
use_fp8_w8a8: Optional[bool] = False,
|
1002 |
):
|
|
|
1004 |
return "fp8_w8a8"
|
1005 |
elif use_int8_w8a16:
|
1006 |
return "int8_w8a16"
|
1007 |
+
elif use_int4_w4a16:
|
1008 |
+
return "int4_w8a16"
|
1009 |
elif dtype == torch.float:
|
1010 |
# avoiding cases where kernel fails when float32 MoE
|
1011 |
# use fp16/bfloat16 configs
|
|
|
1013 |
return None
|
1014 |
|
1015 |
|
1016 |
+
def inplace_fused_experts(
|
1017 |
+
hidden_states: torch.Tensor,
|
1018 |
+
w1: torch.Tensor,
|
1019 |
+
w2: torch.Tensor,
|
1020 |
+
topk_weights: torch.Tensor,
|
1021 |
+
topk_ids: torch.Tensor,
|
1022 |
+
use_fp8_w8a8: bool = False,
|
1023 |
+
use_int8_w8a16: bool = False,
|
1024 |
+
use_int4_w4a16: bool = False,
|
1025 |
+
w1_scale: Optional[torch.Tensor] = None,
|
1026 |
+
w2_scale: Optional[torch.Tensor] = None,
|
1027 |
+
w1_zp: Optional[torch.Tensor] = None,
|
1028 |
+
w2_zp: Optional[torch.Tensor] = None,
|
1029 |
+
a1_scale: Optional[torch.Tensor] = None,
|
1030 |
+
a2_scale: Optional[torch.Tensor] = None,
|
1031 |
+
block_shape: Optional[List[int]] = None,
|
1032 |
+
) -> None:
|
1033 |
+
fused_experts_impl(
|
1034 |
+
hidden_states,
|
1035 |
+
w1,
|
1036 |
+
w2,
|
1037 |
+
topk_weights,
|
1038 |
+
topk_ids,
|
1039 |
+
True,
|
1040 |
+
use_fp8_w8a8,
|
1041 |
+
use_int8_w8a16,
|
1042 |
+
use_int4_w4a16,
|
1043 |
+
w1_scale,
|
1044 |
+
w2_scale,
|
1045 |
+
w1_zp,
|
1046 |
+
w2_zp,
|
1047 |
+
a1_scale,
|
1048 |
+
a2_scale,
|
1049 |
+
block_shape,
|
1050 |
+
)
|
1051 |
+
|
1052 |
+
|
1053 |
+
def outplace_fused_experts(
|
1054 |
+
hidden_states: torch.Tensor,
|
1055 |
+
w1: torch.Tensor,
|
1056 |
+
w2: torch.Tensor,
|
1057 |
+
topk_weights: torch.Tensor,
|
1058 |
+
topk_ids: torch.Tensor,
|
1059 |
+
use_fp8_w8a8: bool = False,
|
1060 |
+
use_int8_w8a16: bool = False,
|
1061 |
+
use_int4_w4a16: bool = False,
|
1062 |
+
w1_scale: Optional[torch.Tensor] = None,
|
1063 |
+
w2_scale: Optional[torch.Tensor] = None,
|
1064 |
+
w1_zp: Optional[torch.Tensor] = None,
|
1065 |
+
w2_zp: Optional[torch.Tensor] = None,
|
1066 |
+
a1_scale: Optional[torch.Tensor] = None,
|
1067 |
+
a2_scale: Optional[torch.Tensor] = None,
|
1068 |
+
block_shape: Optional[List[int]] = None,
|
1069 |
+
) -> torch.Tensor:
|
1070 |
+
return fused_experts_impl(
|
1071 |
+
hidden_states,
|
1072 |
+
w1,
|
1073 |
+
w2,
|
1074 |
+
topk_weights,
|
1075 |
+
topk_ids,
|
1076 |
+
False,
|
1077 |
+
use_fp8_w8a8,
|
1078 |
+
use_int8_w8a16,
|
1079 |
+
use_int4_w4a16,
|
1080 |
+
w1_scale,
|
1081 |
+
w2_scale,
|
1082 |
+
w1_zp,
|
1083 |
+
w2_zp,
|
1084 |
+
a1_scale,
|
1085 |
+
a2_scale,
|
1086 |
+
block_shape,
|
1087 |
+
)
|
1088 |
+
|
1089 |
+
|
1090 |
def fused_experts(
|
1091 |
hidden_states: torch.Tensor,
|
1092 |
w1: torch.Tensor,
|
|
|
1094 |
topk_weights: torch.Tensor,
|
1095 |
topk_ids: torch.Tensor,
|
1096 |
inplace: bool = False,
|
|
|
1097 |
use_fp8_w8a8: bool = False,
|
1098 |
use_int8_w8a16: bool = False,
|
1099 |
+
use_int4_w4a16: bool = False,
|
1100 |
+
w1_scale: Optional[torch.Tensor] = None,
|
1101 |
+
w2_scale: Optional[torch.Tensor] = None,
|
1102 |
+
w1_zp: Optional[torch.Tensor] = None,
|
1103 |
+
w2_zp: Optional[torch.Tensor] = None,
|
1104 |
+
a1_scale: Optional[torch.Tensor] = None,
|
1105 |
+
a2_scale: Optional[torch.Tensor] = None,
|
1106 |
+
block_shape: Optional[List[int]] = None,
|
1107 |
+
):
|
1108 |
+
if inplace:
|
1109 |
+
inplace_fused_experts(
|
1110 |
+
hidden_states,
|
1111 |
+
w1,
|
1112 |
+
w2,
|
1113 |
+
topk_weights,
|
1114 |
+
topk_ids,
|
1115 |
+
use_fp8_w8a8,
|
1116 |
+
use_int8_w8a16,
|
1117 |
+
use_int4_w4a16,
|
1118 |
+
w1_scale,
|
1119 |
+
w2_scale,
|
1120 |
+
w1_zp,
|
1121 |
+
w2_zp,
|
1122 |
+
a1_scale,
|
1123 |
+
a2_scale,
|
1124 |
+
block_shape,
|
1125 |
+
)
|
1126 |
+
return hidden_states
|
1127 |
+
else:
|
1128 |
+
return outplace_fused_experts(
|
1129 |
+
hidden_states,
|
1130 |
+
w1,
|
1131 |
+
w2,
|
1132 |
+
topk_weights,
|
1133 |
+
topk_ids,
|
1134 |
+
use_fp8_w8a8,
|
1135 |
+
use_int8_w8a16,
|
1136 |
+
use_int4_w4a16,
|
1137 |
+
w1_scale,
|
1138 |
+
w2_scale,
|
1139 |
+
w1_zp,
|
1140 |
+
w2_zp,
|
1141 |
+
a1_scale,
|
1142 |
+
a2_scale,
|
1143 |
+
block_shape,
|
1144 |
+
)
|
1145 |
+
|
1146 |
+
|
1147 |
+
def fused_experts_impl(
|
1148 |
+
hidden_states: torch.Tensor,
|
1149 |
+
w1: torch.Tensor,
|
1150 |
+
w2: torch.Tensor,
|
1151 |
+
topk_weights: torch.Tensor,
|
1152 |
+
topk_ids: torch.Tensor,
|
1153 |
+
inplace: bool = False,
|
1154 |
+
use_fp8_w8a8: bool = False,
|
1155 |
+
use_int8_w8a16: bool = False,
|
1156 |
+
use_int4_w4a16: bool = False,
|
1157 |
w1_scale: Optional[torch.Tensor] = None,
|
1158 |
w2_scale: Optional[torch.Tensor] = None,
|
1159 |
+
w1_zp: Optional[torch.Tensor] = None,
|
1160 |
+
w2_zp: Optional[torch.Tensor] = None,
|
1161 |
a1_scale: Optional[torch.Tensor] = None,
|
1162 |
a2_scale: Optional[torch.Tensor] = None,
|
1163 |
+
block_shape: Optional[List[int]] = None,
|
1164 |
):
|
1165 |
# Check constraints.
|
1166 |
+
if use_int4_w4a16:
|
1167 |
+
assert hidden_states.shape[1] // 2 == w1.shape[2], "Hidden size mismatch"
|
1168 |
+
else:
|
1169 |
+
assert hidden_states.shape[1] == w1.shape[2], "Hidden size mismatch"
|
1170 |
+
|
1171 |
assert topk_weights.shape == topk_ids.shape, "topk shape mismatch"
|
1172 |
assert hidden_states.is_contiguous(), "Hidden_states must be contiguous"
|
1173 |
assert w1.is_contiguous(), "Expert weights1 must be contiguous"
|
|
|
1183 |
config_dtype = get_config_dtype_str(
|
1184 |
use_fp8_w8a8=use_fp8_w8a8,
|
1185 |
use_int8_w8a16=use_int8_w8a16,
|
1186 |
+
use_int4_w4a16=use_int4_w4a16,
|
1187 |
dtype=hidden_states.dtype,
|
1188 |
)
|
1189 |
|
|
|
1193 |
w2.shape,
|
1194 |
topk_ids.shape[1],
|
1195 |
config_dtype,
|
1196 |
+
block_shape=block_shape,
|
1197 |
)
|
1198 |
|
1199 |
config = get_config_func(M)
|
|
|
1214 |
dtype=hidden_states.dtype,
|
1215 |
)
|
1216 |
|
1217 |
+
if hidden_states.dtype == torch.bfloat16:
|
1218 |
+
compute_type = tl.bfloat16
|
1219 |
+
elif hidden_states.dtype == torch.float16:
|
1220 |
+
compute_type = tl.float16
|
1221 |
+
elif hidden_states.dtype == torch.float32:
|
1222 |
+
compute_type = tl.float32
|
1223 |
+
else:
|
1224 |
+
raise ValueError(f"Unsupported compute_type: {hidden_states.dtype}")
|
1225 |
|
1226 |
if inplace:
|
1227 |
out_hidden_states = hidden_states
|
|
|
1262 |
intermediate_cache1,
|
1263 |
a1_scale,
|
1264 |
w1_scale,
|
1265 |
+
w1_zp,
|
1266 |
curr_topk_weights,
|
1267 |
curr_topk_ids,
|
1268 |
sorted_token_ids,
|
|
|
1274 |
compute_type=compute_type,
|
1275 |
use_fp8_w8a8=use_fp8_w8a8,
|
1276 |
use_int8_w8a16=use_int8_w8a16,
|
1277 |
+
use_int4_w4a16=use_int4_w4a16,
|
1278 |
+
block_shape=block_shape,
|
1279 |
)
|
1280 |
|
1281 |
ops.silu_and_mul(intermediate_cache2, intermediate_cache1.view(-1, N))
|
|
|
1286 |
intermediate_cache3,
|
1287 |
a2_scale,
|
1288 |
w2_scale,
|
1289 |
+
w2_zp,
|
1290 |
curr_topk_weights,
|
1291 |
curr_topk_ids,
|
1292 |
sorted_token_ids,
|
|
|
1298 |
compute_type=compute_type,
|
1299 |
use_fp8_w8a8=use_fp8_w8a8,
|
1300 |
use_int8_w8a16=use_int8_w8a16,
|
1301 |
+
use_int4_w4a16=use_int4_w4a16,
|
1302 |
+
block_shape=block_shape,
|
1303 |
)
|
1304 |
|
1305 |
ops.moe_sum(
|
|
|
1317 |
topk: int,
|
1318 |
renormalize: bool,
|
1319 |
inplace: bool = False,
|
|
|
1320 |
use_grouped_topk: bool = False,
|
1321 |
num_expert_group: Optional[int] = None,
|
1322 |
topk_group: Optional[int] = None,
|
1323 |
custom_routing_function: Optional[Callable] = None,
|
1324 |
use_fp8_w8a8: bool = False,
|
1325 |
use_int8_w8a16: bool = False,
|
1326 |
+
use_int4_w4a16: bool = False,
|
1327 |
w1_scale: Optional[torch.Tensor] = None,
|
1328 |
w2_scale: Optional[torch.Tensor] = None,
|
1329 |
+
w1_zp: Optional[torch.Tensor] = None,
|
1330 |
+
w2_zp: Optional[torch.Tensor] = None,
|
1331 |
a1_scale: Optional[torch.Tensor] = None,
|
1332 |
a2_scale: Optional[torch.Tensor] = None,
|
1333 |
+
block_shape: Optional[List[int]] = None,
|
1334 |
) -> torch.Tensor:
|
1335 |
"""
|
1336 |
This function computes a Mixture of Experts (MoE) layer using two sets of
|
|
|
1346 |
- renormalize (bool): If True, renormalize the top-k weights to sum to 1.
|
1347 |
- inplace (bool): If True, perform the operation in-place.
|
1348 |
Defaults to False.
|
|
|
|
|
1349 |
- num_expert_group: Optional[int]: additional parameter for grouped_topk
|
1350 |
- topk_group: Optional[int]: additional parameter for grouped_topk
|
1351 |
- use_grouped_topk: If True, use grouped_topk instead of fused_topk
|
1352 |
note: Deepseekv2 model uses grouped_topk
|
1353 |
- use_fp8_w8a8 (bool): If True, use fp8 arithmetic to compute the inner
|
1354 |
products for w1 and w2. Defaults to False.
|
1355 |
+
- use_int8_w8a16 (bool): If True, use matmul of int8 weight and bf16/fp16
|
1356 |
+
activation to compute the inner products for w1 and w2.
|
1357 |
+
Defaults to False.
|
1358 |
+
- use_int4_w4a16 (bool): If True, use matmul of int4 weight and bf16/fp16
|
1359 |
+
activation to compute the inner products for w1 and w2.
|
1360 |
+
Defaults to False.
|
1361 |
- w1_scale (Optional[torch.Tensor]): Optional scale to be used for
|
1362 |
w1.
|
1363 |
- w2_scale (Optional[torch.Tensor]): Optional scale to be used for
|
1364 |
w2.
|
1365 |
+
- a1_scale (Optional[torch.Tensor]): Optional scale to be used for
|
1366 |
+
a1.
|
1367 |
+
- a2_scale (Optional[torch.Tensor]): Optional scale to be used for
|
1368 |
+
a2.
|
1369 |
+
- block_shape: (Optional[List[int]]): Optional block size for block-wise
|
1370 |
+
quantization.
|
1371 |
|
1372 |
Returns:
|
1373 |
- torch.Tensor: The output tensor after applying the MoE layer.
|
|
|
1401 |
topk_weights,
|
1402 |
topk_ids,
|
1403 |
inplace=inplace,
|
|
|
1404 |
use_fp8_w8a8=use_fp8_w8a8,
|
1405 |
use_int8_w8a16=use_int8_w8a16,
|
1406 |
+
use_int4_w4a16=use_int4_w4a16,
|
1407 |
w1_scale=w1_scale,
|
1408 |
w2_scale=w2_scale,
|
1409 |
+
w1_zp=w1_zp,
|
1410 |
+
w2_zp=w2_zp,
|
1411 |
a1_scale=a1_scale,
|
1412 |
a2_scale=a2_scale,
|
1413 |
+
block_shape=block_shape,
|
1414 |
)
|
build/torch25-cxx11-cu118-x86_64-linux/moe/platforms.py
CHANGED
@@ -1,22 +1,32 @@
|
|
1 |
-
from
|
2 |
-
import os
|
3 |
-
from functools import lru_cache, wraps
|
4 |
|
5 |
import torch
|
6 |
|
7 |
IS_ROCM = torch.version.hip is not None
|
8 |
|
9 |
-
|
|
|
|
|
|
|
|
|
|
|
10 |
@classmethod
|
11 |
@lru_cache(maxsize=8)
|
12 |
def get_device_name(cls, device_id: int = 0) -> str:
|
13 |
return torch.cuda.get_device_name(0)
|
14 |
|
15 |
-
|
|
|
|
|
|
|
|
|
16 |
@classmethod
|
17 |
@lru_cache(maxsize=8)
|
18 |
def get_device_name(cls, device_id: int = 0) -> str:
|
19 |
return torch.cuda.get_device_name(device_id)
|
20 |
|
|
|
|
|
|
|
21 |
|
22 |
current_platform = RocmPlatform() if IS_ROCM else CudaPlatform()
|
|
|
1 |
+
from functools import lru_cache
|
|
|
|
|
2 |
|
3 |
import torch
|
4 |
|
5 |
IS_ROCM = torch.version.hip is not None
|
6 |
|
7 |
+
|
8 |
+
class Platform:
|
9 |
+
simple_compile_backend: str = "inductor"
|
10 |
+
|
11 |
+
|
12 |
+
class CudaPlatform(Platform):
|
13 |
@classmethod
|
14 |
@lru_cache(maxsize=8)
|
15 |
def get_device_name(cls, device_id: int = 0) -> str:
|
16 |
return torch.cuda.get_device_name(0)
|
17 |
|
18 |
+
def is_rocm(self):
|
19 |
+
return False
|
20 |
+
|
21 |
+
|
22 |
+
class RocmPlatform(Platform):
|
23 |
@classmethod
|
24 |
@lru_cache(maxsize=8)
|
25 |
def get_device_name(cls, device_id: int = 0) -> str:
|
26 |
return torch.cuda.get_device_name(device_id)
|
27 |
|
28 |
+
def is_rocm(self):
|
29 |
+
return True
|
30 |
+
|
31 |
|
32 |
current_platform = RocmPlatform() if IS_ROCM else CudaPlatform()
|
build/torch25-cxx11-cu121-x86_64-linux/moe/{_moe_pqwfgssq5enn2.abi3.so → _moe_tuji4gj3mmhfo.abi3.so}
RENAMED
@@ -1,3 +1,3 @@
|
|
1 |
version https://git-lfs.github.com/spec/v1
|
2 |
-
oid sha256:
|
3 |
-
size
|
|
|
1 |
version https://git-lfs.github.com/spec/v1
|
2 |
+
oid sha256:7848d33b838158269ee403fbd068b92fae716bfc27a22f393935247b9ad58848
|
3 |
+
size 86034528
|
build/torch25-cxx11-cu121-x86_64-linux/moe/_ops.py
CHANGED
@@ -1,9 +1,9 @@
|
|
1 |
import torch
|
2 |
-
from . import
|
3 |
-
ops = torch.ops.
|
4 |
|
5 |
def add_op_namespace_prefix(op_name: str):
|
6 |
"""
|
7 |
Prefix op by namespace.
|
8 |
"""
|
9 |
-
return f"
|
|
|
1 |
import torch
|
2 |
+
from . import _moe_tuji4gj3mmhfo
|
3 |
+
ops = torch.ops._moe_tuji4gj3mmhfo
|
4 |
|
5 |
def add_op_namespace_prefix(op_name: str):
|
6 |
"""
|
7 |
Prefix op by namespace.
|
8 |
"""
|
9 |
+
return f"_moe_tuji4gj3mmhfo::{op_name}"
|
build/torch25-cxx11-cu121-x86_64-linux/moe/fp8.py
CHANGED
@@ -1,6 +1,11 @@
|
|
|
|
|
|
1 |
import torch
|
|
|
|
|
2 |
|
3 |
-
|
|
|
4 |
|
5 |
|
6 |
def is_hip() -> bool:
|
@@ -49,15 +54,179 @@ def scaled_fp8_quant(
|
|
49 |
if scale is None:
|
50 |
if use_per_token_if_dynamic:
|
51 |
scale = torch.empty((shape[0], 1), device=input.device, dtype=torch.float32)
|
52 |
-
|
53 |
-
output, input, scale, scale_ub
|
54 |
-
)
|
55 |
else:
|
56 |
scale = torch.zeros(1, device=input.device, dtype=torch.float32)
|
57 |
-
|
58 |
else:
|
59 |
# num_token_padding not implemented for this case
|
60 |
assert scale.numel() == 1 or num_token_padding is None
|
61 |
-
|
62 |
|
63 |
return output, scale
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
1 |
+
from typing import Tuple, Optional, Union
|
2 |
+
|
3 |
import torch
|
4 |
+
import triton
|
5 |
+
import triton.language as tl
|
6 |
|
7 |
+
|
8 |
+
from ._ops import ops
|
9 |
|
10 |
|
11 |
def is_hip() -> bool:
|
|
|
54 |
if scale is None:
|
55 |
if use_per_token_if_dynamic:
|
56 |
scale = torch.empty((shape[0], 1), device=input.device, dtype=torch.float32)
|
57 |
+
ops.dynamic_per_token_scaled_fp8_quant(output, input, scale, scale_ub)
|
|
|
|
|
58 |
else:
|
59 |
scale = torch.zeros(1, device=input.device, dtype=torch.float32)
|
60 |
+
ops.dynamic_scaled_fp8_quant(output, input, scale)
|
61 |
else:
|
62 |
# num_token_padding not implemented for this case
|
63 |
assert scale.numel() == 1 or num_token_padding is None
|
64 |
+
ops.static_scaled_fp8_quant(output, input, scale)
|
65 |
|
66 |
return output, scale
|
67 |
+
|
68 |
+
|
69 |
+
@triton.jit
|
70 |
+
def _per_token_group_quant_fp8(
|
71 |
+
# Pointers to inputs and output
|
72 |
+
y_ptr,
|
73 |
+
y_q_ptr,
|
74 |
+
y_s_ptr,
|
75 |
+
group_size,
|
76 |
+
# Avoid to divide zero
|
77 |
+
eps,
|
78 |
+
# Information for float8
|
79 |
+
fp8_min,
|
80 |
+
fp8_max,
|
81 |
+
# Meta-parameters
|
82 |
+
BLOCK: tl.constexpr,
|
83 |
+
):
|
84 |
+
"""A Triton-accelerated function to perform per-token-group
|
85 |
+
quantization on a tensor.
|
86 |
+
This function converts the tensor values into float8 values.
|
87 |
+
"""
|
88 |
+
# Map the program id to the row of X and Y it should compute.
|
89 |
+
g_id = tl.program_id(0)
|
90 |
+
y_ptr += g_id * group_size
|
91 |
+
y_q_ptr += g_id * group_size
|
92 |
+
y_s_ptr += g_id
|
93 |
+
|
94 |
+
cols = tl.arange(0, BLOCK) # N <= BLOCK
|
95 |
+
mask = cols < group_size
|
96 |
+
|
97 |
+
y = tl.load(y_ptr + cols, mask=mask, other=0.0).to(tl.float32)
|
98 |
+
# Quant
|
99 |
+
_absmax = tl.maximum(tl.max(tl.abs(y)), eps)
|
100 |
+
y_s = _absmax / fp8_max
|
101 |
+
y_q = tl.clamp(y / y_s, fp8_min, fp8_max).to(y_q_ptr.dtype.element_ty)
|
102 |
+
|
103 |
+
tl.store(y_q_ptr + cols, y_q, mask=mask)
|
104 |
+
tl.store(y_s_ptr, y_s)
|
105 |
+
|
106 |
+
|
107 |
+
@triton.jit
|
108 |
+
def _per_token_group_quant_fp8_colmajor(
|
109 |
+
# Pointers to inputs and output
|
110 |
+
y_ptr,
|
111 |
+
y_q_ptr,
|
112 |
+
y_s_ptr,
|
113 |
+
group_size,
|
114 |
+
# Num columns of y
|
115 |
+
y_num_columns,
|
116 |
+
# Stride from one column to the next of y_s
|
117 |
+
y_s_col_stride,
|
118 |
+
# Avoid to divide zero
|
119 |
+
eps,
|
120 |
+
# Information for float8
|
121 |
+
fp8_min,
|
122 |
+
fp8_max,
|
123 |
+
# Meta-parameters
|
124 |
+
BLOCK: tl.constexpr,
|
125 |
+
):
|
126 |
+
"""A Triton-accelerated function to perform per-token-group
|
127 |
+
quantization on a tensor.
|
128 |
+
This function converts the tensor values into float8 values.
|
129 |
+
"""
|
130 |
+
# Map the program id to the row of X and Y it should compute.
|
131 |
+
g_id = tl.program_id(0)
|
132 |
+
y_ptr += g_id * group_size
|
133 |
+
y_q_ptr += g_id * group_size
|
134 |
+
|
135 |
+
# Convert g_id the flattened block coordinate to 2D so we can index
|
136 |
+
# into the output y_scales matrix
|
137 |
+
blocks_per_row = y_num_columns // group_size
|
138 |
+
scale_col = g_id % blocks_per_row
|
139 |
+
scale_row = g_id // blocks_per_row
|
140 |
+
y_s_ptr += scale_col * y_s_col_stride + scale_row
|
141 |
+
|
142 |
+
cols = tl.arange(0, BLOCK) # group_size <= BLOCK
|
143 |
+
mask = cols < group_size
|
144 |
+
|
145 |
+
y = tl.load(y_ptr + cols, mask=mask, other=0.0).to(tl.float32)
|
146 |
+
# Quant
|
147 |
+
_absmax = tl.maximum(tl.max(tl.abs(y)), eps)
|
148 |
+
y_s = _absmax / fp8_max
|
149 |
+
y_q = tl.clamp(y / y_s, fp8_min, fp8_max).to(y_q_ptr.dtype.element_ty)
|
150 |
+
|
151 |
+
tl.store(y_q_ptr + cols, y_q, mask=mask)
|
152 |
+
tl.store(y_s_ptr, y_s)
|
153 |
+
|
154 |
+
|
155 |
+
def per_token_group_quant_fp8(
|
156 |
+
x: torch.Tensor,
|
157 |
+
group_size: int,
|
158 |
+
eps: float = 1e-10,
|
159 |
+
dtype: Optional[torch.dtype] = None,
|
160 |
+
column_major_scales: bool = False,
|
161 |
+
) -> Tuple[torch.Tensor, torch.Tensor]:
|
162 |
+
"""Function to perform per-token-group quantization on an input tensor `x`.
|
163 |
+
It converts the tensor values into signed float8 values and returns the
|
164 |
+
quantized tensor along with the scaling factor used for quantization.
|
165 |
+
Args:
|
166 |
+
x: The input tensor with ndim >= 2.
|
167 |
+
group_size: The group size used for quantization.
|
168 |
+
eps: The minimum to avoid dividing zero.
|
169 |
+
dtype: The dype of output tensor. Note that only `torch.float8_e4m3fn`
|
170 |
+
is supported for now.
|
171 |
+
Returns:
|
172 |
+
Tuple[torch.Tensor, torch.Tensor]: The quantized tensor and the
|
173 |
+
scaling factor for quantization.
|
174 |
+
"""
|
175 |
+
if dtype is None:
|
176 |
+
dtype = (
|
177 |
+
torch.float8_e4m3fnuz if current_platform.is_rocm() else torch.float8_e4m3fn
|
178 |
+
)
|
179 |
+
assert x.shape[-1] % group_size == 0, (
|
180 |
+
f"the last dimension of `x` {x.shape[-1]} must be divisible "
|
181 |
+
f"by `group_size` {group_size}"
|
182 |
+
)
|
183 |
+
assert x.is_contiguous(), "`x` must be contiguous"
|
184 |
+
|
185 |
+
finfo = torch.finfo(dtype)
|
186 |
+
fp8_min = finfo.min
|
187 |
+
fp8_max = finfo.max
|
188 |
+
|
189 |
+
x_q = torch.empty_like(x, device=x.device, dtype=dtype)
|
190 |
+
M = x.numel() // group_size
|
191 |
+
N = group_size
|
192 |
+
if column_major_scales:
|
193 |
+
shape = (x.shape[-1] // group_size,) + x.shape[:-1]
|
194 |
+
x_s = torch.empty(shape, device=x.device, dtype=torch.float32).permute(-1, -2)
|
195 |
+
else:
|
196 |
+
shape = x.shape[:-1] + (x.shape[-1] // group_size,)
|
197 |
+
x_s = torch.empty(shape, device=x.device, dtype=torch.float32)
|
198 |
+
|
199 |
+
BLOCK = triton.next_power_of_2(N)
|
200 |
+
# heuristics for number of warps
|
201 |
+
num_warps = min(max(BLOCK // 256, 1), 8)
|
202 |
+
num_stages = 1
|
203 |
+
if column_major_scales:
|
204 |
+
_per_token_group_quant_fp8_colmajor[(M,)](
|
205 |
+
x,
|
206 |
+
x_q,
|
207 |
+
x_s,
|
208 |
+
group_size,
|
209 |
+
x.shape[1],
|
210 |
+
x_s.stride(1),
|
211 |
+
eps,
|
212 |
+
fp8_min=fp8_min,
|
213 |
+
fp8_max=fp8_max,
|
214 |
+
BLOCK=BLOCK,
|
215 |
+
num_warps=num_warps,
|
216 |
+
num_stages=num_stages,
|
217 |
+
)
|
218 |
+
else:
|
219 |
+
_per_token_group_quant_fp8[(M,)](
|
220 |
+
x,
|
221 |
+
x_q,
|
222 |
+
x_s,
|
223 |
+
group_size,
|
224 |
+
eps,
|
225 |
+
fp8_min=fp8_min,
|
226 |
+
fp8_max=fp8_max,
|
227 |
+
BLOCK=BLOCK,
|
228 |
+
num_warps=num_warps,
|
229 |
+
num_stages=num_stages,
|
230 |
+
)
|
231 |
+
|
232 |
+
return x_q, x_s
|
build/torch25-cxx11-cu121-x86_64-linux/moe/fused_marlin_moe.py
CHANGED
@@ -40,7 +40,6 @@ def single_marlin_moe(
|
|
40 |
g_idx: Optional[torch.Tensor] = None,
|
41 |
sort_indices: Optional[torch.Tensor] = None,
|
42 |
w_zeros: Optional[torch.Tensor] = None,
|
43 |
-
override_config: Optional[Dict[str, Any]] = None,
|
44 |
num_bits: int = 8,
|
45 |
is_k_full: bool = True,
|
46 |
) -> torch.Tensor:
|
@@ -61,8 +60,6 @@ def single_marlin_moe(
|
|
61 |
- topk (int): The number of top-k experts to select.
|
62 |
- renormalize (bool): If True, renormalize the top-k weights to sum to 1.
|
63 |
- w_zeros (Optional[torch.Tensor]): Optional zero points to be used for w.
|
64 |
-
- override_config (Optional[Dict[str, Any]]): Optional override
|
65 |
-
for the kernel configuration.
|
66 |
- num_bits (bool): The number of bits in expert weights quantization.
|
67 |
|
68 |
Returns:
|
@@ -90,7 +87,6 @@ def single_marlin_moe(
|
|
90 |
w.shape,
|
91 |
topk_ids.shape[1],
|
92 |
None,
|
93 |
-
override_config=override_config,
|
94 |
is_marlin=True,
|
95 |
)
|
96 |
config = get_config_func(M)
|
@@ -154,6 +150,25 @@ def single_marlin_moe(
|
|
154 |
return torch.sum(intermediate_cache.view(*intermediate_cache.shape), dim=1)
|
155 |
|
156 |
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
157 |
def fused_marlin_moe(
|
158 |
hidden_states: torch.Tensor,
|
159 |
w1: torch.Tensor,
|
@@ -169,7 +184,6 @@ def fused_marlin_moe(
|
|
169 |
sort_indices2: Optional[torch.Tensor] = None,
|
170 |
w1_zeros: Optional[torch.Tensor] = None,
|
171 |
w2_zeros: Optional[torch.Tensor] = None,
|
172 |
-
override_config: Optional[Dict[str, Any]] = None,
|
173 |
num_bits: int = 8,
|
174 |
is_k_full: bool = True,
|
175 |
) -> torch.Tensor:
|
@@ -193,8 +207,6 @@ def fused_marlin_moe(
|
|
193 |
permutation.
|
194 |
- topk_weights (torch.Tensor): Top-k weights.
|
195 |
- topk_ids (torch.Tensor): Indices of topk-k elements.
|
196 |
-
- override_config (Optional[Dict[str, Any]]): Optional override
|
197 |
-
for the kernel configuration.
|
198 |
- w1_zeros (Optional[torch.Tensor]): Optional zero points to be used for w1.
|
199 |
- w2_zeros (Optional[torch.Tensor]): Optional zero points to be used for w2.
|
200 |
- num_bits (bool): The number of bits in expert weights quantization.
|
@@ -248,7 +260,6 @@ def fused_marlin_moe(
|
|
248 |
w2.shape,
|
249 |
topk_ids.shape[1],
|
250 |
None,
|
251 |
-
override_config=override_config,
|
252 |
is_marlin=True,
|
253 |
)
|
254 |
config = get_config_func(M)
|
@@ -350,6 +361,30 @@ def fused_marlin_moe(
|
|
350 |
return torch.sum(intermediate_cache3.view(*intermediate_cache3.shape), dim=1)
|
351 |
|
352 |
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
353 |
if hasattr(ops, "marlin_gemm_moe"):
|
354 |
|
355 |
@register_fake(add_op_namespace_prefix("marlin_gemm_moe"))
|
|
|
40 |
g_idx: Optional[torch.Tensor] = None,
|
41 |
sort_indices: Optional[torch.Tensor] = None,
|
42 |
w_zeros: Optional[torch.Tensor] = None,
|
|
|
43 |
num_bits: int = 8,
|
44 |
is_k_full: bool = True,
|
45 |
) -> torch.Tensor:
|
|
|
60 |
- topk (int): The number of top-k experts to select.
|
61 |
- renormalize (bool): If True, renormalize the top-k weights to sum to 1.
|
62 |
- w_zeros (Optional[torch.Tensor]): Optional zero points to be used for w.
|
|
|
|
|
63 |
- num_bits (bool): The number of bits in expert weights quantization.
|
64 |
|
65 |
Returns:
|
|
|
87 |
w.shape,
|
88 |
topk_ids.shape[1],
|
89 |
None,
|
|
|
90 |
is_marlin=True,
|
91 |
)
|
92 |
config = get_config_func(M)
|
|
|
150 |
return torch.sum(intermediate_cache.view(*intermediate_cache.shape), dim=1)
|
151 |
|
152 |
|
153 |
+
if hasattr(ops, "single_marlin_gemm_moe"):
|
154 |
+
|
155 |
+
@register_fake(add_op_namespace_prefix("single_marlin_gemm_moe"))
|
156 |
+
def single_marlin_moe_fake(
|
157 |
+
hidden_states: torch.Tensor,
|
158 |
+
w: torch.Tensor,
|
159 |
+
scales: torch.Tensor,
|
160 |
+
gating_output: torch.Tensor,
|
161 |
+
topk: int,
|
162 |
+
renormalize: bool,
|
163 |
+
g_idx: Optional[torch.Tensor] = None,
|
164 |
+
sort_indices: Optional[torch.Tensor] = None,
|
165 |
+
w_zeros: Optional[torch.Tensor] = None,
|
166 |
+
num_bits: int = 8,
|
167 |
+
is_k_full: bool = True,
|
168 |
+
) -> torch.Tensor:
|
169 |
+
return torch.empty_like(hidden_states)
|
170 |
+
|
171 |
+
|
172 |
def fused_marlin_moe(
|
173 |
hidden_states: torch.Tensor,
|
174 |
w1: torch.Tensor,
|
|
|
184 |
sort_indices2: Optional[torch.Tensor] = None,
|
185 |
w1_zeros: Optional[torch.Tensor] = None,
|
186 |
w2_zeros: Optional[torch.Tensor] = None,
|
|
|
187 |
num_bits: int = 8,
|
188 |
is_k_full: bool = True,
|
189 |
) -> torch.Tensor:
|
|
|
207 |
permutation.
|
208 |
- topk_weights (torch.Tensor): Top-k weights.
|
209 |
- topk_ids (torch.Tensor): Indices of topk-k elements.
|
|
|
|
|
210 |
- w1_zeros (Optional[torch.Tensor]): Optional zero points to be used for w1.
|
211 |
- w2_zeros (Optional[torch.Tensor]): Optional zero points to be used for w2.
|
212 |
- num_bits (bool): The number of bits in expert weights quantization.
|
|
|
260 |
w2.shape,
|
261 |
topk_ids.shape[1],
|
262 |
None,
|
|
|
263 |
is_marlin=True,
|
264 |
)
|
265 |
config = get_config_func(M)
|
|
|
361 |
return torch.sum(intermediate_cache3.view(*intermediate_cache3.shape), dim=1)
|
362 |
|
363 |
|
364 |
+
if hasattr(ops, "fused_marlin_moe"):
|
365 |
+
|
366 |
+
@register_fake(add_op_namespace_prefix("fused_marlin_moe"))
|
367 |
+
def fused_marlin_moe_fake(
|
368 |
+
hidden_states: torch.Tensor,
|
369 |
+
w1: torch.Tensor,
|
370 |
+
w2: torch.Tensor,
|
371 |
+
w1_scale: torch.Tensor,
|
372 |
+
w2_scale: torch.Tensor,
|
373 |
+
gating_output: torch.Tensor,
|
374 |
+
topk_weights: torch.Tensor,
|
375 |
+
topk_ids: torch.Tensor,
|
376 |
+
g_idx1: Optional[torch.Tensor] = None,
|
377 |
+
g_idx2: Optional[torch.Tensor] = None,
|
378 |
+
sort_indices1: Optional[torch.Tensor] = None,
|
379 |
+
sort_indices2: Optional[torch.Tensor] = None,
|
380 |
+
w1_zeros: Optional[torch.Tensor] = None,
|
381 |
+
w2_zeros: Optional[torch.Tensor] = None,
|
382 |
+
num_bits: int = 8,
|
383 |
+
is_k_full: bool = True,
|
384 |
+
) -> torch.Tensor:
|
385 |
+
return torch.empty_like(hidden_states)
|
386 |
+
|
387 |
+
|
388 |
if hasattr(ops, "marlin_gemm_moe"):
|
389 |
|
390 |
@register_fake(add_op_namespace_prefix("marlin_gemm_moe"))
|
build/torch25-cxx11-cu121-x86_64-linux/moe/fused_moe.py
CHANGED
@@ -1,21 +1,242 @@
|
|
|
|
1 |
"""Fused MoE kernel."""
|
2 |
|
3 |
import functools
|
4 |
import json
|
|
|
5 |
import os
|
6 |
-
from typing import Any, Callable, Dict, Optional, Tuple
|
7 |
|
8 |
import torch
|
9 |
import triton
|
10 |
import triton.language as tl
|
11 |
|
|
|
12 |
from ._ops import ops
|
13 |
-
from .fp8 import scaled_fp8_quant
|
14 |
from .platforms import current_platform
|
15 |
|
|
|
|
|
|
|
16 |
VLLM_FUSED_MOE_CHUNK_SIZE = int(os.getenv("VLLM_FUSED_MOE_CHUNK_SIZE", "32768"))
|
17 |
|
18 |
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
19 |
@triton.jit
|
20 |
def fused_moe_kernel(
|
21 |
# Pointers to matrices
|
@@ -44,8 +265,14 @@ def fused_moe_kernel(
|
|
44 |
stride_bn,
|
45 |
stride_cm,
|
46 |
stride_cn,
|
|
|
|
|
47 |
stride_bse,
|
|
|
48 |
stride_bsn,
|
|
|
|
|
|
|
49 |
# Meta-parameters
|
50 |
BLOCK_SIZE_M: tl.constexpr,
|
51 |
BLOCK_SIZE_N: tl.constexpr,
|
@@ -105,17 +332,17 @@ def fused_moe_kernel(
|
|
105 |
num_tokens_post_padded = tl.load(num_tokens_post_padded_ptr)
|
106 |
if pid_m * BLOCK_SIZE_M >= num_tokens_post_padded:
|
107 |
return
|
108 |
-
offs_token_id = pid_m * BLOCK_SIZE_M + tl.arange(0, BLOCK_SIZE_M)
|
109 |
offs_token = tl.load(sorted_token_ids_ptr + offs_token_id)
|
110 |
token_mask = offs_token < num_valid_tokens
|
111 |
|
112 |
-
offs_bn = (pid_n * BLOCK_SIZE_N + tl.arange(0, BLOCK_SIZE_N)) % N
|
113 |
offs_k = tl.arange(0, BLOCK_SIZE_K)
|
114 |
a_ptrs = a_ptr + (
|
115 |
offs_token[:, None] // top_k * stride_am + offs_k[None, :] * stride_ak
|
116 |
)
|
117 |
|
118 |
-
off_experts = tl.load(expert_ids_ptr + pid_m)
|
119 |
b_ptrs = (
|
120 |
b_ptr
|
121 |
+ off_experts * stride_be
|
@@ -128,8 +355,15 @@ def fused_moe_kernel(
|
|
128 |
b_scale = tl.load(b_scale_ptrs)
|
129 |
|
130 |
if use_fp8_w8a8:
|
131 |
-
|
132 |
-
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
133 |
|
134 |
# -----------------------------------------------------------
|
135 |
# Iterate to compute a block of the C matrix.
|
@@ -151,7 +385,17 @@ def fused_moe_kernel(
|
|
151 |
if use_int8_w8a16:
|
152 |
accumulator = tl.dot(a, b.to(compute_type), acc=accumulator)
|
153 |
elif use_fp8_w8a8:
|
154 |
-
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
155 |
else:
|
156 |
accumulator += tl.dot(a, b)
|
157 |
# Advance the ptrs to the next K block.
|
@@ -164,7 +408,10 @@ def fused_moe_kernel(
|
|
164 |
if use_int8_w8a16:
|
165 |
accumulator = (accumulator * b_scale).to(compute_type)
|
166 |
elif use_fp8_w8a8:
|
167 |
-
|
|
|
|
|
|
|
168 |
else:
|
169 |
accumulator = accumulator.to(compute_type)
|
170 |
# -----------------------------------------------------------
|
@@ -175,6 +422,141 @@ def fused_moe_kernel(
|
|
175 |
tl.store(c_ptrs, accumulator, mask=c_mask)
|
176 |
|
177 |
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
178 |
def moe_align_block_size(
|
179 |
topk_ids: torch.Tensor, block_size: int, num_experts: int
|
180 |
) -> Tuple[torch.Tensor, torch.Tensor, torch.Tensor]:
|
@@ -225,9 +607,34 @@ def moe_align_block_size(
|
|
225 |
(max_num_m_blocks,), dtype=torch.int32, device=topk_ids.device
|
226 |
)
|
227 |
num_tokens_post_pad = torch.empty((1), dtype=torch.int32, device=topk_ids.device)
|
228 |
-
|
229 |
-
|
230 |
-
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
231 |
return sorted_ids, expert_ids, num_tokens_post_pad
|
232 |
|
233 |
|
@@ -237,6 +644,7 @@ def invoke_fused_moe_kernel(
|
|
237 |
C: torch.Tensor,
|
238 |
A_scale: Optional[torch.Tensor],
|
239 |
B_scale: Optional[torch.Tensor],
|
|
|
240 |
topk_weights: torch.Tensor,
|
241 |
topk_ids: torch.Tensor,
|
242 |
sorted_token_ids: torch.Tensor,
|
@@ -248,64 +656,147 @@ def invoke_fused_moe_kernel(
|
|
248 |
compute_type: tl.dtype,
|
249 |
use_fp8_w8a8: bool,
|
250 |
use_int8_w8a16: bool,
|
|
|
|
|
251 |
) -> None:
|
252 |
assert topk_weights.stride(1) == 1
|
253 |
assert sorted_token_ids.stride(0) == 1
|
254 |
|
255 |
if use_fp8_w8a8:
|
256 |
-
A, A_scale = scaled_fp8_quant(A, A_scale)
|
257 |
assert B_scale is not None
|
258 |
-
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
259 |
assert B_scale is not None
|
|
|
260 |
else:
|
261 |
assert A_scale is None
|
262 |
assert B_scale is None
|
263 |
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
264 |
grid = lambda META: (
|
265 |
-
triton.cdiv(
|
266 |
* triton.cdiv(B.shape[1], META["BLOCK_SIZE_N"]),
|
267 |
)
|
268 |
|
269 |
-
|
270 |
-
|
271 |
-
|
272 |
-
|
273 |
-
|
274 |
-
B_scale
|
275 |
-
|
276 |
-
|
277 |
-
|
278 |
-
|
279 |
-
|
280 |
-
|
281 |
-
|
282 |
-
|
283 |
-
|
284 |
-
|
285 |
-
|
286 |
-
|
287 |
-
|
288 |
-
|
289 |
-
|
290 |
-
|
291 |
-
|
292 |
-
|
293 |
-
|
294 |
-
|
295 |
-
|
296 |
-
|
297 |
-
|
298 |
-
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
299 |
|
300 |
|
301 |
-
|
|
|
|
|
|
|
302 |
device_name = current_platform.get_device_name().replace(" ", "_")
|
303 |
dtype_selector = "" if not dtype else f",dtype={dtype}"
|
304 |
-
|
|
|
|
|
|
|
305 |
|
306 |
|
|
|
307 |
@functools.lru_cache
|
308 |
-
def get_moe_configs(
|
|
|
|
|
|
|
|
|
|
|
|
|
309 |
"""
|
310 |
Return optimized configurations for the fused MoE kernel.
|
311 |
|
@@ -317,18 +808,27 @@ def get_moe_configs(E: int, N: int, dtype: Optional[str]) -> Optional[Dict[int,
|
|
317 |
|
318 |
# First look up if an optimized configuration is available in the configs
|
319 |
# directory
|
320 |
-
|
|
|
321 |
|
322 |
config_file_path = os.path.join(
|
323 |
os.path.dirname(os.path.realpath(__file__)), "configs", json_file_name
|
324 |
)
|
325 |
if os.path.exists(config_file_path):
|
326 |
with open(config_file_path) as f:
|
|
|
327 |
# If a configuration has been found, return it
|
328 |
return {int(key): val for key, val in json.load(f).items()}
|
329 |
|
330 |
# If no optimized configuration is available, we will use the default
|
331 |
# configuration
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
332 |
return None
|
333 |
|
334 |
|
@@ -340,21 +840,34 @@ def get_default_config(
|
|
340 |
topk: int,
|
341 |
dtype: Optional[str],
|
342 |
is_marlin: bool,
|
|
|
343 |
) -> Dict[str, int]:
|
344 |
-
|
345 |
-
|
346 |
-
|
347 |
-
"BLOCK_SIZE_K": 32,
|
348 |
-
"GROUP_SIZE_M": 8,
|
349 |
-
}
|
350 |
-
# A heuristic: fused marlin works faster with this config for small M
|
351 |
-
if M <= E or (is_marlin and M <= 32):
|
352 |
config = {
|
353 |
-
"BLOCK_SIZE_M":
|
354 |
-
"BLOCK_SIZE_N":
|
355 |
-
"BLOCK_SIZE_K":
|
356 |
-
"GROUP_SIZE_M":
|
|
|
|
|
357 |
}
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
358 |
return config
|
359 |
|
360 |
|
@@ -364,15 +877,21 @@ def try_get_optimal_moe_config(
|
|
364 |
top_k: int,
|
365 |
dtype: Optional[str],
|
366 |
M: int,
|
367 |
-
override_config: Optional[Dict[str, Any]] = None,
|
368 |
is_marlin: bool = False,
|
|
|
369 |
):
|
|
|
|
|
|
|
|
|
370 |
if override_config:
|
371 |
config = override_config
|
372 |
else:
|
373 |
# First try to load optimal config from the file
|
374 |
E, _, N = w2_shape
|
375 |
-
|
|
|
|
|
376 |
|
377 |
if configs:
|
378 |
# If an optimal configuration map has been found, look up the
|
@@ -380,7 +899,9 @@ def try_get_optimal_moe_config(
|
|
380 |
config = configs[min(configs.keys(), key=lambda x: abs(x - M))]
|
381 |
else:
|
382 |
# Else use the default config
|
383 |
-
config = get_default_config(
|
|
|
|
|
384 |
return config
|
385 |
|
386 |
|
@@ -416,7 +937,8 @@ def fused_topk(
|
|
416 |
return topk_weights, topk_ids
|
417 |
|
418 |
|
419 |
-
# This is used by the Deepseek-V2 model
|
|
|
420 |
def grouped_topk(
|
421 |
hidden_states: torch.Tensor,
|
422 |
gating_output: torch.Tensor,
|
@@ -424,11 +946,25 @@ def grouped_topk(
|
|
424 |
renormalize: bool,
|
425 |
num_expert_group: int = 0,
|
426 |
topk_group: int = 0,
|
|
|
|
|
427 |
):
|
428 |
|
429 |
assert hidden_states.shape[0] == gating_output.shape[0], "Number of tokens mismatch"
|
430 |
|
431 |
-
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
432 |
num_token = scores.shape[0]
|
433 |
group_scores = (
|
434 |
scores.view(num_token, num_expert_group, -1).max(dim=-1).values
|
@@ -444,7 +980,13 @@ def grouped_topk(
|
|
444 |
.reshape(num_token, -1)
|
445 |
) # [n, e]
|
446 |
tmp_scores = scores.masked_fill(~score_mask.bool(), 0.0) # [n, e]
|
447 |
-
|
|
|
|
|
|
|
|
|
|
|
|
|
448 |
|
449 |
if renormalize:
|
450 |
topk_weights = topk_weights / topk_weights.sum(dim=-1, keepdim=True)
|
@@ -454,6 +996,7 @@ def grouped_topk(
|
|
454 |
|
455 |
def get_config_dtype_str(
|
456 |
dtype: torch.dtype,
|
|
|
457 |
use_int8_w8a16: Optional[bool] = False,
|
458 |
use_fp8_w8a8: Optional[bool] = False,
|
459 |
):
|
@@ -461,6 +1004,8 @@ def get_config_dtype_str(
|
|
461 |
return "fp8_w8a8"
|
462 |
elif use_int8_w8a16:
|
463 |
return "int8_w8a16"
|
|
|
|
|
464 |
elif dtype == torch.float:
|
465 |
# avoiding cases where kernel fails when float32 MoE
|
466 |
# use fp16/bfloat16 configs
|
@@ -468,6 +1013,80 @@ def get_config_dtype_str(
|
|
468 |
return None
|
469 |
|
470 |
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
471 |
def fused_experts(
|
472 |
hidden_states: torch.Tensor,
|
473 |
w1: torch.Tensor,
|
@@ -475,16 +1094,80 @@ def fused_experts(
|
|
475 |
topk_weights: torch.Tensor,
|
476 |
topk_ids: torch.Tensor,
|
477 |
inplace: bool = False,
|
478 |
-
override_config: Optional[Dict[str, Any]] = None,
|
479 |
use_fp8_w8a8: bool = False,
|
480 |
use_int8_w8a16: bool = False,
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
481 |
w1_scale: Optional[torch.Tensor] = None,
|
482 |
w2_scale: Optional[torch.Tensor] = None,
|
|
|
|
|
483 |
a1_scale: Optional[torch.Tensor] = None,
|
484 |
a2_scale: Optional[torch.Tensor] = None,
|
|
|
485 |
):
|
486 |
# Check constraints.
|
487 |
-
|
|
|
|
|
|
|
|
|
488 |
assert topk_weights.shape == topk_ids.shape, "topk shape mismatch"
|
489 |
assert hidden_states.is_contiguous(), "Hidden_states must be contiguous"
|
490 |
assert w1.is_contiguous(), "Expert weights1 must be contiguous"
|
@@ -500,6 +1183,7 @@ def fused_experts(
|
|
500 |
config_dtype = get_config_dtype_str(
|
501 |
use_fp8_w8a8=use_fp8_w8a8,
|
502 |
use_int8_w8a16=use_int8_w8a16,
|
|
|
503 |
dtype=hidden_states.dtype,
|
504 |
)
|
505 |
|
@@ -509,7 +1193,7 @@ def fused_experts(
|
|
509 |
w2.shape,
|
510 |
topk_ids.shape[1],
|
511 |
config_dtype,
|
512 |
-
|
513 |
)
|
514 |
|
515 |
config = get_config_func(M)
|
@@ -530,7 +1214,14 @@ def fused_experts(
|
|
530 |
dtype=hidden_states.dtype,
|
531 |
)
|
532 |
|
533 |
-
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
534 |
|
535 |
if inplace:
|
536 |
out_hidden_states = hidden_states
|
@@ -571,6 +1262,7 @@ def fused_experts(
|
|
571 |
intermediate_cache1,
|
572 |
a1_scale,
|
573 |
w1_scale,
|
|
|
574 |
curr_topk_weights,
|
575 |
curr_topk_ids,
|
576 |
sorted_token_ids,
|
@@ -582,6 +1274,8 @@ def fused_experts(
|
|
582 |
compute_type=compute_type,
|
583 |
use_fp8_w8a8=use_fp8_w8a8,
|
584 |
use_int8_w8a16=use_int8_w8a16,
|
|
|
|
|
585 |
)
|
586 |
|
587 |
ops.silu_and_mul(intermediate_cache2, intermediate_cache1.view(-1, N))
|
@@ -592,6 +1286,7 @@ def fused_experts(
|
|
592 |
intermediate_cache3,
|
593 |
a2_scale,
|
594 |
w2_scale,
|
|
|
595 |
curr_topk_weights,
|
596 |
curr_topk_ids,
|
597 |
sorted_token_ids,
|
@@ -603,6 +1298,8 @@ def fused_experts(
|
|
603 |
compute_type=compute_type,
|
604 |
use_fp8_w8a8=use_fp8_w8a8,
|
605 |
use_int8_w8a16=use_int8_w8a16,
|
|
|
|
|
606 |
)
|
607 |
|
608 |
ops.moe_sum(
|
@@ -620,17 +1317,20 @@ def fused_moe(
|
|
620 |
topk: int,
|
621 |
renormalize: bool,
|
622 |
inplace: bool = False,
|
623 |
-
override_config: Optional[Dict[str, Any]] = None,
|
624 |
use_grouped_topk: bool = False,
|
625 |
num_expert_group: Optional[int] = None,
|
626 |
topk_group: Optional[int] = None,
|
627 |
custom_routing_function: Optional[Callable] = None,
|
628 |
use_fp8_w8a8: bool = False,
|
629 |
use_int8_w8a16: bool = False,
|
|
|
630 |
w1_scale: Optional[torch.Tensor] = None,
|
631 |
w2_scale: Optional[torch.Tensor] = None,
|
|
|
|
|
632 |
a1_scale: Optional[torch.Tensor] = None,
|
633 |
a2_scale: Optional[torch.Tensor] = None,
|
|
|
634 |
) -> torch.Tensor:
|
635 |
"""
|
636 |
This function computes a Mixture of Experts (MoE) layer using two sets of
|
@@ -646,20 +1346,28 @@ def fused_moe(
|
|
646 |
- renormalize (bool): If True, renormalize the top-k weights to sum to 1.
|
647 |
- inplace (bool): If True, perform the operation in-place.
|
648 |
Defaults to False.
|
649 |
-
- override_config (Optional[Dict[str, Any]]): Optional override
|
650 |
-
for the kernel configuration.
|
651 |
- num_expert_group: Optional[int]: additional parameter for grouped_topk
|
652 |
- topk_group: Optional[int]: additional parameter for grouped_topk
|
653 |
- use_grouped_topk: If True, use grouped_topk instead of fused_topk
|
654 |
note: Deepseekv2 model uses grouped_topk
|
655 |
- use_fp8_w8a8 (bool): If True, use fp8 arithmetic to compute the inner
|
656 |
products for w1 and w2. Defaults to False.
|
657 |
-
- use_int8_w8a16 (bool): If True, use
|
658 |
-
products for w1 and w2.
|
|
|
|
|
|
|
|
|
659 |
- w1_scale (Optional[torch.Tensor]): Optional scale to be used for
|
660 |
w1.
|
661 |
- w2_scale (Optional[torch.Tensor]): Optional scale to be used for
|
662 |
w2.
|
|
|
|
|
|
|
|
|
|
|
|
|
663 |
|
664 |
Returns:
|
665 |
- torch.Tensor: The output tensor after applying the MoE layer.
|
@@ -693,11 +1401,14 @@ def fused_moe(
|
|
693 |
topk_weights,
|
694 |
topk_ids,
|
695 |
inplace=inplace,
|
696 |
-
override_config=override_config,
|
697 |
use_fp8_w8a8=use_fp8_w8a8,
|
698 |
use_int8_w8a16=use_int8_w8a16,
|
|
|
699 |
w1_scale=w1_scale,
|
700 |
w2_scale=w2_scale,
|
|
|
|
|
701 |
a1_scale=a1_scale,
|
702 |
a2_scale=a2_scale,
|
|
|
703 |
)
|
|
|
1 |
+
# SPDX-License-Identifier: Apache-2.0
|
2 |
"""Fused MoE kernel."""
|
3 |
|
4 |
import functools
|
5 |
import json
|
6 |
+
import logging
|
7 |
import os
|
8 |
+
from typing import Any, Callable, Dict, List, Optional, Tuple
|
9 |
|
10 |
import torch
|
11 |
import triton
|
12 |
import triton.language as tl
|
13 |
|
14 |
+
|
15 |
from ._ops import ops
|
16 |
+
from .fp8 import per_token_group_quant_fp8, scaled_fp8_quant
|
17 |
from .platforms import current_platform
|
18 |
|
19 |
+
logger = logging.getLogger(__name__)
|
20 |
+
|
21 |
+
|
22 |
VLLM_FUSED_MOE_CHUNK_SIZE = int(os.getenv("VLLM_FUSED_MOE_CHUNK_SIZE", "32768"))
|
23 |
|
24 |
|
25 |
+
@triton.jit
|
26 |
+
def fused_moe_kernel_gptq_awq(
|
27 |
+
# Pointers to matrices
|
28 |
+
a_ptr,
|
29 |
+
b_ptr,
|
30 |
+
c_ptr,
|
31 |
+
b_scale_ptr,
|
32 |
+
b_zp_ptr,
|
33 |
+
topk_weights_ptr,
|
34 |
+
sorted_token_ids_ptr,
|
35 |
+
expert_ids_ptr,
|
36 |
+
num_tokens_post_padded_ptr,
|
37 |
+
# Matrix dimensions
|
38 |
+
N: tl.constexpr,
|
39 |
+
K: tl.constexpr,
|
40 |
+
EM,
|
41 |
+
num_valid_tokens,
|
42 |
+
# The stride variables represent how much to increase the ptr by when
|
43 |
+
# moving by 1 element in a particular dimension. E.g. `stride_am` is
|
44 |
+
# how much to increase `a_ptr` by to get the element one row down
|
45 |
+
# (A has M rows).
|
46 |
+
stride_am,
|
47 |
+
stride_ak,
|
48 |
+
stride_be,
|
49 |
+
stride_bk,
|
50 |
+
stride_bn,
|
51 |
+
stride_cm,
|
52 |
+
stride_cn,
|
53 |
+
stride_bse,
|
54 |
+
stride_bsk,
|
55 |
+
stride_bsn,
|
56 |
+
stride_bze,
|
57 |
+
stride_bzk,
|
58 |
+
stride_bzn,
|
59 |
+
block_k_diviable: tl.constexpr,
|
60 |
+
group_size: tl.constexpr,
|
61 |
+
# Meta-parameters
|
62 |
+
BLOCK_SIZE_M: tl.constexpr,
|
63 |
+
BLOCK_SIZE_N: tl.constexpr,
|
64 |
+
BLOCK_SIZE_K: tl.constexpr,
|
65 |
+
GROUP_SIZE_M: tl.constexpr,
|
66 |
+
MUL_ROUTED_WEIGHT: tl.constexpr,
|
67 |
+
top_k: tl.constexpr,
|
68 |
+
compute_type: tl.constexpr,
|
69 |
+
has_zp: tl.constexpr,
|
70 |
+
use_int4_w4a16: tl.constexpr,
|
71 |
+
use_int8_w8a16: tl.constexpr,
|
72 |
+
):
|
73 |
+
"""
|
74 |
+
Implements the fused computation for a Mixture of Experts (MOE) using
|
75 |
+
token and expert matrices.
|
76 |
+
|
77 |
+
Key Parameters:
|
78 |
+
- A: The input tensor representing tokens with shape (*, K), where '*' can
|
79 |
+
be any shape representing batches and K is the feature dimension of
|
80 |
+
each token.
|
81 |
+
- B: The stacked MOE weight tensor with shape (E, N, K), where E is
|
82 |
+
the number of experts, K is the input feature dimension, and N is
|
83 |
+
the output feature dimension.
|
84 |
+
- C: The output cache tensor with shape (M, topk, N), where M is the
|
85 |
+
total number of tokens post padding, topk is the number of times
|
86 |
+
each token is repeated, and N is the output feature dimension.
|
87 |
+
- sorted_token_ids: A tensor containing the sorted indices of tokens,
|
88 |
+
repeated topk times and arranged by the expert index they are
|
89 |
+
assigned to.
|
90 |
+
- expert_ids: A tensor containing the indices of the expert for each
|
91 |
+
block. It determines which expert matrix from B should be used for
|
92 |
+
each block in A.
|
93 |
+
This kernel performs the multiplication of a token by its corresponding
|
94 |
+
expert matrix as determined by `expert_ids`. The sorting of
|
95 |
+
`sorted_token_ids` by expert index and padding ensures divisibility by
|
96 |
+
BLOCK_SIZE_M, which is necessary to maintain consistency in block matrix
|
97 |
+
multiplication across different blocks processed by the same expert.
|
98 |
+
"""
|
99 |
+
# -----------------------------------------------------------
|
100 |
+
# Map program ids `pid` to the block of C it should compute.
|
101 |
+
# This is done in a grouped ordering to promote L2 data reuse.
|
102 |
+
pid = tl.program_id(axis=0)
|
103 |
+
num_pid_m = tl.cdiv(EM, BLOCK_SIZE_M)
|
104 |
+
num_pid_n = tl.cdiv(N, BLOCK_SIZE_N)
|
105 |
+
num_pid_in_group = GROUP_SIZE_M * num_pid_n
|
106 |
+
group_id = pid // num_pid_in_group
|
107 |
+
first_pid_m = group_id * GROUP_SIZE_M
|
108 |
+
group_size_m = min(num_pid_m - first_pid_m, GROUP_SIZE_M)
|
109 |
+
pid_m = first_pid_m + ((pid % num_pid_in_group) % group_size_m)
|
110 |
+
pid_n = (pid % num_pid_in_group) // group_size_m
|
111 |
+
|
112 |
+
# ----------------------------------------------------------
|
113 |
+
# Create pointers for the first blocks of A and B.
|
114 |
+
# We will advance this pointer as we move in the K direction
|
115 |
+
# and accumulate
|
116 |
+
# `a_ptrs` is a block of [BLOCK_SIZE_M, BLOCK_SIZE_K] pointers
|
117 |
+
# `b_ptrs` is a block of [BLOCK_SIZE_K, BLOCK_SIZE_N] pointers
|
118 |
+
num_tokens_post_padded = tl.load(num_tokens_post_padded_ptr)
|
119 |
+
if pid_m * BLOCK_SIZE_M >= num_tokens_post_padded:
|
120 |
+
return
|
121 |
+
offs_token_id = pid_m * BLOCK_SIZE_M + tl.arange(0, BLOCK_SIZE_M).to(tl.int64)
|
122 |
+
offs_token = tl.load(sorted_token_ids_ptr + offs_token_id)
|
123 |
+
token_mask = offs_token < num_valid_tokens
|
124 |
+
|
125 |
+
offs_bn = (pid_n * BLOCK_SIZE_N + tl.arange(0, BLOCK_SIZE_N).to(tl.int64)) % N
|
126 |
+
offs_k = tl.arange(0, BLOCK_SIZE_K)
|
127 |
+
a_ptrs = a_ptr + (
|
128 |
+
offs_token[:, None] // top_k * stride_am + offs_k[None, :] * stride_ak
|
129 |
+
)
|
130 |
+
|
131 |
+
off_experts = tl.load(expert_ids_ptr + pid_m).to(tl.int64)
|
132 |
+
|
133 |
+
if use_int4_w4a16:
|
134 |
+
b_ptrs = (
|
135 |
+
b_ptr
|
136 |
+
+ off_experts * stride_be
|
137 |
+
+ (offs_k[:, None] // 2) * stride_bk
|
138 |
+
+ offs_bn[None, :] * stride_bn
|
139 |
+
)
|
140 |
+
b_shifter = (offs_k[:, None] % 2) * 4
|
141 |
+
elif use_int8_w8a16:
|
142 |
+
b_ptrs = (
|
143 |
+
b_ptr
|
144 |
+
+ off_experts * stride_be
|
145 |
+
+ offs_k[:, None] * stride_bk
|
146 |
+
+ offs_bn[None, :] * stride_bn
|
147 |
+
)
|
148 |
+
|
149 |
+
if not has_zp and use_int4_w4a16:
|
150 |
+
b_zp_num = 8
|
151 |
+
if not has_zp and use_int8_w8a16:
|
152 |
+
b_zp_num = 128
|
153 |
+
elif has_zp and use_int4_w4a16:
|
154 |
+
b_zp_shifter = (offs_bn[None, :] % 2) * 4
|
155 |
+
|
156 |
+
# -----------------------------------------------------------
|
157 |
+
# Iterate to compute a block of the C matrix.
|
158 |
+
# We accumulate into a `[BLOCK_SIZE_M, BLOCK_SIZE_N]` block
|
159 |
+
# of fp32 values for higher accuracy.
|
160 |
+
# `accumulator` will be converted back to fp16 after the loop.
|
161 |
+
accumulator = tl.zeros((BLOCK_SIZE_M, BLOCK_SIZE_N), dtype=tl.float32)
|
162 |
+
for k in range(0, tl.cdiv(K, BLOCK_SIZE_K)):
|
163 |
+
# Load the next block of A and B, generate a mask by checking the
|
164 |
+
# K dimension.
|
165 |
+
|
166 |
+
if not block_k_diviable:
|
167 |
+
k_mask = offs_k[:, None] < K - k * BLOCK_SIZE_K
|
168 |
+
k_other = 0.0
|
169 |
+
else:
|
170 |
+
k_mask = None
|
171 |
+
k_other = None
|
172 |
+
|
173 |
+
a = tl.load(
|
174 |
+
a_ptrs,
|
175 |
+
mask=token_mask[:, None] & (offs_k[None, :] < K - k * BLOCK_SIZE_K),
|
176 |
+
other=0.0,
|
177 |
+
)
|
178 |
+
b = tl.load(b_ptrs)
|
179 |
+
if use_int4_w4a16:
|
180 |
+
b = (b >> b_shifter) & 0xF
|
181 |
+
|
182 |
+
b_scale_ptrs = (
|
183 |
+
b_scale_ptr
|
184 |
+
+ off_experts * stride_bse
|
185 |
+
+ offs_bn[None, :] * stride_bsn
|
186 |
+
+ ((offs_k[:, None] + BLOCK_SIZE_K * k) // group_size) * stride_bsk
|
187 |
+
)
|
188 |
+
b_scale = tl.load(b_scale_ptrs, mask=k_mask, other=k_other)
|
189 |
+
b_scale = b_scale.to(tl.float32)
|
190 |
+
|
191 |
+
if has_zp and use_int4_w4a16:
|
192 |
+
offs_k_true = (offs_k[:, None] + BLOCK_SIZE_K * k) // group_size
|
193 |
+
b_zp_ptrs = (
|
194 |
+
b_zp_ptr
|
195 |
+
+ off_experts * stride_bze
|
196 |
+
+ (offs_bn[None, :] // 2) * stride_bzn
|
197 |
+
+ offs_k_true * stride_bzk
|
198 |
+
)
|
199 |
+
b_zp = tl.load(b_zp_ptrs, mask=k_mask, other=k_other)
|
200 |
+
b_zp = (b_zp >> b_zp_shifter) & 0xF
|
201 |
+
b_zp = b_zp.to(tl.float32)
|
202 |
+
elif has_zp and use_int8_w8a16:
|
203 |
+
offs_k_true = (offs_k[:, None] + BLOCK_SIZE_K * k) // group_size
|
204 |
+
b_zp_ptrs = (
|
205 |
+
b_zp_ptr
|
206 |
+
+ off_experts * stride_bze
|
207 |
+
+ offs_bn[None, :] * stride_bzn
|
208 |
+
+ offs_k_true * stride_bzk
|
209 |
+
)
|
210 |
+
b_zp = tl.load(b_zp_ptrs, mask=k_mask, other=k_other)
|
211 |
+
b_zp = b_zp.to(tl.float32)
|
212 |
+
|
213 |
+
# We accumulate along the K dimension.
|
214 |
+
if has_zp:
|
215 |
+
b = ((b.to(tl.float32) - b_zp) * b_scale).to(compute_type)
|
216 |
+
else:
|
217 |
+
b = ((b.to(tl.float32) - b_zp_num) * b_scale).to(compute_type)
|
218 |
+
accumulator = tl.dot(a, b, acc=accumulator)
|
219 |
+
|
220 |
+
# Advance the ptrs to the next K block.
|
221 |
+
a_ptrs += BLOCK_SIZE_K * stride_ak
|
222 |
+
if use_int4_w4a16:
|
223 |
+
b_ptrs += (BLOCK_SIZE_K // 2) * stride_bk
|
224 |
+
else:
|
225 |
+
b_ptrs += BLOCK_SIZE_K * stride_bk
|
226 |
+
|
227 |
+
if MUL_ROUTED_WEIGHT:
|
228 |
+
moe_weight = tl.load(topk_weights_ptr + offs_token, mask=token_mask, other=0)
|
229 |
+
accumulator = accumulator * moe_weight[:, None]
|
230 |
+
|
231 |
+
accumulator = accumulator.to(compute_type)
|
232 |
+
# -----------------------------------------------------------
|
233 |
+
# Write back the block of the output
|
234 |
+
offs_cn = pid_n * BLOCK_SIZE_N + tl.arange(0, BLOCK_SIZE_N)
|
235 |
+
c_ptrs = c_ptr + stride_cm * offs_token[:, None] + stride_cn * offs_cn[None, :]
|
236 |
+
c_mask = token_mask[:, None] & (offs_cn[None, :] < N)
|
237 |
+
tl.store(c_ptrs, accumulator, mask=c_mask)
|
238 |
+
|
239 |
+
|
240 |
@triton.jit
|
241 |
def fused_moe_kernel(
|
242 |
# Pointers to matrices
|
|
|
265 |
stride_bn,
|
266 |
stride_cm,
|
267 |
stride_cn,
|
268 |
+
stride_asm,
|
269 |
+
stride_ask,
|
270 |
stride_bse,
|
271 |
+
stride_bsk,
|
272 |
stride_bsn,
|
273 |
+
# Block size for block-wise quantization
|
274 |
+
group_n: tl.constexpr,
|
275 |
+
group_k: tl.constexpr,
|
276 |
# Meta-parameters
|
277 |
BLOCK_SIZE_M: tl.constexpr,
|
278 |
BLOCK_SIZE_N: tl.constexpr,
|
|
|
332 |
num_tokens_post_padded = tl.load(num_tokens_post_padded_ptr)
|
333 |
if pid_m * BLOCK_SIZE_M >= num_tokens_post_padded:
|
334 |
return
|
335 |
+
offs_token_id = pid_m * BLOCK_SIZE_M + tl.arange(0, BLOCK_SIZE_M).to(tl.int64)
|
336 |
offs_token = tl.load(sorted_token_ids_ptr + offs_token_id)
|
337 |
token_mask = offs_token < num_valid_tokens
|
338 |
|
339 |
+
offs_bn = (pid_n * BLOCK_SIZE_N + tl.arange(0, BLOCK_SIZE_N).to(tl.int64)) % N
|
340 |
offs_k = tl.arange(0, BLOCK_SIZE_K)
|
341 |
a_ptrs = a_ptr + (
|
342 |
offs_token[:, None] // top_k * stride_am + offs_k[None, :] * stride_ak
|
343 |
)
|
344 |
|
345 |
+
off_experts = tl.load(expert_ids_ptr + pid_m).to(tl.int64)
|
346 |
b_ptrs = (
|
347 |
b_ptr
|
348 |
+ off_experts * stride_be
|
|
|
355 |
b_scale = tl.load(b_scale_ptrs)
|
356 |
|
357 |
if use_fp8_w8a8:
|
358 |
+
if group_k > 0 and group_n > 0:
|
359 |
+
a_scale_ptrs = a_scale_ptr + (offs_token // top_k) * stride_asm
|
360 |
+
offs_bsn = offs_bn // group_n
|
361 |
+
b_scale_ptrs = (
|
362 |
+
b_scale_ptr + off_experts * stride_bse + offs_bsn * stride_bsn
|
363 |
+
)
|
364 |
+
else:
|
365 |
+
a_scale = tl.load(a_scale_ptr)
|
366 |
+
b_scale = tl.load(b_scale_ptr + off_experts)
|
367 |
|
368 |
# -----------------------------------------------------------
|
369 |
# Iterate to compute a block of the C matrix.
|
|
|
385 |
if use_int8_w8a16:
|
386 |
accumulator = tl.dot(a, b.to(compute_type), acc=accumulator)
|
387 |
elif use_fp8_w8a8:
|
388 |
+
if group_k > 0 and group_n > 0:
|
389 |
+
k_start = k * BLOCK_SIZE_K
|
390 |
+
offs_ks = k_start // group_k
|
391 |
+
a_scale = tl.load(
|
392 |
+
a_scale_ptrs + offs_ks * stride_ask, mask=token_mask, other=0.0
|
393 |
+
)
|
394 |
+
b_scale = tl.load(b_scale_ptrs + offs_ks * stride_bsk)
|
395 |
+
|
396 |
+
accumulator += tl.dot(a, b) * a_scale[:, None] * b_scale[None, :]
|
397 |
+
else:
|
398 |
+
accumulator = tl.dot(a, b, acc=accumulator)
|
399 |
else:
|
400 |
accumulator += tl.dot(a, b)
|
401 |
# Advance the ptrs to the next K block.
|
|
|
408 |
if use_int8_w8a16:
|
409 |
accumulator = (accumulator * b_scale).to(compute_type)
|
410 |
elif use_fp8_w8a8:
|
411 |
+
if group_k > 0 and group_n > 0:
|
412 |
+
accumulator = accumulator.to(compute_type)
|
413 |
+
else:
|
414 |
+
accumulator = (accumulator * a_scale * b_scale).to(compute_type)
|
415 |
else:
|
416 |
accumulator = accumulator.to(compute_type)
|
417 |
# -----------------------------------------------------------
|
|
|
422 |
tl.store(c_ptrs, accumulator, mask=c_mask)
|
423 |
|
424 |
|
425 |
+
def ceil_div(a, b):
|
426 |
+
return (a + b - 1) // b
|
427 |
+
|
428 |
+
|
429 |
+
@triton.jit
|
430 |
+
def moe_align_block_size_stage1(
|
431 |
+
topk_ids_ptr,
|
432 |
+
tokens_cnts_ptr,
|
433 |
+
num_experts: tl.constexpr,
|
434 |
+
numel: tl.constexpr,
|
435 |
+
tokens_per_thread: tl.constexpr,
|
436 |
+
):
|
437 |
+
pid = tl.program_id(0)
|
438 |
+
|
439 |
+
start_idx = pid * tokens_per_thread
|
440 |
+
|
441 |
+
off_c = (pid + 1) * num_experts
|
442 |
+
|
443 |
+
for i in range(tokens_per_thread):
|
444 |
+
if start_idx + i < numel:
|
445 |
+
idx = tl.load(topk_ids_ptr + start_idx + i)
|
446 |
+
token_cnt = tl.load(tokens_cnts_ptr + off_c + idx)
|
447 |
+
tl.store(tokens_cnts_ptr + off_c + idx, token_cnt + 1)
|
448 |
+
|
449 |
+
|
450 |
+
@triton.jit
|
451 |
+
def moe_align_block_size_stage2(
|
452 |
+
tokens_cnts_ptr,
|
453 |
+
num_experts: tl.constexpr,
|
454 |
+
):
|
455 |
+
pid = tl.program_id(0)
|
456 |
+
|
457 |
+
last_cnt = 0
|
458 |
+
for i in range(1, num_experts + 1):
|
459 |
+
token_cnt = tl.load(tokens_cnts_ptr + i * num_experts + pid)
|
460 |
+
last_cnt = last_cnt + token_cnt
|
461 |
+
tl.store(tokens_cnts_ptr + i * num_experts + pid, last_cnt)
|
462 |
+
|
463 |
+
|
464 |
+
@triton.jit
|
465 |
+
def moe_align_block_size_stage3(
|
466 |
+
total_tokens_post_pad_ptr,
|
467 |
+
tokens_cnts_ptr,
|
468 |
+
cumsum_ptr,
|
469 |
+
num_experts: tl.constexpr,
|
470 |
+
block_size: tl.constexpr,
|
471 |
+
):
|
472 |
+
last_cumsum = 0
|
473 |
+
off_cnt = num_experts * num_experts
|
474 |
+
for i in range(1, num_experts + 1):
|
475 |
+
token_cnt = tl.load(tokens_cnts_ptr + off_cnt + i - 1)
|
476 |
+
last_cumsum = last_cumsum + tl.cdiv(token_cnt, block_size) * block_size
|
477 |
+
tl.store(cumsum_ptr + i, last_cumsum)
|
478 |
+
tl.store(total_tokens_post_pad_ptr, last_cumsum)
|
479 |
+
|
480 |
+
|
481 |
+
@triton.jit
|
482 |
+
def moe_align_block_size_stage4(
|
483 |
+
topk_ids_ptr,
|
484 |
+
sorted_token_ids_ptr,
|
485 |
+
expert_ids_ptr,
|
486 |
+
tokens_cnts_ptr,
|
487 |
+
cumsum_ptr,
|
488 |
+
num_experts: tl.constexpr,
|
489 |
+
block_size: tl.constexpr,
|
490 |
+
numel: tl.constexpr,
|
491 |
+
tokens_per_thread: tl.constexpr,
|
492 |
+
):
|
493 |
+
pid = tl.program_id(0)
|
494 |
+
start_idx = tl.load(cumsum_ptr + pid)
|
495 |
+
end_idx = tl.load(cumsum_ptr + pid + 1)
|
496 |
+
|
497 |
+
for i in range(start_idx, end_idx, block_size):
|
498 |
+
tl.store(expert_ids_ptr + i // block_size, pid)
|
499 |
+
|
500 |
+
start_idx = pid * tokens_per_thread
|
501 |
+
off_t = pid * num_experts
|
502 |
+
|
503 |
+
for i in range(start_idx, tl.minimum(start_idx + tokens_per_thread, numel)):
|
504 |
+
expert_id = tl.load(topk_ids_ptr + i)
|
505 |
+
token_cnt = tl.load(tokens_cnts_ptr + off_t + expert_id)
|
506 |
+
rank_post_pad = token_cnt + tl.load(cumsum_ptr + expert_id)
|
507 |
+
tl.store(sorted_token_ids_ptr + rank_post_pad, i)
|
508 |
+
tl.store(tokens_cnts_ptr + off_t + expert_id, token_cnt + 1)
|
509 |
+
|
510 |
+
|
511 |
+
# Triton implementation based on:
|
512 |
+
# https://github.com/sgl-project/sglang/commit/ba5112ff691d791a9e38c6c71f59324a5fcb49d0
|
513 |
+
def moe_align_block_size_triton(
|
514 |
+
topk_ids: torch.Tensor,
|
515 |
+
num_experts: int,
|
516 |
+
block_size: int,
|
517 |
+
sorted_token_ids: torch.Tensor,
|
518 |
+
expert_ids: torch.Tensor,
|
519 |
+
num_tokens_post_pad: torch.Tensor,
|
520 |
+
) -> None:
|
521 |
+
numel = topk_ids.numel()
|
522 |
+
grid = (num_experts,)
|
523 |
+
tokens_cnts = torch.zeros(
|
524 |
+
(num_experts + 1, num_experts), dtype=torch.int32, device=topk_ids.device
|
525 |
+
)
|
526 |
+
cumsum = torch.zeros((num_experts + 1,), dtype=torch.int32, device=topk_ids.device)
|
527 |
+
tokens_per_thread = ceil_div(numel, num_experts)
|
528 |
+
|
529 |
+
moe_align_block_size_stage1[grid](
|
530 |
+
topk_ids,
|
531 |
+
tokens_cnts,
|
532 |
+
num_experts,
|
533 |
+
numel,
|
534 |
+
tokens_per_thread,
|
535 |
+
)
|
536 |
+
moe_align_block_size_stage2[grid](
|
537 |
+
tokens_cnts,
|
538 |
+
num_experts,
|
539 |
+
)
|
540 |
+
moe_align_block_size_stage3[(1,)](
|
541 |
+
num_tokens_post_pad,
|
542 |
+
tokens_cnts,
|
543 |
+
cumsum,
|
544 |
+
num_experts,
|
545 |
+
block_size,
|
546 |
+
)
|
547 |
+
moe_align_block_size_stage4[grid](
|
548 |
+
topk_ids,
|
549 |
+
sorted_token_ids,
|
550 |
+
expert_ids,
|
551 |
+
tokens_cnts,
|
552 |
+
cumsum,
|
553 |
+
num_experts,
|
554 |
+
block_size,
|
555 |
+
numel,
|
556 |
+
tokens_per_thread,
|
557 |
+
)
|
558 |
+
|
559 |
+
|
560 |
def moe_align_block_size(
|
561 |
topk_ids: torch.Tensor, block_size: int, num_experts: int
|
562 |
) -> Tuple[torch.Tensor, torch.Tensor, torch.Tensor]:
|
|
|
607 |
(max_num_m_blocks,), dtype=torch.int32, device=topk_ids.device
|
608 |
)
|
609 |
num_tokens_post_pad = torch.empty((1), dtype=torch.int32, device=topk_ids.device)
|
610 |
+
if num_experts >= 224:
|
611 |
+
if VLLM_ENABLE_MOE_ALIGN_BLOCK_SIZE_TRITON:
|
612 |
+
moe_align_block_size_triton(
|
613 |
+
topk_ids,
|
614 |
+
num_experts,
|
615 |
+
block_size,
|
616 |
+
sorted_ids,
|
617 |
+
expert_ids,
|
618 |
+
num_tokens_post_pad,
|
619 |
+
)
|
620 |
+
else:
|
621 |
+
ops.sgl_moe_align_block_size(
|
622 |
+
topk_ids,
|
623 |
+
num_experts,
|
624 |
+
block_size,
|
625 |
+
sorted_ids,
|
626 |
+
expert_ids,
|
627 |
+
num_tokens_post_pad,
|
628 |
+
)
|
629 |
+
else:
|
630 |
+
ops.moe_align_block_size(
|
631 |
+
topk_ids,
|
632 |
+
num_experts,
|
633 |
+
block_size,
|
634 |
+
sorted_ids,
|
635 |
+
expert_ids,
|
636 |
+
num_tokens_post_pad,
|
637 |
+
)
|
638 |
return sorted_ids, expert_ids, num_tokens_post_pad
|
639 |
|
640 |
|
|
|
644 |
C: torch.Tensor,
|
645 |
A_scale: Optional[torch.Tensor],
|
646 |
B_scale: Optional[torch.Tensor],
|
647 |
+
B_zp: Optional[torch.Tensor],
|
648 |
topk_weights: torch.Tensor,
|
649 |
topk_ids: torch.Tensor,
|
650 |
sorted_token_ids: torch.Tensor,
|
|
|
656 |
compute_type: tl.dtype,
|
657 |
use_fp8_w8a8: bool,
|
658 |
use_int8_w8a16: bool,
|
659 |
+
use_int4_w4a16: bool,
|
660 |
+
block_shape: Optional[List[int]] = None,
|
661 |
) -> None:
|
662 |
assert topk_weights.stride(1) == 1
|
663 |
assert sorted_token_ids.stride(0) == 1
|
664 |
|
665 |
if use_fp8_w8a8:
|
|
|
666 |
assert B_scale is not None
|
667 |
+
if block_shape is None:
|
668 |
+
A, A_scale = scaled_fp8_quant(A, A_scale)
|
669 |
+
else:
|
670 |
+
assert len(block_shape) == 2
|
671 |
+
block_n, block_k = block_shape[0], block_shape[1]
|
672 |
+
A, A_scale = per_token_group_quant_fp8(A, block_k)
|
673 |
+
assert triton.cdiv(A.shape[-1], block_k) == A_scale.shape[-1]
|
674 |
+
assert triton.cdiv(B.shape[-2], block_n) == B_scale.shape[-2]
|
675 |
+
assert triton.cdiv(B.shape[-1], block_k) == B_scale.shape[-1]
|
676 |
+
elif use_int8_w8a16 or use_int4_w4a16:
|
677 |
assert B_scale is not None
|
678 |
+
assert block_shape is None or block_shape[0] == 0
|
679 |
else:
|
680 |
assert A_scale is None
|
681 |
assert B_scale is None
|
682 |
|
683 |
+
EM = sorted_token_ids.shape[0]
|
684 |
+
if A.shape[0] < config["BLOCK_SIZE_M"]:
|
685 |
+
# optimize for small batch_size.
|
686 |
+
# We assume that top_ids of each token is unique, so
|
687 |
+
# so num_valid_experts <= batch_size <= BLOCK_SIZE_M,
|
688 |
+
# and we can skip some invalid blocks.
|
689 |
+
EM = min(sorted_token_ids.shape[0], A.shape[0] * top_k * config["BLOCK_SIZE_M"])
|
690 |
grid = lambda META: (
|
691 |
+
triton.cdiv(EM, META["BLOCK_SIZE_M"])
|
692 |
* triton.cdiv(B.shape[1], META["BLOCK_SIZE_N"]),
|
693 |
)
|
694 |
|
695 |
+
if (
|
696 |
+
(use_int8_w8a16 or use_int4_w4a16)
|
697 |
+
and block_shape is not None
|
698 |
+
and block_shape[1] > 0
|
699 |
+
):
|
700 |
+
assert B_scale is not None and B_scale.ndim == 3
|
701 |
+
assert B_zp is None or B_zp.ndim == 3
|
702 |
+
|
703 |
+
fused_moe_kernel_gptq_awq[grid](
|
704 |
+
A,
|
705 |
+
B,
|
706 |
+
C,
|
707 |
+
B_scale,
|
708 |
+
B_zp,
|
709 |
+
topk_weights,
|
710 |
+
sorted_token_ids,
|
711 |
+
expert_ids,
|
712 |
+
num_tokens_post_padded,
|
713 |
+
B.shape[1],
|
714 |
+
A.shape[1],
|
715 |
+
EM,
|
716 |
+
topk_ids.numel(),
|
717 |
+
A.stride(0),
|
718 |
+
A.stride(1),
|
719 |
+
B.stride(0),
|
720 |
+
B.stride(2),
|
721 |
+
B.stride(1),
|
722 |
+
C.stride(1),
|
723 |
+
C.stride(2),
|
724 |
+
B_scale.stride(0),
|
725 |
+
B_scale.stride(2),
|
726 |
+
B_scale.stride(1),
|
727 |
+
B_zp.stride(0) if B_zp is not None else 0,
|
728 |
+
B_zp.stride(2) if B_zp is not None else 0,
|
729 |
+
B_zp.stride(1) if B_zp is not None else 0,
|
730 |
+
block_k_diviable=A.shape[1] % config["BLOCK_SIZE_K"] == 0,
|
731 |
+
group_size=block_shape[1],
|
732 |
+
MUL_ROUTED_WEIGHT=mul_routed_weight,
|
733 |
+
top_k=top_k,
|
734 |
+
compute_type=compute_type,
|
735 |
+
has_zp=B_zp is not None,
|
736 |
+
use_int4_w4a16=use_int4_w4a16,
|
737 |
+
use_int8_w8a16=use_int8_w8a16,
|
738 |
+
**config,
|
739 |
+
)
|
740 |
+
|
741 |
+
else:
|
742 |
+
fused_moe_kernel[grid](
|
743 |
+
A,
|
744 |
+
B,
|
745 |
+
C,
|
746 |
+
A_scale,
|
747 |
+
B_scale,
|
748 |
+
topk_weights,
|
749 |
+
sorted_token_ids,
|
750 |
+
expert_ids,
|
751 |
+
num_tokens_post_padded,
|
752 |
+
B.shape[1],
|
753 |
+
A.shape[1],
|
754 |
+
EM,
|
755 |
+
topk_ids.numel(),
|
756 |
+
A.stride(0),
|
757 |
+
A.stride(1),
|
758 |
+
B.stride(0),
|
759 |
+
B.stride(2),
|
760 |
+
B.stride(1),
|
761 |
+
C.stride(1),
|
762 |
+
C.stride(2),
|
763 |
+
A_scale.stride(0) if A_scale is not None and A_scale.ndim == 2 else 0,
|
764 |
+
A_scale.stride(1) if A_scale is not None and A_scale.ndim == 2 else 0,
|
765 |
+
B_scale.stride(0) if B_scale is not None and B_scale.ndim >= 2 else 0,
|
766 |
+
B_scale.stride(2) if B_scale is not None and B_scale.ndim == 3 else 0,
|
767 |
+
B_scale.stride(1) if B_scale is not None and B_scale.ndim >= 2 else 0,
|
768 |
+
0 if block_shape is None else block_shape[0],
|
769 |
+
0 if block_shape is None else block_shape[1],
|
770 |
+
MUL_ROUTED_WEIGHT=mul_routed_weight,
|
771 |
+
top_k=top_k,
|
772 |
+
compute_type=compute_type,
|
773 |
+
use_fp8_w8a8=use_fp8_w8a8,
|
774 |
+
use_int8_w8a16=use_int8_w8a16,
|
775 |
+
**config,
|
776 |
+
)
|
777 |
|
778 |
|
779 |
+
# Adapted from: https://github.com/sgl-project/sglang/pull/2628
|
780 |
+
def get_config_file_name(
|
781 |
+
E: int, N: int, dtype: Optional[str], block_shape: Optional[List[int]] = None
|
782 |
+
) -> str:
|
783 |
device_name = current_platform.get_device_name().replace(" ", "_")
|
784 |
dtype_selector = "" if not dtype else f",dtype={dtype}"
|
785 |
+
block_shape_selector = (
|
786 |
+
"" if not block_shape or not all(block_shape) else f",block_shape={block_shape}"
|
787 |
+
)
|
788 |
+
return f"E={E},N={N},device_name={device_name}{dtype_selector}{block_shape_selector}.json" # noqa: E501
|
789 |
|
790 |
|
791 |
+
# Adapted from: https://github.com/sgl-project/sglang/pull/2628
|
792 |
@functools.lru_cache
|
793 |
+
def get_moe_configs(
|
794 |
+
E: int,
|
795 |
+
N: int,
|
796 |
+
dtype: Optional[str],
|
797 |
+
block_n: Optional[int] = None,
|
798 |
+
block_k: Optional[int] = None,
|
799 |
+
) -> Optional[Dict[int, Any]]:
|
800 |
"""
|
801 |
Return optimized configurations for the fused MoE kernel.
|
802 |
|
|
|
808 |
|
809 |
# First look up if an optimized configuration is available in the configs
|
810 |
# directory
|
811 |
+
block_shape = [block_n, block_k] if block_n and block_k else None
|
812 |
+
json_file_name = get_config_file_name(E, N, dtype, block_shape)
|
813 |
|
814 |
config_file_path = os.path.join(
|
815 |
os.path.dirname(os.path.realpath(__file__)), "configs", json_file_name
|
816 |
)
|
817 |
if os.path.exists(config_file_path):
|
818 |
with open(config_file_path) as f:
|
819 |
+
logger.info("Using configuration from %s for MoE layer.", config_file_path)
|
820 |
# If a configuration has been found, return it
|
821 |
return {int(key): val for key, val in json.load(f).items()}
|
822 |
|
823 |
# If no optimized configuration is available, we will use the default
|
824 |
# configuration
|
825 |
+
logger.warning(
|
826 |
+
(
|
827 |
+
"Using default MoE config. Performance might be sub-optimal! "
|
828 |
+
"Config file not found at %s"
|
829 |
+
),
|
830 |
+
config_file_path,
|
831 |
+
)
|
832 |
return None
|
833 |
|
834 |
|
|
|
840 |
topk: int,
|
841 |
dtype: Optional[str],
|
842 |
is_marlin: bool,
|
843 |
+
block_shape: Optional[List[int]] = None,
|
844 |
) -> Dict[str, int]:
|
845 |
+
if dtype == "fp8_w8a8" and block_shape is not None:
|
846 |
+
# Block-wise quant: BLOCK_SIZE_N must be divisible by block_shape[0]
|
847 |
+
# BLOCK_SIZE_K must be divisible by block_shape[1]
|
|
|
|
|
|
|
|
|
|
|
848 |
config = {
|
849 |
+
"BLOCK_SIZE_M": 64,
|
850 |
+
"BLOCK_SIZE_N": block_shape[0],
|
851 |
+
"BLOCK_SIZE_K": block_shape[1],
|
852 |
+
"GROUP_SIZE_M": 32,
|
853 |
+
"num_warps": 4,
|
854 |
+
"num_stages": 3,
|
855 |
}
|
856 |
+
else:
|
857 |
+
config = {
|
858 |
+
"BLOCK_SIZE_M": 64,
|
859 |
+
"BLOCK_SIZE_N": 64,
|
860 |
+
"BLOCK_SIZE_K": 32,
|
861 |
+
"GROUP_SIZE_M": 8,
|
862 |
+
}
|
863 |
+
# A heuristic: fused marlin works faster with this config for small M
|
864 |
+
if M <= E or (is_marlin and M <= 32):
|
865 |
+
config = {
|
866 |
+
"BLOCK_SIZE_M": 16,
|
867 |
+
"BLOCK_SIZE_N": 32,
|
868 |
+
"BLOCK_SIZE_K": 64,
|
869 |
+
"GROUP_SIZE_M": 1,
|
870 |
+
}
|
871 |
return config
|
872 |
|
873 |
|
|
|
877 |
top_k: int,
|
878 |
dtype: Optional[str],
|
879 |
M: int,
|
|
|
880 |
is_marlin: bool = False,
|
881 |
+
block_shape: Optional[List[int]] = None,
|
882 |
):
|
883 |
+
# from vllm.model_executor.layers.fused_moe import get_config
|
884 |
+
# TODO: removed when syncing to vLLM, do we need this?
|
885 |
+
# override_config = get_config()
|
886 |
+
override_config = None
|
887 |
if override_config:
|
888 |
config = override_config
|
889 |
else:
|
890 |
# First try to load optimal config from the file
|
891 |
E, _, N = w2_shape
|
892 |
+
block_n = block_shape[0] if block_shape else 0
|
893 |
+
block_k = block_shape[1] if block_shape else 0
|
894 |
+
configs = get_moe_configs(E, N, dtype, block_n, block_k)
|
895 |
|
896 |
if configs:
|
897 |
# If an optimal configuration map has been found, look up the
|
|
|
899 |
config = configs[min(configs.keys(), key=lambda x: abs(x - M))]
|
900 |
else:
|
901 |
# Else use the default config
|
902 |
+
config = get_default_config(
|
903 |
+
M, E, N, w1_shape[2], top_k, dtype, is_marlin, block_shape
|
904 |
+
)
|
905 |
return config
|
906 |
|
907 |
|
|
|
937 |
return topk_weights, topk_ids
|
938 |
|
939 |
|
940 |
+
# This is used by the Deepseek-V2 and Deepseek-V3 model
|
941 |
+
@torch.compile(dynamic=True, backend=current_platform.simple_compile_backend)
|
942 |
def grouped_topk(
|
943 |
hidden_states: torch.Tensor,
|
944 |
gating_output: torch.Tensor,
|
|
|
946 |
renormalize: bool,
|
947 |
num_expert_group: int = 0,
|
948 |
topk_group: int = 0,
|
949 |
+
scoring_func: str = "softmax",
|
950 |
+
e_score_correction_bias: Optional[torch.Tensor] = None,
|
951 |
):
|
952 |
|
953 |
assert hidden_states.shape[0] == gating_output.shape[0], "Number of tokens mismatch"
|
954 |
|
955 |
+
if scoring_func == "softmax":
|
956 |
+
scores = torch.softmax(gating_output, dim=-1)
|
957 |
+
elif scoring_func == "sigmoid":
|
958 |
+
scores = gating_output.sigmoid()
|
959 |
+
else:
|
960 |
+
raise ValueError(f"Unsupported scoring function: {scoring_func}")
|
961 |
+
|
962 |
+
if e_score_correction_bias is not None:
|
963 |
+
# Store original scores before applying correction bias. We use biased
|
964 |
+
# scores for expert selection but original scores for routing weights
|
965 |
+
original_scores = scores
|
966 |
+
scores = scores + e_score_correction_bias.unsqueeze(0)
|
967 |
+
|
968 |
num_token = scores.shape[0]
|
969 |
group_scores = (
|
970 |
scores.view(num_token, num_expert_group, -1).max(dim=-1).values
|
|
|
980 |
.reshape(num_token, -1)
|
981 |
) # [n, e]
|
982 |
tmp_scores = scores.masked_fill(~score_mask.bool(), 0.0) # [n, e]
|
983 |
+
|
984 |
+
if e_score_correction_bias is not None:
|
985 |
+
topk_ids = torch.topk(tmp_scores, k=topk, dim=-1, sorted=False)[1]
|
986 |
+
# Use original unbiased scores for the routing weights
|
987 |
+
topk_weights = original_scores.gather(1, topk_ids)
|
988 |
+
else:
|
989 |
+
topk_weights, topk_ids = torch.topk(tmp_scores, k=topk, dim=-1, sorted=False)
|
990 |
|
991 |
if renormalize:
|
992 |
topk_weights = topk_weights / topk_weights.sum(dim=-1, keepdim=True)
|
|
|
996 |
|
997 |
def get_config_dtype_str(
|
998 |
dtype: torch.dtype,
|
999 |
+
use_int4_w4a16: Optional[bool] = False,
|
1000 |
use_int8_w8a16: Optional[bool] = False,
|
1001 |
use_fp8_w8a8: Optional[bool] = False,
|
1002 |
):
|
|
|
1004 |
return "fp8_w8a8"
|
1005 |
elif use_int8_w8a16:
|
1006 |
return "int8_w8a16"
|
1007 |
+
elif use_int4_w4a16:
|
1008 |
+
return "int4_w8a16"
|
1009 |
elif dtype == torch.float:
|
1010 |
# avoiding cases where kernel fails when float32 MoE
|
1011 |
# use fp16/bfloat16 configs
|
|
|
1013 |
return None
|
1014 |
|
1015 |
|
1016 |
+
def inplace_fused_experts(
|
1017 |
+
hidden_states: torch.Tensor,
|
1018 |
+
w1: torch.Tensor,
|
1019 |
+
w2: torch.Tensor,
|
1020 |
+
topk_weights: torch.Tensor,
|
1021 |
+
topk_ids: torch.Tensor,
|
1022 |
+
use_fp8_w8a8: bool = False,
|
1023 |
+
use_int8_w8a16: bool = False,
|
1024 |
+
use_int4_w4a16: bool = False,
|
1025 |
+
w1_scale: Optional[torch.Tensor] = None,
|
1026 |
+
w2_scale: Optional[torch.Tensor] = None,
|
1027 |
+
w1_zp: Optional[torch.Tensor] = None,
|
1028 |
+
w2_zp: Optional[torch.Tensor] = None,
|
1029 |
+
a1_scale: Optional[torch.Tensor] = None,
|
1030 |
+
a2_scale: Optional[torch.Tensor] = None,
|
1031 |
+
block_shape: Optional[List[int]] = None,
|
1032 |
+
) -> None:
|
1033 |
+
fused_experts_impl(
|
1034 |
+
hidden_states,
|
1035 |
+
w1,
|
1036 |
+
w2,
|
1037 |
+
topk_weights,
|
1038 |
+
topk_ids,
|
1039 |
+
True,
|
1040 |
+
use_fp8_w8a8,
|
1041 |
+
use_int8_w8a16,
|
1042 |
+
use_int4_w4a16,
|
1043 |
+
w1_scale,
|
1044 |
+
w2_scale,
|
1045 |
+
w1_zp,
|
1046 |
+
w2_zp,
|
1047 |
+
a1_scale,
|
1048 |
+
a2_scale,
|
1049 |
+
block_shape,
|
1050 |
+
)
|
1051 |
+
|
1052 |
+
|
1053 |
+
def outplace_fused_experts(
|
1054 |
+
hidden_states: torch.Tensor,
|
1055 |
+
w1: torch.Tensor,
|
1056 |
+
w2: torch.Tensor,
|
1057 |
+
topk_weights: torch.Tensor,
|
1058 |
+
topk_ids: torch.Tensor,
|
1059 |
+
use_fp8_w8a8: bool = False,
|
1060 |
+
use_int8_w8a16: bool = False,
|
1061 |
+
use_int4_w4a16: bool = False,
|
1062 |
+
w1_scale: Optional[torch.Tensor] = None,
|
1063 |
+
w2_scale: Optional[torch.Tensor] = None,
|
1064 |
+
w1_zp: Optional[torch.Tensor] = None,
|
1065 |
+
w2_zp: Optional[torch.Tensor] = None,
|
1066 |
+
a1_scale: Optional[torch.Tensor] = None,
|
1067 |
+
a2_scale: Optional[torch.Tensor] = None,
|
1068 |
+
block_shape: Optional[List[int]] = None,
|
1069 |
+
) -> torch.Tensor:
|
1070 |
+
return fused_experts_impl(
|
1071 |
+
hidden_states,
|
1072 |
+
w1,
|
1073 |
+
w2,
|
1074 |
+
topk_weights,
|
1075 |
+
topk_ids,
|
1076 |
+
False,
|
1077 |
+
use_fp8_w8a8,
|
1078 |
+
use_int8_w8a16,
|
1079 |
+
use_int4_w4a16,
|
1080 |
+
w1_scale,
|
1081 |
+
w2_scale,
|
1082 |
+
w1_zp,
|
1083 |
+
w2_zp,
|
1084 |
+
a1_scale,
|
1085 |
+
a2_scale,
|
1086 |
+
block_shape,
|
1087 |
+
)
|
1088 |
+
|
1089 |
+
|
1090 |
def fused_experts(
|
1091 |
hidden_states: torch.Tensor,
|
1092 |
w1: torch.Tensor,
|
|
|
1094 |
topk_weights: torch.Tensor,
|
1095 |
topk_ids: torch.Tensor,
|
1096 |
inplace: bool = False,
|
|
|
1097 |
use_fp8_w8a8: bool = False,
|
1098 |
use_int8_w8a16: bool = False,
|
1099 |
+
use_int4_w4a16: bool = False,
|
1100 |
+
w1_scale: Optional[torch.Tensor] = None,
|
1101 |
+
w2_scale: Optional[torch.Tensor] = None,
|
1102 |
+
w1_zp: Optional[torch.Tensor] = None,
|
1103 |
+
w2_zp: Optional[torch.Tensor] = None,
|
1104 |
+
a1_scale: Optional[torch.Tensor] = None,
|
1105 |
+
a2_scale: Optional[torch.Tensor] = None,
|
1106 |
+
block_shape: Optional[List[int]] = None,
|
1107 |
+
):
|
1108 |
+
if inplace:
|
1109 |
+
inplace_fused_experts(
|
1110 |
+
hidden_states,
|
1111 |
+
w1,
|
1112 |
+
w2,
|
1113 |
+
topk_weights,
|
1114 |
+
topk_ids,
|
1115 |
+
use_fp8_w8a8,
|
1116 |
+
use_int8_w8a16,
|
1117 |
+
use_int4_w4a16,
|
1118 |
+
w1_scale,
|
1119 |
+
w2_scale,
|
1120 |
+
w1_zp,
|
1121 |
+
w2_zp,
|
1122 |
+
a1_scale,
|
1123 |
+
a2_scale,
|
1124 |
+
block_shape,
|
1125 |
+
)
|
1126 |
+
return hidden_states
|
1127 |
+
else:
|
1128 |
+
return outplace_fused_experts(
|
1129 |
+
hidden_states,
|
1130 |
+
w1,
|
1131 |
+
w2,
|
1132 |
+
topk_weights,
|
1133 |
+
topk_ids,
|
1134 |
+
use_fp8_w8a8,
|
1135 |
+
use_int8_w8a16,
|
1136 |
+
use_int4_w4a16,
|
1137 |
+
w1_scale,
|
1138 |
+
w2_scale,
|
1139 |
+
w1_zp,
|
1140 |
+
w2_zp,
|
1141 |
+
a1_scale,
|
1142 |
+
a2_scale,
|
1143 |
+
block_shape,
|
1144 |
+
)
|
1145 |
+
|
1146 |
+
|
1147 |
+
def fused_experts_impl(
|
1148 |
+
hidden_states: torch.Tensor,
|
1149 |
+
w1: torch.Tensor,
|
1150 |
+
w2: torch.Tensor,
|
1151 |
+
topk_weights: torch.Tensor,
|
1152 |
+
topk_ids: torch.Tensor,
|
1153 |
+
inplace: bool = False,
|
1154 |
+
use_fp8_w8a8: bool = False,
|
1155 |
+
use_int8_w8a16: bool = False,
|
1156 |
+
use_int4_w4a16: bool = False,
|
1157 |
w1_scale: Optional[torch.Tensor] = None,
|
1158 |
w2_scale: Optional[torch.Tensor] = None,
|
1159 |
+
w1_zp: Optional[torch.Tensor] = None,
|
1160 |
+
w2_zp: Optional[torch.Tensor] = None,
|
1161 |
a1_scale: Optional[torch.Tensor] = None,
|
1162 |
a2_scale: Optional[torch.Tensor] = None,
|
1163 |
+
block_shape: Optional[List[int]] = None,
|
1164 |
):
|
1165 |
# Check constraints.
|
1166 |
+
if use_int4_w4a16:
|
1167 |
+
assert hidden_states.shape[1] // 2 == w1.shape[2], "Hidden size mismatch"
|
1168 |
+
else:
|
1169 |
+
assert hidden_states.shape[1] == w1.shape[2], "Hidden size mismatch"
|
1170 |
+
|
1171 |
assert topk_weights.shape == topk_ids.shape, "topk shape mismatch"
|
1172 |
assert hidden_states.is_contiguous(), "Hidden_states must be contiguous"
|
1173 |
assert w1.is_contiguous(), "Expert weights1 must be contiguous"
|
|
|
1183 |
config_dtype = get_config_dtype_str(
|
1184 |
use_fp8_w8a8=use_fp8_w8a8,
|
1185 |
use_int8_w8a16=use_int8_w8a16,
|
1186 |
+
use_int4_w4a16=use_int4_w4a16,
|
1187 |
dtype=hidden_states.dtype,
|
1188 |
)
|
1189 |
|
|
|
1193 |
w2.shape,
|
1194 |
topk_ids.shape[1],
|
1195 |
config_dtype,
|
1196 |
+
block_shape=block_shape,
|
1197 |
)
|
1198 |
|
1199 |
config = get_config_func(M)
|
|
|
1214 |
dtype=hidden_states.dtype,
|
1215 |
)
|
1216 |
|
1217 |
+
if hidden_states.dtype == torch.bfloat16:
|
1218 |
+
compute_type = tl.bfloat16
|
1219 |
+
elif hidden_states.dtype == torch.float16:
|
1220 |
+
compute_type = tl.float16
|
1221 |
+
elif hidden_states.dtype == torch.float32:
|
1222 |
+
compute_type = tl.float32
|
1223 |
+
else:
|
1224 |
+
raise ValueError(f"Unsupported compute_type: {hidden_states.dtype}")
|
1225 |
|
1226 |
if inplace:
|
1227 |
out_hidden_states = hidden_states
|
|
|
1262 |
intermediate_cache1,
|
1263 |
a1_scale,
|
1264 |
w1_scale,
|
1265 |
+
w1_zp,
|
1266 |
curr_topk_weights,
|
1267 |
curr_topk_ids,
|
1268 |
sorted_token_ids,
|
|
|
1274 |
compute_type=compute_type,
|
1275 |
use_fp8_w8a8=use_fp8_w8a8,
|
1276 |
use_int8_w8a16=use_int8_w8a16,
|
1277 |
+
use_int4_w4a16=use_int4_w4a16,
|
1278 |
+
block_shape=block_shape,
|
1279 |
)
|
1280 |
|
1281 |
ops.silu_and_mul(intermediate_cache2, intermediate_cache1.view(-1, N))
|
|
|
1286 |
intermediate_cache3,
|
1287 |
a2_scale,
|
1288 |
w2_scale,
|
1289 |
+
w2_zp,
|
1290 |
curr_topk_weights,
|
1291 |
curr_topk_ids,
|
1292 |
sorted_token_ids,
|
|
|
1298 |
compute_type=compute_type,
|
1299 |
use_fp8_w8a8=use_fp8_w8a8,
|
1300 |
use_int8_w8a16=use_int8_w8a16,
|
1301 |
+
use_int4_w4a16=use_int4_w4a16,
|
1302 |
+
block_shape=block_shape,
|
1303 |
)
|
1304 |
|
1305 |
ops.moe_sum(
|
|
|
1317 |
topk: int,
|
1318 |
renormalize: bool,
|
1319 |
inplace: bool = False,
|
|
|
1320 |
use_grouped_topk: bool = False,
|
1321 |
num_expert_group: Optional[int] = None,
|
1322 |
topk_group: Optional[int] = None,
|
1323 |
custom_routing_function: Optional[Callable] = None,
|
1324 |
use_fp8_w8a8: bool = False,
|
1325 |
use_int8_w8a16: bool = False,
|
1326 |
+
use_int4_w4a16: bool = False,
|
1327 |
w1_scale: Optional[torch.Tensor] = None,
|
1328 |
w2_scale: Optional[torch.Tensor] = None,
|
1329 |
+
w1_zp: Optional[torch.Tensor] = None,
|
1330 |
+
w2_zp: Optional[torch.Tensor] = None,
|
1331 |
a1_scale: Optional[torch.Tensor] = None,
|
1332 |
a2_scale: Optional[torch.Tensor] = None,
|
1333 |
+
block_shape: Optional[List[int]] = None,
|
1334 |
) -> torch.Tensor:
|
1335 |
"""
|
1336 |
This function computes a Mixture of Experts (MoE) layer using two sets of
|
|
|
1346 |
- renormalize (bool): If True, renormalize the top-k weights to sum to 1.
|
1347 |
- inplace (bool): If True, perform the operation in-place.
|
1348 |
Defaults to False.
|
|
|
|
|
1349 |
- num_expert_group: Optional[int]: additional parameter for grouped_topk
|
1350 |
- topk_group: Optional[int]: additional parameter for grouped_topk
|
1351 |
- use_grouped_topk: If True, use grouped_topk instead of fused_topk
|
1352 |
note: Deepseekv2 model uses grouped_topk
|
1353 |
- use_fp8_w8a8 (bool): If True, use fp8 arithmetic to compute the inner
|
1354 |
products for w1 and w2. Defaults to False.
|
1355 |
+
- use_int8_w8a16 (bool): If True, use matmul of int8 weight and bf16/fp16
|
1356 |
+
activation to compute the inner products for w1 and w2.
|
1357 |
+
Defaults to False.
|
1358 |
+
- use_int4_w4a16 (bool): If True, use matmul of int4 weight and bf16/fp16
|
1359 |
+
activation to compute the inner products for w1 and w2.
|
1360 |
+
Defaults to False.
|
1361 |
- w1_scale (Optional[torch.Tensor]): Optional scale to be used for
|
1362 |
w1.
|
1363 |
- w2_scale (Optional[torch.Tensor]): Optional scale to be used for
|
1364 |
w2.
|
1365 |
+
- a1_scale (Optional[torch.Tensor]): Optional scale to be used for
|
1366 |
+
a1.
|
1367 |
+
- a2_scale (Optional[torch.Tensor]): Optional scale to be used for
|
1368 |
+
a2.
|
1369 |
+
- block_shape: (Optional[List[int]]): Optional block size for block-wise
|
1370 |
+
quantization.
|
1371 |
|
1372 |
Returns:
|
1373 |
- torch.Tensor: The output tensor after applying the MoE layer.
|
|
|
1401 |
topk_weights,
|
1402 |
topk_ids,
|
1403 |
inplace=inplace,
|
|
|
1404 |
use_fp8_w8a8=use_fp8_w8a8,
|
1405 |
use_int8_w8a16=use_int8_w8a16,
|
1406 |
+
use_int4_w4a16=use_int4_w4a16,
|
1407 |
w1_scale=w1_scale,
|
1408 |
w2_scale=w2_scale,
|
1409 |
+
w1_zp=w1_zp,
|
1410 |
+
w2_zp=w2_zp,
|
1411 |
a1_scale=a1_scale,
|
1412 |
a2_scale=a2_scale,
|
1413 |
+
block_shape=block_shape,
|
1414 |
)
|
build/torch25-cxx11-cu121-x86_64-linux/moe/platforms.py
CHANGED
@@ -1,22 +1,32 @@
|
|
1 |
-
from
|
2 |
-
import os
|
3 |
-
from functools import lru_cache, wraps
|
4 |
|
5 |
import torch
|
6 |
|
7 |
IS_ROCM = torch.version.hip is not None
|
8 |
|
9 |
-
|
|
|
|
|
|
|
|
|
|
|
10 |
@classmethod
|
11 |
@lru_cache(maxsize=8)
|
12 |
def get_device_name(cls, device_id: int = 0) -> str:
|
13 |
return torch.cuda.get_device_name(0)
|
14 |
|
15 |
-
|
|
|
|
|
|
|
|
|
16 |
@classmethod
|
17 |
@lru_cache(maxsize=8)
|
18 |
def get_device_name(cls, device_id: int = 0) -> str:
|
19 |
return torch.cuda.get_device_name(device_id)
|
20 |
|
|
|
|
|
|
|
21 |
|
22 |
current_platform = RocmPlatform() if IS_ROCM else CudaPlatform()
|
|
|
1 |
+
from functools import lru_cache
|
|
|
|
|
2 |
|
3 |
import torch
|
4 |
|
5 |
IS_ROCM = torch.version.hip is not None
|
6 |
|
7 |
+
|
8 |
+
class Platform:
|
9 |
+
simple_compile_backend: str = "inductor"
|
10 |
+
|
11 |
+
|
12 |
+
class CudaPlatform(Platform):
|
13 |
@classmethod
|
14 |
@lru_cache(maxsize=8)
|
15 |
def get_device_name(cls, device_id: int = 0) -> str:
|
16 |
return torch.cuda.get_device_name(0)
|
17 |
|
18 |
+
def is_rocm(self):
|
19 |
+
return False
|
20 |
+
|
21 |
+
|
22 |
+
class RocmPlatform(Platform):
|
23 |
@classmethod
|
24 |
@lru_cache(maxsize=8)
|
25 |
def get_device_name(cls, device_id: int = 0) -> str:
|
26 |
return torch.cuda.get_device_name(device_id)
|
27 |
|
28 |
+
def is_rocm(self):
|
29 |
+
return True
|
30 |
+
|
31 |
|
32 |
current_platform = RocmPlatform() if IS_ROCM else CudaPlatform()
|
build/torch25-cxx11-cu124-x86_64-linux/moe/{_moe_lwzoz7knnxf4i.abi3.so → _moe_pss5doo675cd4.abi3.so}
RENAMED
@@ -1,3 +1,3 @@
|
|
1 |
version https://git-lfs.github.com/spec/v1
|
2 |
-
oid sha256:
|
3 |
-
size
|
|
|
1 |
version https://git-lfs.github.com/spec/v1
|
2 |
+
oid sha256:041c922d7e435dbc7ca974c331455f02ed43ecd4adcd859dd8ee593cfea676e3
|
3 |
+
size 85733000
|
build/torch25-cxx11-cu124-x86_64-linux/moe/_ops.py
CHANGED
@@ -1,9 +1,9 @@
|
|
1 |
import torch
|
2 |
-
from . import
|
3 |
-
ops = torch.ops.
|
4 |
|
5 |
def add_op_namespace_prefix(op_name: str):
|
6 |
"""
|
7 |
Prefix op by namespace.
|
8 |
"""
|
9 |
-
return f"
|
|
|
1 |
import torch
|
2 |
+
from . import _moe_pss5doo675cd4
|
3 |
+
ops = torch.ops._moe_pss5doo675cd4
|
4 |
|
5 |
def add_op_namespace_prefix(op_name: str):
|
6 |
"""
|
7 |
Prefix op by namespace.
|
8 |
"""
|
9 |
+
return f"_moe_pss5doo675cd4::{op_name}"
|
build/torch25-cxx11-cu124-x86_64-linux/moe/fp8.py
CHANGED
@@ -1,6 +1,11 @@
|
|
|
|
|
|
1 |
import torch
|
|
|
|
|
2 |
|
3 |
-
|
|
|
4 |
|
5 |
|
6 |
def is_hip() -> bool:
|
@@ -49,15 +54,179 @@ def scaled_fp8_quant(
|
|
49 |
if scale is None:
|
50 |
if use_per_token_if_dynamic:
|
51 |
scale = torch.empty((shape[0], 1), device=input.device, dtype=torch.float32)
|
52 |
-
|
53 |
-
output, input, scale, scale_ub
|
54 |
-
)
|
55 |
else:
|
56 |
scale = torch.zeros(1, device=input.device, dtype=torch.float32)
|
57 |
-
|
58 |
else:
|
59 |
# num_token_padding not implemented for this case
|
60 |
assert scale.numel() == 1 or num_token_padding is None
|
61 |
-
|
62 |
|
63 |
return output, scale
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
1 |
+
from typing import Tuple, Optional, Union
|
2 |
+
|
3 |
import torch
|
4 |
+
import triton
|
5 |
+
import triton.language as tl
|
6 |
|
7 |
+
|
8 |
+
from ._ops import ops
|
9 |
|
10 |
|
11 |
def is_hip() -> bool:
|
|
|
54 |
if scale is None:
|
55 |
if use_per_token_if_dynamic:
|
56 |
scale = torch.empty((shape[0], 1), device=input.device, dtype=torch.float32)
|
57 |
+
ops.dynamic_per_token_scaled_fp8_quant(output, input, scale, scale_ub)
|
|
|
|
|
58 |
else:
|
59 |
scale = torch.zeros(1, device=input.device, dtype=torch.float32)
|
60 |
+
ops.dynamic_scaled_fp8_quant(output, input, scale)
|
61 |
else:
|
62 |
# num_token_padding not implemented for this case
|
63 |
assert scale.numel() == 1 or num_token_padding is None
|
64 |
+
ops.static_scaled_fp8_quant(output, input, scale)
|
65 |
|
66 |
return output, scale
|
67 |
+
|
68 |
+
|
69 |
+
@triton.jit
|
70 |
+
def _per_token_group_quant_fp8(
|
71 |
+
# Pointers to inputs and output
|
72 |
+
y_ptr,
|
73 |
+
y_q_ptr,
|
74 |
+
y_s_ptr,
|
75 |
+
group_size,
|
76 |
+
# Avoid to divide zero
|
77 |
+
eps,
|
78 |
+
# Information for float8
|
79 |
+
fp8_min,
|
80 |
+
fp8_max,
|
81 |
+
# Meta-parameters
|
82 |
+
BLOCK: tl.constexpr,
|
83 |
+
):
|
84 |
+
"""A Triton-accelerated function to perform per-token-group
|
85 |
+
quantization on a tensor.
|
86 |
+
This function converts the tensor values into float8 values.
|
87 |
+
"""
|
88 |
+
# Map the program id to the row of X and Y it should compute.
|
89 |
+
g_id = tl.program_id(0)
|
90 |
+
y_ptr += g_id * group_size
|
91 |
+
y_q_ptr += g_id * group_size
|
92 |
+
y_s_ptr += g_id
|
93 |
+
|
94 |
+
cols = tl.arange(0, BLOCK) # N <= BLOCK
|
95 |
+
mask = cols < group_size
|
96 |
+
|
97 |
+
y = tl.load(y_ptr + cols, mask=mask, other=0.0).to(tl.float32)
|
98 |
+
# Quant
|
99 |
+
_absmax = tl.maximum(tl.max(tl.abs(y)), eps)
|
100 |
+
y_s = _absmax / fp8_max
|
101 |
+
y_q = tl.clamp(y / y_s, fp8_min, fp8_max).to(y_q_ptr.dtype.element_ty)
|
102 |
+
|
103 |
+
tl.store(y_q_ptr + cols, y_q, mask=mask)
|
104 |
+
tl.store(y_s_ptr, y_s)
|
105 |
+
|
106 |
+
|
107 |
+
@triton.jit
|
108 |
+
def _per_token_group_quant_fp8_colmajor(
|
109 |
+
# Pointers to inputs and output
|
110 |
+
y_ptr,
|
111 |
+
y_q_ptr,
|
112 |
+
y_s_ptr,
|
113 |
+
group_size,
|
114 |
+
# Num columns of y
|
115 |
+
y_num_columns,
|
116 |
+
# Stride from one column to the next of y_s
|
117 |
+
y_s_col_stride,
|
118 |
+
# Avoid to divide zero
|
119 |
+
eps,
|
120 |
+
# Information for float8
|
121 |
+
fp8_min,
|
122 |
+
fp8_max,
|
123 |
+
# Meta-parameters
|
124 |
+
BLOCK: tl.constexpr,
|
125 |
+
):
|
126 |
+
"""A Triton-accelerated function to perform per-token-group
|
127 |
+
quantization on a tensor.
|
128 |
+
This function converts the tensor values into float8 values.
|
129 |
+
"""
|
130 |
+
# Map the program id to the row of X and Y it should compute.
|
131 |
+
g_id = tl.program_id(0)
|
132 |
+
y_ptr += g_id * group_size
|
133 |
+
y_q_ptr += g_id * group_size
|
134 |
+
|
135 |
+
# Convert g_id the flattened block coordinate to 2D so we can index
|
136 |
+
# into the output y_scales matrix
|
137 |
+
blocks_per_row = y_num_columns // group_size
|
138 |
+
scale_col = g_id % blocks_per_row
|
139 |
+
scale_row = g_id // blocks_per_row
|
140 |
+
y_s_ptr += scale_col * y_s_col_stride + scale_row
|
141 |
+
|
142 |
+
cols = tl.arange(0, BLOCK) # group_size <= BLOCK
|
143 |
+
mask = cols < group_size
|
144 |
+
|
145 |
+
y = tl.load(y_ptr + cols, mask=mask, other=0.0).to(tl.float32)
|
146 |
+
# Quant
|
147 |
+
_absmax = tl.maximum(tl.max(tl.abs(y)), eps)
|
148 |
+
y_s = _absmax / fp8_max
|
149 |
+
y_q = tl.clamp(y / y_s, fp8_min, fp8_max).to(y_q_ptr.dtype.element_ty)
|
150 |
+
|
151 |
+
tl.store(y_q_ptr + cols, y_q, mask=mask)
|
152 |
+
tl.store(y_s_ptr, y_s)
|
153 |
+
|
154 |
+
|
155 |
+
def per_token_group_quant_fp8(
|
156 |
+
x: torch.Tensor,
|
157 |
+
group_size: int,
|
158 |
+
eps: float = 1e-10,
|
159 |
+
dtype: Optional[torch.dtype] = None,
|
160 |
+
column_major_scales: bool = False,
|
161 |
+
) -> Tuple[torch.Tensor, torch.Tensor]:
|
162 |
+
"""Function to perform per-token-group quantization on an input tensor `x`.
|
163 |
+
It converts the tensor values into signed float8 values and returns the
|
164 |
+
quantized tensor along with the scaling factor used for quantization.
|
165 |
+
Args:
|
166 |
+
x: The input tensor with ndim >= 2.
|
167 |
+
group_size: The group size used for quantization.
|
168 |
+
eps: The minimum to avoid dividing zero.
|
169 |
+
dtype: The dype of output tensor. Note that only `torch.float8_e4m3fn`
|
170 |
+
is supported for now.
|
171 |
+
Returns:
|
172 |
+
Tuple[torch.Tensor, torch.Tensor]: The quantized tensor and the
|
173 |
+
scaling factor for quantization.
|
174 |
+
"""
|
175 |
+
if dtype is None:
|
176 |
+
dtype = (
|
177 |
+
torch.float8_e4m3fnuz if current_platform.is_rocm() else torch.float8_e4m3fn
|
178 |
+
)
|
179 |
+
assert x.shape[-1] % group_size == 0, (
|
180 |
+
f"the last dimension of `x` {x.shape[-1]} must be divisible "
|
181 |
+
f"by `group_size` {group_size}"
|
182 |
+
)
|
183 |
+
assert x.is_contiguous(), "`x` must be contiguous"
|
184 |
+
|
185 |
+
finfo = torch.finfo(dtype)
|
186 |
+
fp8_min = finfo.min
|
187 |
+
fp8_max = finfo.max
|
188 |
+
|
189 |
+
x_q = torch.empty_like(x, device=x.device, dtype=dtype)
|
190 |
+
M = x.numel() // group_size
|
191 |
+
N = group_size
|
192 |
+
if column_major_scales:
|
193 |
+
shape = (x.shape[-1] // group_size,) + x.shape[:-1]
|
194 |
+
x_s = torch.empty(shape, device=x.device, dtype=torch.float32).permute(-1, -2)
|
195 |
+
else:
|
196 |
+
shape = x.shape[:-1] + (x.shape[-1] // group_size,)
|
197 |
+
x_s = torch.empty(shape, device=x.device, dtype=torch.float32)
|
198 |
+
|
199 |
+
BLOCK = triton.next_power_of_2(N)
|
200 |
+
# heuristics for number of warps
|
201 |
+
num_warps = min(max(BLOCK // 256, 1), 8)
|
202 |
+
num_stages = 1
|
203 |
+
if column_major_scales:
|
204 |
+
_per_token_group_quant_fp8_colmajor[(M,)](
|
205 |
+
x,
|
206 |
+
x_q,
|
207 |
+
x_s,
|
208 |
+
group_size,
|
209 |
+
x.shape[1],
|
210 |
+
x_s.stride(1),
|
211 |
+
eps,
|
212 |
+
fp8_min=fp8_min,
|
213 |
+
fp8_max=fp8_max,
|
214 |
+
BLOCK=BLOCK,
|
215 |
+
num_warps=num_warps,
|
216 |
+
num_stages=num_stages,
|
217 |
+
)
|
218 |
+
else:
|
219 |
+
_per_token_group_quant_fp8[(M,)](
|
220 |
+
x,
|
221 |
+
x_q,
|
222 |
+
x_s,
|
223 |
+
group_size,
|
224 |
+
eps,
|
225 |
+
fp8_min=fp8_min,
|
226 |
+
fp8_max=fp8_max,
|
227 |
+
BLOCK=BLOCK,
|
228 |
+
num_warps=num_warps,
|
229 |
+
num_stages=num_stages,
|
230 |
+
)
|
231 |
+
|
232 |
+
return x_q, x_s
|
build/torch25-cxx11-cu124-x86_64-linux/moe/fused_marlin_moe.py
CHANGED
@@ -40,7 +40,6 @@ def single_marlin_moe(
|
|
40 |
g_idx: Optional[torch.Tensor] = None,
|
41 |
sort_indices: Optional[torch.Tensor] = None,
|
42 |
w_zeros: Optional[torch.Tensor] = None,
|
43 |
-
override_config: Optional[Dict[str, Any]] = None,
|
44 |
num_bits: int = 8,
|
45 |
is_k_full: bool = True,
|
46 |
) -> torch.Tensor:
|
@@ -61,8 +60,6 @@ def single_marlin_moe(
|
|
61 |
- topk (int): The number of top-k experts to select.
|
62 |
- renormalize (bool): If True, renormalize the top-k weights to sum to 1.
|
63 |
- w_zeros (Optional[torch.Tensor]): Optional zero points to be used for w.
|
64 |
-
- override_config (Optional[Dict[str, Any]]): Optional override
|
65 |
-
for the kernel configuration.
|
66 |
- num_bits (bool): The number of bits in expert weights quantization.
|
67 |
|
68 |
Returns:
|
@@ -90,7 +87,6 @@ def single_marlin_moe(
|
|
90 |
w.shape,
|
91 |
topk_ids.shape[1],
|
92 |
None,
|
93 |
-
override_config=override_config,
|
94 |
is_marlin=True,
|
95 |
)
|
96 |
config = get_config_func(M)
|
@@ -154,6 +150,25 @@ def single_marlin_moe(
|
|
154 |
return torch.sum(intermediate_cache.view(*intermediate_cache.shape), dim=1)
|
155 |
|
156 |
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
157 |
def fused_marlin_moe(
|
158 |
hidden_states: torch.Tensor,
|
159 |
w1: torch.Tensor,
|
@@ -169,7 +184,6 @@ def fused_marlin_moe(
|
|
169 |
sort_indices2: Optional[torch.Tensor] = None,
|
170 |
w1_zeros: Optional[torch.Tensor] = None,
|
171 |
w2_zeros: Optional[torch.Tensor] = None,
|
172 |
-
override_config: Optional[Dict[str, Any]] = None,
|
173 |
num_bits: int = 8,
|
174 |
is_k_full: bool = True,
|
175 |
) -> torch.Tensor:
|
@@ -193,8 +207,6 @@ def fused_marlin_moe(
|
|
193 |
permutation.
|
194 |
- topk_weights (torch.Tensor): Top-k weights.
|
195 |
- topk_ids (torch.Tensor): Indices of topk-k elements.
|
196 |
-
- override_config (Optional[Dict[str, Any]]): Optional override
|
197 |
-
for the kernel configuration.
|
198 |
- w1_zeros (Optional[torch.Tensor]): Optional zero points to be used for w1.
|
199 |
- w2_zeros (Optional[torch.Tensor]): Optional zero points to be used for w2.
|
200 |
- num_bits (bool): The number of bits in expert weights quantization.
|
@@ -248,7 +260,6 @@ def fused_marlin_moe(
|
|
248 |
w2.shape,
|
249 |
topk_ids.shape[1],
|
250 |
None,
|
251 |
-
override_config=override_config,
|
252 |
is_marlin=True,
|
253 |
)
|
254 |
config = get_config_func(M)
|
@@ -350,6 +361,30 @@ def fused_marlin_moe(
|
|
350 |
return torch.sum(intermediate_cache3.view(*intermediate_cache3.shape), dim=1)
|
351 |
|
352 |
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
353 |
if hasattr(ops, "marlin_gemm_moe"):
|
354 |
|
355 |
@register_fake(add_op_namespace_prefix("marlin_gemm_moe"))
|
|
|
40 |
g_idx: Optional[torch.Tensor] = None,
|
41 |
sort_indices: Optional[torch.Tensor] = None,
|
42 |
w_zeros: Optional[torch.Tensor] = None,
|
|
|
43 |
num_bits: int = 8,
|
44 |
is_k_full: bool = True,
|
45 |
) -> torch.Tensor:
|
|
|
60 |
- topk (int): The number of top-k experts to select.
|
61 |
- renormalize (bool): If True, renormalize the top-k weights to sum to 1.
|
62 |
- w_zeros (Optional[torch.Tensor]): Optional zero points to be used for w.
|
|
|
|
|
63 |
- num_bits (bool): The number of bits in expert weights quantization.
|
64 |
|
65 |
Returns:
|
|
|
87 |
w.shape,
|
88 |
topk_ids.shape[1],
|
89 |
None,
|
|
|
90 |
is_marlin=True,
|
91 |
)
|
92 |
config = get_config_func(M)
|
|
|
150 |
return torch.sum(intermediate_cache.view(*intermediate_cache.shape), dim=1)
|
151 |
|
152 |
|
153 |
+
if hasattr(ops, "single_marlin_gemm_moe"):
|
154 |
+
|
155 |
+
@register_fake(add_op_namespace_prefix("single_marlin_gemm_moe"))
|
156 |
+
def single_marlin_moe_fake(
|
157 |
+
hidden_states: torch.Tensor,
|
158 |
+
w: torch.Tensor,
|
159 |
+
scales: torch.Tensor,
|
160 |
+
gating_output: torch.Tensor,
|
161 |
+
topk: int,
|
162 |
+
renormalize: bool,
|
163 |
+
g_idx: Optional[torch.Tensor] = None,
|
164 |
+
sort_indices: Optional[torch.Tensor] = None,
|
165 |
+
w_zeros: Optional[torch.Tensor] = None,
|
166 |
+
num_bits: int = 8,
|
167 |
+
is_k_full: bool = True,
|
168 |
+
) -> torch.Tensor:
|
169 |
+
return torch.empty_like(hidden_states)
|
170 |
+
|
171 |
+
|
172 |
def fused_marlin_moe(
|
173 |
hidden_states: torch.Tensor,
|
174 |
w1: torch.Tensor,
|
|
|
184 |
sort_indices2: Optional[torch.Tensor] = None,
|
185 |
w1_zeros: Optional[torch.Tensor] = None,
|
186 |
w2_zeros: Optional[torch.Tensor] = None,
|
|
|
187 |
num_bits: int = 8,
|
188 |
is_k_full: bool = True,
|
189 |
) -> torch.Tensor:
|
|
|
207 |
permutation.
|
208 |
- topk_weights (torch.Tensor): Top-k weights.
|
209 |
- topk_ids (torch.Tensor): Indices of topk-k elements.
|
|
|
|
|
210 |
- w1_zeros (Optional[torch.Tensor]): Optional zero points to be used for w1.
|
211 |
- w2_zeros (Optional[torch.Tensor]): Optional zero points to be used for w2.
|
212 |
- num_bits (bool): The number of bits in expert weights quantization.
|
|
|
260 |
w2.shape,
|
261 |
topk_ids.shape[1],
|
262 |
None,
|
|
|
263 |
is_marlin=True,
|
264 |
)
|
265 |
config = get_config_func(M)
|
|
|
361 |
return torch.sum(intermediate_cache3.view(*intermediate_cache3.shape), dim=1)
|
362 |
|
363 |
|
364 |
+
if hasattr(ops, "fused_marlin_moe"):
|
365 |
+
|
366 |
+
@register_fake(add_op_namespace_prefix("fused_marlin_moe"))
|
367 |
+
def fused_marlin_moe_fake(
|
368 |
+
hidden_states: torch.Tensor,
|
369 |
+
w1: torch.Tensor,
|
370 |
+
w2: torch.Tensor,
|
371 |
+
w1_scale: torch.Tensor,
|
372 |
+
w2_scale: torch.Tensor,
|
373 |
+
gating_output: torch.Tensor,
|
374 |
+
topk_weights: torch.Tensor,
|
375 |
+
topk_ids: torch.Tensor,
|
376 |
+
g_idx1: Optional[torch.Tensor] = None,
|
377 |
+
g_idx2: Optional[torch.Tensor] = None,
|
378 |
+
sort_indices1: Optional[torch.Tensor] = None,
|
379 |
+
sort_indices2: Optional[torch.Tensor] = None,
|
380 |
+
w1_zeros: Optional[torch.Tensor] = None,
|
381 |
+
w2_zeros: Optional[torch.Tensor] = None,
|
382 |
+
num_bits: int = 8,
|
383 |
+
is_k_full: bool = True,
|
384 |
+
) -> torch.Tensor:
|
385 |
+
return torch.empty_like(hidden_states)
|
386 |
+
|
387 |
+
|
388 |
if hasattr(ops, "marlin_gemm_moe"):
|
389 |
|
390 |
@register_fake(add_op_namespace_prefix("marlin_gemm_moe"))
|
build/torch25-cxx11-cu124-x86_64-linux/moe/fused_moe.py
CHANGED
@@ -1,21 +1,242 @@
|
|
|
|
1 |
"""Fused MoE kernel."""
|
2 |
|
3 |
import functools
|
4 |
import json
|
|
|
5 |
import os
|
6 |
-
from typing import Any, Callable, Dict, Optional, Tuple
|
7 |
|
8 |
import torch
|
9 |
import triton
|
10 |
import triton.language as tl
|
11 |
|
|
|
12 |
from ._ops import ops
|
13 |
-
from .fp8 import scaled_fp8_quant
|
14 |
from .platforms import current_platform
|
15 |
|
|
|
|
|
|
|
16 |
VLLM_FUSED_MOE_CHUNK_SIZE = int(os.getenv("VLLM_FUSED_MOE_CHUNK_SIZE", "32768"))
|
17 |
|
18 |
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
19 |
@triton.jit
|
20 |
def fused_moe_kernel(
|
21 |
# Pointers to matrices
|
@@ -44,8 +265,14 @@ def fused_moe_kernel(
|
|
44 |
stride_bn,
|
45 |
stride_cm,
|
46 |
stride_cn,
|
|
|
|
|
47 |
stride_bse,
|
|
|
48 |
stride_bsn,
|
|
|
|
|
|
|
49 |
# Meta-parameters
|
50 |
BLOCK_SIZE_M: tl.constexpr,
|
51 |
BLOCK_SIZE_N: tl.constexpr,
|
@@ -105,17 +332,17 @@ def fused_moe_kernel(
|
|
105 |
num_tokens_post_padded = tl.load(num_tokens_post_padded_ptr)
|
106 |
if pid_m * BLOCK_SIZE_M >= num_tokens_post_padded:
|
107 |
return
|
108 |
-
offs_token_id = pid_m * BLOCK_SIZE_M + tl.arange(0, BLOCK_SIZE_M)
|
109 |
offs_token = tl.load(sorted_token_ids_ptr + offs_token_id)
|
110 |
token_mask = offs_token < num_valid_tokens
|
111 |
|
112 |
-
offs_bn = (pid_n * BLOCK_SIZE_N + tl.arange(0, BLOCK_SIZE_N)) % N
|
113 |
offs_k = tl.arange(0, BLOCK_SIZE_K)
|
114 |
a_ptrs = a_ptr + (
|
115 |
offs_token[:, None] // top_k * stride_am + offs_k[None, :] * stride_ak
|
116 |
)
|
117 |
|
118 |
-
off_experts = tl.load(expert_ids_ptr + pid_m)
|
119 |
b_ptrs = (
|
120 |
b_ptr
|
121 |
+ off_experts * stride_be
|
@@ -128,8 +355,15 @@ def fused_moe_kernel(
|
|
128 |
b_scale = tl.load(b_scale_ptrs)
|
129 |
|
130 |
if use_fp8_w8a8:
|
131 |
-
|
132 |
-
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
133 |
|
134 |
# -----------------------------------------------------------
|
135 |
# Iterate to compute a block of the C matrix.
|
@@ -151,7 +385,17 @@ def fused_moe_kernel(
|
|
151 |
if use_int8_w8a16:
|
152 |
accumulator = tl.dot(a, b.to(compute_type), acc=accumulator)
|
153 |
elif use_fp8_w8a8:
|
154 |
-
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
155 |
else:
|
156 |
accumulator += tl.dot(a, b)
|
157 |
# Advance the ptrs to the next K block.
|
@@ -164,7 +408,10 @@ def fused_moe_kernel(
|
|
164 |
if use_int8_w8a16:
|
165 |
accumulator = (accumulator * b_scale).to(compute_type)
|
166 |
elif use_fp8_w8a8:
|
167 |
-
|
|
|
|
|
|
|
168 |
else:
|
169 |
accumulator = accumulator.to(compute_type)
|
170 |
# -----------------------------------------------------------
|
@@ -175,6 +422,141 @@ def fused_moe_kernel(
|
|
175 |
tl.store(c_ptrs, accumulator, mask=c_mask)
|
176 |
|
177 |
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
178 |
def moe_align_block_size(
|
179 |
topk_ids: torch.Tensor, block_size: int, num_experts: int
|
180 |
) -> Tuple[torch.Tensor, torch.Tensor, torch.Tensor]:
|
@@ -225,9 +607,34 @@ def moe_align_block_size(
|
|
225 |
(max_num_m_blocks,), dtype=torch.int32, device=topk_ids.device
|
226 |
)
|
227 |
num_tokens_post_pad = torch.empty((1), dtype=torch.int32, device=topk_ids.device)
|
228 |
-
|
229 |
-
|
230 |
-
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
231 |
return sorted_ids, expert_ids, num_tokens_post_pad
|
232 |
|
233 |
|
@@ -237,6 +644,7 @@ def invoke_fused_moe_kernel(
|
|
237 |
C: torch.Tensor,
|
238 |
A_scale: Optional[torch.Tensor],
|
239 |
B_scale: Optional[torch.Tensor],
|
|
|
240 |
topk_weights: torch.Tensor,
|
241 |
topk_ids: torch.Tensor,
|
242 |
sorted_token_ids: torch.Tensor,
|
@@ -248,64 +656,147 @@ def invoke_fused_moe_kernel(
|
|
248 |
compute_type: tl.dtype,
|
249 |
use_fp8_w8a8: bool,
|
250 |
use_int8_w8a16: bool,
|
|
|
|
|
251 |
) -> None:
|
252 |
assert topk_weights.stride(1) == 1
|
253 |
assert sorted_token_ids.stride(0) == 1
|
254 |
|
255 |
if use_fp8_w8a8:
|
256 |
-
A, A_scale = scaled_fp8_quant(A, A_scale)
|
257 |
assert B_scale is not None
|
258 |
-
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
259 |
assert B_scale is not None
|
|
|
260 |
else:
|
261 |
assert A_scale is None
|
262 |
assert B_scale is None
|
263 |
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
264 |
grid = lambda META: (
|
265 |
-
triton.cdiv(
|
266 |
* triton.cdiv(B.shape[1], META["BLOCK_SIZE_N"]),
|
267 |
)
|
268 |
|
269 |
-
|
270 |
-
|
271 |
-
|
272 |
-
|
273 |
-
|
274 |
-
B_scale
|
275 |
-
|
276 |
-
|
277 |
-
|
278 |
-
|
279 |
-
|
280 |
-
|
281 |
-
|
282 |
-
|
283 |
-
|
284 |
-
|
285 |
-
|
286 |
-
|
287 |
-
|
288 |
-
|
289 |
-
|
290 |
-
|
291 |
-
|
292 |
-
|
293 |
-
|
294 |
-
|
295 |
-
|
296 |
-
|
297 |
-
|
298 |
-
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
299 |
|
300 |
|
301 |
-
|
|
|
|
|
|
|
302 |
device_name = current_platform.get_device_name().replace(" ", "_")
|
303 |
dtype_selector = "" if not dtype else f",dtype={dtype}"
|
304 |
-
|
|
|
|
|
|
|
305 |
|
306 |
|
|
|
307 |
@functools.lru_cache
|
308 |
-
def get_moe_configs(
|
|
|
|
|
|
|
|
|
|
|
|
|
309 |
"""
|
310 |
Return optimized configurations for the fused MoE kernel.
|
311 |
|
@@ -317,18 +808,27 @@ def get_moe_configs(E: int, N: int, dtype: Optional[str]) -> Optional[Dict[int,
|
|
317 |
|
318 |
# First look up if an optimized configuration is available in the configs
|
319 |
# directory
|
320 |
-
|
|
|
321 |
|
322 |
config_file_path = os.path.join(
|
323 |
os.path.dirname(os.path.realpath(__file__)), "configs", json_file_name
|
324 |
)
|
325 |
if os.path.exists(config_file_path):
|
326 |
with open(config_file_path) as f:
|
|
|
327 |
# If a configuration has been found, return it
|
328 |
return {int(key): val for key, val in json.load(f).items()}
|
329 |
|
330 |
# If no optimized configuration is available, we will use the default
|
331 |
# configuration
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
332 |
return None
|
333 |
|
334 |
|
@@ -340,21 +840,34 @@ def get_default_config(
|
|
340 |
topk: int,
|
341 |
dtype: Optional[str],
|
342 |
is_marlin: bool,
|
|
|
343 |
) -> Dict[str, int]:
|
344 |
-
|
345 |
-
|
346 |
-
|
347 |
-
"BLOCK_SIZE_K": 32,
|
348 |
-
"GROUP_SIZE_M": 8,
|
349 |
-
}
|
350 |
-
# A heuristic: fused marlin works faster with this config for small M
|
351 |
-
if M <= E or (is_marlin and M <= 32):
|
352 |
config = {
|
353 |
-
"BLOCK_SIZE_M":
|
354 |
-
"BLOCK_SIZE_N":
|
355 |
-
"BLOCK_SIZE_K":
|
356 |
-
"GROUP_SIZE_M":
|
|
|
|
|
357 |
}
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
358 |
return config
|
359 |
|
360 |
|
@@ -364,15 +877,21 @@ def try_get_optimal_moe_config(
|
|
364 |
top_k: int,
|
365 |
dtype: Optional[str],
|
366 |
M: int,
|
367 |
-
override_config: Optional[Dict[str, Any]] = None,
|
368 |
is_marlin: bool = False,
|
|
|
369 |
):
|
|
|
|
|
|
|
|
|
370 |
if override_config:
|
371 |
config = override_config
|
372 |
else:
|
373 |
# First try to load optimal config from the file
|
374 |
E, _, N = w2_shape
|
375 |
-
|
|
|
|
|
376 |
|
377 |
if configs:
|
378 |
# If an optimal configuration map has been found, look up the
|
@@ -380,7 +899,9 @@ def try_get_optimal_moe_config(
|
|
380 |
config = configs[min(configs.keys(), key=lambda x: abs(x - M))]
|
381 |
else:
|
382 |
# Else use the default config
|
383 |
-
config = get_default_config(
|
|
|
|
|
384 |
return config
|
385 |
|
386 |
|
@@ -416,7 +937,8 @@ def fused_topk(
|
|
416 |
return topk_weights, topk_ids
|
417 |
|
418 |
|
419 |
-
# This is used by the Deepseek-V2 model
|
|
|
420 |
def grouped_topk(
|
421 |
hidden_states: torch.Tensor,
|
422 |
gating_output: torch.Tensor,
|
@@ -424,11 +946,25 @@ def grouped_topk(
|
|
424 |
renormalize: bool,
|
425 |
num_expert_group: int = 0,
|
426 |
topk_group: int = 0,
|
|
|
|
|
427 |
):
|
428 |
|
429 |
assert hidden_states.shape[0] == gating_output.shape[0], "Number of tokens mismatch"
|
430 |
|
431 |
-
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
432 |
num_token = scores.shape[0]
|
433 |
group_scores = (
|
434 |
scores.view(num_token, num_expert_group, -1).max(dim=-1).values
|
@@ -444,7 +980,13 @@ def grouped_topk(
|
|
444 |
.reshape(num_token, -1)
|
445 |
) # [n, e]
|
446 |
tmp_scores = scores.masked_fill(~score_mask.bool(), 0.0) # [n, e]
|
447 |
-
|
|
|
|
|
|
|
|
|
|
|
|
|
448 |
|
449 |
if renormalize:
|
450 |
topk_weights = topk_weights / topk_weights.sum(dim=-1, keepdim=True)
|
@@ -454,6 +996,7 @@ def grouped_topk(
|
|
454 |
|
455 |
def get_config_dtype_str(
|
456 |
dtype: torch.dtype,
|
|
|
457 |
use_int8_w8a16: Optional[bool] = False,
|
458 |
use_fp8_w8a8: Optional[bool] = False,
|
459 |
):
|
@@ -461,6 +1004,8 @@ def get_config_dtype_str(
|
|
461 |
return "fp8_w8a8"
|
462 |
elif use_int8_w8a16:
|
463 |
return "int8_w8a16"
|
|
|
|
|
464 |
elif dtype == torch.float:
|
465 |
# avoiding cases where kernel fails when float32 MoE
|
466 |
# use fp16/bfloat16 configs
|
@@ -468,6 +1013,80 @@ def get_config_dtype_str(
|
|
468 |
return None
|
469 |
|
470 |
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
471 |
def fused_experts(
|
472 |
hidden_states: torch.Tensor,
|
473 |
w1: torch.Tensor,
|
@@ -475,16 +1094,80 @@ def fused_experts(
|
|
475 |
topk_weights: torch.Tensor,
|
476 |
topk_ids: torch.Tensor,
|
477 |
inplace: bool = False,
|
478 |
-
override_config: Optional[Dict[str, Any]] = None,
|
479 |
use_fp8_w8a8: bool = False,
|
480 |
use_int8_w8a16: bool = False,
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
481 |
w1_scale: Optional[torch.Tensor] = None,
|
482 |
w2_scale: Optional[torch.Tensor] = None,
|
|
|
|
|
483 |
a1_scale: Optional[torch.Tensor] = None,
|
484 |
a2_scale: Optional[torch.Tensor] = None,
|
|
|
485 |
):
|
486 |
# Check constraints.
|
487 |
-
|
|
|
|
|
|
|
|
|
488 |
assert topk_weights.shape == topk_ids.shape, "topk shape mismatch"
|
489 |
assert hidden_states.is_contiguous(), "Hidden_states must be contiguous"
|
490 |
assert w1.is_contiguous(), "Expert weights1 must be contiguous"
|
@@ -500,6 +1183,7 @@ def fused_experts(
|
|
500 |
config_dtype = get_config_dtype_str(
|
501 |
use_fp8_w8a8=use_fp8_w8a8,
|
502 |
use_int8_w8a16=use_int8_w8a16,
|
|
|
503 |
dtype=hidden_states.dtype,
|
504 |
)
|
505 |
|
@@ -509,7 +1193,7 @@ def fused_experts(
|
|
509 |
w2.shape,
|
510 |
topk_ids.shape[1],
|
511 |
config_dtype,
|
512 |
-
|
513 |
)
|
514 |
|
515 |
config = get_config_func(M)
|
@@ -530,7 +1214,14 @@ def fused_experts(
|
|
530 |
dtype=hidden_states.dtype,
|
531 |
)
|
532 |
|
533 |
-
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
534 |
|
535 |
if inplace:
|
536 |
out_hidden_states = hidden_states
|
@@ -571,6 +1262,7 @@ def fused_experts(
|
|
571 |
intermediate_cache1,
|
572 |
a1_scale,
|
573 |
w1_scale,
|
|
|
574 |
curr_topk_weights,
|
575 |
curr_topk_ids,
|
576 |
sorted_token_ids,
|
@@ -582,6 +1274,8 @@ def fused_experts(
|
|
582 |
compute_type=compute_type,
|
583 |
use_fp8_w8a8=use_fp8_w8a8,
|
584 |
use_int8_w8a16=use_int8_w8a16,
|
|
|
|
|
585 |
)
|
586 |
|
587 |
ops.silu_and_mul(intermediate_cache2, intermediate_cache1.view(-1, N))
|
@@ -592,6 +1286,7 @@ def fused_experts(
|
|
592 |
intermediate_cache3,
|
593 |
a2_scale,
|
594 |
w2_scale,
|
|
|
595 |
curr_topk_weights,
|
596 |
curr_topk_ids,
|
597 |
sorted_token_ids,
|
@@ -603,6 +1298,8 @@ def fused_experts(
|
|
603 |
compute_type=compute_type,
|
604 |
use_fp8_w8a8=use_fp8_w8a8,
|
605 |
use_int8_w8a16=use_int8_w8a16,
|
|
|
|
|
606 |
)
|
607 |
|
608 |
ops.moe_sum(
|
@@ -620,17 +1317,20 @@ def fused_moe(
|
|
620 |
topk: int,
|
621 |
renormalize: bool,
|
622 |
inplace: bool = False,
|
623 |
-
override_config: Optional[Dict[str, Any]] = None,
|
624 |
use_grouped_topk: bool = False,
|
625 |
num_expert_group: Optional[int] = None,
|
626 |
topk_group: Optional[int] = None,
|
627 |
custom_routing_function: Optional[Callable] = None,
|
628 |
use_fp8_w8a8: bool = False,
|
629 |
use_int8_w8a16: bool = False,
|
|
|
630 |
w1_scale: Optional[torch.Tensor] = None,
|
631 |
w2_scale: Optional[torch.Tensor] = None,
|
|
|
|
|
632 |
a1_scale: Optional[torch.Tensor] = None,
|
633 |
a2_scale: Optional[torch.Tensor] = None,
|
|
|
634 |
) -> torch.Tensor:
|
635 |
"""
|
636 |
This function computes a Mixture of Experts (MoE) layer using two sets of
|
@@ -646,20 +1346,28 @@ def fused_moe(
|
|
646 |
- renormalize (bool): If True, renormalize the top-k weights to sum to 1.
|
647 |
- inplace (bool): If True, perform the operation in-place.
|
648 |
Defaults to False.
|
649 |
-
- override_config (Optional[Dict[str, Any]]): Optional override
|
650 |
-
for the kernel configuration.
|
651 |
- num_expert_group: Optional[int]: additional parameter for grouped_topk
|
652 |
- topk_group: Optional[int]: additional parameter for grouped_topk
|
653 |
- use_grouped_topk: If True, use grouped_topk instead of fused_topk
|
654 |
note: Deepseekv2 model uses grouped_topk
|
655 |
- use_fp8_w8a8 (bool): If True, use fp8 arithmetic to compute the inner
|
656 |
products for w1 and w2. Defaults to False.
|
657 |
-
- use_int8_w8a16 (bool): If True, use
|
658 |
-
products for w1 and w2.
|
|
|
|
|
|
|
|
|
659 |
- w1_scale (Optional[torch.Tensor]): Optional scale to be used for
|
660 |
w1.
|
661 |
- w2_scale (Optional[torch.Tensor]): Optional scale to be used for
|
662 |
w2.
|
|
|
|
|
|
|
|
|
|
|
|
|
663 |
|
664 |
Returns:
|
665 |
- torch.Tensor: The output tensor after applying the MoE layer.
|
@@ -693,11 +1401,14 @@ def fused_moe(
|
|
693 |
topk_weights,
|
694 |
topk_ids,
|
695 |
inplace=inplace,
|
696 |
-
override_config=override_config,
|
697 |
use_fp8_w8a8=use_fp8_w8a8,
|
698 |
use_int8_w8a16=use_int8_w8a16,
|
|
|
699 |
w1_scale=w1_scale,
|
700 |
w2_scale=w2_scale,
|
|
|
|
|
701 |
a1_scale=a1_scale,
|
702 |
a2_scale=a2_scale,
|
|
|
703 |
)
|
|
|
1 |
+
# SPDX-License-Identifier: Apache-2.0
|
2 |
"""Fused MoE kernel."""
|
3 |
|
4 |
import functools
|
5 |
import json
|
6 |
+
import logging
|
7 |
import os
|
8 |
+
from typing import Any, Callable, Dict, List, Optional, Tuple
|
9 |
|
10 |
import torch
|
11 |
import triton
|
12 |
import triton.language as tl
|
13 |
|
14 |
+
|
15 |
from ._ops import ops
|
16 |
+
from .fp8 import per_token_group_quant_fp8, scaled_fp8_quant
|
17 |
from .platforms import current_platform
|
18 |
|
19 |
+
logger = logging.getLogger(__name__)
|
20 |
+
|
21 |
+
|
22 |
VLLM_FUSED_MOE_CHUNK_SIZE = int(os.getenv("VLLM_FUSED_MOE_CHUNK_SIZE", "32768"))
|
23 |
|
24 |
|
25 |
+
@triton.jit
|
26 |
+
def fused_moe_kernel_gptq_awq(
|
27 |
+
# Pointers to matrices
|
28 |
+
a_ptr,
|
29 |
+
b_ptr,
|
30 |
+
c_ptr,
|
31 |
+
b_scale_ptr,
|
32 |
+
b_zp_ptr,
|
33 |
+
topk_weights_ptr,
|
34 |
+
sorted_token_ids_ptr,
|
35 |
+
expert_ids_ptr,
|
36 |
+
num_tokens_post_padded_ptr,
|
37 |
+
# Matrix dimensions
|
38 |
+
N: tl.constexpr,
|
39 |
+
K: tl.constexpr,
|
40 |
+
EM,
|
41 |
+
num_valid_tokens,
|
42 |
+
# The stride variables represent how much to increase the ptr by when
|
43 |
+
# moving by 1 element in a particular dimension. E.g. `stride_am` is
|
44 |
+
# how much to increase `a_ptr` by to get the element one row down
|
45 |
+
# (A has M rows).
|
46 |
+
stride_am,
|
47 |
+
stride_ak,
|
48 |
+
stride_be,
|
49 |
+
stride_bk,
|
50 |
+
stride_bn,
|
51 |
+
stride_cm,
|
52 |
+
stride_cn,
|
53 |
+
stride_bse,
|
54 |
+
stride_bsk,
|
55 |
+
stride_bsn,
|
56 |
+
stride_bze,
|
57 |
+
stride_bzk,
|
58 |
+
stride_bzn,
|
59 |
+
block_k_diviable: tl.constexpr,
|
60 |
+
group_size: tl.constexpr,
|
61 |
+
# Meta-parameters
|
62 |
+
BLOCK_SIZE_M: tl.constexpr,
|
63 |
+
BLOCK_SIZE_N: tl.constexpr,
|
64 |
+
BLOCK_SIZE_K: tl.constexpr,
|
65 |
+
GROUP_SIZE_M: tl.constexpr,
|
66 |
+
MUL_ROUTED_WEIGHT: tl.constexpr,
|
67 |
+
top_k: tl.constexpr,
|
68 |
+
compute_type: tl.constexpr,
|
69 |
+
has_zp: tl.constexpr,
|
70 |
+
use_int4_w4a16: tl.constexpr,
|
71 |
+
use_int8_w8a16: tl.constexpr,
|
72 |
+
):
|
73 |
+
"""
|
74 |
+
Implements the fused computation for a Mixture of Experts (MOE) using
|
75 |
+
token and expert matrices.
|
76 |
+
|
77 |
+
Key Parameters:
|
78 |
+
- A: The input tensor representing tokens with shape (*, K), where '*' can
|
79 |
+
be any shape representing batches and K is the feature dimension of
|
80 |
+
each token.
|
81 |
+
- B: The stacked MOE weight tensor with shape (E, N, K), where E is
|
82 |
+
the number of experts, K is the input feature dimension, and N is
|
83 |
+
the output feature dimension.
|
84 |
+
- C: The output cache tensor with shape (M, topk, N), where M is the
|
85 |
+
total number of tokens post padding, topk is the number of times
|
86 |
+
each token is repeated, and N is the output feature dimension.
|
87 |
+
- sorted_token_ids: A tensor containing the sorted indices of tokens,
|
88 |
+
repeated topk times and arranged by the expert index they are
|
89 |
+
assigned to.
|
90 |
+
- expert_ids: A tensor containing the indices of the expert for each
|
91 |
+
block. It determines which expert matrix from B should be used for
|
92 |
+
each block in A.
|
93 |
+
This kernel performs the multiplication of a token by its corresponding
|
94 |
+
expert matrix as determined by `expert_ids`. The sorting of
|
95 |
+
`sorted_token_ids` by expert index and padding ensures divisibility by
|
96 |
+
BLOCK_SIZE_M, which is necessary to maintain consistency in block matrix
|
97 |
+
multiplication across different blocks processed by the same expert.
|
98 |
+
"""
|
99 |
+
# -----------------------------------------------------------
|
100 |
+
# Map program ids `pid` to the block of C it should compute.
|
101 |
+
# This is done in a grouped ordering to promote L2 data reuse.
|
102 |
+
pid = tl.program_id(axis=0)
|
103 |
+
num_pid_m = tl.cdiv(EM, BLOCK_SIZE_M)
|
104 |
+
num_pid_n = tl.cdiv(N, BLOCK_SIZE_N)
|
105 |
+
num_pid_in_group = GROUP_SIZE_M * num_pid_n
|
106 |
+
group_id = pid // num_pid_in_group
|
107 |
+
first_pid_m = group_id * GROUP_SIZE_M
|
108 |
+
group_size_m = min(num_pid_m - first_pid_m, GROUP_SIZE_M)
|
109 |
+
pid_m = first_pid_m + ((pid % num_pid_in_group) % group_size_m)
|
110 |
+
pid_n = (pid % num_pid_in_group) // group_size_m
|
111 |
+
|
112 |
+
# ----------------------------------------------------------
|
113 |
+
# Create pointers for the first blocks of A and B.
|
114 |
+
# We will advance this pointer as we move in the K direction
|
115 |
+
# and accumulate
|
116 |
+
# `a_ptrs` is a block of [BLOCK_SIZE_M, BLOCK_SIZE_K] pointers
|
117 |
+
# `b_ptrs` is a block of [BLOCK_SIZE_K, BLOCK_SIZE_N] pointers
|
118 |
+
num_tokens_post_padded = tl.load(num_tokens_post_padded_ptr)
|
119 |
+
if pid_m * BLOCK_SIZE_M >= num_tokens_post_padded:
|
120 |
+
return
|
121 |
+
offs_token_id = pid_m * BLOCK_SIZE_M + tl.arange(0, BLOCK_SIZE_M).to(tl.int64)
|
122 |
+
offs_token = tl.load(sorted_token_ids_ptr + offs_token_id)
|
123 |
+
token_mask = offs_token < num_valid_tokens
|
124 |
+
|
125 |
+
offs_bn = (pid_n * BLOCK_SIZE_N + tl.arange(0, BLOCK_SIZE_N).to(tl.int64)) % N
|
126 |
+
offs_k = tl.arange(0, BLOCK_SIZE_K)
|
127 |
+
a_ptrs = a_ptr + (
|
128 |
+
offs_token[:, None] // top_k * stride_am + offs_k[None, :] * stride_ak
|
129 |
+
)
|
130 |
+
|
131 |
+
off_experts = tl.load(expert_ids_ptr + pid_m).to(tl.int64)
|
132 |
+
|
133 |
+
if use_int4_w4a16:
|
134 |
+
b_ptrs = (
|
135 |
+
b_ptr
|
136 |
+
+ off_experts * stride_be
|
137 |
+
+ (offs_k[:, None] // 2) * stride_bk
|
138 |
+
+ offs_bn[None, :] * stride_bn
|
139 |
+
)
|
140 |
+
b_shifter = (offs_k[:, None] % 2) * 4
|
141 |
+
elif use_int8_w8a16:
|
142 |
+
b_ptrs = (
|
143 |
+
b_ptr
|
144 |
+
+ off_experts * stride_be
|
145 |
+
+ offs_k[:, None] * stride_bk
|
146 |
+
+ offs_bn[None, :] * stride_bn
|
147 |
+
)
|
148 |
+
|
149 |
+
if not has_zp and use_int4_w4a16:
|
150 |
+
b_zp_num = 8
|
151 |
+
if not has_zp and use_int8_w8a16:
|
152 |
+
b_zp_num = 128
|
153 |
+
elif has_zp and use_int4_w4a16:
|
154 |
+
b_zp_shifter = (offs_bn[None, :] % 2) * 4
|
155 |
+
|
156 |
+
# -----------------------------------------------------------
|
157 |
+
# Iterate to compute a block of the C matrix.
|
158 |
+
# We accumulate into a `[BLOCK_SIZE_M, BLOCK_SIZE_N]` block
|
159 |
+
# of fp32 values for higher accuracy.
|
160 |
+
# `accumulator` will be converted back to fp16 after the loop.
|
161 |
+
accumulator = tl.zeros((BLOCK_SIZE_M, BLOCK_SIZE_N), dtype=tl.float32)
|
162 |
+
for k in range(0, tl.cdiv(K, BLOCK_SIZE_K)):
|
163 |
+
# Load the next block of A and B, generate a mask by checking the
|
164 |
+
# K dimension.
|
165 |
+
|
166 |
+
if not block_k_diviable:
|
167 |
+
k_mask = offs_k[:, None] < K - k * BLOCK_SIZE_K
|
168 |
+
k_other = 0.0
|
169 |
+
else:
|
170 |
+
k_mask = None
|
171 |
+
k_other = None
|
172 |
+
|
173 |
+
a = tl.load(
|
174 |
+
a_ptrs,
|
175 |
+
mask=token_mask[:, None] & (offs_k[None, :] < K - k * BLOCK_SIZE_K),
|
176 |
+
other=0.0,
|
177 |
+
)
|
178 |
+
b = tl.load(b_ptrs)
|
179 |
+
if use_int4_w4a16:
|
180 |
+
b = (b >> b_shifter) & 0xF
|
181 |
+
|
182 |
+
b_scale_ptrs = (
|
183 |
+
b_scale_ptr
|
184 |
+
+ off_experts * stride_bse
|
185 |
+
+ offs_bn[None, :] * stride_bsn
|
186 |
+
+ ((offs_k[:, None] + BLOCK_SIZE_K * k) // group_size) * stride_bsk
|
187 |
+
)
|
188 |
+
b_scale = tl.load(b_scale_ptrs, mask=k_mask, other=k_other)
|
189 |
+
b_scale = b_scale.to(tl.float32)
|
190 |
+
|
191 |
+
if has_zp and use_int4_w4a16:
|
192 |
+
offs_k_true = (offs_k[:, None] + BLOCK_SIZE_K * k) // group_size
|
193 |
+
b_zp_ptrs = (
|
194 |
+
b_zp_ptr
|
195 |
+
+ off_experts * stride_bze
|
196 |
+
+ (offs_bn[None, :] // 2) * stride_bzn
|
197 |
+
+ offs_k_true * stride_bzk
|
198 |
+
)
|
199 |
+
b_zp = tl.load(b_zp_ptrs, mask=k_mask, other=k_other)
|
200 |
+
b_zp = (b_zp >> b_zp_shifter) & 0xF
|
201 |
+
b_zp = b_zp.to(tl.float32)
|
202 |
+
elif has_zp and use_int8_w8a16:
|
203 |
+
offs_k_true = (offs_k[:, None] + BLOCK_SIZE_K * k) // group_size
|
204 |
+
b_zp_ptrs = (
|
205 |
+
b_zp_ptr
|
206 |
+
+ off_experts * stride_bze
|
207 |
+
+ offs_bn[None, :] * stride_bzn
|
208 |
+
+ offs_k_true * stride_bzk
|
209 |
+
)
|
210 |
+
b_zp = tl.load(b_zp_ptrs, mask=k_mask, other=k_other)
|
211 |
+
b_zp = b_zp.to(tl.float32)
|
212 |
+
|
213 |
+
# We accumulate along the K dimension.
|
214 |
+
if has_zp:
|
215 |
+
b = ((b.to(tl.float32) - b_zp) * b_scale).to(compute_type)
|
216 |
+
else:
|
217 |
+
b = ((b.to(tl.float32) - b_zp_num) * b_scale).to(compute_type)
|
218 |
+
accumulator = tl.dot(a, b, acc=accumulator)
|
219 |
+
|
220 |
+
# Advance the ptrs to the next K block.
|
221 |
+
a_ptrs += BLOCK_SIZE_K * stride_ak
|
222 |
+
if use_int4_w4a16:
|
223 |
+
b_ptrs += (BLOCK_SIZE_K // 2) * stride_bk
|
224 |
+
else:
|
225 |
+
b_ptrs += BLOCK_SIZE_K * stride_bk
|
226 |
+
|
227 |
+
if MUL_ROUTED_WEIGHT:
|
228 |
+
moe_weight = tl.load(topk_weights_ptr + offs_token, mask=token_mask, other=0)
|
229 |
+
accumulator = accumulator * moe_weight[:, None]
|
230 |
+
|
231 |
+
accumulator = accumulator.to(compute_type)
|
232 |
+
# -----------------------------------------------------------
|
233 |
+
# Write back the block of the output
|
234 |
+
offs_cn = pid_n * BLOCK_SIZE_N + tl.arange(0, BLOCK_SIZE_N)
|
235 |
+
c_ptrs = c_ptr + stride_cm * offs_token[:, None] + stride_cn * offs_cn[None, :]
|
236 |
+
c_mask = token_mask[:, None] & (offs_cn[None, :] < N)
|
237 |
+
tl.store(c_ptrs, accumulator, mask=c_mask)
|
238 |
+
|
239 |
+
|
240 |
@triton.jit
|
241 |
def fused_moe_kernel(
|
242 |
# Pointers to matrices
|
|
|
265 |
stride_bn,
|
266 |
stride_cm,
|
267 |
stride_cn,
|
268 |
+
stride_asm,
|
269 |
+
stride_ask,
|
270 |
stride_bse,
|
271 |
+
stride_bsk,
|
272 |
stride_bsn,
|
273 |
+
# Block size for block-wise quantization
|
274 |
+
group_n: tl.constexpr,
|
275 |
+
group_k: tl.constexpr,
|
276 |
# Meta-parameters
|
277 |
BLOCK_SIZE_M: tl.constexpr,
|
278 |
BLOCK_SIZE_N: tl.constexpr,
|
|
|
332 |
num_tokens_post_padded = tl.load(num_tokens_post_padded_ptr)
|
333 |
if pid_m * BLOCK_SIZE_M >= num_tokens_post_padded:
|
334 |
return
|
335 |
+
offs_token_id = pid_m * BLOCK_SIZE_M + tl.arange(0, BLOCK_SIZE_M).to(tl.int64)
|
336 |
offs_token = tl.load(sorted_token_ids_ptr + offs_token_id)
|
337 |
token_mask = offs_token < num_valid_tokens
|
338 |
|
339 |
+
offs_bn = (pid_n * BLOCK_SIZE_N + tl.arange(0, BLOCK_SIZE_N).to(tl.int64)) % N
|
340 |
offs_k = tl.arange(0, BLOCK_SIZE_K)
|
341 |
a_ptrs = a_ptr + (
|
342 |
offs_token[:, None] // top_k * stride_am + offs_k[None, :] * stride_ak
|
343 |
)
|
344 |
|
345 |
+
off_experts = tl.load(expert_ids_ptr + pid_m).to(tl.int64)
|
346 |
b_ptrs = (
|
347 |
b_ptr
|
348 |
+ off_experts * stride_be
|
|
|
355 |
b_scale = tl.load(b_scale_ptrs)
|
356 |
|
357 |
if use_fp8_w8a8:
|
358 |
+
if group_k > 0 and group_n > 0:
|
359 |
+
a_scale_ptrs = a_scale_ptr + (offs_token // top_k) * stride_asm
|
360 |
+
offs_bsn = offs_bn // group_n
|
361 |
+
b_scale_ptrs = (
|
362 |
+
b_scale_ptr + off_experts * stride_bse + offs_bsn * stride_bsn
|
363 |
+
)
|
364 |
+
else:
|
365 |
+
a_scale = tl.load(a_scale_ptr)
|
366 |
+
b_scale = tl.load(b_scale_ptr + off_experts)
|
367 |
|
368 |
# -----------------------------------------------------------
|
369 |
# Iterate to compute a block of the C matrix.
|
|
|
385 |
if use_int8_w8a16:
|
386 |
accumulator = tl.dot(a, b.to(compute_type), acc=accumulator)
|
387 |
elif use_fp8_w8a8:
|
388 |
+
if group_k > 0 and group_n > 0:
|
389 |
+
k_start = k * BLOCK_SIZE_K
|
390 |
+
offs_ks = k_start // group_k
|
391 |
+
a_scale = tl.load(
|
392 |
+
a_scale_ptrs + offs_ks * stride_ask, mask=token_mask, other=0.0
|
393 |
+
)
|
394 |
+
b_scale = tl.load(b_scale_ptrs + offs_ks * stride_bsk)
|
395 |
+
|
396 |
+
accumulator += tl.dot(a, b) * a_scale[:, None] * b_scale[None, :]
|
397 |
+
else:
|
398 |
+
accumulator = tl.dot(a, b, acc=accumulator)
|
399 |
else:
|
400 |
accumulator += tl.dot(a, b)
|
401 |
# Advance the ptrs to the next K block.
|
|
|
408 |
if use_int8_w8a16:
|
409 |
accumulator = (accumulator * b_scale).to(compute_type)
|
410 |
elif use_fp8_w8a8:
|
411 |
+
if group_k > 0 and group_n > 0:
|
412 |
+
accumulator = accumulator.to(compute_type)
|
413 |
+
else:
|
414 |
+
accumulator = (accumulator * a_scale * b_scale).to(compute_type)
|
415 |
else:
|
416 |
accumulator = accumulator.to(compute_type)
|
417 |
# -----------------------------------------------------------
|
|
|
422 |
tl.store(c_ptrs, accumulator, mask=c_mask)
|
423 |
|
424 |
|
425 |
+
def ceil_div(a, b):
|
426 |
+
return (a + b - 1) // b
|
427 |
+
|
428 |
+
|
429 |
+
@triton.jit
|
430 |
+
def moe_align_block_size_stage1(
|
431 |
+
topk_ids_ptr,
|
432 |
+
tokens_cnts_ptr,
|
433 |
+
num_experts: tl.constexpr,
|
434 |
+
numel: tl.constexpr,
|
435 |
+
tokens_per_thread: tl.constexpr,
|
436 |
+
):
|
437 |
+
pid = tl.program_id(0)
|
438 |
+
|
439 |
+
start_idx = pid * tokens_per_thread
|
440 |
+
|
441 |
+
off_c = (pid + 1) * num_experts
|
442 |
+
|
443 |
+
for i in range(tokens_per_thread):
|
444 |
+
if start_idx + i < numel:
|
445 |
+
idx = tl.load(topk_ids_ptr + start_idx + i)
|
446 |
+
token_cnt = tl.load(tokens_cnts_ptr + off_c + idx)
|
447 |
+
tl.store(tokens_cnts_ptr + off_c + idx, token_cnt + 1)
|
448 |
+
|
449 |
+
|
450 |
+
@triton.jit
|
451 |
+
def moe_align_block_size_stage2(
|
452 |
+
tokens_cnts_ptr,
|
453 |
+
num_experts: tl.constexpr,
|
454 |
+
):
|
455 |
+
pid = tl.program_id(0)
|
456 |
+
|
457 |
+
last_cnt = 0
|
458 |
+
for i in range(1, num_experts + 1):
|
459 |
+
token_cnt = tl.load(tokens_cnts_ptr + i * num_experts + pid)
|
460 |
+
last_cnt = last_cnt + token_cnt
|
461 |
+
tl.store(tokens_cnts_ptr + i * num_experts + pid, last_cnt)
|
462 |
+
|
463 |
+
|
464 |
+
@triton.jit
|
465 |
+
def moe_align_block_size_stage3(
|
466 |
+
total_tokens_post_pad_ptr,
|
467 |
+
tokens_cnts_ptr,
|
468 |
+
cumsum_ptr,
|
469 |
+
num_experts: tl.constexpr,
|
470 |
+
block_size: tl.constexpr,
|
471 |
+
):
|
472 |
+
last_cumsum = 0
|
473 |
+
off_cnt = num_experts * num_experts
|
474 |
+
for i in range(1, num_experts + 1):
|
475 |
+
token_cnt = tl.load(tokens_cnts_ptr + off_cnt + i - 1)
|
476 |
+
last_cumsum = last_cumsum + tl.cdiv(token_cnt, block_size) * block_size
|
477 |
+
tl.store(cumsum_ptr + i, last_cumsum)
|
478 |
+
tl.store(total_tokens_post_pad_ptr, last_cumsum)
|
479 |
+
|
480 |
+
|
481 |
+
@triton.jit
|
482 |
+
def moe_align_block_size_stage4(
|
483 |
+
topk_ids_ptr,
|
484 |
+
sorted_token_ids_ptr,
|
485 |
+
expert_ids_ptr,
|
486 |
+
tokens_cnts_ptr,
|
487 |
+
cumsum_ptr,
|
488 |
+
num_experts: tl.constexpr,
|
489 |
+
block_size: tl.constexpr,
|
490 |
+
numel: tl.constexpr,
|
491 |
+
tokens_per_thread: tl.constexpr,
|
492 |
+
):
|
493 |
+
pid = tl.program_id(0)
|
494 |
+
start_idx = tl.load(cumsum_ptr + pid)
|
495 |
+
end_idx = tl.load(cumsum_ptr + pid + 1)
|
496 |
+
|
497 |
+
for i in range(start_idx, end_idx, block_size):
|
498 |
+
tl.store(expert_ids_ptr + i // block_size, pid)
|
499 |
+
|
500 |
+
start_idx = pid * tokens_per_thread
|
501 |
+
off_t = pid * num_experts
|
502 |
+
|
503 |
+
for i in range(start_idx, tl.minimum(start_idx + tokens_per_thread, numel)):
|
504 |
+
expert_id = tl.load(topk_ids_ptr + i)
|
505 |
+
token_cnt = tl.load(tokens_cnts_ptr + off_t + expert_id)
|
506 |
+
rank_post_pad = token_cnt + tl.load(cumsum_ptr + expert_id)
|
507 |
+
tl.store(sorted_token_ids_ptr + rank_post_pad, i)
|
508 |
+
tl.store(tokens_cnts_ptr + off_t + expert_id, token_cnt + 1)
|
509 |
+
|
510 |
+
|
511 |
+
# Triton implementation based on:
|
512 |
+
# https://github.com/sgl-project/sglang/commit/ba5112ff691d791a9e38c6c71f59324a5fcb49d0
|
513 |
+
def moe_align_block_size_triton(
|
514 |
+
topk_ids: torch.Tensor,
|
515 |
+
num_experts: int,
|
516 |
+
block_size: int,
|
517 |
+
sorted_token_ids: torch.Tensor,
|
518 |
+
expert_ids: torch.Tensor,
|
519 |
+
num_tokens_post_pad: torch.Tensor,
|
520 |
+
) -> None:
|
521 |
+
numel = topk_ids.numel()
|
522 |
+
grid = (num_experts,)
|
523 |
+
tokens_cnts = torch.zeros(
|
524 |
+
(num_experts + 1, num_experts), dtype=torch.int32, device=topk_ids.device
|
525 |
+
)
|
526 |
+
cumsum = torch.zeros((num_experts + 1,), dtype=torch.int32, device=topk_ids.device)
|
527 |
+
tokens_per_thread = ceil_div(numel, num_experts)
|
528 |
+
|
529 |
+
moe_align_block_size_stage1[grid](
|
530 |
+
topk_ids,
|
531 |
+
tokens_cnts,
|
532 |
+
num_experts,
|
533 |
+
numel,
|
534 |
+
tokens_per_thread,
|
535 |
+
)
|
536 |
+
moe_align_block_size_stage2[grid](
|
537 |
+
tokens_cnts,
|
538 |
+
num_experts,
|
539 |
+
)
|
540 |
+
moe_align_block_size_stage3[(1,)](
|
541 |
+
num_tokens_post_pad,
|
542 |
+
tokens_cnts,
|
543 |
+
cumsum,
|
544 |
+
num_experts,
|
545 |
+
block_size,
|
546 |
+
)
|
547 |
+
moe_align_block_size_stage4[grid](
|
548 |
+
topk_ids,
|
549 |
+
sorted_token_ids,
|
550 |
+
expert_ids,
|
551 |
+
tokens_cnts,
|
552 |
+
cumsum,
|
553 |
+
num_experts,
|
554 |
+
block_size,
|
555 |
+
numel,
|
556 |
+
tokens_per_thread,
|
557 |
+
)
|
558 |
+
|
559 |
+
|
560 |
def moe_align_block_size(
|
561 |
topk_ids: torch.Tensor, block_size: int, num_experts: int
|
562 |
) -> Tuple[torch.Tensor, torch.Tensor, torch.Tensor]:
|
|
|
607 |
(max_num_m_blocks,), dtype=torch.int32, device=topk_ids.device
|
608 |
)
|
609 |
num_tokens_post_pad = torch.empty((1), dtype=torch.int32, device=topk_ids.device)
|
610 |
+
if num_experts >= 224:
|
611 |
+
if VLLM_ENABLE_MOE_ALIGN_BLOCK_SIZE_TRITON:
|
612 |
+
moe_align_block_size_triton(
|
613 |
+
topk_ids,
|
614 |
+
num_experts,
|
615 |
+
block_size,
|
616 |
+
sorted_ids,
|
617 |
+
expert_ids,
|
618 |
+
num_tokens_post_pad,
|
619 |
+
)
|
620 |
+
else:
|
621 |
+
ops.sgl_moe_align_block_size(
|
622 |
+
topk_ids,
|
623 |
+
num_experts,
|
624 |
+
block_size,
|
625 |
+
sorted_ids,
|
626 |
+
expert_ids,
|
627 |
+
num_tokens_post_pad,
|
628 |
+
)
|
629 |
+
else:
|
630 |
+
ops.moe_align_block_size(
|
631 |
+
topk_ids,
|
632 |
+
num_experts,
|
633 |
+
block_size,
|
634 |
+
sorted_ids,
|
635 |
+
expert_ids,
|
636 |
+
num_tokens_post_pad,
|
637 |
+
)
|
638 |
return sorted_ids, expert_ids, num_tokens_post_pad
|
639 |
|
640 |
|
|
|
644 |
C: torch.Tensor,
|
645 |
A_scale: Optional[torch.Tensor],
|
646 |
B_scale: Optional[torch.Tensor],
|
647 |
+
B_zp: Optional[torch.Tensor],
|
648 |
topk_weights: torch.Tensor,
|
649 |
topk_ids: torch.Tensor,
|
650 |
sorted_token_ids: torch.Tensor,
|
|
|
656 |
compute_type: tl.dtype,
|
657 |
use_fp8_w8a8: bool,
|
658 |
use_int8_w8a16: bool,
|
659 |
+
use_int4_w4a16: bool,
|
660 |
+
block_shape: Optional[List[int]] = None,
|
661 |
) -> None:
|
662 |
assert topk_weights.stride(1) == 1
|
663 |
assert sorted_token_ids.stride(0) == 1
|
664 |
|
665 |
if use_fp8_w8a8:
|
|
|
666 |
assert B_scale is not None
|
667 |
+
if block_shape is None:
|
668 |
+
A, A_scale = scaled_fp8_quant(A, A_scale)
|
669 |
+
else:
|
670 |
+
assert len(block_shape) == 2
|
671 |
+
block_n, block_k = block_shape[0], block_shape[1]
|
672 |
+
A, A_scale = per_token_group_quant_fp8(A, block_k)
|
673 |
+
assert triton.cdiv(A.shape[-1], block_k) == A_scale.shape[-1]
|
674 |
+
assert triton.cdiv(B.shape[-2], block_n) == B_scale.shape[-2]
|
675 |
+
assert triton.cdiv(B.shape[-1], block_k) == B_scale.shape[-1]
|
676 |
+
elif use_int8_w8a16 or use_int4_w4a16:
|
677 |
assert B_scale is not None
|
678 |
+
assert block_shape is None or block_shape[0] == 0
|
679 |
else:
|
680 |
assert A_scale is None
|
681 |
assert B_scale is None
|
682 |
|
683 |
+
EM = sorted_token_ids.shape[0]
|
684 |
+
if A.shape[0] < config["BLOCK_SIZE_M"]:
|
685 |
+
# optimize for small batch_size.
|
686 |
+
# We assume that top_ids of each token is unique, so
|
687 |
+
# so num_valid_experts <= batch_size <= BLOCK_SIZE_M,
|
688 |
+
# and we can skip some invalid blocks.
|
689 |
+
EM = min(sorted_token_ids.shape[0], A.shape[0] * top_k * config["BLOCK_SIZE_M"])
|
690 |
grid = lambda META: (
|
691 |
+
triton.cdiv(EM, META["BLOCK_SIZE_M"])
|
692 |
* triton.cdiv(B.shape[1], META["BLOCK_SIZE_N"]),
|
693 |
)
|
694 |
|
695 |
+
if (
|
696 |
+
(use_int8_w8a16 or use_int4_w4a16)
|
697 |
+
and block_shape is not None
|
698 |
+
and block_shape[1] > 0
|
699 |
+
):
|
700 |
+
assert B_scale is not None and B_scale.ndim == 3
|
701 |
+
assert B_zp is None or B_zp.ndim == 3
|
702 |
+
|
703 |
+
fused_moe_kernel_gptq_awq[grid](
|
704 |
+
A,
|
705 |
+
B,
|
706 |
+
C,
|
707 |
+
B_scale,
|
708 |
+
B_zp,
|
709 |
+
topk_weights,
|
710 |
+
sorted_token_ids,
|
711 |
+
expert_ids,
|
712 |
+
num_tokens_post_padded,
|
713 |
+
B.shape[1],
|
714 |
+
A.shape[1],
|
715 |
+
EM,
|
716 |
+
topk_ids.numel(),
|
717 |
+
A.stride(0),
|
718 |
+
A.stride(1),
|
719 |
+
B.stride(0),
|
720 |
+
B.stride(2),
|
721 |
+
B.stride(1),
|
722 |
+
C.stride(1),
|
723 |
+
C.stride(2),
|
724 |
+
B_scale.stride(0),
|
725 |
+
B_scale.stride(2),
|
726 |
+
B_scale.stride(1),
|
727 |
+
B_zp.stride(0) if B_zp is not None else 0,
|
728 |
+
B_zp.stride(2) if B_zp is not None else 0,
|
729 |
+
B_zp.stride(1) if B_zp is not None else 0,
|
730 |
+
block_k_diviable=A.shape[1] % config["BLOCK_SIZE_K"] == 0,
|
731 |
+
group_size=block_shape[1],
|
732 |
+
MUL_ROUTED_WEIGHT=mul_routed_weight,
|
733 |
+
top_k=top_k,
|
734 |
+
compute_type=compute_type,
|
735 |
+
has_zp=B_zp is not None,
|
736 |
+
use_int4_w4a16=use_int4_w4a16,
|
737 |
+
use_int8_w8a16=use_int8_w8a16,
|
738 |
+
**config,
|
739 |
+
)
|
740 |
+
|
741 |
+
else:
|
742 |
+
fused_moe_kernel[grid](
|
743 |
+
A,
|
744 |
+
B,
|
745 |
+
C,
|
746 |
+
A_scale,
|
747 |
+
B_scale,
|
748 |
+
topk_weights,
|
749 |
+
sorted_token_ids,
|
750 |
+
expert_ids,
|
751 |
+
num_tokens_post_padded,
|
752 |
+
B.shape[1],
|
753 |
+
A.shape[1],
|
754 |
+
EM,
|
755 |
+
topk_ids.numel(),
|
756 |
+
A.stride(0),
|
757 |
+
A.stride(1),
|
758 |
+
B.stride(0),
|
759 |
+
B.stride(2),
|
760 |
+
B.stride(1),
|
761 |
+
C.stride(1),
|
762 |
+
C.stride(2),
|
763 |
+
A_scale.stride(0) if A_scale is not None and A_scale.ndim == 2 else 0,
|
764 |
+
A_scale.stride(1) if A_scale is not None and A_scale.ndim == 2 else 0,
|
765 |
+
B_scale.stride(0) if B_scale is not None and B_scale.ndim >= 2 else 0,
|
766 |
+
B_scale.stride(2) if B_scale is not None and B_scale.ndim == 3 else 0,
|
767 |
+
B_scale.stride(1) if B_scale is not None and B_scale.ndim >= 2 else 0,
|
768 |
+
0 if block_shape is None else block_shape[0],
|
769 |
+
0 if block_shape is None else block_shape[1],
|
770 |
+
MUL_ROUTED_WEIGHT=mul_routed_weight,
|
771 |
+
top_k=top_k,
|
772 |
+
compute_type=compute_type,
|
773 |
+
use_fp8_w8a8=use_fp8_w8a8,
|
774 |
+
use_int8_w8a16=use_int8_w8a16,
|
775 |
+
**config,
|
776 |
+
)
|
777 |
|
778 |
|
779 |
+
# Adapted from: https://github.com/sgl-project/sglang/pull/2628
|
780 |
+
def get_config_file_name(
|
781 |
+
E: int, N: int, dtype: Optional[str], block_shape: Optional[List[int]] = None
|
782 |
+
) -> str:
|
783 |
device_name = current_platform.get_device_name().replace(" ", "_")
|
784 |
dtype_selector = "" if not dtype else f",dtype={dtype}"
|
785 |
+
block_shape_selector = (
|
786 |
+
"" if not block_shape or not all(block_shape) else f",block_shape={block_shape}"
|
787 |
+
)
|
788 |
+
return f"E={E},N={N},device_name={device_name}{dtype_selector}{block_shape_selector}.json" # noqa: E501
|
789 |
|
790 |
|
791 |
+
# Adapted from: https://github.com/sgl-project/sglang/pull/2628
|
792 |
@functools.lru_cache
|
793 |
+
def get_moe_configs(
|
794 |
+
E: int,
|
795 |
+
N: int,
|
796 |
+
dtype: Optional[str],
|
797 |
+
block_n: Optional[int] = None,
|
798 |
+
block_k: Optional[int] = None,
|
799 |
+
) -> Optional[Dict[int, Any]]:
|
800 |
"""
|
801 |
Return optimized configurations for the fused MoE kernel.
|
802 |
|
|
|
808 |
|
809 |
# First look up if an optimized configuration is available in the configs
|
810 |
# directory
|
811 |
+
block_shape = [block_n, block_k] if block_n and block_k else None
|
812 |
+
json_file_name = get_config_file_name(E, N, dtype, block_shape)
|
813 |
|
814 |
config_file_path = os.path.join(
|
815 |
os.path.dirname(os.path.realpath(__file__)), "configs", json_file_name
|
816 |
)
|
817 |
if os.path.exists(config_file_path):
|
818 |
with open(config_file_path) as f:
|
819 |
+
logger.info("Using configuration from %s for MoE layer.", config_file_path)
|
820 |
# If a configuration has been found, return it
|
821 |
return {int(key): val for key, val in json.load(f).items()}
|
822 |
|
823 |
# If no optimized configuration is available, we will use the default
|
824 |
# configuration
|
825 |
+
logger.warning(
|
826 |
+
(
|
827 |
+
"Using default MoE config. Performance might be sub-optimal! "
|
828 |
+
"Config file not found at %s"
|
829 |
+
),
|
830 |
+
config_file_path,
|
831 |
+
)
|
832 |
return None
|
833 |
|
834 |
|
|
|
840 |
topk: int,
|
841 |
dtype: Optional[str],
|
842 |
is_marlin: bool,
|
843 |
+
block_shape: Optional[List[int]] = None,
|
844 |
) -> Dict[str, int]:
|
845 |
+
if dtype == "fp8_w8a8" and block_shape is not None:
|
846 |
+
# Block-wise quant: BLOCK_SIZE_N must be divisible by block_shape[0]
|
847 |
+
# BLOCK_SIZE_K must be divisible by block_shape[1]
|
|
|
|
|
|
|
|
|
|
|
848 |
config = {
|
849 |
+
"BLOCK_SIZE_M": 64,
|
850 |
+
"BLOCK_SIZE_N": block_shape[0],
|
851 |
+
"BLOCK_SIZE_K": block_shape[1],
|
852 |
+
"GROUP_SIZE_M": 32,
|
853 |
+
"num_warps": 4,
|
854 |
+
"num_stages": 3,
|
855 |
}
|
856 |
+
else:
|
857 |
+
config = {
|
858 |
+
"BLOCK_SIZE_M": 64,
|
859 |
+
"BLOCK_SIZE_N": 64,
|
860 |
+
"BLOCK_SIZE_K": 32,
|
861 |
+
"GROUP_SIZE_M": 8,
|
862 |
+
}
|
863 |
+
# A heuristic: fused marlin works faster with this config for small M
|
864 |
+
if M <= E or (is_marlin and M <= 32):
|
865 |
+
config = {
|
866 |
+
"BLOCK_SIZE_M": 16,
|
867 |
+
"BLOCK_SIZE_N": 32,
|
868 |
+
"BLOCK_SIZE_K": 64,
|
869 |
+
"GROUP_SIZE_M": 1,
|
870 |
+
}
|
871 |
return config
|
872 |
|
873 |
|
|
|
877 |
top_k: int,
|
878 |
dtype: Optional[str],
|
879 |
M: int,
|
|
|
880 |
is_marlin: bool = False,
|
881 |
+
block_shape: Optional[List[int]] = None,
|
882 |
):
|
883 |
+
# from vllm.model_executor.layers.fused_moe import get_config
|
884 |
+
# TODO: removed when syncing to vLLM, do we need this?
|
885 |
+
# override_config = get_config()
|
886 |
+
override_config = None
|
887 |
if override_config:
|
888 |
config = override_config
|
889 |
else:
|
890 |
# First try to load optimal config from the file
|
891 |
E, _, N = w2_shape
|
892 |
+
block_n = block_shape[0] if block_shape else 0
|
893 |
+
block_k = block_shape[1] if block_shape else 0
|
894 |
+
configs = get_moe_configs(E, N, dtype, block_n, block_k)
|
895 |
|
896 |
if configs:
|
897 |
# If an optimal configuration map has been found, look up the
|
|
|
899 |
config = configs[min(configs.keys(), key=lambda x: abs(x - M))]
|
900 |
else:
|
901 |
# Else use the default config
|
902 |
+
config = get_default_config(
|
903 |
+
M, E, N, w1_shape[2], top_k, dtype, is_marlin, block_shape
|
904 |
+
)
|
905 |
return config
|
906 |
|
907 |
|
|
|
937 |
return topk_weights, topk_ids
|
938 |
|
939 |
|
940 |
+
# This is used by the Deepseek-V2 and Deepseek-V3 model
|
941 |
+
@torch.compile(dynamic=True, backend=current_platform.simple_compile_backend)
|
942 |
def grouped_topk(
|
943 |
hidden_states: torch.Tensor,
|
944 |
gating_output: torch.Tensor,
|
|
|
946 |
renormalize: bool,
|
947 |
num_expert_group: int = 0,
|
948 |
topk_group: int = 0,
|
949 |
+
scoring_func: str = "softmax",
|
950 |
+
e_score_correction_bias: Optional[torch.Tensor] = None,
|
951 |
):
|
952 |
|
953 |
assert hidden_states.shape[0] == gating_output.shape[0], "Number of tokens mismatch"
|
954 |
|
955 |
+
if scoring_func == "softmax":
|
956 |
+
scores = torch.softmax(gating_output, dim=-1)
|
957 |
+
elif scoring_func == "sigmoid":
|
958 |
+
scores = gating_output.sigmoid()
|
959 |
+
else:
|
960 |
+
raise ValueError(f"Unsupported scoring function: {scoring_func}")
|
961 |
+
|
962 |
+
if e_score_correction_bias is not None:
|
963 |
+
# Store original scores before applying correction bias. We use biased
|
964 |
+
# scores for expert selection but original scores for routing weights
|
965 |
+
original_scores = scores
|
966 |
+
scores = scores + e_score_correction_bias.unsqueeze(0)
|
967 |
+
|
968 |
num_token = scores.shape[0]
|
969 |
group_scores = (
|
970 |
scores.view(num_token, num_expert_group, -1).max(dim=-1).values
|
|
|
980 |
.reshape(num_token, -1)
|
981 |
) # [n, e]
|
982 |
tmp_scores = scores.masked_fill(~score_mask.bool(), 0.0) # [n, e]
|
983 |
+
|
984 |
+
if e_score_correction_bias is not None:
|
985 |
+
topk_ids = torch.topk(tmp_scores, k=topk, dim=-1, sorted=False)[1]
|
986 |
+
# Use original unbiased scores for the routing weights
|
987 |
+
topk_weights = original_scores.gather(1, topk_ids)
|
988 |
+
else:
|
989 |
+
topk_weights, topk_ids = torch.topk(tmp_scores, k=topk, dim=-1, sorted=False)
|
990 |
|
991 |
if renormalize:
|
992 |
topk_weights = topk_weights / topk_weights.sum(dim=-1, keepdim=True)
|
|
|
996 |
|
997 |
def get_config_dtype_str(
|
998 |
dtype: torch.dtype,
|
999 |
+
use_int4_w4a16: Optional[bool] = False,
|
1000 |
use_int8_w8a16: Optional[bool] = False,
|
1001 |
use_fp8_w8a8: Optional[bool] = False,
|
1002 |
):
|
|
|
1004 |
return "fp8_w8a8"
|
1005 |
elif use_int8_w8a16:
|
1006 |
return "int8_w8a16"
|
1007 |
+
elif use_int4_w4a16:
|
1008 |
+
return "int4_w8a16"
|
1009 |
elif dtype == torch.float:
|
1010 |
# avoiding cases where kernel fails when float32 MoE
|
1011 |
# use fp16/bfloat16 configs
|
|
|
1013 |
return None
|
1014 |
|
1015 |
|
1016 |
+
def inplace_fused_experts(
|
1017 |
+
hidden_states: torch.Tensor,
|
1018 |
+
w1: torch.Tensor,
|
1019 |
+
w2: torch.Tensor,
|
1020 |
+
topk_weights: torch.Tensor,
|
1021 |
+
topk_ids: torch.Tensor,
|
1022 |
+
use_fp8_w8a8: bool = False,
|
1023 |
+
use_int8_w8a16: bool = False,
|
1024 |
+
use_int4_w4a16: bool = False,
|
1025 |
+
w1_scale: Optional[torch.Tensor] = None,
|
1026 |
+
w2_scale: Optional[torch.Tensor] = None,
|
1027 |
+
w1_zp: Optional[torch.Tensor] = None,
|
1028 |
+
w2_zp: Optional[torch.Tensor] = None,
|
1029 |
+
a1_scale: Optional[torch.Tensor] = None,
|
1030 |
+
a2_scale: Optional[torch.Tensor] = None,
|
1031 |
+
block_shape: Optional[List[int]] = None,
|
1032 |
+
) -> None:
|
1033 |
+
fused_experts_impl(
|
1034 |
+
hidden_states,
|
1035 |
+
w1,
|
1036 |
+
w2,
|
1037 |
+
topk_weights,
|
1038 |
+
topk_ids,
|
1039 |
+
True,
|
1040 |
+
use_fp8_w8a8,
|
1041 |
+
use_int8_w8a16,
|
1042 |
+
use_int4_w4a16,
|
1043 |
+
w1_scale,
|
1044 |
+
w2_scale,
|
1045 |
+
w1_zp,
|
1046 |
+
w2_zp,
|
1047 |
+
a1_scale,
|
1048 |
+
a2_scale,
|
1049 |
+
block_shape,
|
1050 |
+
)
|
1051 |
+
|
1052 |
+
|
1053 |
+
def outplace_fused_experts(
|
1054 |
+
hidden_states: torch.Tensor,
|
1055 |
+
w1: torch.Tensor,
|
1056 |
+
w2: torch.Tensor,
|
1057 |
+
topk_weights: torch.Tensor,
|
1058 |
+
topk_ids: torch.Tensor,
|
1059 |
+
use_fp8_w8a8: bool = False,
|
1060 |
+
use_int8_w8a16: bool = False,
|
1061 |
+
use_int4_w4a16: bool = False,
|
1062 |
+
w1_scale: Optional[torch.Tensor] = None,
|
1063 |
+
w2_scale: Optional[torch.Tensor] = None,
|
1064 |
+
w1_zp: Optional[torch.Tensor] = None,
|
1065 |
+
w2_zp: Optional[torch.Tensor] = None,
|
1066 |
+
a1_scale: Optional[torch.Tensor] = None,
|
1067 |
+
a2_scale: Optional[torch.Tensor] = None,
|
1068 |
+
block_shape: Optional[List[int]] = None,
|
1069 |
+
) -> torch.Tensor:
|
1070 |
+
return fused_experts_impl(
|
1071 |
+
hidden_states,
|
1072 |
+
w1,
|
1073 |
+
w2,
|
1074 |
+
topk_weights,
|
1075 |
+
topk_ids,
|
1076 |
+
False,
|
1077 |
+
use_fp8_w8a8,
|
1078 |
+
use_int8_w8a16,
|
1079 |
+
use_int4_w4a16,
|
1080 |
+
w1_scale,
|
1081 |
+
w2_scale,
|
1082 |
+
w1_zp,
|
1083 |
+
w2_zp,
|
1084 |
+
a1_scale,
|
1085 |
+
a2_scale,
|
1086 |
+
block_shape,
|
1087 |
+
)
|
1088 |
+
|
1089 |
+
|
1090 |
def fused_experts(
|
1091 |
hidden_states: torch.Tensor,
|
1092 |
w1: torch.Tensor,
|
|
|
1094 |
topk_weights: torch.Tensor,
|
1095 |
topk_ids: torch.Tensor,
|
1096 |
inplace: bool = False,
|
|
|
1097 |
use_fp8_w8a8: bool = False,
|
1098 |
use_int8_w8a16: bool = False,
|
1099 |
+
use_int4_w4a16: bool = False,
|
1100 |
+
w1_scale: Optional[torch.Tensor] = None,
|
1101 |
+
w2_scale: Optional[torch.Tensor] = None,
|
1102 |
+
w1_zp: Optional[torch.Tensor] = None,
|
1103 |
+
w2_zp: Optional[torch.Tensor] = None,
|
1104 |
+
a1_scale: Optional[torch.Tensor] = None,
|
1105 |
+
a2_scale: Optional[torch.Tensor] = None,
|
1106 |
+
block_shape: Optional[List[int]] = None,
|
1107 |
+
):
|
1108 |
+
if inplace:
|
1109 |
+
inplace_fused_experts(
|
1110 |
+
hidden_states,
|
1111 |
+
w1,
|
1112 |
+
w2,
|
1113 |
+
topk_weights,
|
1114 |
+
topk_ids,
|
1115 |
+
use_fp8_w8a8,
|
1116 |
+
use_int8_w8a16,
|
1117 |
+
use_int4_w4a16,
|
1118 |
+
w1_scale,
|
1119 |
+
w2_scale,
|
1120 |
+
w1_zp,
|
1121 |
+
w2_zp,
|
1122 |
+
a1_scale,
|
1123 |
+
a2_scale,
|
1124 |
+
block_shape,
|
1125 |
+
)
|
1126 |
+
return hidden_states
|
1127 |
+
else:
|
1128 |
+
return outplace_fused_experts(
|
1129 |
+
hidden_states,
|
1130 |
+
w1,
|
1131 |
+
w2,
|
1132 |
+
topk_weights,
|
1133 |
+
topk_ids,
|
1134 |
+
use_fp8_w8a8,
|
1135 |
+
use_int8_w8a16,
|
1136 |
+
use_int4_w4a16,
|
1137 |
+
w1_scale,
|
1138 |
+
w2_scale,
|
1139 |
+
w1_zp,
|
1140 |
+
w2_zp,
|
1141 |
+
a1_scale,
|
1142 |
+
a2_scale,
|
1143 |
+
block_shape,
|
1144 |
+
)
|
1145 |
+
|
1146 |
+
|
1147 |
+
def fused_experts_impl(
|
1148 |
+
hidden_states: torch.Tensor,
|
1149 |
+
w1: torch.Tensor,
|
1150 |
+
w2: torch.Tensor,
|
1151 |
+
topk_weights: torch.Tensor,
|
1152 |
+
topk_ids: torch.Tensor,
|
1153 |
+
inplace: bool = False,
|
1154 |
+
use_fp8_w8a8: bool = False,
|
1155 |
+
use_int8_w8a16: bool = False,
|
1156 |
+
use_int4_w4a16: bool = False,
|
1157 |
w1_scale: Optional[torch.Tensor] = None,
|
1158 |
w2_scale: Optional[torch.Tensor] = None,
|
1159 |
+
w1_zp: Optional[torch.Tensor] = None,
|
1160 |
+
w2_zp: Optional[torch.Tensor] = None,
|
1161 |
a1_scale: Optional[torch.Tensor] = None,
|
1162 |
a2_scale: Optional[torch.Tensor] = None,
|
1163 |
+
block_shape: Optional[List[int]] = None,
|
1164 |
):
|
1165 |
# Check constraints.
|
1166 |
+
if use_int4_w4a16:
|
1167 |
+
assert hidden_states.shape[1] // 2 == w1.shape[2], "Hidden size mismatch"
|
1168 |
+
else:
|
1169 |
+
assert hidden_states.shape[1] == w1.shape[2], "Hidden size mismatch"
|
1170 |
+
|
1171 |
assert topk_weights.shape == topk_ids.shape, "topk shape mismatch"
|
1172 |
assert hidden_states.is_contiguous(), "Hidden_states must be contiguous"
|
1173 |
assert w1.is_contiguous(), "Expert weights1 must be contiguous"
|
|
|
1183 |
config_dtype = get_config_dtype_str(
|
1184 |
use_fp8_w8a8=use_fp8_w8a8,
|
1185 |
use_int8_w8a16=use_int8_w8a16,
|
1186 |
+
use_int4_w4a16=use_int4_w4a16,
|
1187 |
dtype=hidden_states.dtype,
|
1188 |
)
|
1189 |
|
|
|
1193 |
w2.shape,
|
1194 |
topk_ids.shape[1],
|
1195 |
config_dtype,
|
1196 |
+
block_shape=block_shape,
|
1197 |
)
|
1198 |
|
1199 |
config = get_config_func(M)
|
|
|
1214 |
dtype=hidden_states.dtype,
|
1215 |
)
|
1216 |
|
1217 |
+
if hidden_states.dtype == torch.bfloat16:
|
1218 |
+
compute_type = tl.bfloat16
|
1219 |
+
elif hidden_states.dtype == torch.float16:
|
1220 |
+
compute_type = tl.float16
|
1221 |
+
elif hidden_states.dtype == torch.float32:
|
1222 |
+
compute_type = tl.float32
|
1223 |
+
else:
|
1224 |
+
raise ValueError(f"Unsupported compute_type: {hidden_states.dtype}")
|
1225 |
|
1226 |
if inplace:
|
1227 |
out_hidden_states = hidden_states
|
|
|
1262 |
intermediate_cache1,
|
1263 |
a1_scale,
|
1264 |
w1_scale,
|
1265 |
+
w1_zp,
|
1266 |
curr_topk_weights,
|
1267 |
curr_topk_ids,
|
1268 |
sorted_token_ids,
|
|
|
1274 |
compute_type=compute_type,
|
1275 |
use_fp8_w8a8=use_fp8_w8a8,
|
1276 |
use_int8_w8a16=use_int8_w8a16,
|
1277 |
+
use_int4_w4a16=use_int4_w4a16,
|
1278 |
+
block_shape=block_shape,
|
1279 |
)
|
1280 |
|
1281 |
ops.silu_and_mul(intermediate_cache2, intermediate_cache1.view(-1, N))
|
|
|
1286 |
intermediate_cache3,
|
1287 |
a2_scale,
|
1288 |
w2_scale,
|
1289 |
+
w2_zp,
|
1290 |
curr_topk_weights,
|
1291 |
curr_topk_ids,
|
1292 |
sorted_token_ids,
|
|
|
1298 |
compute_type=compute_type,
|
1299 |
use_fp8_w8a8=use_fp8_w8a8,
|
1300 |
use_int8_w8a16=use_int8_w8a16,
|
1301 |
+
use_int4_w4a16=use_int4_w4a16,
|
1302 |
+
block_shape=block_shape,
|
1303 |
)
|
1304 |
|
1305 |
ops.moe_sum(
|
|
|
1317 |
topk: int,
|
1318 |
renormalize: bool,
|
1319 |
inplace: bool = False,
|
|
|
1320 |
use_grouped_topk: bool = False,
|
1321 |
num_expert_group: Optional[int] = None,
|
1322 |
topk_group: Optional[int] = None,
|
1323 |
custom_routing_function: Optional[Callable] = None,
|
1324 |
use_fp8_w8a8: bool = False,
|
1325 |
use_int8_w8a16: bool = False,
|
1326 |
+
use_int4_w4a16: bool = False,
|
1327 |
w1_scale: Optional[torch.Tensor] = None,
|
1328 |
w2_scale: Optional[torch.Tensor] = None,
|
1329 |
+
w1_zp: Optional[torch.Tensor] = None,
|
1330 |
+
w2_zp: Optional[torch.Tensor] = None,
|
1331 |
a1_scale: Optional[torch.Tensor] = None,
|
1332 |
a2_scale: Optional[torch.Tensor] = None,
|
1333 |
+
block_shape: Optional[List[int]] = None,
|
1334 |
) -> torch.Tensor:
|
1335 |
"""
|
1336 |
This function computes a Mixture of Experts (MoE) layer using two sets of
|
|
|
1346 |
- renormalize (bool): If True, renormalize the top-k weights to sum to 1.
|
1347 |
- inplace (bool): If True, perform the operation in-place.
|
1348 |
Defaults to False.
|
|
|
|
|
1349 |
- num_expert_group: Optional[int]: additional parameter for grouped_topk
|
1350 |
- topk_group: Optional[int]: additional parameter for grouped_topk
|
1351 |
- use_grouped_topk: If True, use grouped_topk instead of fused_topk
|
1352 |
note: Deepseekv2 model uses grouped_topk
|
1353 |
- use_fp8_w8a8 (bool): If True, use fp8 arithmetic to compute the inner
|
1354 |
products for w1 and w2. Defaults to False.
|
1355 |
+
- use_int8_w8a16 (bool): If True, use matmul of int8 weight and bf16/fp16
|
1356 |
+
activation to compute the inner products for w1 and w2.
|
1357 |
+
Defaults to False.
|
1358 |
+
- use_int4_w4a16 (bool): If True, use matmul of int4 weight and bf16/fp16
|
1359 |
+
activation to compute the inner products for w1 and w2.
|
1360 |
+
Defaults to False.
|
1361 |
- w1_scale (Optional[torch.Tensor]): Optional scale to be used for
|
1362 |
w1.
|
1363 |
- w2_scale (Optional[torch.Tensor]): Optional scale to be used for
|
1364 |
w2.
|
1365 |
+
- a1_scale (Optional[torch.Tensor]): Optional scale to be used for
|
1366 |
+
a1.
|
1367 |
+
- a2_scale (Optional[torch.Tensor]): Optional scale to be used for
|
1368 |
+
a2.
|
1369 |
+
- block_shape: (Optional[List[int]]): Optional block size for block-wise
|
1370 |
+
quantization.
|
1371 |
|
1372 |
Returns:
|
1373 |
- torch.Tensor: The output tensor after applying the MoE layer.
|
|
|
1401 |
topk_weights,
|
1402 |
topk_ids,
|
1403 |
inplace=inplace,
|
|
|
1404 |
use_fp8_w8a8=use_fp8_w8a8,
|
1405 |
use_int8_w8a16=use_int8_w8a16,
|
1406 |
+
use_int4_w4a16=use_int4_w4a16,
|
1407 |
w1_scale=w1_scale,
|
1408 |
w2_scale=w2_scale,
|
1409 |
+
w1_zp=w1_zp,
|
1410 |
+
w2_zp=w2_zp,
|
1411 |
a1_scale=a1_scale,
|
1412 |
a2_scale=a2_scale,
|
1413 |
+
block_shape=block_shape,
|
1414 |
)
|
build/torch25-cxx11-cu124-x86_64-linux/moe/platforms.py
CHANGED
@@ -1,22 +1,32 @@
|
|
1 |
-
from
|
2 |
-
import os
|
3 |
-
from functools import lru_cache, wraps
|
4 |
|
5 |
import torch
|
6 |
|
7 |
IS_ROCM = torch.version.hip is not None
|
8 |
|
9 |
-
|
|
|
|
|
|
|
|
|
|
|
10 |
@classmethod
|
11 |
@lru_cache(maxsize=8)
|
12 |
def get_device_name(cls, device_id: int = 0) -> str:
|
13 |
return torch.cuda.get_device_name(0)
|
14 |
|
15 |
-
|
|
|
|
|
|
|
|
|
16 |
@classmethod
|
17 |
@lru_cache(maxsize=8)
|
18 |
def get_device_name(cls, device_id: int = 0) -> str:
|
19 |
return torch.cuda.get_device_name(device_id)
|
20 |
|
|
|
|
|
|
|
21 |
|
22 |
current_platform = RocmPlatform() if IS_ROCM else CudaPlatform()
|
|
|
1 |
+
from functools import lru_cache
|
|
|
|
|
2 |
|
3 |
import torch
|
4 |
|
5 |
IS_ROCM = torch.version.hip is not None
|
6 |
|
7 |
+
|
8 |
+
class Platform:
|
9 |
+
simple_compile_backend: str = "inductor"
|
10 |
+
|
11 |
+
|
12 |
+
class CudaPlatform(Platform):
|
13 |
@classmethod
|
14 |
@lru_cache(maxsize=8)
|
15 |
def get_device_name(cls, device_id: int = 0) -> str:
|
16 |
return torch.cuda.get_device_name(0)
|
17 |
|
18 |
+
def is_rocm(self):
|
19 |
+
return False
|
20 |
+
|
21 |
+
|
22 |
+
class RocmPlatform(Platform):
|
23 |
@classmethod
|
24 |
@lru_cache(maxsize=8)
|
25 |
def get_device_name(cls, device_id: int = 0) -> str:
|
26 |
return torch.cuda.get_device_name(device_id)
|
27 |
|
28 |
+
def is_rocm(self):
|
29 |
+
return True
|
30 |
+
|
31 |
|
32 |
current_platform = RocmPlatform() if IS_ROCM else CudaPlatform()
|
build/torch25-cxx98-cu118-x86_64-linux/moe/{_moe_uhyif3wslpwak.abi3.so → _moe_5uyw6qhdybj5e.abi3.so}
RENAMED
@@ -1,3 +1,3 @@
|
|
1 |
version https://git-lfs.github.com/spec/v1
|
2 |
-
oid sha256:
|
3 |
-
size
|
|
|
1 |
version https://git-lfs.github.com/spec/v1
|
2 |
+
oid sha256:acfcb8be6199c8e08519a1db8ec8122f7ec69a96c798d9c26e681469ba326782
|
3 |
+
size 85815472
|
build/torch25-cxx98-cu118-x86_64-linux/moe/_ops.py
CHANGED
@@ -1,9 +1,9 @@
|
|
1 |
import torch
|
2 |
-
from . import
|
3 |
-
ops = torch.ops.
|
4 |
|
5 |
def add_op_namespace_prefix(op_name: str):
|
6 |
"""
|
7 |
Prefix op by namespace.
|
8 |
"""
|
9 |
-
return f"
|
|
|
1 |
import torch
|
2 |
+
from . import _moe_5uyw6qhdybj5e
|
3 |
+
ops = torch.ops._moe_5uyw6qhdybj5e
|
4 |
|
5 |
def add_op_namespace_prefix(op_name: str):
|
6 |
"""
|
7 |
Prefix op by namespace.
|
8 |
"""
|
9 |
+
return f"_moe_5uyw6qhdybj5e::{op_name}"
|
build/torch25-cxx98-cu118-x86_64-linux/moe/fp8.py
CHANGED
@@ -1,6 +1,11 @@
|
|
|
|
|
|
1 |
import torch
|
|
|
|
|
2 |
|
3 |
-
|
|
|
4 |
|
5 |
|
6 |
def is_hip() -> bool:
|
@@ -49,15 +54,179 @@ def scaled_fp8_quant(
|
|
49 |
if scale is None:
|
50 |
if use_per_token_if_dynamic:
|
51 |
scale = torch.empty((shape[0], 1), device=input.device, dtype=torch.float32)
|
52 |
-
|
53 |
-
output, input, scale, scale_ub
|
54 |
-
)
|
55 |
else:
|
56 |
scale = torch.zeros(1, device=input.device, dtype=torch.float32)
|
57 |
-
|
58 |
else:
|
59 |
# num_token_padding not implemented for this case
|
60 |
assert scale.numel() == 1 or num_token_padding is None
|
61 |
-
|
62 |
|
63 |
return output, scale
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
1 |
+
from typing import Tuple, Optional, Union
|
2 |
+
|
3 |
import torch
|
4 |
+
import triton
|
5 |
+
import triton.language as tl
|
6 |
|
7 |
+
|
8 |
+
from ._ops import ops
|
9 |
|
10 |
|
11 |
def is_hip() -> bool:
|
|
|
54 |
if scale is None:
|
55 |
if use_per_token_if_dynamic:
|
56 |
scale = torch.empty((shape[0], 1), device=input.device, dtype=torch.float32)
|
57 |
+
ops.dynamic_per_token_scaled_fp8_quant(output, input, scale, scale_ub)
|
|
|
|
|
58 |
else:
|
59 |
scale = torch.zeros(1, device=input.device, dtype=torch.float32)
|
60 |
+
ops.dynamic_scaled_fp8_quant(output, input, scale)
|
61 |
else:
|
62 |
# num_token_padding not implemented for this case
|
63 |
assert scale.numel() == 1 or num_token_padding is None
|
64 |
+
ops.static_scaled_fp8_quant(output, input, scale)
|
65 |
|
66 |
return output, scale
|
67 |
+
|
68 |
+
|
69 |
+
@triton.jit
|
70 |
+
def _per_token_group_quant_fp8(
|
71 |
+
# Pointers to inputs and output
|
72 |
+
y_ptr,
|
73 |
+
y_q_ptr,
|
74 |
+
y_s_ptr,
|
75 |
+
group_size,
|
76 |
+
# Avoid to divide zero
|
77 |
+
eps,
|
78 |
+
# Information for float8
|
79 |
+
fp8_min,
|
80 |
+
fp8_max,
|
81 |
+
# Meta-parameters
|
82 |
+
BLOCK: tl.constexpr,
|
83 |
+
):
|
84 |
+
"""A Triton-accelerated function to perform per-token-group
|
85 |
+
quantization on a tensor.
|
86 |
+
This function converts the tensor values into float8 values.
|
87 |
+
"""
|
88 |
+
# Map the program id to the row of X and Y it should compute.
|
89 |
+
g_id = tl.program_id(0)
|
90 |
+
y_ptr += g_id * group_size
|
91 |
+
y_q_ptr += g_id * group_size
|
92 |
+
y_s_ptr += g_id
|
93 |
+
|
94 |
+
cols = tl.arange(0, BLOCK) # N <= BLOCK
|
95 |
+
mask = cols < group_size
|
96 |
+
|
97 |
+
y = tl.load(y_ptr + cols, mask=mask, other=0.0).to(tl.float32)
|
98 |
+
# Quant
|
99 |
+
_absmax = tl.maximum(tl.max(tl.abs(y)), eps)
|
100 |
+
y_s = _absmax / fp8_max
|
101 |
+
y_q = tl.clamp(y / y_s, fp8_min, fp8_max).to(y_q_ptr.dtype.element_ty)
|
102 |
+
|
103 |
+
tl.store(y_q_ptr + cols, y_q, mask=mask)
|
104 |
+
tl.store(y_s_ptr, y_s)
|
105 |
+
|
106 |
+
|
107 |
+
@triton.jit
|
108 |
+
def _per_token_group_quant_fp8_colmajor(
|
109 |
+
# Pointers to inputs and output
|
110 |
+
y_ptr,
|
111 |
+
y_q_ptr,
|
112 |
+
y_s_ptr,
|
113 |
+
group_size,
|
114 |
+
# Num columns of y
|
115 |
+
y_num_columns,
|
116 |
+
# Stride from one column to the next of y_s
|
117 |
+
y_s_col_stride,
|
118 |
+
# Avoid to divide zero
|
119 |
+
eps,
|
120 |
+
# Information for float8
|
121 |
+
fp8_min,
|
122 |
+
fp8_max,
|
123 |
+
# Meta-parameters
|
124 |
+
BLOCK: tl.constexpr,
|
125 |
+
):
|
126 |
+
"""A Triton-accelerated function to perform per-token-group
|
127 |
+
quantization on a tensor.
|
128 |
+
This function converts the tensor values into float8 values.
|
129 |
+
"""
|
130 |
+
# Map the program id to the row of X and Y it should compute.
|
131 |
+
g_id = tl.program_id(0)
|
132 |
+
y_ptr += g_id * group_size
|
133 |
+
y_q_ptr += g_id * group_size
|
134 |
+
|
135 |
+
# Convert g_id the flattened block coordinate to 2D so we can index
|
136 |
+
# into the output y_scales matrix
|
137 |
+
blocks_per_row = y_num_columns // group_size
|
138 |
+
scale_col = g_id % blocks_per_row
|
139 |
+
scale_row = g_id // blocks_per_row
|
140 |
+
y_s_ptr += scale_col * y_s_col_stride + scale_row
|
141 |
+
|
142 |
+
cols = tl.arange(0, BLOCK) # group_size <= BLOCK
|
143 |
+
mask = cols < group_size
|
144 |
+
|
145 |
+
y = tl.load(y_ptr + cols, mask=mask, other=0.0).to(tl.float32)
|
146 |
+
# Quant
|
147 |
+
_absmax = tl.maximum(tl.max(tl.abs(y)), eps)
|
148 |
+
y_s = _absmax / fp8_max
|
149 |
+
y_q = tl.clamp(y / y_s, fp8_min, fp8_max).to(y_q_ptr.dtype.element_ty)
|
150 |
+
|
151 |
+
tl.store(y_q_ptr + cols, y_q, mask=mask)
|
152 |
+
tl.store(y_s_ptr, y_s)
|
153 |
+
|
154 |
+
|
155 |
+
def per_token_group_quant_fp8(
|
156 |
+
x: torch.Tensor,
|
157 |
+
group_size: int,
|
158 |
+
eps: float = 1e-10,
|
159 |
+
dtype: Optional[torch.dtype] = None,
|
160 |
+
column_major_scales: bool = False,
|
161 |
+
) -> Tuple[torch.Tensor, torch.Tensor]:
|
162 |
+
"""Function to perform per-token-group quantization on an input tensor `x`.
|
163 |
+
It converts the tensor values into signed float8 values and returns the
|
164 |
+
quantized tensor along with the scaling factor used for quantization.
|
165 |
+
Args:
|
166 |
+
x: The input tensor with ndim >= 2.
|
167 |
+
group_size: The group size used for quantization.
|
168 |
+
eps: The minimum to avoid dividing zero.
|
169 |
+
dtype: The dype of output tensor. Note that only `torch.float8_e4m3fn`
|
170 |
+
is supported for now.
|
171 |
+
Returns:
|
172 |
+
Tuple[torch.Tensor, torch.Tensor]: The quantized tensor and the
|
173 |
+
scaling factor for quantization.
|
174 |
+
"""
|
175 |
+
if dtype is None:
|
176 |
+
dtype = (
|
177 |
+
torch.float8_e4m3fnuz if current_platform.is_rocm() else torch.float8_e4m3fn
|
178 |
+
)
|
179 |
+
assert x.shape[-1] % group_size == 0, (
|
180 |
+
f"the last dimension of `x` {x.shape[-1]} must be divisible "
|
181 |
+
f"by `group_size` {group_size}"
|
182 |
+
)
|
183 |
+
assert x.is_contiguous(), "`x` must be contiguous"
|
184 |
+
|
185 |
+
finfo = torch.finfo(dtype)
|
186 |
+
fp8_min = finfo.min
|
187 |
+
fp8_max = finfo.max
|
188 |
+
|
189 |
+
x_q = torch.empty_like(x, device=x.device, dtype=dtype)
|
190 |
+
M = x.numel() // group_size
|
191 |
+
N = group_size
|
192 |
+
if column_major_scales:
|
193 |
+
shape = (x.shape[-1] // group_size,) + x.shape[:-1]
|
194 |
+
x_s = torch.empty(shape, device=x.device, dtype=torch.float32).permute(-1, -2)
|
195 |
+
else:
|
196 |
+
shape = x.shape[:-1] + (x.shape[-1] // group_size,)
|
197 |
+
x_s = torch.empty(shape, device=x.device, dtype=torch.float32)
|
198 |
+
|
199 |
+
BLOCK = triton.next_power_of_2(N)
|
200 |
+
# heuristics for number of warps
|
201 |
+
num_warps = min(max(BLOCK // 256, 1), 8)
|
202 |
+
num_stages = 1
|
203 |
+
if column_major_scales:
|
204 |
+
_per_token_group_quant_fp8_colmajor[(M,)](
|
205 |
+
x,
|
206 |
+
x_q,
|
207 |
+
x_s,
|
208 |
+
group_size,
|
209 |
+
x.shape[1],
|
210 |
+
x_s.stride(1),
|
211 |
+
eps,
|
212 |
+
fp8_min=fp8_min,
|
213 |
+
fp8_max=fp8_max,
|
214 |
+
BLOCK=BLOCK,
|
215 |
+
num_warps=num_warps,
|
216 |
+
num_stages=num_stages,
|
217 |
+
)
|
218 |
+
else:
|
219 |
+
_per_token_group_quant_fp8[(M,)](
|
220 |
+
x,
|
221 |
+
x_q,
|
222 |
+
x_s,
|
223 |
+
group_size,
|
224 |
+
eps,
|
225 |
+
fp8_min=fp8_min,
|
226 |
+
fp8_max=fp8_max,
|
227 |
+
BLOCK=BLOCK,
|
228 |
+
num_warps=num_warps,
|
229 |
+
num_stages=num_stages,
|
230 |
+
)
|
231 |
+
|
232 |
+
return x_q, x_s
|
build/torch25-cxx98-cu118-x86_64-linux/moe/fused_marlin_moe.py
CHANGED
@@ -40,7 +40,6 @@ def single_marlin_moe(
|
|
40 |
g_idx: Optional[torch.Tensor] = None,
|
41 |
sort_indices: Optional[torch.Tensor] = None,
|
42 |
w_zeros: Optional[torch.Tensor] = None,
|
43 |
-
override_config: Optional[Dict[str, Any]] = None,
|
44 |
num_bits: int = 8,
|
45 |
is_k_full: bool = True,
|
46 |
) -> torch.Tensor:
|
@@ -61,8 +60,6 @@ def single_marlin_moe(
|
|
61 |
- topk (int): The number of top-k experts to select.
|
62 |
- renormalize (bool): If True, renormalize the top-k weights to sum to 1.
|
63 |
- w_zeros (Optional[torch.Tensor]): Optional zero points to be used for w.
|
64 |
-
- override_config (Optional[Dict[str, Any]]): Optional override
|
65 |
-
for the kernel configuration.
|
66 |
- num_bits (bool): The number of bits in expert weights quantization.
|
67 |
|
68 |
Returns:
|
@@ -90,7 +87,6 @@ def single_marlin_moe(
|
|
90 |
w.shape,
|
91 |
topk_ids.shape[1],
|
92 |
None,
|
93 |
-
override_config=override_config,
|
94 |
is_marlin=True,
|
95 |
)
|
96 |
config = get_config_func(M)
|
@@ -154,6 +150,25 @@ def single_marlin_moe(
|
|
154 |
return torch.sum(intermediate_cache.view(*intermediate_cache.shape), dim=1)
|
155 |
|
156 |
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
157 |
def fused_marlin_moe(
|
158 |
hidden_states: torch.Tensor,
|
159 |
w1: torch.Tensor,
|
@@ -169,7 +184,6 @@ def fused_marlin_moe(
|
|
169 |
sort_indices2: Optional[torch.Tensor] = None,
|
170 |
w1_zeros: Optional[torch.Tensor] = None,
|
171 |
w2_zeros: Optional[torch.Tensor] = None,
|
172 |
-
override_config: Optional[Dict[str, Any]] = None,
|
173 |
num_bits: int = 8,
|
174 |
is_k_full: bool = True,
|
175 |
) -> torch.Tensor:
|
@@ -193,8 +207,6 @@ def fused_marlin_moe(
|
|
193 |
permutation.
|
194 |
- topk_weights (torch.Tensor): Top-k weights.
|
195 |
- topk_ids (torch.Tensor): Indices of topk-k elements.
|
196 |
-
- override_config (Optional[Dict[str, Any]]): Optional override
|
197 |
-
for the kernel configuration.
|
198 |
- w1_zeros (Optional[torch.Tensor]): Optional zero points to be used for w1.
|
199 |
- w2_zeros (Optional[torch.Tensor]): Optional zero points to be used for w2.
|
200 |
- num_bits (bool): The number of bits in expert weights quantization.
|
@@ -248,7 +260,6 @@ def fused_marlin_moe(
|
|
248 |
w2.shape,
|
249 |
topk_ids.shape[1],
|
250 |
None,
|
251 |
-
override_config=override_config,
|
252 |
is_marlin=True,
|
253 |
)
|
254 |
config = get_config_func(M)
|
@@ -350,6 +361,30 @@ def fused_marlin_moe(
|
|
350 |
return torch.sum(intermediate_cache3.view(*intermediate_cache3.shape), dim=1)
|
351 |
|
352 |
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
353 |
if hasattr(ops, "marlin_gemm_moe"):
|
354 |
|
355 |
@register_fake(add_op_namespace_prefix("marlin_gemm_moe"))
|
|
|
40 |
g_idx: Optional[torch.Tensor] = None,
|
41 |
sort_indices: Optional[torch.Tensor] = None,
|
42 |
w_zeros: Optional[torch.Tensor] = None,
|
|
|
43 |
num_bits: int = 8,
|
44 |
is_k_full: bool = True,
|
45 |
) -> torch.Tensor:
|
|
|
60 |
- topk (int): The number of top-k experts to select.
|
61 |
- renormalize (bool): If True, renormalize the top-k weights to sum to 1.
|
62 |
- w_zeros (Optional[torch.Tensor]): Optional zero points to be used for w.
|
|
|
|
|
63 |
- num_bits (bool): The number of bits in expert weights quantization.
|
64 |
|
65 |
Returns:
|
|
|
87 |
w.shape,
|
88 |
topk_ids.shape[1],
|
89 |
None,
|
|
|
90 |
is_marlin=True,
|
91 |
)
|
92 |
config = get_config_func(M)
|
|
|
150 |
return torch.sum(intermediate_cache.view(*intermediate_cache.shape), dim=1)
|
151 |
|
152 |
|
153 |
+
if hasattr(ops, "single_marlin_gemm_moe"):
|
154 |
+
|
155 |
+
@register_fake(add_op_namespace_prefix("single_marlin_gemm_moe"))
|
156 |
+
def single_marlin_moe_fake(
|
157 |
+
hidden_states: torch.Tensor,
|
158 |
+
w: torch.Tensor,
|
159 |
+
scales: torch.Tensor,
|
160 |
+
gating_output: torch.Tensor,
|
161 |
+
topk: int,
|
162 |
+
renormalize: bool,
|
163 |
+
g_idx: Optional[torch.Tensor] = None,
|
164 |
+
sort_indices: Optional[torch.Tensor] = None,
|
165 |
+
w_zeros: Optional[torch.Tensor] = None,
|
166 |
+
num_bits: int = 8,
|
167 |
+
is_k_full: bool = True,
|
168 |
+
) -> torch.Tensor:
|
169 |
+
return torch.empty_like(hidden_states)
|
170 |
+
|
171 |
+
|
172 |
def fused_marlin_moe(
|
173 |
hidden_states: torch.Tensor,
|
174 |
w1: torch.Tensor,
|
|
|
184 |
sort_indices2: Optional[torch.Tensor] = None,
|
185 |
w1_zeros: Optional[torch.Tensor] = None,
|
186 |
w2_zeros: Optional[torch.Tensor] = None,
|
|
|
187 |
num_bits: int = 8,
|
188 |
is_k_full: bool = True,
|
189 |
) -> torch.Tensor:
|
|
|
207 |
permutation.
|
208 |
- topk_weights (torch.Tensor): Top-k weights.
|
209 |
- topk_ids (torch.Tensor): Indices of topk-k elements.
|
|
|
|
|
210 |
- w1_zeros (Optional[torch.Tensor]): Optional zero points to be used for w1.
|
211 |
- w2_zeros (Optional[torch.Tensor]): Optional zero points to be used for w2.
|
212 |
- num_bits (bool): The number of bits in expert weights quantization.
|
|
|
260 |
w2.shape,
|
261 |
topk_ids.shape[1],
|
262 |
None,
|
|
|
263 |
is_marlin=True,
|
264 |
)
|
265 |
config = get_config_func(M)
|
|
|
361 |
return torch.sum(intermediate_cache3.view(*intermediate_cache3.shape), dim=1)
|
362 |
|
363 |
|
364 |
+
if hasattr(ops, "fused_marlin_moe"):
|
365 |
+
|
366 |
+
@register_fake(add_op_namespace_prefix("fused_marlin_moe"))
|
367 |
+
def fused_marlin_moe_fake(
|
368 |
+
hidden_states: torch.Tensor,
|
369 |
+
w1: torch.Tensor,
|
370 |
+
w2: torch.Tensor,
|
371 |
+
w1_scale: torch.Tensor,
|
372 |
+
w2_scale: torch.Tensor,
|
373 |
+
gating_output: torch.Tensor,
|
374 |
+
topk_weights: torch.Tensor,
|
375 |
+
topk_ids: torch.Tensor,
|
376 |
+
g_idx1: Optional[torch.Tensor] = None,
|
377 |
+
g_idx2: Optional[torch.Tensor] = None,
|
378 |
+
sort_indices1: Optional[torch.Tensor] = None,
|
379 |
+
sort_indices2: Optional[torch.Tensor] = None,
|
380 |
+
w1_zeros: Optional[torch.Tensor] = None,
|
381 |
+
w2_zeros: Optional[torch.Tensor] = None,
|
382 |
+
num_bits: int = 8,
|
383 |
+
is_k_full: bool = True,
|
384 |
+
) -> torch.Tensor:
|
385 |
+
return torch.empty_like(hidden_states)
|
386 |
+
|
387 |
+
|
388 |
if hasattr(ops, "marlin_gemm_moe"):
|
389 |
|
390 |
@register_fake(add_op_namespace_prefix("marlin_gemm_moe"))
|
build/torch25-cxx98-cu118-x86_64-linux/moe/fused_moe.py
CHANGED
@@ -1,21 +1,242 @@
|
|
|
|
1 |
"""Fused MoE kernel."""
|
2 |
|
3 |
import functools
|
4 |
import json
|
|
|
5 |
import os
|
6 |
-
from typing import Any, Callable, Dict, Optional, Tuple
|
7 |
|
8 |
import torch
|
9 |
import triton
|
10 |
import triton.language as tl
|
11 |
|
|
|
12 |
from ._ops import ops
|
13 |
-
from .fp8 import scaled_fp8_quant
|
14 |
from .platforms import current_platform
|
15 |
|
|
|
|
|
|
|
16 |
VLLM_FUSED_MOE_CHUNK_SIZE = int(os.getenv("VLLM_FUSED_MOE_CHUNK_SIZE", "32768"))
|
17 |
|
18 |
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
19 |
@triton.jit
|
20 |
def fused_moe_kernel(
|
21 |
# Pointers to matrices
|
@@ -44,8 +265,14 @@ def fused_moe_kernel(
|
|
44 |
stride_bn,
|
45 |
stride_cm,
|
46 |
stride_cn,
|
|
|
|
|
47 |
stride_bse,
|
|
|
48 |
stride_bsn,
|
|
|
|
|
|
|
49 |
# Meta-parameters
|
50 |
BLOCK_SIZE_M: tl.constexpr,
|
51 |
BLOCK_SIZE_N: tl.constexpr,
|
@@ -105,17 +332,17 @@ def fused_moe_kernel(
|
|
105 |
num_tokens_post_padded = tl.load(num_tokens_post_padded_ptr)
|
106 |
if pid_m * BLOCK_SIZE_M >= num_tokens_post_padded:
|
107 |
return
|
108 |
-
offs_token_id = pid_m * BLOCK_SIZE_M + tl.arange(0, BLOCK_SIZE_M)
|
109 |
offs_token = tl.load(sorted_token_ids_ptr + offs_token_id)
|
110 |
token_mask = offs_token < num_valid_tokens
|
111 |
|
112 |
-
offs_bn = (pid_n * BLOCK_SIZE_N + tl.arange(0, BLOCK_SIZE_N)) % N
|
113 |
offs_k = tl.arange(0, BLOCK_SIZE_K)
|
114 |
a_ptrs = a_ptr + (
|
115 |
offs_token[:, None] // top_k * stride_am + offs_k[None, :] * stride_ak
|
116 |
)
|
117 |
|
118 |
-
off_experts = tl.load(expert_ids_ptr + pid_m)
|
119 |
b_ptrs = (
|
120 |
b_ptr
|
121 |
+ off_experts * stride_be
|
@@ -128,8 +355,15 @@ def fused_moe_kernel(
|
|
128 |
b_scale = tl.load(b_scale_ptrs)
|
129 |
|
130 |
if use_fp8_w8a8:
|
131 |
-
|
132 |
-
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
133 |
|
134 |
# -----------------------------------------------------------
|
135 |
# Iterate to compute a block of the C matrix.
|
@@ -151,7 +385,17 @@ def fused_moe_kernel(
|
|
151 |
if use_int8_w8a16:
|
152 |
accumulator = tl.dot(a, b.to(compute_type), acc=accumulator)
|
153 |
elif use_fp8_w8a8:
|
154 |
-
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
155 |
else:
|
156 |
accumulator += tl.dot(a, b)
|
157 |
# Advance the ptrs to the next K block.
|
@@ -164,7 +408,10 @@ def fused_moe_kernel(
|
|
164 |
if use_int8_w8a16:
|
165 |
accumulator = (accumulator * b_scale).to(compute_type)
|
166 |
elif use_fp8_w8a8:
|
167 |
-
|
|
|
|
|
|
|
168 |
else:
|
169 |
accumulator = accumulator.to(compute_type)
|
170 |
# -----------------------------------------------------------
|
@@ -175,6 +422,141 @@ def fused_moe_kernel(
|
|
175 |
tl.store(c_ptrs, accumulator, mask=c_mask)
|
176 |
|
177 |
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
178 |
def moe_align_block_size(
|
179 |
topk_ids: torch.Tensor, block_size: int, num_experts: int
|
180 |
) -> Tuple[torch.Tensor, torch.Tensor, torch.Tensor]:
|
@@ -225,9 +607,34 @@ def moe_align_block_size(
|
|
225 |
(max_num_m_blocks,), dtype=torch.int32, device=topk_ids.device
|
226 |
)
|
227 |
num_tokens_post_pad = torch.empty((1), dtype=torch.int32, device=topk_ids.device)
|
228 |
-
|
229 |
-
|
230 |
-
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
231 |
return sorted_ids, expert_ids, num_tokens_post_pad
|
232 |
|
233 |
|
@@ -237,6 +644,7 @@ def invoke_fused_moe_kernel(
|
|
237 |
C: torch.Tensor,
|
238 |
A_scale: Optional[torch.Tensor],
|
239 |
B_scale: Optional[torch.Tensor],
|
|
|
240 |
topk_weights: torch.Tensor,
|
241 |
topk_ids: torch.Tensor,
|
242 |
sorted_token_ids: torch.Tensor,
|
@@ -248,64 +656,147 @@ def invoke_fused_moe_kernel(
|
|
248 |
compute_type: tl.dtype,
|
249 |
use_fp8_w8a8: bool,
|
250 |
use_int8_w8a16: bool,
|
|
|
|
|
251 |
) -> None:
|
252 |
assert topk_weights.stride(1) == 1
|
253 |
assert sorted_token_ids.stride(0) == 1
|
254 |
|
255 |
if use_fp8_w8a8:
|
256 |
-
A, A_scale = scaled_fp8_quant(A, A_scale)
|
257 |
assert B_scale is not None
|
258 |
-
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
259 |
assert B_scale is not None
|
|
|
260 |
else:
|
261 |
assert A_scale is None
|
262 |
assert B_scale is None
|
263 |
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
264 |
grid = lambda META: (
|
265 |
-
triton.cdiv(
|
266 |
* triton.cdiv(B.shape[1], META["BLOCK_SIZE_N"]),
|
267 |
)
|
268 |
|
269 |
-
|
270 |
-
|
271 |
-
|
272 |
-
|
273 |
-
|
274 |
-
B_scale
|
275 |
-
|
276 |
-
|
277 |
-
|
278 |
-
|
279 |
-
|
280 |
-
|
281 |
-
|
282 |
-
|
283 |
-
|
284 |
-
|
285 |
-
|
286 |
-
|
287 |
-
|
288 |
-
|
289 |
-
|
290 |
-
|
291 |
-
|
292 |
-
|
293 |
-
|
294 |
-
|
295 |
-
|
296 |
-
|
297 |
-
|
298 |
-
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
299 |
|
300 |
|
301 |
-
|
|
|
|
|
|
|
302 |
device_name = current_platform.get_device_name().replace(" ", "_")
|
303 |
dtype_selector = "" if not dtype else f",dtype={dtype}"
|
304 |
-
|
|
|
|
|
|
|
305 |
|
306 |
|
|
|
307 |
@functools.lru_cache
|
308 |
-
def get_moe_configs(
|
|
|
|
|
|
|
|
|
|
|
|
|
309 |
"""
|
310 |
Return optimized configurations for the fused MoE kernel.
|
311 |
|
@@ -317,18 +808,27 @@ def get_moe_configs(E: int, N: int, dtype: Optional[str]) -> Optional[Dict[int,
|
|
317 |
|
318 |
# First look up if an optimized configuration is available in the configs
|
319 |
# directory
|
320 |
-
|
|
|
321 |
|
322 |
config_file_path = os.path.join(
|
323 |
os.path.dirname(os.path.realpath(__file__)), "configs", json_file_name
|
324 |
)
|
325 |
if os.path.exists(config_file_path):
|
326 |
with open(config_file_path) as f:
|
|
|
327 |
# If a configuration has been found, return it
|
328 |
return {int(key): val for key, val in json.load(f).items()}
|
329 |
|
330 |
# If no optimized configuration is available, we will use the default
|
331 |
# configuration
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
332 |
return None
|
333 |
|
334 |
|
@@ -340,21 +840,34 @@ def get_default_config(
|
|
340 |
topk: int,
|
341 |
dtype: Optional[str],
|
342 |
is_marlin: bool,
|
|
|
343 |
) -> Dict[str, int]:
|
344 |
-
|
345 |
-
|
346 |
-
|
347 |
-
"BLOCK_SIZE_K": 32,
|
348 |
-
"GROUP_SIZE_M": 8,
|
349 |
-
}
|
350 |
-
# A heuristic: fused marlin works faster with this config for small M
|
351 |
-
if M <= E or (is_marlin and M <= 32):
|
352 |
config = {
|
353 |
-
"BLOCK_SIZE_M":
|
354 |
-
"BLOCK_SIZE_N":
|
355 |
-
"BLOCK_SIZE_K":
|
356 |
-
"GROUP_SIZE_M":
|
|
|
|
|
357 |
}
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
358 |
return config
|
359 |
|
360 |
|
@@ -364,15 +877,21 @@ def try_get_optimal_moe_config(
|
|
364 |
top_k: int,
|
365 |
dtype: Optional[str],
|
366 |
M: int,
|
367 |
-
override_config: Optional[Dict[str, Any]] = None,
|
368 |
is_marlin: bool = False,
|
|
|
369 |
):
|
|
|
|
|
|
|
|
|
370 |
if override_config:
|
371 |
config = override_config
|
372 |
else:
|
373 |
# First try to load optimal config from the file
|
374 |
E, _, N = w2_shape
|
375 |
-
|
|
|
|
|
376 |
|
377 |
if configs:
|
378 |
# If an optimal configuration map has been found, look up the
|
@@ -380,7 +899,9 @@ def try_get_optimal_moe_config(
|
|
380 |
config = configs[min(configs.keys(), key=lambda x: abs(x - M))]
|
381 |
else:
|
382 |
# Else use the default config
|
383 |
-
config = get_default_config(
|
|
|
|
|
384 |
return config
|
385 |
|
386 |
|
@@ -416,7 +937,8 @@ def fused_topk(
|
|
416 |
return topk_weights, topk_ids
|
417 |
|
418 |
|
419 |
-
# This is used by the Deepseek-V2 model
|
|
|
420 |
def grouped_topk(
|
421 |
hidden_states: torch.Tensor,
|
422 |
gating_output: torch.Tensor,
|
@@ -424,11 +946,25 @@ def grouped_topk(
|
|
424 |
renormalize: bool,
|
425 |
num_expert_group: int = 0,
|
426 |
topk_group: int = 0,
|
|
|
|
|
427 |
):
|
428 |
|
429 |
assert hidden_states.shape[0] == gating_output.shape[0], "Number of tokens mismatch"
|
430 |
|
431 |
-
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
432 |
num_token = scores.shape[0]
|
433 |
group_scores = (
|
434 |
scores.view(num_token, num_expert_group, -1).max(dim=-1).values
|
@@ -444,7 +980,13 @@ def grouped_topk(
|
|
444 |
.reshape(num_token, -1)
|
445 |
) # [n, e]
|
446 |
tmp_scores = scores.masked_fill(~score_mask.bool(), 0.0) # [n, e]
|
447 |
-
|
|
|
|
|
|
|
|
|
|
|
|
|
448 |
|
449 |
if renormalize:
|
450 |
topk_weights = topk_weights / topk_weights.sum(dim=-1, keepdim=True)
|
@@ -454,6 +996,7 @@ def grouped_topk(
|
|
454 |
|
455 |
def get_config_dtype_str(
|
456 |
dtype: torch.dtype,
|
|
|
457 |
use_int8_w8a16: Optional[bool] = False,
|
458 |
use_fp8_w8a8: Optional[bool] = False,
|
459 |
):
|
@@ -461,6 +1004,8 @@ def get_config_dtype_str(
|
|
461 |
return "fp8_w8a8"
|
462 |
elif use_int8_w8a16:
|
463 |
return "int8_w8a16"
|
|
|
|
|
464 |
elif dtype == torch.float:
|
465 |
# avoiding cases where kernel fails when float32 MoE
|
466 |
# use fp16/bfloat16 configs
|
@@ -468,6 +1013,80 @@ def get_config_dtype_str(
|
|
468 |
return None
|
469 |
|
470 |
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
471 |
def fused_experts(
|
472 |
hidden_states: torch.Tensor,
|
473 |
w1: torch.Tensor,
|
@@ -475,16 +1094,80 @@ def fused_experts(
|
|
475 |
topk_weights: torch.Tensor,
|
476 |
topk_ids: torch.Tensor,
|
477 |
inplace: bool = False,
|
478 |
-
override_config: Optional[Dict[str, Any]] = None,
|
479 |
use_fp8_w8a8: bool = False,
|
480 |
use_int8_w8a16: bool = False,
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
481 |
w1_scale: Optional[torch.Tensor] = None,
|
482 |
w2_scale: Optional[torch.Tensor] = None,
|
|
|
|
|
483 |
a1_scale: Optional[torch.Tensor] = None,
|
484 |
a2_scale: Optional[torch.Tensor] = None,
|
|
|
485 |
):
|
486 |
# Check constraints.
|
487 |
-
|
|
|
|
|
|
|
|
|
488 |
assert topk_weights.shape == topk_ids.shape, "topk shape mismatch"
|
489 |
assert hidden_states.is_contiguous(), "Hidden_states must be contiguous"
|
490 |
assert w1.is_contiguous(), "Expert weights1 must be contiguous"
|
@@ -500,6 +1183,7 @@ def fused_experts(
|
|
500 |
config_dtype = get_config_dtype_str(
|
501 |
use_fp8_w8a8=use_fp8_w8a8,
|
502 |
use_int8_w8a16=use_int8_w8a16,
|
|
|
503 |
dtype=hidden_states.dtype,
|
504 |
)
|
505 |
|
@@ -509,7 +1193,7 @@ def fused_experts(
|
|
509 |
w2.shape,
|
510 |
topk_ids.shape[1],
|
511 |
config_dtype,
|
512 |
-
|
513 |
)
|
514 |
|
515 |
config = get_config_func(M)
|
@@ -530,7 +1214,14 @@ def fused_experts(
|
|
530 |
dtype=hidden_states.dtype,
|
531 |
)
|
532 |
|
533 |
-
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
534 |
|
535 |
if inplace:
|
536 |
out_hidden_states = hidden_states
|
@@ -571,6 +1262,7 @@ def fused_experts(
|
|
571 |
intermediate_cache1,
|
572 |
a1_scale,
|
573 |
w1_scale,
|
|
|
574 |
curr_topk_weights,
|
575 |
curr_topk_ids,
|
576 |
sorted_token_ids,
|
@@ -582,6 +1274,8 @@ def fused_experts(
|
|
582 |
compute_type=compute_type,
|
583 |
use_fp8_w8a8=use_fp8_w8a8,
|
584 |
use_int8_w8a16=use_int8_w8a16,
|
|
|
|
|
585 |
)
|
586 |
|
587 |
ops.silu_and_mul(intermediate_cache2, intermediate_cache1.view(-1, N))
|
@@ -592,6 +1286,7 @@ def fused_experts(
|
|
592 |
intermediate_cache3,
|
593 |
a2_scale,
|
594 |
w2_scale,
|
|
|
595 |
curr_topk_weights,
|
596 |
curr_topk_ids,
|
597 |
sorted_token_ids,
|
@@ -603,6 +1298,8 @@ def fused_experts(
|
|
603 |
compute_type=compute_type,
|
604 |
use_fp8_w8a8=use_fp8_w8a8,
|
605 |
use_int8_w8a16=use_int8_w8a16,
|
|
|
|
|
606 |
)
|
607 |
|
608 |
ops.moe_sum(
|
@@ -620,17 +1317,20 @@ def fused_moe(
|
|
620 |
topk: int,
|
621 |
renormalize: bool,
|
622 |
inplace: bool = False,
|
623 |
-
override_config: Optional[Dict[str, Any]] = None,
|
624 |
use_grouped_topk: bool = False,
|
625 |
num_expert_group: Optional[int] = None,
|
626 |
topk_group: Optional[int] = None,
|
627 |
custom_routing_function: Optional[Callable] = None,
|
628 |
use_fp8_w8a8: bool = False,
|
629 |
use_int8_w8a16: bool = False,
|
|
|
630 |
w1_scale: Optional[torch.Tensor] = None,
|
631 |
w2_scale: Optional[torch.Tensor] = None,
|
|
|
|
|
632 |
a1_scale: Optional[torch.Tensor] = None,
|
633 |
a2_scale: Optional[torch.Tensor] = None,
|
|
|
634 |
) -> torch.Tensor:
|
635 |
"""
|
636 |
This function computes a Mixture of Experts (MoE) layer using two sets of
|
@@ -646,20 +1346,28 @@ def fused_moe(
|
|
646 |
- renormalize (bool): If True, renormalize the top-k weights to sum to 1.
|
647 |
- inplace (bool): If True, perform the operation in-place.
|
648 |
Defaults to False.
|
649 |
-
- override_config (Optional[Dict[str, Any]]): Optional override
|
650 |
-
for the kernel configuration.
|
651 |
- num_expert_group: Optional[int]: additional parameter for grouped_topk
|
652 |
- topk_group: Optional[int]: additional parameter for grouped_topk
|
653 |
- use_grouped_topk: If True, use grouped_topk instead of fused_topk
|
654 |
note: Deepseekv2 model uses grouped_topk
|
655 |
- use_fp8_w8a8 (bool): If True, use fp8 arithmetic to compute the inner
|
656 |
products for w1 and w2. Defaults to False.
|
657 |
-
- use_int8_w8a16 (bool): If True, use
|
658 |
-
products for w1 and w2.
|
|
|
|
|
|
|
|
|
659 |
- w1_scale (Optional[torch.Tensor]): Optional scale to be used for
|
660 |
w1.
|
661 |
- w2_scale (Optional[torch.Tensor]): Optional scale to be used for
|
662 |
w2.
|
|
|
|
|
|
|
|
|
|
|
|
|
663 |
|
664 |
Returns:
|
665 |
- torch.Tensor: The output tensor after applying the MoE layer.
|
@@ -693,11 +1401,14 @@ def fused_moe(
|
|
693 |
topk_weights,
|
694 |
topk_ids,
|
695 |
inplace=inplace,
|
696 |
-
override_config=override_config,
|
697 |
use_fp8_w8a8=use_fp8_w8a8,
|
698 |
use_int8_w8a16=use_int8_w8a16,
|
|
|
699 |
w1_scale=w1_scale,
|
700 |
w2_scale=w2_scale,
|
|
|
|
|
701 |
a1_scale=a1_scale,
|
702 |
a2_scale=a2_scale,
|
|
|
703 |
)
|
|
|
1 |
+
# SPDX-License-Identifier: Apache-2.0
|
2 |
"""Fused MoE kernel."""
|
3 |
|
4 |
import functools
|
5 |
import json
|
6 |
+
import logging
|
7 |
import os
|
8 |
+
from typing import Any, Callable, Dict, List, Optional, Tuple
|
9 |
|
10 |
import torch
|
11 |
import triton
|
12 |
import triton.language as tl
|
13 |
|
14 |
+
|
15 |
from ._ops import ops
|
16 |
+
from .fp8 import per_token_group_quant_fp8, scaled_fp8_quant
|
17 |
from .platforms import current_platform
|
18 |
|
19 |
+
logger = logging.getLogger(__name__)
|
20 |
+
|
21 |
+
|
22 |
VLLM_FUSED_MOE_CHUNK_SIZE = int(os.getenv("VLLM_FUSED_MOE_CHUNK_SIZE", "32768"))
|
23 |
|
24 |
|
25 |
+
@triton.jit
|
26 |
+
def fused_moe_kernel_gptq_awq(
|
27 |
+
# Pointers to matrices
|
28 |
+
a_ptr,
|
29 |
+
b_ptr,
|
30 |
+
c_ptr,
|
31 |
+
b_scale_ptr,
|
32 |
+
b_zp_ptr,
|
33 |
+
topk_weights_ptr,
|
34 |
+
sorted_token_ids_ptr,
|
35 |
+
expert_ids_ptr,
|
36 |
+
num_tokens_post_padded_ptr,
|
37 |
+
# Matrix dimensions
|
38 |
+
N: tl.constexpr,
|
39 |
+
K: tl.constexpr,
|
40 |
+
EM,
|
41 |
+
num_valid_tokens,
|
42 |
+
# The stride variables represent how much to increase the ptr by when
|
43 |
+
# moving by 1 element in a particular dimension. E.g. `stride_am` is
|
44 |
+
# how much to increase `a_ptr` by to get the element one row down
|
45 |
+
# (A has M rows).
|
46 |
+
stride_am,
|
47 |
+
stride_ak,
|
48 |
+
stride_be,
|
49 |
+
stride_bk,
|
50 |
+
stride_bn,
|
51 |
+
stride_cm,
|
52 |
+
stride_cn,
|
53 |
+
stride_bse,
|
54 |
+
stride_bsk,
|
55 |
+
stride_bsn,
|
56 |
+
stride_bze,
|
57 |
+
stride_bzk,
|
58 |
+
stride_bzn,
|
59 |
+
block_k_diviable: tl.constexpr,
|
60 |
+
group_size: tl.constexpr,
|
61 |
+
# Meta-parameters
|
62 |
+
BLOCK_SIZE_M: tl.constexpr,
|
63 |
+
BLOCK_SIZE_N: tl.constexpr,
|
64 |
+
BLOCK_SIZE_K: tl.constexpr,
|
65 |
+
GROUP_SIZE_M: tl.constexpr,
|
66 |
+
MUL_ROUTED_WEIGHT: tl.constexpr,
|
67 |
+
top_k: tl.constexpr,
|
68 |
+
compute_type: tl.constexpr,
|
69 |
+
has_zp: tl.constexpr,
|
70 |
+
use_int4_w4a16: tl.constexpr,
|
71 |
+
use_int8_w8a16: tl.constexpr,
|
72 |
+
):
|
73 |
+
"""
|
74 |
+
Implements the fused computation for a Mixture of Experts (MOE) using
|
75 |
+
token and expert matrices.
|
76 |
+
|
77 |
+
Key Parameters:
|
78 |
+
- A: The input tensor representing tokens with shape (*, K), where '*' can
|
79 |
+
be any shape representing batches and K is the feature dimension of
|
80 |
+
each token.
|
81 |
+
- B: The stacked MOE weight tensor with shape (E, N, K), where E is
|
82 |
+
the number of experts, K is the input feature dimension, and N is
|
83 |
+
the output feature dimension.
|
84 |
+
- C: The output cache tensor with shape (M, topk, N), where M is the
|
85 |
+
total number of tokens post padding, topk is the number of times
|
86 |
+
each token is repeated, and N is the output feature dimension.
|
87 |
+
- sorted_token_ids: A tensor containing the sorted indices of tokens,
|
88 |
+
repeated topk times and arranged by the expert index they are
|
89 |
+
assigned to.
|
90 |
+
- expert_ids: A tensor containing the indices of the expert for each
|
91 |
+
block. It determines which expert matrix from B should be used for
|
92 |
+
each block in A.
|
93 |
+
This kernel performs the multiplication of a token by its corresponding
|
94 |
+
expert matrix as determined by `expert_ids`. The sorting of
|
95 |
+
`sorted_token_ids` by expert index and padding ensures divisibility by
|
96 |
+
BLOCK_SIZE_M, which is necessary to maintain consistency in block matrix
|
97 |
+
multiplication across different blocks processed by the same expert.
|
98 |
+
"""
|
99 |
+
# -----------------------------------------------------------
|
100 |
+
# Map program ids `pid` to the block of C it should compute.
|
101 |
+
# This is done in a grouped ordering to promote L2 data reuse.
|
102 |
+
pid = tl.program_id(axis=0)
|
103 |
+
num_pid_m = tl.cdiv(EM, BLOCK_SIZE_M)
|
104 |
+
num_pid_n = tl.cdiv(N, BLOCK_SIZE_N)
|
105 |
+
num_pid_in_group = GROUP_SIZE_M * num_pid_n
|
106 |
+
group_id = pid // num_pid_in_group
|
107 |
+
first_pid_m = group_id * GROUP_SIZE_M
|
108 |
+
group_size_m = min(num_pid_m - first_pid_m, GROUP_SIZE_M)
|
109 |
+
pid_m = first_pid_m + ((pid % num_pid_in_group) % group_size_m)
|
110 |
+
pid_n = (pid % num_pid_in_group) // group_size_m
|
111 |
+
|
112 |
+
# ----------------------------------------------------------
|
113 |
+
# Create pointers for the first blocks of A and B.
|
114 |
+
# We will advance this pointer as we move in the K direction
|
115 |
+
# and accumulate
|
116 |
+
# `a_ptrs` is a block of [BLOCK_SIZE_M, BLOCK_SIZE_K] pointers
|
117 |
+
# `b_ptrs` is a block of [BLOCK_SIZE_K, BLOCK_SIZE_N] pointers
|
118 |
+
num_tokens_post_padded = tl.load(num_tokens_post_padded_ptr)
|
119 |
+
if pid_m * BLOCK_SIZE_M >= num_tokens_post_padded:
|
120 |
+
return
|
121 |
+
offs_token_id = pid_m * BLOCK_SIZE_M + tl.arange(0, BLOCK_SIZE_M).to(tl.int64)
|
122 |
+
offs_token = tl.load(sorted_token_ids_ptr + offs_token_id)
|
123 |
+
token_mask = offs_token < num_valid_tokens
|
124 |
+
|
125 |
+
offs_bn = (pid_n * BLOCK_SIZE_N + tl.arange(0, BLOCK_SIZE_N).to(tl.int64)) % N
|
126 |
+
offs_k = tl.arange(0, BLOCK_SIZE_K)
|
127 |
+
a_ptrs = a_ptr + (
|
128 |
+
offs_token[:, None] // top_k * stride_am + offs_k[None, :] * stride_ak
|
129 |
+
)
|
130 |
+
|
131 |
+
off_experts = tl.load(expert_ids_ptr + pid_m).to(tl.int64)
|
132 |
+
|
133 |
+
if use_int4_w4a16:
|
134 |
+
b_ptrs = (
|
135 |
+
b_ptr
|
136 |
+
+ off_experts * stride_be
|
137 |
+
+ (offs_k[:, None] // 2) * stride_bk
|
138 |
+
+ offs_bn[None, :] * stride_bn
|
139 |
+
)
|
140 |
+
b_shifter = (offs_k[:, None] % 2) * 4
|
141 |
+
elif use_int8_w8a16:
|
142 |
+
b_ptrs = (
|
143 |
+
b_ptr
|
144 |
+
+ off_experts * stride_be
|
145 |
+
+ offs_k[:, None] * stride_bk
|
146 |
+
+ offs_bn[None, :] * stride_bn
|
147 |
+
)
|
148 |
+
|
149 |
+
if not has_zp and use_int4_w4a16:
|
150 |
+
b_zp_num = 8
|
151 |
+
if not has_zp and use_int8_w8a16:
|
152 |
+
b_zp_num = 128
|
153 |
+
elif has_zp and use_int4_w4a16:
|
154 |
+
b_zp_shifter = (offs_bn[None, :] % 2) * 4
|
155 |
+
|
156 |
+
# -----------------------------------------------------------
|
157 |
+
# Iterate to compute a block of the C matrix.
|
158 |
+
# We accumulate into a `[BLOCK_SIZE_M, BLOCK_SIZE_N]` block
|
159 |
+
# of fp32 values for higher accuracy.
|
160 |
+
# `accumulator` will be converted back to fp16 after the loop.
|
161 |
+
accumulator = tl.zeros((BLOCK_SIZE_M, BLOCK_SIZE_N), dtype=tl.float32)
|
162 |
+
for k in range(0, tl.cdiv(K, BLOCK_SIZE_K)):
|
163 |
+
# Load the next block of A and B, generate a mask by checking the
|
164 |
+
# K dimension.
|
165 |
+
|
166 |
+
if not block_k_diviable:
|
167 |
+
k_mask = offs_k[:, None] < K - k * BLOCK_SIZE_K
|
168 |
+
k_other = 0.0
|
169 |
+
else:
|
170 |
+
k_mask = None
|
171 |
+
k_other = None
|
172 |
+
|
173 |
+
a = tl.load(
|
174 |
+
a_ptrs,
|
175 |
+
mask=token_mask[:, None] & (offs_k[None, :] < K - k * BLOCK_SIZE_K),
|
176 |
+
other=0.0,
|
177 |
+
)
|
178 |
+
b = tl.load(b_ptrs)
|
179 |
+
if use_int4_w4a16:
|
180 |
+
b = (b >> b_shifter) & 0xF
|
181 |
+
|
182 |
+
b_scale_ptrs = (
|
183 |
+
b_scale_ptr
|
184 |
+
+ off_experts * stride_bse
|
185 |
+
+ offs_bn[None, :] * stride_bsn
|
186 |
+
+ ((offs_k[:, None] + BLOCK_SIZE_K * k) // group_size) * stride_bsk
|
187 |
+
)
|
188 |
+
b_scale = tl.load(b_scale_ptrs, mask=k_mask, other=k_other)
|
189 |
+
b_scale = b_scale.to(tl.float32)
|
190 |
+
|
191 |
+
if has_zp and use_int4_w4a16:
|
192 |
+
offs_k_true = (offs_k[:, None] + BLOCK_SIZE_K * k) // group_size
|
193 |
+
b_zp_ptrs = (
|
194 |
+
b_zp_ptr
|
195 |
+
+ off_experts * stride_bze
|
196 |
+
+ (offs_bn[None, :] // 2) * stride_bzn
|
197 |
+
+ offs_k_true * stride_bzk
|
198 |
+
)
|
199 |
+
b_zp = tl.load(b_zp_ptrs, mask=k_mask, other=k_other)
|
200 |
+
b_zp = (b_zp >> b_zp_shifter) & 0xF
|
201 |
+
b_zp = b_zp.to(tl.float32)
|
202 |
+
elif has_zp and use_int8_w8a16:
|
203 |
+
offs_k_true = (offs_k[:, None] + BLOCK_SIZE_K * k) // group_size
|
204 |
+
b_zp_ptrs = (
|
205 |
+
b_zp_ptr
|
206 |
+
+ off_experts * stride_bze
|
207 |
+
+ offs_bn[None, :] * stride_bzn
|
208 |
+
+ offs_k_true * stride_bzk
|
209 |
+
)
|
210 |
+
b_zp = tl.load(b_zp_ptrs, mask=k_mask, other=k_other)
|
211 |
+
b_zp = b_zp.to(tl.float32)
|
212 |
+
|
213 |
+
# We accumulate along the K dimension.
|
214 |
+
if has_zp:
|
215 |
+
b = ((b.to(tl.float32) - b_zp) * b_scale).to(compute_type)
|
216 |
+
else:
|
217 |
+
b = ((b.to(tl.float32) - b_zp_num) * b_scale).to(compute_type)
|
218 |
+
accumulator = tl.dot(a, b, acc=accumulator)
|
219 |
+
|
220 |
+
# Advance the ptrs to the next K block.
|
221 |
+
a_ptrs += BLOCK_SIZE_K * stride_ak
|
222 |
+
if use_int4_w4a16:
|
223 |
+
b_ptrs += (BLOCK_SIZE_K // 2) * stride_bk
|
224 |
+
else:
|
225 |
+
b_ptrs += BLOCK_SIZE_K * stride_bk
|
226 |
+
|
227 |
+
if MUL_ROUTED_WEIGHT:
|
228 |
+
moe_weight = tl.load(topk_weights_ptr + offs_token, mask=token_mask, other=0)
|
229 |
+
accumulator = accumulator * moe_weight[:, None]
|
230 |
+
|
231 |
+
accumulator = accumulator.to(compute_type)
|
232 |
+
# -----------------------------------------------------------
|
233 |
+
# Write back the block of the output
|
234 |
+
offs_cn = pid_n * BLOCK_SIZE_N + tl.arange(0, BLOCK_SIZE_N)
|
235 |
+
c_ptrs = c_ptr + stride_cm * offs_token[:, None] + stride_cn * offs_cn[None, :]
|
236 |
+
c_mask = token_mask[:, None] & (offs_cn[None, :] < N)
|
237 |
+
tl.store(c_ptrs, accumulator, mask=c_mask)
|
238 |
+
|
239 |
+
|
240 |
@triton.jit
|
241 |
def fused_moe_kernel(
|
242 |
# Pointers to matrices
|
|
|
265 |
stride_bn,
|
266 |
stride_cm,
|
267 |
stride_cn,
|
268 |
+
stride_asm,
|
269 |
+
stride_ask,
|
270 |
stride_bse,
|
271 |
+
stride_bsk,
|
272 |
stride_bsn,
|
273 |
+
# Block size for block-wise quantization
|
274 |
+
group_n: tl.constexpr,
|
275 |
+
group_k: tl.constexpr,
|
276 |
# Meta-parameters
|
277 |
BLOCK_SIZE_M: tl.constexpr,
|
278 |
BLOCK_SIZE_N: tl.constexpr,
|
|
|
332 |
num_tokens_post_padded = tl.load(num_tokens_post_padded_ptr)
|
333 |
if pid_m * BLOCK_SIZE_M >= num_tokens_post_padded:
|
334 |
return
|
335 |
+
offs_token_id = pid_m * BLOCK_SIZE_M + tl.arange(0, BLOCK_SIZE_M).to(tl.int64)
|
336 |
offs_token = tl.load(sorted_token_ids_ptr + offs_token_id)
|
337 |
token_mask = offs_token < num_valid_tokens
|
338 |
|
339 |
+
offs_bn = (pid_n * BLOCK_SIZE_N + tl.arange(0, BLOCK_SIZE_N).to(tl.int64)) % N
|
340 |
offs_k = tl.arange(0, BLOCK_SIZE_K)
|
341 |
a_ptrs = a_ptr + (
|
342 |
offs_token[:, None] // top_k * stride_am + offs_k[None, :] * stride_ak
|
343 |
)
|
344 |
|
345 |
+
off_experts = tl.load(expert_ids_ptr + pid_m).to(tl.int64)
|
346 |
b_ptrs = (
|
347 |
b_ptr
|
348 |
+ off_experts * stride_be
|
|
|
355 |
b_scale = tl.load(b_scale_ptrs)
|
356 |
|
357 |
if use_fp8_w8a8:
|
358 |
+
if group_k > 0 and group_n > 0:
|
359 |
+
a_scale_ptrs = a_scale_ptr + (offs_token // top_k) * stride_asm
|
360 |
+
offs_bsn = offs_bn // group_n
|
361 |
+
b_scale_ptrs = (
|
362 |
+
b_scale_ptr + off_experts * stride_bse + offs_bsn * stride_bsn
|
363 |
+
)
|
364 |
+
else:
|
365 |
+
a_scale = tl.load(a_scale_ptr)
|
366 |
+
b_scale = tl.load(b_scale_ptr + off_experts)
|
367 |
|
368 |
# -----------------------------------------------------------
|
369 |
# Iterate to compute a block of the C matrix.
|
|
|
385 |
if use_int8_w8a16:
|
386 |
accumulator = tl.dot(a, b.to(compute_type), acc=accumulator)
|
387 |
elif use_fp8_w8a8:
|
388 |
+
if group_k > 0 and group_n > 0:
|
389 |
+
k_start = k * BLOCK_SIZE_K
|
390 |
+
offs_ks = k_start // group_k
|
391 |
+
a_scale = tl.load(
|
392 |
+
a_scale_ptrs + offs_ks * stride_ask, mask=token_mask, other=0.0
|
393 |
+
)
|
394 |
+
b_scale = tl.load(b_scale_ptrs + offs_ks * stride_bsk)
|
395 |
+
|
396 |
+
accumulator += tl.dot(a, b) * a_scale[:, None] * b_scale[None, :]
|
397 |
+
else:
|
398 |
+
accumulator = tl.dot(a, b, acc=accumulator)
|
399 |
else:
|
400 |
accumulator += tl.dot(a, b)
|
401 |
# Advance the ptrs to the next K block.
|
|
|
408 |
if use_int8_w8a16:
|
409 |
accumulator = (accumulator * b_scale).to(compute_type)
|
410 |
elif use_fp8_w8a8:
|
411 |
+
if group_k > 0 and group_n > 0:
|
412 |
+
accumulator = accumulator.to(compute_type)
|
413 |
+
else:
|
414 |
+
accumulator = (accumulator * a_scale * b_scale).to(compute_type)
|
415 |
else:
|
416 |
accumulator = accumulator.to(compute_type)
|
417 |
# -----------------------------------------------------------
|
|
|
422 |
tl.store(c_ptrs, accumulator, mask=c_mask)
|
423 |
|
424 |
|
425 |
+
def ceil_div(a, b):
|
426 |
+
return (a + b - 1) // b
|
427 |
+
|
428 |
+
|
429 |
+
@triton.jit
|
430 |
+
def moe_align_block_size_stage1(
|
431 |
+
topk_ids_ptr,
|
432 |
+
tokens_cnts_ptr,
|
433 |
+
num_experts: tl.constexpr,
|
434 |
+
numel: tl.constexpr,
|
435 |
+
tokens_per_thread: tl.constexpr,
|
436 |
+
):
|
437 |
+
pid = tl.program_id(0)
|
438 |
+
|
439 |
+
start_idx = pid * tokens_per_thread
|
440 |
+
|
441 |
+
off_c = (pid + 1) * num_experts
|
442 |
+
|
443 |
+
for i in range(tokens_per_thread):
|
444 |
+
if start_idx + i < numel:
|
445 |
+
idx = tl.load(topk_ids_ptr + start_idx + i)
|
446 |
+
token_cnt = tl.load(tokens_cnts_ptr + off_c + idx)
|
447 |
+
tl.store(tokens_cnts_ptr + off_c + idx, token_cnt + 1)
|
448 |
+
|
449 |
+
|
450 |
+
@triton.jit
|
451 |
+
def moe_align_block_size_stage2(
|
452 |
+
tokens_cnts_ptr,
|
453 |
+
num_experts: tl.constexpr,
|
454 |
+
):
|
455 |
+
pid = tl.program_id(0)
|
456 |
+
|
457 |
+
last_cnt = 0
|
458 |
+
for i in range(1, num_experts + 1):
|
459 |
+
token_cnt = tl.load(tokens_cnts_ptr + i * num_experts + pid)
|
460 |
+
last_cnt = last_cnt + token_cnt
|
461 |
+
tl.store(tokens_cnts_ptr + i * num_experts + pid, last_cnt)
|
462 |
+
|
463 |
+
|
464 |
+
@triton.jit
|
465 |
+
def moe_align_block_size_stage3(
|
466 |
+
total_tokens_post_pad_ptr,
|
467 |
+
tokens_cnts_ptr,
|
468 |
+
cumsum_ptr,
|
469 |
+
num_experts: tl.constexpr,
|
470 |
+
block_size: tl.constexpr,
|
471 |
+
):
|
472 |
+
last_cumsum = 0
|
473 |
+
off_cnt = num_experts * num_experts
|
474 |
+
for i in range(1, num_experts + 1):
|
475 |
+
token_cnt = tl.load(tokens_cnts_ptr + off_cnt + i - 1)
|
476 |
+
last_cumsum = last_cumsum + tl.cdiv(token_cnt, block_size) * block_size
|
477 |
+
tl.store(cumsum_ptr + i, last_cumsum)
|
478 |
+
tl.store(total_tokens_post_pad_ptr, last_cumsum)
|
479 |
+
|
480 |
+
|
481 |
+
@triton.jit
|
482 |
+
def moe_align_block_size_stage4(
|
483 |
+
topk_ids_ptr,
|
484 |
+
sorted_token_ids_ptr,
|
485 |
+
expert_ids_ptr,
|
486 |
+
tokens_cnts_ptr,
|
487 |
+
cumsum_ptr,
|
488 |
+
num_experts: tl.constexpr,
|
489 |
+
block_size: tl.constexpr,
|
490 |
+
numel: tl.constexpr,
|
491 |
+
tokens_per_thread: tl.constexpr,
|
492 |
+
):
|
493 |
+
pid = tl.program_id(0)
|
494 |
+
start_idx = tl.load(cumsum_ptr + pid)
|
495 |
+
end_idx = tl.load(cumsum_ptr + pid + 1)
|
496 |
+
|
497 |
+
for i in range(start_idx, end_idx, block_size):
|
498 |
+
tl.store(expert_ids_ptr + i // block_size, pid)
|
499 |
+
|
500 |
+
start_idx = pid * tokens_per_thread
|
501 |
+
off_t = pid * num_experts
|
502 |
+
|
503 |
+
for i in range(start_idx, tl.minimum(start_idx + tokens_per_thread, numel)):
|
504 |
+
expert_id = tl.load(topk_ids_ptr + i)
|
505 |
+
token_cnt = tl.load(tokens_cnts_ptr + off_t + expert_id)
|
506 |
+
rank_post_pad = token_cnt + tl.load(cumsum_ptr + expert_id)
|
507 |
+
tl.store(sorted_token_ids_ptr + rank_post_pad, i)
|
508 |
+
tl.store(tokens_cnts_ptr + off_t + expert_id, token_cnt + 1)
|
509 |
+
|
510 |
+
|
511 |
+
# Triton implementation based on:
|
512 |
+
# https://github.com/sgl-project/sglang/commit/ba5112ff691d791a9e38c6c71f59324a5fcb49d0
|
513 |
+
def moe_align_block_size_triton(
|
514 |
+
topk_ids: torch.Tensor,
|
515 |
+
num_experts: int,
|
516 |
+
block_size: int,
|
517 |
+
sorted_token_ids: torch.Tensor,
|
518 |
+
expert_ids: torch.Tensor,
|
519 |
+
num_tokens_post_pad: torch.Tensor,
|
520 |
+
) -> None:
|
521 |
+
numel = topk_ids.numel()
|
522 |
+
grid = (num_experts,)
|
523 |
+
tokens_cnts = torch.zeros(
|
524 |
+
(num_experts + 1, num_experts), dtype=torch.int32, device=topk_ids.device
|
525 |
+
)
|
526 |
+
cumsum = torch.zeros((num_experts + 1,), dtype=torch.int32, device=topk_ids.device)
|
527 |
+
tokens_per_thread = ceil_div(numel, num_experts)
|
528 |
+
|
529 |
+
moe_align_block_size_stage1[grid](
|
530 |
+
topk_ids,
|
531 |
+
tokens_cnts,
|
532 |
+
num_experts,
|
533 |
+
numel,
|
534 |
+
tokens_per_thread,
|
535 |
+
)
|
536 |
+
moe_align_block_size_stage2[grid](
|
537 |
+
tokens_cnts,
|
538 |
+
num_experts,
|
539 |
+
)
|
540 |
+
moe_align_block_size_stage3[(1,)](
|
541 |
+
num_tokens_post_pad,
|
542 |
+
tokens_cnts,
|
543 |
+
cumsum,
|
544 |
+
num_experts,
|
545 |
+
block_size,
|
546 |
+
)
|
547 |
+
moe_align_block_size_stage4[grid](
|
548 |
+
topk_ids,
|
549 |
+
sorted_token_ids,
|
550 |
+
expert_ids,
|
551 |
+
tokens_cnts,
|
552 |
+
cumsum,
|
553 |
+
num_experts,
|
554 |
+
block_size,
|
555 |
+
numel,
|
556 |
+
tokens_per_thread,
|
557 |
+
)
|
558 |
+
|
559 |
+
|
560 |
def moe_align_block_size(
|
561 |
topk_ids: torch.Tensor, block_size: int, num_experts: int
|
562 |
) -> Tuple[torch.Tensor, torch.Tensor, torch.Tensor]:
|
|
|
607 |
(max_num_m_blocks,), dtype=torch.int32, device=topk_ids.device
|
608 |
)
|
609 |
num_tokens_post_pad = torch.empty((1), dtype=torch.int32, device=topk_ids.device)
|
610 |
+
if num_experts >= 224:
|
611 |
+
if VLLM_ENABLE_MOE_ALIGN_BLOCK_SIZE_TRITON:
|
612 |
+
moe_align_block_size_triton(
|
613 |
+
topk_ids,
|
614 |
+
num_experts,
|
615 |
+
block_size,
|
616 |
+
sorted_ids,
|
617 |
+
expert_ids,
|
618 |
+
num_tokens_post_pad,
|
619 |
+
)
|
620 |
+
else:
|
621 |
+
ops.sgl_moe_align_block_size(
|
622 |
+
topk_ids,
|
623 |
+
num_experts,
|
624 |
+
block_size,
|
625 |
+
sorted_ids,
|
626 |
+
expert_ids,
|
627 |
+
num_tokens_post_pad,
|
628 |
+
)
|
629 |
+
else:
|
630 |
+
ops.moe_align_block_size(
|
631 |
+
topk_ids,
|
632 |
+
num_experts,
|
633 |
+
block_size,
|
634 |
+
sorted_ids,
|
635 |
+
expert_ids,
|
636 |
+
num_tokens_post_pad,
|
637 |
+
)
|
638 |
return sorted_ids, expert_ids, num_tokens_post_pad
|
639 |
|
640 |
|
|
|
644 |
C: torch.Tensor,
|
645 |
A_scale: Optional[torch.Tensor],
|
646 |
B_scale: Optional[torch.Tensor],
|
647 |
+
B_zp: Optional[torch.Tensor],
|
648 |
topk_weights: torch.Tensor,
|
649 |
topk_ids: torch.Tensor,
|
650 |
sorted_token_ids: torch.Tensor,
|
|
|
656 |
compute_type: tl.dtype,
|
657 |
use_fp8_w8a8: bool,
|
658 |
use_int8_w8a16: bool,
|
659 |
+
use_int4_w4a16: bool,
|
660 |
+
block_shape: Optional[List[int]] = None,
|
661 |
) -> None:
|
662 |
assert topk_weights.stride(1) == 1
|
663 |
assert sorted_token_ids.stride(0) == 1
|
664 |
|
665 |
if use_fp8_w8a8:
|
|
|
666 |
assert B_scale is not None
|
667 |
+
if block_shape is None:
|
668 |
+
A, A_scale = scaled_fp8_quant(A, A_scale)
|
669 |
+
else:
|
670 |
+
assert len(block_shape) == 2
|
671 |
+
block_n, block_k = block_shape[0], block_shape[1]
|
672 |
+
A, A_scale = per_token_group_quant_fp8(A, block_k)
|
673 |
+
assert triton.cdiv(A.shape[-1], block_k) == A_scale.shape[-1]
|
674 |
+
assert triton.cdiv(B.shape[-2], block_n) == B_scale.shape[-2]
|
675 |
+
assert triton.cdiv(B.shape[-1], block_k) == B_scale.shape[-1]
|
676 |
+
elif use_int8_w8a16 or use_int4_w4a16:
|
677 |
assert B_scale is not None
|
678 |
+
assert block_shape is None or block_shape[0] == 0
|
679 |
else:
|
680 |
assert A_scale is None
|
681 |
assert B_scale is None
|
682 |
|
683 |
+
EM = sorted_token_ids.shape[0]
|
684 |
+
if A.shape[0] < config["BLOCK_SIZE_M"]:
|
685 |
+
# optimize for small batch_size.
|
686 |
+
# We assume that top_ids of each token is unique, so
|
687 |
+
# so num_valid_experts <= batch_size <= BLOCK_SIZE_M,
|
688 |
+
# and we can skip some invalid blocks.
|
689 |
+
EM = min(sorted_token_ids.shape[0], A.shape[0] * top_k * config["BLOCK_SIZE_M"])
|
690 |
grid = lambda META: (
|
691 |
+
triton.cdiv(EM, META["BLOCK_SIZE_M"])
|
692 |
* triton.cdiv(B.shape[1], META["BLOCK_SIZE_N"]),
|
693 |
)
|
694 |
|
695 |
+
if (
|
696 |
+
(use_int8_w8a16 or use_int4_w4a16)
|
697 |
+
and block_shape is not None
|
698 |
+
and block_shape[1] > 0
|
699 |
+
):
|
700 |
+
assert B_scale is not None and B_scale.ndim == 3
|
701 |
+
assert B_zp is None or B_zp.ndim == 3
|
702 |
+
|
703 |
+
fused_moe_kernel_gptq_awq[grid](
|
704 |
+
A,
|
705 |
+
B,
|
706 |
+
C,
|
707 |
+
B_scale,
|
708 |
+
B_zp,
|
709 |
+
topk_weights,
|
710 |
+
sorted_token_ids,
|
711 |
+
expert_ids,
|
712 |
+
num_tokens_post_padded,
|
713 |
+
B.shape[1],
|
714 |
+
A.shape[1],
|
715 |
+
EM,
|
716 |
+
topk_ids.numel(),
|
717 |
+
A.stride(0),
|
718 |
+
A.stride(1),
|
719 |
+
B.stride(0),
|
720 |
+
B.stride(2),
|
721 |
+
B.stride(1),
|
722 |
+
C.stride(1),
|
723 |
+
C.stride(2),
|
724 |
+
B_scale.stride(0),
|
725 |
+
B_scale.stride(2),
|
726 |
+
B_scale.stride(1),
|
727 |
+
B_zp.stride(0) if B_zp is not None else 0,
|
728 |
+
B_zp.stride(2) if B_zp is not None else 0,
|
729 |
+
B_zp.stride(1) if B_zp is not None else 0,
|
730 |
+
block_k_diviable=A.shape[1] % config["BLOCK_SIZE_K"] == 0,
|
731 |
+
group_size=block_shape[1],
|
732 |
+
MUL_ROUTED_WEIGHT=mul_routed_weight,
|
733 |
+
top_k=top_k,
|
734 |
+
compute_type=compute_type,
|
735 |
+
has_zp=B_zp is not None,
|
736 |
+
use_int4_w4a16=use_int4_w4a16,
|
737 |
+
use_int8_w8a16=use_int8_w8a16,
|
738 |
+
**config,
|
739 |
+
)
|
740 |
+
|
741 |
+
else:
|
742 |
+
fused_moe_kernel[grid](
|
743 |
+
A,
|
744 |
+
B,
|
745 |
+
C,
|
746 |
+
A_scale,
|
747 |
+
B_scale,
|
748 |
+
topk_weights,
|
749 |
+
sorted_token_ids,
|
750 |
+
expert_ids,
|
751 |
+
num_tokens_post_padded,
|
752 |
+
B.shape[1],
|
753 |
+
A.shape[1],
|
754 |
+
EM,
|
755 |
+
topk_ids.numel(),
|
756 |
+
A.stride(0),
|
757 |
+
A.stride(1),
|
758 |
+
B.stride(0),
|
759 |
+
B.stride(2),
|
760 |
+
B.stride(1),
|
761 |
+
C.stride(1),
|
762 |
+
C.stride(2),
|
763 |
+
A_scale.stride(0) if A_scale is not None and A_scale.ndim == 2 else 0,
|
764 |
+
A_scale.stride(1) if A_scale is not None and A_scale.ndim == 2 else 0,
|
765 |
+
B_scale.stride(0) if B_scale is not None and B_scale.ndim >= 2 else 0,
|
766 |
+
B_scale.stride(2) if B_scale is not None and B_scale.ndim == 3 else 0,
|
767 |
+
B_scale.stride(1) if B_scale is not None and B_scale.ndim >= 2 else 0,
|
768 |
+
0 if block_shape is None else block_shape[0],
|
769 |
+
0 if block_shape is None else block_shape[1],
|
770 |
+
MUL_ROUTED_WEIGHT=mul_routed_weight,
|
771 |
+
top_k=top_k,
|
772 |
+
compute_type=compute_type,
|
773 |
+
use_fp8_w8a8=use_fp8_w8a8,
|
774 |
+
use_int8_w8a16=use_int8_w8a16,
|
775 |
+
**config,
|
776 |
+
)
|
777 |
|
778 |
|
779 |
+
# Adapted from: https://github.com/sgl-project/sglang/pull/2628
|
780 |
+
def get_config_file_name(
|
781 |
+
E: int, N: int, dtype: Optional[str], block_shape: Optional[List[int]] = None
|
782 |
+
) -> str:
|
783 |
device_name = current_platform.get_device_name().replace(" ", "_")
|
784 |
dtype_selector = "" if not dtype else f",dtype={dtype}"
|
785 |
+
block_shape_selector = (
|
786 |
+
"" if not block_shape or not all(block_shape) else f",block_shape={block_shape}"
|
787 |
+
)
|
788 |
+
return f"E={E},N={N},device_name={device_name}{dtype_selector}{block_shape_selector}.json" # noqa: E501
|
789 |
|
790 |
|
791 |
+
# Adapted from: https://github.com/sgl-project/sglang/pull/2628
|
792 |
@functools.lru_cache
|
793 |
+
def get_moe_configs(
|
794 |
+
E: int,
|
795 |
+
N: int,
|
796 |
+
dtype: Optional[str],
|
797 |
+
block_n: Optional[int] = None,
|
798 |
+
block_k: Optional[int] = None,
|
799 |
+
) -> Optional[Dict[int, Any]]:
|
800 |
"""
|
801 |
Return optimized configurations for the fused MoE kernel.
|
802 |
|
|
|
808 |
|
809 |
# First look up if an optimized configuration is available in the configs
|
810 |
# directory
|
811 |
+
block_shape = [block_n, block_k] if block_n and block_k else None
|
812 |
+
json_file_name = get_config_file_name(E, N, dtype, block_shape)
|
813 |
|
814 |
config_file_path = os.path.join(
|
815 |
os.path.dirname(os.path.realpath(__file__)), "configs", json_file_name
|
816 |
)
|
817 |
if os.path.exists(config_file_path):
|
818 |
with open(config_file_path) as f:
|
819 |
+
logger.info("Using configuration from %s for MoE layer.", config_file_path)
|
820 |
# If a configuration has been found, return it
|
821 |
return {int(key): val for key, val in json.load(f).items()}
|
822 |
|
823 |
# If no optimized configuration is available, we will use the default
|
824 |
# configuration
|
825 |
+
logger.warning(
|
826 |
+
(
|
827 |
+
"Using default MoE config. Performance might be sub-optimal! "
|
828 |
+
"Config file not found at %s"
|
829 |
+
),
|
830 |
+
config_file_path,
|
831 |
+
)
|
832 |
return None
|
833 |
|
834 |
|
|
|
840 |
topk: int,
|
841 |
dtype: Optional[str],
|
842 |
is_marlin: bool,
|
843 |
+
block_shape: Optional[List[int]] = None,
|
844 |
) -> Dict[str, int]:
|
845 |
+
if dtype == "fp8_w8a8" and block_shape is not None:
|
846 |
+
# Block-wise quant: BLOCK_SIZE_N must be divisible by block_shape[0]
|
847 |
+
# BLOCK_SIZE_K must be divisible by block_shape[1]
|
|
|
|
|
|
|
|
|
|
|
848 |
config = {
|
849 |
+
"BLOCK_SIZE_M": 64,
|
850 |
+
"BLOCK_SIZE_N": block_shape[0],
|
851 |
+
"BLOCK_SIZE_K": block_shape[1],
|
852 |
+
"GROUP_SIZE_M": 32,
|
853 |
+
"num_warps": 4,
|
854 |
+
"num_stages": 3,
|
855 |
}
|
856 |
+
else:
|
857 |
+
config = {
|
858 |
+
"BLOCK_SIZE_M": 64,
|
859 |
+
"BLOCK_SIZE_N": 64,
|
860 |
+
"BLOCK_SIZE_K": 32,
|
861 |
+
"GROUP_SIZE_M": 8,
|
862 |
+
}
|
863 |
+
# A heuristic: fused marlin works faster with this config for small M
|
864 |
+
if M <= E or (is_marlin and M <= 32):
|
865 |
+
config = {
|
866 |
+
"BLOCK_SIZE_M": 16,
|
867 |
+
"BLOCK_SIZE_N": 32,
|
868 |
+
"BLOCK_SIZE_K": 64,
|
869 |
+
"GROUP_SIZE_M": 1,
|
870 |
+
}
|
871 |
return config
|
872 |
|
873 |
|
|
|
877 |
top_k: int,
|
878 |
dtype: Optional[str],
|
879 |
M: int,
|
|
|
880 |
is_marlin: bool = False,
|
881 |
+
block_shape: Optional[List[int]] = None,
|
882 |
):
|
883 |
+
# from vllm.model_executor.layers.fused_moe import get_config
|
884 |
+
# TODO: removed when syncing to vLLM, do we need this?
|
885 |
+
# override_config = get_config()
|
886 |
+
override_config = None
|
887 |
if override_config:
|
888 |
config = override_config
|
889 |
else:
|
890 |
# First try to load optimal config from the file
|
891 |
E, _, N = w2_shape
|
892 |
+
block_n = block_shape[0] if block_shape else 0
|
893 |
+
block_k = block_shape[1] if block_shape else 0
|
894 |
+
configs = get_moe_configs(E, N, dtype, block_n, block_k)
|
895 |
|
896 |
if configs:
|
897 |
# If an optimal configuration map has been found, look up the
|
|
|
899 |
config = configs[min(configs.keys(), key=lambda x: abs(x - M))]
|
900 |
else:
|
901 |
# Else use the default config
|
902 |
+
config = get_default_config(
|
903 |
+
M, E, N, w1_shape[2], top_k, dtype, is_marlin, block_shape
|
904 |
+
)
|
905 |
return config
|
906 |
|
907 |
|
|
|
937 |
return topk_weights, topk_ids
|
938 |
|
939 |
|
940 |
+
# This is used by the Deepseek-V2 and Deepseek-V3 model
|
941 |
+
@torch.compile(dynamic=True, backend=current_platform.simple_compile_backend)
|
942 |
def grouped_topk(
|
943 |
hidden_states: torch.Tensor,
|
944 |
gating_output: torch.Tensor,
|
|
|
946 |
renormalize: bool,
|
947 |
num_expert_group: int = 0,
|
948 |
topk_group: int = 0,
|
949 |
+
scoring_func: str = "softmax",
|
950 |
+
e_score_correction_bias: Optional[torch.Tensor] = None,
|
951 |
):
|
952 |
|
953 |
assert hidden_states.shape[0] == gating_output.shape[0], "Number of tokens mismatch"
|
954 |
|
955 |
+
if scoring_func == "softmax":
|
956 |
+
scores = torch.softmax(gating_output, dim=-1)
|
957 |
+
elif scoring_func == "sigmoid":
|
958 |
+
scores = gating_output.sigmoid()
|
959 |
+
else:
|
960 |
+
raise ValueError(f"Unsupported scoring function: {scoring_func}")
|
961 |
+
|
962 |
+
if e_score_correction_bias is not None:
|
963 |
+
# Store original scores before applying correction bias. We use biased
|
964 |
+
# scores for expert selection but original scores for routing weights
|
965 |
+
original_scores = scores
|
966 |
+
scores = scores + e_score_correction_bias.unsqueeze(0)
|
967 |
+
|
968 |
num_token = scores.shape[0]
|
969 |
group_scores = (
|
970 |
scores.view(num_token, num_expert_group, -1).max(dim=-1).values
|
|
|
980 |
.reshape(num_token, -1)
|
981 |
) # [n, e]
|
982 |
tmp_scores = scores.masked_fill(~score_mask.bool(), 0.0) # [n, e]
|
983 |
+
|
984 |
+
if e_score_correction_bias is not None:
|
985 |
+
topk_ids = torch.topk(tmp_scores, k=topk, dim=-1, sorted=False)[1]
|
986 |
+
# Use original unbiased scores for the routing weights
|
987 |
+
topk_weights = original_scores.gather(1, topk_ids)
|
988 |
+
else:
|
989 |
+
topk_weights, topk_ids = torch.topk(tmp_scores, k=topk, dim=-1, sorted=False)
|
990 |
|
991 |
if renormalize:
|
992 |
topk_weights = topk_weights / topk_weights.sum(dim=-1, keepdim=True)
|
|
|
996 |
|
997 |
def get_config_dtype_str(
|
998 |
dtype: torch.dtype,
|
999 |
+
use_int4_w4a16: Optional[bool] = False,
|
1000 |
use_int8_w8a16: Optional[bool] = False,
|
1001 |
use_fp8_w8a8: Optional[bool] = False,
|
1002 |
):
|
|
|
1004 |
return "fp8_w8a8"
|
1005 |
elif use_int8_w8a16:
|
1006 |
return "int8_w8a16"
|
1007 |
+
elif use_int4_w4a16:
|
1008 |
+
return "int4_w8a16"
|
1009 |
elif dtype == torch.float:
|
1010 |
# avoiding cases where kernel fails when float32 MoE
|
1011 |
# use fp16/bfloat16 configs
|
|
|
1013 |
return None
|
1014 |
|
1015 |
|
1016 |
+
def inplace_fused_experts(
|
1017 |
+
hidden_states: torch.Tensor,
|
1018 |
+
w1: torch.Tensor,
|
1019 |
+
w2: torch.Tensor,
|
1020 |
+
topk_weights: torch.Tensor,
|
1021 |
+
topk_ids: torch.Tensor,
|
1022 |
+
use_fp8_w8a8: bool = False,
|
1023 |
+
use_int8_w8a16: bool = False,
|
1024 |
+
use_int4_w4a16: bool = False,
|
1025 |
+
w1_scale: Optional[torch.Tensor] = None,
|
1026 |
+
w2_scale: Optional[torch.Tensor] = None,
|
1027 |
+
w1_zp: Optional[torch.Tensor] = None,
|
1028 |
+
w2_zp: Optional[torch.Tensor] = None,
|
1029 |
+
a1_scale: Optional[torch.Tensor] = None,
|
1030 |
+
a2_scale: Optional[torch.Tensor] = None,
|
1031 |
+
block_shape: Optional[List[int]] = None,
|
1032 |
+
) -> None:
|
1033 |
+
fused_experts_impl(
|
1034 |
+
hidden_states,
|
1035 |
+
w1,
|
1036 |
+
w2,
|
1037 |
+
topk_weights,
|
1038 |
+
topk_ids,
|
1039 |
+
True,
|
1040 |
+
use_fp8_w8a8,
|
1041 |
+
use_int8_w8a16,
|
1042 |
+
use_int4_w4a16,
|
1043 |
+
w1_scale,
|
1044 |
+
w2_scale,
|
1045 |
+
w1_zp,
|
1046 |
+
w2_zp,
|
1047 |
+
a1_scale,
|
1048 |
+
a2_scale,
|
1049 |
+
block_shape,
|
1050 |
+
)
|
1051 |
+
|
1052 |
+
|
1053 |
+
def outplace_fused_experts(
|
1054 |
+
hidden_states: torch.Tensor,
|
1055 |
+
w1: torch.Tensor,
|
1056 |
+
w2: torch.Tensor,
|
1057 |
+
topk_weights: torch.Tensor,
|
1058 |
+
topk_ids: torch.Tensor,
|
1059 |
+
use_fp8_w8a8: bool = False,
|
1060 |
+
use_int8_w8a16: bool = False,
|
1061 |
+
use_int4_w4a16: bool = False,
|
1062 |
+
w1_scale: Optional[torch.Tensor] = None,
|
1063 |
+
w2_scale: Optional[torch.Tensor] = None,
|
1064 |
+
w1_zp: Optional[torch.Tensor] = None,
|
1065 |
+
w2_zp: Optional[torch.Tensor] = None,
|
1066 |
+
a1_scale: Optional[torch.Tensor] = None,
|
1067 |
+
a2_scale: Optional[torch.Tensor] = None,
|
1068 |
+
block_shape: Optional[List[int]] = None,
|
1069 |
+
) -> torch.Tensor:
|
1070 |
+
return fused_experts_impl(
|
1071 |
+
hidden_states,
|
1072 |
+
w1,
|
1073 |
+
w2,
|
1074 |
+
topk_weights,
|
1075 |
+
topk_ids,
|
1076 |
+
False,
|
1077 |
+
use_fp8_w8a8,
|
1078 |
+
use_int8_w8a16,
|
1079 |
+
use_int4_w4a16,
|
1080 |
+
w1_scale,
|
1081 |
+
w2_scale,
|
1082 |
+
w1_zp,
|
1083 |
+
w2_zp,
|
1084 |
+
a1_scale,
|
1085 |
+
a2_scale,
|
1086 |
+
block_shape,
|
1087 |
+
)
|
1088 |
+
|
1089 |
+
|
1090 |
def fused_experts(
|
1091 |
hidden_states: torch.Tensor,
|
1092 |
w1: torch.Tensor,
|
|
|
1094 |
topk_weights: torch.Tensor,
|
1095 |
topk_ids: torch.Tensor,
|
1096 |
inplace: bool = False,
|
|
|
1097 |
use_fp8_w8a8: bool = False,
|
1098 |
use_int8_w8a16: bool = False,
|
1099 |
+
use_int4_w4a16: bool = False,
|
1100 |
+
w1_scale: Optional[torch.Tensor] = None,
|
1101 |
+
w2_scale: Optional[torch.Tensor] = None,
|
1102 |
+
w1_zp: Optional[torch.Tensor] = None,
|
1103 |
+
w2_zp: Optional[torch.Tensor] = None,
|
1104 |
+
a1_scale: Optional[torch.Tensor] = None,
|
1105 |
+
a2_scale: Optional[torch.Tensor] = None,
|
1106 |
+
block_shape: Optional[List[int]] = None,
|
1107 |
+
):
|
1108 |
+
if inplace:
|
1109 |
+
inplace_fused_experts(
|
1110 |
+
hidden_states,
|
1111 |
+
w1,
|
1112 |
+
w2,
|
1113 |
+
topk_weights,
|
1114 |
+
topk_ids,
|
1115 |
+
use_fp8_w8a8,
|
1116 |
+
use_int8_w8a16,
|
1117 |
+
use_int4_w4a16,
|
1118 |
+
w1_scale,
|
1119 |
+
w2_scale,
|
1120 |
+
w1_zp,
|
1121 |
+
w2_zp,
|
1122 |
+
a1_scale,
|
1123 |
+
a2_scale,
|
1124 |
+
block_shape,
|
1125 |
+
)
|
1126 |
+
return hidden_states
|
1127 |
+
else:
|
1128 |
+
return outplace_fused_experts(
|
1129 |
+
hidden_states,
|
1130 |
+
w1,
|
1131 |
+
w2,
|
1132 |
+
topk_weights,
|
1133 |
+
topk_ids,
|
1134 |
+
use_fp8_w8a8,
|
1135 |
+
use_int8_w8a16,
|
1136 |
+
use_int4_w4a16,
|
1137 |
+
w1_scale,
|
1138 |
+
w2_scale,
|
1139 |
+
w1_zp,
|
1140 |
+
w2_zp,
|
1141 |
+
a1_scale,
|
1142 |
+
a2_scale,
|
1143 |
+
block_shape,
|
1144 |
+
)
|
1145 |
+
|
1146 |
+
|
1147 |
+
def fused_experts_impl(
|
1148 |
+
hidden_states: torch.Tensor,
|
1149 |
+
w1: torch.Tensor,
|
1150 |
+
w2: torch.Tensor,
|
1151 |
+
topk_weights: torch.Tensor,
|
1152 |
+
topk_ids: torch.Tensor,
|
1153 |
+
inplace: bool = False,
|
1154 |
+
use_fp8_w8a8: bool = False,
|
1155 |
+
use_int8_w8a16: bool = False,
|
1156 |
+
use_int4_w4a16: bool = False,
|
1157 |
w1_scale: Optional[torch.Tensor] = None,
|
1158 |
w2_scale: Optional[torch.Tensor] = None,
|
1159 |
+
w1_zp: Optional[torch.Tensor] = None,
|
1160 |
+
w2_zp: Optional[torch.Tensor] = None,
|
1161 |
a1_scale: Optional[torch.Tensor] = None,
|
1162 |
a2_scale: Optional[torch.Tensor] = None,
|
1163 |
+
block_shape: Optional[List[int]] = None,
|
1164 |
):
|
1165 |
# Check constraints.
|
1166 |
+
if use_int4_w4a16:
|
1167 |
+
assert hidden_states.shape[1] // 2 == w1.shape[2], "Hidden size mismatch"
|
1168 |
+
else:
|
1169 |
+
assert hidden_states.shape[1] == w1.shape[2], "Hidden size mismatch"
|
1170 |
+
|
1171 |
assert topk_weights.shape == topk_ids.shape, "topk shape mismatch"
|
1172 |
assert hidden_states.is_contiguous(), "Hidden_states must be contiguous"
|
1173 |
assert w1.is_contiguous(), "Expert weights1 must be contiguous"
|
|
|
1183 |
config_dtype = get_config_dtype_str(
|
1184 |
use_fp8_w8a8=use_fp8_w8a8,
|
1185 |
use_int8_w8a16=use_int8_w8a16,
|
1186 |
+
use_int4_w4a16=use_int4_w4a16,
|
1187 |
dtype=hidden_states.dtype,
|
1188 |
)
|
1189 |
|
|
|
1193 |
w2.shape,
|
1194 |
topk_ids.shape[1],
|
1195 |
config_dtype,
|
1196 |
+
block_shape=block_shape,
|
1197 |
)
|
1198 |
|
1199 |
config = get_config_func(M)
|
|
|
1214 |
dtype=hidden_states.dtype,
|
1215 |
)
|
1216 |
|
1217 |
+
if hidden_states.dtype == torch.bfloat16:
|
1218 |
+
compute_type = tl.bfloat16
|
1219 |
+
elif hidden_states.dtype == torch.float16:
|
1220 |
+
compute_type = tl.float16
|
1221 |
+
elif hidden_states.dtype == torch.float32:
|
1222 |
+
compute_type = tl.float32
|
1223 |
+
else:
|
1224 |
+
raise ValueError(f"Unsupported compute_type: {hidden_states.dtype}")
|
1225 |
|
1226 |
if inplace:
|
1227 |
out_hidden_states = hidden_states
|
|
|
1262 |
intermediate_cache1,
|
1263 |
a1_scale,
|
1264 |
w1_scale,
|
1265 |
+
w1_zp,
|
1266 |
curr_topk_weights,
|
1267 |
curr_topk_ids,
|
1268 |
sorted_token_ids,
|
|
|
1274 |
compute_type=compute_type,
|
1275 |
use_fp8_w8a8=use_fp8_w8a8,
|
1276 |
use_int8_w8a16=use_int8_w8a16,
|
1277 |
+
use_int4_w4a16=use_int4_w4a16,
|
1278 |
+
block_shape=block_shape,
|
1279 |
)
|
1280 |
|
1281 |
ops.silu_and_mul(intermediate_cache2, intermediate_cache1.view(-1, N))
|
|
|
1286 |
intermediate_cache3,
|
1287 |
a2_scale,
|
1288 |
w2_scale,
|
1289 |
+
w2_zp,
|
1290 |
curr_topk_weights,
|
1291 |
curr_topk_ids,
|
1292 |
sorted_token_ids,
|
|
|
1298 |
compute_type=compute_type,
|
1299 |
use_fp8_w8a8=use_fp8_w8a8,
|
1300 |
use_int8_w8a16=use_int8_w8a16,
|
1301 |
+
use_int4_w4a16=use_int4_w4a16,
|
1302 |
+
block_shape=block_shape,
|
1303 |
)
|
1304 |
|
1305 |
ops.moe_sum(
|
|
|
1317 |
topk: int,
|
1318 |
renormalize: bool,
|
1319 |
inplace: bool = False,
|
|
|
1320 |
use_grouped_topk: bool = False,
|
1321 |
num_expert_group: Optional[int] = None,
|
1322 |
topk_group: Optional[int] = None,
|
1323 |
custom_routing_function: Optional[Callable] = None,
|
1324 |
use_fp8_w8a8: bool = False,
|
1325 |
use_int8_w8a16: bool = False,
|
1326 |
+
use_int4_w4a16: bool = False,
|
1327 |
w1_scale: Optional[torch.Tensor] = None,
|
1328 |
w2_scale: Optional[torch.Tensor] = None,
|
1329 |
+
w1_zp: Optional[torch.Tensor] = None,
|
1330 |
+
w2_zp: Optional[torch.Tensor] = None,
|
1331 |
a1_scale: Optional[torch.Tensor] = None,
|
1332 |
a2_scale: Optional[torch.Tensor] = None,
|
1333 |
+
block_shape: Optional[List[int]] = None,
|
1334 |
) -> torch.Tensor:
|
1335 |
"""
|
1336 |
This function computes a Mixture of Experts (MoE) layer using two sets of
|
|
|
1346 |
- renormalize (bool): If True, renormalize the top-k weights to sum to 1.
|
1347 |
- inplace (bool): If True, perform the operation in-place.
|
1348 |
Defaults to False.
|
|
|
|
|
1349 |
- num_expert_group: Optional[int]: additional parameter for grouped_topk
|
1350 |
- topk_group: Optional[int]: additional parameter for grouped_topk
|
1351 |
- use_grouped_topk: If True, use grouped_topk instead of fused_topk
|
1352 |
note: Deepseekv2 model uses grouped_topk
|
1353 |
- use_fp8_w8a8 (bool): If True, use fp8 arithmetic to compute the inner
|
1354 |
products for w1 and w2. Defaults to False.
|
1355 |
+
- use_int8_w8a16 (bool): If True, use matmul of int8 weight and bf16/fp16
|
1356 |
+
activation to compute the inner products for w1 and w2.
|
1357 |
+
Defaults to False.
|
1358 |
+
- use_int4_w4a16 (bool): If True, use matmul of int4 weight and bf16/fp16
|
1359 |
+
activation to compute the inner products for w1 and w2.
|
1360 |
+
Defaults to False.
|
1361 |
- w1_scale (Optional[torch.Tensor]): Optional scale to be used for
|
1362 |
w1.
|
1363 |
- w2_scale (Optional[torch.Tensor]): Optional scale to be used for
|
1364 |
w2.
|
1365 |
+
- a1_scale (Optional[torch.Tensor]): Optional scale to be used for
|
1366 |
+
a1.
|
1367 |
+
- a2_scale (Optional[torch.Tensor]): Optional scale to be used for
|
1368 |
+
a2.
|
1369 |
+
- block_shape: (Optional[List[int]]): Optional block size for block-wise
|
1370 |
+
quantization.
|
1371 |
|
1372 |
Returns:
|
1373 |
- torch.Tensor: The output tensor after applying the MoE layer.
|
|
|
1401 |
topk_weights,
|
1402 |
topk_ids,
|
1403 |
inplace=inplace,
|
|
|
1404 |
use_fp8_w8a8=use_fp8_w8a8,
|
1405 |
use_int8_w8a16=use_int8_w8a16,
|
1406 |
+
use_int4_w4a16=use_int4_w4a16,
|
1407 |
w1_scale=w1_scale,
|
1408 |
w2_scale=w2_scale,
|
1409 |
+
w1_zp=w1_zp,
|
1410 |
+
w2_zp=w2_zp,
|
1411 |
a1_scale=a1_scale,
|
1412 |
a2_scale=a2_scale,
|
1413 |
+
block_shape=block_shape,
|
1414 |
)
|
build/torch25-cxx98-cu118-x86_64-linux/moe/platforms.py
CHANGED
@@ -1,22 +1,32 @@
|
|
1 |
-
from
|
2 |
-
import os
|
3 |
-
from functools import lru_cache, wraps
|
4 |
|
5 |
import torch
|
6 |
|
7 |
IS_ROCM = torch.version.hip is not None
|
8 |
|
9 |
-
|
|
|
|
|
|
|
|
|
|
|
10 |
@classmethod
|
11 |
@lru_cache(maxsize=8)
|
12 |
def get_device_name(cls, device_id: int = 0) -> str:
|
13 |
return torch.cuda.get_device_name(0)
|
14 |
|
15 |
-
|
|
|
|
|
|
|
|
|
16 |
@classmethod
|
17 |
@lru_cache(maxsize=8)
|
18 |
def get_device_name(cls, device_id: int = 0) -> str:
|
19 |
return torch.cuda.get_device_name(device_id)
|
20 |
|
|
|
|
|
|
|
21 |
|
22 |
current_platform = RocmPlatform() if IS_ROCM else CudaPlatform()
|
|
|
1 |
+
from functools import lru_cache
|
|
|
|
|
2 |
|
3 |
import torch
|
4 |
|
5 |
IS_ROCM = torch.version.hip is not None
|
6 |
|
7 |
+
|
8 |
+
class Platform:
|
9 |
+
simple_compile_backend: str = "inductor"
|
10 |
+
|
11 |
+
|
12 |
+
class CudaPlatform(Platform):
|
13 |
@classmethod
|
14 |
@lru_cache(maxsize=8)
|
15 |
def get_device_name(cls, device_id: int = 0) -> str:
|
16 |
return torch.cuda.get_device_name(0)
|
17 |
|
18 |
+
def is_rocm(self):
|
19 |
+
return False
|
20 |
+
|
21 |
+
|
22 |
+
class RocmPlatform(Platform):
|
23 |
@classmethod
|
24 |
@lru_cache(maxsize=8)
|
25 |
def get_device_name(cls, device_id: int = 0) -> str:
|
26 |
return torch.cuda.get_device_name(device_id)
|
27 |
|
28 |
+
def is_rocm(self):
|
29 |
+
return True
|
30 |
+
|
31 |
|
32 |
current_platform = RocmPlatform() if IS_ROCM else CudaPlatform()
|
build/torch25-cxx98-cu121-x86_64-linux/moe/_moe_tj3osoay2niyk.abi3.so
ADDED
@@ -0,0 +1,3 @@
|
|
|
|
|
|
|
|
|
1 |
+
version https://git-lfs.github.com/spec/v1
|
2 |
+
oid sha256:55b0eed6d5e4f8ef44d2f5baea4466cc633ae561aefd48dc54d648b9dc4742f3
|
3 |
+
size 86026776
|
build/torch25-cxx98-cu121-x86_64-linux/moe/_moe_xsk7dxl7fy4pk.abi3.so
DELETED
@@ -1,3 +0,0 @@
|
|
1 |
-
version https://git-lfs.github.com/spec/v1
|
2 |
-
oid sha256:53bd3b3d77a869ea6325993ff091433f370925006947f7a8218c02c6b24fddf9
|
3 |
-
size 84360992
|
|
|
|
|
|
|
|
build/torch25-cxx98-cu121-x86_64-linux/moe/_ops.py
CHANGED
@@ -1,9 +1,9 @@
|
|
1 |
import torch
|
2 |
-
from . import
|
3 |
-
ops = torch.ops.
|
4 |
|
5 |
def add_op_namespace_prefix(op_name: str):
|
6 |
"""
|
7 |
Prefix op by namespace.
|
8 |
"""
|
9 |
-
return f"
|
|
|
1 |
import torch
|
2 |
+
from . import _moe_tj3osoay2niyk
|
3 |
+
ops = torch.ops._moe_tj3osoay2niyk
|
4 |
|
5 |
def add_op_namespace_prefix(op_name: str):
|
6 |
"""
|
7 |
Prefix op by namespace.
|
8 |
"""
|
9 |
+
return f"_moe_tj3osoay2niyk::{op_name}"
|
build/torch25-cxx98-cu121-x86_64-linux/moe/fp8.py
CHANGED
@@ -1,6 +1,11 @@
|
|
|
|
|
|
1 |
import torch
|
|
|
|
|
2 |
|
3 |
-
|
|
|
4 |
|
5 |
|
6 |
def is_hip() -> bool:
|
@@ -49,15 +54,179 @@ def scaled_fp8_quant(
|
|
49 |
if scale is None:
|
50 |
if use_per_token_if_dynamic:
|
51 |
scale = torch.empty((shape[0], 1), device=input.device, dtype=torch.float32)
|
52 |
-
|
53 |
-
output, input, scale, scale_ub
|
54 |
-
)
|
55 |
else:
|
56 |
scale = torch.zeros(1, device=input.device, dtype=torch.float32)
|
57 |
-
|
58 |
else:
|
59 |
# num_token_padding not implemented for this case
|
60 |
assert scale.numel() == 1 or num_token_padding is None
|
61 |
-
|
62 |
|
63 |
return output, scale
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
1 |
+
from typing import Tuple, Optional, Union
|
2 |
+
|
3 |
import torch
|
4 |
+
import triton
|
5 |
+
import triton.language as tl
|
6 |
|
7 |
+
|
8 |
+
from ._ops import ops
|
9 |
|
10 |
|
11 |
def is_hip() -> bool:
|
|
|
54 |
if scale is None:
|
55 |
if use_per_token_if_dynamic:
|
56 |
scale = torch.empty((shape[0], 1), device=input.device, dtype=torch.float32)
|
57 |
+
ops.dynamic_per_token_scaled_fp8_quant(output, input, scale, scale_ub)
|
|
|
|
|
58 |
else:
|
59 |
scale = torch.zeros(1, device=input.device, dtype=torch.float32)
|
60 |
+
ops.dynamic_scaled_fp8_quant(output, input, scale)
|
61 |
else:
|
62 |
# num_token_padding not implemented for this case
|
63 |
assert scale.numel() == 1 or num_token_padding is None
|
64 |
+
ops.static_scaled_fp8_quant(output, input, scale)
|
65 |
|
66 |
return output, scale
|
67 |
+
|
68 |
+
|
69 |
+
@triton.jit
|
70 |
+
def _per_token_group_quant_fp8(
|
71 |
+
# Pointers to inputs and output
|
72 |
+
y_ptr,
|
73 |
+
y_q_ptr,
|
74 |
+
y_s_ptr,
|
75 |
+
group_size,
|
76 |
+
# Avoid to divide zero
|
77 |
+
eps,
|
78 |
+
# Information for float8
|
79 |
+
fp8_min,
|
80 |
+
fp8_max,
|
81 |
+
# Meta-parameters
|
82 |
+
BLOCK: tl.constexpr,
|
83 |
+
):
|
84 |
+
"""A Triton-accelerated function to perform per-token-group
|
85 |
+
quantization on a tensor.
|
86 |
+
This function converts the tensor values into float8 values.
|
87 |
+
"""
|
88 |
+
# Map the program id to the row of X and Y it should compute.
|
89 |
+
g_id = tl.program_id(0)
|
90 |
+
y_ptr += g_id * group_size
|
91 |
+
y_q_ptr += g_id * group_size
|
92 |
+
y_s_ptr += g_id
|
93 |
+
|
94 |
+
cols = tl.arange(0, BLOCK) # N <= BLOCK
|
95 |
+
mask = cols < group_size
|
96 |
+
|
97 |
+
y = tl.load(y_ptr + cols, mask=mask, other=0.0).to(tl.float32)
|
98 |
+
# Quant
|
99 |
+
_absmax = tl.maximum(tl.max(tl.abs(y)), eps)
|
100 |
+
y_s = _absmax / fp8_max
|
101 |
+
y_q = tl.clamp(y / y_s, fp8_min, fp8_max).to(y_q_ptr.dtype.element_ty)
|
102 |
+
|
103 |
+
tl.store(y_q_ptr + cols, y_q, mask=mask)
|
104 |
+
tl.store(y_s_ptr, y_s)
|
105 |
+
|
106 |
+
|
107 |
+
@triton.jit
|
108 |
+
def _per_token_group_quant_fp8_colmajor(
|
109 |
+
# Pointers to inputs and output
|
110 |
+
y_ptr,
|
111 |
+
y_q_ptr,
|
112 |
+
y_s_ptr,
|
113 |
+
group_size,
|
114 |
+
# Num columns of y
|
115 |
+
y_num_columns,
|
116 |
+
# Stride from one column to the next of y_s
|
117 |
+
y_s_col_stride,
|
118 |
+
# Avoid to divide zero
|
119 |
+
eps,
|
120 |
+
# Information for float8
|
121 |
+
fp8_min,
|
122 |
+
fp8_max,
|
123 |
+
# Meta-parameters
|
124 |
+
BLOCK: tl.constexpr,
|
125 |
+
):
|
126 |
+
"""A Triton-accelerated function to perform per-token-group
|
127 |
+
quantization on a tensor.
|
128 |
+
This function converts the tensor values into float8 values.
|
129 |
+
"""
|
130 |
+
# Map the program id to the row of X and Y it should compute.
|
131 |
+
g_id = tl.program_id(0)
|
132 |
+
y_ptr += g_id * group_size
|
133 |
+
y_q_ptr += g_id * group_size
|
134 |
+
|
135 |
+
# Convert g_id the flattened block coordinate to 2D so we can index
|
136 |
+
# into the output y_scales matrix
|
137 |
+
blocks_per_row = y_num_columns // group_size
|
138 |
+
scale_col = g_id % blocks_per_row
|
139 |
+
scale_row = g_id // blocks_per_row
|
140 |
+
y_s_ptr += scale_col * y_s_col_stride + scale_row
|
141 |
+
|
142 |
+
cols = tl.arange(0, BLOCK) # group_size <= BLOCK
|
143 |
+
mask = cols < group_size
|
144 |
+
|
145 |
+
y = tl.load(y_ptr + cols, mask=mask, other=0.0).to(tl.float32)
|
146 |
+
# Quant
|
147 |
+
_absmax = tl.maximum(tl.max(tl.abs(y)), eps)
|
148 |
+
y_s = _absmax / fp8_max
|
149 |
+
y_q = tl.clamp(y / y_s, fp8_min, fp8_max).to(y_q_ptr.dtype.element_ty)
|
150 |
+
|
151 |
+
tl.store(y_q_ptr + cols, y_q, mask=mask)
|
152 |
+
tl.store(y_s_ptr, y_s)
|
153 |
+
|
154 |
+
|
155 |
+
def per_token_group_quant_fp8(
|
156 |
+
x: torch.Tensor,
|
157 |
+
group_size: int,
|
158 |
+
eps: float = 1e-10,
|
159 |
+
dtype: Optional[torch.dtype] = None,
|
160 |
+
column_major_scales: bool = False,
|
161 |
+
) -> Tuple[torch.Tensor, torch.Tensor]:
|
162 |
+
"""Function to perform per-token-group quantization on an input tensor `x`.
|
163 |
+
It converts the tensor values into signed float8 values and returns the
|
164 |
+
quantized tensor along with the scaling factor used for quantization.
|
165 |
+
Args:
|
166 |
+
x: The input tensor with ndim >= 2.
|
167 |
+
group_size: The group size used for quantization.
|
168 |
+
eps: The minimum to avoid dividing zero.
|
169 |
+
dtype: The dype of output tensor. Note that only `torch.float8_e4m3fn`
|
170 |
+
is supported for now.
|
171 |
+
Returns:
|
172 |
+
Tuple[torch.Tensor, torch.Tensor]: The quantized tensor and the
|
173 |
+
scaling factor for quantization.
|
174 |
+
"""
|
175 |
+
if dtype is None:
|
176 |
+
dtype = (
|
177 |
+
torch.float8_e4m3fnuz if current_platform.is_rocm() else torch.float8_e4m3fn
|
178 |
+
)
|
179 |
+
assert x.shape[-1] % group_size == 0, (
|
180 |
+
f"the last dimension of `x` {x.shape[-1]} must be divisible "
|
181 |
+
f"by `group_size` {group_size}"
|
182 |
+
)
|
183 |
+
assert x.is_contiguous(), "`x` must be contiguous"
|
184 |
+
|
185 |
+
finfo = torch.finfo(dtype)
|
186 |
+
fp8_min = finfo.min
|
187 |
+
fp8_max = finfo.max
|
188 |
+
|
189 |
+
x_q = torch.empty_like(x, device=x.device, dtype=dtype)
|
190 |
+
M = x.numel() // group_size
|
191 |
+
N = group_size
|
192 |
+
if column_major_scales:
|
193 |
+
shape = (x.shape[-1] // group_size,) + x.shape[:-1]
|
194 |
+
x_s = torch.empty(shape, device=x.device, dtype=torch.float32).permute(-1, -2)
|
195 |
+
else:
|
196 |
+
shape = x.shape[:-1] + (x.shape[-1] // group_size,)
|
197 |
+
x_s = torch.empty(shape, device=x.device, dtype=torch.float32)
|
198 |
+
|
199 |
+
BLOCK = triton.next_power_of_2(N)
|
200 |
+
# heuristics for number of warps
|
201 |
+
num_warps = min(max(BLOCK // 256, 1), 8)
|
202 |
+
num_stages = 1
|
203 |
+
if column_major_scales:
|
204 |
+
_per_token_group_quant_fp8_colmajor[(M,)](
|
205 |
+
x,
|
206 |
+
x_q,
|
207 |
+
x_s,
|
208 |
+
group_size,
|
209 |
+
x.shape[1],
|
210 |
+
x_s.stride(1),
|
211 |
+
eps,
|
212 |
+
fp8_min=fp8_min,
|
213 |
+
fp8_max=fp8_max,
|
214 |
+
BLOCK=BLOCK,
|
215 |
+
num_warps=num_warps,
|
216 |
+
num_stages=num_stages,
|
217 |
+
)
|
218 |
+
else:
|
219 |
+
_per_token_group_quant_fp8[(M,)](
|
220 |
+
x,
|
221 |
+
x_q,
|
222 |
+
x_s,
|
223 |
+
group_size,
|
224 |
+
eps,
|
225 |
+
fp8_min=fp8_min,
|
226 |
+
fp8_max=fp8_max,
|
227 |
+
BLOCK=BLOCK,
|
228 |
+
num_warps=num_warps,
|
229 |
+
num_stages=num_stages,
|
230 |
+
)
|
231 |
+
|
232 |
+
return x_q, x_s
|
build/torch25-cxx98-cu121-x86_64-linux/moe/fused_marlin_moe.py
CHANGED
@@ -40,7 +40,6 @@ def single_marlin_moe(
|
|
40 |
g_idx: Optional[torch.Tensor] = None,
|
41 |
sort_indices: Optional[torch.Tensor] = None,
|
42 |
w_zeros: Optional[torch.Tensor] = None,
|
43 |
-
override_config: Optional[Dict[str, Any]] = None,
|
44 |
num_bits: int = 8,
|
45 |
is_k_full: bool = True,
|
46 |
) -> torch.Tensor:
|
@@ -61,8 +60,6 @@ def single_marlin_moe(
|
|
61 |
- topk (int): The number of top-k experts to select.
|
62 |
- renormalize (bool): If True, renormalize the top-k weights to sum to 1.
|
63 |
- w_zeros (Optional[torch.Tensor]): Optional zero points to be used for w.
|
64 |
-
- override_config (Optional[Dict[str, Any]]): Optional override
|
65 |
-
for the kernel configuration.
|
66 |
- num_bits (bool): The number of bits in expert weights quantization.
|
67 |
|
68 |
Returns:
|
@@ -90,7 +87,6 @@ def single_marlin_moe(
|
|
90 |
w.shape,
|
91 |
topk_ids.shape[1],
|
92 |
None,
|
93 |
-
override_config=override_config,
|
94 |
is_marlin=True,
|
95 |
)
|
96 |
config = get_config_func(M)
|
@@ -154,6 +150,25 @@ def single_marlin_moe(
|
|
154 |
return torch.sum(intermediate_cache.view(*intermediate_cache.shape), dim=1)
|
155 |
|
156 |
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
157 |
def fused_marlin_moe(
|
158 |
hidden_states: torch.Tensor,
|
159 |
w1: torch.Tensor,
|
@@ -169,7 +184,6 @@ def fused_marlin_moe(
|
|
169 |
sort_indices2: Optional[torch.Tensor] = None,
|
170 |
w1_zeros: Optional[torch.Tensor] = None,
|
171 |
w2_zeros: Optional[torch.Tensor] = None,
|
172 |
-
override_config: Optional[Dict[str, Any]] = None,
|
173 |
num_bits: int = 8,
|
174 |
is_k_full: bool = True,
|
175 |
) -> torch.Tensor:
|
@@ -193,8 +207,6 @@ def fused_marlin_moe(
|
|
193 |
permutation.
|
194 |
- topk_weights (torch.Tensor): Top-k weights.
|
195 |
- topk_ids (torch.Tensor): Indices of topk-k elements.
|
196 |
-
- override_config (Optional[Dict[str, Any]]): Optional override
|
197 |
-
for the kernel configuration.
|
198 |
- w1_zeros (Optional[torch.Tensor]): Optional zero points to be used for w1.
|
199 |
- w2_zeros (Optional[torch.Tensor]): Optional zero points to be used for w2.
|
200 |
- num_bits (bool): The number of bits in expert weights quantization.
|
@@ -248,7 +260,6 @@ def fused_marlin_moe(
|
|
248 |
w2.shape,
|
249 |
topk_ids.shape[1],
|
250 |
None,
|
251 |
-
override_config=override_config,
|
252 |
is_marlin=True,
|
253 |
)
|
254 |
config = get_config_func(M)
|
@@ -350,6 +361,30 @@ def fused_marlin_moe(
|
|
350 |
return torch.sum(intermediate_cache3.view(*intermediate_cache3.shape), dim=1)
|
351 |
|
352 |
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
353 |
if hasattr(ops, "marlin_gemm_moe"):
|
354 |
|
355 |
@register_fake(add_op_namespace_prefix("marlin_gemm_moe"))
|
|
|
40 |
g_idx: Optional[torch.Tensor] = None,
|
41 |
sort_indices: Optional[torch.Tensor] = None,
|
42 |
w_zeros: Optional[torch.Tensor] = None,
|
|
|
43 |
num_bits: int = 8,
|
44 |
is_k_full: bool = True,
|
45 |
) -> torch.Tensor:
|
|
|
60 |
- topk (int): The number of top-k experts to select.
|
61 |
- renormalize (bool): If True, renormalize the top-k weights to sum to 1.
|
62 |
- w_zeros (Optional[torch.Tensor]): Optional zero points to be used for w.
|
|
|
|
|
63 |
- num_bits (bool): The number of bits in expert weights quantization.
|
64 |
|
65 |
Returns:
|
|
|
87 |
w.shape,
|
88 |
topk_ids.shape[1],
|
89 |
None,
|
|
|
90 |
is_marlin=True,
|
91 |
)
|
92 |
config = get_config_func(M)
|
|
|
150 |
return torch.sum(intermediate_cache.view(*intermediate_cache.shape), dim=1)
|
151 |
|
152 |
|
153 |
+
if hasattr(ops, "single_marlin_gemm_moe"):
|
154 |
+
|
155 |
+
@register_fake(add_op_namespace_prefix("single_marlin_gemm_moe"))
|
156 |
+
def single_marlin_moe_fake(
|
157 |
+
hidden_states: torch.Tensor,
|
158 |
+
w: torch.Tensor,
|
159 |
+
scales: torch.Tensor,
|
160 |
+
gating_output: torch.Tensor,
|
161 |
+
topk: int,
|
162 |
+
renormalize: bool,
|
163 |
+
g_idx: Optional[torch.Tensor] = None,
|
164 |
+
sort_indices: Optional[torch.Tensor] = None,
|
165 |
+
w_zeros: Optional[torch.Tensor] = None,
|
166 |
+
num_bits: int = 8,
|
167 |
+
is_k_full: bool = True,
|
168 |
+
) -> torch.Tensor:
|
169 |
+
return torch.empty_like(hidden_states)
|
170 |
+
|
171 |
+
|
172 |
def fused_marlin_moe(
|
173 |
hidden_states: torch.Tensor,
|
174 |
w1: torch.Tensor,
|
|
|
184 |
sort_indices2: Optional[torch.Tensor] = None,
|
185 |
w1_zeros: Optional[torch.Tensor] = None,
|
186 |
w2_zeros: Optional[torch.Tensor] = None,
|
|
|
187 |
num_bits: int = 8,
|
188 |
is_k_full: bool = True,
|
189 |
) -> torch.Tensor:
|
|
|
207 |
permutation.
|
208 |
- topk_weights (torch.Tensor): Top-k weights.
|
209 |
- topk_ids (torch.Tensor): Indices of topk-k elements.
|
|
|
|
|
210 |
- w1_zeros (Optional[torch.Tensor]): Optional zero points to be used for w1.
|
211 |
- w2_zeros (Optional[torch.Tensor]): Optional zero points to be used for w2.
|
212 |
- num_bits (bool): The number of bits in expert weights quantization.
|
|
|
260 |
w2.shape,
|
261 |
topk_ids.shape[1],
|
262 |
None,
|
|
|
263 |
is_marlin=True,
|
264 |
)
|
265 |
config = get_config_func(M)
|
|
|
361 |
return torch.sum(intermediate_cache3.view(*intermediate_cache3.shape), dim=1)
|
362 |
|
363 |
|
364 |
+
if hasattr(ops, "fused_marlin_moe"):
|
365 |
+
|
366 |
+
@register_fake(add_op_namespace_prefix("fused_marlin_moe"))
|
367 |
+
def fused_marlin_moe_fake(
|
368 |
+
hidden_states: torch.Tensor,
|
369 |
+
w1: torch.Tensor,
|
370 |
+
w2: torch.Tensor,
|
371 |
+
w1_scale: torch.Tensor,
|
372 |
+
w2_scale: torch.Tensor,
|
373 |
+
gating_output: torch.Tensor,
|
374 |
+
topk_weights: torch.Tensor,
|
375 |
+
topk_ids: torch.Tensor,
|
376 |
+
g_idx1: Optional[torch.Tensor] = None,
|
377 |
+
g_idx2: Optional[torch.Tensor] = None,
|
378 |
+
sort_indices1: Optional[torch.Tensor] = None,
|
379 |
+
sort_indices2: Optional[torch.Tensor] = None,
|
380 |
+
w1_zeros: Optional[torch.Tensor] = None,
|
381 |
+
w2_zeros: Optional[torch.Tensor] = None,
|
382 |
+
num_bits: int = 8,
|
383 |
+
is_k_full: bool = True,
|
384 |
+
) -> torch.Tensor:
|
385 |
+
return torch.empty_like(hidden_states)
|
386 |
+
|
387 |
+
|
388 |
if hasattr(ops, "marlin_gemm_moe"):
|
389 |
|
390 |
@register_fake(add_op_namespace_prefix("marlin_gemm_moe"))
|
build/torch25-cxx98-cu121-x86_64-linux/moe/fused_moe.py
CHANGED
@@ -1,21 +1,242 @@
|
|
|
|
1 |
"""Fused MoE kernel."""
|
2 |
|
3 |
import functools
|
4 |
import json
|
|
|
5 |
import os
|
6 |
-
from typing import Any, Callable, Dict, Optional, Tuple
|
7 |
|
8 |
import torch
|
9 |
import triton
|
10 |
import triton.language as tl
|
11 |
|
|
|
12 |
from ._ops import ops
|
13 |
-
from .fp8 import scaled_fp8_quant
|
14 |
from .platforms import current_platform
|
15 |
|
|
|
|
|
|
|
16 |
VLLM_FUSED_MOE_CHUNK_SIZE = int(os.getenv("VLLM_FUSED_MOE_CHUNK_SIZE", "32768"))
|
17 |
|
18 |
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
19 |
@triton.jit
|
20 |
def fused_moe_kernel(
|
21 |
# Pointers to matrices
|
@@ -44,8 +265,14 @@ def fused_moe_kernel(
|
|
44 |
stride_bn,
|
45 |
stride_cm,
|
46 |
stride_cn,
|
|
|
|
|
47 |
stride_bse,
|
|
|
48 |
stride_bsn,
|
|
|
|
|
|
|
49 |
# Meta-parameters
|
50 |
BLOCK_SIZE_M: tl.constexpr,
|
51 |
BLOCK_SIZE_N: tl.constexpr,
|
@@ -105,17 +332,17 @@ def fused_moe_kernel(
|
|
105 |
num_tokens_post_padded = tl.load(num_tokens_post_padded_ptr)
|
106 |
if pid_m * BLOCK_SIZE_M >= num_tokens_post_padded:
|
107 |
return
|
108 |
-
offs_token_id = pid_m * BLOCK_SIZE_M + tl.arange(0, BLOCK_SIZE_M)
|
109 |
offs_token = tl.load(sorted_token_ids_ptr + offs_token_id)
|
110 |
token_mask = offs_token < num_valid_tokens
|
111 |
|
112 |
-
offs_bn = (pid_n * BLOCK_SIZE_N + tl.arange(0, BLOCK_SIZE_N)) % N
|
113 |
offs_k = tl.arange(0, BLOCK_SIZE_K)
|
114 |
a_ptrs = a_ptr + (
|
115 |
offs_token[:, None] // top_k * stride_am + offs_k[None, :] * stride_ak
|
116 |
)
|
117 |
|
118 |
-
off_experts = tl.load(expert_ids_ptr + pid_m)
|
119 |
b_ptrs = (
|
120 |
b_ptr
|
121 |
+ off_experts * stride_be
|
@@ -128,8 +355,15 @@ def fused_moe_kernel(
|
|
128 |
b_scale = tl.load(b_scale_ptrs)
|
129 |
|
130 |
if use_fp8_w8a8:
|
131 |
-
|
132 |
-
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
133 |
|
134 |
# -----------------------------------------------------------
|
135 |
# Iterate to compute a block of the C matrix.
|
@@ -151,7 +385,17 @@ def fused_moe_kernel(
|
|
151 |
if use_int8_w8a16:
|
152 |
accumulator = tl.dot(a, b.to(compute_type), acc=accumulator)
|
153 |
elif use_fp8_w8a8:
|
154 |
-
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
155 |
else:
|
156 |
accumulator += tl.dot(a, b)
|
157 |
# Advance the ptrs to the next K block.
|
@@ -164,7 +408,10 @@ def fused_moe_kernel(
|
|
164 |
if use_int8_w8a16:
|
165 |
accumulator = (accumulator * b_scale).to(compute_type)
|
166 |
elif use_fp8_w8a8:
|
167 |
-
|
|
|
|
|
|
|
168 |
else:
|
169 |
accumulator = accumulator.to(compute_type)
|
170 |
# -----------------------------------------------------------
|
@@ -175,6 +422,141 @@ def fused_moe_kernel(
|
|
175 |
tl.store(c_ptrs, accumulator, mask=c_mask)
|
176 |
|
177 |
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
178 |
def moe_align_block_size(
|
179 |
topk_ids: torch.Tensor, block_size: int, num_experts: int
|
180 |
) -> Tuple[torch.Tensor, torch.Tensor, torch.Tensor]:
|
@@ -225,9 +607,34 @@ def moe_align_block_size(
|
|
225 |
(max_num_m_blocks,), dtype=torch.int32, device=topk_ids.device
|
226 |
)
|
227 |
num_tokens_post_pad = torch.empty((1), dtype=torch.int32, device=topk_ids.device)
|
228 |
-
|
229 |
-
|
230 |
-
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
231 |
return sorted_ids, expert_ids, num_tokens_post_pad
|
232 |
|
233 |
|
@@ -237,6 +644,7 @@ def invoke_fused_moe_kernel(
|
|
237 |
C: torch.Tensor,
|
238 |
A_scale: Optional[torch.Tensor],
|
239 |
B_scale: Optional[torch.Tensor],
|
|
|
240 |
topk_weights: torch.Tensor,
|
241 |
topk_ids: torch.Tensor,
|
242 |
sorted_token_ids: torch.Tensor,
|
@@ -248,64 +656,147 @@ def invoke_fused_moe_kernel(
|
|
248 |
compute_type: tl.dtype,
|
249 |
use_fp8_w8a8: bool,
|
250 |
use_int8_w8a16: bool,
|
|
|
|
|
251 |
) -> None:
|
252 |
assert topk_weights.stride(1) == 1
|
253 |
assert sorted_token_ids.stride(0) == 1
|
254 |
|
255 |
if use_fp8_w8a8:
|
256 |
-
A, A_scale = scaled_fp8_quant(A, A_scale)
|
257 |
assert B_scale is not None
|
258 |
-
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
259 |
assert B_scale is not None
|
|
|
260 |
else:
|
261 |
assert A_scale is None
|
262 |
assert B_scale is None
|
263 |
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
264 |
grid = lambda META: (
|
265 |
-
triton.cdiv(
|
266 |
* triton.cdiv(B.shape[1], META["BLOCK_SIZE_N"]),
|
267 |
)
|
268 |
|
269 |
-
|
270 |
-
|
271 |
-
|
272 |
-
|
273 |
-
|
274 |
-
B_scale
|
275 |
-
|
276 |
-
|
277 |
-
|
278 |
-
|
279 |
-
|
280 |
-
|
281 |
-
|
282 |
-
|
283 |
-
|
284 |
-
|
285 |
-
|
286 |
-
|
287 |
-
|
288 |
-
|
289 |
-
|
290 |
-
|
291 |
-
|
292 |
-
|
293 |
-
|
294 |
-
|
295 |
-
|
296 |
-
|
297 |
-
|
298 |
-
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
299 |
|
300 |
|
301 |
-
|
|
|
|
|
|
|
302 |
device_name = current_platform.get_device_name().replace(" ", "_")
|
303 |
dtype_selector = "" if not dtype else f",dtype={dtype}"
|
304 |
-
|
|
|
|
|
|
|
305 |
|
306 |
|
|
|
307 |
@functools.lru_cache
|
308 |
-
def get_moe_configs(
|
|
|
|
|
|
|
|
|
|
|
|
|
309 |
"""
|
310 |
Return optimized configurations for the fused MoE kernel.
|
311 |
|
@@ -317,18 +808,27 @@ def get_moe_configs(E: int, N: int, dtype: Optional[str]) -> Optional[Dict[int,
|
|
317 |
|
318 |
# First look up if an optimized configuration is available in the configs
|
319 |
# directory
|
320 |
-
|
|
|
321 |
|
322 |
config_file_path = os.path.join(
|
323 |
os.path.dirname(os.path.realpath(__file__)), "configs", json_file_name
|
324 |
)
|
325 |
if os.path.exists(config_file_path):
|
326 |
with open(config_file_path) as f:
|
|
|
327 |
# If a configuration has been found, return it
|
328 |
return {int(key): val for key, val in json.load(f).items()}
|
329 |
|
330 |
# If no optimized configuration is available, we will use the default
|
331 |
# configuration
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
332 |
return None
|
333 |
|
334 |
|
@@ -340,21 +840,34 @@ def get_default_config(
|
|
340 |
topk: int,
|
341 |
dtype: Optional[str],
|
342 |
is_marlin: bool,
|
|
|
343 |
) -> Dict[str, int]:
|
344 |
-
|
345 |
-
|
346 |
-
|
347 |
-
"BLOCK_SIZE_K": 32,
|
348 |
-
"GROUP_SIZE_M": 8,
|
349 |
-
}
|
350 |
-
# A heuristic: fused marlin works faster with this config for small M
|
351 |
-
if M <= E or (is_marlin and M <= 32):
|
352 |
config = {
|
353 |
-
"BLOCK_SIZE_M":
|
354 |
-
"BLOCK_SIZE_N":
|
355 |
-
"BLOCK_SIZE_K":
|
356 |
-
"GROUP_SIZE_M":
|
|
|
|
|
357 |
}
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
358 |
return config
|
359 |
|
360 |
|
@@ -364,15 +877,21 @@ def try_get_optimal_moe_config(
|
|
364 |
top_k: int,
|
365 |
dtype: Optional[str],
|
366 |
M: int,
|
367 |
-
override_config: Optional[Dict[str, Any]] = None,
|
368 |
is_marlin: bool = False,
|
|
|
369 |
):
|
|
|
|
|
|
|
|
|
370 |
if override_config:
|
371 |
config = override_config
|
372 |
else:
|
373 |
# First try to load optimal config from the file
|
374 |
E, _, N = w2_shape
|
375 |
-
|
|
|
|
|
376 |
|
377 |
if configs:
|
378 |
# If an optimal configuration map has been found, look up the
|
@@ -380,7 +899,9 @@ def try_get_optimal_moe_config(
|
|
380 |
config = configs[min(configs.keys(), key=lambda x: abs(x - M))]
|
381 |
else:
|
382 |
# Else use the default config
|
383 |
-
config = get_default_config(
|
|
|
|
|
384 |
return config
|
385 |
|
386 |
|
@@ -416,7 +937,8 @@ def fused_topk(
|
|
416 |
return topk_weights, topk_ids
|
417 |
|
418 |
|
419 |
-
# This is used by the Deepseek-V2 model
|
|
|
420 |
def grouped_topk(
|
421 |
hidden_states: torch.Tensor,
|
422 |
gating_output: torch.Tensor,
|
@@ -424,11 +946,25 @@ def grouped_topk(
|
|
424 |
renormalize: bool,
|
425 |
num_expert_group: int = 0,
|
426 |
topk_group: int = 0,
|
|
|
|
|
427 |
):
|
428 |
|
429 |
assert hidden_states.shape[0] == gating_output.shape[0], "Number of tokens mismatch"
|
430 |
|
431 |
-
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
432 |
num_token = scores.shape[0]
|
433 |
group_scores = (
|
434 |
scores.view(num_token, num_expert_group, -1).max(dim=-1).values
|
@@ -444,7 +980,13 @@ def grouped_topk(
|
|
444 |
.reshape(num_token, -1)
|
445 |
) # [n, e]
|
446 |
tmp_scores = scores.masked_fill(~score_mask.bool(), 0.0) # [n, e]
|
447 |
-
|
|
|
|
|
|
|
|
|
|
|
|
|
448 |
|
449 |
if renormalize:
|
450 |
topk_weights = topk_weights / topk_weights.sum(dim=-1, keepdim=True)
|
@@ -454,6 +996,7 @@ def grouped_topk(
|
|
454 |
|
455 |
def get_config_dtype_str(
|
456 |
dtype: torch.dtype,
|
|
|
457 |
use_int8_w8a16: Optional[bool] = False,
|
458 |
use_fp8_w8a8: Optional[bool] = False,
|
459 |
):
|
@@ -461,6 +1004,8 @@ def get_config_dtype_str(
|
|
461 |
return "fp8_w8a8"
|
462 |
elif use_int8_w8a16:
|
463 |
return "int8_w8a16"
|
|
|
|
|
464 |
elif dtype == torch.float:
|
465 |
# avoiding cases where kernel fails when float32 MoE
|
466 |
# use fp16/bfloat16 configs
|
@@ -468,6 +1013,80 @@ def get_config_dtype_str(
|
|
468 |
return None
|
469 |
|
470 |
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
471 |
def fused_experts(
|
472 |
hidden_states: torch.Tensor,
|
473 |
w1: torch.Tensor,
|
@@ -475,16 +1094,80 @@ def fused_experts(
|
|
475 |
topk_weights: torch.Tensor,
|
476 |
topk_ids: torch.Tensor,
|
477 |
inplace: bool = False,
|
478 |
-
override_config: Optional[Dict[str, Any]] = None,
|
479 |
use_fp8_w8a8: bool = False,
|
480 |
use_int8_w8a16: bool = False,
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
481 |
w1_scale: Optional[torch.Tensor] = None,
|
482 |
w2_scale: Optional[torch.Tensor] = None,
|
|
|
|
|
483 |
a1_scale: Optional[torch.Tensor] = None,
|
484 |
a2_scale: Optional[torch.Tensor] = None,
|
|
|
485 |
):
|
486 |
# Check constraints.
|
487 |
-
|
|
|
|
|
|
|
|
|
488 |
assert topk_weights.shape == topk_ids.shape, "topk shape mismatch"
|
489 |
assert hidden_states.is_contiguous(), "Hidden_states must be contiguous"
|
490 |
assert w1.is_contiguous(), "Expert weights1 must be contiguous"
|
@@ -500,6 +1183,7 @@ def fused_experts(
|
|
500 |
config_dtype = get_config_dtype_str(
|
501 |
use_fp8_w8a8=use_fp8_w8a8,
|
502 |
use_int8_w8a16=use_int8_w8a16,
|
|
|
503 |
dtype=hidden_states.dtype,
|
504 |
)
|
505 |
|
@@ -509,7 +1193,7 @@ def fused_experts(
|
|
509 |
w2.shape,
|
510 |
topk_ids.shape[1],
|
511 |
config_dtype,
|
512 |
-
|
513 |
)
|
514 |
|
515 |
config = get_config_func(M)
|
@@ -530,7 +1214,14 @@ def fused_experts(
|
|
530 |
dtype=hidden_states.dtype,
|
531 |
)
|
532 |
|
533 |
-
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
534 |
|
535 |
if inplace:
|
536 |
out_hidden_states = hidden_states
|
@@ -571,6 +1262,7 @@ def fused_experts(
|
|
571 |
intermediate_cache1,
|
572 |
a1_scale,
|
573 |
w1_scale,
|
|
|
574 |
curr_topk_weights,
|
575 |
curr_topk_ids,
|
576 |
sorted_token_ids,
|
@@ -582,6 +1274,8 @@ def fused_experts(
|
|
582 |
compute_type=compute_type,
|
583 |
use_fp8_w8a8=use_fp8_w8a8,
|
584 |
use_int8_w8a16=use_int8_w8a16,
|
|
|
|
|
585 |
)
|
586 |
|
587 |
ops.silu_and_mul(intermediate_cache2, intermediate_cache1.view(-1, N))
|
@@ -592,6 +1286,7 @@ def fused_experts(
|
|
592 |
intermediate_cache3,
|
593 |
a2_scale,
|
594 |
w2_scale,
|
|
|
595 |
curr_topk_weights,
|
596 |
curr_topk_ids,
|
597 |
sorted_token_ids,
|
@@ -603,6 +1298,8 @@ def fused_experts(
|
|
603 |
compute_type=compute_type,
|
604 |
use_fp8_w8a8=use_fp8_w8a8,
|
605 |
use_int8_w8a16=use_int8_w8a16,
|
|
|
|
|
606 |
)
|
607 |
|
608 |
ops.moe_sum(
|
@@ -620,17 +1317,20 @@ def fused_moe(
|
|
620 |
topk: int,
|
621 |
renormalize: bool,
|
622 |
inplace: bool = False,
|
623 |
-
override_config: Optional[Dict[str, Any]] = None,
|
624 |
use_grouped_topk: bool = False,
|
625 |
num_expert_group: Optional[int] = None,
|
626 |
topk_group: Optional[int] = None,
|
627 |
custom_routing_function: Optional[Callable] = None,
|
628 |
use_fp8_w8a8: bool = False,
|
629 |
use_int8_w8a16: bool = False,
|
|
|
630 |
w1_scale: Optional[torch.Tensor] = None,
|
631 |
w2_scale: Optional[torch.Tensor] = None,
|
|
|
|
|
632 |
a1_scale: Optional[torch.Tensor] = None,
|
633 |
a2_scale: Optional[torch.Tensor] = None,
|
|
|
634 |
) -> torch.Tensor:
|
635 |
"""
|
636 |
This function computes a Mixture of Experts (MoE) layer using two sets of
|
@@ -646,20 +1346,28 @@ def fused_moe(
|
|
646 |
- renormalize (bool): If True, renormalize the top-k weights to sum to 1.
|
647 |
- inplace (bool): If True, perform the operation in-place.
|
648 |
Defaults to False.
|
649 |
-
- override_config (Optional[Dict[str, Any]]): Optional override
|
650 |
-
for the kernel configuration.
|
651 |
- num_expert_group: Optional[int]: additional parameter for grouped_topk
|
652 |
- topk_group: Optional[int]: additional parameter for grouped_topk
|
653 |
- use_grouped_topk: If True, use grouped_topk instead of fused_topk
|
654 |
note: Deepseekv2 model uses grouped_topk
|
655 |
- use_fp8_w8a8 (bool): If True, use fp8 arithmetic to compute the inner
|
656 |
products for w1 and w2. Defaults to False.
|
657 |
-
- use_int8_w8a16 (bool): If True, use
|
658 |
-
products for w1 and w2.
|
|
|
|
|
|
|
|
|
659 |
- w1_scale (Optional[torch.Tensor]): Optional scale to be used for
|
660 |
w1.
|
661 |
- w2_scale (Optional[torch.Tensor]): Optional scale to be used for
|
662 |
w2.
|
|
|
|
|
|
|
|
|
|
|
|
|
663 |
|
664 |
Returns:
|
665 |
- torch.Tensor: The output tensor after applying the MoE layer.
|
@@ -693,11 +1401,14 @@ def fused_moe(
|
|
693 |
topk_weights,
|
694 |
topk_ids,
|
695 |
inplace=inplace,
|
696 |
-
override_config=override_config,
|
697 |
use_fp8_w8a8=use_fp8_w8a8,
|
698 |
use_int8_w8a16=use_int8_w8a16,
|
|
|
699 |
w1_scale=w1_scale,
|
700 |
w2_scale=w2_scale,
|
|
|
|
|
701 |
a1_scale=a1_scale,
|
702 |
a2_scale=a2_scale,
|
|
|
703 |
)
|
|
|
1 |
+
# SPDX-License-Identifier: Apache-2.0
|
2 |
"""Fused MoE kernel."""
|
3 |
|
4 |
import functools
|
5 |
import json
|
6 |
+
import logging
|
7 |
import os
|
8 |
+
from typing import Any, Callable, Dict, List, Optional, Tuple
|
9 |
|
10 |
import torch
|
11 |
import triton
|
12 |
import triton.language as tl
|
13 |
|
14 |
+
|
15 |
from ._ops import ops
|
16 |
+
from .fp8 import per_token_group_quant_fp8, scaled_fp8_quant
|
17 |
from .platforms import current_platform
|
18 |
|
19 |
+
logger = logging.getLogger(__name__)
|
20 |
+
|
21 |
+
|
22 |
VLLM_FUSED_MOE_CHUNK_SIZE = int(os.getenv("VLLM_FUSED_MOE_CHUNK_SIZE", "32768"))
|
23 |
|
24 |
|
25 |
+
@triton.jit
|
26 |
+
def fused_moe_kernel_gptq_awq(
|
27 |
+
# Pointers to matrices
|
28 |
+
a_ptr,
|
29 |
+
b_ptr,
|
30 |
+
c_ptr,
|
31 |
+
b_scale_ptr,
|
32 |
+
b_zp_ptr,
|
33 |
+
topk_weights_ptr,
|
34 |
+
sorted_token_ids_ptr,
|
35 |
+
expert_ids_ptr,
|
36 |
+
num_tokens_post_padded_ptr,
|
37 |
+
# Matrix dimensions
|
38 |
+
N: tl.constexpr,
|
39 |
+
K: tl.constexpr,
|
40 |
+
EM,
|
41 |
+
num_valid_tokens,
|
42 |
+
# The stride variables represent how much to increase the ptr by when
|
43 |
+
# moving by 1 element in a particular dimension. E.g. `stride_am` is
|
44 |
+
# how much to increase `a_ptr` by to get the element one row down
|
45 |
+
# (A has M rows).
|
46 |
+
stride_am,
|
47 |
+
stride_ak,
|
48 |
+
stride_be,
|
49 |
+
stride_bk,
|
50 |
+
stride_bn,
|
51 |
+
stride_cm,
|
52 |
+
stride_cn,
|
53 |
+
stride_bse,
|
54 |
+
stride_bsk,
|
55 |
+
stride_bsn,
|
56 |
+
stride_bze,
|
57 |
+
stride_bzk,
|
58 |
+
stride_bzn,
|
59 |
+
block_k_diviable: tl.constexpr,
|
60 |
+
group_size: tl.constexpr,
|
61 |
+
# Meta-parameters
|
62 |
+
BLOCK_SIZE_M: tl.constexpr,
|
63 |
+
BLOCK_SIZE_N: tl.constexpr,
|
64 |
+
BLOCK_SIZE_K: tl.constexpr,
|
65 |
+
GROUP_SIZE_M: tl.constexpr,
|
66 |
+
MUL_ROUTED_WEIGHT: tl.constexpr,
|
67 |
+
top_k: tl.constexpr,
|
68 |
+
compute_type: tl.constexpr,
|
69 |
+
has_zp: tl.constexpr,
|
70 |
+
use_int4_w4a16: tl.constexpr,
|
71 |
+
use_int8_w8a16: tl.constexpr,
|
72 |
+
):
|
73 |
+
"""
|
74 |
+
Implements the fused computation for a Mixture of Experts (MOE) using
|
75 |
+
token and expert matrices.
|
76 |
+
|
77 |
+
Key Parameters:
|
78 |
+
- A: The input tensor representing tokens with shape (*, K), where '*' can
|
79 |
+
be any shape representing batches and K is the feature dimension of
|
80 |
+
each token.
|
81 |
+
- B: The stacked MOE weight tensor with shape (E, N, K), where E is
|
82 |
+
the number of experts, K is the input feature dimension, and N is
|
83 |
+
the output feature dimension.
|
84 |
+
- C: The output cache tensor with shape (M, topk, N), where M is the
|
85 |
+
total number of tokens post padding, topk is the number of times
|
86 |
+
each token is repeated, and N is the output feature dimension.
|
87 |
+
- sorted_token_ids: A tensor containing the sorted indices of tokens,
|
88 |
+
repeated topk times and arranged by the expert index they are
|
89 |
+
assigned to.
|
90 |
+
- expert_ids: A tensor containing the indices of the expert for each
|
91 |
+
block. It determines which expert matrix from B should be used for
|
92 |
+
each block in A.
|
93 |
+
This kernel performs the multiplication of a token by its corresponding
|
94 |
+
expert matrix as determined by `expert_ids`. The sorting of
|
95 |
+
`sorted_token_ids` by expert index and padding ensures divisibility by
|
96 |
+
BLOCK_SIZE_M, which is necessary to maintain consistency in block matrix
|
97 |
+
multiplication across different blocks processed by the same expert.
|
98 |
+
"""
|
99 |
+
# -----------------------------------------------------------
|
100 |
+
# Map program ids `pid` to the block of C it should compute.
|
101 |
+
# This is done in a grouped ordering to promote L2 data reuse.
|
102 |
+
pid = tl.program_id(axis=0)
|
103 |
+
num_pid_m = tl.cdiv(EM, BLOCK_SIZE_M)
|
104 |
+
num_pid_n = tl.cdiv(N, BLOCK_SIZE_N)
|
105 |
+
num_pid_in_group = GROUP_SIZE_M * num_pid_n
|
106 |
+
group_id = pid // num_pid_in_group
|
107 |
+
first_pid_m = group_id * GROUP_SIZE_M
|
108 |
+
group_size_m = min(num_pid_m - first_pid_m, GROUP_SIZE_M)
|
109 |
+
pid_m = first_pid_m + ((pid % num_pid_in_group) % group_size_m)
|
110 |
+
pid_n = (pid % num_pid_in_group) // group_size_m
|
111 |
+
|
112 |
+
# ----------------------------------------------------------
|
113 |
+
# Create pointers for the first blocks of A and B.
|
114 |
+
# We will advance this pointer as we move in the K direction
|
115 |
+
# and accumulate
|
116 |
+
# `a_ptrs` is a block of [BLOCK_SIZE_M, BLOCK_SIZE_K] pointers
|
117 |
+
# `b_ptrs` is a block of [BLOCK_SIZE_K, BLOCK_SIZE_N] pointers
|
118 |
+
num_tokens_post_padded = tl.load(num_tokens_post_padded_ptr)
|
119 |
+
if pid_m * BLOCK_SIZE_M >= num_tokens_post_padded:
|
120 |
+
return
|
121 |
+
offs_token_id = pid_m * BLOCK_SIZE_M + tl.arange(0, BLOCK_SIZE_M).to(tl.int64)
|
122 |
+
offs_token = tl.load(sorted_token_ids_ptr + offs_token_id)
|
123 |
+
token_mask = offs_token < num_valid_tokens
|
124 |
+
|
125 |
+
offs_bn = (pid_n * BLOCK_SIZE_N + tl.arange(0, BLOCK_SIZE_N).to(tl.int64)) % N
|
126 |
+
offs_k = tl.arange(0, BLOCK_SIZE_K)
|
127 |
+
a_ptrs = a_ptr + (
|
128 |
+
offs_token[:, None] // top_k * stride_am + offs_k[None, :] * stride_ak
|
129 |
+
)
|
130 |
+
|
131 |
+
off_experts = tl.load(expert_ids_ptr + pid_m).to(tl.int64)
|
132 |
+
|
133 |
+
if use_int4_w4a16:
|
134 |
+
b_ptrs = (
|
135 |
+
b_ptr
|
136 |
+
+ off_experts * stride_be
|
137 |
+
+ (offs_k[:, None] // 2) * stride_bk
|
138 |
+
+ offs_bn[None, :] * stride_bn
|
139 |
+
)
|
140 |
+
b_shifter = (offs_k[:, None] % 2) * 4
|
141 |
+
elif use_int8_w8a16:
|
142 |
+
b_ptrs = (
|
143 |
+
b_ptr
|
144 |
+
+ off_experts * stride_be
|
145 |
+
+ offs_k[:, None] * stride_bk
|
146 |
+
+ offs_bn[None, :] * stride_bn
|
147 |
+
)
|
148 |
+
|
149 |
+
if not has_zp and use_int4_w4a16:
|
150 |
+
b_zp_num = 8
|
151 |
+
if not has_zp and use_int8_w8a16:
|
152 |
+
b_zp_num = 128
|
153 |
+
elif has_zp and use_int4_w4a16:
|
154 |
+
b_zp_shifter = (offs_bn[None, :] % 2) * 4
|
155 |
+
|
156 |
+
# -----------------------------------------------------------
|
157 |
+
# Iterate to compute a block of the C matrix.
|
158 |
+
# We accumulate into a `[BLOCK_SIZE_M, BLOCK_SIZE_N]` block
|
159 |
+
# of fp32 values for higher accuracy.
|
160 |
+
# `accumulator` will be converted back to fp16 after the loop.
|
161 |
+
accumulator = tl.zeros((BLOCK_SIZE_M, BLOCK_SIZE_N), dtype=tl.float32)
|
162 |
+
for k in range(0, tl.cdiv(K, BLOCK_SIZE_K)):
|
163 |
+
# Load the next block of A and B, generate a mask by checking the
|
164 |
+
# K dimension.
|
165 |
+
|
166 |
+
if not block_k_diviable:
|
167 |
+
k_mask = offs_k[:, None] < K - k * BLOCK_SIZE_K
|
168 |
+
k_other = 0.0
|
169 |
+
else:
|
170 |
+
k_mask = None
|
171 |
+
k_other = None
|
172 |
+
|
173 |
+
a = tl.load(
|
174 |
+
a_ptrs,
|
175 |
+
mask=token_mask[:, None] & (offs_k[None, :] < K - k * BLOCK_SIZE_K),
|
176 |
+
other=0.0,
|
177 |
+
)
|
178 |
+
b = tl.load(b_ptrs)
|
179 |
+
if use_int4_w4a16:
|
180 |
+
b = (b >> b_shifter) & 0xF
|
181 |
+
|
182 |
+
b_scale_ptrs = (
|
183 |
+
b_scale_ptr
|
184 |
+
+ off_experts * stride_bse
|
185 |
+
+ offs_bn[None, :] * stride_bsn
|
186 |
+
+ ((offs_k[:, None] + BLOCK_SIZE_K * k) // group_size) * stride_bsk
|
187 |
+
)
|
188 |
+
b_scale = tl.load(b_scale_ptrs, mask=k_mask, other=k_other)
|
189 |
+
b_scale = b_scale.to(tl.float32)
|
190 |
+
|
191 |
+
if has_zp and use_int4_w4a16:
|
192 |
+
offs_k_true = (offs_k[:, None] + BLOCK_SIZE_K * k) // group_size
|
193 |
+
b_zp_ptrs = (
|
194 |
+
b_zp_ptr
|
195 |
+
+ off_experts * stride_bze
|
196 |
+
+ (offs_bn[None, :] // 2) * stride_bzn
|
197 |
+
+ offs_k_true * stride_bzk
|
198 |
+
)
|
199 |
+
b_zp = tl.load(b_zp_ptrs, mask=k_mask, other=k_other)
|
200 |
+
b_zp = (b_zp >> b_zp_shifter) & 0xF
|
201 |
+
b_zp = b_zp.to(tl.float32)
|
202 |
+
elif has_zp and use_int8_w8a16:
|
203 |
+
offs_k_true = (offs_k[:, None] + BLOCK_SIZE_K * k) // group_size
|
204 |
+
b_zp_ptrs = (
|
205 |
+
b_zp_ptr
|
206 |
+
+ off_experts * stride_bze
|
207 |
+
+ offs_bn[None, :] * stride_bzn
|
208 |
+
+ offs_k_true * stride_bzk
|
209 |
+
)
|
210 |
+
b_zp = tl.load(b_zp_ptrs, mask=k_mask, other=k_other)
|
211 |
+
b_zp = b_zp.to(tl.float32)
|
212 |
+
|
213 |
+
# We accumulate along the K dimension.
|
214 |
+
if has_zp:
|
215 |
+
b = ((b.to(tl.float32) - b_zp) * b_scale).to(compute_type)
|
216 |
+
else:
|
217 |
+
b = ((b.to(tl.float32) - b_zp_num) * b_scale).to(compute_type)
|
218 |
+
accumulator = tl.dot(a, b, acc=accumulator)
|
219 |
+
|
220 |
+
# Advance the ptrs to the next K block.
|
221 |
+
a_ptrs += BLOCK_SIZE_K * stride_ak
|
222 |
+
if use_int4_w4a16:
|
223 |
+
b_ptrs += (BLOCK_SIZE_K // 2) * stride_bk
|
224 |
+
else:
|
225 |
+
b_ptrs += BLOCK_SIZE_K * stride_bk
|
226 |
+
|
227 |
+
if MUL_ROUTED_WEIGHT:
|
228 |
+
moe_weight = tl.load(topk_weights_ptr + offs_token, mask=token_mask, other=0)
|
229 |
+
accumulator = accumulator * moe_weight[:, None]
|
230 |
+
|
231 |
+
accumulator = accumulator.to(compute_type)
|
232 |
+
# -----------------------------------------------------------
|
233 |
+
# Write back the block of the output
|
234 |
+
offs_cn = pid_n * BLOCK_SIZE_N + tl.arange(0, BLOCK_SIZE_N)
|
235 |
+
c_ptrs = c_ptr + stride_cm * offs_token[:, None] + stride_cn * offs_cn[None, :]
|
236 |
+
c_mask = token_mask[:, None] & (offs_cn[None, :] < N)
|
237 |
+
tl.store(c_ptrs, accumulator, mask=c_mask)
|
238 |
+
|
239 |
+
|
240 |
@triton.jit
|
241 |
def fused_moe_kernel(
|
242 |
# Pointers to matrices
|
|
|
265 |
stride_bn,
|
266 |
stride_cm,
|
267 |
stride_cn,
|
268 |
+
stride_asm,
|
269 |
+
stride_ask,
|
270 |
stride_bse,
|
271 |
+
stride_bsk,
|
272 |
stride_bsn,
|
273 |
+
# Block size for block-wise quantization
|
274 |
+
group_n: tl.constexpr,
|
275 |
+
group_k: tl.constexpr,
|
276 |
# Meta-parameters
|
277 |
BLOCK_SIZE_M: tl.constexpr,
|
278 |
BLOCK_SIZE_N: tl.constexpr,
|
|
|
332 |
num_tokens_post_padded = tl.load(num_tokens_post_padded_ptr)
|
333 |
if pid_m * BLOCK_SIZE_M >= num_tokens_post_padded:
|
334 |
return
|
335 |
+
offs_token_id = pid_m * BLOCK_SIZE_M + tl.arange(0, BLOCK_SIZE_M).to(tl.int64)
|
336 |
offs_token = tl.load(sorted_token_ids_ptr + offs_token_id)
|
337 |
token_mask = offs_token < num_valid_tokens
|
338 |
|
339 |
+
offs_bn = (pid_n * BLOCK_SIZE_N + tl.arange(0, BLOCK_SIZE_N).to(tl.int64)) % N
|
340 |
offs_k = tl.arange(0, BLOCK_SIZE_K)
|
341 |
a_ptrs = a_ptr + (
|
342 |
offs_token[:, None] // top_k * stride_am + offs_k[None, :] * stride_ak
|
343 |
)
|
344 |
|
345 |
+
off_experts = tl.load(expert_ids_ptr + pid_m).to(tl.int64)
|
346 |
b_ptrs = (
|
347 |
b_ptr
|
348 |
+ off_experts * stride_be
|
|
|
355 |
b_scale = tl.load(b_scale_ptrs)
|
356 |
|
357 |
if use_fp8_w8a8:
|
358 |
+
if group_k > 0 and group_n > 0:
|
359 |
+
a_scale_ptrs = a_scale_ptr + (offs_token // top_k) * stride_asm
|
360 |
+
offs_bsn = offs_bn // group_n
|
361 |
+
b_scale_ptrs = (
|
362 |
+
b_scale_ptr + off_experts * stride_bse + offs_bsn * stride_bsn
|
363 |
+
)
|
364 |
+
else:
|
365 |
+
a_scale = tl.load(a_scale_ptr)
|
366 |
+
b_scale = tl.load(b_scale_ptr + off_experts)
|
367 |
|
368 |
# -----------------------------------------------------------
|
369 |
# Iterate to compute a block of the C matrix.
|
|
|
385 |
if use_int8_w8a16:
|
386 |
accumulator = tl.dot(a, b.to(compute_type), acc=accumulator)
|
387 |
elif use_fp8_w8a8:
|
388 |
+
if group_k > 0 and group_n > 0:
|
389 |
+
k_start = k * BLOCK_SIZE_K
|
390 |
+
offs_ks = k_start // group_k
|
391 |
+
a_scale = tl.load(
|
392 |
+
a_scale_ptrs + offs_ks * stride_ask, mask=token_mask, other=0.0
|
393 |
+
)
|
394 |
+
b_scale = tl.load(b_scale_ptrs + offs_ks * stride_bsk)
|
395 |
+
|
396 |
+
accumulator += tl.dot(a, b) * a_scale[:, None] * b_scale[None, :]
|
397 |
+
else:
|
398 |
+
accumulator = tl.dot(a, b, acc=accumulator)
|
399 |
else:
|
400 |
accumulator += tl.dot(a, b)
|
401 |
# Advance the ptrs to the next K block.
|
|
|
408 |
if use_int8_w8a16:
|
409 |
accumulator = (accumulator * b_scale).to(compute_type)
|
410 |
elif use_fp8_w8a8:
|
411 |
+
if group_k > 0 and group_n > 0:
|
412 |
+
accumulator = accumulator.to(compute_type)
|
413 |
+
else:
|
414 |
+
accumulator = (accumulator * a_scale * b_scale).to(compute_type)
|
415 |
else:
|
416 |
accumulator = accumulator.to(compute_type)
|
417 |
# -----------------------------------------------------------
|
|
|
422 |
tl.store(c_ptrs, accumulator, mask=c_mask)
|
423 |
|
424 |
|
425 |
+
def ceil_div(a, b):
|
426 |
+
return (a + b - 1) // b
|
427 |
+
|
428 |
+
|
429 |
+
@triton.jit
|
430 |
+
def moe_align_block_size_stage1(
|
431 |
+
topk_ids_ptr,
|
432 |
+
tokens_cnts_ptr,
|
433 |
+
num_experts: tl.constexpr,
|
434 |
+
numel: tl.constexpr,
|
435 |
+
tokens_per_thread: tl.constexpr,
|
436 |
+
):
|
437 |
+
pid = tl.program_id(0)
|
438 |
+
|
439 |
+
start_idx = pid * tokens_per_thread
|
440 |
+
|
441 |
+
off_c = (pid + 1) * num_experts
|
442 |
+
|
443 |
+
for i in range(tokens_per_thread):
|
444 |
+
if start_idx + i < numel:
|
445 |
+
idx = tl.load(topk_ids_ptr + start_idx + i)
|
446 |
+
token_cnt = tl.load(tokens_cnts_ptr + off_c + idx)
|
447 |
+
tl.store(tokens_cnts_ptr + off_c + idx, token_cnt + 1)
|
448 |
+
|
449 |
+
|
450 |
+
@triton.jit
|
451 |
+
def moe_align_block_size_stage2(
|
452 |
+
tokens_cnts_ptr,
|
453 |
+
num_experts: tl.constexpr,
|
454 |
+
):
|
455 |
+
pid = tl.program_id(0)
|
456 |
+
|
457 |
+
last_cnt = 0
|
458 |
+
for i in range(1, num_experts + 1):
|
459 |
+
token_cnt = tl.load(tokens_cnts_ptr + i * num_experts + pid)
|
460 |
+
last_cnt = last_cnt + token_cnt
|
461 |
+
tl.store(tokens_cnts_ptr + i * num_experts + pid, last_cnt)
|
462 |
+
|
463 |
+
|
464 |
+
@triton.jit
|
465 |
+
def moe_align_block_size_stage3(
|
466 |
+
total_tokens_post_pad_ptr,
|
467 |
+
tokens_cnts_ptr,
|
468 |
+
cumsum_ptr,
|
469 |
+
num_experts: tl.constexpr,
|
470 |
+
block_size: tl.constexpr,
|
471 |
+
):
|
472 |
+
last_cumsum = 0
|
473 |
+
off_cnt = num_experts * num_experts
|
474 |
+
for i in range(1, num_experts + 1):
|
475 |
+
token_cnt = tl.load(tokens_cnts_ptr + off_cnt + i - 1)
|
476 |
+
last_cumsum = last_cumsum + tl.cdiv(token_cnt, block_size) * block_size
|
477 |
+
tl.store(cumsum_ptr + i, last_cumsum)
|
478 |
+
tl.store(total_tokens_post_pad_ptr, last_cumsum)
|
479 |
+
|
480 |
+
|
481 |
+
@triton.jit
|
482 |
+
def moe_align_block_size_stage4(
|
483 |
+
topk_ids_ptr,
|
484 |
+
sorted_token_ids_ptr,
|
485 |
+
expert_ids_ptr,
|
486 |
+
tokens_cnts_ptr,
|
487 |
+
cumsum_ptr,
|
488 |
+
num_experts: tl.constexpr,
|
489 |
+
block_size: tl.constexpr,
|
490 |
+
numel: tl.constexpr,
|
491 |
+
tokens_per_thread: tl.constexpr,
|
492 |
+
):
|
493 |
+
pid = tl.program_id(0)
|
494 |
+
start_idx = tl.load(cumsum_ptr + pid)
|
495 |
+
end_idx = tl.load(cumsum_ptr + pid + 1)
|
496 |
+
|
497 |
+
for i in range(start_idx, end_idx, block_size):
|
498 |
+
tl.store(expert_ids_ptr + i // block_size, pid)
|
499 |
+
|
500 |
+
start_idx = pid * tokens_per_thread
|
501 |
+
off_t = pid * num_experts
|
502 |
+
|
503 |
+
for i in range(start_idx, tl.minimum(start_idx + tokens_per_thread, numel)):
|
504 |
+
expert_id = tl.load(topk_ids_ptr + i)
|
505 |
+
token_cnt = tl.load(tokens_cnts_ptr + off_t + expert_id)
|
506 |
+
rank_post_pad = token_cnt + tl.load(cumsum_ptr + expert_id)
|
507 |
+
tl.store(sorted_token_ids_ptr + rank_post_pad, i)
|
508 |
+
tl.store(tokens_cnts_ptr + off_t + expert_id, token_cnt + 1)
|
509 |
+
|
510 |
+
|
511 |
+
# Triton implementation based on:
|
512 |
+
# https://github.com/sgl-project/sglang/commit/ba5112ff691d791a9e38c6c71f59324a5fcb49d0
|
513 |
+
def moe_align_block_size_triton(
|
514 |
+
topk_ids: torch.Tensor,
|
515 |
+
num_experts: int,
|
516 |
+
block_size: int,
|
517 |
+
sorted_token_ids: torch.Tensor,
|
518 |
+
expert_ids: torch.Tensor,
|
519 |
+
num_tokens_post_pad: torch.Tensor,
|
520 |
+
) -> None:
|
521 |
+
numel = topk_ids.numel()
|
522 |
+
grid = (num_experts,)
|
523 |
+
tokens_cnts = torch.zeros(
|
524 |
+
(num_experts + 1, num_experts), dtype=torch.int32, device=topk_ids.device
|
525 |
+
)
|
526 |
+
cumsum = torch.zeros((num_experts + 1,), dtype=torch.int32, device=topk_ids.device)
|
527 |
+
tokens_per_thread = ceil_div(numel, num_experts)
|
528 |
+
|
529 |
+
moe_align_block_size_stage1[grid](
|
530 |
+
topk_ids,
|
531 |
+
tokens_cnts,
|
532 |
+
num_experts,
|
533 |
+
numel,
|
534 |
+
tokens_per_thread,
|
535 |
+
)
|
536 |
+
moe_align_block_size_stage2[grid](
|
537 |
+
tokens_cnts,
|
538 |
+
num_experts,
|
539 |
+
)
|
540 |
+
moe_align_block_size_stage3[(1,)](
|
541 |
+
num_tokens_post_pad,
|
542 |
+
tokens_cnts,
|
543 |
+
cumsum,
|
544 |
+
num_experts,
|
545 |
+
block_size,
|
546 |
+
)
|
547 |
+
moe_align_block_size_stage4[grid](
|
548 |
+
topk_ids,
|
549 |
+
sorted_token_ids,
|
550 |
+
expert_ids,
|
551 |
+
tokens_cnts,
|
552 |
+
cumsum,
|
553 |
+
num_experts,
|
554 |
+
block_size,
|
555 |
+
numel,
|
556 |
+
tokens_per_thread,
|
557 |
+
)
|
558 |
+
|
559 |
+
|
560 |
def moe_align_block_size(
|
561 |
topk_ids: torch.Tensor, block_size: int, num_experts: int
|
562 |
) -> Tuple[torch.Tensor, torch.Tensor, torch.Tensor]:
|
|
|
607 |
(max_num_m_blocks,), dtype=torch.int32, device=topk_ids.device
|
608 |
)
|
609 |
num_tokens_post_pad = torch.empty((1), dtype=torch.int32, device=topk_ids.device)
|
610 |
+
if num_experts >= 224:
|
611 |
+
if VLLM_ENABLE_MOE_ALIGN_BLOCK_SIZE_TRITON:
|
612 |
+
moe_align_block_size_triton(
|
613 |
+
topk_ids,
|
614 |
+
num_experts,
|
615 |
+
block_size,
|
616 |
+
sorted_ids,
|
617 |
+
expert_ids,
|
618 |
+
num_tokens_post_pad,
|
619 |
+
)
|
620 |
+
else:
|
621 |
+
ops.sgl_moe_align_block_size(
|
622 |
+
topk_ids,
|
623 |
+
num_experts,
|
624 |
+
block_size,
|
625 |
+
sorted_ids,
|
626 |
+
expert_ids,
|
627 |
+
num_tokens_post_pad,
|
628 |
+
)
|
629 |
+
else:
|
630 |
+
ops.moe_align_block_size(
|
631 |
+
topk_ids,
|
632 |
+
num_experts,
|
633 |
+
block_size,
|
634 |
+
sorted_ids,
|
635 |
+
expert_ids,
|
636 |
+
num_tokens_post_pad,
|
637 |
+
)
|
638 |
return sorted_ids, expert_ids, num_tokens_post_pad
|
639 |
|
640 |
|
|
|
644 |
C: torch.Tensor,
|
645 |
A_scale: Optional[torch.Tensor],
|
646 |
B_scale: Optional[torch.Tensor],
|
647 |
+
B_zp: Optional[torch.Tensor],
|
648 |
topk_weights: torch.Tensor,
|
649 |
topk_ids: torch.Tensor,
|
650 |
sorted_token_ids: torch.Tensor,
|
|
|
656 |
compute_type: tl.dtype,
|
657 |
use_fp8_w8a8: bool,
|
658 |
use_int8_w8a16: bool,
|
659 |
+
use_int4_w4a16: bool,
|
660 |
+
block_shape: Optional[List[int]] = None,
|
661 |
) -> None:
|
662 |
assert topk_weights.stride(1) == 1
|
663 |
assert sorted_token_ids.stride(0) == 1
|
664 |
|
665 |
if use_fp8_w8a8:
|
|
|
666 |
assert B_scale is not None
|
667 |
+
if block_shape is None:
|
668 |
+
A, A_scale = scaled_fp8_quant(A, A_scale)
|
669 |
+
else:
|
670 |
+
assert len(block_shape) == 2
|
671 |
+
block_n, block_k = block_shape[0], block_shape[1]
|
672 |
+
A, A_scale = per_token_group_quant_fp8(A, block_k)
|
673 |
+
assert triton.cdiv(A.shape[-1], block_k) == A_scale.shape[-1]
|
674 |
+
assert triton.cdiv(B.shape[-2], block_n) == B_scale.shape[-2]
|
675 |
+
assert triton.cdiv(B.shape[-1], block_k) == B_scale.shape[-1]
|
676 |
+
elif use_int8_w8a16 or use_int4_w4a16:
|
677 |
assert B_scale is not None
|
678 |
+
assert block_shape is None or block_shape[0] == 0
|
679 |
else:
|
680 |
assert A_scale is None
|
681 |
assert B_scale is None
|
682 |
|
683 |
+
EM = sorted_token_ids.shape[0]
|
684 |
+
if A.shape[0] < config["BLOCK_SIZE_M"]:
|
685 |
+
# optimize for small batch_size.
|
686 |
+
# We assume that top_ids of each token is unique, so
|
687 |
+
# so num_valid_experts <= batch_size <= BLOCK_SIZE_M,
|
688 |
+
# and we can skip some invalid blocks.
|
689 |
+
EM = min(sorted_token_ids.shape[0], A.shape[0] * top_k * config["BLOCK_SIZE_M"])
|
690 |
grid = lambda META: (
|
691 |
+
triton.cdiv(EM, META["BLOCK_SIZE_M"])
|
692 |
* triton.cdiv(B.shape[1], META["BLOCK_SIZE_N"]),
|
693 |
)
|
694 |
|
695 |
+
if (
|
696 |
+
(use_int8_w8a16 or use_int4_w4a16)
|
697 |
+
and block_shape is not None
|
698 |
+
and block_shape[1] > 0
|
699 |
+
):
|
700 |
+
assert B_scale is not None and B_scale.ndim == 3
|
701 |
+
assert B_zp is None or B_zp.ndim == 3
|
702 |
+
|
703 |
+
fused_moe_kernel_gptq_awq[grid](
|
704 |
+
A,
|
705 |
+
B,
|
706 |
+
C,
|
707 |
+
B_scale,
|
708 |
+
B_zp,
|
709 |
+
topk_weights,
|
710 |
+
sorted_token_ids,
|
711 |
+
expert_ids,
|
712 |
+
num_tokens_post_padded,
|
713 |
+
B.shape[1],
|
714 |
+
A.shape[1],
|
715 |
+
EM,
|
716 |
+
topk_ids.numel(),
|
717 |
+
A.stride(0),
|
718 |
+
A.stride(1),
|
719 |
+
B.stride(0),
|
720 |
+
B.stride(2),
|
721 |
+
B.stride(1),
|
722 |
+
C.stride(1),
|
723 |
+
C.stride(2),
|
724 |
+
B_scale.stride(0),
|
725 |
+
B_scale.stride(2),
|
726 |
+
B_scale.stride(1),
|
727 |
+
B_zp.stride(0) if B_zp is not None else 0,
|
728 |
+
B_zp.stride(2) if B_zp is not None else 0,
|
729 |
+
B_zp.stride(1) if B_zp is not None else 0,
|
730 |
+
block_k_diviable=A.shape[1] % config["BLOCK_SIZE_K"] == 0,
|
731 |
+
group_size=block_shape[1],
|
732 |
+
MUL_ROUTED_WEIGHT=mul_routed_weight,
|
733 |
+
top_k=top_k,
|
734 |
+
compute_type=compute_type,
|
735 |
+
has_zp=B_zp is not None,
|
736 |
+
use_int4_w4a16=use_int4_w4a16,
|
737 |
+
use_int8_w8a16=use_int8_w8a16,
|
738 |
+
**config,
|
739 |
+
)
|
740 |
+
|
741 |
+
else:
|
742 |
+
fused_moe_kernel[grid](
|
743 |
+
A,
|
744 |
+
B,
|
745 |
+
C,
|
746 |
+
A_scale,
|
747 |
+
B_scale,
|
748 |
+
topk_weights,
|
749 |
+
sorted_token_ids,
|
750 |
+
expert_ids,
|
751 |
+
num_tokens_post_padded,
|
752 |
+
B.shape[1],
|
753 |
+
A.shape[1],
|
754 |
+
EM,
|
755 |
+
topk_ids.numel(),
|
756 |
+
A.stride(0),
|
757 |
+
A.stride(1),
|
758 |
+
B.stride(0),
|
759 |
+
B.stride(2),
|
760 |
+
B.stride(1),
|
761 |
+
C.stride(1),
|
762 |
+
C.stride(2),
|
763 |
+
A_scale.stride(0) if A_scale is not None and A_scale.ndim == 2 else 0,
|
764 |
+
A_scale.stride(1) if A_scale is not None and A_scale.ndim == 2 else 0,
|
765 |
+
B_scale.stride(0) if B_scale is not None and B_scale.ndim >= 2 else 0,
|
766 |
+
B_scale.stride(2) if B_scale is not None and B_scale.ndim == 3 else 0,
|
767 |
+
B_scale.stride(1) if B_scale is not None and B_scale.ndim >= 2 else 0,
|
768 |
+
0 if block_shape is None else block_shape[0],
|
769 |
+
0 if block_shape is None else block_shape[1],
|
770 |
+
MUL_ROUTED_WEIGHT=mul_routed_weight,
|
771 |
+
top_k=top_k,
|
772 |
+
compute_type=compute_type,
|
773 |
+
use_fp8_w8a8=use_fp8_w8a8,
|
774 |
+
use_int8_w8a16=use_int8_w8a16,
|
775 |
+
**config,
|
776 |
+
)
|
777 |
|
778 |
|
779 |
+
# Adapted from: https://github.com/sgl-project/sglang/pull/2628
|
780 |
+
def get_config_file_name(
|
781 |
+
E: int, N: int, dtype: Optional[str], block_shape: Optional[List[int]] = None
|
782 |
+
) -> str:
|
783 |
device_name = current_platform.get_device_name().replace(" ", "_")
|
784 |
dtype_selector = "" if not dtype else f",dtype={dtype}"
|
785 |
+
block_shape_selector = (
|
786 |
+
"" if not block_shape or not all(block_shape) else f",block_shape={block_shape}"
|
787 |
+
)
|
788 |
+
return f"E={E},N={N},device_name={device_name}{dtype_selector}{block_shape_selector}.json" # noqa: E501
|
789 |
|
790 |
|
791 |
+
# Adapted from: https://github.com/sgl-project/sglang/pull/2628
|
792 |
@functools.lru_cache
|
793 |
+
def get_moe_configs(
|
794 |
+
E: int,
|
795 |
+
N: int,
|
796 |
+
dtype: Optional[str],
|
797 |
+
block_n: Optional[int] = None,
|
798 |
+
block_k: Optional[int] = None,
|
799 |
+
) -> Optional[Dict[int, Any]]:
|
800 |
"""
|
801 |
Return optimized configurations for the fused MoE kernel.
|
802 |
|
|
|
808 |
|
809 |
# First look up if an optimized configuration is available in the configs
|
810 |
# directory
|
811 |
+
block_shape = [block_n, block_k] if block_n and block_k else None
|
812 |
+
json_file_name = get_config_file_name(E, N, dtype, block_shape)
|
813 |
|
814 |
config_file_path = os.path.join(
|
815 |
os.path.dirname(os.path.realpath(__file__)), "configs", json_file_name
|
816 |
)
|
817 |
if os.path.exists(config_file_path):
|
818 |
with open(config_file_path) as f:
|
819 |
+
logger.info("Using configuration from %s for MoE layer.", config_file_path)
|
820 |
# If a configuration has been found, return it
|
821 |
return {int(key): val for key, val in json.load(f).items()}
|
822 |
|
823 |
# If no optimized configuration is available, we will use the default
|
824 |
# configuration
|
825 |
+
logger.warning(
|
826 |
+
(
|
827 |
+
"Using default MoE config. Performance might be sub-optimal! "
|
828 |
+
"Config file not found at %s"
|
829 |
+
),
|
830 |
+
config_file_path,
|
831 |
+
)
|
832 |
return None
|
833 |
|
834 |
|
|
|
840 |
topk: int,
|
841 |
dtype: Optional[str],
|
842 |
is_marlin: bool,
|
843 |
+
block_shape: Optional[List[int]] = None,
|
844 |
) -> Dict[str, int]:
|
845 |
+
if dtype == "fp8_w8a8" and block_shape is not None:
|
846 |
+
# Block-wise quant: BLOCK_SIZE_N must be divisible by block_shape[0]
|
847 |
+
# BLOCK_SIZE_K must be divisible by block_shape[1]
|
|
|
|
|
|
|
|
|
|
|
848 |
config = {
|
849 |
+
"BLOCK_SIZE_M": 64,
|
850 |
+
"BLOCK_SIZE_N": block_shape[0],
|
851 |
+
"BLOCK_SIZE_K": block_shape[1],
|
852 |
+
"GROUP_SIZE_M": 32,
|
853 |
+
"num_warps": 4,
|
854 |
+
"num_stages": 3,
|
855 |
}
|
856 |
+
else:
|
857 |
+
config = {
|
858 |
+
"BLOCK_SIZE_M": 64,
|
859 |
+
"BLOCK_SIZE_N": 64,
|
860 |
+
"BLOCK_SIZE_K": 32,
|
861 |
+
"GROUP_SIZE_M": 8,
|
862 |
+
}
|
863 |
+
# A heuristic: fused marlin works faster with this config for small M
|
864 |
+
if M <= E or (is_marlin and M <= 32):
|
865 |
+
config = {
|
866 |
+
"BLOCK_SIZE_M": 16,
|
867 |
+
"BLOCK_SIZE_N": 32,
|
868 |
+
"BLOCK_SIZE_K": 64,
|
869 |
+
"GROUP_SIZE_M": 1,
|
870 |
+
}
|
871 |
return config
|
872 |
|
873 |
|
|
|
877 |
top_k: int,
|
878 |
dtype: Optional[str],
|
879 |
M: int,
|
|
|
880 |
is_marlin: bool = False,
|
881 |
+
block_shape: Optional[List[int]] = None,
|
882 |
):
|
883 |
+
# from vllm.model_executor.layers.fused_moe import get_config
|
884 |
+
# TODO: removed when syncing to vLLM, do we need this?
|
885 |
+
# override_config = get_config()
|
886 |
+
override_config = None
|
887 |
if override_config:
|
888 |
config = override_config
|
889 |
else:
|
890 |
# First try to load optimal config from the file
|
891 |
E, _, N = w2_shape
|
892 |
+
block_n = block_shape[0] if block_shape else 0
|
893 |
+
block_k = block_shape[1] if block_shape else 0
|
894 |
+
configs = get_moe_configs(E, N, dtype, block_n, block_k)
|
895 |
|
896 |
if configs:
|
897 |
# If an optimal configuration map has been found, look up the
|
|
|
899 |
config = configs[min(configs.keys(), key=lambda x: abs(x - M))]
|
900 |
else:
|
901 |
# Else use the default config
|
902 |
+
config = get_default_config(
|
903 |
+
M, E, N, w1_shape[2], top_k, dtype, is_marlin, block_shape
|
904 |
+
)
|
905 |
return config
|
906 |
|
907 |
|
|
|
937 |
return topk_weights, topk_ids
|
938 |
|
939 |
|
940 |
+
# This is used by the Deepseek-V2 and Deepseek-V3 model
|
941 |
+
@torch.compile(dynamic=True, backend=current_platform.simple_compile_backend)
|
942 |
def grouped_topk(
|
943 |
hidden_states: torch.Tensor,
|
944 |
gating_output: torch.Tensor,
|
|
|
946 |
renormalize: bool,
|
947 |
num_expert_group: int = 0,
|
948 |
topk_group: int = 0,
|
949 |
+
scoring_func: str = "softmax",
|
950 |
+
e_score_correction_bias: Optional[torch.Tensor] = None,
|
951 |
):
|
952 |
|
953 |
assert hidden_states.shape[0] == gating_output.shape[0], "Number of tokens mismatch"
|
954 |
|
955 |
+
if scoring_func == "softmax":
|
956 |
+
scores = torch.softmax(gating_output, dim=-1)
|
957 |
+
elif scoring_func == "sigmoid":
|
958 |
+
scores = gating_output.sigmoid()
|
959 |
+
else:
|
960 |
+
raise ValueError(f"Unsupported scoring function: {scoring_func}")
|
961 |
+
|
962 |
+
if e_score_correction_bias is not None:
|
963 |
+
# Store original scores before applying correction bias. We use biased
|
964 |
+
# scores for expert selection but original scores for routing weights
|
965 |
+
original_scores = scores
|
966 |
+
scores = scores + e_score_correction_bias.unsqueeze(0)
|
967 |
+
|
968 |
num_token = scores.shape[0]
|
969 |
group_scores = (
|
970 |
scores.view(num_token, num_expert_group, -1).max(dim=-1).values
|
|
|
980 |
.reshape(num_token, -1)
|
981 |
) # [n, e]
|
982 |
tmp_scores = scores.masked_fill(~score_mask.bool(), 0.0) # [n, e]
|
983 |
+
|
984 |
+
if e_score_correction_bias is not None:
|
985 |
+
topk_ids = torch.topk(tmp_scores, k=topk, dim=-1, sorted=False)[1]
|
986 |
+
# Use original unbiased scores for the routing weights
|
987 |
+
topk_weights = original_scores.gather(1, topk_ids)
|
988 |
+
else:
|
989 |
+
topk_weights, topk_ids = torch.topk(tmp_scores, k=topk, dim=-1, sorted=False)
|
990 |
|
991 |
if renormalize:
|
992 |
topk_weights = topk_weights / topk_weights.sum(dim=-1, keepdim=True)
|
|
|
996 |
|
997 |
def get_config_dtype_str(
|
998 |
dtype: torch.dtype,
|
999 |
+
use_int4_w4a16: Optional[bool] = False,
|
1000 |
use_int8_w8a16: Optional[bool] = False,
|
1001 |
use_fp8_w8a8: Optional[bool] = False,
|
1002 |
):
|
|
|
1004 |
return "fp8_w8a8"
|
1005 |
elif use_int8_w8a16:
|
1006 |
return "int8_w8a16"
|
1007 |
+
elif use_int4_w4a16:
|
1008 |
+
return "int4_w8a16"
|
1009 |
elif dtype == torch.float:
|
1010 |
# avoiding cases where kernel fails when float32 MoE
|
1011 |
# use fp16/bfloat16 configs
|
|
|
1013 |
return None
|
1014 |
|
1015 |
|
1016 |
+
def inplace_fused_experts(
|
1017 |
+
hidden_states: torch.Tensor,
|
1018 |
+
w1: torch.Tensor,
|
1019 |
+
w2: torch.Tensor,
|
1020 |
+
topk_weights: torch.Tensor,
|
1021 |
+
topk_ids: torch.Tensor,
|
1022 |
+
use_fp8_w8a8: bool = False,
|
1023 |
+
use_int8_w8a16: bool = False,
|
1024 |
+
use_int4_w4a16: bool = False,
|
1025 |
+
w1_scale: Optional[torch.Tensor] = None,
|
1026 |
+
w2_scale: Optional[torch.Tensor] = None,
|
1027 |
+
w1_zp: Optional[torch.Tensor] = None,
|
1028 |
+
w2_zp: Optional[torch.Tensor] = None,
|
1029 |
+
a1_scale: Optional[torch.Tensor] = None,
|
1030 |
+
a2_scale: Optional[torch.Tensor] = None,
|
1031 |
+
block_shape: Optional[List[int]] = None,
|
1032 |
+
) -> None:
|
1033 |
+
fused_experts_impl(
|
1034 |
+
hidden_states,
|
1035 |
+
w1,
|
1036 |
+
w2,
|
1037 |
+
topk_weights,
|
1038 |
+
topk_ids,
|
1039 |
+
True,
|
1040 |
+
use_fp8_w8a8,
|
1041 |
+
use_int8_w8a16,
|
1042 |
+
use_int4_w4a16,
|
1043 |
+
w1_scale,
|
1044 |
+
w2_scale,
|
1045 |
+
w1_zp,
|
1046 |
+
w2_zp,
|
1047 |
+
a1_scale,
|
1048 |
+
a2_scale,
|
1049 |
+
block_shape,
|
1050 |
+
)
|
1051 |
+
|
1052 |
+
|
1053 |
+
def outplace_fused_experts(
|
1054 |
+
hidden_states: torch.Tensor,
|
1055 |
+
w1: torch.Tensor,
|
1056 |
+
w2: torch.Tensor,
|
1057 |
+
topk_weights: torch.Tensor,
|
1058 |
+
topk_ids: torch.Tensor,
|
1059 |
+
use_fp8_w8a8: bool = False,
|
1060 |
+
use_int8_w8a16: bool = False,
|
1061 |
+
use_int4_w4a16: bool = False,
|
1062 |
+
w1_scale: Optional[torch.Tensor] = None,
|
1063 |
+
w2_scale: Optional[torch.Tensor] = None,
|
1064 |
+
w1_zp: Optional[torch.Tensor] = None,
|
1065 |
+
w2_zp: Optional[torch.Tensor] = None,
|
1066 |
+
a1_scale: Optional[torch.Tensor] = None,
|
1067 |
+
a2_scale: Optional[torch.Tensor] = None,
|
1068 |
+
block_shape: Optional[List[int]] = None,
|
1069 |
+
) -> torch.Tensor:
|
1070 |
+
return fused_experts_impl(
|
1071 |
+
hidden_states,
|
1072 |
+
w1,
|
1073 |
+
w2,
|
1074 |
+
topk_weights,
|
1075 |
+
topk_ids,
|
1076 |
+
False,
|
1077 |
+
use_fp8_w8a8,
|
1078 |
+
use_int8_w8a16,
|
1079 |
+
use_int4_w4a16,
|
1080 |
+
w1_scale,
|
1081 |
+
w2_scale,
|
1082 |
+
w1_zp,
|
1083 |
+
w2_zp,
|
1084 |
+
a1_scale,
|
1085 |
+
a2_scale,
|
1086 |
+
block_shape,
|
1087 |
+
)
|
1088 |
+
|
1089 |
+
|
1090 |
def fused_experts(
|
1091 |
hidden_states: torch.Tensor,
|
1092 |
w1: torch.Tensor,
|
|
|
1094 |
topk_weights: torch.Tensor,
|
1095 |
topk_ids: torch.Tensor,
|
1096 |
inplace: bool = False,
|
|
|
1097 |
use_fp8_w8a8: bool = False,
|
1098 |
use_int8_w8a16: bool = False,
|
1099 |
+
use_int4_w4a16: bool = False,
|
1100 |
+
w1_scale: Optional[torch.Tensor] = None,
|
1101 |
+
w2_scale: Optional[torch.Tensor] = None,
|
1102 |
+
w1_zp: Optional[torch.Tensor] = None,
|
1103 |
+
w2_zp: Optional[torch.Tensor] = None,
|
1104 |
+
a1_scale: Optional[torch.Tensor] = None,
|
1105 |
+
a2_scale: Optional[torch.Tensor] = None,
|
1106 |
+
block_shape: Optional[List[int]] = None,
|
1107 |
+
):
|
1108 |
+
if inplace:
|
1109 |
+
inplace_fused_experts(
|
1110 |
+
hidden_states,
|
1111 |
+
w1,
|
1112 |
+
w2,
|
1113 |
+
topk_weights,
|
1114 |
+
topk_ids,
|
1115 |
+
use_fp8_w8a8,
|
1116 |
+
use_int8_w8a16,
|
1117 |
+
use_int4_w4a16,
|
1118 |
+
w1_scale,
|
1119 |
+
w2_scale,
|
1120 |
+
w1_zp,
|
1121 |
+
w2_zp,
|
1122 |
+
a1_scale,
|
1123 |
+
a2_scale,
|
1124 |
+
block_shape,
|
1125 |
+
)
|
1126 |
+
return hidden_states
|
1127 |
+
else:
|
1128 |
+
return outplace_fused_experts(
|
1129 |
+
hidden_states,
|
1130 |
+
w1,
|
1131 |
+
w2,
|
1132 |
+
topk_weights,
|
1133 |
+
topk_ids,
|
1134 |
+
use_fp8_w8a8,
|
1135 |
+
use_int8_w8a16,
|
1136 |
+
use_int4_w4a16,
|
1137 |
+
w1_scale,
|
1138 |
+
w2_scale,
|
1139 |
+
w1_zp,
|
1140 |
+
w2_zp,
|
1141 |
+
a1_scale,
|
1142 |
+
a2_scale,
|
1143 |
+
block_shape,
|
1144 |
+
)
|
1145 |
+
|
1146 |
+
|
1147 |
+
def fused_experts_impl(
|
1148 |
+
hidden_states: torch.Tensor,
|
1149 |
+
w1: torch.Tensor,
|
1150 |
+
w2: torch.Tensor,
|
1151 |
+
topk_weights: torch.Tensor,
|
1152 |
+
topk_ids: torch.Tensor,
|
1153 |
+
inplace: bool = False,
|
1154 |
+
use_fp8_w8a8: bool = False,
|
1155 |
+
use_int8_w8a16: bool = False,
|
1156 |
+
use_int4_w4a16: bool = False,
|
1157 |
w1_scale: Optional[torch.Tensor] = None,
|
1158 |
w2_scale: Optional[torch.Tensor] = None,
|
1159 |
+
w1_zp: Optional[torch.Tensor] = None,
|
1160 |
+
w2_zp: Optional[torch.Tensor] = None,
|
1161 |
a1_scale: Optional[torch.Tensor] = None,
|
1162 |
a2_scale: Optional[torch.Tensor] = None,
|
1163 |
+
block_shape: Optional[List[int]] = None,
|
1164 |
):
|
1165 |
# Check constraints.
|
1166 |
+
if use_int4_w4a16:
|
1167 |
+
assert hidden_states.shape[1] // 2 == w1.shape[2], "Hidden size mismatch"
|
1168 |
+
else:
|
1169 |
+
assert hidden_states.shape[1] == w1.shape[2], "Hidden size mismatch"
|
1170 |
+
|
1171 |
assert topk_weights.shape == topk_ids.shape, "topk shape mismatch"
|
1172 |
assert hidden_states.is_contiguous(), "Hidden_states must be contiguous"
|
1173 |
assert w1.is_contiguous(), "Expert weights1 must be contiguous"
|
|
|
1183 |
config_dtype = get_config_dtype_str(
|
1184 |
use_fp8_w8a8=use_fp8_w8a8,
|
1185 |
use_int8_w8a16=use_int8_w8a16,
|
1186 |
+
use_int4_w4a16=use_int4_w4a16,
|
1187 |
dtype=hidden_states.dtype,
|
1188 |
)
|
1189 |
|
|
|
1193 |
w2.shape,
|
1194 |
topk_ids.shape[1],
|
1195 |
config_dtype,
|
1196 |
+
block_shape=block_shape,
|
1197 |
)
|
1198 |
|
1199 |
config = get_config_func(M)
|
|
|
1214 |
dtype=hidden_states.dtype,
|
1215 |
)
|
1216 |
|
1217 |
+
if hidden_states.dtype == torch.bfloat16:
|
1218 |
+
compute_type = tl.bfloat16
|
1219 |
+
elif hidden_states.dtype == torch.float16:
|
1220 |
+
compute_type = tl.float16
|
1221 |
+
elif hidden_states.dtype == torch.float32:
|
1222 |
+
compute_type = tl.float32
|
1223 |
+
else:
|
1224 |
+
raise ValueError(f"Unsupported compute_type: {hidden_states.dtype}")
|
1225 |
|
1226 |
if inplace:
|
1227 |
out_hidden_states = hidden_states
|
|
|
1262 |
intermediate_cache1,
|
1263 |
a1_scale,
|
1264 |
w1_scale,
|
1265 |
+
w1_zp,
|
1266 |
curr_topk_weights,
|
1267 |
curr_topk_ids,
|
1268 |
sorted_token_ids,
|
|
|
1274 |
compute_type=compute_type,
|
1275 |
use_fp8_w8a8=use_fp8_w8a8,
|
1276 |
use_int8_w8a16=use_int8_w8a16,
|
1277 |
+
use_int4_w4a16=use_int4_w4a16,
|
1278 |
+
block_shape=block_shape,
|
1279 |
)
|
1280 |
|
1281 |
ops.silu_and_mul(intermediate_cache2, intermediate_cache1.view(-1, N))
|
|
|
1286 |
intermediate_cache3,
|
1287 |
a2_scale,
|
1288 |
w2_scale,
|
1289 |
+
w2_zp,
|
1290 |
curr_topk_weights,
|
1291 |
curr_topk_ids,
|
1292 |
sorted_token_ids,
|
|
|
1298 |
compute_type=compute_type,
|
1299 |
use_fp8_w8a8=use_fp8_w8a8,
|
1300 |
use_int8_w8a16=use_int8_w8a16,
|
1301 |
+
use_int4_w4a16=use_int4_w4a16,
|
1302 |
+
block_shape=block_shape,
|
1303 |
)
|
1304 |
|
1305 |
ops.moe_sum(
|
|
|
1317 |
topk: int,
|
1318 |
renormalize: bool,
|
1319 |
inplace: bool = False,
|
|
|
1320 |
use_grouped_topk: bool = False,
|
1321 |
num_expert_group: Optional[int] = None,
|
1322 |
topk_group: Optional[int] = None,
|
1323 |
custom_routing_function: Optional[Callable] = None,
|
1324 |
use_fp8_w8a8: bool = False,
|
1325 |
use_int8_w8a16: bool = False,
|
1326 |
+
use_int4_w4a16: bool = False,
|
1327 |
w1_scale: Optional[torch.Tensor] = None,
|
1328 |
w2_scale: Optional[torch.Tensor] = None,
|
1329 |
+
w1_zp: Optional[torch.Tensor] = None,
|
1330 |
+
w2_zp: Optional[torch.Tensor] = None,
|
1331 |
a1_scale: Optional[torch.Tensor] = None,
|
1332 |
a2_scale: Optional[torch.Tensor] = None,
|
1333 |
+
block_shape: Optional[List[int]] = None,
|
1334 |
) -> torch.Tensor:
|
1335 |
"""
|
1336 |
This function computes a Mixture of Experts (MoE) layer using two sets of
|
|
|
1346 |
- renormalize (bool): If True, renormalize the top-k weights to sum to 1.
|
1347 |
- inplace (bool): If True, perform the operation in-place.
|
1348 |
Defaults to False.
|
|
|
|
|
1349 |
- num_expert_group: Optional[int]: additional parameter for grouped_topk
|
1350 |
- topk_group: Optional[int]: additional parameter for grouped_topk
|
1351 |
- use_grouped_topk: If True, use grouped_topk instead of fused_topk
|
1352 |
note: Deepseekv2 model uses grouped_topk
|
1353 |
- use_fp8_w8a8 (bool): If True, use fp8 arithmetic to compute the inner
|
1354 |
products for w1 and w2. Defaults to False.
|
1355 |
+
- use_int8_w8a16 (bool): If True, use matmul of int8 weight and bf16/fp16
|
1356 |
+
activation to compute the inner products for w1 and w2.
|
1357 |
+
Defaults to False.
|
1358 |
+
- use_int4_w4a16 (bool): If True, use matmul of int4 weight and bf16/fp16
|
1359 |
+
activation to compute the inner products for w1 and w2.
|
1360 |
+
Defaults to False.
|
1361 |
- w1_scale (Optional[torch.Tensor]): Optional scale to be used for
|
1362 |
w1.
|
1363 |
- w2_scale (Optional[torch.Tensor]): Optional scale to be used for
|
1364 |
w2.
|
1365 |
+
- a1_scale (Optional[torch.Tensor]): Optional scale to be used for
|
1366 |
+
a1.
|
1367 |
+
- a2_scale (Optional[torch.Tensor]): Optional scale to be used for
|
1368 |
+
a2.
|
1369 |
+
- block_shape: (Optional[List[int]]): Optional block size for block-wise
|
1370 |
+
quantization.
|
1371 |
|
1372 |
Returns:
|
1373 |
- torch.Tensor: The output tensor after applying the MoE layer.
|
|
|
1401 |
topk_weights,
|
1402 |
topk_ids,
|
1403 |
inplace=inplace,
|
|
|
1404 |
use_fp8_w8a8=use_fp8_w8a8,
|
1405 |
use_int8_w8a16=use_int8_w8a16,
|
1406 |
+
use_int4_w4a16=use_int4_w4a16,
|
1407 |
w1_scale=w1_scale,
|
1408 |
w2_scale=w2_scale,
|
1409 |
+
w1_zp=w1_zp,
|
1410 |
+
w2_zp=w2_zp,
|
1411 |
a1_scale=a1_scale,
|
1412 |
a2_scale=a2_scale,
|
1413 |
+
block_shape=block_shape,
|
1414 |
)
|
build/torch25-cxx98-cu121-x86_64-linux/moe/platforms.py
CHANGED
@@ -1,22 +1,32 @@
|
|
1 |
-
from
|
2 |
-
import os
|
3 |
-
from functools import lru_cache, wraps
|
4 |
|
5 |
import torch
|
6 |
|
7 |
IS_ROCM = torch.version.hip is not None
|
8 |
|
9 |
-
|
|
|
|
|
|
|
|
|
|
|
10 |
@classmethod
|
11 |
@lru_cache(maxsize=8)
|
12 |
def get_device_name(cls, device_id: int = 0) -> str:
|
13 |
return torch.cuda.get_device_name(0)
|
14 |
|
15 |
-
|
|
|
|
|
|
|
|
|
16 |
@classmethod
|
17 |
@lru_cache(maxsize=8)
|
18 |
def get_device_name(cls, device_id: int = 0) -> str:
|
19 |
return torch.cuda.get_device_name(device_id)
|
20 |
|
|
|
|
|
|
|
21 |
|
22 |
current_platform = RocmPlatform() if IS_ROCM else CudaPlatform()
|
|
|
1 |
+
from functools import lru_cache
|
|
|
|
|
2 |
|
3 |
import torch
|
4 |
|
5 |
IS_ROCM = torch.version.hip is not None
|
6 |
|
7 |
+
|
8 |
+
class Platform:
|
9 |
+
simple_compile_backend: str = "inductor"
|
10 |
+
|
11 |
+
|
12 |
+
class CudaPlatform(Platform):
|
13 |
@classmethod
|
14 |
@lru_cache(maxsize=8)
|
15 |
def get_device_name(cls, device_id: int = 0) -> str:
|
16 |
return torch.cuda.get_device_name(0)
|
17 |
|
18 |
+
def is_rocm(self):
|
19 |
+
return False
|
20 |
+
|
21 |
+
|
22 |
+
class RocmPlatform(Platform):
|
23 |
@classmethod
|
24 |
@lru_cache(maxsize=8)
|
25 |
def get_device_name(cls, device_id: int = 0) -> str:
|
26 |
return torch.cuda.get_device_name(device_id)
|
27 |
|
28 |
+
def is_rocm(self):
|
29 |
+
return True
|
30 |
+
|
31 |
|
32 |
current_platform = RocmPlatform() if IS_ROCM else CudaPlatform()
|
build/torch25-cxx98-cu124-x86_64-linux/moe/_moe_b25pgchg5o5pa.abi3.so
DELETED
@@ -1,3 +0,0 @@
|
|
1 |
-
version https://git-lfs.github.com/spec/v1
|
2 |
-
oid sha256:1b5e6a3b584873f4b48185c810e8cc1045b000e45269f2490a2e2fc3a45e144b
|
3 |
-
size 84059584
|
|
|
|
|
|
|
|
build/torch25-cxx98-cu124-x86_64-linux/moe/_moe_phlujktdbqekw.abi3.so
ADDED
@@ -0,0 +1,3 @@
|
|
|
|
|
|
|
|
|
1 |
+
version https://git-lfs.github.com/spec/v1
|
2 |
+
oid sha256:7c3b1cc57c3f73b7c43aec3aa6c0673bc8e24827a0338ef8beeb431392e9ac3e
|
3 |
+
size 85733416
|
build/torch25-cxx98-cu124-x86_64-linux/moe/_ops.py
CHANGED
@@ -1,9 +1,9 @@
|
|
1 |
import torch
|
2 |
-
from . import
|
3 |
-
ops = torch.ops.
|
4 |
|
5 |
def add_op_namespace_prefix(op_name: str):
|
6 |
"""
|
7 |
Prefix op by namespace.
|
8 |
"""
|
9 |
-
return f"
|
|
|
1 |
import torch
|
2 |
+
from . import _moe_phlujktdbqekw
|
3 |
+
ops = torch.ops._moe_phlujktdbqekw
|
4 |
|
5 |
def add_op_namespace_prefix(op_name: str):
|
6 |
"""
|
7 |
Prefix op by namespace.
|
8 |
"""
|
9 |
+
return f"_moe_phlujktdbqekw::{op_name}"
|
build/torch25-cxx98-cu124-x86_64-linux/moe/fp8.py
CHANGED
@@ -1,6 +1,11 @@
|
|
|
|
|
|
1 |
import torch
|
|
|
|
|
2 |
|
3 |
-
|
|
|
4 |
|
5 |
|
6 |
def is_hip() -> bool:
|
@@ -49,15 +54,179 @@ def scaled_fp8_quant(
|
|
49 |
if scale is None:
|
50 |
if use_per_token_if_dynamic:
|
51 |
scale = torch.empty((shape[0], 1), device=input.device, dtype=torch.float32)
|
52 |
-
|
53 |
-
output, input, scale, scale_ub
|
54 |
-
)
|
55 |
else:
|
56 |
scale = torch.zeros(1, device=input.device, dtype=torch.float32)
|
57 |
-
|
58 |
else:
|
59 |
# num_token_padding not implemented for this case
|
60 |
assert scale.numel() == 1 or num_token_padding is None
|
61 |
-
|
62 |
|
63 |
return output, scale
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
1 |
+
from typing import Tuple, Optional, Union
|
2 |
+
|
3 |
import torch
|
4 |
+
import triton
|
5 |
+
import triton.language as tl
|
6 |
|
7 |
+
|
8 |
+
from ._ops import ops
|
9 |
|
10 |
|
11 |
def is_hip() -> bool:
|
|
|
54 |
if scale is None:
|
55 |
if use_per_token_if_dynamic:
|
56 |
scale = torch.empty((shape[0], 1), device=input.device, dtype=torch.float32)
|
57 |
+
ops.dynamic_per_token_scaled_fp8_quant(output, input, scale, scale_ub)
|
|
|
|
|
58 |
else:
|
59 |
scale = torch.zeros(1, device=input.device, dtype=torch.float32)
|
60 |
+
ops.dynamic_scaled_fp8_quant(output, input, scale)
|
61 |
else:
|
62 |
# num_token_padding not implemented for this case
|
63 |
assert scale.numel() == 1 or num_token_padding is None
|
64 |
+
ops.static_scaled_fp8_quant(output, input, scale)
|
65 |
|
66 |
return output, scale
|
67 |
+
|
68 |
+
|
69 |
+
@triton.jit
|
70 |
+
def _per_token_group_quant_fp8(
|
71 |
+
# Pointers to inputs and output
|
72 |
+
y_ptr,
|
73 |
+
y_q_ptr,
|
74 |
+
y_s_ptr,
|
75 |
+
group_size,
|
76 |
+
# Avoid to divide zero
|
77 |
+
eps,
|
78 |
+
# Information for float8
|
79 |
+
fp8_min,
|
80 |
+
fp8_max,
|
81 |
+
# Meta-parameters
|
82 |
+
BLOCK: tl.constexpr,
|
83 |
+
):
|
84 |
+
"""A Triton-accelerated function to perform per-token-group
|
85 |
+
quantization on a tensor.
|
86 |
+
This function converts the tensor values into float8 values.
|
87 |
+
"""
|
88 |
+
# Map the program id to the row of X and Y it should compute.
|
89 |
+
g_id = tl.program_id(0)
|
90 |
+
y_ptr += g_id * group_size
|
91 |
+
y_q_ptr += g_id * group_size
|
92 |
+
y_s_ptr += g_id
|
93 |
+
|
94 |
+
cols = tl.arange(0, BLOCK) # N <= BLOCK
|
95 |
+
mask = cols < group_size
|
96 |
+
|
97 |
+
y = tl.load(y_ptr + cols, mask=mask, other=0.0).to(tl.float32)
|
98 |
+
# Quant
|
99 |
+
_absmax = tl.maximum(tl.max(tl.abs(y)), eps)
|
100 |
+
y_s = _absmax / fp8_max
|
101 |
+
y_q = tl.clamp(y / y_s, fp8_min, fp8_max).to(y_q_ptr.dtype.element_ty)
|
102 |
+
|
103 |
+
tl.store(y_q_ptr + cols, y_q, mask=mask)
|
104 |
+
tl.store(y_s_ptr, y_s)
|
105 |
+
|
106 |
+
|
107 |
+
@triton.jit
|
108 |
+
def _per_token_group_quant_fp8_colmajor(
|
109 |
+
# Pointers to inputs and output
|
110 |
+
y_ptr,
|
111 |
+
y_q_ptr,
|
112 |
+
y_s_ptr,
|
113 |
+
group_size,
|
114 |
+
# Num columns of y
|
115 |
+
y_num_columns,
|
116 |
+
# Stride from one column to the next of y_s
|
117 |
+
y_s_col_stride,
|
118 |
+
# Avoid to divide zero
|
119 |
+
eps,
|
120 |
+
# Information for float8
|
121 |
+
fp8_min,
|
122 |
+
fp8_max,
|
123 |
+
# Meta-parameters
|
124 |
+
BLOCK: tl.constexpr,
|
125 |
+
):
|
126 |
+
"""A Triton-accelerated function to perform per-token-group
|
127 |
+
quantization on a tensor.
|
128 |
+
This function converts the tensor values into float8 values.
|
129 |
+
"""
|
130 |
+
# Map the program id to the row of X and Y it should compute.
|
131 |
+
g_id = tl.program_id(0)
|
132 |
+
y_ptr += g_id * group_size
|
133 |
+
y_q_ptr += g_id * group_size
|
134 |
+
|
135 |
+
# Convert g_id the flattened block coordinate to 2D so we can index
|
136 |
+
# into the output y_scales matrix
|
137 |
+
blocks_per_row = y_num_columns // group_size
|
138 |
+
scale_col = g_id % blocks_per_row
|
139 |
+
scale_row = g_id // blocks_per_row
|
140 |
+
y_s_ptr += scale_col * y_s_col_stride + scale_row
|
141 |
+
|
142 |
+
cols = tl.arange(0, BLOCK) # group_size <= BLOCK
|
143 |
+
mask = cols < group_size
|
144 |
+
|
145 |
+
y = tl.load(y_ptr + cols, mask=mask, other=0.0).to(tl.float32)
|
146 |
+
# Quant
|
147 |
+
_absmax = tl.maximum(tl.max(tl.abs(y)), eps)
|
148 |
+
y_s = _absmax / fp8_max
|
149 |
+
y_q = tl.clamp(y / y_s, fp8_min, fp8_max).to(y_q_ptr.dtype.element_ty)
|
150 |
+
|
151 |
+
tl.store(y_q_ptr + cols, y_q, mask=mask)
|
152 |
+
tl.store(y_s_ptr, y_s)
|
153 |
+
|
154 |
+
|
155 |
+
def per_token_group_quant_fp8(
|
156 |
+
x: torch.Tensor,
|
157 |
+
group_size: int,
|
158 |
+
eps: float = 1e-10,
|
159 |
+
dtype: Optional[torch.dtype] = None,
|
160 |
+
column_major_scales: bool = False,
|
161 |
+
) -> Tuple[torch.Tensor, torch.Tensor]:
|
162 |
+
"""Function to perform per-token-group quantization on an input tensor `x`.
|
163 |
+
It converts the tensor values into signed float8 values and returns the
|
164 |
+
quantized tensor along with the scaling factor used for quantization.
|
165 |
+
Args:
|
166 |
+
x: The input tensor with ndim >= 2.
|
167 |
+
group_size: The group size used for quantization.
|
168 |
+
eps: The minimum to avoid dividing zero.
|
169 |
+
dtype: The dype of output tensor. Note that only `torch.float8_e4m3fn`
|
170 |
+
is supported for now.
|
171 |
+
Returns:
|
172 |
+
Tuple[torch.Tensor, torch.Tensor]: The quantized tensor and the
|
173 |
+
scaling factor for quantization.
|
174 |
+
"""
|
175 |
+
if dtype is None:
|
176 |
+
dtype = (
|
177 |
+
torch.float8_e4m3fnuz if current_platform.is_rocm() else torch.float8_e4m3fn
|
178 |
+
)
|
179 |
+
assert x.shape[-1] % group_size == 0, (
|
180 |
+
f"the last dimension of `x` {x.shape[-1]} must be divisible "
|
181 |
+
f"by `group_size` {group_size}"
|
182 |
+
)
|
183 |
+
assert x.is_contiguous(), "`x` must be contiguous"
|
184 |
+
|
185 |
+
finfo = torch.finfo(dtype)
|
186 |
+
fp8_min = finfo.min
|
187 |
+
fp8_max = finfo.max
|
188 |
+
|
189 |
+
x_q = torch.empty_like(x, device=x.device, dtype=dtype)
|
190 |
+
M = x.numel() // group_size
|
191 |
+
N = group_size
|
192 |
+
if column_major_scales:
|
193 |
+
shape = (x.shape[-1] // group_size,) + x.shape[:-1]
|
194 |
+
x_s = torch.empty(shape, device=x.device, dtype=torch.float32).permute(-1, -2)
|
195 |
+
else:
|
196 |
+
shape = x.shape[:-1] + (x.shape[-1] // group_size,)
|
197 |
+
x_s = torch.empty(shape, device=x.device, dtype=torch.float32)
|
198 |
+
|
199 |
+
BLOCK = triton.next_power_of_2(N)
|
200 |
+
# heuristics for number of warps
|
201 |
+
num_warps = min(max(BLOCK // 256, 1), 8)
|
202 |
+
num_stages = 1
|
203 |
+
if column_major_scales:
|
204 |
+
_per_token_group_quant_fp8_colmajor[(M,)](
|
205 |
+
x,
|
206 |
+
x_q,
|
207 |
+
x_s,
|
208 |
+
group_size,
|
209 |
+
x.shape[1],
|
210 |
+
x_s.stride(1),
|
211 |
+
eps,
|
212 |
+
fp8_min=fp8_min,
|
213 |
+
fp8_max=fp8_max,
|
214 |
+
BLOCK=BLOCK,
|
215 |
+
num_warps=num_warps,
|
216 |
+
num_stages=num_stages,
|
217 |
+
)
|
218 |
+
else:
|
219 |
+
_per_token_group_quant_fp8[(M,)](
|
220 |
+
x,
|
221 |
+
x_q,
|
222 |
+
x_s,
|
223 |
+
group_size,
|
224 |
+
eps,
|
225 |
+
fp8_min=fp8_min,
|
226 |
+
fp8_max=fp8_max,
|
227 |
+
BLOCK=BLOCK,
|
228 |
+
num_warps=num_warps,
|
229 |
+
num_stages=num_stages,
|
230 |
+
)
|
231 |
+
|
232 |
+
return x_q, x_s
|
build/torch25-cxx98-cu124-x86_64-linux/moe/fused_marlin_moe.py
CHANGED
@@ -40,7 +40,6 @@ def single_marlin_moe(
|
|
40 |
g_idx: Optional[torch.Tensor] = None,
|
41 |
sort_indices: Optional[torch.Tensor] = None,
|
42 |
w_zeros: Optional[torch.Tensor] = None,
|
43 |
-
override_config: Optional[Dict[str, Any]] = None,
|
44 |
num_bits: int = 8,
|
45 |
is_k_full: bool = True,
|
46 |
) -> torch.Tensor:
|
@@ -61,8 +60,6 @@ def single_marlin_moe(
|
|
61 |
- topk (int): The number of top-k experts to select.
|
62 |
- renormalize (bool): If True, renormalize the top-k weights to sum to 1.
|
63 |
- w_zeros (Optional[torch.Tensor]): Optional zero points to be used for w.
|
64 |
-
- override_config (Optional[Dict[str, Any]]): Optional override
|
65 |
-
for the kernel configuration.
|
66 |
- num_bits (bool): The number of bits in expert weights quantization.
|
67 |
|
68 |
Returns:
|
@@ -90,7 +87,6 @@ def single_marlin_moe(
|
|
90 |
w.shape,
|
91 |
topk_ids.shape[1],
|
92 |
None,
|
93 |
-
override_config=override_config,
|
94 |
is_marlin=True,
|
95 |
)
|
96 |
config = get_config_func(M)
|
@@ -154,6 +150,25 @@ def single_marlin_moe(
|
|
154 |
return torch.sum(intermediate_cache.view(*intermediate_cache.shape), dim=1)
|
155 |
|
156 |
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
157 |
def fused_marlin_moe(
|
158 |
hidden_states: torch.Tensor,
|
159 |
w1: torch.Tensor,
|
@@ -169,7 +184,6 @@ def fused_marlin_moe(
|
|
169 |
sort_indices2: Optional[torch.Tensor] = None,
|
170 |
w1_zeros: Optional[torch.Tensor] = None,
|
171 |
w2_zeros: Optional[torch.Tensor] = None,
|
172 |
-
override_config: Optional[Dict[str, Any]] = None,
|
173 |
num_bits: int = 8,
|
174 |
is_k_full: bool = True,
|
175 |
) -> torch.Tensor:
|
@@ -193,8 +207,6 @@ def fused_marlin_moe(
|
|
193 |
permutation.
|
194 |
- topk_weights (torch.Tensor): Top-k weights.
|
195 |
- topk_ids (torch.Tensor): Indices of topk-k elements.
|
196 |
-
- override_config (Optional[Dict[str, Any]]): Optional override
|
197 |
-
for the kernel configuration.
|
198 |
- w1_zeros (Optional[torch.Tensor]): Optional zero points to be used for w1.
|
199 |
- w2_zeros (Optional[torch.Tensor]): Optional zero points to be used for w2.
|
200 |
- num_bits (bool): The number of bits in expert weights quantization.
|
@@ -248,7 +260,6 @@ def fused_marlin_moe(
|
|
248 |
w2.shape,
|
249 |
topk_ids.shape[1],
|
250 |
None,
|
251 |
-
override_config=override_config,
|
252 |
is_marlin=True,
|
253 |
)
|
254 |
config = get_config_func(M)
|
@@ -350,6 +361,30 @@ def fused_marlin_moe(
|
|
350 |
return torch.sum(intermediate_cache3.view(*intermediate_cache3.shape), dim=1)
|
351 |
|
352 |
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
353 |
if hasattr(ops, "marlin_gemm_moe"):
|
354 |
|
355 |
@register_fake(add_op_namespace_prefix("marlin_gemm_moe"))
|
|
|
40 |
g_idx: Optional[torch.Tensor] = None,
|
41 |
sort_indices: Optional[torch.Tensor] = None,
|
42 |
w_zeros: Optional[torch.Tensor] = None,
|
|
|
43 |
num_bits: int = 8,
|
44 |
is_k_full: bool = True,
|
45 |
) -> torch.Tensor:
|
|
|
60 |
- topk (int): The number of top-k experts to select.
|
61 |
- renormalize (bool): If True, renormalize the top-k weights to sum to 1.
|
62 |
- w_zeros (Optional[torch.Tensor]): Optional zero points to be used for w.
|
|
|
|
|
63 |
- num_bits (bool): The number of bits in expert weights quantization.
|
64 |
|
65 |
Returns:
|
|
|
87 |
w.shape,
|
88 |
topk_ids.shape[1],
|
89 |
None,
|
|
|
90 |
is_marlin=True,
|
91 |
)
|
92 |
config = get_config_func(M)
|
|
|
150 |
return torch.sum(intermediate_cache.view(*intermediate_cache.shape), dim=1)
|
151 |
|
152 |
|
153 |
+
if hasattr(ops, "single_marlin_gemm_moe"):
|
154 |
+
|
155 |
+
@register_fake(add_op_namespace_prefix("single_marlin_gemm_moe"))
|
156 |
+
def single_marlin_moe_fake(
|
157 |
+
hidden_states: torch.Tensor,
|
158 |
+
w: torch.Tensor,
|
159 |
+
scales: torch.Tensor,
|
160 |
+
gating_output: torch.Tensor,
|
161 |
+
topk: int,
|
162 |
+
renormalize: bool,
|
163 |
+
g_idx: Optional[torch.Tensor] = None,
|
164 |
+
sort_indices: Optional[torch.Tensor] = None,
|
165 |
+
w_zeros: Optional[torch.Tensor] = None,
|
166 |
+
num_bits: int = 8,
|
167 |
+
is_k_full: bool = True,
|
168 |
+
) -> torch.Tensor:
|
169 |
+
return torch.empty_like(hidden_states)
|
170 |
+
|
171 |
+
|
172 |
def fused_marlin_moe(
|
173 |
hidden_states: torch.Tensor,
|
174 |
w1: torch.Tensor,
|
|
|
184 |
sort_indices2: Optional[torch.Tensor] = None,
|
185 |
w1_zeros: Optional[torch.Tensor] = None,
|
186 |
w2_zeros: Optional[torch.Tensor] = None,
|
|
|
187 |
num_bits: int = 8,
|
188 |
is_k_full: bool = True,
|
189 |
) -> torch.Tensor:
|
|
|
207 |
permutation.
|
208 |
- topk_weights (torch.Tensor): Top-k weights.
|
209 |
- topk_ids (torch.Tensor): Indices of topk-k elements.
|
|
|
|
|
210 |
- w1_zeros (Optional[torch.Tensor]): Optional zero points to be used for w1.
|
211 |
- w2_zeros (Optional[torch.Tensor]): Optional zero points to be used for w2.
|
212 |
- num_bits (bool): The number of bits in expert weights quantization.
|
|
|
260 |
w2.shape,
|
261 |
topk_ids.shape[1],
|
262 |
None,
|
|
|
263 |
is_marlin=True,
|
264 |
)
|
265 |
config = get_config_func(M)
|
|
|
361 |
return torch.sum(intermediate_cache3.view(*intermediate_cache3.shape), dim=1)
|
362 |
|
363 |
|
364 |
+
if hasattr(ops, "fused_marlin_moe"):
|
365 |
+
|
366 |
+
@register_fake(add_op_namespace_prefix("fused_marlin_moe"))
|
367 |
+
def fused_marlin_moe_fake(
|
368 |
+
hidden_states: torch.Tensor,
|
369 |
+
w1: torch.Tensor,
|
370 |
+
w2: torch.Tensor,
|
371 |
+
w1_scale: torch.Tensor,
|
372 |
+
w2_scale: torch.Tensor,
|
373 |
+
gating_output: torch.Tensor,
|
374 |
+
topk_weights: torch.Tensor,
|
375 |
+
topk_ids: torch.Tensor,
|
376 |
+
g_idx1: Optional[torch.Tensor] = None,
|
377 |
+
g_idx2: Optional[torch.Tensor] = None,
|
378 |
+
sort_indices1: Optional[torch.Tensor] = None,
|
379 |
+
sort_indices2: Optional[torch.Tensor] = None,
|
380 |
+
w1_zeros: Optional[torch.Tensor] = None,
|
381 |
+
w2_zeros: Optional[torch.Tensor] = None,
|
382 |
+
num_bits: int = 8,
|
383 |
+
is_k_full: bool = True,
|
384 |
+
) -> torch.Tensor:
|
385 |
+
return torch.empty_like(hidden_states)
|
386 |
+
|
387 |
+
|
388 |
if hasattr(ops, "marlin_gemm_moe"):
|
389 |
|
390 |
@register_fake(add_op_namespace_prefix("marlin_gemm_moe"))
|
build/torch25-cxx98-cu124-x86_64-linux/moe/fused_moe.py
CHANGED
@@ -1,21 +1,242 @@
|
|
|
|
1 |
"""Fused MoE kernel."""
|
2 |
|
3 |
import functools
|
4 |
import json
|
|
|
5 |
import os
|
6 |
-
from typing import Any, Callable, Dict, Optional, Tuple
|
7 |
|
8 |
import torch
|
9 |
import triton
|
10 |
import triton.language as tl
|
11 |
|
|
|
12 |
from ._ops import ops
|
13 |
-
from .fp8 import scaled_fp8_quant
|
14 |
from .platforms import current_platform
|
15 |
|
|
|
|
|
|
|
16 |
VLLM_FUSED_MOE_CHUNK_SIZE = int(os.getenv("VLLM_FUSED_MOE_CHUNK_SIZE", "32768"))
|
17 |
|
18 |
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
19 |
@triton.jit
|
20 |
def fused_moe_kernel(
|
21 |
# Pointers to matrices
|
@@ -44,8 +265,14 @@ def fused_moe_kernel(
|
|
44 |
stride_bn,
|
45 |
stride_cm,
|
46 |
stride_cn,
|
|
|
|
|
47 |
stride_bse,
|
|
|
48 |
stride_bsn,
|
|
|
|
|
|
|
49 |
# Meta-parameters
|
50 |
BLOCK_SIZE_M: tl.constexpr,
|
51 |
BLOCK_SIZE_N: tl.constexpr,
|
@@ -105,17 +332,17 @@ def fused_moe_kernel(
|
|
105 |
num_tokens_post_padded = tl.load(num_tokens_post_padded_ptr)
|
106 |
if pid_m * BLOCK_SIZE_M >= num_tokens_post_padded:
|
107 |
return
|
108 |
-
offs_token_id = pid_m * BLOCK_SIZE_M + tl.arange(0, BLOCK_SIZE_M)
|
109 |
offs_token = tl.load(sorted_token_ids_ptr + offs_token_id)
|
110 |
token_mask = offs_token < num_valid_tokens
|
111 |
|
112 |
-
offs_bn = (pid_n * BLOCK_SIZE_N + tl.arange(0, BLOCK_SIZE_N)) % N
|
113 |
offs_k = tl.arange(0, BLOCK_SIZE_K)
|
114 |
a_ptrs = a_ptr + (
|
115 |
offs_token[:, None] // top_k * stride_am + offs_k[None, :] * stride_ak
|
116 |
)
|
117 |
|
118 |
-
off_experts = tl.load(expert_ids_ptr + pid_m)
|
119 |
b_ptrs = (
|
120 |
b_ptr
|
121 |
+ off_experts * stride_be
|
@@ -128,8 +355,15 @@ def fused_moe_kernel(
|
|
128 |
b_scale = tl.load(b_scale_ptrs)
|
129 |
|
130 |
if use_fp8_w8a8:
|
131 |
-
|
132 |
-
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
133 |
|
134 |
# -----------------------------------------------------------
|
135 |
# Iterate to compute a block of the C matrix.
|
@@ -151,7 +385,17 @@ def fused_moe_kernel(
|
|
151 |
if use_int8_w8a16:
|
152 |
accumulator = tl.dot(a, b.to(compute_type), acc=accumulator)
|
153 |
elif use_fp8_w8a8:
|
154 |
-
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
155 |
else:
|
156 |
accumulator += tl.dot(a, b)
|
157 |
# Advance the ptrs to the next K block.
|
@@ -164,7 +408,10 @@ def fused_moe_kernel(
|
|
164 |
if use_int8_w8a16:
|
165 |
accumulator = (accumulator * b_scale).to(compute_type)
|
166 |
elif use_fp8_w8a8:
|
167 |
-
|
|
|
|
|
|
|
168 |
else:
|
169 |
accumulator = accumulator.to(compute_type)
|
170 |
# -----------------------------------------------------------
|
@@ -175,6 +422,141 @@ def fused_moe_kernel(
|
|
175 |
tl.store(c_ptrs, accumulator, mask=c_mask)
|
176 |
|
177 |
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
178 |
def moe_align_block_size(
|
179 |
topk_ids: torch.Tensor, block_size: int, num_experts: int
|
180 |
) -> Tuple[torch.Tensor, torch.Tensor, torch.Tensor]:
|
@@ -225,9 +607,34 @@ def moe_align_block_size(
|
|
225 |
(max_num_m_blocks,), dtype=torch.int32, device=topk_ids.device
|
226 |
)
|
227 |
num_tokens_post_pad = torch.empty((1), dtype=torch.int32, device=topk_ids.device)
|
228 |
-
|
229 |
-
|
230 |
-
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
231 |
return sorted_ids, expert_ids, num_tokens_post_pad
|
232 |
|
233 |
|
@@ -237,6 +644,7 @@ def invoke_fused_moe_kernel(
|
|
237 |
C: torch.Tensor,
|
238 |
A_scale: Optional[torch.Tensor],
|
239 |
B_scale: Optional[torch.Tensor],
|
|
|
240 |
topk_weights: torch.Tensor,
|
241 |
topk_ids: torch.Tensor,
|
242 |
sorted_token_ids: torch.Tensor,
|
@@ -248,64 +656,147 @@ def invoke_fused_moe_kernel(
|
|
248 |
compute_type: tl.dtype,
|
249 |
use_fp8_w8a8: bool,
|
250 |
use_int8_w8a16: bool,
|
|
|
|
|
251 |
) -> None:
|
252 |
assert topk_weights.stride(1) == 1
|
253 |
assert sorted_token_ids.stride(0) == 1
|
254 |
|
255 |
if use_fp8_w8a8:
|
256 |
-
A, A_scale = scaled_fp8_quant(A, A_scale)
|
257 |
assert B_scale is not None
|
258 |
-
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
259 |
assert B_scale is not None
|
|
|
260 |
else:
|
261 |
assert A_scale is None
|
262 |
assert B_scale is None
|
263 |
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
264 |
grid = lambda META: (
|
265 |
-
triton.cdiv(
|
266 |
* triton.cdiv(B.shape[1], META["BLOCK_SIZE_N"]),
|
267 |
)
|
268 |
|
269 |
-
|
270 |
-
|
271 |
-
|
272 |
-
|
273 |
-
|
274 |
-
B_scale
|
275 |
-
|
276 |
-
|
277 |
-
|
278 |
-
|
279 |
-
|
280 |
-
|
281 |
-
|
282 |
-
|
283 |
-
|
284 |
-
|
285 |
-
|
286 |
-
|
287 |
-
|
288 |
-
|
289 |
-
|
290 |
-
|
291 |
-
|
292 |
-
|
293 |
-
|
294 |
-
|
295 |
-
|
296 |
-
|
297 |
-
|
298 |
-
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
299 |
|
300 |
|
301 |
-
|
|
|
|
|
|
|
302 |
device_name = current_platform.get_device_name().replace(" ", "_")
|
303 |
dtype_selector = "" if not dtype else f",dtype={dtype}"
|
304 |
-
|
|
|
|
|
|
|
305 |
|
306 |
|
|
|
307 |
@functools.lru_cache
|
308 |
-
def get_moe_configs(
|
|
|
|
|
|
|
|
|
|
|
|
|
309 |
"""
|
310 |
Return optimized configurations for the fused MoE kernel.
|
311 |
|
@@ -317,18 +808,27 @@ def get_moe_configs(E: int, N: int, dtype: Optional[str]) -> Optional[Dict[int,
|
|
317 |
|
318 |
# First look up if an optimized configuration is available in the configs
|
319 |
# directory
|
320 |
-
|
|
|
321 |
|
322 |
config_file_path = os.path.join(
|
323 |
os.path.dirname(os.path.realpath(__file__)), "configs", json_file_name
|
324 |
)
|
325 |
if os.path.exists(config_file_path):
|
326 |
with open(config_file_path) as f:
|
|
|
327 |
# If a configuration has been found, return it
|
328 |
return {int(key): val for key, val in json.load(f).items()}
|
329 |
|
330 |
# If no optimized configuration is available, we will use the default
|
331 |
# configuration
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
332 |
return None
|
333 |
|
334 |
|
@@ -340,21 +840,34 @@ def get_default_config(
|
|
340 |
topk: int,
|
341 |
dtype: Optional[str],
|
342 |
is_marlin: bool,
|
|
|
343 |
) -> Dict[str, int]:
|
344 |
-
|
345 |
-
|
346 |
-
|
347 |
-
"BLOCK_SIZE_K": 32,
|
348 |
-
"GROUP_SIZE_M": 8,
|
349 |
-
}
|
350 |
-
# A heuristic: fused marlin works faster with this config for small M
|
351 |
-
if M <= E or (is_marlin and M <= 32):
|
352 |
config = {
|
353 |
-
"BLOCK_SIZE_M":
|
354 |
-
"BLOCK_SIZE_N":
|
355 |
-
"BLOCK_SIZE_K":
|
356 |
-
"GROUP_SIZE_M":
|
|
|
|
|
357 |
}
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
358 |
return config
|
359 |
|
360 |
|
@@ -364,15 +877,21 @@ def try_get_optimal_moe_config(
|
|
364 |
top_k: int,
|
365 |
dtype: Optional[str],
|
366 |
M: int,
|
367 |
-
override_config: Optional[Dict[str, Any]] = None,
|
368 |
is_marlin: bool = False,
|
|
|
369 |
):
|
|
|
|
|
|
|
|
|
370 |
if override_config:
|
371 |
config = override_config
|
372 |
else:
|
373 |
# First try to load optimal config from the file
|
374 |
E, _, N = w2_shape
|
375 |
-
|
|
|
|
|
376 |
|
377 |
if configs:
|
378 |
# If an optimal configuration map has been found, look up the
|
@@ -380,7 +899,9 @@ def try_get_optimal_moe_config(
|
|
380 |
config = configs[min(configs.keys(), key=lambda x: abs(x - M))]
|
381 |
else:
|
382 |
# Else use the default config
|
383 |
-
config = get_default_config(
|
|
|
|
|
384 |
return config
|
385 |
|
386 |
|
@@ -416,7 +937,8 @@ def fused_topk(
|
|
416 |
return topk_weights, topk_ids
|
417 |
|
418 |
|
419 |
-
# This is used by the Deepseek-V2 model
|
|
|
420 |
def grouped_topk(
|
421 |
hidden_states: torch.Tensor,
|
422 |
gating_output: torch.Tensor,
|
@@ -424,11 +946,25 @@ def grouped_topk(
|
|
424 |
renormalize: bool,
|
425 |
num_expert_group: int = 0,
|
426 |
topk_group: int = 0,
|
|
|
|
|
427 |
):
|
428 |
|
429 |
assert hidden_states.shape[0] == gating_output.shape[0], "Number of tokens mismatch"
|
430 |
|
431 |
-
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
432 |
num_token = scores.shape[0]
|
433 |
group_scores = (
|
434 |
scores.view(num_token, num_expert_group, -1).max(dim=-1).values
|
@@ -444,7 +980,13 @@ def grouped_topk(
|
|
444 |
.reshape(num_token, -1)
|
445 |
) # [n, e]
|
446 |
tmp_scores = scores.masked_fill(~score_mask.bool(), 0.0) # [n, e]
|
447 |
-
|
|
|
|
|
|
|
|
|
|
|
|
|
448 |
|
449 |
if renormalize:
|
450 |
topk_weights = topk_weights / topk_weights.sum(dim=-1, keepdim=True)
|
@@ -454,6 +996,7 @@ def grouped_topk(
|
|
454 |
|
455 |
def get_config_dtype_str(
|
456 |
dtype: torch.dtype,
|
|
|
457 |
use_int8_w8a16: Optional[bool] = False,
|
458 |
use_fp8_w8a8: Optional[bool] = False,
|
459 |
):
|
@@ -461,6 +1004,8 @@ def get_config_dtype_str(
|
|
461 |
return "fp8_w8a8"
|
462 |
elif use_int8_w8a16:
|
463 |
return "int8_w8a16"
|
|
|
|
|
464 |
elif dtype == torch.float:
|
465 |
# avoiding cases where kernel fails when float32 MoE
|
466 |
# use fp16/bfloat16 configs
|
@@ -468,6 +1013,80 @@ def get_config_dtype_str(
|
|
468 |
return None
|
469 |
|
470 |
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
471 |
def fused_experts(
|
472 |
hidden_states: torch.Tensor,
|
473 |
w1: torch.Tensor,
|
@@ -475,16 +1094,80 @@ def fused_experts(
|
|
475 |
topk_weights: torch.Tensor,
|
476 |
topk_ids: torch.Tensor,
|
477 |
inplace: bool = False,
|
478 |
-
override_config: Optional[Dict[str, Any]] = None,
|
479 |
use_fp8_w8a8: bool = False,
|
480 |
use_int8_w8a16: bool = False,
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
481 |
w1_scale: Optional[torch.Tensor] = None,
|
482 |
w2_scale: Optional[torch.Tensor] = None,
|
|
|
|
|
483 |
a1_scale: Optional[torch.Tensor] = None,
|
484 |
a2_scale: Optional[torch.Tensor] = None,
|
|
|
485 |
):
|
486 |
# Check constraints.
|
487 |
-
|
|
|
|
|
|
|
|
|
488 |
assert topk_weights.shape == topk_ids.shape, "topk shape mismatch"
|
489 |
assert hidden_states.is_contiguous(), "Hidden_states must be contiguous"
|
490 |
assert w1.is_contiguous(), "Expert weights1 must be contiguous"
|
@@ -500,6 +1183,7 @@ def fused_experts(
|
|
500 |
config_dtype = get_config_dtype_str(
|
501 |
use_fp8_w8a8=use_fp8_w8a8,
|
502 |
use_int8_w8a16=use_int8_w8a16,
|
|
|
503 |
dtype=hidden_states.dtype,
|
504 |
)
|
505 |
|
@@ -509,7 +1193,7 @@ def fused_experts(
|
|
509 |
w2.shape,
|
510 |
topk_ids.shape[1],
|
511 |
config_dtype,
|
512 |
-
|
513 |
)
|
514 |
|
515 |
config = get_config_func(M)
|
@@ -530,7 +1214,14 @@ def fused_experts(
|
|
530 |
dtype=hidden_states.dtype,
|
531 |
)
|
532 |
|
533 |
-
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
534 |
|
535 |
if inplace:
|
536 |
out_hidden_states = hidden_states
|
@@ -571,6 +1262,7 @@ def fused_experts(
|
|
571 |
intermediate_cache1,
|
572 |
a1_scale,
|
573 |
w1_scale,
|
|
|
574 |
curr_topk_weights,
|
575 |
curr_topk_ids,
|
576 |
sorted_token_ids,
|
@@ -582,6 +1274,8 @@ def fused_experts(
|
|
582 |
compute_type=compute_type,
|
583 |
use_fp8_w8a8=use_fp8_w8a8,
|
584 |
use_int8_w8a16=use_int8_w8a16,
|
|
|
|
|
585 |
)
|
586 |
|
587 |
ops.silu_and_mul(intermediate_cache2, intermediate_cache1.view(-1, N))
|
@@ -592,6 +1286,7 @@ def fused_experts(
|
|
592 |
intermediate_cache3,
|
593 |
a2_scale,
|
594 |
w2_scale,
|
|
|
595 |
curr_topk_weights,
|
596 |
curr_topk_ids,
|
597 |
sorted_token_ids,
|
@@ -603,6 +1298,8 @@ def fused_experts(
|
|
603 |
compute_type=compute_type,
|
604 |
use_fp8_w8a8=use_fp8_w8a8,
|
605 |
use_int8_w8a16=use_int8_w8a16,
|
|
|
|
|
606 |
)
|
607 |
|
608 |
ops.moe_sum(
|
@@ -620,17 +1317,20 @@ def fused_moe(
|
|
620 |
topk: int,
|
621 |
renormalize: bool,
|
622 |
inplace: bool = False,
|
623 |
-
override_config: Optional[Dict[str, Any]] = None,
|
624 |
use_grouped_topk: bool = False,
|
625 |
num_expert_group: Optional[int] = None,
|
626 |
topk_group: Optional[int] = None,
|
627 |
custom_routing_function: Optional[Callable] = None,
|
628 |
use_fp8_w8a8: bool = False,
|
629 |
use_int8_w8a16: bool = False,
|
|
|
630 |
w1_scale: Optional[torch.Tensor] = None,
|
631 |
w2_scale: Optional[torch.Tensor] = None,
|
|
|
|
|
632 |
a1_scale: Optional[torch.Tensor] = None,
|
633 |
a2_scale: Optional[torch.Tensor] = None,
|
|
|
634 |
) -> torch.Tensor:
|
635 |
"""
|
636 |
This function computes a Mixture of Experts (MoE) layer using two sets of
|
@@ -646,20 +1346,28 @@ def fused_moe(
|
|
646 |
- renormalize (bool): If True, renormalize the top-k weights to sum to 1.
|
647 |
- inplace (bool): If True, perform the operation in-place.
|
648 |
Defaults to False.
|
649 |
-
- override_config (Optional[Dict[str, Any]]): Optional override
|
650 |
-
for the kernel configuration.
|
651 |
- num_expert_group: Optional[int]: additional parameter for grouped_topk
|
652 |
- topk_group: Optional[int]: additional parameter for grouped_topk
|
653 |
- use_grouped_topk: If True, use grouped_topk instead of fused_topk
|
654 |
note: Deepseekv2 model uses grouped_topk
|
655 |
- use_fp8_w8a8 (bool): If True, use fp8 arithmetic to compute the inner
|
656 |
products for w1 and w2. Defaults to False.
|
657 |
-
- use_int8_w8a16 (bool): If True, use
|
658 |
-
products for w1 and w2.
|
|
|
|
|
|
|
|
|
659 |
- w1_scale (Optional[torch.Tensor]): Optional scale to be used for
|
660 |
w1.
|
661 |
- w2_scale (Optional[torch.Tensor]): Optional scale to be used for
|
662 |
w2.
|
|
|
|
|
|
|
|
|
|
|
|
|
663 |
|
664 |
Returns:
|
665 |
- torch.Tensor: The output tensor after applying the MoE layer.
|
@@ -693,11 +1401,14 @@ def fused_moe(
|
|
693 |
topk_weights,
|
694 |
topk_ids,
|
695 |
inplace=inplace,
|
696 |
-
override_config=override_config,
|
697 |
use_fp8_w8a8=use_fp8_w8a8,
|
698 |
use_int8_w8a16=use_int8_w8a16,
|
|
|
699 |
w1_scale=w1_scale,
|
700 |
w2_scale=w2_scale,
|
|
|
|
|
701 |
a1_scale=a1_scale,
|
702 |
a2_scale=a2_scale,
|
|
|
703 |
)
|
|
|
1 |
+
# SPDX-License-Identifier: Apache-2.0
|
2 |
"""Fused MoE kernel."""
|
3 |
|
4 |
import functools
|
5 |
import json
|
6 |
+
import logging
|
7 |
import os
|
8 |
+
from typing import Any, Callable, Dict, List, Optional, Tuple
|
9 |
|
10 |
import torch
|
11 |
import triton
|
12 |
import triton.language as tl
|
13 |
|
14 |
+
|
15 |
from ._ops import ops
|
16 |
+
from .fp8 import per_token_group_quant_fp8, scaled_fp8_quant
|
17 |
from .platforms import current_platform
|
18 |
|
19 |
+
logger = logging.getLogger(__name__)
|
20 |
+
|
21 |
+
|
22 |
VLLM_FUSED_MOE_CHUNK_SIZE = int(os.getenv("VLLM_FUSED_MOE_CHUNK_SIZE", "32768"))
|
23 |
|
24 |
|
25 |
+
@triton.jit
|
26 |
+
def fused_moe_kernel_gptq_awq(
|
27 |
+
# Pointers to matrices
|
28 |
+
a_ptr,
|
29 |
+
b_ptr,
|
30 |
+
c_ptr,
|
31 |
+
b_scale_ptr,
|
32 |
+
b_zp_ptr,
|
33 |
+
topk_weights_ptr,
|
34 |
+
sorted_token_ids_ptr,
|
35 |
+
expert_ids_ptr,
|
36 |
+
num_tokens_post_padded_ptr,
|
37 |
+
# Matrix dimensions
|
38 |
+
N: tl.constexpr,
|
39 |
+
K: tl.constexpr,
|
40 |
+
EM,
|
41 |
+
num_valid_tokens,
|
42 |
+
# The stride variables represent how much to increase the ptr by when
|
43 |
+
# moving by 1 element in a particular dimension. E.g. `stride_am` is
|
44 |
+
# how much to increase `a_ptr` by to get the element one row down
|
45 |
+
# (A has M rows).
|
46 |
+
stride_am,
|
47 |
+
stride_ak,
|
48 |
+
stride_be,
|
49 |
+
stride_bk,
|
50 |
+
stride_bn,
|
51 |
+
stride_cm,
|
52 |
+
stride_cn,
|
53 |
+
stride_bse,
|
54 |
+
stride_bsk,
|
55 |
+
stride_bsn,
|
56 |
+
stride_bze,
|
57 |
+
stride_bzk,
|
58 |
+
stride_bzn,
|
59 |
+
block_k_diviable: tl.constexpr,
|
60 |
+
group_size: tl.constexpr,
|
61 |
+
# Meta-parameters
|
62 |
+
BLOCK_SIZE_M: tl.constexpr,
|
63 |
+
BLOCK_SIZE_N: tl.constexpr,
|
64 |
+
BLOCK_SIZE_K: tl.constexpr,
|
65 |
+
GROUP_SIZE_M: tl.constexpr,
|
66 |
+
MUL_ROUTED_WEIGHT: tl.constexpr,
|
67 |
+
top_k: tl.constexpr,
|
68 |
+
compute_type: tl.constexpr,
|
69 |
+
has_zp: tl.constexpr,
|
70 |
+
use_int4_w4a16: tl.constexpr,
|
71 |
+
use_int8_w8a16: tl.constexpr,
|
72 |
+
):
|
73 |
+
"""
|
74 |
+
Implements the fused computation for a Mixture of Experts (MOE) using
|
75 |
+
token and expert matrices.
|
76 |
+
|
77 |
+
Key Parameters:
|
78 |
+
- A: The input tensor representing tokens with shape (*, K), where '*' can
|
79 |
+
be any shape representing batches and K is the feature dimension of
|
80 |
+
each token.
|
81 |
+
- B: The stacked MOE weight tensor with shape (E, N, K), where E is
|
82 |
+
the number of experts, K is the input feature dimension, and N is
|
83 |
+
the output feature dimension.
|
84 |
+
- C: The output cache tensor with shape (M, topk, N), where M is the
|
85 |
+
total number of tokens post padding, topk is the number of times
|
86 |
+
each token is repeated, and N is the output feature dimension.
|
87 |
+
- sorted_token_ids: A tensor containing the sorted indices of tokens,
|
88 |
+
repeated topk times and arranged by the expert index they are
|
89 |
+
assigned to.
|
90 |
+
- expert_ids: A tensor containing the indices of the expert for each
|
91 |
+
block. It determines which expert matrix from B should be used for
|
92 |
+
each block in A.
|
93 |
+
This kernel performs the multiplication of a token by its corresponding
|
94 |
+
expert matrix as determined by `expert_ids`. The sorting of
|
95 |
+
`sorted_token_ids` by expert index and padding ensures divisibility by
|
96 |
+
BLOCK_SIZE_M, which is necessary to maintain consistency in block matrix
|
97 |
+
multiplication across different blocks processed by the same expert.
|
98 |
+
"""
|
99 |
+
# -----------------------------------------------------------
|
100 |
+
# Map program ids `pid` to the block of C it should compute.
|
101 |
+
# This is done in a grouped ordering to promote L2 data reuse.
|
102 |
+
pid = tl.program_id(axis=0)
|
103 |
+
num_pid_m = tl.cdiv(EM, BLOCK_SIZE_M)
|
104 |
+
num_pid_n = tl.cdiv(N, BLOCK_SIZE_N)
|
105 |
+
num_pid_in_group = GROUP_SIZE_M * num_pid_n
|
106 |
+
group_id = pid // num_pid_in_group
|
107 |
+
first_pid_m = group_id * GROUP_SIZE_M
|
108 |
+
group_size_m = min(num_pid_m - first_pid_m, GROUP_SIZE_M)
|
109 |
+
pid_m = first_pid_m + ((pid % num_pid_in_group) % group_size_m)
|
110 |
+
pid_n = (pid % num_pid_in_group) // group_size_m
|
111 |
+
|
112 |
+
# ----------------------------------------------------------
|
113 |
+
# Create pointers for the first blocks of A and B.
|
114 |
+
# We will advance this pointer as we move in the K direction
|
115 |
+
# and accumulate
|
116 |
+
# `a_ptrs` is a block of [BLOCK_SIZE_M, BLOCK_SIZE_K] pointers
|
117 |
+
# `b_ptrs` is a block of [BLOCK_SIZE_K, BLOCK_SIZE_N] pointers
|
118 |
+
num_tokens_post_padded = tl.load(num_tokens_post_padded_ptr)
|
119 |
+
if pid_m * BLOCK_SIZE_M >= num_tokens_post_padded:
|
120 |
+
return
|
121 |
+
offs_token_id = pid_m * BLOCK_SIZE_M + tl.arange(0, BLOCK_SIZE_M).to(tl.int64)
|
122 |
+
offs_token = tl.load(sorted_token_ids_ptr + offs_token_id)
|
123 |
+
token_mask = offs_token < num_valid_tokens
|
124 |
+
|
125 |
+
offs_bn = (pid_n * BLOCK_SIZE_N + tl.arange(0, BLOCK_SIZE_N).to(tl.int64)) % N
|
126 |
+
offs_k = tl.arange(0, BLOCK_SIZE_K)
|
127 |
+
a_ptrs = a_ptr + (
|
128 |
+
offs_token[:, None] // top_k * stride_am + offs_k[None, :] * stride_ak
|
129 |
+
)
|
130 |
+
|
131 |
+
off_experts = tl.load(expert_ids_ptr + pid_m).to(tl.int64)
|
132 |
+
|
133 |
+
if use_int4_w4a16:
|
134 |
+
b_ptrs = (
|
135 |
+
b_ptr
|
136 |
+
+ off_experts * stride_be
|
137 |
+
+ (offs_k[:, None] // 2) * stride_bk
|
138 |
+
+ offs_bn[None, :] * stride_bn
|
139 |
+
)
|
140 |
+
b_shifter = (offs_k[:, None] % 2) * 4
|
141 |
+
elif use_int8_w8a16:
|
142 |
+
b_ptrs = (
|
143 |
+
b_ptr
|
144 |
+
+ off_experts * stride_be
|
145 |
+
+ offs_k[:, None] * stride_bk
|
146 |
+
+ offs_bn[None, :] * stride_bn
|
147 |
+
)
|
148 |
+
|
149 |
+
if not has_zp and use_int4_w4a16:
|
150 |
+
b_zp_num = 8
|
151 |
+
if not has_zp and use_int8_w8a16:
|
152 |
+
b_zp_num = 128
|
153 |
+
elif has_zp and use_int4_w4a16:
|
154 |
+
b_zp_shifter = (offs_bn[None, :] % 2) * 4
|
155 |
+
|
156 |
+
# -----------------------------------------------------------
|
157 |
+
# Iterate to compute a block of the C matrix.
|
158 |
+
# We accumulate into a `[BLOCK_SIZE_M, BLOCK_SIZE_N]` block
|
159 |
+
# of fp32 values for higher accuracy.
|
160 |
+
# `accumulator` will be converted back to fp16 after the loop.
|
161 |
+
accumulator = tl.zeros((BLOCK_SIZE_M, BLOCK_SIZE_N), dtype=tl.float32)
|
162 |
+
for k in range(0, tl.cdiv(K, BLOCK_SIZE_K)):
|
163 |
+
# Load the next block of A and B, generate a mask by checking the
|
164 |
+
# K dimension.
|
165 |
+
|
166 |
+
if not block_k_diviable:
|
167 |
+
k_mask = offs_k[:, None] < K - k * BLOCK_SIZE_K
|
168 |
+
k_other = 0.0
|
169 |
+
else:
|
170 |
+
k_mask = None
|
171 |
+
k_other = None
|
172 |
+
|
173 |
+
a = tl.load(
|
174 |
+
a_ptrs,
|
175 |
+
mask=token_mask[:, None] & (offs_k[None, :] < K - k * BLOCK_SIZE_K),
|
176 |
+
other=0.0,
|
177 |
+
)
|
178 |
+
b = tl.load(b_ptrs)
|
179 |
+
if use_int4_w4a16:
|
180 |
+
b = (b >> b_shifter) & 0xF
|
181 |
+
|
182 |
+
b_scale_ptrs = (
|
183 |
+
b_scale_ptr
|
184 |
+
+ off_experts * stride_bse
|
185 |
+
+ offs_bn[None, :] * stride_bsn
|
186 |
+
+ ((offs_k[:, None] + BLOCK_SIZE_K * k) // group_size) * stride_bsk
|
187 |
+
)
|
188 |
+
b_scale = tl.load(b_scale_ptrs, mask=k_mask, other=k_other)
|
189 |
+
b_scale = b_scale.to(tl.float32)
|
190 |
+
|
191 |
+
if has_zp and use_int4_w4a16:
|
192 |
+
offs_k_true = (offs_k[:, None] + BLOCK_SIZE_K * k) // group_size
|
193 |
+
b_zp_ptrs = (
|
194 |
+
b_zp_ptr
|
195 |
+
+ off_experts * stride_bze
|
196 |
+
+ (offs_bn[None, :] // 2) * stride_bzn
|
197 |
+
+ offs_k_true * stride_bzk
|
198 |
+
)
|
199 |
+
b_zp = tl.load(b_zp_ptrs, mask=k_mask, other=k_other)
|
200 |
+
b_zp = (b_zp >> b_zp_shifter) & 0xF
|
201 |
+
b_zp = b_zp.to(tl.float32)
|
202 |
+
elif has_zp and use_int8_w8a16:
|
203 |
+
offs_k_true = (offs_k[:, None] + BLOCK_SIZE_K * k) // group_size
|
204 |
+
b_zp_ptrs = (
|
205 |
+
b_zp_ptr
|
206 |
+
+ off_experts * stride_bze
|
207 |
+
+ offs_bn[None, :] * stride_bzn
|
208 |
+
+ offs_k_true * stride_bzk
|
209 |
+
)
|
210 |
+
b_zp = tl.load(b_zp_ptrs, mask=k_mask, other=k_other)
|
211 |
+
b_zp = b_zp.to(tl.float32)
|
212 |
+
|
213 |
+
# We accumulate along the K dimension.
|
214 |
+
if has_zp:
|
215 |
+
b = ((b.to(tl.float32) - b_zp) * b_scale).to(compute_type)
|
216 |
+
else:
|
217 |
+
b = ((b.to(tl.float32) - b_zp_num) * b_scale).to(compute_type)
|
218 |
+
accumulator = tl.dot(a, b, acc=accumulator)
|
219 |
+
|
220 |
+
# Advance the ptrs to the next K block.
|
221 |
+
a_ptrs += BLOCK_SIZE_K * stride_ak
|
222 |
+
if use_int4_w4a16:
|
223 |
+
b_ptrs += (BLOCK_SIZE_K // 2) * stride_bk
|
224 |
+
else:
|
225 |
+
b_ptrs += BLOCK_SIZE_K * stride_bk
|
226 |
+
|
227 |
+
if MUL_ROUTED_WEIGHT:
|
228 |
+
moe_weight = tl.load(topk_weights_ptr + offs_token, mask=token_mask, other=0)
|
229 |
+
accumulator = accumulator * moe_weight[:, None]
|
230 |
+
|
231 |
+
accumulator = accumulator.to(compute_type)
|
232 |
+
# -----------------------------------------------------------
|
233 |
+
# Write back the block of the output
|
234 |
+
offs_cn = pid_n * BLOCK_SIZE_N + tl.arange(0, BLOCK_SIZE_N)
|
235 |
+
c_ptrs = c_ptr + stride_cm * offs_token[:, None] + stride_cn * offs_cn[None, :]
|
236 |
+
c_mask = token_mask[:, None] & (offs_cn[None, :] < N)
|
237 |
+
tl.store(c_ptrs, accumulator, mask=c_mask)
|
238 |
+
|
239 |
+
|
240 |
@triton.jit
|
241 |
def fused_moe_kernel(
|
242 |
# Pointers to matrices
|
|
|
265 |
stride_bn,
|
266 |
stride_cm,
|
267 |
stride_cn,
|
268 |
+
stride_asm,
|
269 |
+
stride_ask,
|
270 |
stride_bse,
|
271 |
+
stride_bsk,
|
272 |
stride_bsn,
|
273 |
+
# Block size for block-wise quantization
|
274 |
+
group_n: tl.constexpr,
|
275 |
+
group_k: tl.constexpr,
|
276 |
# Meta-parameters
|
277 |
BLOCK_SIZE_M: tl.constexpr,
|
278 |
BLOCK_SIZE_N: tl.constexpr,
|
|
|
332 |
num_tokens_post_padded = tl.load(num_tokens_post_padded_ptr)
|
333 |
if pid_m * BLOCK_SIZE_M >= num_tokens_post_padded:
|
334 |
return
|
335 |
+
offs_token_id = pid_m * BLOCK_SIZE_M + tl.arange(0, BLOCK_SIZE_M).to(tl.int64)
|
336 |
offs_token = tl.load(sorted_token_ids_ptr + offs_token_id)
|
337 |
token_mask = offs_token < num_valid_tokens
|
338 |
|
339 |
+
offs_bn = (pid_n * BLOCK_SIZE_N + tl.arange(0, BLOCK_SIZE_N).to(tl.int64)) % N
|
340 |
offs_k = tl.arange(0, BLOCK_SIZE_K)
|
341 |
a_ptrs = a_ptr + (
|
342 |
offs_token[:, None] // top_k * stride_am + offs_k[None, :] * stride_ak
|
343 |
)
|
344 |
|
345 |
+
off_experts = tl.load(expert_ids_ptr + pid_m).to(tl.int64)
|
346 |
b_ptrs = (
|
347 |
b_ptr
|
348 |
+ off_experts * stride_be
|
|
|
355 |
b_scale = tl.load(b_scale_ptrs)
|
356 |
|
357 |
if use_fp8_w8a8:
|
358 |
+
if group_k > 0 and group_n > 0:
|
359 |
+
a_scale_ptrs = a_scale_ptr + (offs_token // top_k) * stride_asm
|
360 |
+
offs_bsn = offs_bn // group_n
|
361 |
+
b_scale_ptrs = (
|
362 |
+
b_scale_ptr + off_experts * stride_bse + offs_bsn * stride_bsn
|
363 |
+
)
|
364 |
+
else:
|
365 |
+
a_scale = tl.load(a_scale_ptr)
|
366 |
+
b_scale = tl.load(b_scale_ptr + off_experts)
|
367 |
|
368 |
# -----------------------------------------------------------
|
369 |
# Iterate to compute a block of the C matrix.
|
|
|
385 |
if use_int8_w8a16:
|
386 |
accumulator = tl.dot(a, b.to(compute_type), acc=accumulator)
|
387 |
elif use_fp8_w8a8:
|
388 |
+
if group_k > 0 and group_n > 0:
|
389 |
+
k_start = k * BLOCK_SIZE_K
|
390 |
+
offs_ks = k_start // group_k
|
391 |
+
a_scale = tl.load(
|
392 |
+
a_scale_ptrs + offs_ks * stride_ask, mask=token_mask, other=0.0
|
393 |
+
)
|
394 |
+
b_scale = tl.load(b_scale_ptrs + offs_ks * stride_bsk)
|
395 |
+
|
396 |
+
accumulator += tl.dot(a, b) * a_scale[:, None] * b_scale[None, :]
|
397 |
+
else:
|
398 |
+
accumulator = tl.dot(a, b, acc=accumulator)
|
399 |
else:
|
400 |
accumulator += tl.dot(a, b)
|
401 |
# Advance the ptrs to the next K block.
|
|
|
408 |
if use_int8_w8a16:
|
409 |
accumulator = (accumulator * b_scale).to(compute_type)
|
410 |
elif use_fp8_w8a8:
|
411 |
+
if group_k > 0 and group_n > 0:
|
412 |
+
accumulator = accumulator.to(compute_type)
|
413 |
+
else:
|
414 |
+
accumulator = (accumulator * a_scale * b_scale).to(compute_type)
|
415 |
else:
|
416 |
accumulator = accumulator.to(compute_type)
|
417 |
# -----------------------------------------------------------
|
|
|
422 |
tl.store(c_ptrs, accumulator, mask=c_mask)
|
423 |
|
424 |
|
425 |
+
def ceil_div(a, b):
|
426 |
+
return (a + b - 1) // b
|
427 |
+
|
428 |
+
|
429 |
+
@triton.jit
|
430 |
+
def moe_align_block_size_stage1(
|
431 |
+
topk_ids_ptr,
|
432 |
+
tokens_cnts_ptr,
|
433 |
+
num_experts: tl.constexpr,
|
434 |
+
numel: tl.constexpr,
|
435 |
+
tokens_per_thread: tl.constexpr,
|
436 |
+
):
|
437 |
+
pid = tl.program_id(0)
|
438 |
+
|
439 |
+
start_idx = pid * tokens_per_thread
|
440 |
+
|
441 |
+
off_c = (pid + 1) * num_experts
|
442 |
+
|
443 |
+
for i in range(tokens_per_thread):
|
444 |
+
if start_idx + i < numel:
|
445 |
+
idx = tl.load(topk_ids_ptr + start_idx + i)
|
446 |
+
token_cnt = tl.load(tokens_cnts_ptr + off_c + idx)
|
447 |
+
tl.store(tokens_cnts_ptr + off_c + idx, token_cnt + 1)
|
448 |
+
|
449 |
+
|
450 |
+
@triton.jit
|
451 |
+
def moe_align_block_size_stage2(
|
452 |
+
tokens_cnts_ptr,
|
453 |
+
num_experts: tl.constexpr,
|
454 |
+
):
|
455 |
+
pid = tl.program_id(0)
|
456 |
+
|
457 |
+
last_cnt = 0
|
458 |
+
for i in range(1, num_experts + 1):
|
459 |
+
token_cnt = tl.load(tokens_cnts_ptr + i * num_experts + pid)
|
460 |
+
last_cnt = last_cnt + token_cnt
|
461 |
+
tl.store(tokens_cnts_ptr + i * num_experts + pid, last_cnt)
|
462 |
+
|
463 |
+
|
464 |
+
@triton.jit
|
465 |
+
def moe_align_block_size_stage3(
|
466 |
+
total_tokens_post_pad_ptr,
|
467 |
+
tokens_cnts_ptr,
|
468 |
+
cumsum_ptr,
|
469 |
+
num_experts: tl.constexpr,
|
470 |
+
block_size: tl.constexpr,
|
471 |
+
):
|
472 |
+
last_cumsum = 0
|
473 |
+
off_cnt = num_experts * num_experts
|
474 |
+
for i in range(1, num_experts + 1):
|
475 |
+
token_cnt = tl.load(tokens_cnts_ptr + off_cnt + i - 1)
|
476 |
+
last_cumsum = last_cumsum + tl.cdiv(token_cnt, block_size) * block_size
|
477 |
+
tl.store(cumsum_ptr + i, last_cumsum)
|
478 |
+
tl.store(total_tokens_post_pad_ptr, last_cumsum)
|
479 |
+
|
480 |
+
|
481 |
+
@triton.jit
|
482 |
+
def moe_align_block_size_stage4(
|
483 |
+
topk_ids_ptr,
|
484 |
+
sorted_token_ids_ptr,
|
485 |
+
expert_ids_ptr,
|
486 |
+
tokens_cnts_ptr,
|
487 |
+
cumsum_ptr,
|
488 |
+
num_experts: tl.constexpr,
|
489 |
+
block_size: tl.constexpr,
|
490 |
+
numel: tl.constexpr,
|
491 |
+
tokens_per_thread: tl.constexpr,
|
492 |
+
):
|
493 |
+
pid = tl.program_id(0)
|
494 |
+
start_idx = tl.load(cumsum_ptr + pid)
|
495 |
+
end_idx = tl.load(cumsum_ptr + pid + 1)
|
496 |
+
|
497 |
+
for i in range(start_idx, end_idx, block_size):
|
498 |
+
tl.store(expert_ids_ptr + i // block_size, pid)
|
499 |
+
|
500 |
+
start_idx = pid * tokens_per_thread
|
501 |
+
off_t = pid * num_experts
|
502 |
+
|
503 |
+
for i in range(start_idx, tl.minimum(start_idx + tokens_per_thread, numel)):
|
504 |
+
expert_id = tl.load(topk_ids_ptr + i)
|
505 |
+
token_cnt = tl.load(tokens_cnts_ptr + off_t + expert_id)
|
506 |
+
rank_post_pad = token_cnt + tl.load(cumsum_ptr + expert_id)
|
507 |
+
tl.store(sorted_token_ids_ptr + rank_post_pad, i)
|
508 |
+
tl.store(tokens_cnts_ptr + off_t + expert_id, token_cnt + 1)
|
509 |
+
|
510 |
+
|
511 |
+
# Triton implementation based on:
|
512 |
+
# https://github.com/sgl-project/sglang/commit/ba5112ff691d791a9e38c6c71f59324a5fcb49d0
|
513 |
+
def moe_align_block_size_triton(
|
514 |
+
topk_ids: torch.Tensor,
|
515 |
+
num_experts: int,
|
516 |
+
block_size: int,
|
517 |
+
sorted_token_ids: torch.Tensor,
|
518 |
+
expert_ids: torch.Tensor,
|
519 |
+
num_tokens_post_pad: torch.Tensor,
|
520 |
+
) -> None:
|
521 |
+
numel = topk_ids.numel()
|
522 |
+
grid = (num_experts,)
|
523 |
+
tokens_cnts = torch.zeros(
|
524 |
+
(num_experts + 1, num_experts), dtype=torch.int32, device=topk_ids.device
|
525 |
+
)
|
526 |
+
cumsum = torch.zeros((num_experts + 1,), dtype=torch.int32, device=topk_ids.device)
|
527 |
+
tokens_per_thread = ceil_div(numel, num_experts)
|
528 |
+
|
529 |
+
moe_align_block_size_stage1[grid](
|
530 |
+
topk_ids,
|
531 |
+
tokens_cnts,
|
532 |
+
num_experts,
|
533 |
+
numel,
|
534 |
+
tokens_per_thread,
|
535 |
+
)
|
536 |
+
moe_align_block_size_stage2[grid](
|
537 |
+
tokens_cnts,
|
538 |
+
num_experts,
|
539 |
+
)
|
540 |
+
moe_align_block_size_stage3[(1,)](
|
541 |
+
num_tokens_post_pad,
|
542 |
+
tokens_cnts,
|
543 |
+
cumsum,
|
544 |
+
num_experts,
|
545 |
+
block_size,
|
546 |
+
)
|
547 |
+
moe_align_block_size_stage4[grid](
|
548 |
+
topk_ids,
|
549 |
+
sorted_token_ids,
|
550 |
+
expert_ids,
|
551 |
+
tokens_cnts,
|
552 |
+
cumsum,
|
553 |
+
num_experts,
|
554 |
+
block_size,
|
555 |
+
numel,
|
556 |
+
tokens_per_thread,
|
557 |
+
)
|
558 |
+
|
559 |
+
|
560 |
def moe_align_block_size(
|
561 |
topk_ids: torch.Tensor, block_size: int, num_experts: int
|
562 |
) -> Tuple[torch.Tensor, torch.Tensor, torch.Tensor]:
|
|
|
607 |
(max_num_m_blocks,), dtype=torch.int32, device=topk_ids.device
|
608 |
)
|
609 |
num_tokens_post_pad = torch.empty((1), dtype=torch.int32, device=topk_ids.device)
|
610 |
+
if num_experts >= 224:
|
611 |
+
if VLLM_ENABLE_MOE_ALIGN_BLOCK_SIZE_TRITON:
|
612 |
+
moe_align_block_size_triton(
|
613 |
+
topk_ids,
|
614 |
+
num_experts,
|
615 |
+
block_size,
|
616 |
+
sorted_ids,
|
617 |
+
expert_ids,
|
618 |
+
num_tokens_post_pad,
|
619 |
+
)
|
620 |
+
else:
|
621 |
+
ops.sgl_moe_align_block_size(
|
622 |
+
topk_ids,
|
623 |
+
num_experts,
|
624 |
+
block_size,
|
625 |
+
sorted_ids,
|
626 |
+
expert_ids,
|
627 |
+
num_tokens_post_pad,
|
628 |
+
)
|
629 |
+
else:
|
630 |
+
ops.moe_align_block_size(
|
631 |
+
topk_ids,
|
632 |
+
num_experts,
|
633 |
+
block_size,
|
634 |
+
sorted_ids,
|
635 |
+
expert_ids,
|
636 |
+
num_tokens_post_pad,
|
637 |
+
)
|
638 |
return sorted_ids, expert_ids, num_tokens_post_pad
|
639 |
|
640 |
|
|
|
644 |
C: torch.Tensor,
|
645 |
A_scale: Optional[torch.Tensor],
|
646 |
B_scale: Optional[torch.Tensor],
|
647 |
+
B_zp: Optional[torch.Tensor],
|
648 |
topk_weights: torch.Tensor,
|
649 |
topk_ids: torch.Tensor,
|
650 |
sorted_token_ids: torch.Tensor,
|
|
|
656 |
compute_type: tl.dtype,
|
657 |
use_fp8_w8a8: bool,
|
658 |
use_int8_w8a16: bool,
|
659 |
+
use_int4_w4a16: bool,
|
660 |
+
block_shape: Optional[List[int]] = None,
|
661 |
) -> None:
|
662 |
assert topk_weights.stride(1) == 1
|
663 |
assert sorted_token_ids.stride(0) == 1
|
664 |
|
665 |
if use_fp8_w8a8:
|
|
|
666 |
assert B_scale is not None
|
667 |
+
if block_shape is None:
|
668 |
+
A, A_scale = scaled_fp8_quant(A, A_scale)
|
669 |
+
else:
|
670 |
+
assert len(block_shape) == 2
|
671 |
+
block_n, block_k = block_shape[0], block_shape[1]
|
672 |
+
A, A_scale = per_token_group_quant_fp8(A, block_k)
|
673 |
+
assert triton.cdiv(A.shape[-1], block_k) == A_scale.shape[-1]
|
674 |
+
assert triton.cdiv(B.shape[-2], block_n) == B_scale.shape[-2]
|
675 |
+
assert triton.cdiv(B.shape[-1], block_k) == B_scale.shape[-1]
|
676 |
+
elif use_int8_w8a16 or use_int4_w4a16:
|
677 |
assert B_scale is not None
|
678 |
+
assert block_shape is None or block_shape[0] == 0
|
679 |
else:
|
680 |
assert A_scale is None
|
681 |
assert B_scale is None
|
682 |
|
683 |
+
EM = sorted_token_ids.shape[0]
|
684 |
+
if A.shape[0] < config["BLOCK_SIZE_M"]:
|
685 |
+
# optimize for small batch_size.
|
686 |
+
# We assume that top_ids of each token is unique, so
|
687 |
+
# so num_valid_experts <= batch_size <= BLOCK_SIZE_M,
|
688 |
+
# and we can skip some invalid blocks.
|
689 |
+
EM = min(sorted_token_ids.shape[0], A.shape[0] * top_k * config["BLOCK_SIZE_M"])
|
690 |
grid = lambda META: (
|
691 |
+
triton.cdiv(EM, META["BLOCK_SIZE_M"])
|
692 |
* triton.cdiv(B.shape[1], META["BLOCK_SIZE_N"]),
|
693 |
)
|
694 |
|
695 |
+
if (
|
696 |
+
(use_int8_w8a16 or use_int4_w4a16)
|
697 |
+
and block_shape is not None
|
698 |
+
and block_shape[1] > 0
|
699 |
+
):
|
700 |
+
assert B_scale is not None and B_scale.ndim == 3
|
701 |
+
assert B_zp is None or B_zp.ndim == 3
|
702 |
+
|
703 |
+
fused_moe_kernel_gptq_awq[grid](
|
704 |
+
A,
|
705 |
+
B,
|
706 |
+
C,
|
707 |
+
B_scale,
|
708 |
+
B_zp,
|
709 |
+
topk_weights,
|
710 |
+
sorted_token_ids,
|
711 |
+
expert_ids,
|
712 |
+
num_tokens_post_padded,
|
713 |
+
B.shape[1],
|
714 |
+
A.shape[1],
|
715 |
+
EM,
|
716 |
+
topk_ids.numel(),
|
717 |
+
A.stride(0),
|
718 |
+
A.stride(1),
|
719 |
+
B.stride(0),
|
720 |
+
B.stride(2),
|
721 |
+
B.stride(1),
|
722 |
+
C.stride(1),
|
723 |
+
C.stride(2),
|
724 |
+
B_scale.stride(0),
|
725 |
+
B_scale.stride(2),
|
726 |
+
B_scale.stride(1),
|
727 |
+
B_zp.stride(0) if B_zp is not None else 0,
|
728 |
+
B_zp.stride(2) if B_zp is not None else 0,
|
729 |
+
B_zp.stride(1) if B_zp is not None else 0,
|
730 |
+
block_k_diviable=A.shape[1] % config["BLOCK_SIZE_K"] == 0,
|
731 |
+
group_size=block_shape[1],
|
732 |
+
MUL_ROUTED_WEIGHT=mul_routed_weight,
|
733 |
+
top_k=top_k,
|
734 |
+
compute_type=compute_type,
|
735 |
+
has_zp=B_zp is not None,
|
736 |
+
use_int4_w4a16=use_int4_w4a16,
|
737 |
+
use_int8_w8a16=use_int8_w8a16,
|
738 |
+
**config,
|
739 |
+
)
|
740 |
+
|
741 |
+
else:
|
742 |
+
fused_moe_kernel[grid](
|
743 |
+
A,
|
744 |
+
B,
|
745 |
+
C,
|
746 |
+
A_scale,
|
747 |
+
B_scale,
|
748 |
+
topk_weights,
|
749 |
+
sorted_token_ids,
|
750 |
+
expert_ids,
|
751 |
+
num_tokens_post_padded,
|
752 |
+
B.shape[1],
|
753 |
+
A.shape[1],
|
754 |
+
EM,
|
755 |
+
topk_ids.numel(),
|
756 |
+
A.stride(0),
|
757 |
+
A.stride(1),
|
758 |
+
B.stride(0),
|
759 |
+
B.stride(2),
|
760 |
+
B.stride(1),
|
761 |
+
C.stride(1),
|
762 |
+
C.stride(2),
|
763 |
+
A_scale.stride(0) if A_scale is not None and A_scale.ndim == 2 else 0,
|
764 |
+
A_scale.stride(1) if A_scale is not None and A_scale.ndim == 2 else 0,
|
765 |
+
B_scale.stride(0) if B_scale is not None and B_scale.ndim >= 2 else 0,
|
766 |
+
B_scale.stride(2) if B_scale is not None and B_scale.ndim == 3 else 0,
|
767 |
+
B_scale.stride(1) if B_scale is not None and B_scale.ndim >= 2 else 0,
|
768 |
+
0 if block_shape is None else block_shape[0],
|
769 |
+
0 if block_shape is None else block_shape[1],
|
770 |
+
MUL_ROUTED_WEIGHT=mul_routed_weight,
|
771 |
+
top_k=top_k,
|
772 |
+
compute_type=compute_type,
|
773 |
+
use_fp8_w8a8=use_fp8_w8a8,
|
774 |
+
use_int8_w8a16=use_int8_w8a16,
|
775 |
+
**config,
|
776 |
+
)
|
777 |
|
778 |
|
779 |
+
# Adapted from: https://github.com/sgl-project/sglang/pull/2628
|
780 |
+
def get_config_file_name(
|
781 |
+
E: int, N: int, dtype: Optional[str], block_shape: Optional[List[int]] = None
|
782 |
+
) -> str:
|
783 |
device_name = current_platform.get_device_name().replace(" ", "_")
|
784 |
dtype_selector = "" if not dtype else f",dtype={dtype}"
|
785 |
+
block_shape_selector = (
|
786 |
+
"" if not block_shape or not all(block_shape) else f",block_shape={block_shape}"
|
787 |
+
)
|
788 |
+
return f"E={E},N={N},device_name={device_name}{dtype_selector}{block_shape_selector}.json" # noqa: E501
|
789 |
|
790 |
|
791 |
+
# Adapted from: https://github.com/sgl-project/sglang/pull/2628
|
792 |
@functools.lru_cache
|
793 |
+
def get_moe_configs(
|
794 |
+
E: int,
|
795 |
+
N: int,
|
796 |
+
dtype: Optional[str],
|
797 |
+
block_n: Optional[int] = None,
|
798 |
+
block_k: Optional[int] = None,
|
799 |
+
) -> Optional[Dict[int, Any]]:
|
800 |
"""
|
801 |
Return optimized configurations for the fused MoE kernel.
|
802 |
|
|
|
808 |
|
809 |
# First look up if an optimized configuration is available in the configs
|
810 |
# directory
|
811 |
+
block_shape = [block_n, block_k] if block_n and block_k else None
|
812 |
+
json_file_name = get_config_file_name(E, N, dtype, block_shape)
|
813 |
|
814 |
config_file_path = os.path.join(
|
815 |
os.path.dirname(os.path.realpath(__file__)), "configs", json_file_name
|
816 |
)
|
817 |
if os.path.exists(config_file_path):
|
818 |
with open(config_file_path) as f:
|
819 |
+
logger.info("Using configuration from %s for MoE layer.", config_file_path)
|
820 |
# If a configuration has been found, return it
|
821 |
return {int(key): val for key, val in json.load(f).items()}
|
822 |
|
823 |
# If no optimized configuration is available, we will use the default
|
824 |
# configuration
|
825 |
+
logger.warning(
|
826 |
+
(
|
827 |
+
"Using default MoE config. Performance might be sub-optimal! "
|
828 |
+
"Config file not found at %s"
|
829 |
+
),
|
830 |
+
config_file_path,
|
831 |
+
)
|
832 |
return None
|
833 |
|
834 |
|
|
|
840 |
topk: int,
|
841 |
dtype: Optional[str],
|
842 |
is_marlin: bool,
|
843 |
+
block_shape: Optional[List[int]] = None,
|
844 |
) -> Dict[str, int]:
|
845 |
+
if dtype == "fp8_w8a8" and block_shape is not None:
|
846 |
+
# Block-wise quant: BLOCK_SIZE_N must be divisible by block_shape[0]
|
847 |
+
# BLOCK_SIZE_K must be divisible by block_shape[1]
|
|
|
|
|
|
|
|
|
|
|
848 |
config = {
|
849 |
+
"BLOCK_SIZE_M": 64,
|
850 |
+
"BLOCK_SIZE_N": block_shape[0],
|
851 |
+
"BLOCK_SIZE_K": block_shape[1],
|
852 |
+
"GROUP_SIZE_M": 32,
|
853 |
+
"num_warps": 4,
|
854 |
+
"num_stages": 3,
|
855 |
}
|
856 |
+
else:
|
857 |
+
config = {
|
858 |
+
"BLOCK_SIZE_M": 64,
|
859 |
+
"BLOCK_SIZE_N": 64,
|
860 |
+
"BLOCK_SIZE_K": 32,
|
861 |
+
"GROUP_SIZE_M": 8,
|
862 |
+
}
|
863 |
+
# A heuristic: fused marlin works faster with this config for small M
|
864 |
+
if M <= E or (is_marlin and M <= 32):
|
865 |
+
config = {
|
866 |
+
"BLOCK_SIZE_M": 16,
|
867 |
+
"BLOCK_SIZE_N": 32,
|
868 |
+
"BLOCK_SIZE_K": 64,
|
869 |
+
"GROUP_SIZE_M": 1,
|
870 |
+
}
|
871 |
return config
|
872 |
|
873 |
|
|
|
877 |
top_k: int,
|
878 |
dtype: Optional[str],
|
879 |
M: int,
|
|
|
880 |
is_marlin: bool = False,
|
881 |
+
block_shape: Optional[List[int]] = None,
|
882 |
):
|
883 |
+
# from vllm.model_executor.layers.fused_moe import get_config
|
884 |
+
# TODO: removed when syncing to vLLM, do we need this?
|
885 |
+
# override_config = get_config()
|
886 |
+
override_config = None
|
887 |
if override_config:
|
888 |
config = override_config
|
889 |
else:
|
890 |
# First try to load optimal config from the file
|
891 |
E, _, N = w2_shape
|
892 |
+
block_n = block_shape[0] if block_shape else 0
|
893 |
+
block_k = block_shape[1] if block_shape else 0
|
894 |
+
configs = get_moe_configs(E, N, dtype, block_n, block_k)
|
895 |
|
896 |
if configs:
|
897 |
# If an optimal configuration map has been found, look up the
|
|
|
899 |
config = configs[min(configs.keys(), key=lambda x: abs(x - M))]
|
900 |
else:
|
901 |
# Else use the default config
|
902 |
+
config = get_default_config(
|
903 |
+
M, E, N, w1_shape[2], top_k, dtype, is_marlin, block_shape
|
904 |
+
)
|
905 |
return config
|
906 |
|
907 |
|
|
|
937 |
return topk_weights, topk_ids
|
938 |
|
939 |
|
940 |
+
# This is used by the Deepseek-V2 and Deepseek-V3 model
|
941 |
+
@torch.compile(dynamic=True, backend=current_platform.simple_compile_backend)
|
942 |
def grouped_topk(
|
943 |
hidden_states: torch.Tensor,
|
944 |
gating_output: torch.Tensor,
|
|
|
946 |
renormalize: bool,
|
947 |
num_expert_group: int = 0,
|
948 |
topk_group: int = 0,
|
949 |
+
scoring_func: str = "softmax",
|
950 |
+
e_score_correction_bias: Optional[torch.Tensor] = None,
|
951 |
):
|
952 |
|
953 |
assert hidden_states.shape[0] == gating_output.shape[0], "Number of tokens mismatch"
|
954 |
|
955 |
+
if scoring_func == "softmax":
|
956 |
+
scores = torch.softmax(gating_output, dim=-1)
|
957 |
+
elif scoring_func == "sigmoid":
|
958 |
+
scores = gating_output.sigmoid()
|
959 |
+
else:
|
960 |
+
raise ValueError(f"Unsupported scoring function: {scoring_func}")
|
961 |
+
|
962 |
+
if e_score_correction_bias is not None:
|
963 |
+
# Store original scores before applying correction bias. We use biased
|
964 |
+
# scores for expert selection but original scores for routing weights
|
965 |
+
original_scores = scores
|
966 |
+
scores = scores + e_score_correction_bias.unsqueeze(0)
|
967 |
+
|
968 |
num_token = scores.shape[0]
|
969 |
group_scores = (
|
970 |
scores.view(num_token, num_expert_group, -1).max(dim=-1).values
|
|
|
980 |
.reshape(num_token, -1)
|
981 |
) # [n, e]
|
982 |
tmp_scores = scores.masked_fill(~score_mask.bool(), 0.0) # [n, e]
|
983 |
+
|
984 |
+
if e_score_correction_bias is not None:
|
985 |
+
topk_ids = torch.topk(tmp_scores, k=topk, dim=-1, sorted=False)[1]
|
986 |
+
# Use original unbiased scores for the routing weights
|
987 |
+
topk_weights = original_scores.gather(1, topk_ids)
|
988 |
+
else:
|
989 |
+
topk_weights, topk_ids = torch.topk(tmp_scores, k=topk, dim=-1, sorted=False)
|
990 |
|
991 |
if renormalize:
|
992 |
topk_weights = topk_weights / topk_weights.sum(dim=-1, keepdim=True)
|
|
|
996 |
|
997 |
def get_config_dtype_str(
|
998 |
dtype: torch.dtype,
|
999 |
+
use_int4_w4a16: Optional[bool] = False,
|
1000 |
use_int8_w8a16: Optional[bool] = False,
|
1001 |
use_fp8_w8a8: Optional[bool] = False,
|
1002 |
):
|
|
|
1004 |
return "fp8_w8a8"
|
1005 |
elif use_int8_w8a16:
|
1006 |
return "int8_w8a16"
|
1007 |
+
elif use_int4_w4a16:
|
1008 |
+
return "int4_w8a16"
|
1009 |
elif dtype == torch.float:
|
1010 |
# avoiding cases where kernel fails when float32 MoE
|
1011 |
# use fp16/bfloat16 configs
|
|
|
1013 |
return None
|
1014 |
|
1015 |
|
1016 |
+
def inplace_fused_experts(
|
1017 |
+
hidden_states: torch.Tensor,
|
1018 |
+
w1: torch.Tensor,
|
1019 |
+
w2: torch.Tensor,
|
1020 |
+
topk_weights: torch.Tensor,
|
1021 |
+
topk_ids: torch.Tensor,
|
1022 |
+
use_fp8_w8a8: bool = False,
|
1023 |
+
use_int8_w8a16: bool = False,
|
1024 |
+
use_int4_w4a16: bool = False,
|
1025 |
+
w1_scale: Optional[torch.Tensor] = None,
|
1026 |
+
w2_scale: Optional[torch.Tensor] = None,
|
1027 |
+
w1_zp: Optional[torch.Tensor] = None,
|
1028 |
+
w2_zp: Optional[torch.Tensor] = None,
|
1029 |
+
a1_scale: Optional[torch.Tensor] = None,
|
1030 |
+
a2_scale: Optional[torch.Tensor] = None,
|
1031 |
+
block_shape: Optional[List[int]] = None,
|
1032 |
+
) -> None:
|
1033 |
+
fused_experts_impl(
|
1034 |
+
hidden_states,
|
1035 |
+
w1,
|
1036 |
+
w2,
|
1037 |
+
topk_weights,
|
1038 |
+
topk_ids,
|
1039 |
+
True,
|
1040 |
+
use_fp8_w8a8,
|
1041 |
+
use_int8_w8a16,
|
1042 |
+
use_int4_w4a16,
|
1043 |
+
w1_scale,
|
1044 |
+
w2_scale,
|
1045 |
+
w1_zp,
|
1046 |
+
w2_zp,
|
1047 |
+
a1_scale,
|
1048 |
+
a2_scale,
|
1049 |
+
block_shape,
|
1050 |
+
)
|
1051 |
+
|
1052 |
+
|
1053 |
+
def outplace_fused_experts(
|
1054 |
+
hidden_states: torch.Tensor,
|
1055 |
+
w1: torch.Tensor,
|
1056 |
+
w2: torch.Tensor,
|
1057 |
+
topk_weights: torch.Tensor,
|
1058 |
+
topk_ids: torch.Tensor,
|
1059 |
+
use_fp8_w8a8: bool = False,
|
1060 |
+
use_int8_w8a16: bool = False,
|
1061 |
+
use_int4_w4a16: bool = False,
|
1062 |
+
w1_scale: Optional[torch.Tensor] = None,
|
1063 |
+
w2_scale: Optional[torch.Tensor] = None,
|
1064 |
+
w1_zp: Optional[torch.Tensor] = None,
|
1065 |
+
w2_zp: Optional[torch.Tensor] = None,
|
1066 |
+
a1_scale: Optional[torch.Tensor] = None,
|
1067 |
+
a2_scale: Optional[torch.Tensor] = None,
|
1068 |
+
block_shape: Optional[List[int]] = None,
|
1069 |
+
) -> torch.Tensor:
|
1070 |
+
return fused_experts_impl(
|
1071 |
+
hidden_states,
|
1072 |
+
w1,
|
1073 |
+
w2,
|
1074 |
+
topk_weights,
|
1075 |
+
topk_ids,
|
1076 |
+
False,
|
1077 |
+
use_fp8_w8a8,
|
1078 |
+
use_int8_w8a16,
|
1079 |
+
use_int4_w4a16,
|
1080 |
+
w1_scale,
|
1081 |
+
w2_scale,
|
1082 |
+
w1_zp,
|
1083 |
+
w2_zp,
|
1084 |
+
a1_scale,
|
1085 |
+
a2_scale,
|
1086 |
+
block_shape,
|
1087 |
+
)
|
1088 |
+
|
1089 |
+
|
1090 |
def fused_experts(
|
1091 |
hidden_states: torch.Tensor,
|
1092 |
w1: torch.Tensor,
|
|
|
1094 |
topk_weights: torch.Tensor,
|
1095 |
topk_ids: torch.Tensor,
|
1096 |
inplace: bool = False,
|
|
|
1097 |
use_fp8_w8a8: bool = False,
|
1098 |
use_int8_w8a16: bool = False,
|
1099 |
+
use_int4_w4a16: bool = False,
|
1100 |
+
w1_scale: Optional[torch.Tensor] = None,
|
1101 |
+
w2_scale: Optional[torch.Tensor] = None,
|
1102 |
+
w1_zp: Optional[torch.Tensor] = None,
|
1103 |
+
w2_zp: Optional[torch.Tensor] = None,
|
1104 |
+
a1_scale: Optional[torch.Tensor] = None,
|
1105 |
+
a2_scale: Optional[torch.Tensor] = None,
|
1106 |
+
block_shape: Optional[List[int]] = None,
|
1107 |
+
):
|
1108 |
+
if inplace:
|
1109 |
+
inplace_fused_experts(
|
1110 |
+
hidden_states,
|
1111 |
+
w1,
|
1112 |
+
w2,
|
1113 |
+
topk_weights,
|
1114 |
+
topk_ids,
|
1115 |
+
use_fp8_w8a8,
|
1116 |
+
use_int8_w8a16,
|
1117 |
+
use_int4_w4a16,
|
1118 |
+
w1_scale,
|
1119 |
+
w2_scale,
|
1120 |
+
w1_zp,
|
1121 |
+
w2_zp,
|
1122 |
+
a1_scale,
|
1123 |
+
a2_scale,
|
1124 |
+
block_shape,
|
1125 |
+
)
|
1126 |
+
return hidden_states
|
1127 |
+
else:
|
1128 |
+
return outplace_fused_experts(
|
1129 |
+
hidden_states,
|
1130 |
+
w1,
|
1131 |
+
w2,
|
1132 |
+
topk_weights,
|
1133 |
+
topk_ids,
|
1134 |
+
use_fp8_w8a8,
|
1135 |
+
use_int8_w8a16,
|
1136 |
+
use_int4_w4a16,
|
1137 |
+
w1_scale,
|
1138 |
+
w2_scale,
|
1139 |
+
w1_zp,
|
1140 |
+
w2_zp,
|
1141 |
+
a1_scale,
|
1142 |
+
a2_scale,
|
1143 |
+
block_shape,
|
1144 |
+
)
|
1145 |
+
|
1146 |
+
|
1147 |
+
def fused_experts_impl(
|
1148 |
+
hidden_states: torch.Tensor,
|
1149 |
+
w1: torch.Tensor,
|
1150 |
+
w2: torch.Tensor,
|
1151 |
+
topk_weights: torch.Tensor,
|
1152 |
+
topk_ids: torch.Tensor,
|
1153 |
+
inplace: bool = False,
|
1154 |
+
use_fp8_w8a8: bool = False,
|
1155 |
+
use_int8_w8a16: bool = False,
|
1156 |
+
use_int4_w4a16: bool = False,
|
1157 |
w1_scale: Optional[torch.Tensor] = None,
|
1158 |
w2_scale: Optional[torch.Tensor] = None,
|
1159 |
+
w1_zp: Optional[torch.Tensor] = None,
|
1160 |
+
w2_zp: Optional[torch.Tensor] = None,
|
1161 |
a1_scale: Optional[torch.Tensor] = None,
|
1162 |
a2_scale: Optional[torch.Tensor] = None,
|
1163 |
+
block_shape: Optional[List[int]] = None,
|
1164 |
):
|
1165 |
# Check constraints.
|
1166 |
+
if use_int4_w4a16:
|
1167 |
+
assert hidden_states.shape[1] // 2 == w1.shape[2], "Hidden size mismatch"
|
1168 |
+
else:
|
1169 |
+
assert hidden_states.shape[1] == w1.shape[2], "Hidden size mismatch"
|
1170 |
+
|
1171 |
assert topk_weights.shape == topk_ids.shape, "topk shape mismatch"
|
1172 |
assert hidden_states.is_contiguous(), "Hidden_states must be contiguous"
|
1173 |
assert w1.is_contiguous(), "Expert weights1 must be contiguous"
|
|
|
1183 |
config_dtype = get_config_dtype_str(
|
1184 |
use_fp8_w8a8=use_fp8_w8a8,
|
1185 |
use_int8_w8a16=use_int8_w8a16,
|
1186 |
+
use_int4_w4a16=use_int4_w4a16,
|
1187 |
dtype=hidden_states.dtype,
|
1188 |
)
|
1189 |
|
|
|
1193 |
w2.shape,
|
1194 |
topk_ids.shape[1],
|
1195 |
config_dtype,
|
1196 |
+
block_shape=block_shape,
|
1197 |
)
|
1198 |
|
1199 |
config = get_config_func(M)
|
|
|
1214 |
dtype=hidden_states.dtype,
|
1215 |
)
|
1216 |
|
1217 |
+
if hidden_states.dtype == torch.bfloat16:
|
1218 |
+
compute_type = tl.bfloat16
|
1219 |
+
elif hidden_states.dtype == torch.float16:
|
1220 |
+
compute_type = tl.float16
|
1221 |
+
elif hidden_states.dtype == torch.float32:
|
1222 |
+
compute_type = tl.float32
|
1223 |
+
else:
|
1224 |
+
raise ValueError(f"Unsupported compute_type: {hidden_states.dtype}")
|
1225 |
|
1226 |
if inplace:
|
1227 |
out_hidden_states = hidden_states
|
|
|
1262 |
intermediate_cache1,
|
1263 |
a1_scale,
|
1264 |
w1_scale,
|
1265 |
+
w1_zp,
|
1266 |
curr_topk_weights,
|
1267 |
curr_topk_ids,
|
1268 |
sorted_token_ids,
|
|
|
1274 |
compute_type=compute_type,
|
1275 |
use_fp8_w8a8=use_fp8_w8a8,
|
1276 |
use_int8_w8a16=use_int8_w8a16,
|
1277 |
+
use_int4_w4a16=use_int4_w4a16,
|
1278 |
+
block_shape=block_shape,
|
1279 |
)
|
1280 |
|
1281 |
ops.silu_and_mul(intermediate_cache2, intermediate_cache1.view(-1, N))
|
|
|
1286 |
intermediate_cache3,
|
1287 |
a2_scale,
|
1288 |
w2_scale,
|
1289 |
+
w2_zp,
|
1290 |
curr_topk_weights,
|
1291 |
curr_topk_ids,
|
1292 |
sorted_token_ids,
|
|
|
1298 |
compute_type=compute_type,
|
1299 |
use_fp8_w8a8=use_fp8_w8a8,
|
1300 |
use_int8_w8a16=use_int8_w8a16,
|
1301 |
+
use_int4_w4a16=use_int4_w4a16,
|
1302 |
+
block_shape=block_shape,
|
1303 |
)
|
1304 |
|
1305 |
ops.moe_sum(
|
|
|
1317 |
topk: int,
|
1318 |
renormalize: bool,
|
1319 |
inplace: bool = False,
|
|
|
1320 |
use_grouped_topk: bool = False,
|
1321 |
num_expert_group: Optional[int] = None,
|
1322 |
topk_group: Optional[int] = None,
|
1323 |
custom_routing_function: Optional[Callable] = None,
|
1324 |
use_fp8_w8a8: bool = False,
|
1325 |
use_int8_w8a16: bool = False,
|
1326 |
+
use_int4_w4a16: bool = False,
|
1327 |
w1_scale: Optional[torch.Tensor] = None,
|
1328 |
w2_scale: Optional[torch.Tensor] = None,
|
1329 |
+
w1_zp: Optional[torch.Tensor] = None,
|
1330 |
+
w2_zp: Optional[torch.Tensor] = None,
|
1331 |
a1_scale: Optional[torch.Tensor] = None,
|
1332 |
a2_scale: Optional[torch.Tensor] = None,
|
1333 |
+
block_shape: Optional[List[int]] = None,
|
1334 |
) -> torch.Tensor:
|
1335 |
"""
|
1336 |
This function computes a Mixture of Experts (MoE) layer using two sets of
|
|
|
1346 |
- renormalize (bool): If True, renormalize the top-k weights to sum to 1.
|
1347 |
- inplace (bool): If True, perform the operation in-place.
|
1348 |
Defaults to False.
|
|
|
|
|
1349 |
- num_expert_group: Optional[int]: additional parameter for grouped_topk
|
1350 |
- topk_group: Optional[int]: additional parameter for grouped_topk
|
1351 |
- use_grouped_topk: If True, use grouped_topk instead of fused_topk
|
1352 |
note: Deepseekv2 model uses grouped_topk
|
1353 |
- use_fp8_w8a8 (bool): If True, use fp8 arithmetic to compute the inner
|
1354 |
products for w1 and w2. Defaults to False.
|
1355 |
+
- use_int8_w8a16 (bool): If True, use matmul of int8 weight and bf16/fp16
|
1356 |
+
activation to compute the inner products for w1 and w2.
|
1357 |
+
Defaults to False.
|
1358 |
+
- use_int4_w4a16 (bool): If True, use matmul of int4 weight and bf16/fp16
|
1359 |
+
activation to compute the inner products for w1 and w2.
|
1360 |
+
Defaults to False.
|
1361 |
- w1_scale (Optional[torch.Tensor]): Optional scale to be used for
|
1362 |
w1.
|
1363 |
- w2_scale (Optional[torch.Tensor]): Optional scale to be used for
|
1364 |
w2.
|
1365 |
+
- a1_scale (Optional[torch.Tensor]): Optional scale to be used for
|
1366 |
+
a1.
|
1367 |
+
- a2_scale (Optional[torch.Tensor]): Optional scale to be used for
|
1368 |
+
a2.
|
1369 |
+
- block_shape: (Optional[List[int]]): Optional block size for block-wise
|
1370 |
+
quantization.
|
1371 |
|
1372 |
Returns:
|
1373 |
- torch.Tensor: The output tensor after applying the MoE layer.
|
|
|
1401 |
topk_weights,
|
1402 |
topk_ids,
|
1403 |
inplace=inplace,
|
|
|
1404 |
use_fp8_w8a8=use_fp8_w8a8,
|
1405 |
use_int8_w8a16=use_int8_w8a16,
|
1406 |
+
use_int4_w4a16=use_int4_w4a16,
|
1407 |
w1_scale=w1_scale,
|
1408 |
w2_scale=w2_scale,
|
1409 |
+
w1_zp=w1_zp,
|
1410 |
+
w2_zp=w2_zp,
|
1411 |
a1_scale=a1_scale,
|
1412 |
a2_scale=a2_scale,
|
1413 |
+
block_shape=block_shape,
|
1414 |
)
|
build/torch25-cxx98-cu124-x86_64-linux/moe/platforms.py
CHANGED
@@ -1,22 +1,32 @@
|
|
1 |
-
from
|
2 |
-
import os
|
3 |
-
from functools import lru_cache, wraps
|
4 |
|
5 |
import torch
|
6 |
|
7 |
IS_ROCM = torch.version.hip is not None
|
8 |
|
9 |
-
|
|
|
|
|
|
|
|
|
|
|
10 |
@classmethod
|
11 |
@lru_cache(maxsize=8)
|
12 |
def get_device_name(cls, device_id: int = 0) -> str:
|
13 |
return torch.cuda.get_device_name(0)
|
14 |
|
15 |
-
|
|
|
|
|
|
|
|
|
16 |
@classmethod
|
17 |
@lru_cache(maxsize=8)
|
18 |
def get_device_name(cls, device_id: int = 0) -> str:
|
19 |
return torch.cuda.get_device_name(device_id)
|
20 |
|
|
|
|
|
|
|
21 |
|
22 |
current_platform = RocmPlatform() if IS_ROCM else CudaPlatform()
|
|
|
1 |
+
from functools import lru_cache
|
|
|
|
|
2 |
|
3 |
import torch
|
4 |
|
5 |
IS_ROCM = torch.version.hip is not None
|
6 |
|
7 |
+
|
8 |
+
class Platform:
|
9 |
+
simple_compile_backend: str = "inductor"
|
10 |
+
|
11 |
+
|
12 |
+
class CudaPlatform(Platform):
|
13 |
@classmethod
|
14 |
@lru_cache(maxsize=8)
|
15 |
def get_device_name(cls, device_id: int = 0) -> str:
|
16 |
return torch.cuda.get_device_name(0)
|
17 |
|
18 |
+
def is_rocm(self):
|
19 |
+
return False
|
20 |
+
|
21 |
+
|
22 |
+
class RocmPlatform(Platform):
|
23 |
@classmethod
|
24 |
@lru_cache(maxsize=8)
|
25 |
def get_device_name(cls, device_id: int = 0) -> str:
|
26 |
return torch.cuda.get_device_name(device_id)
|
27 |
|
28 |
+
def is_rocm(self):
|
29 |
+
return True
|
30 |
+
|
31 |
|
32 |
current_platform = RocmPlatform() if IS_ROCM else CudaPlatform()
|
build/torch26-cxx11-cu118-x86_64-linux/moe/_moe_ooomuvan6f6yy.abi3.so
DELETED
@@ -1,3 +0,0 @@
|
|
1 |
-
version https://git-lfs.github.com/spec/v1
|
2 |
-
oid sha256:1de7247bc801effbb2c8698bb47eddb97a57baeea9fb7bb05f70f42d0db0ab7f
|
3 |
-
size 84165848
|
|
|
|
|
|
|
|
build/torch26-cxx11-cu118-x86_64-linux/moe/_moe_zlz7rpd2goyn2.abi3.so
ADDED
@@ -0,0 +1,3 @@
|
|
|
|
|
|
|
|
|
1 |
+
version https://git-lfs.github.com/spec/v1
|
2 |
+
oid sha256:658fb6f129cf6ba0ea172ccfd1f115c0a03e5574122456ab9ecd35122908369a
|
3 |
+
size 85823776
|
build/torch26-cxx11-cu118-x86_64-linux/moe/_ops.py
CHANGED
@@ -1,9 +1,9 @@
|
|
1 |
import torch
|
2 |
-
from . import
|
3 |
-
ops = torch.ops.
|
4 |
|
5 |
def add_op_namespace_prefix(op_name: str):
|
6 |
"""
|
7 |
Prefix op by namespace.
|
8 |
"""
|
9 |
-
return f"
|
|
|
1 |
import torch
|
2 |
+
from . import _moe_zlz7rpd2goyn2
|
3 |
+
ops = torch.ops._moe_zlz7rpd2goyn2
|
4 |
|
5 |
def add_op_namespace_prefix(op_name: str):
|
6 |
"""
|
7 |
Prefix op by namespace.
|
8 |
"""
|
9 |
+
return f"_moe_zlz7rpd2goyn2::{op_name}"
|
build/torch26-cxx11-cu118-x86_64-linux/moe/fp8.py
CHANGED
@@ -1,6 +1,11 @@
|
|
|
|
|
|
1 |
import torch
|
|
|
|
|
2 |
|
3 |
-
|
|
|
4 |
|
5 |
|
6 |
def is_hip() -> bool:
|
@@ -49,15 +54,179 @@ def scaled_fp8_quant(
|
|
49 |
if scale is None:
|
50 |
if use_per_token_if_dynamic:
|
51 |
scale = torch.empty((shape[0], 1), device=input.device, dtype=torch.float32)
|
52 |
-
|
53 |
-
output, input, scale, scale_ub
|
54 |
-
)
|
55 |
else:
|
56 |
scale = torch.zeros(1, device=input.device, dtype=torch.float32)
|
57 |
-
|
58 |
else:
|
59 |
# num_token_padding not implemented for this case
|
60 |
assert scale.numel() == 1 or num_token_padding is None
|
61 |
-
|
62 |
|
63 |
return output, scale
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
1 |
+
from typing import Tuple, Optional, Union
|
2 |
+
|
3 |
import torch
|
4 |
+
import triton
|
5 |
+
import triton.language as tl
|
6 |
|
7 |
+
|
8 |
+
from ._ops import ops
|
9 |
|
10 |
|
11 |
def is_hip() -> bool:
|
|
|
54 |
if scale is None:
|
55 |
if use_per_token_if_dynamic:
|
56 |
scale = torch.empty((shape[0], 1), device=input.device, dtype=torch.float32)
|
57 |
+
ops.dynamic_per_token_scaled_fp8_quant(output, input, scale, scale_ub)
|
|
|
|
|
58 |
else:
|
59 |
scale = torch.zeros(1, device=input.device, dtype=torch.float32)
|
60 |
+
ops.dynamic_scaled_fp8_quant(output, input, scale)
|
61 |
else:
|
62 |
# num_token_padding not implemented for this case
|
63 |
assert scale.numel() == 1 or num_token_padding is None
|
64 |
+
ops.static_scaled_fp8_quant(output, input, scale)
|
65 |
|
66 |
return output, scale
|
67 |
+
|
68 |
+
|
69 |
+
@triton.jit
|
70 |
+
def _per_token_group_quant_fp8(
|
71 |
+
# Pointers to inputs and output
|
72 |
+
y_ptr,
|
73 |
+
y_q_ptr,
|
74 |
+
y_s_ptr,
|
75 |
+
group_size,
|
76 |
+
# Avoid to divide zero
|
77 |
+
eps,
|
78 |
+
# Information for float8
|
79 |
+
fp8_min,
|
80 |
+
fp8_max,
|
81 |
+
# Meta-parameters
|
82 |
+
BLOCK: tl.constexpr,
|
83 |
+
):
|
84 |
+
"""A Triton-accelerated function to perform per-token-group
|
85 |
+
quantization on a tensor.
|
86 |
+
This function converts the tensor values into float8 values.
|
87 |
+
"""
|
88 |
+
# Map the program id to the row of X and Y it should compute.
|
89 |
+
g_id = tl.program_id(0)
|
90 |
+
y_ptr += g_id * group_size
|
91 |
+
y_q_ptr += g_id * group_size
|
92 |
+
y_s_ptr += g_id
|
93 |
+
|
94 |
+
cols = tl.arange(0, BLOCK) # N <= BLOCK
|
95 |
+
mask = cols < group_size
|
96 |
+
|
97 |
+
y = tl.load(y_ptr + cols, mask=mask, other=0.0).to(tl.float32)
|
98 |
+
# Quant
|
99 |
+
_absmax = tl.maximum(tl.max(tl.abs(y)), eps)
|
100 |
+
y_s = _absmax / fp8_max
|
101 |
+
y_q = tl.clamp(y / y_s, fp8_min, fp8_max).to(y_q_ptr.dtype.element_ty)
|
102 |
+
|
103 |
+
tl.store(y_q_ptr + cols, y_q, mask=mask)
|
104 |
+
tl.store(y_s_ptr, y_s)
|
105 |
+
|
106 |
+
|
107 |
+
@triton.jit
|
108 |
+
def _per_token_group_quant_fp8_colmajor(
|
109 |
+
# Pointers to inputs and output
|
110 |
+
y_ptr,
|
111 |
+
y_q_ptr,
|
112 |
+
y_s_ptr,
|
113 |
+
group_size,
|
114 |
+
# Num columns of y
|
115 |
+
y_num_columns,
|
116 |
+
# Stride from one column to the next of y_s
|
117 |
+
y_s_col_stride,
|
118 |
+
# Avoid to divide zero
|
119 |
+
eps,
|
120 |
+
# Information for float8
|
121 |
+
fp8_min,
|
122 |
+
fp8_max,
|
123 |
+
# Meta-parameters
|
124 |
+
BLOCK: tl.constexpr,
|
125 |
+
):
|
126 |
+
"""A Triton-accelerated function to perform per-token-group
|
127 |
+
quantization on a tensor.
|
128 |
+
This function converts the tensor values into float8 values.
|
129 |
+
"""
|
130 |
+
# Map the program id to the row of X and Y it should compute.
|
131 |
+
g_id = tl.program_id(0)
|
132 |
+
y_ptr += g_id * group_size
|
133 |
+
y_q_ptr += g_id * group_size
|
134 |
+
|
135 |
+
# Convert g_id the flattened block coordinate to 2D so we can index
|
136 |
+
# into the output y_scales matrix
|
137 |
+
blocks_per_row = y_num_columns // group_size
|
138 |
+
scale_col = g_id % blocks_per_row
|
139 |
+
scale_row = g_id // blocks_per_row
|
140 |
+
y_s_ptr += scale_col * y_s_col_stride + scale_row
|
141 |
+
|
142 |
+
cols = tl.arange(0, BLOCK) # group_size <= BLOCK
|
143 |
+
mask = cols < group_size
|
144 |
+
|
145 |
+
y = tl.load(y_ptr + cols, mask=mask, other=0.0).to(tl.float32)
|
146 |
+
# Quant
|
147 |
+
_absmax = tl.maximum(tl.max(tl.abs(y)), eps)
|
148 |
+
y_s = _absmax / fp8_max
|
149 |
+
y_q = tl.clamp(y / y_s, fp8_min, fp8_max).to(y_q_ptr.dtype.element_ty)
|
150 |
+
|
151 |
+
tl.store(y_q_ptr + cols, y_q, mask=mask)
|
152 |
+
tl.store(y_s_ptr, y_s)
|
153 |
+
|
154 |
+
|
155 |
+
def per_token_group_quant_fp8(
|
156 |
+
x: torch.Tensor,
|
157 |
+
group_size: int,
|
158 |
+
eps: float = 1e-10,
|
159 |
+
dtype: Optional[torch.dtype] = None,
|
160 |
+
column_major_scales: bool = False,
|
161 |
+
) -> Tuple[torch.Tensor, torch.Tensor]:
|
162 |
+
"""Function to perform per-token-group quantization on an input tensor `x`.
|
163 |
+
It converts the tensor values into signed float8 values and returns the
|
164 |
+
quantized tensor along with the scaling factor used for quantization.
|
165 |
+
Args:
|
166 |
+
x: The input tensor with ndim >= 2.
|
167 |
+
group_size: The group size used for quantization.
|
168 |
+
eps: The minimum to avoid dividing zero.
|
169 |
+
dtype: The dype of output tensor. Note that only `torch.float8_e4m3fn`
|
170 |
+
is supported for now.
|
171 |
+
Returns:
|
172 |
+
Tuple[torch.Tensor, torch.Tensor]: The quantized tensor and the
|
173 |
+
scaling factor for quantization.
|
174 |
+
"""
|
175 |
+
if dtype is None:
|
176 |
+
dtype = (
|
177 |
+
torch.float8_e4m3fnuz if current_platform.is_rocm() else torch.float8_e4m3fn
|
178 |
+
)
|
179 |
+
assert x.shape[-1] % group_size == 0, (
|
180 |
+
f"the last dimension of `x` {x.shape[-1]} must be divisible "
|
181 |
+
f"by `group_size` {group_size}"
|
182 |
+
)
|
183 |
+
assert x.is_contiguous(), "`x` must be contiguous"
|
184 |
+
|
185 |
+
finfo = torch.finfo(dtype)
|
186 |
+
fp8_min = finfo.min
|
187 |
+
fp8_max = finfo.max
|
188 |
+
|
189 |
+
x_q = torch.empty_like(x, device=x.device, dtype=dtype)
|
190 |
+
M = x.numel() // group_size
|
191 |
+
N = group_size
|
192 |
+
if column_major_scales:
|
193 |
+
shape = (x.shape[-1] // group_size,) + x.shape[:-1]
|
194 |
+
x_s = torch.empty(shape, device=x.device, dtype=torch.float32).permute(-1, -2)
|
195 |
+
else:
|
196 |
+
shape = x.shape[:-1] + (x.shape[-1] // group_size,)
|
197 |
+
x_s = torch.empty(shape, device=x.device, dtype=torch.float32)
|
198 |
+
|
199 |
+
BLOCK = triton.next_power_of_2(N)
|
200 |
+
# heuristics for number of warps
|
201 |
+
num_warps = min(max(BLOCK // 256, 1), 8)
|
202 |
+
num_stages = 1
|
203 |
+
if column_major_scales:
|
204 |
+
_per_token_group_quant_fp8_colmajor[(M,)](
|
205 |
+
x,
|
206 |
+
x_q,
|
207 |
+
x_s,
|
208 |
+
group_size,
|
209 |
+
x.shape[1],
|
210 |
+
x_s.stride(1),
|
211 |
+
eps,
|
212 |
+
fp8_min=fp8_min,
|
213 |
+
fp8_max=fp8_max,
|
214 |
+
BLOCK=BLOCK,
|
215 |
+
num_warps=num_warps,
|
216 |
+
num_stages=num_stages,
|
217 |
+
)
|
218 |
+
else:
|
219 |
+
_per_token_group_quant_fp8[(M,)](
|
220 |
+
x,
|
221 |
+
x_q,
|
222 |
+
x_s,
|
223 |
+
group_size,
|
224 |
+
eps,
|
225 |
+
fp8_min=fp8_min,
|
226 |
+
fp8_max=fp8_max,
|
227 |
+
BLOCK=BLOCK,
|
228 |
+
num_warps=num_warps,
|
229 |
+
num_stages=num_stages,
|
230 |
+
)
|
231 |
+
|
232 |
+
return x_q, x_s
|
build/torch26-cxx11-cu118-x86_64-linux/moe/fused_marlin_moe.py
CHANGED
@@ -40,7 +40,6 @@ def single_marlin_moe(
|
|
40 |
g_idx: Optional[torch.Tensor] = None,
|
41 |
sort_indices: Optional[torch.Tensor] = None,
|
42 |
w_zeros: Optional[torch.Tensor] = None,
|
43 |
-
override_config: Optional[Dict[str, Any]] = None,
|
44 |
num_bits: int = 8,
|
45 |
is_k_full: bool = True,
|
46 |
) -> torch.Tensor:
|
@@ -61,8 +60,6 @@ def single_marlin_moe(
|
|
61 |
- topk (int): The number of top-k experts to select.
|
62 |
- renormalize (bool): If True, renormalize the top-k weights to sum to 1.
|
63 |
- w_zeros (Optional[torch.Tensor]): Optional zero points to be used for w.
|
64 |
-
- override_config (Optional[Dict[str, Any]]): Optional override
|
65 |
-
for the kernel configuration.
|
66 |
- num_bits (bool): The number of bits in expert weights quantization.
|
67 |
|
68 |
Returns:
|
@@ -90,7 +87,6 @@ def single_marlin_moe(
|
|
90 |
w.shape,
|
91 |
topk_ids.shape[1],
|
92 |
None,
|
93 |
-
override_config=override_config,
|
94 |
is_marlin=True,
|
95 |
)
|
96 |
config = get_config_func(M)
|
@@ -154,6 +150,25 @@ def single_marlin_moe(
|
|
154 |
return torch.sum(intermediate_cache.view(*intermediate_cache.shape), dim=1)
|
155 |
|
156 |
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
157 |
def fused_marlin_moe(
|
158 |
hidden_states: torch.Tensor,
|
159 |
w1: torch.Tensor,
|
@@ -169,7 +184,6 @@ def fused_marlin_moe(
|
|
169 |
sort_indices2: Optional[torch.Tensor] = None,
|
170 |
w1_zeros: Optional[torch.Tensor] = None,
|
171 |
w2_zeros: Optional[torch.Tensor] = None,
|
172 |
-
override_config: Optional[Dict[str, Any]] = None,
|
173 |
num_bits: int = 8,
|
174 |
is_k_full: bool = True,
|
175 |
) -> torch.Tensor:
|
@@ -193,8 +207,6 @@ def fused_marlin_moe(
|
|
193 |
permutation.
|
194 |
- topk_weights (torch.Tensor): Top-k weights.
|
195 |
- topk_ids (torch.Tensor): Indices of topk-k elements.
|
196 |
-
- override_config (Optional[Dict[str, Any]]): Optional override
|
197 |
-
for the kernel configuration.
|
198 |
- w1_zeros (Optional[torch.Tensor]): Optional zero points to be used for w1.
|
199 |
- w2_zeros (Optional[torch.Tensor]): Optional zero points to be used for w2.
|
200 |
- num_bits (bool): The number of bits in expert weights quantization.
|
@@ -248,7 +260,6 @@ def fused_marlin_moe(
|
|
248 |
w2.shape,
|
249 |
topk_ids.shape[1],
|
250 |
None,
|
251 |
-
override_config=override_config,
|
252 |
is_marlin=True,
|
253 |
)
|
254 |
config = get_config_func(M)
|
@@ -350,6 +361,30 @@ def fused_marlin_moe(
|
|
350 |
return torch.sum(intermediate_cache3.view(*intermediate_cache3.shape), dim=1)
|
351 |
|
352 |
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
353 |
if hasattr(ops, "marlin_gemm_moe"):
|
354 |
|
355 |
@register_fake(add_op_namespace_prefix("marlin_gemm_moe"))
|
|
|
40 |
g_idx: Optional[torch.Tensor] = None,
|
41 |
sort_indices: Optional[torch.Tensor] = None,
|
42 |
w_zeros: Optional[torch.Tensor] = None,
|
|
|
43 |
num_bits: int = 8,
|
44 |
is_k_full: bool = True,
|
45 |
) -> torch.Tensor:
|
|
|
60 |
- topk (int): The number of top-k experts to select.
|
61 |
- renormalize (bool): If True, renormalize the top-k weights to sum to 1.
|
62 |
- w_zeros (Optional[torch.Tensor]): Optional zero points to be used for w.
|
|
|
|
|
63 |
- num_bits (bool): The number of bits in expert weights quantization.
|
64 |
|
65 |
Returns:
|
|
|
87 |
w.shape,
|
88 |
topk_ids.shape[1],
|
89 |
None,
|
|
|
90 |
is_marlin=True,
|
91 |
)
|
92 |
config = get_config_func(M)
|
|
|
150 |
return torch.sum(intermediate_cache.view(*intermediate_cache.shape), dim=1)
|
151 |
|
152 |
|
153 |
+
if hasattr(ops, "single_marlin_gemm_moe"):
|
154 |
+
|
155 |
+
@register_fake(add_op_namespace_prefix("single_marlin_gemm_moe"))
|
156 |
+
def single_marlin_moe_fake(
|
157 |
+
hidden_states: torch.Tensor,
|
158 |
+
w: torch.Tensor,
|
159 |
+
scales: torch.Tensor,
|
160 |
+
gating_output: torch.Tensor,
|
161 |
+
topk: int,
|
162 |
+
renormalize: bool,
|
163 |
+
g_idx: Optional[torch.Tensor] = None,
|
164 |
+
sort_indices: Optional[torch.Tensor] = None,
|
165 |
+
w_zeros: Optional[torch.Tensor] = None,
|
166 |
+
num_bits: int = 8,
|
167 |
+
is_k_full: bool = True,
|
168 |
+
) -> torch.Tensor:
|
169 |
+
return torch.empty_like(hidden_states)
|
170 |
+
|
171 |
+
|
172 |
def fused_marlin_moe(
|
173 |
hidden_states: torch.Tensor,
|
174 |
w1: torch.Tensor,
|
|
|
184 |
sort_indices2: Optional[torch.Tensor] = None,
|
185 |
w1_zeros: Optional[torch.Tensor] = None,
|
186 |
w2_zeros: Optional[torch.Tensor] = None,
|
|
|
187 |
num_bits: int = 8,
|
188 |
is_k_full: bool = True,
|
189 |
) -> torch.Tensor:
|
|
|
207 |
permutation.
|
208 |
- topk_weights (torch.Tensor): Top-k weights.
|
209 |
- topk_ids (torch.Tensor): Indices of topk-k elements.
|
|
|
|
|
210 |
- w1_zeros (Optional[torch.Tensor]): Optional zero points to be used for w1.
|
211 |
- w2_zeros (Optional[torch.Tensor]): Optional zero points to be used for w2.
|
212 |
- num_bits (bool): The number of bits in expert weights quantization.
|
|
|
260 |
w2.shape,
|
261 |
topk_ids.shape[1],
|
262 |
None,
|
|
|
263 |
is_marlin=True,
|
264 |
)
|
265 |
config = get_config_func(M)
|
|
|
361 |
return torch.sum(intermediate_cache3.view(*intermediate_cache3.shape), dim=1)
|
362 |
|
363 |
|
364 |
+
if hasattr(ops, "fused_marlin_moe"):
|
365 |
+
|
366 |
+
@register_fake(add_op_namespace_prefix("fused_marlin_moe"))
|
367 |
+
def fused_marlin_moe_fake(
|
368 |
+
hidden_states: torch.Tensor,
|
369 |
+
w1: torch.Tensor,
|
370 |
+
w2: torch.Tensor,
|
371 |
+
w1_scale: torch.Tensor,
|
372 |
+
w2_scale: torch.Tensor,
|
373 |
+
gating_output: torch.Tensor,
|
374 |
+
topk_weights: torch.Tensor,
|
375 |
+
topk_ids: torch.Tensor,
|
376 |
+
g_idx1: Optional[torch.Tensor] = None,
|
377 |
+
g_idx2: Optional[torch.Tensor] = None,
|
378 |
+
sort_indices1: Optional[torch.Tensor] = None,
|
379 |
+
sort_indices2: Optional[torch.Tensor] = None,
|
380 |
+
w1_zeros: Optional[torch.Tensor] = None,
|
381 |
+
w2_zeros: Optional[torch.Tensor] = None,
|
382 |
+
num_bits: int = 8,
|
383 |
+
is_k_full: bool = True,
|
384 |
+
) -> torch.Tensor:
|
385 |
+
return torch.empty_like(hidden_states)
|
386 |
+
|
387 |
+
|
388 |
if hasattr(ops, "marlin_gemm_moe"):
|
389 |
|
390 |
@register_fake(add_op_namespace_prefix("marlin_gemm_moe"))
|
build/torch26-cxx11-cu118-x86_64-linux/moe/fused_moe.py
CHANGED
@@ -1,21 +1,242 @@
|
|
|
|
1 |
"""Fused MoE kernel."""
|
2 |
|
3 |
import functools
|
4 |
import json
|
|
|
5 |
import os
|
6 |
-
from typing import Any, Callable, Dict, Optional, Tuple
|
7 |
|
8 |
import torch
|
9 |
import triton
|
10 |
import triton.language as tl
|
11 |
|
|
|
12 |
from ._ops import ops
|
13 |
-
from .fp8 import scaled_fp8_quant
|
14 |
from .platforms import current_platform
|
15 |
|
|
|
|
|
|
|
16 |
VLLM_FUSED_MOE_CHUNK_SIZE = int(os.getenv("VLLM_FUSED_MOE_CHUNK_SIZE", "32768"))
|
17 |
|
18 |
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
19 |
@triton.jit
|
20 |
def fused_moe_kernel(
|
21 |
# Pointers to matrices
|
@@ -44,8 +265,14 @@ def fused_moe_kernel(
|
|
44 |
stride_bn,
|
45 |
stride_cm,
|
46 |
stride_cn,
|
|
|
|
|
47 |
stride_bse,
|
|
|
48 |
stride_bsn,
|
|
|
|
|
|
|
49 |
# Meta-parameters
|
50 |
BLOCK_SIZE_M: tl.constexpr,
|
51 |
BLOCK_SIZE_N: tl.constexpr,
|
@@ -105,17 +332,17 @@ def fused_moe_kernel(
|
|
105 |
num_tokens_post_padded = tl.load(num_tokens_post_padded_ptr)
|
106 |
if pid_m * BLOCK_SIZE_M >= num_tokens_post_padded:
|
107 |
return
|
108 |
-
offs_token_id = pid_m * BLOCK_SIZE_M + tl.arange(0, BLOCK_SIZE_M)
|
109 |
offs_token = tl.load(sorted_token_ids_ptr + offs_token_id)
|
110 |
token_mask = offs_token < num_valid_tokens
|
111 |
|
112 |
-
offs_bn = (pid_n * BLOCK_SIZE_N + tl.arange(0, BLOCK_SIZE_N)) % N
|
113 |
offs_k = tl.arange(0, BLOCK_SIZE_K)
|
114 |
a_ptrs = a_ptr + (
|
115 |
offs_token[:, None] // top_k * stride_am + offs_k[None, :] * stride_ak
|
116 |
)
|
117 |
|
118 |
-
off_experts = tl.load(expert_ids_ptr + pid_m)
|
119 |
b_ptrs = (
|
120 |
b_ptr
|
121 |
+ off_experts * stride_be
|
@@ -128,8 +355,15 @@ def fused_moe_kernel(
|
|
128 |
b_scale = tl.load(b_scale_ptrs)
|
129 |
|
130 |
if use_fp8_w8a8:
|
131 |
-
|
132 |
-
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
133 |
|
134 |
# -----------------------------------------------------------
|
135 |
# Iterate to compute a block of the C matrix.
|
@@ -151,7 +385,17 @@ def fused_moe_kernel(
|
|
151 |
if use_int8_w8a16:
|
152 |
accumulator = tl.dot(a, b.to(compute_type), acc=accumulator)
|
153 |
elif use_fp8_w8a8:
|
154 |
-
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
155 |
else:
|
156 |
accumulator += tl.dot(a, b)
|
157 |
# Advance the ptrs to the next K block.
|
@@ -164,7 +408,10 @@ def fused_moe_kernel(
|
|
164 |
if use_int8_w8a16:
|
165 |
accumulator = (accumulator * b_scale).to(compute_type)
|
166 |
elif use_fp8_w8a8:
|
167 |
-
|
|
|
|
|
|
|
168 |
else:
|
169 |
accumulator = accumulator.to(compute_type)
|
170 |
# -----------------------------------------------------------
|
@@ -175,6 +422,141 @@ def fused_moe_kernel(
|
|
175 |
tl.store(c_ptrs, accumulator, mask=c_mask)
|
176 |
|
177 |
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
178 |
def moe_align_block_size(
|
179 |
topk_ids: torch.Tensor, block_size: int, num_experts: int
|
180 |
) -> Tuple[torch.Tensor, torch.Tensor, torch.Tensor]:
|
@@ -225,9 +607,34 @@ def moe_align_block_size(
|
|
225 |
(max_num_m_blocks,), dtype=torch.int32, device=topk_ids.device
|
226 |
)
|
227 |
num_tokens_post_pad = torch.empty((1), dtype=torch.int32, device=topk_ids.device)
|
228 |
-
|
229 |
-
|
230 |
-
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
231 |
return sorted_ids, expert_ids, num_tokens_post_pad
|
232 |
|
233 |
|
@@ -237,6 +644,7 @@ def invoke_fused_moe_kernel(
|
|
237 |
C: torch.Tensor,
|
238 |
A_scale: Optional[torch.Tensor],
|
239 |
B_scale: Optional[torch.Tensor],
|
|
|
240 |
topk_weights: torch.Tensor,
|
241 |
topk_ids: torch.Tensor,
|
242 |
sorted_token_ids: torch.Tensor,
|
@@ -248,64 +656,147 @@ def invoke_fused_moe_kernel(
|
|
248 |
compute_type: tl.dtype,
|
249 |
use_fp8_w8a8: bool,
|
250 |
use_int8_w8a16: bool,
|
|
|
|
|
251 |
) -> None:
|
252 |
assert topk_weights.stride(1) == 1
|
253 |
assert sorted_token_ids.stride(0) == 1
|
254 |
|
255 |
if use_fp8_w8a8:
|
256 |
-
A, A_scale = scaled_fp8_quant(A, A_scale)
|
257 |
assert B_scale is not None
|
258 |
-
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
259 |
assert B_scale is not None
|
|
|
260 |
else:
|
261 |
assert A_scale is None
|
262 |
assert B_scale is None
|
263 |
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
264 |
grid = lambda META: (
|
265 |
-
triton.cdiv(
|
266 |
* triton.cdiv(B.shape[1], META["BLOCK_SIZE_N"]),
|
267 |
)
|
268 |
|
269 |
-
|
270 |
-
|
271 |
-
|
272 |
-
|
273 |
-
|
274 |
-
B_scale
|
275 |
-
|
276 |
-
|
277 |
-
|
278 |
-
|
279 |
-
|
280 |
-
|
281 |
-
|
282 |
-
|
283 |
-
|
284 |
-
|
285 |
-
|
286 |
-
|
287 |
-
|
288 |
-
|
289 |
-
|
290 |
-
|
291 |
-
|
292 |
-
|
293 |
-
|
294 |
-
|
295 |
-
|
296 |
-
|
297 |
-
|
298 |
-
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
299 |
|
300 |
|
301 |
-
|
|
|
|
|
|
|
302 |
device_name = current_platform.get_device_name().replace(" ", "_")
|
303 |
dtype_selector = "" if not dtype else f",dtype={dtype}"
|
304 |
-
|
|
|
|
|
|
|
305 |
|
306 |
|
|
|
307 |
@functools.lru_cache
|
308 |
-
def get_moe_configs(
|
|
|
|
|
|
|
|
|
|
|
|
|
309 |
"""
|
310 |
Return optimized configurations for the fused MoE kernel.
|
311 |
|
@@ -317,18 +808,27 @@ def get_moe_configs(E: int, N: int, dtype: Optional[str]) -> Optional[Dict[int,
|
|
317 |
|
318 |
# First look up if an optimized configuration is available in the configs
|
319 |
# directory
|
320 |
-
|
|
|
321 |
|
322 |
config_file_path = os.path.join(
|
323 |
os.path.dirname(os.path.realpath(__file__)), "configs", json_file_name
|
324 |
)
|
325 |
if os.path.exists(config_file_path):
|
326 |
with open(config_file_path) as f:
|
|
|
327 |
# If a configuration has been found, return it
|
328 |
return {int(key): val for key, val in json.load(f).items()}
|
329 |
|
330 |
# If no optimized configuration is available, we will use the default
|
331 |
# configuration
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
332 |
return None
|
333 |
|
334 |
|
@@ -340,21 +840,34 @@ def get_default_config(
|
|
340 |
topk: int,
|
341 |
dtype: Optional[str],
|
342 |
is_marlin: bool,
|
|
|
343 |
) -> Dict[str, int]:
|
344 |
-
|
345 |
-
|
346 |
-
|
347 |
-
"BLOCK_SIZE_K": 32,
|
348 |
-
"GROUP_SIZE_M": 8,
|
349 |
-
}
|
350 |
-
# A heuristic: fused marlin works faster with this config for small M
|
351 |
-
if M <= E or (is_marlin and M <= 32):
|
352 |
config = {
|
353 |
-
"BLOCK_SIZE_M":
|
354 |
-
"BLOCK_SIZE_N":
|
355 |
-
"BLOCK_SIZE_K":
|
356 |
-
"GROUP_SIZE_M":
|
|
|
|
|
357 |
}
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
358 |
return config
|
359 |
|
360 |
|
@@ -364,15 +877,21 @@ def try_get_optimal_moe_config(
|
|
364 |
top_k: int,
|
365 |
dtype: Optional[str],
|
366 |
M: int,
|
367 |
-
override_config: Optional[Dict[str, Any]] = None,
|
368 |
is_marlin: bool = False,
|
|
|
369 |
):
|
|
|
|
|
|
|
|
|
370 |
if override_config:
|
371 |
config = override_config
|
372 |
else:
|
373 |
# First try to load optimal config from the file
|
374 |
E, _, N = w2_shape
|
375 |
-
|
|
|
|
|
376 |
|
377 |
if configs:
|
378 |
# If an optimal configuration map has been found, look up the
|
@@ -380,7 +899,9 @@ def try_get_optimal_moe_config(
|
|
380 |
config = configs[min(configs.keys(), key=lambda x: abs(x - M))]
|
381 |
else:
|
382 |
# Else use the default config
|
383 |
-
config = get_default_config(
|
|
|
|
|
384 |
return config
|
385 |
|
386 |
|
@@ -416,7 +937,8 @@ def fused_topk(
|
|
416 |
return topk_weights, topk_ids
|
417 |
|
418 |
|
419 |
-
# This is used by the Deepseek-V2 model
|
|
|
420 |
def grouped_topk(
|
421 |
hidden_states: torch.Tensor,
|
422 |
gating_output: torch.Tensor,
|
@@ -424,11 +946,25 @@ def grouped_topk(
|
|
424 |
renormalize: bool,
|
425 |
num_expert_group: int = 0,
|
426 |
topk_group: int = 0,
|
|
|
|
|
427 |
):
|
428 |
|
429 |
assert hidden_states.shape[0] == gating_output.shape[0], "Number of tokens mismatch"
|
430 |
|
431 |
-
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
432 |
num_token = scores.shape[0]
|
433 |
group_scores = (
|
434 |
scores.view(num_token, num_expert_group, -1).max(dim=-1).values
|
@@ -444,7 +980,13 @@ def grouped_topk(
|
|
444 |
.reshape(num_token, -1)
|
445 |
) # [n, e]
|
446 |
tmp_scores = scores.masked_fill(~score_mask.bool(), 0.0) # [n, e]
|
447 |
-
|
|
|
|
|
|
|
|
|
|
|
|
|
448 |
|
449 |
if renormalize:
|
450 |
topk_weights = topk_weights / topk_weights.sum(dim=-1, keepdim=True)
|
@@ -454,6 +996,7 @@ def grouped_topk(
|
|
454 |
|
455 |
def get_config_dtype_str(
|
456 |
dtype: torch.dtype,
|
|
|
457 |
use_int8_w8a16: Optional[bool] = False,
|
458 |
use_fp8_w8a8: Optional[bool] = False,
|
459 |
):
|
@@ -461,6 +1004,8 @@ def get_config_dtype_str(
|
|
461 |
return "fp8_w8a8"
|
462 |
elif use_int8_w8a16:
|
463 |
return "int8_w8a16"
|
|
|
|
|
464 |
elif dtype == torch.float:
|
465 |
# avoiding cases where kernel fails when float32 MoE
|
466 |
# use fp16/bfloat16 configs
|
@@ -468,6 +1013,80 @@ def get_config_dtype_str(
|
|
468 |
return None
|
469 |
|
470 |
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
471 |
def fused_experts(
|
472 |
hidden_states: torch.Tensor,
|
473 |
w1: torch.Tensor,
|
@@ -475,16 +1094,80 @@ def fused_experts(
|
|
475 |
topk_weights: torch.Tensor,
|
476 |
topk_ids: torch.Tensor,
|
477 |
inplace: bool = False,
|
478 |
-
override_config: Optional[Dict[str, Any]] = None,
|
479 |
use_fp8_w8a8: bool = False,
|
480 |
use_int8_w8a16: bool = False,
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
481 |
w1_scale: Optional[torch.Tensor] = None,
|
482 |
w2_scale: Optional[torch.Tensor] = None,
|
|
|
|
|
483 |
a1_scale: Optional[torch.Tensor] = None,
|
484 |
a2_scale: Optional[torch.Tensor] = None,
|
|
|
485 |
):
|
486 |
# Check constraints.
|
487 |
-
|
|
|
|
|
|
|
|
|
488 |
assert topk_weights.shape == topk_ids.shape, "topk shape mismatch"
|
489 |
assert hidden_states.is_contiguous(), "Hidden_states must be contiguous"
|
490 |
assert w1.is_contiguous(), "Expert weights1 must be contiguous"
|
@@ -500,6 +1183,7 @@ def fused_experts(
|
|
500 |
config_dtype = get_config_dtype_str(
|
501 |
use_fp8_w8a8=use_fp8_w8a8,
|
502 |
use_int8_w8a16=use_int8_w8a16,
|
|
|
503 |
dtype=hidden_states.dtype,
|
504 |
)
|
505 |
|
@@ -509,7 +1193,7 @@ def fused_experts(
|
|
509 |
w2.shape,
|
510 |
topk_ids.shape[1],
|
511 |
config_dtype,
|
512 |
-
|
513 |
)
|
514 |
|
515 |
config = get_config_func(M)
|
@@ -530,7 +1214,14 @@ def fused_experts(
|
|
530 |
dtype=hidden_states.dtype,
|
531 |
)
|
532 |
|
533 |
-
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
534 |
|
535 |
if inplace:
|
536 |
out_hidden_states = hidden_states
|
@@ -571,6 +1262,7 @@ def fused_experts(
|
|
571 |
intermediate_cache1,
|
572 |
a1_scale,
|
573 |
w1_scale,
|
|
|
574 |
curr_topk_weights,
|
575 |
curr_topk_ids,
|
576 |
sorted_token_ids,
|
@@ -582,6 +1274,8 @@ def fused_experts(
|
|
582 |
compute_type=compute_type,
|
583 |
use_fp8_w8a8=use_fp8_w8a8,
|
584 |
use_int8_w8a16=use_int8_w8a16,
|
|
|
|
|
585 |
)
|
586 |
|
587 |
ops.silu_and_mul(intermediate_cache2, intermediate_cache1.view(-1, N))
|
@@ -592,6 +1286,7 @@ def fused_experts(
|
|
592 |
intermediate_cache3,
|
593 |
a2_scale,
|
594 |
w2_scale,
|
|
|
595 |
curr_topk_weights,
|
596 |
curr_topk_ids,
|
597 |
sorted_token_ids,
|
@@ -603,6 +1298,8 @@ def fused_experts(
|
|
603 |
compute_type=compute_type,
|
604 |
use_fp8_w8a8=use_fp8_w8a8,
|
605 |
use_int8_w8a16=use_int8_w8a16,
|
|
|
|
|
606 |
)
|
607 |
|
608 |
ops.moe_sum(
|
@@ -620,17 +1317,20 @@ def fused_moe(
|
|
620 |
topk: int,
|
621 |
renormalize: bool,
|
622 |
inplace: bool = False,
|
623 |
-
override_config: Optional[Dict[str, Any]] = None,
|
624 |
use_grouped_topk: bool = False,
|
625 |
num_expert_group: Optional[int] = None,
|
626 |
topk_group: Optional[int] = None,
|
627 |
custom_routing_function: Optional[Callable] = None,
|
628 |
use_fp8_w8a8: bool = False,
|
629 |
use_int8_w8a16: bool = False,
|
|
|
630 |
w1_scale: Optional[torch.Tensor] = None,
|
631 |
w2_scale: Optional[torch.Tensor] = None,
|
|
|
|
|
632 |
a1_scale: Optional[torch.Tensor] = None,
|
633 |
a2_scale: Optional[torch.Tensor] = None,
|
|
|
634 |
) -> torch.Tensor:
|
635 |
"""
|
636 |
This function computes a Mixture of Experts (MoE) layer using two sets of
|
@@ -646,20 +1346,28 @@ def fused_moe(
|
|
646 |
- renormalize (bool): If True, renormalize the top-k weights to sum to 1.
|
647 |
- inplace (bool): If True, perform the operation in-place.
|
648 |
Defaults to False.
|
649 |
-
- override_config (Optional[Dict[str, Any]]): Optional override
|
650 |
-
for the kernel configuration.
|
651 |
- num_expert_group: Optional[int]: additional parameter for grouped_topk
|
652 |
- topk_group: Optional[int]: additional parameter for grouped_topk
|
653 |
- use_grouped_topk: If True, use grouped_topk instead of fused_topk
|
654 |
note: Deepseekv2 model uses grouped_topk
|
655 |
- use_fp8_w8a8 (bool): If True, use fp8 arithmetic to compute the inner
|
656 |
products for w1 and w2. Defaults to False.
|
657 |
-
- use_int8_w8a16 (bool): If True, use
|
658 |
-
products for w1 and w2.
|
|
|
|
|
|
|
|
|
659 |
- w1_scale (Optional[torch.Tensor]): Optional scale to be used for
|
660 |
w1.
|
661 |
- w2_scale (Optional[torch.Tensor]): Optional scale to be used for
|
662 |
w2.
|
|
|
|
|
|
|
|
|
|
|
|
|
663 |
|
664 |
Returns:
|
665 |
- torch.Tensor: The output tensor after applying the MoE layer.
|
@@ -693,11 +1401,14 @@ def fused_moe(
|
|
693 |
topk_weights,
|
694 |
topk_ids,
|
695 |
inplace=inplace,
|
696 |
-
override_config=override_config,
|
697 |
use_fp8_w8a8=use_fp8_w8a8,
|
698 |
use_int8_w8a16=use_int8_w8a16,
|
|
|
699 |
w1_scale=w1_scale,
|
700 |
w2_scale=w2_scale,
|
|
|
|
|
701 |
a1_scale=a1_scale,
|
702 |
a2_scale=a2_scale,
|
|
|
703 |
)
|
|
|
1 |
+
# SPDX-License-Identifier: Apache-2.0
|
2 |
"""Fused MoE kernel."""
|
3 |
|
4 |
import functools
|
5 |
import json
|
6 |
+
import logging
|
7 |
import os
|
8 |
+
from typing import Any, Callable, Dict, List, Optional, Tuple
|
9 |
|
10 |
import torch
|
11 |
import triton
|
12 |
import triton.language as tl
|
13 |
|
14 |
+
|
15 |
from ._ops import ops
|
16 |
+
from .fp8 import per_token_group_quant_fp8, scaled_fp8_quant
|
17 |
from .platforms import current_platform
|
18 |
|
19 |
+
logger = logging.getLogger(__name__)
|
20 |
+
|
21 |
+
|
22 |
VLLM_FUSED_MOE_CHUNK_SIZE = int(os.getenv("VLLM_FUSED_MOE_CHUNK_SIZE", "32768"))
|
23 |
|
24 |
|
25 |
+
@triton.jit
|
26 |
+
def fused_moe_kernel_gptq_awq(
|
27 |
+
# Pointers to matrices
|
28 |
+
a_ptr,
|
29 |
+
b_ptr,
|
30 |
+
c_ptr,
|
31 |
+
b_scale_ptr,
|
32 |
+
b_zp_ptr,
|
33 |
+
topk_weights_ptr,
|
34 |
+
sorted_token_ids_ptr,
|
35 |
+
expert_ids_ptr,
|
36 |
+
num_tokens_post_padded_ptr,
|
37 |
+
# Matrix dimensions
|
38 |
+
N: tl.constexpr,
|
39 |
+
K: tl.constexpr,
|
40 |
+
EM,
|
41 |
+
num_valid_tokens,
|
42 |
+
# The stride variables represent how much to increase the ptr by when
|
43 |
+
# moving by 1 element in a particular dimension. E.g. `stride_am` is
|
44 |
+
# how much to increase `a_ptr` by to get the element one row down
|
45 |
+
# (A has M rows).
|
46 |
+
stride_am,
|
47 |
+
stride_ak,
|
48 |
+
stride_be,
|
49 |
+
stride_bk,
|
50 |
+
stride_bn,
|
51 |
+
stride_cm,
|
52 |
+
stride_cn,
|
53 |
+
stride_bse,
|
54 |
+
stride_bsk,
|
55 |
+
stride_bsn,
|
56 |
+
stride_bze,
|
57 |
+
stride_bzk,
|
58 |
+
stride_bzn,
|
59 |
+
block_k_diviable: tl.constexpr,
|
60 |
+
group_size: tl.constexpr,
|
61 |
+
# Meta-parameters
|
62 |
+
BLOCK_SIZE_M: tl.constexpr,
|
63 |
+
BLOCK_SIZE_N: tl.constexpr,
|
64 |
+
BLOCK_SIZE_K: tl.constexpr,
|
65 |
+
GROUP_SIZE_M: tl.constexpr,
|
66 |
+
MUL_ROUTED_WEIGHT: tl.constexpr,
|
67 |
+
top_k: tl.constexpr,
|
68 |
+
compute_type: tl.constexpr,
|
69 |
+
has_zp: tl.constexpr,
|
70 |
+
use_int4_w4a16: tl.constexpr,
|
71 |
+
use_int8_w8a16: tl.constexpr,
|
72 |
+
):
|
73 |
+
"""
|
74 |
+
Implements the fused computation for a Mixture of Experts (MOE) using
|
75 |
+
token and expert matrices.
|
76 |
+
|
77 |
+
Key Parameters:
|
78 |
+
- A: The input tensor representing tokens with shape (*, K), where '*' can
|
79 |
+
be any shape representing batches and K is the feature dimension of
|
80 |
+
each token.
|
81 |
+
- B: The stacked MOE weight tensor with shape (E, N, K), where E is
|
82 |
+
the number of experts, K is the input feature dimension, and N is
|
83 |
+
the output feature dimension.
|
84 |
+
- C: The output cache tensor with shape (M, topk, N), where M is the
|
85 |
+
total number of tokens post padding, topk is the number of times
|
86 |
+
each token is repeated, and N is the output feature dimension.
|
87 |
+
- sorted_token_ids: A tensor containing the sorted indices of tokens,
|
88 |
+
repeated topk times and arranged by the expert index they are
|
89 |
+
assigned to.
|
90 |
+
- expert_ids: A tensor containing the indices of the expert for each
|
91 |
+
block. It determines which expert matrix from B should be used for
|
92 |
+
each block in A.
|
93 |
+
This kernel performs the multiplication of a token by its corresponding
|
94 |
+
expert matrix as determined by `expert_ids`. The sorting of
|
95 |
+
`sorted_token_ids` by expert index and padding ensures divisibility by
|
96 |
+
BLOCK_SIZE_M, which is necessary to maintain consistency in block matrix
|
97 |
+
multiplication across different blocks processed by the same expert.
|
98 |
+
"""
|
99 |
+
# -----------------------------------------------------------
|
100 |
+
# Map program ids `pid` to the block of C it should compute.
|
101 |
+
# This is done in a grouped ordering to promote L2 data reuse.
|
102 |
+
pid = tl.program_id(axis=0)
|
103 |
+
num_pid_m = tl.cdiv(EM, BLOCK_SIZE_M)
|
104 |
+
num_pid_n = tl.cdiv(N, BLOCK_SIZE_N)
|
105 |
+
num_pid_in_group = GROUP_SIZE_M * num_pid_n
|
106 |
+
group_id = pid // num_pid_in_group
|
107 |
+
first_pid_m = group_id * GROUP_SIZE_M
|
108 |
+
group_size_m = min(num_pid_m - first_pid_m, GROUP_SIZE_M)
|
109 |
+
pid_m = first_pid_m + ((pid % num_pid_in_group) % group_size_m)
|
110 |
+
pid_n = (pid % num_pid_in_group) // group_size_m
|
111 |
+
|
112 |
+
# ----------------------------------------------------------
|
113 |
+
# Create pointers for the first blocks of A and B.
|
114 |
+
# We will advance this pointer as we move in the K direction
|
115 |
+
# and accumulate
|
116 |
+
# `a_ptrs` is a block of [BLOCK_SIZE_M, BLOCK_SIZE_K] pointers
|
117 |
+
# `b_ptrs` is a block of [BLOCK_SIZE_K, BLOCK_SIZE_N] pointers
|
118 |
+
num_tokens_post_padded = tl.load(num_tokens_post_padded_ptr)
|
119 |
+
if pid_m * BLOCK_SIZE_M >= num_tokens_post_padded:
|
120 |
+
return
|
121 |
+
offs_token_id = pid_m * BLOCK_SIZE_M + tl.arange(0, BLOCK_SIZE_M).to(tl.int64)
|
122 |
+
offs_token = tl.load(sorted_token_ids_ptr + offs_token_id)
|
123 |
+
token_mask = offs_token < num_valid_tokens
|
124 |
+
|
125 |
+
offs_bn = (pid_n * BLOCK_SIZE_N + tl.arange(0, BLOCK_SIZE_N).to(tl.int64)) % N
|
126 |
+
offs_k = tl.arange(0, BLOCK_SIZE_K)
|
127 |
+
a_ptrs = a_ptr + (
|
128 |
+
offs_token[:, None] // top_k * stride_am + offs_k[None, :] * stride_ak
|
129 |
+
)
|
130 |
+
|
131 |
+
off_experts = tl.load(expert_ids_ptr + pid_m).to(tl.int64)
|
132 |
+
|
133 |
+
if use_int4_w4a16:
|
134 |
+
b_ptrs = (
|
135 |
+
b_ptr
|
136 |
+
+ off_experts * stride_be
|
137 |
+
+ (offs_k[:, None] // 2) * stride_bk
|
138 |
+
+ offs_bn[None, :] * stride_bn
|
139 |
+
)
|
140 |
+
b_shifter = (offs_k[:, None] % 2) * 4
|
141 |
+
elif use_int8_w8a16:
|
142 |
+
b_ptrs = (
|
143 |
+
b_ptr
|
144 |
+
+ off_experts * stride_be
|
145 |
+
+ offs_k[:, None] * stride_bk
|
146 |
+
+ offs_bn[None, :] * stride_bn
|
147 |
+
)
|
148 |
+
|
149 |
+
if not has_zp and use_int4_w4a16:
|
150 |
+
b_zp_num = 8
|
151 |
+
if not has_zp and use_int8_w8a16:
|
152 |
+
b_zp_num = 128
|
153 |
+
elif has_zp and use_int4_w4a16:
|
154 |
+
b_zp_shifter = (offs_bn[None, :] % 2) * 4
|
155 |
+
|
156 |
+
# -----------------------------------------------------------
|
157 |
+
# Iterate to compute a block of the C matrix.
|
158 |
+
# We accumulate into a `[BLOCK_SIZE_M, BLOCK_SIZE_N]` block
|
159 |
+
# of fp32 values for higher accuracy.
|
160 |
+
# `accumulator` will be converted back to fp16 after the loop.
|
161 |
+
accumulator = tl.zeros((BLOCK_SIZE_M, BLOCK_SIZE_N), dtype=tl.float32)
|
162 |
+
for k in range(0, tl.cdiv(K, BLOCK_SIZE_K)):
|
163 |
+
# Load the next block of A and B, generate a mask by checking the
|
164 |
+
# K dimension.
|
165 |
+
|
166 |
+
if not block_k_diviable:
|
167 |
+
k_mask = offs_k[:, None] < K - k * BLOCK_SIZE_K
|
168 |
+
k_other = 0.0
|
169 |
+
else:
|
170 |
+
k_mask = None
|
171 |
+
k_other = None
|
172 |
+
|
173 |
+
a = tl.load(
|
174 |
+
a_ptrs,
|
175 |
+
mask=token_mask[:, None] & (offs_k[None, :] < K - k * BLOCK_SIZE_K),
|
176 |
+
other=0.0,
|
177 |
+
)
|
178 |
+
b = tl.load(b_ptrs)
|
179 |
+
if use_int4_w4a16:
|
180 |
+
b = (b >> b_shifter) & 0xF
|
181 |
+
|
182 |
+
b_scale_ptrs = (
|
183 |
+
b_scale_ptr
|
184 |
+
+ off_experts * stride_bse
|
185 |
+
+ offs_bn[None, :] * stride_bsn
|
186 |
+
+ ((offs_k[:, None] + BLOCK_SIZE_K * k) // group_size) * stride_bsk
|
187 |
+
)
|
188 |
+
b_scale = tl.load(b_scale_ptrs, mask=k_mask, other=k_other)
|
189 |
+
b_scale = b_scale.to(tl.float32)
|
190 |
+
|
191 |
+
if has_zp and use_int4_w4a16:
|
192 |
+
offs_k_true = (offs_k[:, None] + BLOCK_SIZE_K * k) // group_size
|
193 |
+
b_zp_ptrs = (
|
194 |
+
b_zp_ptr
|
195 |
+
+ off_experts * stride_bze
|
196 |
+
+ (offs_bn[None, :] // 2) * stride_bzn
|
197 |
+
+ offs_k_true * stride_bzk
|
198 |
+
)
|
199 |
+
b_zp = tl.load(b_zp_ptrs, mask=k_mask, other=k_other)
|
200 |
+
b_zp = (b_zp >> b_zp_shifter) & 0xF
|
201 |
+
b_zp = b_zp.to(tl.float32)
|
202 |
+
elif has_zp and use_int8_w8a16:
|
203 |
+
offs_k_true = (offs_k[:, None] + BLOCK_SIZE_K * k) // group_size
|
204 |
+
b_zp_ptrs = (
|
205 |
+
b_zp_ptr
|
206 |
+
+ off_experts * stride_bze
|
207 |
+
+ offs_bn[None, :] * stride_bzn
|
208 |
+
+ offs_k_true * stride_bzk
|
209 |
+
)
|
210 |
+
b_zp = tl.load(b_zp_ptrs, mask=k_mask, other=k_other)
|
211 |
+
b_zp = b_zp.to(tl.float32)
|
212 |
+
|
213 |
+
# We accumulate along the K dimension.
|
214 |
+
if has_zp:
|
215 |
+
b = ((b.to(tl.float32) - b_zp) * b_scale).to(compute_type)
|
216 |
+
else:
|
217 |
+
b = ((b.to(tl.float32) - b_zp_num) * b_scale).to(compute_type)
|
218 |
+
accumulator = tl.dot(a, b, acc=accumulator)
|
219 |
+
|
220 |
+
# Advance the ptrs to the next K block.
|
221 |
+
a_ptrs += BLOCK_SIZE_K * stride_ak
|
222 |
+
if use_int4_w4a16:
|
223 |
+
b_ptrs += (BLOCK_SIZE_K // 2) * stride_bk
|
224 |
+
else:
|
225 |
+
b_ptrs += BLOCK_SIZE_K * stride_bk
|
226 |
+
|
227 |
+
if MUL_ROUTED_WEIGHT:
|
228 |
+
moe_weight = tl.load(topk_weights_ptr + offs_token, mask=token_mask, other=0)
|
229 |
+
accumulator = accumulator * moe_weight[:, None]
|
230 |
+
|
231 |
+
accumulator = accumulator.to(compute_type)
|
232 |
+
# -----------------------------------------------------------
|
233 |
+
# Write back the block of the output
|
234 |
+
offs_cn = pid_n * BLOCK_SIZE_N + tl.arange(0, BLOCK_SIZE_N)
|
235 |
+
c_ptrs = c_ptr + stride_cm * offs_token[:, None] + stride_cn * offs_cn[None, :]
|
236 |
+
c_mask = token_mask[:, None] & (offs_cn[None, :] < N)
|
237 |
+
tl.store(c_ptrs, accumulator, mask=c_mask)
|
238 |
+
|
239 |
+
|
240 |
@triton.jit
|
241 |
def fused_moe_kernel(
|
242 |
# Pointers to matrices
|
|
|
265 |
stride_bn,
|
266 |
stride_cm,
|
267 |
stride_cn,
|
268 |
+
stride_asm,
|
269 |
+
stride_ask,
|
270 |
stride_bse,
|
271 |
+
stride_bsk,
|
272 |
stride_bsn,
|
273 |
+
# Block size for block-wise quantization
|
274 |
+
group_n: tl.constexpr,
|
275 |
+
group_k: tl.constexpr,
|
276 |
# Meta-parameters
|
277 |
BLOCK_SIZE_M: tl.constexpr,
|
278 |
BLOCK_SIZE_N: tl.constexpr,
|
|
|
332 |
num_tokens_post_padded = tl.load(num_tokens_post_padded_ptr)
|
333 |
if pid_m * BLOCK_SIZE_M >= num_tokens_post_padded:
|
334 |
return
|
335 |
+
offs_token_id = pid_m * BLOCK_SIZE_M + tl.arange(0, BLOCK_SIZE_M).to(tl.int64)
|
336 |
offs_token = tl.load(sorted_token_ids_ptr + offs_token_id)
|
337 |
token_mask = offs_token < num_valid_tokens
|
338 |
|
339 |
+
offs_bn = (pid_n * BLOCK_SIZE_N + tl.arange(0, BLOCK_SIZE_N).to(tl.int64)) % N
|
340 |
offs_k = tl.arange(0, BLOCK_SIZE_K)
|
341 |
a_ptrs = a_ptr + (
|
342 |
offs_token[:, None] // top_k * stride_am + offs_k[None, :] * stride_ak
|
343 |
)
|
344 |
|
345 |
+
off_experts = tl.load(expert_ids_ptr + pid_m).to(tl.int64)
|
346 |
b_ptrs = (
|
347 |
b_ptr
|
348 |
+ off_experts * stride_be
|
|
|
355 |
b_scale = tl.load(b_scale_ptrs)
|
356 |
|
357 |
if use_fp8_w8a8:
|
358 |
+
if group_k > 0 and group_n > 0:
|
359 |
+
a_scale_ptrs = a_scale_ptr + (offs_token // top_k) * stride_asm
|
360 |
+
offs_bsn = offs_bn // group_n
|
361 |
+
b_scale_ptrs = (
|
362 |
+
b_scale_ptr + off_experts * stride_bse + offs_bsn * stride_bsn
|
363 |
+
)
|
364 |
+
else:
|
365 |
+
a_scale = tl.load(a_scale_ptr)
|
366 |
+
b_scale = tl.load(b_scale_ptr + off_experts)
|
367 |
|
368 |
# -----------------------------------------------------------
|
369 |
# Iterate to compute a block of the C matrix.
|
|
|
385 |
if use_int8_w8a16:
|
386 |
accumulator = tl.dot(a, b.to(compute_type), acc=accumulator)
|
387 |
elif use_fp8_w8a8:
|
388 |
+
if group_k > 0 and group_n > 0:
|
389 |
+
k_start = k * BLOCK_SIZE_K
|
390 |
+
offs_ks = k_start // group_k
|
391 |
+
a_scale = tl.load(
|
392 |
+
a_scale_ptrs + offs_ks * stride_ask, mask=token_mask, other=0.0
|
393 |
+
)
|
394 |
+
b_scale = tl.load(b_scale_ptrs + offs_ks * stride_bsk)
|
395 |
+
|
396 |
+
accumulator += tl.dot(a, b) * a_scale[:, None] * b_scale[None, :]
|
397 |
+
else:
|
398 |
+
accumulator = tl.dot(a, b, acc=accumulator)
|
399 |
else:
|
400 |
accumulator += tl.dot(a, b)
|
401 |
# Advance the ptrs to the next K block.
|
|
|
408 |
if use_int8_w8a16:
|
409 |
accumulator = (accumulator * b_scale).to(compute_type)
|
410 |
elif use_fp8_w8a8:
|
411 |
+
if group_k > 0 and group_n > 0:
|
412 |
+
accumulator = accumulator.to(compute_type)
|
413 |
+
else:
|
414 |
+
accumulator = (accumulator * a_scale * b_scale).to(compute_type)
|
415 |
else:
|
416 |
accumulator = accumulator.to(compute_type)
|
417 |
# -----------------------------------------------------------
|
|
|
422 |
tl.store(c_ptrs, accumulator, mask=c_mask)
|
423 |
|
424 |
|
425 |
+
def ceil_div(a, b):
|
426 |
+
return (a + b - 1) // b
|
427 |
+
|
428 |
+
|
429 |
+
@triton.jit
|
430 |
+
def moe_align_block_size_stage1(
|
431 |
+
topk_ids_ptr,
|
432 |
+
tokens_cnts_ptr,
|
433 |
+
num_experts: tl.constexpr,
|
434 |
+
numel: tl.constexpr,
|
435 |
+
tokens_per_thread: tl.constexpr,
|
436 |
+
):
|
437 |
+
pid = tl.program_id(0)
|
438 |
+
|
439 |
+
start_idx = pid * tokens_per_thread
|
440 |
+
|
441 |
+
off_c = (pid + 1) * num_experts
|
442 |
+
|
443 |
+
for i in range(tokens_per_thread):
|
444 |
+
if start_idx + i < numel:
|
445 |
+
idx = tl.load(topk_ids_ptr + start_idx + i)
|
446 |
+
token_cnt = tl.load(tokens_cnts_ptr + off_c + idx)
|
447 |
+
tl.store(tokens_cnts_ptr + off_c + idx, token_cnt + 1)
|
448 |
+
|
449 |
+
|
450 |
+
@triton.jit
|
451 |
+
def moe_align_block_size_stage2(
|
452 |
+
tokens_cnts_ptr,
|
453 |
+
num_experts: tl.constexpr,
|
454 |
+
):
|
455 |
+
pid = tl.program_id(0)
|
456 |
+
|
457 |
+
last_cnt = 0
|
458 |
+
for i in range(1, num_experts + 1):
|
459 |
+
token_cnt = tl.load(tokens_cnts_ptr + i * num_experts + pid)
|
460 |
+
last_cnt = last_cnt + token_cnt
|
461 |
+
tl.store(tokens_cnts_ptr + i * num_experts + pid, last_cnt)
|
462 |
+
|
463 |
+
|
464 |
+
@triton.jit
|
465 |
+
def moe_align_block_size_stage3(
|
466 |
+
total_tokens_post_pad_ptr,
|
467 |
+
tokens_cnts_ptr,
|
468 |
+
cumsum_ptr,
|
469 |
+
num_experts: tl.constexpr,
|
470 |
+
block_size: tl.constexpr,
|
471 |
+
):
|
472 |
+
last_cumsum = 0
|
473 |
+
off_cnt = num_experts * num_experts
|
474 |
+
for i in range(1, num_experts + 1):
|
475 |
+
token_cnt = tl.load(tokens_cnts_ptr + off_cnt + i - 1)
|
476 |
+
last_cumsum = last_cumsum + tl.cdiv(token_cnt, block_size) * block_size
|
477 |
+
tl.store(cumsum_ptr + i, last_cumsum)
|
478 |
+
tl.store(total_tokens_post_pad_ptr, last_cumsum)
|
479 |
+
|
480 |
+
|
481 |
+
@triton.jit
|
482 |
+
def moe_align_block_size_stage4(
|
483 |
+
topk_ids_ptr,
|
484 |
+
sorted_token_ids_ptr,
|
485 |
+
expert_ids_ptr,
|
486 |
+
tokens_cnts_ptr,
|
487 |
+
cumsum_ptr,
|
488 |
+
num_experts: tl.constexpr,
|
489 |
+
block_size: tl.constexpr,
|
490 |
+
numel: tl.constexpr,
|
491 |
+
tokens_per_thread: tl.constexpr,
|
492 |
+
):
|
493 |
+
pid = tl.program_id(0)
|
494 |
+
start_idx = tl.load(cumsum_ptr + pid)
|
495 |
+
end_idx = tl.load(cumsum_ptr + pid + 1)
|
496 |
+
|
497 |
+
for i in range(start_idx, end_idx, block_size):
|
498 |
+
tl.store(expert_ids_ptr + i // block_size, pid)
|
499 |
+
|
500 |
+
start_idx = pid * tokens_per_thread
|
501 |
+
off_t = pid * num_experts
|
502 |
+
|
503 |
+
for i in range(start_idx, tl.minimum(start_idx + tokens_per_thread, numel)):
|
504 |
+
expert_id = tl.load(topk_ids_ptr + i)
|
505 |
+
token_cnt = tl.load(tokens_cnts_ptr + off_t + expert_id)
|
506 |
+
rank_post_pad = token_cnt + tl.load(cumsum_ptr + expert_id)
|
507 |
+
tl.store(sorted_token_ids_ptr + rank_post_pad, i)
|
508 |
+
tl.store(tokens_cnts_ptr + off_t + expert_id, token_cnt + 1)
|
509 |
+
|
510 |
+
|
511 |
+
# Triton implementation based on:
|
512 |
+
# https://github.com/sgl-project/sglang/commit/ba5112ff691d791a9e38c6c71f59324a5fcb49d0
|
513 |
+
def moe_align_block_size_triton(
|
514 |
+
topk_ids: torch.Tensor,
|
515 |
+
num_experts: int,
|
516 |
+
block_size: int,
|
517 |
+
sorted_token_ids: torch.Tensor,
|
518 |
+
expert_ids: torch.Tensor,
|
519 |
+
num_tokens_post_pad: torch.Tensor,
|
520 |
+
) -> None:
|
521 |
+
numel = topk_ids.numel()
|
522 |
+
grid = (num_experts,)
|
523 |
+
tokens_cnts = torch.zeros(
|
524 |
+
(num_experts + 1, num_experts), dtype=torch.int32, device=topk_ids.device
|
525 |
+
)
|
526 |
+
cumsum = torch.zeros((num_experts + 1,), dtype=torch.int32, device=topk_ids.device)
|
527 |
+
tokens_per_thread = ceil_div(numel, num_experts)
|
528 |
+
|
529 |
+
moe_align_block_size_stage1[grid](
|
530 |
+
topk_ids,
|
531 |
+
tokens_cnts,
|
532 |
+
num_experts,
|
533 |
+
numel,
|
534 |
+
tokens_per_thread,
|
535 |
+
)
|
536 |
+
moe_align_block_size_stage2[grid](
|
537 |
+
tokens_cnts,
|
538 |
+
num_experts,
|
539 |
+
)
|
540 |
+
moe_align_block_size_stage3[(1,)](
|
541 |
+
num_tokens_post_pad,
|
542 |
+
tokens_cnts,
|
543 |
+
cumsum,
|
544 |
+
num_experts,
|
545 |
+
block_size,
|
546 |
+
)
|
547 |
+
moe_align_block_size_stage4[grid](
|
548 |
+
topk_ids,
|
549 |
+
sorted_token_ids,
|
550 |
+
expert_ids,
|
551 |
+
tokens_cnts,
|
552 |
+
cumsum,
|
553 |
+
num_experts,
|
554 |
+
block_size,
|
555 |
+
numel,
|
556 |
+
tokens_per_thread,
|
557 |
+
)
|
558 |
+
|
559 |
+
|
560 |
def moe_align_block_size(
|
561 |
topk_ids: torch.Tensor, block_size: int, num_experts: int
|
562 |
) -> Tuple[torch.Tensor, torch.Tensor, torch.Tensor]:
|
|
|
607 |
(max_num_m_blocks,), dtype=torch.int32, device=topk_ids.device
|
608 |
)
|
609 |
num_tokens_post_pad = torch.empty((1), dtype=torch.int32, device=topk_ids.device)
|
610 |
+
if num_experts >= 224:
|
611 |
+
if VLLM_ENABLE_MOE_ALIGN_BLOCK_SIZE_TRITON:
|
612 |
+
moe_align_block_size_triton(
|
613 |
+
topk_ids,
|
614 |
+
num_experts,
|
615 |
+
block_size,
|
616 |
+
sorted_ids,
|
617 |
+
expert_ids,
|
618 |
+
num_tokens_post_pad,
|
619 |
+
)
|
620 |
+
else:
|
621 |
+
ops.sgl_moe_align_block_size(
|
622 |
+
topk_ids,
|
623 |
+
num_experts,
|
624 |
+
block_size,
|
625 |
+
sorted_ids,
|
626 |
+
expert_ids,
|
627 |
+
num_tokens_post_pad,
|
628 |
+
)
|
629 |
+
else:
|
630 |
+
ops.moe_align_block_size(
|
631 |
+
topk_ids,
|
632 |
+
num_experts,
|
633 |
+
block_size,
|
634 |
+
sorted_ids,
|
635 |
+
expert_ids,
|
636 |
+
num_tokens_post_pad,
|
637 |
+
)
|
638 |
return sorted_ids, expert_ids, num_tokens_post_pad
|
639 |
|
640 |
|
|
|
644 |
C: torch.Tensor,
|
645 |
A_scale: Optional[torch.Tensor],
|
646 |
B_scale: Optional[torch.Tensor],
|
647 |
+
B_zp: Optional[torch.Tensor],
|
648 |
topk_weights: torch.Tensor,
|
649 |
topk_ids: torch.Tensor,
|
650 |
sorted_token_ids: torch.Tensor,
|
|
|
656 |
compute_type: tl.dtype,
|
657 |
use_fp8_w8a8: bool,
|
658 |
use_int8_w8a16: bool,
|
659 |
+
use_int4_w4a16: bool,
|
660 |
+
block_shape: Optional[List[int]] = None,
|
661 |
) -> None:
|
662 |
assert topk_weights.stride(1) == 1
|
663 |
assert sorted_token_ids.stride(0) == 1
|
664 |
|
665 |
if use_fp8_w8a8:
|
|
|
666 |
assert B_scale is not None
|
667 |
+
if block_shape is None:
|
668 |
+
A, A_scale = scaled_fp8_quant(A, A_scale)
|
669 |
+
else:
|
670 |
+
assert len(block_shape) == 2
|
671 |
+
block_n, block_k = block_shape[0], block_shape[1]
|
672 |
+
A, A_scale = per_token_group_quant_fp8(A, block_k)
|
673 |
+
assert triton.cdiv(A.shape[-1], block_k) == A_scale.shape[-1]
|
674 |
+
assert triton.cdiv(B.shape[-2], block_n) == B_scale.shape[-2]
|
675 |
+
assert triton.cdiv(B.shape[-1], block_k) == B_scale.shape[-1]
|
676 |
+
elif use_int8_w8a16 or use_int4_w4a16:
|
677 |
assert B_scale is not None
|
678 |
+
assert block_shape is None or block_shape[0] == 0
|
679 |
else:
|
680 |
assert A_scale is None
|
681 |
assert B_scale is None
|
682 |
|
683 |
+
EM = sorted_token_ids.shape[0]
|
684 |
+
if A.shape[0] < config["BLOCK_SIZE_M"]:
|
685 |
+
# optimize for small batch_size.
|
686 |
+
# We assume that top_ids of each token is unique, so
|
687 |
+
# so num_valid_experts <= batch_size <= BLOCK_SIZE_M,
|
688 |
+
# and we can skip some invalid blocks.
|
689 |
+
EM = min(sorted_token_ids.shape[0], A.shape[0] * top_k * config["BLOCK_SIZE_M"])
|
690 |
grid = lambda META: (
|
691 |
+
triton.cdiv(EM, META["BLOCK_SIZE_M"])
|
692 |
* triton.cdiv(B.shape[1], META["BLOCK_SIZE_N"]),
|
693 |
)
|
694 |
|
695 |
+
if (
|
696 |
+
(use_int8_w8a16 or use_int4_w4a16)
|
697 |
+
and block_shape is not None
|
698 |
+
and block_shape[1] > 0
|
699 |
+
):
|
700 |
+
assert B_scale is not None and B_scale.ndim == 3
|
701 |
+
assert B_zp is None or B_zp.ndim == 3
|
702 |
+
|
703 |
+
fused_moe_kernel_gptq_awq[grid](
|
704 |
+
A,
|
705 |
+
B,
|
706 |
+
C,
|
707 |
+
B_scale,
|
708 |
+
B_zp,
|
709 |
+
topk_weights,
|
710 |
+
sorted_token_ids,
|
711 |
+
expert_ids,
|
712 |
+
num_tokens_post_padded,
|
713 |
+
B.shape[1],
|
714 |
+
A.shape[1],
|
715 |
+
EM,
|
716 |
+
topk_ids.numel(),
|
717 |
+
A.stride(0),
|
718 |
+
A.stride(1),
|
719 |
+
B.stride(0),
|
720 |
+
B.stride(2),
|
721 |
+
B.stride(1),
|
722 |
+
C.stride(1),
|
723 |
+
C.stride(2),
|
724 |
+
B_scale.stride(0),
|
725 |
+
B_scale.stride(2),
|
726 |
+
B_scale.stride(1),
|
727 |
+
B_zp.stride(0) if B_zp is not None else 0,
|
728 |
+
B_zp.stride(2) if B_zp is not None else 0,
|
729 |
+
B_zp.stride(1) if B_zp is not None else 0,
|
730 |
+
block_k_diviable=A.shape[1] % config["BLOCK_SIZE_K"] == 0,
|
731 |
+
group_size=block_shape[1],
|
732 |
+
MUL_ROUTED_WEIGHT=mul_routed_weight,
|
733 |
+
top_k=top_k,
|
734 |
+
compute_type=compute_type,
|
735 |
+
has_zp=B_zp is not None,
|
736 |
+
use_int4_w4a16=use_int4_w4a16,
|
737 |
+
use_int8_w8a16=use_int8_w8a16,
|
738 |
+
**config,
|
739 |
+
)
|
740 |
+
|
741 |
+
else:
|
742 |
+
fused_moe_kernel[grid](
|
743 |
+
A,
|
744 |
+
B,
|
745 |
+
C,
|
746 |
+
A_scale,
|
747 |
+
B_scale,
|
748 |
+
topk_weights,
|
749 |
+
sorted_token_ids,
|
750 |
+
expert_ids,
|
751 |
+
num_tokens_post_padded,
|
752 |
+
B.shape[1],
|
753 |
+
A.shape[1],
|
754 |
+
EM,
|
755 |
+
topk_ids.numel(),
|
756 |
+
A.stride(0),
|
757 |
+
A.stride(1),
|
758 |
+
B.stride(0),
|
759 |
+
B.stride(2),
|
760 |
+
B.stride(1),
|
761 |
+
C.stride(1),
|
762 |
+
C.stride(2),
|
763 |
+
A_scale.stride(0) if A_scale is not None and A_scale.ndim == 2 else 0,
|
764 |
+
A_scale.stride(1) if A_scale is not None and A_scale.ndim == 2 else 0,
|
765 |
+
B_scale.stride(0) if B_scale is not None and B_scale.ndim >= 2 else 0,
|
766 |
+
B_scale.stride(2) if B_scale is not None and B_scale.ndim == 3 else 0,
|
767 |
+
B_scale.stride(1) if B_scale is not None and B_scale.ndim >= 2 else 0,
|
768 |
+
0 if block_shape is None else block_shape[0],
|
769 |
+
0 if block_shape is None else block_shape[1],
|
770 |
+
MUL_ROUTED_WEIGHT=mul_routed_weight,
|
771 |
+
top_k=top_k,
|
772 |
+
compute_type=compute_type,
|
773 |
+
use_fp8_w8a8=use_fp8_w8a8,
|
774 |
+
use_int8_w8a16=use_int8_w8a16,
|
775 |
+
**config,
|
776 |
+
)
|
777 |
|
778 |
|
779 |
+
# Adapted from: https://github.com/sgl-project/sglang/pull/2628
|
780 |
+
def get_config_file_name(
|
781 |
+
E: int, N: int, dtype: Optional[str], block_shape: Optional[List[int]] = None
|
782 |
+
) -> str:
|
783 |
device_name = current_platform.get_device_name().replace(" ", "_")
|
784 |
dtype_selector = "" if not dtype else f",dtype={dtype}"
|
785 |
+
block_shape_selector = (
|
786 |
+
"" if not block_shape or not all(block_shape) else f",block_shape={block_shape}"
|
787 |
+
)
|
788 |
+
return f"E={E},N={N},device_name={device_name}{dtype_selector}{block_shape_selector}.json" # noqa: E501
|
789 |
|
790 |
|
791 |
+
# Adapted from: https://github.com/sgl-project/sglang/pull/2628
|
792 |
@functools.lru_cache
|
793 |
+
def get_moe_configs(
|
794 |
+
E: int,
|
795 |
+
N: int,
|
796 |
+
dtype: Optional[str],
|
797 |
+
block_n: Optional[int] = None,
|
798 |
+
block_k: Optional[int] = None,
|
799 |
+
) -> Optional[Dict[int, Any]]:
|
800 |
"""
|
801 |
Return optimized configurations for the fused MoE kernel.
|
802 |
|
|
|
808 |
|
809 |
# First look up if an optimized configuration is available in the configs
|
810 |
# directory
|
811 |
+
block_shape = [block_n, block_k] if block_n and block_k else None
|
812 |
+
json_file_name = get_config_file_name(E, N, dtype, block_shape)
|
813 |
|
814 |
config_file_path = os.path.join(
|
815 |
os.path.dirname(os.path.realpath(__file__)), "configs", json_file_name
|
816 |
)
|
817 |
if os.path.exists(config_file_path):
|
818 |
with open(config_file_path) as f:
|
819 |
+
logger.info("Using configuration from %s for MoE layer.", config_file_path)
|
820 |
# If a configuration has been found, return it
|
821 |
return {int(key): val for key, val in json.load(f).items()}
|
822 |
|
823 |
# If no optimized configuration is available, we will use the default
|
824 |
# configuration
|
825 |
+
logger.warning(
|
826 |
+
(
|
827 |
+
"Using default MoE config. Performance might be sub-optimal! "
|
828 |
+
"Config file not found at %s"
|
829 |
+
),
|
830 |
+
config_file_path,
|
831 |
+
)
|
832 |
return None
|
833 |
|
834 |
|
|
|
840 |
topk: int,
|
841 |
dtype: Optional[str],
|
842 |
is_marlin: bool,
|
843 |
+
block_shape: Optional[List[int]] = None,
|
844 |
) -> Dict[str, int]:
|
845 |
+
if dtype == "fp8_w8a8" and block_shape is not None:
|
846 |
+
# Block-wise quant: BLOCK_SIZE_N must be divisible by block_shape[0]
|
847 |
+
# BLOCK_SIZE_K must be divisible by block_shape[1]
|
|
|
|
|
|
|
|
|
|
|
848 |
config = {
|
849 |
+
"BLOCK_SIZE_M": 64,
|
850 |
+
"BLOCK_SIZE_N": block_shape[0],
|
851 |
+
"BLOCK_SIZE_K": block_shape[1],
|
852 |
+
"GROUP_SIZE_M": 32,
|
853 |
+
"num_warps": 4,
|
854 |
+
"num_stages": 3,
|
855 |
}
|
856 |
+
else:
|
857 |
+
config = {
|
858 |
+
"BLOCK_SIZE_M": 64,
|
859 |
+
"BLOCK_SIZE_N": 64,
|
860 |
+
"BLOCK_SIZE_K": 32,
|
861 |
+
"GROUP_SIZE_M": 8,
|
862 |
+
}
|
863 |
+
# A heuristic: fused marlin works faster with this config for small M
|
864 |
+
if M <= E or (is_marlin and M <= 32):
|
865 |
+
config = {
|
866 |
+
"BLOCK_SIZE_M": 16,
|
867 |
+
"BLOCK_SIZE_N": 32,
|
868 |
+
"BLOCK_SIZE_K": 64,
|
869 |
+
"GROUP_SIZE_M": 1,
|
870 |
+
}
|
871 |
return config
|
872 |
|
873 |
|
|
|
877 |
top_k: int,
|
878 |
dtype: Optional[str],
|
879 |
M: int,
|
|
|
880 |
is_marlin: bool = False,
|
881 |
+
block_shape: Optional[List[int]] = None,
|
882 |
):
|
883 |
+
# from vllm.model_executor.layers.fused_moe import get_config
|
884 |
+
# TODO: removed when syncing to vLLM, do we need this?
|
885 |
+
# override_config = get_config()
|
886 |
+
override_config = None
|
887 |
if override_config:
|
888 |
config = override_config
|
889 |
else:
|
890 |
# First try to load optimal config from the file
|
891 |
E, _, N = w2_shape
|
892 |
+
block_n = block_shape[0] if block_shape else 0
|
893 |
+
block_k = block_shape[1] if block_shape else 0
|
894 |
+
configs = get_moe_configs(E, N, dtype, block_n, block_k)
|
895 |
|
896 |
if configs:
|
897 |
# If an optimal configuration map has been found, look up the
|
|
|
899 |
config = configs[min(configs.keys(), key=lambda x: abs(x - M))]
|
900 |
else:
|
901 |
# Else use the default config
|
902 |
+
config = get_default_config(
|
903 |
+
M, E, N, w1_shape[2], top_k, dtype, is_marlin, block_shape
|
904 |
+
)
|
905 |
return config
|
906 |
|
907 |
|
|
|
937 |
return topk_weights, topk_ids
|
938 |
|
939 |
|
940 |
+
# This is used by the Deepseek-V2 and Deepseek-V3 model
|
941 |
+
@torch.compile(dynamic=True, backend=current_platform.simple_compile_backend)
|
942 |
def grouped_topk(
|
943 |
hidden_states: torch.Tensor,
|
944 |
gating_output: torch.Tensor,
|
|
|
946 |
renormalize: bool,
|
947 |
num_expert_group: int = 0,
|
948 |
topk_group: int = 0,
|
949 |
+
scoring_func: str = "softmax",
|
950 |
+
e_score_correction_bias: Optional[torch.Tensor] = None,
|
951 |
):
|
952 |
|
953 |
assert hidden_states.shape[0] == gating_output.shape[0], "Number of tokens mismatch"
|
954 |
|
955 |
+
if scoring_func == "softmax":
|
956 |
+
scores = torch.softmax(gating_output, dim=-1)
|
957 |
+
elif scoring_func == "sigmoid":
|
958 |
+
scores = gating_output.sigmoid()
|
959 |
+
else:
|
960 |
+
raise ValueError(f"Unsupported scoring function: {scoring_func}")
|
961 |
+
|
962 |
+
if e_score_correction_bias is not None:
|
963 |
+
# Store original scores before applying correction bias. We use biased
|
964 |
+
# scores for expert selection but original scores for routing weights
|
965 |
+
original_scores = scores
|
966 |
+
scores = scores + e_score_correction_bias.unsqueeze(0)
|
967 |
+
|
968 |
num_token = scores.shape[0]
|
969 |
group_scores = (
|
970 |
scores.view(num_token, num_expert_group, -1).max(dim=-1).values
|
|
|
980 |
.reshape(num_token, -1)
|
981 |
) # [n, e]
|
982 |
tmp_scores = scores.masked_fill(~score_mask.bool(), 0.0) # [n, e]
|
983 |
+
|
984 |
+
if e_score_correction_bias is not None:
|
985 |
+
topk_ids = torch.topk(tmp_scores, k=topk, dim=-1, sorted=False)[1]
|
986 |
+
# Use original unbiased scores for the routing weights
|
987 |
+
topk_weights = original_scores.gather(1, topk_ids)
|
988 |
+
else:
|
989 |
+
topk_weights, topk_ids = torch.topk(tmp_scores, k=topk, dim=-1, sorted=False)
|
990 |
|
991 |
if renormalize:
|
992 |
topk_weights = topk_weights / topk_weights.sum(dim=-1, keepdim=True)
|
|
|
996 |
|
997 |
def get_config_dtype_str(
|
998 |
dtype: torch.dtype,
|
999 |
+
use_int4_w4a16: Optional[bool] = False,
|
1000 |
use_int8_w8a16: Optional[bool] = False,
|
1001 |
use_fp8_w8a8: Optional[bool] = False,
|
1002 |
):
|
|
|
1004 |
return "fp8_w8a8"
|
1005 |
elif use_int8_w8a16:
|
1006 |
return "int8_w8a16"
|
1007 |
+
elif use_int4_w4a16:
|
1008 |
+
return "int4_w8a16"
|
1009 |
elif dtype == torch.float:
|
1010 |
# avoiding cases where kernel fails when float32 MoE
|
1011 |
# use fp16/bfloat16 configs
|
|
|
1013 |
return None
|
1014 |
|
1015 |
|
1016 |
+
def inplace_fused_experts(
|
1017 |
+
hidden_states: torch.Tensor,
|
1018 |
+
w1: torch.Tensor,
|
1019 |
+
w2: torch.Tensor,
|
1020 |
+
topk_weights: torch.Tensor,
|
1021 |
+
topk_ids: torch.Tensor,
|
1022 |
+
use_fp8_w8a8: bool = False,
|
1023 |
+
use_int8_w8a16: bool = False,
|
1024 |
+
use_int4_w4a16: bool = False,
|
1025 |
+
w1_scale: Optional[torch.Tensor] = None,
|
1026 |
+
w2_scale: Optional[torch.Tensor] = None,
|
1027 |
+
w1_zp: Optional[torch.Tensor] = None,
|
1028 |
+
w2_zp: Optional[torch.Tensor] = None,
|
1029 |
+
a1_scale: Optional[torch.Tensor] = None,
|
1030 |
+
a2_scale: Optional[torch.Tensor] = None,
|
1031 |
+
block_shape: Optional[List[int]] = None,
|
1032 |
+
) -> None:
|
1033 |
+
fused_experts_impl(
|
1034 |
+
hidden_states,
|
1035 |
+
w1,
|
1036 |
+
w2,
|
1037 |
+
topk_weights,
|
1038 |
+
topk_ids,
|
1039 |
+
True,
|
1040 |
+
use_fp8_w8a8,
|
1041 |
+
use_int8_w8a16,
|
1042 |
+
use_int4_w4a16,
|
1043 |
+
w1_scale,
|
1044 |
+
w2_scale,
|
1045 |
+
w1_zp,
|
1046 |
+
w2_zp,
|
1047 |
+
a1_scale,
|
1048 |
+
a2_scale,
|
1049 |
+
block_shape,
|
1050 |
+
)
|
1051 |
+
|
1052 |
+
|
1053 |
+
def outplace_fused_experts(
|
1054 |
+
hidden_states: torch.Tensor,
|
1055 |
+
w1: torch.Tensor,
|
1056 |
+
w2: torch.Tensor,
|
1057 |
+
topk_weights: torch.Tensor,
|
1058 |
+
topk_ids: torch.Tensor,
|
1059 |
+
use_fp8_w8a8: bool = False,
|
1060 |
+
use_int8_w8a16: bool = False,
|
1061 |
+
use_int4_w4a16: bool = False,
|
1062 |
+
w1_scale: Optional[torch.Tensor] = None,
|
1063 |
+
w2_scale: Optional[torch.Tensor] = None,
|
1064 |
+
w1_zp: Optional[torch.Tensor] = None,
|
1065 |
+
w2_zp: Optional[torch.Tensor] = None,
|
1066 |
+
a1_scale: Optional[torch.Tensor] = None,
|
1067 |
+
a2_scale: Optional[torch.Tensor] = None,
|
1068 |
+
block_shape: Optional[List[int]] = None,
|
1069 |
+
) -> torch.Tensor:
|
1070 |
+
return fused_experts_impl(
|
1071 |
+
hidden_states,
|
1072 |
+
w1,
|
1073 |
+
w2,
|
1074 |
+
topk_weights,
|
1075 |
+
topk_ids,
|
1076 |
+
False,
|
1077 |
+
use_fp8_w8a8,
|
1078 |
+
use_int8_w8a16,
|
1079 |
+
use_int4_w4a16,
|
1080 |
+
w1_scale,
|
1081 |
+
w2_scale,
|
1082 |
+
w1_zp,
|
1083 |
+
w2_zp,
|
1084 |
+
a1_scale,
|
1085 |
+
a2_scale,
|
1086 |
+
block_shape,
|
1087 |
+
)
|
1088 |
+
|
1089 |
+
|
1090 |
def fused_experts(
|
1091 |
hidden_states: torch.Tensor,
|
1092 |
w1: torch.Tensor,
|
|
|
1094 |
topk_weights: torch.Tensor,
|
1095 |
topk_ids: torch.Tensor,
|
1096 |
inplace: bool = False,
|
|
|
1097 |
use_fp8_w8a8: bool = False,
|
1098 |
use_int8_w8a16: bool = False,
|
1099 |
+
use_int4_w4a16: bool = False,
|
1100 |
+
w1_scale: Optional[torch.Tensor] = None,
|
1101 |
+
w2_scale: Optional[torch.Tensor] = None,
|
1102 |
+
w1_zp: Optional[torch.Tensor] = None,
|
1103 |
+
w2_zp: Optional[torch.Tensor] = None,
|
1104 |
+
a1_scale: Optional[torch.Tensor] = None,
|
1105 |
+
a2_scale: Optional[torch.Tensor] = None,
|
1106 |
+
block_shape: Optional[List[int]] = None,
|
1107 |
+
):
|
1108 |
+
if inplace:
|
1109 |
+
inplace_fused_experts(
|
1110 |
+
hidden_states,
|
1111 |
+
w1,
|
1112 |
+
w2,
|
1113 |
+
topk_weights,
|
1114 |
+
topk_ids,
|
1115 |
+
use_fp8_w8a8,
|
1116 |
+
use_int8_w8a16,
|
1117 |
+
use_int4_w4a16,
|
1118 |
+
w1_scale,
|
1119 |
+
w2_scale,
|
1120 |
+
w1_zp,
|
1121 |
+
w2_zp,
|
1122 |
+
a1_scale,
|
1123 |
+
a2_scale,
|
1124 |
+
block_shape,
|
1125 |
+
)
|
1126 |
+
return hidden_states
|
1127 |
+
else:
|
1128 |
+
return outplace_fused_experts(
|
1129 |
+
hidden_states,
|
1130 |
+
w1,
|
1131 |
+
w2,
|
1132 |
+
topk_weights,
|
1133 |
+
topk_ids,
|
1134 |
+
use_fp8_w8a8,
|
1135 |
+
use_int8_w8a16,
|
1136 |
+
use_int4_w4a16,
|
1137 |
+
w1_scale,
|
1138 |
+
w2_scale,
|
1139 |
+
w1_zp,
|
1140 |
+
w2_zp,
|
1141 |
+
a1_scale,
|
1142 |
+
a2_scale,
|
1143 |
+
block_shape,
|
1144 |
+
)
|
1145 |
+
|
1146 |
+
|
1147 |
+
def fused_experts_impl(
|
1148 |
+
hidden_states: torch.Tensor,
|
1149 |
+
w1: torch.Tensor,
|
1150 |
+
w2: torch.Tensor,
|
1151 |
+
topk_weights: torch.Tensor,
|
1152 |
+
topk_ids: torch.Tensor,
|
1153 |
+
inplace: bool = False,
|
1154 |
+
use_fp8_w8a8: bool = False,
|
1155 |
+
use_int8_w8a16: bool = False,
|
1156 |
+
use_int4_w4a16: bool = False,
|
1157 |
w1_scale: Optional[torch.Tensor] = None,
|
1158 |
w2_scale: Optional[torch.Tensor] = None,
|
1159 |
+
w1_zp: Optional[torch.Tensor] = None,
|
1160 |
+
w2_zp: Optional[torch.Tensor] = None,
|
1161 |
a1_scale: Optional[torch.Tensor] = None,
|
1162 |
a2_scale: Optional[torch.Tensor] = None,
|
1163 |
+
block_shape: Optional[List[int]] = None,
|
1164 |
):
|
1165 |
# Check constraints.
|
1166 |
+
if use_int4_w4a16:
|
1167 |
+
assert hidden_states.shape[1] // 2 == w1.shape[2], "Hidden size mismatch"
|
1168 |
+
else:
|
1169 |
+
assert hidden_states.shape[1] == w1.shape[2], "Hidden size mismatch"
|
1170 |
+
|
1171 |
assert topk_weights.shape == topk_ids.shape, "topk shape mismatch"
|
1172 |
assert hidden_states.is_contiguous(), "Hidden_states must be contiguous"
|
1173 |
assert w1.is_contiguous(), "Expert weights1 must be contiguous"
|
|
|
1183 |
config_dtype = get_config_dtype_str(
|
1184 |
use_fp8_w8a8=use_fp8_w8a8,
|
1185 |
use_int8_w8a16=use_int8_w8a16,
|
1186 |
+
use_int4_w4a16=use_int4_w4a16,
|
1187 |
dtype=hidden_states.dtype,
|
1188 |
)
|
1189 |
|
|
|
1193 |
w2.shape,
|
1194 |
topk_ids.shape[1],
|
1195 |
config_dtype,
|
1196 |
+
block_shape=block_shape,
|
1197 |
)
|
1198 |
|
1199 |
config = get_config_func(M)
|
|
|
1214 |
dtype=hidden_states.dtype,
|
1215 |
)
|
1216 |
|
1217 |
+
if hidden_states.dtype == torch.bfloat16:
|
1218 |
+
compute_type = tl.bfloat16
|
1219 |
+
elif hidden_states.dtype == torch.float16:
|
1220 |
+
compute_type = tl.float16
|
1221 |
+
elif hidden_states.dtype == torch.float32:
|
1222 |
+
compute_type = tl.float32
|
1223 |
+
else:
|
1224 |
+
raise ValueError(f"Unsupported compute_type: {hidden_states.dtype}")
|
1225 |
|
1226 |
if inplace:
|
1227 |
out_hidden_states = hidden_states
|
|
|
1262 |
intermediate_cache1,
|
1263 |
a1_scale,
|
1264 |
w1_scale,
|
1265 |
+
w1_zp,
|
1266 |
curr_topk_weights,
|
1267 |
curr_topk_ids,
|
1268 |
sorted_token_ids,
|
|
|
1274 |
compute_type=compute_type,
|
1275 |
use_fp8_w8a8=use_fp8_w8a8,
|
1276 |
use_int8_w8a16=use_int8_w8a16,
|
1277 |
+
use_int4_w4a16=use_int4_w4a16,
|
1278 |
+
block_shape=block_shape,
|
1279 |
)
|
1280 |
|
1281 |
ops.silu_and_mul(intermediate_cache2, intermediate_cache1.view(-1, N))
|
|
|
1286 |
intermediate_cache3,
|
1287 |
a2_scale,
|
1288 |
w2_scale,
|
1289 |
+
w2_zp,
|
1290 |
curr_topk_weights,
|
1291 |
curr_topk_ids,
|
1292 |
sorted_token_ids,
|
|
|
1298 |
compute_type=compute_type,
|
1299 |
use_fp8_w8a8=use_fp8_w8a8,
|
1300 |
use_int8_w8a16=use_int8_w8a16,
|
1301 |
+
use_int4_w4a16=use_int4_w4a16,
|
1302 |
+
block_shape=block_shape,
|
1303 |
)
|
1304 |
|
1305 |
ops.moe_sum(
|
|
|
1317 |
topk: int,
|
1318 |
renormalize: bool,
|
1319 |
inplace: bool = False,
|
|
|
1320 |
use_grouped_topk: bool = False,
|
1321 |
num_expert_group: Optional[int] = None,
|
1322 |
topk_group: Optional[int] = None,
|
1323 |
custom_routing_function: Optional[Callable] = None,
|
1324 |
use_fp8_w8a8: bool = False,
|
1325 |
use_int8_w8a16: bool = False,
|
1326 |
+
use_int4_w4a16: bool = False,
|
1327 |
w1_scale: Optional[torch.Tensor] = None,
|
1328 |
w2_scale: Optional[torch.Tensor] = None,
|
1329 |
+
w1_zp: Optional[torch.Tensor] = None,
|
1330 |
+
w2_zp: Optional[torch.Tensor] = None,
|
1331 |
a1_scale: Optional[torch.Tensor] = None,
|
1332 |
a2_scale: Optional[torch.Tensor] = None,
|
1333 |
+
block_shape: Optional[List[int]] = None,
|
1334 |
) -> torch.Tensor:
|
1335 |
"""
|
1336 |
This function computes a Mixture of Experts (MoE) layer using two sets of
|
|
|
1346 |
- renormalize (bool): If True, renormalize the top-k weights to sum to 1.
|
1347 |
- inplace (bool): If True, perform the operation in-place.
|
1348 |
Defaults to False.
|
|
|
|
|
1349 |
- num_expert_group: Optional[int]: additional parameter for grouped_topk
|
1350 |
- topk_group: Optional[int]: additional parameter for grouped_topk
|
1351 |
- use_grouped_topk: If True, use grouped_topk instead of fused_topk
|
1352 |
note: Deepseekv2 model uses grouped_topk
|
1353 |
- use_fp8_w8a8 (bool): If True, use fp8 arithmetic to compute the inner
|
1354 |
products for w1 and w2. Defaults to False.
|
1355 |
+
- use_int8_w8a16 (bool): If True, use matmul of int8 weight and bf16/fp16
|
1356 |
+
activation to compute the inner products for w1 and w2.
|
1357 |
+
Defaults to False.
|
1358 |
+
- use_int4_w4a16 (bool): If True, use matmul of int4 weight and bf16/fp16
|
1359 |
+
activation to compute the inner products for w1 and w2.
|
1360 |
+
Defaults to False.
|
1361 |
- w1_scale (Optional[torch.Tensor]): Optional scale to be used for
|
1362 |
w1.
|
1363 |
- w2_scale (Optional[torch.Tensor]): Optional scale to be used for
|
1364 |
w2.
|
1365 |
+
- a1_scale (Optional[torch.Tensor]): Optional scale to be used for
|
1366 |
+
a1.
|
1367 |
+
- a2_scale (Optional[torch.Tensor]): Optional scale to be used for
|
1368 |
+
a2.
|
1369 |
+
- block_shape: (Optional[List[int]]): Optional block size for block-wise
|
1370 |
+
quantization.
|
1371 |
|
1372 |
Returns:
|
1373 |
- torch.Tensor: The output tensor after applying the MoE layer.
|
|
|
1401 |
topk_weights,
|
1402 |
topk_ids,
|
1403 |
inplace=inplace,
|
|
|
1404 |
use_fp8_w8a8=use_fp8_w8a8,
|
1405 |
use_int8_w8a16=use_int8_w8a16,
|
1406 |
+
use_int4_w4a16=use_int4_w4a16,
|
1407 |
w1_scale=w1_scale,
|
1408 |
w2_scale=w2_scale,
|
1409 |
+
w1_zp=w1_zp,
|
1410 |
+
w2_zp=w2_zp,
|
1411 |
a1_scale=a1_scale,
|
1412 |
a2_scale=a2_scale,
|
1413 |
+
block_shape=block_shape,
|
1414 |
)
|
build/torch26-cxx11-cu118-x86_64-linux/moe/platforms.py
CHANGED
@@ -1,22 +1,32 @@
|
|
1 |
-
from
|
2 |
-
import os
|
3 |
-
from functools import lru_cache, wraps
|
4 |
|
5 |
import torch
|
6 |
|
7 |
IS_ROCM = torch.version.hip is not None
|
8 |
|
9 |
-
|
|
|
|
|
|
|
|
|
|
|
10 |
@classmethod
|
11 |
@lru_cache(maxsize=8)
|
12 |
def get_device_name(cls, device_id: int = 0) -> str:
|
13 |
return torch.cuda.get_device_name(0)
|
14 |
|
15 |
-
|
|
|
|
|
|
|
|
|
16 |
@classmethod
|
17 |
@lru_cache(maxsize=8)
|
18 |
def get_device_name(cls, device_id: int = 0) -> str:
|
19 |
return torch.cuda.get_device_name(device_id)
|
20 |
|
|
|
|
|
|
|
21 |
|
22 |
current_platform = RocmPlatform() if IS_ROCM else CudaPlatform()
|
|
|
1 |
+
from functools import lru_cache
|
|
|
|
|
2 |
|
3 |
import torch
|
4 |
|
5 |
IS_ROCM = torch.version.hip is not None
|
6 |
|
7 |
+
|
8 |
+
class Platform:
|
9 |
+
simple_compile_backend: str = "inductor"
|
10 |
+
|
11 |
+
|
12 |
+
class CudaPlatform(Platform):
|
13 |
@classmethod
|
14 |
@lru_cache(maxsize=8)
|
15 |
def get_device_name(cls, device_id: int = 0) -> str:
|
16 |
return torch.cuda.get_device_name(0)
|
17 |
|
18 |
+
def is_rocm(self):
|
19 |
+
return False
|
20 |
+
|
21 |
+
|
22 |
+
class RocmPlatform(Platform):
|
23 |
@classmethod
|
24 |
@lru_cache(maxsize=8)
|
25 |
def get_device_name(cls, device_id: int = 0) -> str:
|
26 |
return torch.cuda.get_device_name(device_id)
|
27 |
|
28 |
+
def is_rocm(self):
|
29 |
+
return True
|
30 |
+
|
31 |
|
32 |
current_platform = RocmPlatform() if IS_ROCM else CudaPlatform()
|
build/torch26-cxx11-cu124-x86_64-linux/moe/_moe_h5rxhm5fum47w.abi3.so
DELETED
@@ -1,3 +0,0 @@
|
|
1 |
-
version https://git-lfs.github.com/spec/v1
|
2 |
-
oid sha256:82358e87c49352e80bf23b7cbb9e52ed655be254b7da552ebdaa5af172a8625f
|
3 |
-
size 84063432
|
|
|
|
|
|
|
|
build/torch26-cxx11-cu124-x86_64-linux/moe/_moe_wua27hyvpwmli.abi3.so
ADDED
@@ -0,0 +1,3 @@
|
|
|
|
|
|
|
|
|
1 |
+
version https://git-lfs.github.com/spec/v1
|
2 |
+
oid sha256:d3f7f1fa2f76004fba0e0d4eb8cbc3e35a7182538c83261f4a01a8e7401bfa81
|
3 |
+
size 85737400
|
build/torch26-cxx11-cu124-x86_64-linux/moe/_ops.py
CHANGED
@@ -1,9 +1,9 @@
|
|
1 |
import torch
|
2 |
-
from . import
|
3 |
-
ops = torch.ops.
|
4 |
|
5 |
def add_op_namespace_prefix(op_name: str):
|
6 |
"""
|
7 |
Prefix op by namespace.
|
8 |
"""
|
9 |
-
return f"
|
|
|
1 |
import torch
|
2 |
+
from . import _moe_wua27hyvpwmli
|
3 |
+
ops = torch.ops._moe_wua27hyvpwmli
|
4 |
|
5 |
def add_op_namespace_prefix(op_name: str):
|
6 |
"""
|
7 |
Prefix op by namespace.
|
8 |
"""
|
9 |
+
return f"_moe_wua27hyvpwmli::{op_name}"
|
build/torch26-cxx11-cu124-x86_64-linux/moe/fp8.py
CHANGED
@@ -1,6 +1,11 @@
|
|
|
|
|
|
1 |
import torch
|
|
|
|
|
2 |
|
3 |
-
|
|
|
4 |
|
5 |
|
6 |
def is_hip() -> bool:
|
@@ -49,15 +54,179 @@ def scaled_fp8_quant(
|
|
49 |
if scale is None:
|
50 |
if use_per_token_if_dynamic:
|
51 |
scale = torch.empty((shape[0], 1), device=input.device, dtype=torch.float32)
|
52 |
-
|
53 |
-
output, input, scale, scale_ub
|
54 |
-
)
|
55 |
else:
|
56 |
scale = torch.zeros(1, device=input.device, dtype=torch.float32)
|
57 |
-
|
58 |
else:
|
59 |
# num_token_padding not implemented for this case
|
60 |
assert scale.numel() == 1 or num_token_padding is None
|
61 |
-
|
62 |
|
63 |
return output, scale
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
1 |
+
from typing import Tuple, Optional, Union
|
2 |
+
|
3 |
import torch
|
4 |
+
import triton
|
5 |
+
import triton.language as tl
|
6 |
|
7 |
+
|
8 |
+
from ._ops import ops
|
9 |
|
10 |
|
11 |
def is_hip() -> bool:
|
|
|
54 |
if scale is None:
|
55 |
if use_per_token_if_dynamic:
|
56 |
scale = torch.empty((shape[0], 1), device=input.device, dtype=torch.float32)
|
57 |
+
ops.dynamic_per_token_scaled_fp8_quant(output, input, scale, scale_ub)
|
|
|
|
|
58 |
else:
|
59 |
scale = torch.zeros(1, device=input.device, dtype=torch.float32)
|
60 |
+
ops.dynamic_scaled_fp8_quant(output, input, scale)
|
61 |
else:
|
62 |
# num_token_padding not implemented for this case
|
63 |
assert scale.numel() == 1 or num_token_padding is None
|
64 |
+
ops.static_scaled_fp8_quant(output, input, scale)
|
65 |
|
66 |
return output, scale
|
67 |
+
|
68 |
+
|
69 |
+
@triton.jit
|
70 |
+
def _per_token_group_quant_fp8(
|
71 |
+
# Pointers to inputs and output
|
72 |
+
y_ptr,
|
73 |
+
y_q_ptr,
|
74 |
+
y_s_ptr,
|
75 |
+
group_size,
|
76 |
+
# Avoid to divide zero
|
77 |
+
eps,
|
78 |
+
# Information for float8
|
79 |
+
fp8_min,
|
80 |
+
fp8_max,
|
81 |
+
# Meta-parameters
|
82 |
+
BLOCK: tl.constexpr,
|
83 |
+
):
|
84 |
+
"""A Triton-accelerated function to perform per-token-group
|
85 |
+
quantization on a tensor.
|
86 |
+
This function converts the tensor values into float8 values.
|
87 |
+
"""
|
88 |
+
# Map the program id to the row of X and Y it should compute.
|
89 |
+
g_id = tl.program_id(0)
|
90 |
+
y_ptr += g_id * group_size
|
91 |
+
y_q_ptr += g_id * group_size
|
92 |
+
y_s_ptr += g_id
|
93 |
+
|
94 |
+
cols = tl.arange(0, BLOCK) # N <= BLOCK
|
95 |
+
mask = cols < group_size
|
96 |
+
|
97 |
+
y = tl.load(y_ptr + cols, mask=mask, other=0.0).to(tl.float32)
|
98 |
+
# Quant
|
99 |
+
_absmax = tl.maximum(tl.max(tl.abs(y)), eps)
|
100 |
+
y_s = _absmax / fp8_max
|
101 |
+
y_q = tl.clamp(y / y_s, fp8_min, fp8_max).to(y_q_ptr.dtype.element_ty)
|
102 |
+
|
103 |
+
tl.store(y_q_ptr + cols, y_q, mask=mask)
|
104 |
+
tl.store(y_s_ptr, y_s)
|
105 |
+
|
106 |
+
|
107 |
+
@triton.jit
|
108 |
+
def _per_token_group_quant_fp8_colmajor(
|
109 |
+
# Pointers to inputs and output
|
110 |
+
y_ptr,
|
111 |
+
y_q_ptr,
|
112 |
+
y_s_ptr,
|
113 |
+
group_size,
|
114 |
+
# Num columns of y
|
115 |
+
y_num_columns,
|
116 |
+
# Stride from one column to the next of y_s
|
117 |
+
y_s_col_stride,
|
118 |
+
# Avoid to divide zero
|
119 |
+
eps,
|
120 |
+
# Information for float8
|
121 |
+
fp8_min,
|
122 |
+
fp8_max,
|
123 |
+
# Meta-parameters
|
124 |
+
BLOCK: tl.constexpr,
|
125 |
+
):
|
126 |
+
"""A Triton-accelerated function to perform per-token-group
|
127 |
+
quantization on a tensor.
|
128 |
+
This function converts the tensor values into float8 values.
|
129 |
+
"""
|
130 |
+
# Map the program id to the row of X and Y it should compute.
|
131 |
+
g_id = tl.program_id(0)
|
132 |
+
y_ptr += g_id * group_size
|
133 |
+
y_q_ptr += g_id * group_size
|
134 |
+
|
135 |
+
# Convert g_id the flattened block coordinate to 2D so we can index
|
136 |
+
# into the output y_scales matrix
|
137 |
+
blocks_per_row = y_num_columns // group_size
|
138 |
+
scale_col = g_id % blocks_per_row
|
139 |
+
scale_row = g_id // blocks_per_row
|
140 |
+
y_s_ptr += scale_col * y_s_col_stride + scale_row
|
141 |
+
|
142 |
+
cols = tl.arange(0, BLOCK) # group_size <= BLOCK
|
143 |
+
mask = cols < group_size
|
144 |
+
|
145 |
+
y = tl.load(y_ptr + cols, mask=mask, other=0.0).to(tl.float32)
|
146 |
+
# Quant
|
147 |
+
_absmax = tl.maximum(tl.max(tl.abs(y)), eps)
|
148 |
+
y_s = _absmax / fp8_max
|
149 |
+
y_q = tl.clamp(y / y_s, fp8_min, fp8_max).to(y_q_ptr.dtype.element_ty)
|
150 |
+
|
151 |
+
tl.store(y_q_ptr + cols, y_q, mask=mask)
|
152 |
+
tl.store(y_s_ptr, y_s)
|
153 |
+
|
154 |
+
|
155 |
+
def per_token_group_quant_fp8(
|
156 |
+
x: torch.Tensor,
|
157 |
+
group_size: int,
|
158 |
+
eps: float = 1e-10,
|
159 |
+
dtype: Optional[torch.dtype] = None,
|
160 |
+
column_major_scales: bool = False,
|
161 |
+
) -> Tuple[torch.Tensor, torch.Tensor]:
|
162 |
+
"""Function to perform per-token-group quantization on an input tensor `x`.
|
163 |
+
It converts the tensor values into signed float8 values and returns the
|
164 |
+
quantized tensor along with the scaling factor used for quantization.
|
165 |
+
Args:
|
166 |
+
x: The input tensor with ndim >= 2.
|
167 |
+
group_size: The group size used for quantization.
|
168 |
+
eps: The minimum to avoid dividing zero.
|
169 |
+
dtype: The dype of output tensor. Note that only `torch.float8_e4m3fn`
|
170 |
+
is supported for now.
|
171 |
+
Returns:
|
172 |
+
Tuple[torch.Tensor, torch.Tensor]: The quantized tensor and the
|
173 |
+
scaling factor for quantization.
|
174 |
+
"""
|
175 |
+
if dtype is None:
|
176 |
+
dtype = (
|
177 |
+
torch.float8_e4m3fnuz if current_platform.is_rocm() else torch.float8_e4m3fn
|
178 |
+
)
|
179 |
+
assert x.shape[-1] % group_size == 0, (
|
180 |
+
f"the last dimension of `x` {x.shape[-1]} must be divisible "
|
181 |
+
f"by `group_size` {group_size}"
|
182 |
+
)
|
183 |
+
assert x.is_contiguous(), "`x` must be contiguous"
|
184 |
+
|
185 |
+
finfo = torch.finfo(dtype)
|
186 |
+
fp8_min = finfo.min
|
187 |
+
fp8_max = finfo.max
|
188 |
+
|
189 |
+
x_q = torch.empty_like(x, device=x.device, dtype=dtype)
|
190 |
+
M = x.numel() // group_size
|
191 |
+
N = group_size
|
192 |
+
if column_major_scales:
|
193 |
+
shape = (x.shape[-1] // group_size,) + x.shape[:-1]
|
194 |
+
x_s = torch.empty(shape, device=x.device, dtype=torch.float32).permute(-1, -2)
|
195 |
+
else:
|
196 |
+
shape = x.shape[:-1] + (x.shape[-1] // group_size,)
|
197 |
+
x_s = torch.empty(shape, device=x.device, dtype=torch.float32)
|
198 |
+
|
199 |
+
BLOCK = triton.next_power_of_2(N)
|
200 |
+
# heuristics for number of warps
|
201 |
+
num_warps = min(max(BLOCK // 256, 1), 8)
|
202 |
+
num_stages = 1
|
203 |
+
if column_major_scales:
|
204 |
+
_per_token_group_quant_fp8_colmajor[(M,)](
|
205 |
+
x,
|
206 |
+
x_q,
|
207 |
+
x_s,
|
208 |
+
group_size,
|
209 |
+
x.shape[1],
|
210 |
+
x_s.stride(1),
|
211 |
+
eps,
|
212 |
+
fp8_min=fp8_min,
|
213 |
+
fp8_max=fp8_max,
|
214 |
+
BLOCK=BLOCK,
|
215 |
+
num_warps=num_warps,
|
216 |
+
num_stages=num_stages,
|
217 |
+
)
|
218 |
+
else:
|
219 |
+
_per_token_group_quant_fp8[(M,)](
|
220 |
+
x,
|
221 |
+
x_q,
|
222 |
+
x_s,
|
223 |
+
group_size,
|
224 |
+
eps,
|
225 |
+
fp8_min=fp8_min,
|
226 |
+
fp8_max=fp8_max,
|
227 |
+
BLOCK=BLOCK,
|
228 |
+
num_warps=num_warps,
|
229 |
+
num_stages=num_stages,
|
230 |
+
)
|
231 |
+
|
232 |
+
return x_q, x_s
|
build/torch26-cxx11-cu124-x86_64-linux/moe/fused_marlin_moe.py
CHANGED
@@ -40,7 +40,6 @@ def single_marlin_moe(
|
|
40 |
g_idx: Optional[torch.Tensor] = None,
|
41 |
sort_indices: Optional[torch.Tensor] = None,
|
42 |
w_zeros: Optional[torch.Tensor] = None,
|
43 |
-
override_config: Optional[Dict[str, Any]] = None,
|
44 |
num_bits: int = 8,
|
45 |
is_k_full: bool = True,
|
46 |
) -> torch.Tensor:
|
@@ -61,8 +60,6 @@ def single_marlin_moe(
|
|
61 |
- topk (int): The number of top-k experts to select.
|
62 |
- renormalize (bool): If True, renormalize the top-k weights to sum to 1.
|
63 |
- w_zeros (Optional[torch.Tensor]): Optional zero points to be used for w.
|
64 |
-
- override_config (Optional[Dict[str, Any]]): Optional override
|
65 |
-
for the kernel configuration.
|
66 |
- num_bits (bool): The number of bits in expert weights quantization.
|
67 |
|
68 |
Returns:
|
@@ -90,7 +87,6 @@ def single_marlin_moe(
|
|
90 |
w.shape,
|
91 |
topk_ids.shape[1],
|
92 |
None,
|
93 |
-
override_config=override_config,
|
94 |
is_marlin=True,
|
95 |
)
|
96 |
config = get_config_func(M)
|
@@ -154,6 +150,25 @@ def single_marlin_moe(
|
|
154 |
return torch.sum(intermediate_cache.view(*intermediate_cache.shape), dim=1)
|
155 |
|
156 |
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
157 |
def fused_marlin_moe(
|
158 |
hidden_states: torch.Tensor,
|
159 |
w1: torch.Tensor,
|
@@ -169,7 +184,6 @@ def fused_marlin_moe(
|
|
169 |
sort_indices2: Optional[torch.Tensor] = None,
|
170 |
w1_zeros: Optional[torch.Tensor] = None,
|
171 |
w2_zeros: Optional[torch.Tensor] = None,
|
172 |
-
override_config: Optional[Dict[str, Any]] = None,
|
173 |
num_bits: int = 8,
|
174 |
is_k_full: bool = True,
|
175 |
) -> torch.Tensor:
|
@@ -193,8 +207,6 @@ def fused_marlin_moe(
|
|
193 |
permutation.
|
194 |
- topk_weights (torch.Tensor): Top-k weights.
|
195 |
- topk_ids (torch.Tensor): Indices of topk-k elements.
|
196 |
-
- override_config (Optional[Dict[str, Any]]): Optional override
|
197 |
-
for the kernel configuration.
|
198 |
- w1_zeros (Optional[torch.Tensor]): Optional zero points to be used for w1.
|
199 |
- w2_zeros (Optional[torch.Tensor]): Optional zero points to be used for w2.
|
200 |
- num_bits (bool): The number of bits in expert weights quantization.
|
@@ -248,7 +260,6 @@ def fused_marlin_moe(
|
|
248 |
w2.shape,
|
249 |
topk_ids.shape[1],
|
250 |
None,
|
251 |
-
override_config=override_config,
|
252 |
is_marlin=True,
|
253 |
)
|
254 |
config = get_config_func(M)
|
@@ -350,6 +361,30 @@ def fused_marlin_moe(
|
|
350 |
return torch.sum(intermediate_cache3.view(*intermediate_cache3.shape), dim=1)
|
351 |
|
352 |
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
353 |
if hasattr(ops, "marlin_gemm_moe"):
|
354 |
|
355 |
@register_fake(add_op_namespace_prefix("marlin_gemm_moe"))
|
|
|
40 |
g_idx: Optional[torch.Tensor] = None,
|
41 |
sort_indices: Optional[torch.Tensor] = None,
|
42 |
w_zeros: Optional[torch.Tensor] = None,
|
|
|
43 |
num_bits: int = 8,
|
44 |
is_k_full: bool = True,
|
45 |
) -> torch.Tensor:
|
|
|
60 |
- topk (int): The number of top-k experts to select.
|
61 |
- renormalize (bool): If True, renormalize the top-k weights to sum to 1.
|
62 |
- w_zeros (Optional[torch.Tensor]): Optional zero points to be used for w.
|
|
|
|
|
63 |
- num_bits (bool): The number of bits in expert weights quantization.
|
64 |
|
65 |
Returns:
|
|
|
87 |
w.shape,
|
88 |
topk_ids.shape[1],
|
89 |
None,
|
|
|
90 |
is_marlin=True,
|
91 |
)
|
92 |
config = get_config_func(M)
|
|
|
150 |
return torch.sum(intermediate_cache.view(*intermediate_cache.shape), dim=1)
|
151 |
|
152 |
|
153 |
+
if hasattr(ops, "single_marlin_gemm_moe"):
|
154 |
+
|
155 |
+
@register_fake(add_op_namespace_prefix("single_marlin_gemm_moe"))
|
156 |
+
def single_marlin_moe_fake(
|
157 |
+
hidden_states: torch.Tensor,
|
158 |
+
w: torch.Tensor,
|
159 |
+
scales: torch.Tensor,
|
160 |
+
gating_output: torch.Tensor,
|
161 |
+
topk: int,
|
162 |
+
renormalize: bool,
|
163 |
+
g_idx: Optional[torch.Tensor] = None,
|
164 |
+
sort_indices: Optional[torch.Tensor] = None,
|
165 |
+
w_zeros: Optional[torch.Tensor] = None,
|
166 |
+
num_bits: int = 8,
|
167 |
+
is_k_full: bool = True,
|
168 |
+
) -> torch.Tensor:
|
169 |
+
return torch.empty_like(hidden_states)
|
170 |
+
|
171 |
+
|
172 |
def fused_marlin_moe(
|
173 |
hidden_states: torch.Tensor,
|
174 |
w1: torch.Tensor,
|
|
|
184 |
sort_indices2: Optional[torch.Tensor] = None,
|
185 |
w1_zeros: Optional[torch.Tensor] = None,
|
186 |
w2_zeros: Optional[torch.Tensor] = None,
|
|
|
187 |
num_bits: int = 8,
|
188 |
is_k_full: bool = True,
|
189 |
) -> torch.Tensor:
|
|
|
207 |
permutation.
|
208 |
- topk_weights (torch.Tensor): Top-k weights.
|
209 |
- topk_ids (torch.Tensor): Indices of topk-k elements.
|
|
|
|
|
210 |
- w1_zeros (Optional[torch.Tensor]): Optional zero points to be used for w1.
|
211 |
- w2_zeros (Optional[torch.Tensor]): Optional zero points to be used for w2.
|
212 |
- num_bits (bool): The number of bits in expert weights quantization.
|
|
|
260 |
w2.shape,
|
261 |
topk_ids.shape[1],
|
262 |
None,
|
|
|
263 |
is_marlin=True,
|
264 |
)
|
265 |
config = get_config_func(M)
|
|
|
361 |
return torch.sum(intermediate_cache3.view(*intermediate_cache3.shape), dim=1)
|
362 |
|
363 |
|
364 |
+
if hasattr(ops, "fused_marlin_moe"):
|
365 |
+
|
366 |
+
@register_fake(add_op_namespace_prefix("fused_marlin_moe"))
|
367 |
+
def fused_marlin_moe_fake(
|
368 |
+
hidden_states: torch.Tensor,
|
369 |
+
w1: torch.Tensor,
|
370 |
+
w2: torch.Tensor,
|
371 |
+
w1_scale: torch.Tensor,
|
372 |
+
w2_scale: torch.Tensor,
|
373 |
+
gating_output: torch.Tensor,
|
374 |
+
topk_weights: torch.Tensor,
|
375 |
+
topk_ids: torch.Tensor,
|
376 |
+
g_idx1: Optional[torch.Tensor] = None,
|
377 |
+
g_idx2: Optional[torch.Tensor] = None,
|
378 |
+
sort_indices1: Optional[torch.Tensor] = None,
|
379 |
+
sort_indices2: Optional[torch.Tensor] = None,
|
380 |
+
w1_zeros: Optional[torch.Tensor] = None,
|
381 |
+
w2_zeros: Optional[torch.Tensor] = None,
|
382 |
+
num_bits: int = 8,
|
383 |
+
is_k_full: bool = True,
|
384 |
+
) -> torch.Tensor:
|
385 |
+
return torch.empty_like(hidden_states)
|
386 |
+
|
387 |
+
|
388 |
if hasattr(ops, "marlin_gemm_moe"):
|
389 |
|
390 |
@register_fake(add_op_namespace_prefix("marlin_gemm_moe"))
|