Build
Browse files- build/torch24-cxx11-cu118-x86_64-linux/quantization/__init__.py +107 -1
- build/torch24-cxx11-cu118-x86_64-linux/quantization/_quantization_0_0_1.abi3.so +2 -2
- build/torch24-cxx11-cu121-x86_64-linux/quantization/__init__.py +107 -1
- build/torch24-cxx11-cu121-x86_64-linux/quantization/_quantization_0_0_1.abi3.so +2 -2
- build/torch24-cxx11-cu124-x86_64-linux/quantization/__init__.py +107 -1
- build/torch24-cxx11-cu124-x86_64-linux/quantization/_quantization_0_0_1.abi3.so +2 -2
- build/torch24-cxx98-cu118-x86_64-linux/quantization/__init__.py +107 -1
- build/torch24-cxx98-cu118-x86_64-linux/quantization/_quantization_0_0_1.abi3.so +2 -2
- build/torch24-cxx98-cu121-x86_64-linux/quantization/__init__.py +107 -1
- build/torch24-cxx98-cu121-x86_64-linux/quantization/_quantization_0_0_1.abi3.so +2 -2
- build/torch24-cxx98-cu124-x86_64-linux/quantization/__init__.py +107 -1
- build/torch24-cxx98-cu124-x86_64-linux/quantization/_quantization_0_0_1.abi3.so +2 -2
- build/torch25-cxx11-cu118-x86_64-linux/quantization/__init__.py +107 -1
- build/torch25-cxx11-cu118-x86_64-linux/quantization/_quantization_0_0_1.abi3.so +2 -2
- build/torch25-cxx11-cu121-x86_64-linux/quantization/__init__.py +107 -1
- build/torch25-cxx11-cu121-x86_64-linux/quantization/_quantization_0_0_1.abi3.so +2 -2
- build/torch25-cxx11-cu124-x86_64-linux/quantization/__init__.py +107 -1
- build/torch25-cxx11-cu124-x86_64-linux/quantization/_quantization_0_0_1.abi3.so +2 -2
- build/torch25-cxx98-cu118-x86_64-linux/quantization/__init__.py +107 -1
- build/torch25-cxx98-cu118-x86_64-linux/quantization/_quantization_0_0_1.abi3.so +2 -2
- build/torch25-cxx98-cu121-x86_64-linux/quantization/__init__.py +107 -1
- build/torch25-cxx98-cu121-x86_64-linux/quantization/_quantization_0_0_1.abi3.so +2 -2
- build/torch25-cxx98-cu124-x86_64-linux/quantization/__init__.py +107 -1
- build/torch25-cxx98-cu124-x86_64-linux/quantization/_quantization_0_0_1.abi3.so +2 -2
build/torch24-cxx11-cu118-x86_64-linux/quantization/__init__.py
CHANGED
@@ -1,4 +1,4 @@
|
|
1 |
-
from typing import Optional
|
2 |
|
3 |
import torch
|
4 |
|
@@ -42,3 +42,109 @@ def cutlass_scaled_mm(a: torch.Tensor,
|
|
42 |
|
43 |
return out
|
44 |
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
1 |
+
from typing import Optional, Tuple
|
2 |
|
3 |
import torch
|
4 |
|
|
|
42 |
|
43 |
return out
|
44 |
|
45 |
+
# fp8
|
46 |
+
def scaled_fp8_quant(
|
47 |
+
input: torch.Tensor,
|
48 |
+
scale: Optional[torch.Tensor] = None,
|
49 |
+
num_token_padding: Optional[int] = None,
|
50 |
+
scale_ub: Optional[torch.Tensor] = None,
|
51 |
+
use_per_token_if_dynamic: bool = False,
|
52 |
+
) -> Tuple[torch.Tensor, torch.Tensor]:
|
53 |
+
"""
|
54 |
+
Quantize input tensor to FP8 and return quantized tensor and scale.
|
55 |
+
|
56 |
+
This function supports both static and dynamic quantization: If you
|
57 |
+
provide the scale, it will use static scaling and if you omit it,
|
58 |
+
the scale will be determined dynamically. The function also allows
|
59 |
+
optional padding of the output tensors for downstream kernels that
|
60 |
+
will benefit from padding.
|
61 |
+
|
62 |
+
Args:
|
63 |
+
input: The input tensor to be quantized to FP8
|
64 |
+
scale: Optional scaling factor for the FP8 quantization
|
65 |
+
scale_ub: Optional upper bound for scaling factor in dynamic
|
66 |
+
per token case
|
67 |
+
num_token_padding: If specified, pad the first dimension
|
68 |
+
of the output to at least this value.
|
69 |
+
use_per_token_if_dynamic: Whether to do per_tensor or per_token
|
70 |
+
in the dynamic quantization case.
|
71 |
+
|
72 |
+
Returns:
|
73 |
+
Tuple[torch.Tensor, torch.Tensor]: The output tensor in FP8 and
|
74 |
+
scaling factor.
|
75 |
+
"""
|
76 |
+
# This code assumes batch_dim and num_tokens are flattened
|
77 |
+
assert (input.ndim == 2)
|
78 |
+
shape: Union[Tuple[int, int], torch.Size] = input.shape
|
79 |
+
# For rocm, the output fp8 dtype is torch.float_e3m3fnuz
|
80 |
+
#out_dtype: torch.dtype = torch.float8_e4m3fnuz \
|
81 |
+
# if current_platform.is_rocm() else torch.float8_e4m3fn
|
82 |
+
out_dtype = torch.float8_e4m3fn
|
83 |
+
if num_token_padding:
|
84 |
+
shape = (max(num_token_padding, input.shape[0]), shape[1])
|
85 |
+
output = torch.empty(shape, device=input.device, dtype=out_dtype)
|
86 |
+
|
87 |
+
if scale is None:
|
88 |
+
if use_per_token_if_dynamic:
|
89 |
+
scale = torch.empty((shape[0], 1),
|
90 |
+
device=input.device,
|
91 |
+
dtype=torch.float32)
|
92 |
+
ops.dynamic_per_token_scaled_fp8_quant(
|
93 |
+
output, input, scale, scale_ub)
|
94 |
+
else:
|
95 |
+
scale = torch.zeros(1, device=input.device, dtype=torch.float32)
|
96 |
+
ops.dynamic_scaled_fp8_quant(output, input, scale)
|
97 |
+
else:
|
98 |
+
# num_token_padding not implemented for this case
|
99 |
+
assert (scale.numel() == 1 or num_token_padding is None)
|
100 |
+
ops.static_scaled_fp8_quant(output, input, scale)
|
101 |
+
|
102 |
+
return output, scale
|
103 |
+
|
104 |
+
# int8
|
105 |
+
def scaled_int8_quant(
|
106 |
+
input: torch.Tensor,
|
107 |
+
scale: Optional[torch.Tensor] = None,
|
108 |
+
azp: Optional[torch.Tensor] = None,
|
109 |
+
symmetric: bool = True
|
110 |
+
) -> Tuple[torch.Tensor, torch.Tensor, Optional[torch.Tensor]]:
|
111 |
+
"""
|
112 |
+
Quantize the input tensor to int8 and return the quantized tensor and scale, and maybe azp.
|
113 |
+
|
114 |
+
Args:
|
115 |
+
input: The input tensor to be quantized to int8.
|
116 |
+
scale: Optional scaling factor for the int8 quantization.
|
117 |
+
When not provided, we invoke dynamic-per-token quantization.
|
118 |
+
azp: Optional zero-point for the int8 quantization.
|
119 |
+
Must be provided for asymmetric quantization if `scale` is provided.
|
120 |
+
symmetric: Whether to use symmetric quantization (scale only, azp ignored).
|
121 |
+
|
122 |
+
Returns:
|
123 |
+
Tuple[torch.Tensor, torch.Tensor, Optional[torch.Tensor]] : Output int8 tensor, scales, and optionally azp.
|
124 |
+
"""
|
125 |
+
output = torch.empty_like(input, dtype=torch.int8)
|
126 |
+
if scale is not None:
|
127 |
+
# static-per-tensor quantization.
|
128 |
+
assert symmetric == (
|
129 |
+
azp is
|
130 |
+
None), "azp must only be provided for asymmetric quantization."
|
131 |
+
ops.static_scaled_int8_quant(output, input, scale, azp)
|
132 |
+
return output, scale, azp
|
133 |
+
|
134 |
+
# dynamic-per-token quantization.
|
135 |
+
input_scales = torch.empty((input.numel() // input.shape[-1], 1),
|
136 |
+
device=input.device,
|
137 |
+
dtype=torch.float32)
|
138 |
+
input_azp = None if symmetric else torch.empty_like(input_scales,
|
139 |
+
dtype=torch.int32)
|
140 |
+
ops.dynamic_scaled_int8_quant(output, input, input_scales,
|
141 |
+
input_azp)
|
142 |
+
return output, input_scales, input_azp
|
143 |
+
|
144 |
+
# fp8 marlin
|
145 |
+
def fp8_marlin_gemm(a: torch.Tensor, b_q_weight: torch.Tensor,
|
146 |
+
b_scales: torch.Tensor, workspace: torch.Tensor,
|
147 |
+
num_bits: int, size_m: int, size_n: int,
|
148 |
+
size_k: int) -> torch.Tensor:
|
149 |
+
return ops.fp8_marlin_gemm(a, b_q_weight, b_scales, workspace,
|
150 |
+
num_bits, size_m, size_n, size_k)
|
build/torch24-cxx11-cu118-x86_64-linux/quantization/_quantization_0_0_1.abi3.so
CHANGED
@@ -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 sha256:1ff4df94b3de0caab5c2c7584f21eb3898d495bcf92a731fb1fd9a46ba0dff50
|
3 |
+
size 39178896
|
build/torch24-cxx11-cu121-x86_64-linux/quantization/__init__.py
CHANGED
@@ -1,4 +1,4 @@
|
|
1 |
-
from typing import Optional
|
2 |
|
3 |
import torch
|
4 |
|
@@ -42,3 +42,109 @@ def cutlass_scaled_mm(a: torch.Tensor,
|
|
42 |
|
43 |
return out
|
44 |
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
1 |
+
from typing import Optional, Tuple
|
2 |
|
3 |
import torch
|
4 |
|
|
|
42 |
|
43 |
return out
|
44 |
|
45 |
+
# fp8
|
46 |
+
def scaled_fp8_quant(
|
47 |
+
input: torch.Tensor,
|
48 |
+
scale: Optional[torch.Tensor] = None,
|
49 |
+
num_token_padding: Optional[int] = None,
|
50 |
+
scale_ub: Optional[torch.Tensor] = None,
|
51 |
+
use_per_token_if_dynamic: bool = False,
|
52 |
+
) -> Tuple[torch.Tensor, torch.Tensor]:
|
53 |
+
"""
|
54 |
+
Quantize input tensor to FP8 and return quantized tensor and scale.
|
55 |
+
|
56 |
+
This function supports both static and dynamic quantization: If you
|
57 |
+
provide the scale, it will use static scaling and if you omit it,
|
58 |
+
the scale will be determined dynamically. The function also allows
|
59 |
+
optional padding of the output tensors for downstream kernels that
|
60 |
+
will benefit from padding.
|
61 |
+
|
62 |
+
Args:
|
63 |
+
input: The input tensor to be quantized to FP8
|
64 |
+
scale: Optional scaling factor for the FP8 quantization
|
65 |
+
scale_ub: Optional upper bound for scaling factor in dynamic
|
66 |
+
per token case
|
67 |
+
num_token_padding: If specified, pad the first dimension
|
68 |
+
of the output to at least this value.
|
69 |
+
use_per_token_if_dynamic: Whether to do per_tensor or per_token
|
70 |
+
in the dynamic quantization case.
|
71 |
+
|
72 |
+
Returns:
|
73 |
+
Tuple[torch.Tensor, torch.Tensor]: The output tensor in FP8 and
|
74 |
+
scaling factor.
|
75 |
+
"""
|
76 |
+
# This code assumes batch_dim and num_tokens are flattened
|
77 |
+
assert (input.ndim == 2)
|
78 |
+
shape: Union[Tuple[int, int], torch.Size] = input.shape
|
79 |
+
# For rocm, the output fp8 dtype is torch.float_e3m3fnuz
|
80 |
+
#out_dtype: torch.dtype = torch.float8_e4m3fnuz \
|
81 |
+
# if current_platform.is_rocm() else torch.float8_e4m3fn
|
82 |
+
out_dtype = torch.float8_e4m3fn
|
83 |
+
if num_token_padding:
|
84 |
+
shape = (max(num_token_padding, input.shape[0]), shape[1])
|
85 |
+
output = torch.empty(shape, device=input.device, dtype=out_dtype)
|
86 |
+
|
87 |
+
if scale is None:
|
88 |
+
if use_per_token_if_dynamic:
|
89 |
+
scale = torch.empty((shape[0], 1),
|
90 |
+
device=input.device,
|
91 |
+
dtype=torch.float32)
|
92 |
+
ops.dynamic_per_token_scaled_fp8_quant(
|
93 |
+
output, input, scale, scale_ub)
|
94 |
+
else:
|
95 |
+
scale = torch.zeros(1, device=input.device, dtype=torch.float32)
|
96 |
+
ops.dynamic_scaled_fp8_quant(output, input, scale)
|
97 |
+
else:
|
98 |
+
# num_token_padding not implemented for this case
|
99 |
+
assert (scale.numel() == 1 or num_token_padding is None)
|
100 |
+
ops.static_scaled_fp8_quant(output, input, scale)
|
101 |
+
|
102 |
+
return output, scale
|
103 |
+
|
104 |
+
# int8
|
105 |
+
def scaled_int8_quant(
|
106 |
+
input: torch.Tensor,
|
107 |
+
scale: Optional[torch.Tensor] = None,
|
108 |
+
azp: Optional[torch.Tensor] = None,
|
109 |
+
symmetric: bool = True
|
110 |
+
) -> Tuple[torch.Tensor, torch.Tensor, Optional[torch.Tensor]]:
|
111 |
+
"""
|
112 |
+
Quantize the input tensor to int8 and return the quantized tensor and scale, and maybe azp.
|
113 |
+
|
114 |
+
Args:
|
115 |
+
input: The input tensor to be quantized to int8.
|
116 |
+
scale: Optional scaling factor for the int8 quantization.
|
117 |
+
When not provided, we invoke dynamic-per-token quantization.
|
118 |
+
azp: Optional zero-point for the int8 quantization.
|
119 |
+
Must be provided for asymmetric quantization if `scale` is provided.
|
120 |
+
symmetric: Whether to use symmetric quantization (scale only, azp ignored).
|
121 |
+
|
122 |
+
Returns:
|
123 |
+
Tuple[torch.Tensor, torch.Tensor, Optional[torch.Tensor]] : Output int8 tensor, scales, and optionally azp.
|
124 |
+
"""
|
125 |
+
output = torch.empty_like(input, dtype=torch.int8)
|
126 |
+
if scale is not None:
|
127 |
+
# static-per-tensor quantization.
|
128 |
+
assert symmetric == (
|
129 |
+
azp is
|
130 |
+
None), "azp must only be provided for asymmetric quantization."
|
131 |
+
ops.static_scaled_int8_quant(output, input, scale, azp)
|
132 |
+
return output, scale, azp
|
133 |
+
|
134 |
+
# dynamic-per-token quantization.
|
135 |
+
input_scales = torch.empty((input.numel() // input.shape[-1], 1),
|
136 |
+
device=input.device,
|
137 |
+
dtype=torch.float32)
|
138 |
+
input_azp = None if symmetric else torch.empty_like(input_scales,
|
139 |
+
dtype=torch.int32)
|
140 |
+
ops.dynamic_scaled_int8_quant(output, input, input_scales,
|
141 |
+
input_azp)
|
142 |
+
return output, input_scales, input_azp
|
143 |
+
|
144 |
+
# fp8 marlin
|
145 |
+
def fp8_marlin_gemm(a: torch.Tensor, b_q_weight: torch.Tensor,
|
146 |
+
b_scales: torch.Tensor, workspace: torch.Tensor,
|
147 |
+
num_bits: int, size_m: int, size_n: int,
|
148 |
+
size_k: int) -> torch.Tensor:
|
149 |
+
return ops.fp8_marlin_gemm(a, b_q_weight, b_scales, workspace,
|
150 |
+
num_bits, size_m, size_n, size_k)
|
build/torch24-cxx11-cu121-x86_64-linux/quantization/_quantization_0_0_1.abi3.so
CHANGED
@@ -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 sha256:0dd82ba302527185d214e508371d75778524c412934d0f2399bd3d00402b89c5
|
3 |
+
size 46540064
|
build/torch24-cxx11-cu124-x86_64-linux/quantization/__init__.py
CHANGED
@@ -1,4 +1,4 @@
|
|
1 |
-
from typing import Optional
|
2 |
|
3 |
import torch
|
4 |
|
@@ -42,3 +42,109 @@ def cutlass_scaled_mm(a: torch.Tensor,
|
|
42 |
|
43 |
return out
|
44 |
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
1 |
+
from typing import Optional, Tuple
|
2 |
|
3 |
import torch
|
4 |
|
|
|
42 |
|
43 |
return out
|
44 |
|
45 |
+
# fp8
|
46 |
+
def scaled_fp8_quant(
|
47 |
+
input: torch.Tensor,
|
48 |
+
scale: Optional[torch.Tensor] = None,
|
49 |
+
num_token_padding: Optional[int] = None,
|
50 |
+
scale_ub: Optional[torch.Tensor] = None,
|
51 |
+
use_per_token_if_dynamic: bool = False,
|
52 |
+
) -> Tuple[torch.Tensor, torch.Tensor]:
|
53 |
+
"""
|
54 |
+
Quantize input tensor to FP8 and return quantized tensor and scale.
|
55 |
+
|
56 |
+
This function supports both static and dynamic quantization: If you
|
57 |
+
provide the scale, it will use static scaling and if you omit it,
|
58 |
+
the scale will be determined dynamically. The function also allows
|
59 |
+
optional padding of the output tensors for downstream kernels that
|
60 |
+
will benefit from padding.
|
61 |
+
|
62 |
+
Args:
|
63 |
+
input: The input tensor to be quantized to FP8
|
64 |
+
scale: Optional scaling factor for the FP8 quantization
|
65 |
+
scale_ub: Optional upper bound for scaling factor in dynamic
|
66 |
+
per token case
|
67 |
+
num_token_padding: If specified, pad the first dimension
|
68 |
+
of the output to at least this value.
|
69 |
+
use_per_token_if_dynamic: Whether to do per_tensor or per_token
|
70 |
+
in the dynamic quantization case.
|
71 |
+
|
72 |
+
Returns:
|
73 |
+
Tuple[torch.Tensor, torch.Tensor]: The output tensor in FP8 and
|
74 |
+
scaling factor.
|
75 |
+
"""
|
76 |
+
# This code assumes batch_dim and num_tokens are flattened
|
77 |
+
assert (input.ndim == 2)
|
78 |
+
shape: Union[Tuple[int, int], torch.Size] = input.shape
|
79 |
+
# For rocm, the output fp8 dtype is torch.float_e3m3fnuz
|
80 |
+
#out_dtype: torch.dtype = torch.float8_e4m3fnuz \
|
81 |
+
# if current_platform.is_rocm() else torch.float8_e4m3fn
|
82 |
+
out_dtype = torch.float8_e4m3fn
|
83 |
+
if num_token_padding:
|
84 |
+
shape = (max(num_token_padding, input.shape[0]), shape[1])
|
85 |
+
output = torch.empty(shape, device=input.device, dtype=out_dtype)
|
86 |
+
|
87 |
+
if scale is None:
|
88 |
+
if use_per_token_if_dynamic:
|
89 |
+
scale = torch.empty((shape[0], 1),
|
90 |
+
device=input.device,
|
91 |
+
dtype=torch.float32)
|
92 |
+
ops.dynamic_per_token_scaled_fp8_quant(
|
93 |
+
output, input, scale, scale_ub)
|
94 |
+
else:
|
95 |
+
scale = torch.zeros(1, device=input.device, dtype=torch.float32)
|
96 |
+
ops.dynamic_scaled_fp8_quant(output, input, scale)
|
97 |
+
else:
|
98 |
+
# num_token_padding not implemented for this case
|
99 |
+
assert (scale.numel() == 1 or num_token_padding is None)
|
100 |
+
ops.static_scaled_fp8_quant(output, input, scale)
|
101 |
+
|
102 |
+
return output, scale
|
103 |
+
|
104 |
+
# int8
|
105 |
+
def scaled_int8_quant(
|
106 |
+
input: torch.Tensor,
|
107 |
+
scale: Optional[torch.Tensor] = None,
|
108 |
+
azp: Optional[torch.Tensor] = None,
|
109 |
+
symmetric: bool = True
|
110 |
+
) -> Tuple[torch.Tensor, torch.Tensor, Optional[torch.Tensor]]:
|
111 |
+
"""
|
112 |
+
Quantize the input tensor to int8 and return the quantized tensor and scale, and maybe azp.
|
113 |
+
|
114 |
+
Args:
|
115 |
+
input: The input tensor to be quantized to int8.
|
116 |
+
scale: Optional scaling factor for the int8 quantization.
|
117 |
+
When not provided, we invoke dynamic-per-token quantization.
|
118 |
+
azp: Optional zero-point for the int8 quantization.
|
119 |
+
Must be provided for asymmetric quantization if `scale` is provided.
|
120 |
+
symmetric: Whether to use symmetric quantization (scale only, azp ignored).
|
121 |
+
|
122 |
+
Returns:
|
123 |
+
Tuple[torch.Tensor, torch.Tensor, Optional[torch.Tensor]] : Output int8 tensor, scales, and optionally azp.
|
124 |
+
"""
|
125 |
+
output = torch.empty_like(input, dtype=torch.int8)
|
126 |
+
if scale is not None:
|
127 |
+
# static-per-tensor quantization.
|
128 |
+
assert symmetric == (
|
129 |
+
azp is
|
130 |
+
None), "azp must only be provided for asymmetric quantization."
|
131 |
+
ops.static_scaled_int8_quant(output, input, scale, azp)
|
132 |
+
return output, scale, azp
|
133 |
+
|
134 |
+
# dynamic-per-token quantization.
|
135 |
+
input_scales = torch.empty((input.numel() // input.shape[-1], 1),
|
136 |
+
device=input.device,
|
137 |
+
dtype=torch.float32)
|
138 |
+
input_azp = None if symmetric else torch.empty_like(input_scales,
|
139 |
+
dtype=torch.int32)
|
140 |
+
ops.dynamic_scaled_int8_quant(output, input, input_scales,
|
141 |
+
input_azp)
|
142 |
+
return output, input_scales, input_azp
|
143 |
+
|
144 |
+
# fp8 marlin
|
145 |
+
def fp8_marlin_gemm(a: torch.Tensor, b_q_weight: torch.Tensor,
|
146 |
+
b_scales: torch.Tensor, workspace: torch.Tensor,
|
147 |
+
num_bits: int, size_m: int, size_n: int,
|
148 |
+
size_k: int) -> torch.Tensor:
|
149 |
+
return ops.fp8_marlin_gemm(a, b_q_weight, b_scales, workspace,
|
150 |
+
num_bits, size_m, size_n, size_k)
|
build/torch24-cxx11-cu124-x86_64-linux/quantization/_quantization_0_0_1.abi3.so
CHANGED
@@ -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 sha256:0cac14195f6181d145e9f1cc0c1e532f8cfa2914fe7ff59fdf3194c85fd28b9c
|
3 |
+
size 47413592
|
build/torch24-cxx98-cu118-x86_64-linux/quantization/__init__.py
CHANGED
@@ -1,4 +1,4 @@
|
|
1 |
-
from typing import Optional
|
2 |
|
3 |
import torch
|
4 |
|
@@ -42,3 +42,109 @@ def cutlass_scaled_mm(a: torch.Tensor,
|
|
42 |
|
43 |
return out
|
44 |
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
1 |
+
from typing import Optional, Tuple
|
2 |
|
3 |
import torch
|
4 |
|
|
|
42 |
|
43 |
return out
|
44 |
|
45 |
+
# fp8
|
46 |
+
def scaled_fp8_quant(
|
47 |
+
input: torch.Tensor,
|
48 |
+
scale: Optional[torch.Tensor] = None,
|
49 |
+
num_token_padding: Optional[int] = None,
|
50 |
+
scale_ub: Optional[torch.Tensor] = None,
|
51 |
+
use_per_token_if_dynamic: bool = False,
|
52 |
+
) -> Tuple[torch.Tensor, torch.Tensor]:
|
53 |
+
"""
|
54 |
+
Quantize input tensor to FP8 and return quantized tensor and scale.
|
55 |
+
|
56 |
+
This function supports both static and dynamic quantization: If you
|
57 |
+
provide the scale, it will use static scaling and if you omit it,
|
58 |
+
the scale will be determined dynamically. The function also allows
|
59 |
+
optional padding of the output tensors for downstream kernels that
|
60 |
+
will benefit from padding.
|
61 |
+
|
62 |
+
Args:
|
63 |
+
input: The input tensor to be quantized to FP8
|
64 |
+
scale: Optional scaling factor for the FP8 quantization
|
65 |
+
scale_ub: Optional upper bound for scaling factor in dynamic
|
66 |
+
per token case
|
67 |
+
num_token_padding: If specified, pad the first dimension
|
68 |
+
of the output to at least this value.
|
69 |
+
use_per_token_if_dynamic: Whether to do per_tensor or per_token
|
70 |
+
in the dynamic quantization case.
|
71 |
+
|
72 |
+
Returns:
|
73 |
+
Tuple[torch.Tensor, torch.Tensor]: The output tensor in FP8 and
|
74 |
+
scaling factor.
|
75 |
+
"""
|
76 |
+
# This code assumes batch_dim and num_tokens are flattened
|
77 |
+
assert (input.ndim == 2)
|
78 |
+
shape: Union[Tuple[int, int], torch.Size] = input.shape
|
79 |
+
# For rocm, the output fp8 dtype is torch.float_e3m3fnuz
|
80 |
+
#out_dtype: torch.dtype = torch.float8_e4m3fnuz \
|
81 |
+
# if current_platform.is_rocm() else torch.float8_e4m3fn
|
82 |
+
out_dtype = torch.float8_e4m3fn
|
83 |
+
if num_token_padding:
|
84 |
+
shape = (max(num_token_padding, input.shape[0]), shape[1])
|
85 |
+
output = torch.empty(shape, device=input.device, dtype=out_dtype)
|
86 |
+
|
87 |
+
if scale is None:
|
88 |
+
if use_per_token_if_dynamic:
|
89 |
+
scale = torch.empty((shape[0], 1),
|
90 |
+
device=input.device,
|
91 |
+
dtype=torch.float32)
|
92 |
+
ops.dynamic_per_token_scaled_fp8_quant(
|
93 |
+
output, input, scale, scale_ub)
|
94 |
+
else:
|
95 |
+
scale = torch.zeros(1, device=input.device, dtype=torch.float32)
|
96 |
+
ops.dynamic_scaled_fp8_quant(output, input, scale)
|
97 |
+
else:
|
98 |
+
# num_token_padding not implemented for this case
|
99 |
+
assert (scale.numel() == 1 or num_token_padding is None)
|
100 |
+
ops.static_scaled_fp8_quant(output, input, scale)
|
101 |
+
|
102 |
+
return output, scale
|
103 |
+
|
104 |
+
# int8
|
105 |
+
def scaled_int8_quant(
|
106 |
+
input: torch.Tensor,
|
107 |
+
scale: Optional[torch.Tensor] = None,
|
108 |
+
azp: Optional[torch.Tensor] = None,
|
109 |
+
symmetric: bool = True
|
110 |
+
) -> Tuple[torch.Tensor, torch.Tensor, Optional[torch.Tensor]]:
|
111 |
+
"""
|
112 |
+
Quantize the input tensor to int8 and return the quantized tensor and scale, and maybe azp.
|
113 |
+
|
114 |
+
Args:
|
115 |
+
input: The input tensor to be quantized to int8.
|
116 |
+
scale: Optional scaling factor for the int8 quantization.
|
117 |
+
When not provided, we invoke dynamic-per-token quantization.
|
118 |
+
azp: Optional zero-point for the int8 quantization.
|
119 |
+
Must be provided for asymmetric quantization if `scale` is provided.
|
120 |
+
symmetric: Whether to use symmetric quantization (scale only, azp ignored).
|
121 |
+
|
122 |
+
Returns:
|
123 |
+
Tuple[torch.Tensor, torch.Tensor, Optional[torch.Tensor]] : Output int8 tensor, scales, and optionally azp.
|
124 |
+
"""
|
125 |
+
output = torch.empty_like(input, dtype=torch.int8)
|
126 |
+
if scale is not None:
|
127 |
+
# static-per-tensor quantization.
|
128 |
+
assert symmetric == (
|
129 |
+
azp is
|
130 |
+
None), "azp must only be provided for asymmetric quantization."
|
131 |
+
ops.static_scaled_int8_quant(output, input, scale, azp)
|
132 |
+
return output, scale, azp
|
133 |
+
|
134 |
+
# dynamic-per-token quantization.
|
135 |
+
input_scales = torch.empty((input.numel() // input.shape[-1], 1),
|
136 |
+
device=input.device,
|
137 |
+
dtype=torch.float32)
|
138 |
+
input_azp = None if symmetric else torch.empty_like(input_scales,
|
139 |
+
dtype=torch.int32)
|
140 |
+
ops.dynamic_scaled_int8_quant(output, input, input_scales,
|
141 |
+
input_azp)
|
142 |
+
return output, input_scales, input_azp
|
143 |
+
|
144 |
+
# fp8 marlin
|
145 |
+
def fp8_marlin_gemm(a: torch.Tensor, b_q_weight: torch.Tensor,
|
146 |
+
b_scales: torch.Tensor, workspace: torch.Tensor,
|
147 |
+
num_bits: int, size_m: int, size_n: int,
|
148 |
+
size_k: int) -> torch.Tensor:
|
149 |
+
return ops.fp8_marlin_gemm(a, b_q_weight, b_scales, workspace,
|
150 |
+
num_bits, size_m, size_n, size_k)
|
build/torch24-cxx98-cu118-x86_64-linux/quantization/_quantization_0_0_1.abi3.so
CHANGED
@@ -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 sha256:0054660979a4c0a273f32b6293294d02048b14a90af3b1f3e5cb226504cffe15
|
3 |
+
size 39166248
|
build/torch24-cxx98-cu121-x86_64-linux/quantization/__init__.py
CHANGED
@@ -1,4 +1,4 @@
|
|
1 |
-
from typing import Optional
|
2 |
|
3 |
import torch
|
4 |
|
@@ -42,3 +42,109 @@ def cutlass_scaled_mm(a: torch.Tensor,
|
|
42 |
|
43 |
return out
|
44 |
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
1 |
+
from typing import Optional, Tuple
|
2 |
|
3 |
import torch
|
4 |
|
|
|
42 |
|
43 |
return out
|
44 |
|
45 |
+
# fp8
|
46 |
+
def scaled_fp8_quant(
|
47 |
+
input: torch.Tensor,
|
48 |
+
scale: Optional[torch.Tensor] = None,
|
49 |
+
num_token_padding: Optional[int] = None,
|
50 |
+
scale_ub: Optional[torch.Tensor] = None,
|
51 |
+
use_per_token_if_dynamic: bool = False,
|
52 |
+
) -> Tuple[torch.Tensor, torch.Tensor]:
|
53 |
+
"""
|
54 |
+
Quantize input tensor to FP8 and return quantized tensor and scale.
|
55 |
+
|
56 |
+
This function supports both static and dynamic quantization: If you
|
57 |
+
provide the scale, it will use static scaling and if you omit it,
|
58 |
+
the scale will be determined dynamically. The function also allows
|
59 |
+
optional padding of the output tensors for downstream kernels that
|
60 |
+
will benefit from padding.
|
61 |
+
|
62 |
+
Args:
|
63 |
+
input: The input tensor to be quantized to FP8
|
64 |
+
scale: Optional scaling factor for the FP8 quantization
|
65 |
+
scale_ub: Optional upper bound for scaling factor in dynamic
|
66 |
+
per token case
|
67 |
+
num_token_padding: If specified, pad the first dimension
|
68 |
+
of the output to at least this value.
|
69 |
+
use_per_token_if_dynamic: Whether to do per_tensor or per_token
|
70 |
+
in the dynamic quantization case.
|
71 |
+
|
72 |
+
Returns:
|
73 |
+
Tuple[torch.Tensor, torch.Tensor]: The output tensor in FP8 and
|
74 |
+
scaling factor.
|
75 |
+
"""
|
76 |
+
# This code assumes batch_dim and num_tokens are flattened
|
77 |
+
assert (input.ndim == 2)
|
78 |
+
shape: Union[Tuple[int, int], torch.Size] = input.shape
|
79 |
+
# For rocm, the output fp8 dtype is torch.float_e3m3fnuz
|
80 |
+
#out_dtype: torch.dtype = torch.float8_e4m3fnuz \
|
81 |
+
# if current_platform.is_rocm() else torch.float8_e4m3fn
|
82 |
+
out_dtype = torch.float8_e4m3fn
|
83 |
+
if num_token_padding:
|
84 |
+
shape = (max(num_token_padding, input.shape[0]), shape[1])
|
85 |
+
output = torch.empty(shape, device=input.device, dtype=out_dtype)
|
86 |
+
|
87 |
+
if scale is None:
|
88 |
+
if use_per_token_if_dynamic:
|
89 |
+
scale = torch.empty((shape[0], 1),
|
90 |
+
device=input.device,
|
91 |
+
dtype=torch.float32)
|
92 |
+
ops.dynamic_per_token_scaled_fp8_quant(
|
93 |
+
output, input, scale, scale_ub)
|
94 |
+
else:
|
95 |
+
scale = torch.zeros(1, device=input.device, dtype=torch.float32)
|
96 |
+
ops.dynamic_scaled_fp8_quant(output, input, scale)
|
97 |
+
else:
|
98 |
+
# num_token_padding not implemented for this case
|
99 |
+
assert (scale.numel() == 1 or num_token_padding is None)
|
100 |
+
ops.static_scaled_fp8_quant(output, input, scale)
|
101 |
+
|
102 |
+
return output, scale
|
103 |
+
|
104 |
+
# int8
|
105 |
+
def scaled_int8_quant(
|
106 |
+
input: torch.Tensor,
|
107 |
+
scale: Optional[torch.Tensor] = None,
|
108 |
+
azp: Optional[torch.Tensor] = None,
|
109 |
+
symmetric: bool = True
|
110 |
+
) -> Tuple[torch.Tensor, torch.Tensor, Optional[torch.Tensor]]:
|
111 |
+
"""
|
112 |
+
Quantize the input tensor to int8 and return the quantized tensor and scale, and maybe azp.
|
113 |
+
|
114 |
+
Args:
|
115 |
+
input: The input tensor to be quantized to int8.
|
116 |
+
scale: Optional scaling factor for the int8 quantization.
|
117 |
+
When not provided, we invoke dynamic-per-token quantization.
|
118 |
+
azp: Optional zero-point for the int8 quantization.
|
119 |
+
Must be provided for asymmetric quantization if `scale` is provided.
|
120 |
+
symmetric: Whether to use symmetric quantization (scale only, azp ignored).
|
121 |
+
|
122 |
+
Returns:
|
123 |
+
Tuple[torch.Tensor, torch.Tensor, Optional[torch.Tensor]] : Output int8 tensor, scales, and optionally azp.
|
124 |
+
"""
|
125 |
+
output = torch.empty_like(input, dtype=torch.int8)
|
126 |
+
if scale is not None:
|
127 |
+
# static-per-tensor quantization.
|
128 |
+
assert symmetric == (
|
129 |
+
azp is
|
130 |
+
None), "azp must only be provided for asymmetric quantization."
|
131 |
+
ops.static_scaled_int8_quant(output, input, scale, azp)
|
132 |
+
return output, scale, azp
|
133 |
+
|
134 |
+
# dynamic-per-token quantization.
|
135 |
+
input_scales = torch.empty((input.numel() // input.shape[-1], 1),
|
136 |
+
device=input.device,
|
137 |
+
dtype=torch.float32)
|
138 |
+
input_azp = None if symmetric else torch.empty_like(input_scales,
|
139 |
+
dtype=torch.int32)
|
140 |
+
ops.dynamic_scaled_int8_quant(output, input, input_scales,
|
141 |
+
input_azp)
|
142 |
+
return output, input_scales, input_azp
|
143 |
+
|
144 |
+
# fp8 marlin
|
145 |
+
def fp8_marlin_gemm(a: torch.Tensor, b_q_weight: torch.Tensor,
|
146 |
+
b_scales: torch.Tensor, workspace: torch.Tensor,
|
147 |
+
num_bits: int, size_m: int, size_n: int,
|
148 |
+
size_k: int) -> torch.Tensor:
|
149 |
+
return ops.fp8_marlin_gemm(a, b_q_weight, b_scales, workspace,
|
150 |
+
num_bits, size_m, size_n, size_k)
|
build/torch24-cxx98-cu121-x86_64-linux/quantization/_quantization_0_0_1.abi3.so
CHANGED
@@ -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 sha256:735a3a1fe3aea0065a7378f11611378923331d7633000441fa5d0d5e03d0d481
|
3 |
+
size 46534608
|
build/torch24-cxx98-cu124-x86_64-linux/quantization/__init__.py
CHANGED
@@ -1,4 +1,4 @@
|
|
1 |
-
from typing import Optional
|
2 |
|
3 |
import torch
|
4 |
|
@@ -42,3 +42,109 @@ def cutlass_scaled_mm(a: torch.Tensor,
|
|
42 |
|
43 |
return out
|
44 |
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
1 |
+
from typing import Optional, Tuple
|
2 |
|
3 |
import torch
|
4 |
|
|
|
42 |
|
43 |
return out
|
44 |
|
45 |
+
# fp8
|
46 |
+
def scaled_fp8_quant(
|
47 |
+
input: torch.Tensor,
|
48 |
+
scale: Optional[torch.Tensor] = None,
|
49 |
+
num_token_padding: Optional[int] = None,
|
50 |
+
scale_ub: Optional[torch.Tensor] = None,
|
51 |
+
use_per_token_if_dynamic: bool = False,
|
52 |
+
) -> Tuple[torch.Tensor, torch.Tensor]:
|
53 |
+
"""
|
54 |
+
Quantize input tensor to FP8 and return quantized tensor and scale.
|
55 |
+
|
56 |
+
This function supports both static and dynamic quantization: If you
|
57 |
+
provide the scale, it will use static scaling and if you omit it,
|
58 |
+
the scale will be determined dynamically. The function also allows
|
59 |
+
optional padding of the output tensors for downstream kernels that
|
60 |
+
will benefit from padding.
|
61 |
+
|
62 |
+
Args:
|
63 |
+
input: The input tensor to be quantized to FP8
|
64 |
+
scale: Optional scaling factor for the FP8 quantization
|
65 |
+
scale_ub: Optional upper bound for scaling factor in dynamic
|
66 |
+
per token case
|
67 |
+
num_token_padding: If specified, pad the first dimension
|
68 |
+
of the output to at least this value.
|
69 |
+
use_per_token_if_dynamic: Whether to do per_tensor or per_token
|
70 |
+
in the dynamic quantization case.
|
71 |
+
|
72 |
+
Returns:
|
73 |
+
Tuple[torch.Tensor, torch.Tensor]: The output tensor in FP8 and
|
74 |
+
scaling factor.
|
75 |
+
"""
|
76 |
+
# This code assumes batch_dim and num_tokens are flattened
|
77 |
+
assert (input.ndim == 2)
|
78 |
+
shape: Union[Tuple[int, int], torch.Size] = input.shape
|
79 |
+
# For rocm, the output fp8 dtype is torch.float_e3m3fnuz
|
80 |
+
#out_dtype: torch.dtype = torch.float8_e4m3fnuz \
|
81 |
+
# if current_platform.is_rocm() else torch.float8_e4m3fn
|
82 |
+
out_dtype = torch.float8_e4m3fn
|
83 |
+
if num_token_padding:
|
84 |
+
shape = (max(num_token_padding, input.shape[0]), shape[1])
|
85 |
+
output = torch.empty(shape, device=input.device, dtype=out_dtype)
|
86 |
+
|
87 |
+
if scale is None:
|
88 |
+
if use_per_token_if_dynamic:
|
89 |
+
scale = torch.empty((shape[0], 1),
|
90 |
+
device=input.device,
|
91 |
+
dtype=torch.float32)
|
92 |
+
ops.dynamic_per_token_scaled_fp8_quant(
|
93 |
+
output, input, scale, scale_ub)
|
94 |
+
else:
|
95 |
+
scale = torch.zeros(1, device=input.device, dtype=torch.float32)
|
96 |
+
ops.dynamic_scaled_fp8_quant(output, input, scale)
|
97 |
+
else:
|
98 |
+
# num_token_padding not implemented for this case
|
99 |
+
assert (scale.numel() == 1 or num_token_padding is None)
|
100 |
+
ops.static_scaled_fp8_quant(output, input, scale)
|
101 |
+
|
102 |
+
return output, scale
|
103 |
+
|
104 |
+
# int8
|
105 |
+
def scaled_int8_quant(
|
106 |
+
input: torch.Tensor,
|
107 |
+
scale: Optional[torch.Tensor] = None,
|
108 |
+
azp: Optional[torch.Tensor] = None,
|
109 |
+
symmetric: bool = True
|
110 |
+
) -> Tuple[torch.Tensor, torch.Tensor, Optional[torch.Tensor]]:
|
111 |
+
"""
|
112 |
+
Quantize the input tensor to int8 and return the quantized tensor and scale, and maybe azp.
|
113 |
+
|
114 |
+
Args:
|
115 |
+
input: The input tensor to be quantized to int8.
|
116 |
+
scale: Optional scaling factor for the int8 quantization.
|
117 |
+
When not provided, we invoke dynamic-per-token quantization.
|
118 |
+
azp: Optional zero-point for the int8 quantization.
|
119 |
+
Must be provided for asymmetric quantization if `scale` is provided.
|
120 |
+
symmetric: Whether to use symmetric quantization (scale only, azp ignored).
|
121 |
+
|
122 |
+
Returns:
|
123 |
+
Tuple[torch.Tensor, torch.Tensor, Optional[torch.Tensor]] : Output int8 tensor, scales, and optionally azp.
|
124 |
+
"""
|
125 |
+
output = torch.empty_like(input, dtype=torch.int8)
|
126 |
+
if scale is not None:
|
127 |
+
# static-per-tensor quantization.
|
128 |
+
assert symmetric == (
|
129 |
+
azp is
|
130 |
+
None), "azp must only be provided for asymmetric quantization."
|
131 |
+
ops.static_scaled_int8_quant(output, input, scale, azp)
|
132 |
+
return output, scale, azp
|
133 |
+
|
134 |
+
# dynamic-per-token quantization.
|
135 |
+
input_scales = torch.empty((input.numel() // input.shape[-1], 1),
|
136 |
+
device=input.device,
|
137 |
+
dtype=torch.float32)
|
138 |
+
input_azp = None if symmetric else torch.empty_like(input_scales,
|
139 |
+
dtype=torch.int32)
|
140 |
+
ops.dynamic_scaled_int8_quant(output, input, input_scales,
|
141 |
+
input_azp)
|
142 |
+
return output, input_scales, input_azp
|
143 |
+
|
144 |
+
# fp8 marlin
|
145 |
+
def fp8_marlin_gemm(a: torch.Tensor, b_q_weight: torch.Tensor,
|
146 |
+
b_scales: torch.Tensor, workspace: torch.Tensor,
|
147 |
+
num_bits: int, size_m: int, size_n: int,
|
148 |
+
size_k: int) -> torch.Tensor:
|
149 |
+
return ops.fp8_marlin_gemm(a, b_q_weight, b_scales, workspace,
|
150 |
+
num_bits, size_m, size_n, size_k)
|
build/torch24-cxx98-cu124-x86_64-linux/quantization/_quantization_0_0_1.abi3.so
CHANGED
@@ -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 sha256:0b30e0d4e60669a253ed18f92b5219fdc3b40c2f0e792bf275c62a058998ebad
|
3 |
+
size 47404040
|
build/torch25-cxx11-cu118-x86_64-linux/quantization/__init__.py
CHANGED
@@ -1,4 +1,4 @@
|
|
1 |
-
from typing import Optional
|
2 |
|
3 |
import torch
|
4 |
|
@@ -42,3 +42,109 @@ def cutlass_scaled_mm(a: torch.Tensor,
|
|
42 |
|
43 |
return out
|
44 |
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
1 |
+
from typing import Optional, Tuple
|
2 |
|
3 |
import torch
|
4 |
|
|
|
42 |
|
43 |
return out
|
44 |
|
45 |
+
# fp8
|
46 |
+
def scaled_fp8_quant(
|
47 |
+
input: torch.Tensor,
|
48 |
+
scale: Optional[torch.Tensor] = None,
|
49 |
+
num_token_padding: Optional[int] = None,
|
50 |
+
scale_ub: Optional[torch.Tensor] = None,
|
51 |
+
use_per_token_if_dynamic: bool = False,
|
52 |
+
) -> Tuple[torch.Tensor, torch.Tensor]:
|
53 |
+
"""
|
54 |
+
Quantize input tensor to FP8 and return quantized tensor and scale.
|
55 |
+
|
56 |
+
This function supports both static and dynamic quantization: If you
|
57 |
+
provide the scale, it will use static scaling and if you omit it,
|
58 |
+
the scale will be determined dynamically. The function also allows
|
59 |
+
optional padding of the output tensors for downstream kernels that
|
60 |
+
will benefit from padding.
|
61 |
+
|
62 |
+
Args:
|
63 |
+
input: The input tensor to be quantized to FP8
|
64 |
+
scale: Optional scaling factor for the FP8 quantization
|
65 |
+
scale_ub: Optional upper bound for scaling factor in dynamic
|
66 |
+
per token case
|
67 |
+
num_token_padding: If specified, pad the first dimension
|
68 |
+
of the output to at least this value.
|
69 |
+
use_per_token_if_dynamic: Whether to do per_tensor or per_token
|
70 |
+
in the dynamic quantization case.
|
71 |
+
|
72 |
+
Returns:
|
73 |
+
Tuple[torch.Tensor, torch.Tensor]: The output tensor in FP8 and
|
74 |
+
scaling factor.
|
75 |
+
"""
|
76 |
+
# This code assumes batch_dim and num_tokens are flattened
|
77 |
+
assert (input.ndim == 2)
|
78 |
+
shape: Union[Tuple[int, int], torch.Size] = input.shape
|
79 |
+
# For rocm, the output fp8 dtype is torch.float_e3m3fnuz
|
80 |
+
#out_dtype: torch.dtype = torch.float8_e4m3fnuz \
|
81 |
+
# if current_platform.is_rocm() else torch.float8_e4m3fn
|
82 |
+
out_dtype = torch.float8_e4m3fn
|
83 |
+
if num_token_padding:
|
84 |
+
shape = (max(num_token_padding, input.shape[0]), shape[1])
|
85 |
+
output = torch.empty(shape, device=input.device, dtype=out_dtype)
|
86 |
+
|
87 |
+
if scale is None:
|
88 |
+
if use_per_token_if_dynamic:
|
89 |
+
scale = torch.empty((shape[0], 1),
|
90 |
+
device=input.device,
|
91 |
+
dtype=torch.float32)
|
92 |
+
ops.dynamic_per_token_scaled_fp8_quant(
|
93 |
+
output, input, scale, scale_ub)
|
94 |
+
else:
|
95 |
+
scale = torch.zeros(1, device=input.device, dtype=torch.float32)
|
96 |
+
ops.dynamic_scaled_fp8_quant(output, input, scale)
|
97 |
+
else:
|
98 |
+
# num_token_padding not implemented for this case
|
99 |
+
assert (scale.numel() == 1 or num_token_padding is None)
|
100 |
+
ops.static_scaled_fp8_quant(output, input, scale)
|
101 |
+
|
102 |
+
return output, scale
|
103 |
+
|
104 |
+
# int8
|
105 |
+
def scaled_int8_quant(
|
106 |
+
input: torch.Tensor,
|
107 |
+
scale: Optional[torch.Tensor] = None,
|
108 |
+
azp: Optional[torch.Tensor] = None,
|
109 |
+
symmetric: bool = True
|
110 |
+
) -> Tuple[torch.Tensor, torch.Tensor, Optional[torch.Tensor]]:
|
111 |
+
"""
|
112 |
+
Quantize the input tensor to int8 and return the quantized tensor and scale, and maybe azp.
|
113 |
+
|
114 |
+
Args:
|
115 |
+
input: The input tensor to be quantized to int8.
|
116 |
+
scale: Optional scaling factor for the int8 quantization.
|
117 |
+
When not provided, we invoke dynamic-per-token quantization.
|
118 |
+
azp: Optional zero-point for the int8 quantization.
|
119 |
+
Must be provided for asymmetric quantization if `scale` is provided.
|
120 |
+
symmetric: Whether to use symmetric quantization (scale only, azp ignored).
|
121 |
+
|
122 |
+
Returns:
|
123 |
+
Tuple[torch.Tensor, torch.Tensor, Optional[torch.Tensor]] : Output int8 tensor, scales, and optionally azp.
|
124 |
+
"""
|
125 |
+
output = torch.empty_like(input, dtype=torch.int8)
|
126 |
+
if scale is not None:
|
127 |
+
# static-per-tensor quantization.
|
128 |
+
assert symmetric == (
|
129 |
+
azp is
|
130 |
+
None), "azp must only be provided for asymmetric quantization."
|
131 |
+
ops.static_scaled_int8_quant(output, input, scale, azp)
|
132 |
+
return output, scale, azp
|
133 |
+
|
134 |
+
# dynamic-per-token quantization.
|
135 |
+
input_scales = torch.empty((input.numel() // input.shape[-1], 1),
|
136 |
+
device=input.device,
|
137 |
+
dtype=torch.float32)
|
138 |
+
input_azp = None if symmetric else torch.empty_like(input_scales,
|
139 |
+
dtype=torch.int32)
|
140 |
+
ops.dynamic_scaled_int8_quant(output, input, input_scales,
|
141 |
+
input_azp)
|
142 |
+
return output, input_scales, input_azp
|
143 |
+
|
144 |
+
# fp8 marlin
|
145 |
+
def fp8_marlin_gemm(a: torch.Tensor, b_q_weight: torch.Tensor,
|
146 |
+
b_scales: torch.Tensor, workspace: torch.Tensor,
|
147 |
+
num_bits: int, size_m: int, size_n: int,
|
148 |
+
size_k: int) -> torch.Tensor:
|
149 |
+
return ops.fp8_marlin_gemm(a, b_q_weight, b_scales, workspace,
|
150 |
+
num_bits, size_m, size_n, size_k)
|
build/torch25-cxx11-cu118-x86_64-linux/quantization/_quantization_0_0_1.abi3.so
CHANGED
@@ -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:cd1b0623532cab4059a48c8e8e7417df9ba309968adcd03705012fa79b04776d
|
3 |
+
size 39178896
|
build/torch25-cxx11-cu121-x86_64-linux/quantization/__init__.py
CHANGED
@@ -1,4 +1,4 @@
|
|
1 |
-
from typing import Optional
|
2 |
|
3 |
import torch
|
4 |
|
@@ -42,3 +42,109 @@ def cutlass_scaled_mm(a: torch.Tensor,
|
|
42 |
|
43 |
return out
|
44 |
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
1 |
+
from typing import Optional, Tuple
|
2 |
|
3 |
import torch
|
4 |
|
|
|
42 |
|
43 |
return out
|
44 |
|
45 |
+
# fp8
|
46 |
+
def scaled_fp8_quant(
|
47 |
+
input: torch.Tensor,
|
48 |
+
scale: Optional[torch.Tensor] = None,
|
49 |
+
num_token_padding: Optional[int] = None,
|
50 |
+
scale_ub: Optional[torch.Tensor] = None,
|
51 |
+
use_per_token_if_dynamic: bool = False,
|
52 |
+
) -> Tuple[torch.Tensor, torch.Tensor]:
|
53 |
+
"""
|
54 |
+
Quantize input tensor to FP8 and return quantized tensor and scale.
|
55 |
+
|
56 |
+
This function supports both static and dynamic quantization: If you
|
57 |
+
provide the scale, it will use static scaling and if you omit it,
|
58 |
+
the scale will be determined dynamically. The function also allows
|
59 |
+
optional padding of the output tensors for downstream kernels that
|
60 |
+
will benefit from padding.
|
61 |
+
|
62 |
+
Args:
|
63 |
+
input: The input tensor to be quantized to FP8
|
64 |
+
scale: Optional scaling factor for the FP8 quantization
|
65 |
+
scale_ub: Optional upper bound for scaling factor in dynamic
|
66 |
+
per token case
|
67 |
+
num_token_padding: If specified, pad the first dimension
|
68 |
+
of the output to at least this value.
|
69 |
+
use_per_token_if_dynamic: Whether to do per_tensor or per_token
|
70 |
+
in the dynamic quantization case.
|
71 |
+
|
72 |
+
Returns:
|
73 |
+
Tuple[torch.Tensor, torch.Tensor]: The output tensor in FP8 and
|
74 |
+
scaling factor.
|
75 |
+
"""
|
76 |
+
# This code assumes batch_dim and num_tokens are flattened
|
77 |
+
assert (input.ndim == 2)
|
78 |
+
shape: Union[Tuple[int, int], torch.Size] = input.shape
|
79 |
+
# For rocm, the output fp8 dtype is torch.float_e3m3fnuz
|
80 |
+
#out_dtype: torch.dtype = torch.float8_e4m3fnuz \
|
81 |
+
# if current_platform.is_rocm() else torch.float8_e4m3fn
|
82 |
+
out_dtype = torch.float8_e4m3fn
|
83 |
+
if num_token_padding:
|
84 |
+
shape = (max(num_token_padding, input.shape[0]), shape[1])
|
85 |
+
output = torch.empty(shape, device=input.device, dtype=out_dtype)
|
86 |
+
|
87 |
+
if scale is None:
|
88 |
+
if use_per_token_if_dynamic:
|
89 |
+
scale = torch.empty((shape[0], 1),
|
90 |
+
device=input.device,
|
91 |
+
dtype=torch.float32)
|
92 |
+
ops.dynamic_per_token_scaled_fp8_quant(
|
93 |
+
output, input, scale, scale_ub)
|
94 |
+
else:
|
95 |
+
scale = torch.zeros(1, device=input.device, dtype=torch.float32)
|
96 |
+
ops.dynamic_scaled_fp8_quant(output, input, scale)
|
97 |
+
else:
|
98 |
+
# num_token_padding not implemented for this case
|
99 |
+
assert (scale.numel() == 1 or num_token_padding is None)
|
100 |
+
ops.static_scaled_fp8_quant(output, input, scale)
|
101 |
+
|
102 |
+
return output, scale
|
103 |
+
|
104 |
+
# int8
|
105 |
+
def scaled_int8_quant(
|
106 |
+
input: torch.Tensor,
|
107 |
+
scale: Optional[torch.Tensor] = None,
|
108 |
+
azp: Optional[torch.Tensor] = None,
|
109 |
+
symmetric: bool = True
|
110 |
+
) -> Tuple[torch.Tensor, torch.Tensor, Optional[torch.Tensor]]:
|
111 |
+
"""
|
112 |
+
Quantize the input tensor to int8 and return the quantized tensor and scale, and maybe azp.
|
113 |
+
|
114 |
+
Args:
|
115 |
+
input: The input tensor to be quantized to int8.
|
116 |
+
scale: Optional scaling factor for the int8 quantization.
|
117 |
+
When not provided, we invoke dynamic-per-token quantization.
|
118 |
+
azp: Optional zero-point for the int8 quantization.
|
119 |
+
Must be provided for asymmetric quantization if `scale` is provided.
|
120 |
+
symmetric: Whether to use symmetric quantization (scale only, azp ignored).
|
121 |
+
|
122 |
+
Returns:
|
123 |
+
Tuple[torch.Tensor, torch.Tensor, Optional[torch.Tensor]] : Output int8 tensor, scales, and optionally azp.
|
124 |
+
"""
|
125 |
+
output = torch.empty_like(input, dtype=torch.int8)
|
126 |
+
if scale is not None:
|
127 |
+
# static-per-tensor quantization.
|
128 |
+
assert symmetric == (
|
129 |
+
azp is
|
130 |
+
None), "azp must only be provided for asymmetric quantization."
|
131 |
+
ops.static_scaled_int8_quant(output, input, scale, azp)
|
132 |
+
return output, scale, azp
|
133 |
+
|
134 |
+
# dynamic-per-token quantization.
|
135 |
+
input_scales = torch.empty((input.numel() // input.shape[-1], 1),
|
136 |
+
device=input.device,
|
137 |
+
dtype=torch.float32)
|
138 |
+
input_azp = None if symmetric else torch.empty_like(input_scales,
|
139 |
+
dtype=torch.int32)
|
140 |
+
ops.dynamic_scaled_int8_quant(output, input, input_scales,
|
141 |
+
input_azp)
|
142 |
+
return output, input_scales, input_azp
|
143 |
+
|
144 |
+
# fp8 marlin
|
145 |
+
def fp8_marlin_gemm(a: torch.Tensor, b_q_weight: torch.Tensor,
|
146 |
+
b_scales: torch.Tensor, workspace: torch.Tensor,
|
147 |
+
num_bits: int, size_m: int, size_n: int,
|
148 |
+
size_k: int) -> torch.Tensor:
|
149 |
+
return ops.fp8_marlin_gemm(a, b_q_weight, b_scales, workspace,
|
150 |
+
num_bits, size_m, size_n, size_k)
|
build/torch25-cxx11-cu121-x86_64-linux/quantization/_quantization_0_0_1.abi3.so
CHANGED
@@ -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:c129629d41aedd275b5a5545f86d968f0e9cc63085e54b09f0daff930af8f48c
|
3 |
+
size 46540064
|
build/torch25-cxx11-cu124-x86_64-linux/quantization/__init__.py
CHANGED
@@ -1,4 +1,4 @@
|
|
1 |
-
from typing import Optional
|
2 |
|
3 |
import torch
|
4 |
|
@@ -42,3 +42,109 @@ def cutlass_scaled_mm(a: torch.Tensor,
|
|
42 |
|
43 |
return out
|
44 |
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
1 |
+
from typing import Optional, Tuple
|
2 |
|
3 |
import torch
|
4 |
|
|
|
42 |
|
43 |
return out
|
44 |
|
45 |
+
# fp8
|
46 |
+
def scaled_fp8_quant(
|
47 |
+
input: torch.Tensor,
|
48 |
+
scale: Optional[torch.Tensor] = None,
|
49 |
+
num_token_padding: Optional[int] = None,
|
50 |
+
scale_ub: Optional[torch.Tensor] = None,
|
51 |
+
use_per_token_if_dynamic: bool = False,
|
52 |
+
) -> Tuple[torch.Tensor, torch.Tensor]:
|
53 |
+
"""
|
54 |
+
Quantize input tensor to FP8 and return quantized tensor and scale.
|
55 |
+
|
56 |
+
This function supports both static and dynamic quantization: If you
|
57 |
+
provide the scale, it will use static scaling and if you omit it,
|
58 |
+
the scale will be determined dynamically. The function also allows
|
59 |
+
optional padding of the output tensors for downstream kernels that
|
60 |
+
will benefit from padding.
|
61 |
+
|
62 |
+
Args:
|
63 |
+
input: The input tensor to be quantized to FP8
|
64 |
+
scale: Optional scaling factor for the FP8 quantization
|
65 |
+
scale_ub: Optional upper bound for scaling factor in dynamic
|
66 |
+
per token case
|
67 |
+
num_token_padding: If specified, pad the first dimension
|
68 |
+
of the output to at least this value.
|
69 |
+
use_per_token_if_dynamic: Whether to do per_tensor or per_token
|
70 |
+
in the dynamic quantization case.
|
71 |
+
|
72 |
+
Returns:
|
73 |
+
Tuple[torch.Tensor, torch.Tensor]: The output tensor in FP8 and
|
74 |
+
scaling factor.
|
75 |
+
"""
|
76 |
+
# This code assumes batch_dim and num_tokens are flattened
|
77 |
+
assert (input.ndim == 2)
|
78 |
+
shape: Union[Tuple[int, int], torch.Size] = input.shape
|
79 |
+
# For rocm, the output fp8 dtype is torch.float_e3m3fnuz
|
80 |
+
#out_dtype: torch.dtype = torch.float8_e4m3fnuz \
|
81 |
+
# if current_platform.is_rocm() else torch.float8_e4m3fn
|
82 |
+
out_dtype = torch.float8_e4m3fn
|
83 |
+
if num_token_padding:
|
84 |
+
shape = (max(num_token_padding, input.shape[0]), shape[1])
|
85 |
+
output = torch.empty(shape, device=input.device, dtype=out_dtype)
|
86 |
+
|
87 |
+
if scale is None:
|
88 |
+
if use_per_token_if_dynamic:
|
89 |
+
scale = torch.empty((shape[0], 1),
|
90 |
+
device=input.device,
|
91 |
+
dtype=torch.float32)
|
92 |
+
ops.dynamic_per_token_scaled_fp8_quant(
|
93 |
+
output, input, scale, scale_ub)
|
94 |
+
else:
|
95 |
+
scale = torch.zeros(1, device=input.device, dtype=torch.float32)
|
96 |
+
ops.dynamic_scaled_fp8_quant(output, input, scale)
|
97 |
+
else:
|
98 |
+
# num_token_padding not implemented for this case
|
99 |
+
assert (scale.numel() == 1 or num_token_padding is None)
|
100 |
+
ops.static_scaled_fp8_quant(output, input, scale)
|
101 |
+
|
102 |
+
return output, scale
|
103 |
+
|
104 |
+
# int8
|
105 |
+
def scaled_int8_quant(
|
106 |
+
input: torch.Tensor,
|
107 |
+
scale: Optional[torch.Tensor] = None,
|
108 |
+
azp: Optional[torch.Tensor] = None,
|
109 |
+
symmetric: bool = True
|
110 |
+
) -> Tuple[torch.Tensor, torch.Tensor, Optional[torch.Tensor]]:
|
111 |
+
"""
|
112 |
+
Quantize the input tensor to int8 and return the quantized tensor and scale, and maybe azp.
|
113 |
+
|
114 |
+
Args:
|
115 |
+
input: The input tensor to be quantized to int8.
|
116 |
+
scale: Optional scaling factor for the int8 quantization.
|
117 |
+
When not provided, we invoke dynamic-per-token quantization.
|
118 |
+
azp: Optional zero-point for the int8 quantization.
|
119 |
+
Must be provided for asymmetric quantization if `scale` is provided.
|
120 |
+
symmetric: Whether to use symmetric quantization (scale only, azp ignored).
|
121 |
+
|
122 |
+
Returns:
|
123 |
+
Tuple[torch.Tensor, torch.Tensor, Optional[torch.Tensor]] : Output int8 tensor, scales, and optionally azp.
|
124 |
+
"""
|
125 |
+
output = torch.empty_like(input, dtype=torch.int8)
|
126 |
+
if scale is not None:
|
127 |
+
# static-per-tensor quantization.
|
128 |
+
assert symmetric == (
|
129 |
+
azp is
|
130 |
+
None), "azp must only be provided for asymmetric quantization."
|
131 |
+
ops.static_scaled_int8_quant(output, input, scale, azp)
|
132 |
+
return output, scale, azp
|
133 |
+
|
134 |
+
# dynamic-per-token quantization.
|
135 |
+
input_scales = torch.empty((input.numel() // input.shape[-1], 1),
|
136 |
+
device=input.device,
|
137 |
+
dtype=torch.float32)
|
138 |
+
input_azp = None if symmetric else torch.empty_like(input_scales,
|
139 |
+
dtype=torch.int32)
|
140 |
+
ops.dynamic_scaled_int8_quant(output, input, input_scales,
|
141 |
+
input_azp)
|
142 |
+
return output, input_scales, input_azp
|
143 |
+
|
144 |
+
# fp8 marlin
|
145 |
+
def fp8_marlin_gemm(a: torch.Tensor, b_q_weight: torch.Tensor,
|
146 |
+
b_scales: torch.Tensor, workspace: torch.Tensor,
|
147 |
+
num_bits: int, size_m: int, size_n: int,
|
148 |
+
size_k: int) -> torch.Tensor:
|
149 |
+
return ops.fp8_marlin_gemm(a, b_q_weight, b_scales, workspace,
|
150 |
+
num_bits, size_m, size_n, size_k)
|
build/torch25-cxx11-cu124-x86_64-linux/quantization/_quantization_0_0_1.abi3.so
CHANGED
@@ -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:d691489858bbbac2d9031f49378d34d6c1ec3ed9592c933916cdb0fd470b4e54
|
3 |
+
size 47413592
|
build/torch25-cxx98-cu118-x86_64-linux/quantization/__init__.py
CHANGED
@@ -1,4 +1,4 @@
|
|
1 |
-
from typing import Optional
|
2 |
|
3 |
import torch
|
4 |
|
@@ -42,3 +42,109 @@ def cutlass_scaled_mm(a: torch.Tensor,
|
|
42 |
|
43 |
return out
|
44 |
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
1 |
+
from typing import Optional, Tuple
|
2 |
|
3 |
import torch
|
4 |
|
|
|
42 |
|
43 |
return out
|
44 |
|
45 |
+
# fp8
|
46 |
+
def scaled_fp8_quant(
|
47 |
+
input: torch.Tensor,
|
48 |
+
scale: Optional[torch.Tensor] = None,
|
49 |
+
num_token_padding: Optional[int] = None,
|
50 |
+
scale_ub: Optional[torch.Tensor] = None,
|
51 |
+
use_per_token_if_dynamic: bool = False,
|
52 |
+
) -> Tuple[torch.Tensor, torch.Tensor]:
|
53 |
+
"""
|
54 |
+
Quantize input tensor to FP8 and return quantized tensor and scale.
|
55 |
+
|
56 |
+
This function supports both static and dynamic quantization: If you
|
57 |
+
provide the scale, it will use static scaling and if you omit it,
|
58 |
+
the scale will be determined dynamically. The function also allows
|
59 |
+
optional padding of the output tensors for downstream kernels that
|
60 |
+
will benefit from padding.
|
61 |
+
|
62 |
+
Args:
|
63 |
+
input: The input tensor to be quantized to FP8
|
64 |
+
scale: Optional scaling factor for the FP8 quantization
|
65 |
+
scale_ub: Optional upper bound for scaling factor in dynamic
|
66 |
+
per token case
|
67 |
+
num_token_padding: If specified, pad the first dimension
|
68 |
+
of the output to at least this value.
|
69 |
+
use_per_token_if_dynamic: Whether to do per_tensor or per_token
|
70 |
+
in the dynamic quantization case.
|
71 |
+
|
72 |
+
Returns:
|
73 |
+
Tuple[torch.Tensor, torch.Tensor]: The output tensor in FP8 and
|
74 |
+
scaling factor.
|
75 |
+
"""
|
76 |
+
# This code assumes batch_dim and num_tokens are flattened
|
77 |
+
assert (input.ndim == 2)
|
78 |
+
shape: Union[Tuple[int, int], torch.Size] = input.shape
|
79 |
+
# For rocm, the output fp8 dtype is torch.float_e3m3fnuz
|
80 |
+
#out_dtype: torch.dtype = torch.float8_e4m3fnuz \
|
81 |
+
# if current_platform.is_rocm() else torch.float8_e4m3fn
|
82 |
+
out_dtype = torch.float8_e4m3fn
|
83 |
+
if num_token_padding:
|
84 |
+
shape = (max(num_token_padding, input.shape[0]), shape[1])
|
85 |
+
output = torch.empty(shape, device=input.device, dtype=out_dtype)
|
86 |
+
|
87 |
+
if scale is None:
|
88 |
+
if use_per_token_if_dynamic:
|
89 |
+
scale = torch.empty((shape[0], 1),
|
90 |
+
device=input.device,
|
91 |
+
dtype=torch.float32)
|
92 |
+
ops.dynamic_per_token_scaled_fp8_quant(
|
93 |
+
output, input, scale, scale_ub)
|
94 |
+
else:
|
95 |
+
scale = torch.zeros(1, device=input.device, dtype=torch.float32)
|
96 |
+
ops.dynamic_scaled_fp8_quant(output, input, scale)
|
97 |
+
else:
|
98 |
+
# num_token_padding not implemented for this case
|
99 |
+
assert (scale.numel() == 1 or num_token_padding is None)
|
100 |
+
ops.static_scaled_fp8_quant(output, input, scale)
|
101 |
+
|
102 |
+
return output, scale
|
103 |
+
|
104 |
+
# int8
|
105 |
+
def scaled_int8_quant(
|
106 |
+
input: torch.Tensor,
|
107 |
+
scale: Optional[torch.Tensor] = None,
|
108 |
+
azp: Optional[torch.Tensor] = None,
|
109 |
+
symmetric: bool = True
|
110 |
+
) -> Tuple[torch.Tensor, torch.Tensor, Optional[torch.Tensor]]:
|
111 |
+
"""
|
112 |
+
Quantize the input tensor to int8 and return the quantized tensor and scale, and maybe azp.
|
113 |
+
|
114 |
+
Args:
|
115 |
+
input: The input tensor to be quantized to int8.
|
116 |
+
scale: Optional scaling factor for the int8 quantization.
|
117 |
+
When not provided, we invoke dynamic-per-token quantization.
|
118 |
+
azp: Optional zero-point for the int8 quantization.
|
119 |
+
Must be provided for asymmetric quantization if `scale` is provided.
|
120 |
+
symmetric: Whether to use symmetric quantization (scale only, azp ignored).
|
121 |
+
|
122 |
+
Returns:
|
123 |
+
Tuple[torch.Tensor, torch.Tensor, Optional[torch.Tensor]] : Output int8 tensor, scales, and optionally azp.
|
124 |
+
"""
|
125 |
+
output = torch.empty_like(input, dtype=torch.int8)
|
126 |
+
if scale is not None:
|
127 |
+
# static-per-tensor quantization.
|
128 |
+
assert symmetric == (
|
129 |
+
azp is
|
130 |
+
None), "azp must only be provided for asymmetric quantization."
|
131 |
+
ops.static_scaled_int8_quant(output, input, scale, azp)
|
132 |
+
return output, scale, azp
|
133 |
+
|
134 |
+
# dynamic-per-token quantization.
|
135 |
+
input_scales = torch.empty((input.numel() // input.shape[-1], 1),
|
136 |
+
device=input.device,
|
137 |
+
dtype=torch.float32)
|
138 |
+
input_azp = None if symmetric else torch.empty_like(input_scales,
|
139 |
+
dtype=torch.int32)
|
140 |
+
ops.dynamic_scaled_int8_quant(output, input, input_scales,
|
141 |
+
input_azp)
|
142 |
+
return output, input_scales, input_azp
|
143 |
+
|
144 |
+
# fp8 marlin
|
145 |
+
def fp8_marlin_gemm(a: torch.Tensor, b_q_weight: torch.Tensor,
|
146 |
+
b_scales: torch.Tensor, workspace: torch.Tensor,
|
147 |
+
num_bits: int, size_m: int, size_n: int,
|
148 |
+
size_k: int) -> torch.Tensor:
|
149 |
+
return ops.fp8_marlin_gemm(a, b_q_weight, b_scales, workspace,
|
150 |
+
num_bits, size_m, size_n, size_k)
|
build/torch25-cxx98-cu118-x86_64-linux/quantization/_quantization_0_0_1.abi3.so
CHANGED
@@ -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:bc0a49e4a96613598d16cb392b3e5580c1461e2cb6ca291876aeeb4c1afeabf7
|
3 |
+
size 39166248
|
build/torch25-cxx98-cu121-x86_64-linux/quantization/__init__.py
CHANGED
@@ -1,4 +1,4 @@
|
|
1 |
-
from typing import Optional
|
2 |
|
3 |
import torch
|
4 |
|
@@ -42,3 +42,109 @@ def cutlass_scaled_mm(a: torch.Tensor,
|
|
42 |
|
43 |
return out
|
44 |
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
1 |
+
from typing import Optional, Tuple
|
2 |
|
3 |
import torch
|
4 |
|
|
|
42 |
|
43 |
return out
|
44 |
|
45 |
+
# fp8
|
46 |
+
def scaled_fp8_quant(
|
47 |
+
input: torch.Tensor,
|
48 |
+
scale: Optional[torch.Tensor] = None,
|
49 |
+
num_token_padding: Optional[int] = None,
|
50 |
+
scale_ub: Optional[torch.Tensor] = None,
|
51 |
+
use_per_token_if_dynamic: bool = False,
|
52 |
+
) -> Tuple[torch.Tensor, torch.Tensor]:
|
53 |
+
"""
|
54 |
+
Quantize input tensor to FP8 and return quantized tensor and scale.
|
55 |
+
|
56 |
+
This function supports both static and dynamic quantization: If you
|
57 |
+
provide the scale, it will use static scaling and if you omit it,
|
58 |
+
the scale will be determined dynamically. The function also allows
|
59 |
+
optional padding of the output tensors for downstream kernels that
|
60 |
+
will benefit from padding.
|
61 |
+
|
62 |
+
Args:
|
63 |
+
input: The input tensor to be quantized to FP8
|
64 |
+
scale: Optional scaling factor for the FP8 quantization
|
65 |
+
scale_ub: Optional upper bound for scaling factor in dynamic
|
66 |
+
per token case
|
67 |
+
num_token_padding: If specified, pad the first dimension
|
68 |
+
of the output to at least this value.
|
69 |
+
use_per_token_if_dynamic: Whether to do per_tensor or per_token
|
70 |
+
in the dynamic quantization case.
|
71 |
+
|
72 |
+
Returns:
|
73 |
+
Tuple[torch.Tensor, torch.Tensor]: The output tensor in FP8 and
|
74 |
+
scaling factor.
|
75 |
+
"""
|
76 |
+
# This code assumes batch_dim and num_tokens are flattened
|
77 |
+
assert (input.ndim == 2)
|
78 |
+
shape: Union[Tuple[int, int], torch.Size] = input.shape
|
79 |
+
# For rocm, the output fp8 dtype is torch.float_e3m3fnuz
|
80 |
+
#out_dtype: torch.dtype = torch.float8_e4m3fnuz \
|
81 |
+
# if current_platform.is_rocm() else torch.float8_e4m3fn
|
82 |
+
out_dtype = torch.float8_e4m3fn
|
83 |
+
if num_token_padding:
|
84 |
+
shape = (max(num_token_padding, input.shape[0]), shape[1])
|
85 |
+
output = torch.empty(shape, device=input.device, dtype=out_dtype)
|
86 |
+
|
87 |
+
if scale is None:
|
88 |
+
if use_per_token_if_dynamic:
|
89 |
+
scale = torch.empty((shape[0], 1),
|
90 |
+
device=input.device,
|
91 |
+
dtype=torch.float32)
|
92 |
+
ops.dynamic_per_token_scaled_fp8_quant(
|
93 |
+
output, input, scale, scale_ub)
|
94 |
+
else:
|
95 |
+
scale = torch.zeros(1, device=input.device, dtype=torch.float32)
|
96 |
+
ops.dynamic_scaled_fp8_quant(output, input, scale)
|
97 |
+
else:
|
98 |
+
# num_token_padding not implemented for this case
|
99 |
+
assert (scale.numel() == 1 or num_token_padding is None)
|
100 |
+
ops.static_scaled_fp8_quant(output, input, scale)
|
101 |
+
|
102 |
+
return output, scale
|
103 |
+
|
104 |
+
# int8
|
105 |
+
def scaled_int8_quant(
|
106 |
+
input: torch.Tensor,
|
107 |
+
scale: Optional[torch.Tensor] = None,
|
108 |
+
azp: Optional[torch.Tensor] = None,
|
109 |
+
symmetric: bool = True
|
110 |
+
) -> Tuple[torch.Tensor, torch.Tensor, Optional[torch.Tensor]]:
|
111 |
+
"""
|
112 |
+
Quantize the input tensor to int8 and return the quantized tensor and scale, and maybe azp.
|
113 |
+
|
114 |
+
Args:
|
115 |
+
input: The input tensor to be quantized to int8.
|
116 |
+
scale: Optional scaling factor for the int8 quantization.
|
117 |
+
When not provided, we invoke dynamic-per-token quantization.
|
118 |
+
azp: Optional zero-point for the int8 quantization.
|
119 |
+
Must be provided for asymmetric quantization if `scale` is provided.
|
120 |
+
symmetric: Whether to use symmetric quantization (scale only, azp ignored).
|
121 |
+
|
122 |
+
Returns:
|
123 |
+
Tuple[torch.Tensor, torch.Tensor, Optional[torch.Tensor]] : Output int8 tensor, scales, and optionally azp.
|
124 |
+
"""
|
125 |
+
output = torch.empty_like(input, dtype=torch.int8)
|
126 |
+
if scale is not None:
|
127 |
+
# static-per-tensor quantization.
|
128 |
+
assert symmetric == (
|
129 |
+
azp is
|
130 |
+
None), "azp must only be provided for asymmetric quantization."
|
131 |
+
ops.static_scaled_int8_quant(output, input, scale, azp)
|
132 |
+
return output, scale, azp
|
133 |
+
|
134 |
+
# dynamic-per-token quantization.
|
135 |
+
input_scales = torch.empty((input.numel() // input.shape[-1], 1),
|
136 |
+
device=input.device,
|
137 |
+
dtype=torch.float32)
|
138 |
+
input_azp = None if symmetric else torch.empty_like(input_scales,
|
139 |
+
dtype=torch.int32)
|
140 |
+
ops.dynamic_scaled_int8_quant(output, input, input_scales,
|
141 |
+
input_azp)
|
142 |
+
return output, input_scales, input_azp
|
143 |
+
|
144 |
+
# fp8 marlin
|
145 |
+
def fp8_marlin_gemm(a: torch.Tensor, b_q_weight: torch.Tensor,
|
146 |
+
b_scales: torch.Tensor, workspace: torch.Tensor,
|
147 |
+
num_bits: int, size_m: int, size_n: int,
|
148 |
+
size_k: int) -> torch.Tensor:
|
149 |
+
return ops.fp8_marlin_gemm(a, b_q_weight, b_scales, workspace,
|
150 |
+
num_bits, size_m, size_n, size_k)
|
build/torch25-cxx98-cu121-x86_64-linux/quantization/_quantization_0_0_1.abi3.so
CHANGED
@@ -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:c8b055cc65c6680f15caf698d12f7e0b87332132e7489b689381643873a518a0
|
3 |
+
size 46534608
|
build/torch25-cxx98-cu124-x86_64-linux/quantization/__init__.py
CHANGED
@@ -1,4 +1,4 @@
|
|
1 |
-
from typing import Optional
|
2 |
|
3 |
import torch
|
4 |
|
@@ -42,3 +42,109 @@ def cutlass_scaled_mm(a: torch.Tensor,
|
|
42 |
|
43 |
return out
|
44 |
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
1 |
+
from typing import Optional, Tuple
|
2 |
|
3 |
import torch
|
4 |
|
|
|
42 |
|
43 |
return out
|
44 |
|
45 |
+
# fp8
|
46 |
+
def scaled_fp8_quant(
|
47 |
+
input: torch.Tensor,
|
48 |
+
scale: Optional[torch.Tensor] = None,
|
49 |
+
num_token_padding: Optional[int] = None,
|
50 |
+
scale_ub: Optional[torch.Tensor] = None,
|
51 |
+
use_per_token_if_dynamic: bool = False,
|
52 |
+
) -> Tuple[torch.Tensor, torch.Tensor]:
|
53 |
+
"""
|
54 |
+
Quantize input tensor to FP8 and return quantized tensor and scale.
|
55 |
+
|
56 |
+
This function supports both static and dynamic quantization: If you
|
57 |
+
provide the scale, it will use static scaling and if you omit it,
|
58 |
+
the scale will be determined dynamically. The function also allows
|
59 |
+
optional padding of the output tensors for downstream kernels that
|
60 |
+
will benefit from padding.
|
61 |
+
|
62 |
+
Args:
|
63 |
+
input: The input tensor to be quantized to FP8
|
64 |
+
scale: Optional scaling factor for the FP8 quantization
|
65 |
+
scale_ub: Optional upper bound for scaling factor in dynamic
|
66 |
+
per token case
|
67 |
+
num_token_padding: If specified, pad the first dimension
|
68 |
+
of the output to at least this value.
|
69 |
+
use_per_token_if_dynamic: Whether to do per_tensor or per_token
|
70 |
+
in the dynamic quantization case.
|
71 |
+
|
72 |
+
Returns:
|
73 |
+
Tuple[torch.Tensor, torch.Tensor]: The output tensor in FP8 and
|
74 |
+
scaling factor.
|
75 |
+
"""
|
76 |
+
# This code assumes batch_dim and num_tokens are flattened
|
77 |
+
assert (input.ndim == 2)
|
78 |
+
shape: Union[Tuple[int, int], torch.Size] = input.shape
|
79 |
+
# For rocm, the output fp8 dtype is torch.float_e3m3fnuz
|
80 |
+
#out_dtype: torch.dtype = torch.float8_e4m3fnuz \
|
81 |
+
# if current_platform.is_rocm() else torch.float8_e4m3fn
|
82 |
+
out_dtype = torch.float8_e4m3fn
|
83 |
+
if num_token_padding:
|
84 |
+
shape = (max(num_token_padding, input.shape[0]), shape[1])
|
85 |
+
output = torch.empty(shape, device=input.device, dtype=out_dtype)
|
86 |
+
|
87 |
+
if scale is None:
|
88 |
+
if use_per_token_if_dynamic:
|
89 |
+
scale = torch.empty((shape[0], 1),
|
90 |
+
device=input.device,
|
91 |
+
dtype=torch.float32)
|
92 |
+
ops.dynamic_per_token_scaled_fp8_quant(
|
93 |
+
output, input, scale, scale_ub)
|
94 |
+
else:
|
95 |
+
scale = torch.zeros(1, device=input.device, dtype=torch.float32)
|
96 |
+
ops.dynamic_scaled_fp8_quant(output, input, scale)
|
97 |
+
else:
|
98 |
+
# num_token_padding not implemented for this case
|
99 |
+
assert (scale.numel() == 1 or num_token_padding is None)
|
100 |
+
ops.static_scaled_fp8_quant(output, input, scale)
|
101 |
+
|
102 |
+
return output, scale
|
103 |
+
|
104 |
+
# int8
|
105 |
+
def scaled_int8_quant(
|
106 |
+
input: torch.Tensor,
|
107 |
+
scale: Optional[torch.Tensor] = None,
|
108 |
+
azp: Optional[torch.Tensor] = None,
|
109 |
+
symmetric: bool = True
|
110 |
+
) -> Tuple[torch.Tensor, torch.Tensor, Optional[torch.Tensor]]:
|
111 |
+
"""
|
112 |
+
Quantize the input tensor to int8 and return the quantized tensor and scale, and maybe azp.
|
113 |
+
|
114 |
+
Args:
|
115 |
+
input: The input tensor to be quantized to int8.
|
116 |
+
scale: Optional scaling factor for the int8 quantization.
|
117 |
+
When not provided, we invoke dynamic-per-token quantization.
|
118 |
+
azp: Optional zero-point for the int8 quantization.
|
119 |
+
Must be provided for asymmetric quantization if `scale` is provided.
|
120 |
+
symmetric: Whether to use symmetric quantization (scale only, azp ignored).
|
121 |
+
|
122 |
+
Returns:
|
123 |
+
Tuple[torch.Tensor, torch.Tensor, Optional[torch.Tensor]] : Output int8 tensor, scales, and optionally azp.
|
124 |
+
"""
|
125 |
+
output = torch.empty_like(input, dtype=torch.int8)
|
126 |
+
if scale is not None:
|
127 |
+
# static-per-tensor quantization.
|
128 |
+
assert symmetric == (
|
129 |
+
azp is
|
130 |
+
None), "azp must only be provided for asymmetric quantization."
|
131 |
+
ops.static_scaled_int8_quant(output, input, scale, azp)
|
132 |
+
return output, scale, azp
|
133 |
+
|
134 |
+
# dynamic-per-token quantization.
|
135 |
+
input_scales = torch.empty((input.numel() // input.shape[-1], 1),
|
136 |
+
device=input.device,
|
137 |
+
dtype=torch.float32)
|
138 |
+
input_azp = None if symmetric else torch.empty_like(input_scales,
|
139 |
+
dtype=torch.int32)
|
140 |
+
ops.dynamic_scaled_int8_quant(output, input, input_scales,
|
141 |
+
input_azp)
|
142 |
+
return output, input_scales, input_azp
|
143 |
+
|
144 |
+
# fp8 marlin
|
145 |
+
def fp8_marlin_gemm(a: torch.Tensor, b_q_weight: torch.Tensor,
|
146 |
+
b_scales: torch.Tensor, workspace: torch.Tensor,
|
147 |
+
num_bits: int, size_m: int, size_n: int,
|
148 |
+
size_k: int) -> torch.Tensor:
|
149 |
+
return ops.fp8_marlin_gemm(a, b_q_weight, b_scales, workspace,
|
150 |
+
num_bits, size_m, size_n, size_k)
|
build/torch25-cxx98-cu124-x86_64-linux/quantization/_quantization_0_0_1.abi3.so
CHANGED
@@ -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 sha256:5e9f624b3f8585ca9c688a08b7191f1a2212cd1b64701ba9b0075deedd3fe3d4
|
3 |
+
size 47404040
|