File size: 11,631 Bytes
d643072 |
1 2 3 4 5 6 7 8 9 10 11 12 13 14 15 16 17 18 19 20 21 22 23 24 25 26 27 28 29 30 31 32 33 34 35 36 37 38 39 40 41 42 43 44 45 46 47 48 49 50 51 52 53 54 55 56 57 58 59 60 61 62 63 64 65 66 67 68 69 70 71 72 73 74 75 76 77 78 79 80 81 82 83 84 85 86 87 88 89 90 91 92 93 94 95 96 97 98 99 100 101 102 103 104 105 106 107 108 109 110 111 112 113 114 115 116 117 118 119 120 121 122 123 124 125 126 127 128 129 130 131 132 133 134 135 136 137 138 139 140 141 142 143 144 145 146 147 148 149 150 151 152 153 154 155 156 157 158 159 160 161 162 163 164 165 166 167 168 169 170 171 172 173 174 175 176 177 178 179 180 181 182 183 184 185 186 187 188 189 190 191 192 193 194 195 196 197 198 199 200 201 202 203 204 205 206 207 208 209 210 211 212 213 214 215 216 217 218 219 220 221 222 223 224 225 226 227 228 229 230 231 232 233 234 235 236 237 238 239 240 241 242 243 244 245 246 247 248 249 250 251 252 253 254 255 256 257 258 259 260 261 262 263 264 265 266 267 268 269 270 271 272 273 274 275 276 277 278 279 280 281 282 283 284 285 286 287 288 289 290 291 292 293 294 295 296 297 298 299 300 301 302 303 304 305 306 307 308 309 310 311 312 313 314 315 316 317 318 319 320 321 322 |
# Copyright 2024 MIT Han Lab
#
# Licensed under the Apache License, Version 2.0 (the "License");
# you may not use this file except in compliance with the License.
# You may obtain a copy of the License at
#
# http://www.apache.org/licenses/LICENSE-2.0
#
# Unless required by applicable law or agreed to in writing, software
# distributed under the License is distributed on an "AS IS" BASIS,
# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
# See the License for the specific language governing permissions and
# limitations under the License.
#
# SPDX-License-Identifier: Apache-2.0
import time
from dataclasses import dataclass
import ipdb
import torch
from modules.flash_attn import FlashAttention
from modules.lite_mla import LiteMLA
from modules.triton_lite_mla import TritonLiteMLA
from modules.triton_lite_mla_fwd import TritonLiteMLAFwd
from modules.utils.dtype import get_dtype_from_str
from modules.utils.export_onnx import export_onnx
from omegaconf import OmegaConf
from torch import nn
from torch.nn import functional as F
from torchprofile import profile_macs
@dataclass
class DevelopTritonLiteMLAConfig:
batch_size: int = 16
input_size: int = 1024 // 8 // 2
num_channels: int = 1152
num_heads: int = 36
attn_type: str = "LiteMLA"
device: str = "cuda"
dtype: str = "fp16"
profile_macs: bool = False
test_correctness: bool = False
warmup_iterations: int = 50
iterations: int = 1000
random_weight: bool = True
backward: bool = False
autocast: bool = False
use_cuda_graph: bool = False
export_model: bool = False
opset: int = 17
export_path: str = ""
export_dtype: str = "fp32"
export_device: str = "cuda"
def simulate_litemla(
x: torch.Tensor,
qkv_weight: torch.Tensor,
proj_weight: torch.Tensor,
proj_bias: torch.Tensor,
num_heads: int,
head_dim: int,
eps: float,
backward: bool,
):
B, N, C = x.shape
qkv = F.linear(x, qkv_weight).reshape(B, N, 3, C).permute(0, 2, 3, 1)
q, k, v = qkv.unbind(1) # B, 3, C, N --> B, C, N
q = q.reshape(B, C // head_dim, head_dim, N) # b, h, h_d, N
k = k.reshape(B, C // head_dim, head_dim, N).transpose(-1, -2) # b, h, N, h_d
v = v.reshape(B, C // head_dim, head_dim, N) # b, h, h_d, N
q = F.relu(q) # B, h, h_d, N
k = F.relu(k)
q, k, v = q.float(), k.float(), v.float()
if backward:
k.retain_grad()
v.retain_grad()
q.retain_grad()
v_pad = F.pad(v, (0, 0, 0, 1), mode="constant", value=1)
vk = torch.matmul(v_pad, k)
if backward:
vk.retain_grad()
vk_q = torch.matmul(vk, q)
vk_q_numerator, vk_q_denominator = vk_q[:, :, :-1], vk_q[:, :, -1:]
if backward:
vk_q_numerator.retain_grad()
vk_q_denominator.retain_grad()
vk_q_divide = (vk_q_numerator / (vk_q_denominator + eps)).to(x.dtype)
proj_input = vk_q_divide.view(B, C, N).permute(0, 2, 1) # B, N, C
if backward:
proj_input.retain_grad()
y = F.linear(proj_input, proj_weight, proj_bias)
output_dict = {
"q": q,
"k": k,
"v": v,
"vk": vk,
"proj_input": proj_input,
"vk_q_numerator": vk_q_numerator,
"vk_q_denominator": vk_q_denominator,
"vk_q_divide": vk_q_divide,
"y": y,
}
return output_dict
def main():
torch.backends.cuda.matmul.allow_tf32 = True
torch.backends.cudnn.allow_tf32 = True
LiteMLA.fp32_attention = True
torch.cuda.manual_seed(0)
torch.manual_seed(0)
cfg = OmegaConf.structured(DevelopTritonLiteMLAConfig)
cli_cfg = OmegaConf.from_cli()
cfg = OmegaConf.merge(cfg, OmegaConf.masked_copy(cli_cfg, cfg.keys()))
cfg: DevelopTritonLiteMLAConfig = OmegaConf.to_object(cfg)
torch.set_grad_enabled(cfg.backward)
device = torch.device("cuda")
if cfg.autocast:
dtype = torch.float32
autocast_dtype = get_dtype_from_str(cfg.dtype)
else:
dtype = get_dtype_from_str(cfg.dtype)
autocast_dtype = None
if cfg.attn_type == "LiteMLA":
block = LiteMLA(cfg.num_channels, cfg.num_channels, dim=cfg.num_channels // cfg.num_heads, eps=1e-8)
elif cfg.attn_type == "TritonLiteMLA":
block = TritonLiteMLA(cfg.num_channels, cfg.num_heads, eps=1e-8)
elif cfg.attn_type == "TritonLiteMLAFwd":
block = TritonLiteMLAFwd(cfg.num_channels, cfg.num_heads, eps=1e-8)
elif cfg.attn_type == "FlashAttention":
block = FlashAttention(cfg.num_channels, cfg.num_heads)
else:
raise NotImplementedError
if not cfg.backward:
block = block.eval()
block = block.to(device=device, dtype=dtype, memory_format=torch.channels_last)
if cfg.random_weight:
for param in block.parameters():
nn.init.trunc_normal_(param, std=0.001)
if cfg.profile_macs:
macs = profile_macs(block, x)
print(f"macs: {macs}")
if cfg.export_model:
export_dtype = get_dtype_from_str(cfg.export_dtype)
export_device = torch.device(cfg.export_device)
assert cfg.export_path != ""
export_onnx(
block.to(device=export_device, dtype=export_dtype),
(1, cfg.input_size**2, cfg.num_channels),
cfg.export_path,
cfg.opset,
export_dtype,
export_device,
)
if cfg.test_correctness:
ref_block = (
LiteMLA(cfg.num_channels, cfg.num_channels, dim=cfg.num_channels // cfg.num_heads, eps=1e-8)
.eval()
.to(device=device, memory_format=torch.channels_last)
)
block.load_state_dict(ref_block.state_dict())
correct = True
for i in range(10):
ref_x = torch.randn(
cfg.batch_size, cfg.input_size**2, cfg.num_channels, device=device, requires_grad=cfg.backward
)
x = ref_x.clone().detach().to(dtype=dtype).requires_grad_(cfg.backward)
with torch.autocast(device_type="cuda", dtype=autocast_dtype, enabled=cfg.autocast):
output = block(x)
ref_output_dict = simulate_litemla(
ref_x,
ref_block.qkv.weight,
ref_block.proj.weight,
ref_block.proj.bias,
ref_block.in_dim // ref_block.dim,
ref_block.dim,
ref_block.eps,
cfg.backward,
)
ref_output = ref_output_dict["y"]
if cfg.backward:
dy = 0.1 * torch.randn_like(output)
output.backward(dy)
ref_output.backward(dy.float())
# ipdb.set_trace()
ref_output_1 = ref_block(ref_x)
assert torch.allclose(ref_output, ref_output_1)
output_float = output.float()
if not torch.allclose(output_float, ref_output):
correct = False
max_error_pos = (output_float - ref_output).abs().view(-1).argmax()
print(f"comparing forward results")
print(
f"max error: {(output_float - ref_output).abs().max()}, mean error: {(output_float - ref_output).abs().mean()}"
)
print(f"max error pos: {ref_output.view(-1)[max_error_pos]} {output_float.view(-1)[max_error_pos]}")
if cfg.backward:
for name, grad, ref_grad in [
("proj_weight", block.proj.weight.grad, ref_block.proj.weight.grad),
("proj_bias", block.proj.bias.grad, ref_block.proj.bias.grad),
("qkv_weight", block.qkv.weight.grad, ref_block.qkv.weight.grad),
("x", x.grad, ref_x.grad),
]:
print(f"comparing {name}")
grad_float = grad.float()
max_error_pos = (grad_float - ref_grad).abs().view(-1).argmax()
print(
f"max error: {(grad_float - ref_grad).abs().max()}, mean error: {(grad_float - ref_grad).abs().mean()}"
)
print(f"max error pos: {ref_grad.view(-1)[max_error_pos]} {grad_float.view(-1)[max_error_pos]}")
# ipdb.set_trace()
if correct:
print("correct!")
elif cfg.use_cuda_graph:
x = torch.randn(
cfg.batch_size,
cfg.input_size**2,
cfg.num_channels,
device=device,
dtype=dtype,
requires_grad=cfg.backward,
)
grad_y = 0.1 * torch.randn_like(x)
s = torch.cuda.Stream()
s.wait_stream(torch.cuda.current_stream())
with torch.cuda.stream(s):
for i in range(cfg.warmup_iterations):
with torch.autocast(device_type="cuda", dtype=autocast_dtype, enabled=cfg.autocast):
y = block(x)
if cfg.backward:
y.backward(grad_y)
torch.cuda.current_stream().wait_stream(s)
g = torch.cuda.CUDAGraph()
# Sets grads to None before capture, so backward() will create
# .grad attributes with allocations from the graph's private pool
with torch.cuda.graph(g):
with torch.autocast(device_type="cuda", dtype=autocast_dtype, enabled=cfg.autocast):
y = block(x)
if cfg.backward:
y.backward(grad_y)
torch.cuda.synchronize()
start_time = time.time()
for i in range(cfg.iterations):
g.replay()
torch.cuda.synchronize()
end_time = time.time()
print(f"using cuda graph:")
print(f"each step takes {(end_time-start_time)*1000/cfg.iterations:.2f} ms")
print(f"max memory allocated: {torch.cuda.max_memory_allocated()/1024**3:.4f} GB")
else:
x = torch.randn(
cfg.batch_size,
cfg.input_size**2,
cfg.num_channels,
device=device,
dtype=dtype,
requires_grad=cfg.backward,
)
grad_y = 0.1 * torch.randn_like(x)
for i in range(cfg.warmup_iterations):
# ipdb.set_trace()
with torch.autocast(device_type="cuda", dtype=autocast_dtype, enabled=cfg.autocast):
y = block(x)
if cfg.backward:
y.backward(grad_y)
torch.cuda.synchronize()
start_time = time.time()
for i in range(cfg.iterations):
with torch.autocast(device_type="cuda", dtype=autocast_dtype, enabled=cfg.autocast):
y = block(x)
if cfg.backward:
y.backward(grad_y)
torch.cuda.synchronize()
end_time = time.time()
print(f"each step takes {(end_time - start_time) * 1000 / cfg.iterations:.2f} ms")
# ipdb.set_trace()
print(f"max memory allocated: {torch.cuda.max_memory_allocated() / 1024 ** 3:.4f} GB")
# x = torch.randn(cfg.batch_size*2, (cfg.input_size*2)**2, cfg.num_channels, device=device, dtype=dtype, requires_grad=cfg.backward)
# grad_y = 0.1*torch.randn_like(x)
# with torch.autocast(device_type="cuda", dtype=autocast_dtype, enabled=cfg.autocast):
# y = block(x)
# if cfg.backward:
# y.backward(grad_y)
if __name__ == "__main__":
main()
"""
# 64x64 fp16
python -m develop_triton_litemla attn_type=LiteMLA test_correctness=True
each step takes 10.81 ms
max memory allocated: 2.2984 GB
python -m develop_triton_litemla attn_type=TritonLiteMLA test_correctness=True
each step takes 4.70 ms
max memory allocated: 1.6480 GB
"""
|