File size: 13,819 Bytes
78360e7 |
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 99 100 101 102 103 104 105 106 107 108 109 110 111 112 113 114 115 116 117 118 119 120 121 122 123 124 125 126 127 128 129 130 131 132 133 134 135 136 137 138 139 140 141 142 143 144 145 146 147 148 149 150 151 152 153 154 155 156 157 158 159 160 161 162 163 164 165 166 167 168 169 170 171 172 173 174 175 176 177 178 179 180 181 182 183 184 185 186 187 188 189 190 191 192 193 194 195 196 197 198 199 200 201 202 203 204 205 206 207 208 209 210 211 212 213 214 215 216 217 218 219 220 221 222 223 224 225 226 227 228 229 230 231 232 233 234 235 236 237 238 239 240 241 242 243 244 245 246 247 248 249 250 251 252 253 254 255 256 257 258 259 260 261 262 263 264 265 266 267 268 269 270 271 272 273 274 275 276 277 278 279 280 281 282 283 284 285 286 287 288 289 290 291 292 293 294 295 296 297 298 299 300 301 302 303 304 305 306 307 308 309 310 311 312 313 314 315 316 317 318 319 320 321 322 323 324 325 326 327 328 329 330 331 332 333 334 335 336 337 338 339 340 341 342 343 344 345 346 347 348 349 350 351 352 353 354 355 356 357 358 359 360 361 362 363 364 365 366 367 368 369 370 371 372 373 374 375 376 377 378 379 380 381 382 383 384 385 386 387 388 389 390 391 392 393 394 395 396 397 398 399 400 401 402 403 404 |
# Copyright 2024-2025 The Alibaba Wan Team Authors. All rights reserved.
import torch
from importlib.metadata import version
from mmgp import offload
import torch.nn.functional as F
major, minor = torch.cuda.get_device_capability(None)
bfloat16_supported = major >= 8
try:
from xformers.ops import memory_efficient_attention
except ImportError:
memory_efficient_attention = None
try:
import flash_attn_interface
FLASH_ATTN_3_AVAILABLE = True
except ModuleNotFoundError:
FLASH_ATTN_3_AVAILABLE = False
try:
import flash_attn
FLASH_ATTN_2_AVAILABLE = True
except ModuleNotFoundError:
FLASH_ATTN_2_AVAILABLE = False
flash_attn = None
try:
from sageattention import sageattn_varlen
def sageattn_varlen_wrapper(
q,
k,
v,
cu_seqlens_q,
cu_seqlens_kv,
max_seqlen_q,
max_seqlen_kv,
):
return sageattn_varlen(q, k, v, cu_seqlens_q, cu_seqlens_kv, max_seqlen_q, max_seqlen_kv)
except ImportError:
sageattn_varlen_wrapper = None
import warnings
try:
from sageattention import sageattn
from .sage2_core import sageattn as alt_sageattn, is_sage2_supported
sage2_supported = is_sage2_supported()
except ImportError:
sageattn = None
alt_sageattn = None
sage2_supported = False
# @torch.compiler.disable()
def sageattn_wrapper(
qkv_list,
attention_length
):
q,k, v = qkv_list
if True:
qkv_list = [q,k,v]
del q, k ,v
o = alt_sageattn(qkv_list, tensor_layout="NHD")
else:
o = sageattn(q, k, v, tensor_layout="NHD")
del q, k ,v
qkv_list.clear()
return o
# try:
# if True:
# from .sage2_core import sageattn_qk_int8_pv_fp8_window_cuda
# @torch.compiler.disable()
# def sageattn_window_wrapper(
# qkv_list,
# attention_length,
# window
# ):
# q,k, v = qkv_list
# padding_length = q.shape[0] -attention_length
# q = q[:attention_length, :, : ].unsqueeze(0)
# k = k[:attention_length, :, : ].unsqueeze(0)
# v = v[:attention_length, :, : ].unsqueeze(0)
# qkvl_list = [q, k , v]
# del q, k ,v
# o = sageattn_qk_int8_pv_fp8_window_cuda(qkvl_list, tensor_layout="NHD", window = window).squeeze(0)
# qkv_list.clear()
# if padding_length > 0:
# o = torch.cat([o, torch.empty( (padding_length, *o.shape[-2:]), dtype= o.dtype, device=o.device ) ], 0)
# return o
# except ImportError:
# sageattn = sageattn_qk_int8_pv_fp8_window_cuda
@torch.compiler.disable()
def sdpa_wrapper(
qkv_list,
attention_length,
attention_mask = None
):
q, k, v = qkv_list
q = q.transpose(1,2)
k = k.transpose(1,2)
v = v.transpose(1,2)
if attention_mask != None:
attention_mask = attention_mask.transpose(1,2)
o = F.scaled_dot_product_attention( q, k, v, attn_mask=attention_mask, is_causal=False).transpose(1,2)
del q, k ,v
qkv_list.clear()
return o
def get_attention_modes():
ret = ["sdpa", "auto"]
if flash_attn != None:
ret.append("flash")
if memory_efficient_attention != None:
ret.append("xformers")
if sageattn_varlen_wrapper != None:
ret.append("sage")
if sageattn != None and version("sageattention").startswith("2") :
ret.append("sage2")
return ret
def get_supported_attention_modes():
ret = get_attention_modes()
if not sage2_supported:
if "sage2" in ret:
ret.remove("sage2")
major, minor = torch.cuda.get_device_capability()
if major < 7:
if "sage" in ret:
ret.remove("sage")
return ret
__all__ = [
'pay_attention',
'attention',
]
def get_cu_seqlens(batch_size, lens, max_len):
cu_seqlens = torch.zeros([2 * batch_size + 1], dtype=torch.int32, device="cuda")
for i in range(batch_size):
s = lens[i]
s1 = i * max_len + s
s2 = (i + 1) * max_len
cu_seqlens[2 * i + 1] = s1
cu_seqlens[2 * i + 2] = s2
return cu_seqlens
@torch.compiler.disable()
def pay_attention(
qkv_list,
dropout_p=0.,
softmax_scale=None,
causal=False,
window_size=(-1, -1),
deterministic=False,
version=None,
force_attention= None,
attention_mask = None,
cross_attn= False,
q_lens = None,
k_lens = None,
):
# format : torch.Size([batches, tokens, heads, head_features])
# assume if q_lens is non null, each q is padded up to lq (one q out of two will need to be discarded or ignored)
# assume if k_lens is non null, each k is padded up to lk (one k out of two will need to be discarded or ignored)
if attention_mask != None:
force_attention = "sdpa"
if attention_mask.dtype == torch.bfloat16 and not bfloat16_supported:
attention_mask = attention_mask.to(torch.float16)
attn = offload.shared_state["_attention"] if force_attention== None else force_attention
q,k,v = qkv_list
qkv_list.clear()
out_dtype = q.dtype
if q.dtype == torch.bfloat16 and not bfloat16_supported:
q = q.to(torch.float16)
k = k.to(torch.float16)
v = v.to(torch.float16)
final_padding = 0
b, lq, lk = q.size(0), q.size(1), k.size(1)
q = q.to(v.dtype)
k = k.to(v.dtype)
if attn == "chipmunk":
from src.chipmunk.modules import SparseDiffMlp, SparseDiffAttn
from src.chipmunk.util import LayerCounter, GLOBAL_CONFIG
if b > 1 and k_lens != None and attn in ("sage2", "sdpa"):
assert attention_mask == None
# Poor's man var k len attention
assert q_lens == None
chunk_sizes = []
k_sizes = []
current_size = k_lens[0]
current_count= 1
for k_len in k_lens[1:]:
if k_len == current_size:
current_count += 1
else:
chunk_sizes.append(current_count)
k_sizes.append(current_size)
current_count = 1
current_size = k_len
chunk_sizes.append(current_count)
k_sizes.append(k_len)
if len(chunk_sizes) > 1 or k_lens[0] != k.shape[1]:
q_chunks =torch.split(q, chunk_sizes)
k_chunks =torch.split(k, chunk_sizes)
v_chunks =torch.split(v, chunk_sizes)
q, k, v = None, None, None
k_chunks = [ u[:, :sz] for u, sz in zip(k_chunks, k_sizes)]
v_chunks = [ u[:, :sz] for u, sz in zip(v_chunks, k_sizes)]
o = []
for sub_q, sub_k, sub_v in zip(q_chunks, k_chunks, v_chunks):
qkv_list = [sub_q, sub_k, sub_v]
sub_q, sub_k, sub_v = None, None, None
o.append( pay_attention(qkv_list) )
q_chunks, k_chunks, v_chunks = None, None, None
o = torch.cat(o, dim = 0)
return o
elif (q_lens != None or k_lens != None) and attn in ("sage2", "sdpa"):
assert b == 1
szq = q_lens[0].item() if q_lens != None else lq
szk = k_lens[0].item() if k_lens != None else lk
final_padding = lq - szq
q = q[:, :szq]
k = k[:, :szk]
v = v[:, :szk]
if version is not None and version == 3 and not FLASH_ATTN_3_AVAILABLE:
warnings.warn(
'Flash attention 3 is not available, use flash attention 2 instead.'
)
if attn=="sage" or attn=="flash":
if b != 1 :
if k_lens == None:
k_lens = torch.tensor( [lk] * b, dtype=torch.int32).to(device=q.device, non_blocking=True)
if q_lens == None:
q_lens = torch.tensor([lq] * b, dtype=torch.int32).to(device=q.device, non_blocking=True)
k = k.reshape(-1, *k.shape[-2:])
v = v.reshape(-1, *v.shape[-2:])
q = q.reshape(-1, *q.shape[-2:])
cu_seqlens_q=get_cu_seqlens(b, q_lens, lq)
cu_seqlens_k=get_cu_seqlens(b, k_lens, lk)
else:
szq = q_lens[0].item() if q_lens != None else lq
szk = k_lens[0].item() if k_lens != None else lk
if szq != lq or szk != lk:
cu_seqlens_q = torch.tensor([0, szq, lq], dtype=torch.int32, device="cuda")
cu_seqlens_k = torch.tensor([0, szk, lk], dtype=torch.int32, device="cuda")
else:
cu_seqlens_q = torch.tensor([0, lq], dtype=torch.int32, device="cuda")
cu_seqlens_k = torch.tensor([0, lk], dtype=torch.int32, device="cuda")
q = q.squeeze(0)
k = k.squeeze(0)
v = v.squeeze(0)
# apply attention
if attn=="sage":
x = sageattn_varlen_wrapper(
q=q,
k=k,
v=v,
cu_seqlens_q= cu_seqlens_q,
cu_seqlens_kv= cu_seqlens_k,
max_seqlen_q=lq,
max_seqlen_kv=lk,
).unflatten(0, (b, lq))
elif attn=="sage2":
import math
if cross_attn or True:
qkv_list = [q,k,v]
del q,k,v
x = sageattn_wrapper(qkv_list, lq) #.unsqueeze(0)
# else:
# layer = offload.shared_state["layer"]
# embed_sizes = offload.shared_state["embed_sizes"]
# current_step = offload.shared_state["step_no"]
# max_steps = offload.shared_state["max_steps"]
# nb_latents = embed_sizes[0] * embed_sizes[1]* embed_sizes[2]
# window = 0
# start_window_step = int(max_steps * 0.3)
# start_layer = 10
# end_layer = 30
# if (layer < start_layer or layer > end_layer ) or current_step <start_window_step:
# window = 0
# else:
# # coef = min((max_steps - current_step)/(max_steps-start_window_step),1)*max(min((25 - layer)/(25-start_layer),1),0) * 0.7 + 0.3
# coef = 0.3
# print(f"step: {current_step}, layer: {layer}, coef:{coef:0.1f}]")
# window = math.ceil(coef* nb_latents)
# invert_spaces = (layer + current_step) % 2 == 0 and window > 0
# invert_spaces = False
# def flip(q):
# q = q.reshape(*embed_sizes, *q.shape[-2:])
# q = q.transpose(0,2)
# q = q.contiguous()
# q = q.transpose(0,2)
# q = q.reshape( -1, *q.shape[-2:])
# return q
# def flop(q):
# q = q.reshape(embed_sizes[2], embed_sizes[1], embed_sizes[0] , *q.shape[-2:])
# q = q.transpose(0,2)
# q = q.contiguous()
# q = q.transpose(0,2)
# q = q.reshape( -1, *q.shape[-2:])
# return q
# if invert_spaces:
# q = flip(q)
# k = flip(k)
# v = flip(v)
# qkv_list = [q,k,v]
# del q,k,v
# x = sageattn_window_wrapper(qkv_list, lq, window= window) #.unsqueeze(0)
# if invert_spaces:
# x = flop(x)
# x = x.unsqueeze(0)
elif attn=="sdpa":
qkv_list = [q, k, v]
del q ,k ,v
x = sdpa_wrapper( qkv_list, lq, attention_mask = attention_mask) #.unsqueeze(0)
elif attn=="flash" and version == 3:
# Note: dropout_p, window_size are not supported in FA3 now.
x = flash_attn_interface.flash_attn_varlen_func(
q=q,
k=k,
v=v,
cu_seqlens_q= cu_seqlens_q,
cu_seqlens_k= cu_seqlens_k,
seqused_q=None,
seqused_k=None,
max_seqlen_q=lq,
max_seqlen_k=lk,
softmax_scale=softmax_scale,
causal=causal,
deterministic=deterministic)[0].unflatten(0, (b, lq))
elif attn=="flash":
x = flash_attn.flash_attn_varlen_func(
q=q,
k=k,
v=v,
cu_seqlens_q= cu_seqlens_q,
cu_seqlens_k= cu_seqlens_k,
max_seqlen_q=lq,
max_seqlen_k=lk,
dropout_p=dropout_p,
softmax_scale=softmax_scale,
causal=causal,
window_size=window_size,
deterministic=deterministic).unflatten(0, (b, lq))
# output
elif attn=="xformers":
from xformers.ops.fmha.attn_bias import BlockDiagonalPaddedKeysMask
if k_lens == None and q_lens == None:
x = memory_efficient_attention(q, k, v )
elif k_lens != None and q_lens == None:
attn_mask = BlockDiagonalPaddedKeysMask.from_seqlens([lq] * b , lk , list(k_lens) )
x = memory_efficient_attention(q, k, v, attn_bias= attn_mask )
elif b == 1:
szq = q_lens[0].item() if q_lens != None else lq
szk = k_lens[0].item() if k_lens != None else lk
attn_mask = BlockDiagonalPaddedKeysMask.from_seqlens([szq, lq - szq ] , lk , [szk, 0] )
x = memory_efficient_attention(q, k, v, attn_bias= attn_mask )
else:
assert False
x = x.type(out_dtype)
if final_padding > 0:
x = torch.cat([x, torch.empty( (x.shape[0], final_padding, *x.shape[-2:]), dtype= x.dtype, device=x.device ) ], 1)
return x |