File size: 10,249 Bytes
b66f552 | 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 | # Copyright (c) 2023-2025, Songlin Yang, Yu Zhang
import torch
from fla.modules.l2norm import l2norm_bwd, l2norm_fwd
from fla.ops.common.chunk_delta_h import chunk_gated_delta_rule_bwd_dhu, chunk_gated_delta_rule_fwd_h
from fla.ops.common.chunk_o import chunk_bwd_dv_local
from fla.ops.gla.chunk import chunk_gla_bwd_dA, chunk_gla_fwd_o_gk
from fla.ops.kda.chunk_inter import chunk_kda_bwd_dqkwg
from fla.ops.kda.chunk_intra import chunk_kda_bwd_intra, chunk_kda_fwd_intra
from fla.ops.kda.wy_fast import prepare_wy_repr_bwd, recompute_w_u_fwd
from fla.ops.utils import chunk_local_cumsum
from fla.utils import autocast_custom_bwd, autocast_custom_fwd, input_guard
def chunk_kda_fwd(
q: torch.Tensor,
k: torch.Tensor,
v: torch.Tensor,
g: torch.Tensor,
beta: torch.Tensor,
scale: float,
initial_state: torch.Tensor,
output_final_state: bool,
cu_seqlens: torch.LongTensor | None = None,
):
chunk_size = 64
g = chunk_local_cumsum(g, chunk_size=chunk_size, cu_seqlens=cu_seqlens)
# the intra Aqk is kept in fp32
# the computation has very marginal effect on the entire throughput
Aqk, Akk = chunk_kda_fwd_intra(
q=q,
k=k,
gk=g,
beta=beta,
scale=scale,
cu_seqlens=cu_seqlens,
output_dtype=torch.float32,
)
w, u, _, kg = recompute_w_u_fwd(
k=k,
v=v,
beta=beta,
A=Akk,
gk=g,
cu_seqlens=cu_seqlens,
)
h, v_new, final_state = chunk_gated_delta_rule_fwd_h(
k=kg,
w=w,
u=u,
gk=g,
initial_state=initial_state,
output_final_state=output_final_state,
cu_seqlens=cu_seqlens,
)
o = chunk_gla_fwd_o_gk(
q=q,
v=v_new,
g=g,
A=Aqk,
h=h,
scale=scale,
cu_seqlens=cu_seqlens,
chunk_size=chunk_size,
)
return g, o, Aqk, Akk, final_state
def chunk_kda_bwd(
q: torch.Tensor,
k: torch.Tensor,
v: torch.Tensor,
g: torch.Tensor,
beta: torch.Tensor,
Aqk: torch.Tensor,
Akk: torch.Tensor,
scale: float,
initial_state: torch.Tensor,
do: torch.Tensor,
dht: torch.Tensor,
cu_seqlens: torch.LongTensor | None = None,
):
chunk_size = 64
w, u, qg, kg = recompute_w_u_fwd(
q=q,
k=k,
v=v,
beta=beta,
A=Akk,
gk=g,
cu_seqlens=cu_seqlens,
)
h, v_new, _ = chunk_gated_delta_rule_fwd_h(
k=kg,
w=w,
u=u,
gk=g,
initial_state=initial_state,
output_final_state=False,
cu_seqlens=cu_seqlens,
)
dv = chunk_bwd_dv_local(
q=q,
k=k,
do=do,
A=Aqk,
scale=scale,
cu_seqlens=cu_seqlens,
chunk_size=chunk_size,
)
dh, dh0, dv = chunk_gated_delta_rule_bwd_dhu(
q=qg,
k=kg,
w=w,
gk=g,
h0=initial_state,
dht=dht,
do=do,
dv=dv,
scale=scale,
cu_seqlens=cu_seqlens,
)
# dq dk in fp32
dAqk = chunk_gla_bwd_dA(
v=v_new,
do=do,
scale=scale,
cu_seqlens=cu_seqlens,
chunk_size=chunk_size,
)
dq, dk, dw, dg = chunk_kda_bwd_dqkwg(
q=q,
k=k,
v=v_new,
w=w,
g=g,
h=h,
dv=dv,
do=do,
dh=dh,
scale=scale,
cu_seqlens=cu_seqlens,
chunk_size=chunk_size,
)
dk2, dv, db, dg2, dAkk = prepare_wy_repr_bwd(
k=k,
v=v,
beta=beta,
gk=g,
A=Akk,
dw=dw,
du=dv,
cu_seqlens=cu_seqlens,
)
dq, dk2, db, dg2 = chunk_kda_bwd_intra(
q=q,
k=k,
g=g,
beta=beta,
dAqk=dAqk,
dAkk=dAkk,
dq=dq,
dk=dk2,
db=db,
dg=dg2,
cu_seqlens=cu_seqlens,
chunk_size=chunk_size,
)
dk.add_(dk2)
dg.add_(dg2)
return dq, dk, dv, db, dg, dh0
class ChunkKDAFunction(torch.autograd.Function):
@staticmethod
@input_guard
@autocast_custom_fwd
def forward(
ctx,
q: torch.Tensor,
k: torch.Tensor,
v: torch.Tensor,
g: torch.Tensor,
beta: torch.Tensor,
scale: float,
initial_state: torch.Tensor,
output_final_state: bool = False,
use_qk_l2norm_in_kernel: bool = False,
cu_seqlens: torch.LongTensor | None = None,
):
q_rstd, k_rstd = None, None
if use_qk_l2norm_in_kernel:
q, q_rstd = l2norm_fwd(q)
k, k_rstd = l2norm_fwd(k)
g, o, Aqk, Akk, final_state = chunk_kda_fwd(
q=q,
k=k,
v=v,
g=g,
beta=beta,
scale=scale,
initial_state=initial_state,
output_final_state=output_final_state,
cu_seqlens=cu_seqlens,
)
ctx.save_for_backward(q, q_rstd, k, k_rstd, v, g, beta, Aqk, Akk, initial_state, cu_seqlens)
ctx.scale = scale
ctx.use_qk_l2norm_in_kernel = use_qk_l2norm_in_kernel
return o.to(q.dtype), final_state
@staticmethod
@input_guard
@autocast_custom_bwd
def backward(
ctx,
do: torch.Tensor,
dht: torch.Tensor,
):
q, q_rstd, k, k_rstd, v, g, beta, Aqk, Akk, initial_state, cu_seqlens = ctx.saved_tensors
dq, dk, dv, db, dg, dh0 = chunk_kda_bwd(
q=q,
k=k,
v=v,
g=g,
beta=beta,
Aqk=Aqk,
Akk=Akk,
scale=ctx.scale,
initial_state=initial_state,
do=do,
dht=dht,
cu_seqlens=cu_seqlens,
)
if ctx.use_qk_l2norm_in_kernel:
dq = l2norm_bwd(q, q_rstd, dq)
dk = l2norm_bwd(k, k_rstd, dk)
return dq.to(q), dk.to(k), dv.to(v), dg.to(g), db.to(beta), None, dh0, None, None, None
@torch.compiler.disable
def chunk_kda(
q: torch.Tensor,
k: torch.Tensor,
v: torch.Tensor,
g: torch.Tensor,
beta: torch.Tensor,
scale: float = None,
initial_state: torch.Tensor = None,
output_final_state: bool = False,
use_qk_l2norm_in_kernel: bool = False,
cu_seqlens: torch.LongTensor | None = None,
**kwargs,
):
r"""
Args:
q (torch.Tensor):
queries of shape `[B, T, H, K]`.
k (torch.Tensor):
keys of shape `[B, T, H, K]`.
v (torch.Tensor):
values of shape `[B, T, H, V]`.
g (torch.Tensor):
(forget) gating tensor (in log space!) of shape `[B, T, H, K]`.
beta (torch.Tensor):
betas of shape `[B, T, H]`.
scale (Optional[float]):
Scale factor for the RetNet attention scores.
If not provided, it will default to `1 / sqrt(K)`. Default: `None`.
initial_state (Optional[torch.Tensor]):
Initial state of shape `[N, H, K, V]` for `N` input sequences.
For equal-length input sequences, `N` equals the batch size `B`.
Default: `None`.
output_final_state (Optional[bool]):
Whether to output the final state of shape `[N, H, K, V]`. Default: `False`.
use_qk_l2norm_in_kernel (bool):
Whether to apply L2norm to the q,k tensor internally. Default: `False`.
cu_seqlens (torch.LongTensor):
Cumulative sequence lengths of shape `[N+1]` used for variable-length training,
consistent with the FlashAttention API.
Returns:
o (torch.Tensor):
Outputs of shape `[B, T, H, V]`.
final_state (torch.Tensor):
Final state of shape `[N, H, K, V]` if `output_final_state=True` else `None`.
Examples::
>>> import torch
>>> import torch.nn.functional as F
>>> from einops import rearrange
>>> from fla.ops.kda import chunk_kda
# inputs with equal lengths
>>> B, T, H, K, V = 4, 2048, 4, 512, 512
>>> q = torch.randn(B, T, H, K, dtype=torch.bfloat16, device='cuda')
>>> k = torch.randn(B, T, H, K, dtype=torch.bfloat16, device='cuda')
>>> v = torch.randn(B, T, H, V, dtype=torch.bfloat16, device='cuda')
>>> beta = torch.rand(B, T, H, dtype=torch.bfloat16, device='cuda').sigmoid()
>>> g = F.logsigmoid(torch.rand(B, T, H, dtype=torch.bfloat16, device='cuda'))
>>> h0 = torch.randn(B, H, K, V, dtype=torch.bfloat16, device='cuda')
>>> o, ht = chunk_kda(
q, k, v, g, beta,
use_qk_l2norm_in_kernel=True,
initial_state=h0,
output_final_state=True
)
# for variable-length inputs, the batch size `B` is expected to be 1 and `cu_seqlens` is required
>>> q, k, v, beta, g = map(lambda x: rearrange(x, 'b t ... -> 1 (b t) ...'), (q, k, v, beta, g))
# for a batch with 4 sequences, `cu_seqlens` with 5 start/end positions are expected
>>> cu_seqlens = q.new_tensor([0, 2048, 4096, 6144, 8192], dtype=torch.long)
>>> o, ht = chunk_kda(
q, k, v, g, beta,
use_qk_l2norm_in_kernel=True,
initial_state=h0,
output_final_state=True,
cu_seqlens=cu_seqlens
)
"""
if cu_seqlens is not None:
if q.shape[0] != 1:
raise ValueError(
f"The batch size is expected to be 1 rather than {q.shape[0]} when using `cu_seqlens`."
f"Please flatten variable-length inputs before processing.",
)
if initial_state is not None and initial_state.shape[0] != len(cu_seqlens) - 1:
raise ValueError(
f"The number of initial states is expected to be equal to the number of input sequences, "
f"i.e., {len(cu_seqlens) - 1} rather than {initial_state.shape[0]}.",
)
if scale is None:
scale = k.shape[-1] ** -0.5
o, final_state = ChunkKDAFunction.apply(
q,
k,
v,
g,
beta,
scale,
initial_state,
output_final_state,
use_qk_l2norm_in_kernel,
cu_seqlens,
)
return o, final_state
|