File size: 13,492 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 99 100 101 102 103 104 105 106 107 108 109 110 111 112 113 114 115 116 117 118 119 120 121 122 123 124 125 126 127 128 129 130 131 132 133 134 135 136 137 138 139 140 141 142 143 144 145 146 147 148 149 150 151 152 153 154 155 156 157 158 159 160 161 162 163 164 165 166 167 168 169 170 171 172 173 174 175 176 177 178 179 180 181 182 183 184 185 186 187 188 189 190 191 192 193 194 195 196 197 198 199 200 201 202 203 204 205 206 207 208 209 210 211 212 213 214 215 216 217 218 219 220 221 222 223 224 225 226 227 228 229 230 231 232 233 234 235 236 237 238 239 240 241 242 243 244 245 246 247 248 249 250 251 252 253 254 255 256 257 258 259 260 261 262 263 264 265 266 267 268 269 270 271 272 273 274 275 276 277 278 279 280 281 282 283 284 285 286 287 288 289 290 291 292 293 294 295 296 297 298 299 300 301 302 303 304 305 306 307 308 309 310 311 312 313 314 315 316 317 318 319 320 321 322 323 324 325 326 327 328 329 330 331 332 333 334 335 336 337 338 339 340 341 342 343 344 345 346 347 348 349 350 351 352 353 354 355 356 357 358 359 360 361 362 363 364 365 366 367 368 369 370 | # Copyright (c) 2023-2025, Songlin Yang, Yu Zhang
import torch
from fla.modules.l2norm import l2norm_bwd, l2norm_fwd
from fla.ops.common.chunk_h import chunk_bwd_dh
from fla.ops.mesa_net.chunk_cg_solver_bwd import chunk_mesa_cg_bwd
from fla.ops.mesa_net.chunk_cg_solver_fwd import chunk_mesa_cg_fwd
from fla.ops.mesa_net.chunk_h_fwd import chunk_mesa_fwd_h
from fla.ops.mesa_net.chunk_h_kk_intra_bwd import chunk_mesa_net_h_kk_bwd_intra_fn
from fla.ops.mesa_net.chunk_h_kv_intra_bwd import chunk_mesa_net_h_kv_bwd_intra_fn
from fla.ops.utils import chunk_local_cumsum
from fla.utils import autocast_custom_bwd, autocast_custom_fwd, input_guard
def chunk_fwd_mesa_net_fwd(
q: torch.Tensor,
k: torch.Tensor,
v: torch.Tensor,
g: torch.Tensor,
beta: torch.Tensor,
lamb: torch.Tensor,
cu_seqlens: torch.Tensor,
max_CG_iteration: int = 30,
chunk_size: int = 64,
h_kk_init: torch.Tensor | None = None,
h_kv_init: torch.Tensor | None = None,
output_final_state: bool = False,
) -> torch.Tensor:
g = chunk_local_cumsum(g, chunk_size=chunk_size, cu_seqlens=cu_seqlens) if g is not None else None
h_kk, h_kv, h_kk_final, h_kv_final = chunk_mesa_fwd_h(
k=k,
v=v,
g=g,
beta=beta,
h_init=h_kk_init,
h_kv_init=h_kv_init,
output_final_state=output_final_state,
states_in_fp32=False,
cu_seqlens=cu_seqlens,
chunk_size=chunk_size,
)
q_star, o = chunk_mesa_cg_fwd(
q=q,
k=k,
h=h_kk,
h_kv=h_kv,
v=v,
g_local_cumsum=g,
beta=beta,
lamb=lamb,
cu_seqlens=cu_seqlens,
chunk_size=chunk_size,
max_CG_iteration=max_CG_iteration,
)
return g, q_star, o, (h_kk_final, h_kv_final)
def chunk_fwd_mesa_net_bwd(
q: torch.Tensor,
k: torch.Tensor,
v: torch.Tensor,
g: torch.Tensor,
beta: torch.Tensor,
lamb: torch.Tensor,
q_star: torch.Tensor, # should be cached in the forward pass
do: torch.Tensor,
cu_seqlens: torch.Tensor,
max_CG_iteration: int = 30,
chunk_size: int = 64,
h_kk_init: torch.Tensor | None = None,
h_kv_init: torch.Tensor | None = None,
dh_kv_final: torch.Tensor | None = None,
dh_kk_final: torch.Tensor | None = None,
) -> torch.Tensor:
# recompute the hidden states, which is quite cheap
h_kk, h_kv, _, _ = chunk_mesa_fwd_h(
k=k,
v=v,
g=g,
beta=beta,
h_init=h_kk_init,
h_kv_init=h_kv_init,
output_final_state=False,
states_in_fp32=False,
cu_seqlens=cu_seqlens,
chunk_size=chunk_size,
)
dh_kv, dh0_kv = chunk_bwd_dh(
q=q_star,
k=k,
v=v,
g=g,
gk=None,
gv=None,
do=do,
h0=h_kv_init,
dht=dh_kv_final,
states_in_fp32=False,
cu_seqlens=cu_seqlens,
chunk_size=chunk_size,
scale=1,
)
dq, dk_beta, dv, dg = chunk_mesa_net_h_kv_bwd_intra_fn(
q_star=q_star,
k=k,
v=v,
beta=beta,
h_kv=h_kv,
dh_kv=dh_kv,
g=g,
do=do,
cu_seqlens=cu_seqlens,
chunk_size=chunk_size,
)
dq = chunk_mesa_cg_bwd(
dq=dq,
k=k,
h=h_kk,
g_local_cumsum=g,
beta=beta,
lamb=lamb,
cu_seqlens=cu_seqlens,
chunk_size=chunk_size,
max_CG_iteration=max_CG_iteration,
output_dtype=torch.float16,
)
dh_kk, dh0_kk = chunk_bwd_dh(
q=dq,
k=k,
v=k,
g=g,
gk=None,
gv=None,
do=q_star,
h0=h_kk_init,
dht=-dh_kk_final if dh_kk_final is not None else None,
states_in_fp32=False,
cu_seqlens=cu_seqlens,
chunk_size=chunk_size,
scale=1,
)
dk, dg2, dlamb, dbeta = chunk_mesa_net_h_kk_bwd_intra_fn(
k=k,
g=g,
beta=beta,
h=h_kk,
dh=dh_kk,
dk_beta=dk_beta,
q_star=q_star,
dq=dq,
cu_seqlens=cu_seqlens,
chunk_size=chunk_size,
)
dg.add_(dg2)
dg = chunk_local_cumsum(dg, chunk_size=chunk_size, reverse=True, cu_seqlens=cu_seqlens).to(g)
return dq, dk, dv, dg, dbeta, dlamb, -dh0_kk if dh0_kk is not None else None, dh0_kv if dh0_kv is not None else None
class ChunkMesaNetFunction(torch.autograd.Function):
@staticmethod
@input_guard
@autocast_custom_fwd
def forward(
ctx,
q,
k,
v,
g,
beta,
lamb,
cu_seqlens,
max_CG_iteration,
h_kk_init,
h_kv_init,
output_final_state,
use_qk_l2norm_in_kernel,
):
chunk_size = 64
if use_qk_l2norm_in_kernel:
q, q_rstd = l2norm_fwd(q, output_dtype=torch.float16)
k, k_rstd = l2norm_fwd(k, output_dtype=torch.float16)
else:
q_rstd, k_rstd = None, None
q = q.to(torch.float16)
k = k.to(torch.float16)
g_cumsum, q_star, o, (h_kk_final, h_kv_final) = chunk_fwd_mesa_net_fwd(
q=q,
k=k,
v=v,
g=g,
beta=beta,
lamb=lamb,
cu_seqlens=cu_seqlens,
max_CG_iteration=max_CG_iteration,
chunk_size=chunk_size,
h_kk_init=h_kk_init,
h_kv_init=h_kv_init,
output_final_state=output_final_state,
)
ctx.max_CG_iteration = max_CG_iteration
ctx.chunk_size = chunk_size
ctx.cu_seqlens = cu_seqlens
ctx.use_qk_l2norm_in_kernel = use_qk_l2norm_in_kernel
ctx.save_for_backward(q, q_rstd, k, k_rstd, v, g_cumsum, beta, lamb, h_kk_init, h_kv_init, q_star, o)
return o, h_kk_final, h_kv_final
@staticmethod
@input_guard
@autocast_custom_bwd
def backward(ctx, do, dh_kk_final=None, dh_kv_final=None):
q, q_rstd, k, k_rstd, v, g, beta, lamb, h_kk_init, h_kv_init, q_star, o = ctx.saved_tensors
max_CG_iteration = ctx.max_CG_iteration
chunk_size = ctx.chunk_size
cu_seqlens = ctx.cu_seqlens
dq, dk, dv, dg, dbeta, dlamb, dh0_kk, dh0_kv = chunk_fwd_mesa_net_bwd(
q=q, k=k, v=v, g=g, beta=beta, lamb=lamb, q_star=q_star, do=do,
cu_seqlens=cu_seqlens, max_CG_iteration=max_CG_iteration, chunk_size=chunk_size,
h_kk_init=h_kk_init, h_kv_init=h_kv_init, dh_kv_final=dh_kv_final, dh_kk_final=dh_kk_final,
)
if ctx.use_qk_l2norm_in_kernel:
dq = l2norm_bwd(q, q_rstd, dq)
dk = l2norm_bwd(k, k_rstd, dk)
return dq, dk, dv.to(v), dg.to(g), dbeta.to(beta), dlamb.to(lamb), None, None, dh0_kk, dh0_kv, None, None
@torch.compiler.disable
def chunk_mesa_net(
q: torch.Tensor,
k: torch.Tensor,
v: torch.Tensor,
g: torch.Tensor,
beta: torch.Tensor,
lamb: torch.Tensor,
h_kk_init: torch.Tensor | None = None,
h_kv_init: torch.Tensor | None = None,
output_final_state: bool = False,
max_CG_iteration: int = 30,
use_qk_l2norm_in_kernel: bool = False,
cu_seqlens: torch.LongTensor | None = None,
):
r"""
Args:
q (torch.Tensor):
queries of shape `[B, T, H, K]`
k (torch.Tensor):
keys of shape `[B, T, H, K]`. Should be l2-normalized before passing in.
v (torch.Tensor):
values of shape `[B, T, H, V]`.
g (torch.Tensor):
decay factors of shape `[B, T, H]`. Note that `g` should be in log space, that is, `g = log(decay_factor) < 0`.
Recommended input dtype: `torch.float32`.
beta (torch.Tensor):
betas of shape `[B, T, H]`. Recommended input dtype: `torch.float32`.
lamb (torch.Tensor):
lambdas of shape `[B, T, H]`. Recommended input dtype: `torch.float32`.
h_kk_init (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`.
h_kv_init (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`.
max_CG_iteration (int):
Maximum number of conjugate gradient iterations for solving the linear system. Default: `30`.
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):
Do l2 normalization on Q and K in the kernel for saving GPU memory. 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_states_kk, final_states_kv) (Tuple[torch.Tensor, torch.Tensor]):
Final states of shape `[N, H, K, K]` and `[N, H, K, V]` if `output_final_state=True` else `(None, None)`.
Recall that MesaNet has two states, `h_kk` and `h_kv`!
Examples::
>>> import torch
>>> import torch.nn.functional as F
>>> from einops import rearrange
>>> from fla.ops.mesa_net import chunk_mesa_net
# inputs with equal lengths
>>> B, T, H, K, V = 4, 2048, 16, 128, 128
>>> 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')
>>> g = F.logsigmoid(torch.randn(B, T, H, dtype=torch.float32, device='cuda'))
>>> beta = torch.rand(B, T, H, dtype=torch.float32, device='cuda').sigmoid()
# lower bound is 0.25 for numerical stability
>>> lamb = F.softplus(torch.rand(H, K, dtype=torch.float32, device='cuda')) + 0.25
>>> init_state_kk = torch.randn(B, H, K, V, dtype=torch.float32, device='cuda')
>>> init_state_kv = torch.randn(B, H, K, V, dtype=torch.float32, device='cuda')
>>> o, (final_state_kk, final_state_kv) = chunk_mesa_net(
q, k, v, beta, lamb,
h_kk_init=init_state_kk,
h_kv_init=init_state_kv,
max_CG_iteration=30,
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 = map(lambda x: rearrange(x, 'b t ... -> 1 (b t) ...'), (q, k, v, beta))
# 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_var, (final_state_kk_var, final_state_kv_var) = chunk_mesa_net(
q, k, v, beta, lamb,
h_kk_init=init_state_kk,
h_kv_init=init_state_kv,
max_CG_iteration=30,
output_final_state=True,
cu_seqlens=cu_seqlens
)
"""
B, T, H, K = q.shape
assert k.shape == (B, T, H, K), "k must be of shape (batch size, seq len, num head, head dim)."
assert v.shape == (B, T, H, K), "v must be of shape (batch size, seq len, num head, head dim)."
assert g.shape == (B, T, H), "g must be of shape (batch size, seq len, num head)."
assert beta.shape == (B, T, H), "beta must be of shape (batch size, seq len, num head)."
assert lamb.shape == (H, K), "lamb must be of shape (num head, key dim)."
if h_kv_init is not None:
assert h_kv_init.dtype == torch.float32, "h_kv_init must be in float32."
if cu_seqlens is None:
assert h_kv_init.shape == (B, H, K, K), "h_kv_init must be of shape (batch size, num head, head dim, head dim)."
if h_kk_init is not None:
assert h_kk_init.dtype == torch.float32, "h_kk_init must be in float32."
if cu_seqlens is None:
assert h_kk_init.shape == (B, H, K, K), "h_kk_init must be of shape (batch size, num head, head dim, head dim)."
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 h_kk_init is not None and h_kk_init.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 {h_kk_init.shape[0]}.",
)
if h_kv_init is not None and h_kv_init.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 {h_kv_init.shape[0]}.",
)
o, final_state_kk, final_state_kv = ChunkMesaNetFunction.apply(
q,
k,
v,
g,
beta,
lamb,
cu_seqlens,
max_CG_iteration,
h_kk_init,
h_kv_init,
output_final_state,
use_qk_l2norm_in_kernel,
)
return o, final_state_kk, final_state_kv
|