File size: 24,009 Bytes
37351b6 | 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 99 100 101 102 103 104 105 106 107 108 109 110 111 112 113 114 115 116 117 118 119 120 121 122 123 124 125 126 127 128 129 130 131 132 133 134 135 136 137 138 139 140 141 142 143 144 145 146 147 148 149 150 151 152 153 154 155 156 157 158 159 160 161 162 163 164 165 166 167 168 169 170 171 172 173 174 175 176 177 178 179 180 181 182 183 184 185 186 187 188 189 190 191 192 193 194 195 196 197 198 199 200 201 202 203 204 205 206 207 208 209 210 211 212 213 214 215 216 217 218 219 220 221 222 223 224 225 226 227 228 229 230 231 232 233 234 235 236 237 238 239 240 241 242 243 244 245 246 247 248 249 250 251 252 253 254 255 256 257 258 259 260 261 262 263 264 265 266 267 268 269 270 271 272 273 274 275 276 277 278 279 280 281 282 283 284 285 286 287 288 289 290 291 292 293 294 295 296 297 298 299 300 301 302 303 304 305 306 307 308 309 310 311 312 313 314 315 316 317 318 319 320 321 322 323 324 325 326 327 328 329 330 331 332 333 334 335 336 337 338 339 340 341 342 343 344 345 346 347 348 349 350 351 352 353 354 355 356 357 358 359 360 361 362 363 364 365 366 367 368 369 370 371 372 373 374 375 376 377 378 379 380 381 382 383 384 385 386 387 388 389 390 391 392 393 394 395 396 397 398 399 400 401 402 403 404 405 406 407 408 409 410 411 412 413 414 415 416 417 418 419 420 421 422 423 424 425 426 427 428 429 430 431 432 433 434 435 436 437 438 439 440 441 442 443 444 445 446 447 448 449 450 451 452 453 454 455 456 457 458 459 460 461 462 463 464 465 466 467 468 469 470 471 472 473 474 475 476 477 478 479 480 481 482 483 484 485 486 487 488 489 490 491 492 493 494 495 496 497 498 499 500 501 502 503 504 505 506 507 508 509 510 511 512 513 514 515 516 517 518 519 520 521 522 523 524 525 526 527 528 529 530 531 532 533 534 535 536 537 538 539 540 541 542 543 544 545 546 547 548 549 550 551 552 553 554 555 556 557 558 559 560 561 562 563 564 565 566 567 568 569 570 571 572 573 574 575 576 577 578 579 580 581 582 583 584 585 586 587 588 589 590 591 592 593 594 595 596 597 598 599 600 601 602 603 604 605 606 607 608 609 610 611 612 613 614 615 616 617 618 619 620 621 622 | from __future__ import annotations
import importlib
import math
from importlib import metadata as importlib_metadata
from typing import Any, Dict, Iterable, List, Optional, Sequence, Tuple, Union
import torch
import torch.nn.functional as F
from env_utils import configure_flashinfer_runtime, format_flashinfer_exception
FA2_PAPER_TARGET_VERSION = "2.7.4.post1"
FLASHINFER_PAPER_TARGET_VERSION = "0.6.6"
class BackendError(RuntimeError):
def __init__(self, message: str, *, details: Optional[str] = None) -> None:
super().__init__(message)
self.details = details
class DecodeAttentionBackend:
name: str = "backend"
paper_role: str = "auxiliary"
exact_attention: Optional[bool] = None
supports_true_batch_decode: Optional[bool] = None
def decode(self, q_bhd: torch.Tensor, k_nhd: torch.Tensor, v_nhd: torch.Tensor, valid_len: int) -> torch.Tensor:
raise NotImplementedError
def info(self) -> Dict[str, Any]:
return {
"backend": self.name,
"paper_role": self.paper_role,
"exact_attention": self.exact_attention,
"supports_true_batch_decode": self.supports_true_batch_decode,
}
def smoke_test(
self,
*,
device: torch.device,
dtype: torch.dtype,
num_heads: int,
num_kv_heads: int,
head_dim: int,
valid_len: int,
) -> Dict[str, Any]:
batch_size = 2
q = torch.randn(batch_size, num_heads, head_dim, device=device, dtype=dtype)
k = torch.randn(batch_size, valid_len, num_kv_heads, head_dim, device=device, dtype=dtype)
v = torch.randn(batch_size, valid_len, num_kv_heads, head_dim, device=device, dtype=dtype)
out = self.decode(q, k, v, valid_len)
expected_shape = (batch_size, num_heads, head_dim)
if out.shape != expected_shape:
raise BackendError(
f"Smoke test returned wrong shape for {self.name}: got {tuple(out.shape)}, expected {expected_shape}"
)
return {
"backend": self.name,
"batch_size": batch_size,
"valid_len": int(valid_len),
"output_shape": list(out.shape),
**self.info(),
}
class SDPABackend(DecodeAttentionBackend):
name = "sdpa"
paper_role = "debug_exact_reference"
exact_attention = True
supports_true_batch_decode = True
def decode(self, q_bhd: torch.Tensor, k_nhd: torch.Tensor, v_nhd: torch.Tensor, valid_len: int) -> torch.Tensor:
q = q_bhd.unsqueeze(2) # [B,H,1,D]
k = k_nhd[:, :valid_len].permute(0, 2, 1, 3)
v = v_nhd[:, :valid_len].permute(0, 2, 1, 3)
out = F.scaled_dot_product_attention(
q,
k,
v,
attn_mask=None,
dropout_p=0.0,
is_causal=False,
enable_gqa=(q.shape[1] != k.shape[1]),
)
return out[:, :, 0, :].contiguous()
class FA2Backend(DecodeAttentionBackend):
name = "fa2"
paper_role = "main_exact_dense_batched_baseline"
exact_attention = True
supports_true_batch_decode = True
def __init__(
self,
*,
expected_version: str = FA2_PAPER_TARGET_VERSION,
version_policy: str = "warn",
) -> None:
self.expected_version = str(expected_version)
self.version_policy = str(version_policy).lower()
self.actual_mode = "kvcache_batched"
self.module_name: Optional[str] = None
self.module: Optional[Any] = None
self.decode_fn = None
self.version_warning: Optional[str] = None
self._cache_seqlens_buffers: Dict[Tuple[str, int], torch.Tensor] = {}
import_errors: List[str] = []
for candidate in ("flash_attn", "flash_attn.flash_attn_interface"):
try:
module = importlib.import_module(candidate)
except Exception as exc:
import_errors.append(f"{candidate}: {type(exc).__name__}: {exc}")
continue
fn = getattr(module, "flash_attn_with_kvcache", None)
if fn is None:
import_errors.append(f"{candidate}: missing flash_attn_with_kvcache")
continue
self.module_name = candidate
self.module = module
self.decode_fn = fn
break
if self.decode_fn is None:
raise BackendError(
"Failed to import FlashAttention-2 decode KV-cache API.",
details=(
"Expected flash_attn_with_kvcache in flash_attn or flash_attn.flash_attn_interface. "
f"Tried: {' | '.join(import_errors)}"
),
)
self.version = self._resolve_version(self.module)
if self.expected_version and self.version != self.expected_version:
self.version_warning = (
f"flash-attn version mismatch: expected {self.expected_version}, got {self.version}. "
f"This can change API or performance behavior."
)
if self.version_policy == "error":
raise BackendError(self.version_warning)
if self.version_policy == "warn":
print(f"[fa2 warning] {self.version_warning}")
@staticmethod
def _resolve_version(module: Any) -> str:
for dist_name in ("flash-attn", "flash_attn"):
try:
return str(importlib_metadata.version(dist_name))
except Exception:
pass
version = getattr(module, "__version__", None)
return str(version) if version is not None else "unknown"
@staticmethod
def _unwrap_tensor_out(x: Any) -> torch.Tensor:
if isinstance(x, (tuple, list)):
return x[0]
return x
def _cache_seqlens(self, batch_size: int, device: torch.device, valid_len: int) -> torch.Tensor:
key = (str(device), int(batch_size))
buf = self._cache_seqlens_buffers.get(key)
if buf is None or buf.numel() != batch_size or buf.device != device:
buf = torch.empty((batch_size,), device=device, dtype=torch.int32)
self._cache_seqlens_buffers[key] = buf
buf.fill_(int(valid_len))
return buf
def _call_decode(
self,
q_b1hd: torch.Tensor,
k_nhd: torch.Tensor,
v_nhd: torch.Tensor,
cache_seqlens: torch.Tensor,
) -> torch.Tensor:
sm_scale = 1.0 / math.sqrt(float(q_b1hd.shape[-1]))
attempts = [
lambda: self.decode_fn(
q=q_b1hd,
k_cache=k_nhd,
v_cache=v_nhd,
cache_seqlens=cache_seqlens,
softmax_scale=sm_scale,
causal=True,
),
lambda: self.decode_fn(
q_b1hd,
k_nhd,
v_nhd,
cache_seqlens=cache_seqlens,
softmax_scale=sm_scale,
causal=True,
),
lambda: self.decode_fn(
q_b1hd,
k_nhd,
v_nhd,
cache_seqlens=cache_seqlens,
causal=True,
),
]
last_type_error: Optional[TypeError] = None
for attempt in attempts:
try:
return attempt()
except TypeError as exc:
last_type_error = exc
continue
except Exception as exc:
raise BackendError(
f"FlashAttention-2 decode call failed: {type(exc).__name__}: {exc}",
details=(
f"module={self.module_name} version={self.version} "
f"q_shape={tuple(q_b1hd.shape)} k_cache_shape={tuple(k_nhd.shape)} "
f"v_cache_shape={tuple(v_nhd.shape)} valid_len={int(cache_seqlens[0].item())}"
),
) from exc
raise BackendError(
f"FlashAttention-2 decode call failed: {type(last_type_error).__name__}: {last_type_error}",
details=(
f"module={self.module_name} version={self.version} "
f"q_shape={tuple(q_b1hd.shape)} k_cache_shape={tuple(k_nhd.shape)} "
f"v_cache_shape={tuple(v_nhd.shape)} valid_len={int(cache_seqlens[0].item())}"
),
)
def decode(self, q_bhd: torch.Tensor, k_nhd: torch.Tensor, v_nhd: torch.Tensor, valid_len: int) -> torch.Tensor:
if q_bhd.device.type != "cuda":
raise BackendError("FA2 backend requires CUDA tensors.")
if q_bhd.dim() != 3:
raise BackendError(f"Expected q_bhd rank 3, got shape {tuple(q_bhd.shape)}")
batch_size = int(q_bhd.shape[0])
q_b1hd = q_bhd.unsqueeze(1).contiguous() # [B,1,H,D]
cache_seqlens = self._cache_seqlens(batch_size, q_bhd.device, valid_len)
out = self._unwrap_tensor_out(self._call_decode(q_b1hd, k_nhd, v_nhd, cache_seqlens))
if out.dim() != 4 or out.shape[1] != 1:
raise BackendError(
f"Unexpected FA2 decode output shape: got {tuple(out.shape)}, expected [B,1,H,D]"
)
return out[:, 0, :, :].contiguous()
def info(self) -> Dict[str, Any]:
return {
**super().info(),
"module": self.module_name,
"version": self.version,
"expected_version": self.expected_version,
"version_warning": self.version_warning,
"decode_kernel": "flash_attn_with_kvcache",
"actual_mode": self.actual_mode,
"cache_layout": "NHD_contiguous",
"cache_seqlens_mode": "uniform_int32_tensor",
}
class FlashInferBackend(DecodeAttentionBackend):
name = "flashinfer"
paper_role = "legacy_reference_not_main_batched_baseline"
exact_attention = True
def __init__(
self,
*,
mode: str = "single_loop",
use_tensor_cores: bool = True,
expected_version: str = FLASHINFER_PAPER_TARGET_VERSION,
version_policy: str = "warn",
jit_mode: str = "auto",
preload_libstdcpp: str = "auto",
) -> None:
self.mode = str(mode)
self.use_tensor_cores = bool(use_tensor_cores)
self.actual_mode = self.mode
self.expected_version = str(expected_version)
self.version_policy = str(version_policy).lower()
self.runtime = configure_flashinfer_runtime(jit_mode=jit_mode, preload_libstdcpp=preload_libstdcpp)
self.supports_true_batch_decode = self.mode == "batch_compact"
try:
self.flashinfer = importlib.import_module("flashinfer")
except Exception as exc:
raise BackendError(
"Failed to import flashinfer.",
details=format_flashinfer_exception(exc, self.runtime),
) from exc
self.version = getattr(self.flashinfer, "__version__", "unknown")
self.version_warning: Optional[str] = None
if self.expected_version and self.version != self.expected_version:
self.version_warning = (
f"flashinfer version mismatch: expected {self.expected_version}, got {self.version}. "
f"This can change API and performance behavior."
)
if self.version_policy == "error":
raise BackendError(self.version_warning)
if self.version_policy == "warn":
print(f"[flashinfer warning] {self.version_warning}")
self.single_fn = getattr(self.flashinfer, "single_decode_with_kv_cache", None)
self.batch_fn = getattr(self.flashinfer, "batch_decode_with_padded_kv_cache", None)
decode_mod = getattr(self.flashinfer, "decode", None)
if self.single_fn is None and decode_mod is not None:
self.single_fn = getattr(decode_mod, "single_decode_with_kv_cache", None)
if self.batch_fn is None and decode_mod is not None:
self.batch_fn = getattr(decode_mod, "batch_decode_with_padded_kv_cache", None)
if self.single_fn is None:
raise BackendError(
"Could not find flashinfer single decode API.",
details=(
"Expected flashinfer.single_decode_with_kv_cache or "
"flashinfer.decode.single_decode_with_kv_cache to exist."
),
)
def _call_single(self, q_h_d: torch.Tensor, k: torch.Tensor, v: torch.Tensor) -> torch.Tensor:
sm_scale = 1.0 / math.sqrt(float(q_h_d.shape[-1]))
attempts = [
lambda: self.single_fn(
q_h_d,
k,
v,
kv_layout="NHD",
use_tensor_cores=self.use_tensor_cores,
sm_scale=sm_scale,
),
lambda: self.single_fn(q_h_d, k, v, kv_layout="NHD", sm_scale=sm_scale),
lambda: self.single_fn(q_h_d, k, v),
]
last_type_error: Optional[TypeError] = None
for attempt in attempts:
try:
return attempt()
except TypeError as exc:
last_type_error = exc
continue
except Exception as exc:
raise BackendError(
f"FlashInfer single decode call failed: {type(exc).__name__}: {exc}",
details=format_flashinfer_exception(exc, self.runtime),
) from exc
raise BackendError(
f"FlashInfer single decode call failed: {type(last_type_error).__name__}: {last_type_error}",
details=format_flashinfer_exception(last_type_error, self.runtime) if last_type_error is not None else None,
)
def _call_batch_compact(self, q_bhd: torch.Tensor, k_nhd: torch.Tensor, v_nhd: torch.Tensor, valid_len: int) -> torch.Tensor:
if self.batch_fn is None:
raise BackendError(
"FlashInfer batch padded decode function was requested, but the installed flashinfer package does not expose it."
)
sm_scale = 1.0 / math.sqrt(float(q_bhd.shape[-1]))
k_compact = k_nhd[:, :valid_len].contiguous()
v_compact = v_nhd[:, :valid_len].contiguous()
q_compact = q_bhd.contiguous()
attempts = [
lambda: self.batch_fn(
q_compact,
k_compact,
v_compact,
kv_layout="NHD",
use_tensor_cores=self.use_tensor_cores,
sm_scale=sm_scale,
),
lambda: self.batch_fn(q_compact, k_compact, v_compact, kv_layout="NHD", sm_scale=sm_scale),
lambda: self.batch_fn(q_compact, k_compact, v_compact),
]
last_type_error: Optional[TypeError] = None
for attempt in attempts:
try:
return attempt()
except TypeError as exc:
last_type_error = exc
continue
except Exception as exc:
raise BackendError(
f"FlashInfer batch decode call failed: {type(exc).__name__}: {exc}",
details=format_flashinfer_exception(exc, self.runtime),
) from exc
raise BackendError(
f"FlashInfer batch decode call failed: {type(last_type_error).__name__}: {last_type_error}",
details=format_flashinfer_exception(last_type_error, self.runtime) if last_type_error is not None else None,
)
def decode(self, q_bhd: torch.Tensor, k_nhd: torch.Tensor, v_nhd: torch.Tensor, valid_len: int) -> torch.Tensor:
if self.mode == "batch_compact":
self.actual_mode = "batch_compact"
return self._call_batch_compact(q_bhd, k_nhd, v_nhd, valid_len).contiguous()
self.actual_mode = "single_loop"
outputs = []
batch_size = int(q_bhd.shape[0])
for b in range(batch_size):
k_b = k_nhd[b, :valid_len]
v_b = v_nhd[b, :valid_len]
out_b = self._call_single(q_bhd[b].contiguous(), k_b, v_b)
outputs.append(out_b)
return torch.stack(outputs, dim=0).contiguous()
def info(self) -> Dict[str, Any]:
return {
**super().info(),
"version": self.version,
"expected_version": self.expected_version,
"version_warning": self.version_warning,
"requested_mode": self.mode,
"actual_mode": self.actual_mode,
"has_batch_padded_api": self.batch_fn is not None,
"use_tensor_cores": self.use_tensor_cores,
}
class _SantaBaseBackend(DecodeAttentionBackend):
module_candidates: Sequence[str] = ()
name = "santa"
paper_role = "main_batched_sparse_baseline"
exact_attention = False
def __init__(self, *, S: int, seed: int, block_n: Optional[int]) -> None:
self.S = int(S)
self.seed = int(seed)
self.block_n = None if block_n is None else int(block_n)
self.actual_mode = "unknown"
self.module = None
self.module_name = None
self.supports_true_batch_decode = False
import_errors: List[str] = []
for candidate in self.module_candidates:
try:
self.module = importlib.import_module(candidate)
self.module_name = candidate
break
except Exception as exc:
import_errors.append(f"{candidate}: {exc}")
if self.module is None:
raise BackendError(
f"Could not import any {self.name} extension module. Tried: {', '.join(self.module_candidates)}.",
details=" | ".join(import_errors),
)
self.supports_true_batch_decode = hasattr(self.module, "decode_systematic_batched")
@staticmethod
def _unwrap_tensor_out(x: Any) -> torch.Tensor:
if isinstance(x, (tuple, list)):
return x[0]
return x
def _get_decode_fn(self):
if hasattr(self.module, "decode_systematic_batched"):
return getattr(self.module, "decode_systematic_batched")
if hasattr(self.module, "decode_systematic_scalar"):
return getattr(self.module, "decode_systematic_scalar")
raise BackendError(f"Module {self.module_name} does not expose a decode entry point.")
def decode(self, q_bhd: torch.Tensor, k_nhd: torch.Tensor, v_nhd: torch.Tensor, valid_len: int) -> torch.Tensor:
fn = self._get_decode_fn()
kwargs = {
"seed": self.seed,
"want_vrows": False,
"valid_len": int(valid_len),
}
if self.block_n is not None:
kwargs["block_n"] = int(self.block_n)
try:
out = fn(q_bhd.contiguous(), k_nhd, v_nhd, self.S, **kwargs)
self.actual_mode = "batch"
return self._unwrap_tensor_out(out).contiguous()
except Exception as exc:
msg = str(exc)
should_fallback = (
q_bhd.dim() == 3
and (
"Expected q_h [H,D]" in msg
or "Expected q_h [H, D]" in msg
or "valid_len" in msg
or "rank" in msg
or "shape" in msg
)
)
if not should_fallback:
raise
kwargs.pop("valid_len", None)
self.actual_mode = "single_loop_fallback"
outputs = []
batch_size = int(q_bhd.shape[0])
for b in range(batch_size):
k_b = k_nhd[b, :valid_len].contiguous()
v_b = v_nhd[b, :valid_len].contiguous()
out_b = fn(q_bhd[b].contiguous(), k_b, v_b, self.S, **kwargs)
outputs.append(self._unwrap_tensor_out(out_b))
return torch.stack(outputs, dim=0).contiguous()
def info(self) -> Dict[str, Any]:
return {
**super().info(),
"module": self.module_name,
"requested_S": self.S,
"seed": self.seed,
"block_n": self.block_n,
"actual_mode": self.actual_mode,
}
class SantaFlashBackend(_SantaBaseBackend):
name = "santa_flash"
module_candidates = ("santa_flash_batch_cuda", "santa_cuda")
class SantaPropBackend(_SantaBaseBackend):
name = "santa_prop"
module_candidates = ("santa_prop_batch_cuda",)
def _make_one_backend(
raw_name: str,
*,
fa2_expected_version: str,
fa2_version_policy: str,
santa_s: int,
santa_seed: int,
santa_block_n: Optional[int],
flashinfer_mode: str,
flashinfer_use_tensor_cores: bool,
flashinfer_expected_version: str,
flashinfer_version_policy: str,
flashinfer_jit_mode: str,
flashinfer_preload_libstdcpp: str,
) -> DecodeAttentionBackend:
name = raw_name.lower().strip()
if name == "sdpa":
return SDPABackend()
if name in ("fa2", "flashattn2", "flash_attention_2", "flash_attn"):
return FA2Backend(
expected_version=fa2_expected_version,
version_policy=fa2_version_policy,
)
if name == "flashinfer":
return FlashInferBackend(
mode=flashinfer_mode,
use_tensor_cores=flashinfer_use_tensor_cores,
expected_version=flashinfer_expected_version,
version_policy=flashinfer_version_policy,
jit_mode=flashinfer_jit_mode,
preload_libstdcpp=flashinfer_preload_libstdcpp,
)
if name in ("santa", "santa_flash"):
return SantaFlashBackend(
S=santa_s,
seed=santa_seed,
block_n=santa_block_n,
)
if name == "santa_prop":
return SantaPropBackend(
S=santa_s,
seed=santa_seed,
block_n=santa_block_n,
)
raise ValueError(f"Unsupported backend name: {raw_name}")
def build_backends(
backend_names: Iterable[str],
*,
fa2_expected_version: str = FA2_PAPER_TARGET_VERSION,
fa2_version_policy: str = "warn",
santa_s: int,
santa_seed: int,
santa_block_n: Optional[int],
flashinfer_mode: str,
flashinfer_use_tensor_cores: bool,
flashinfer_expected_version: str = FLASHINFER_PAPER_TARGET_VERSION,
flashinfer_version_policy: str = "warn",
flashinfer_jit_mode: str = "auto",
flashinfer_preload_libstdcpp: str = "auto",
skip_init_failures: Optional[bool] = None,
) -> Union[List[DecodeAttentionBackend], Tuple[List[DecodeAttentionBackend], List[Dict[str, Any]]]]:
backends: List[DecodeAttentionBackend] = []
init_errors: List[Dict[str, Any]] = []
for raw_name in backend_names:
try:
backend = _make_one_backend(
raw_name,
fa2_expected_version=fa2_expected_version,
fa2_version_policy=fa2_version_policy,
santa_s=santa_s,
santa_seed=santa_seed,
santa_block_n=santa_block_n,
flashinfer_mode=flashinfer_mode,
flashinfer_use_tensor_cores=flashinfer_use_tensor_cores,
flashinfer_expected_version=flashinfer_expected_version,
flashinfer_version_policy=flashinfer_version_policy,
flashinfer_jit_mode=flashinfer_jit_mode,
flashinfer_preload_libstdcpp=flashinfer_preload_libstdcpp,
)
backends.append(backend)
except Exception as exc:
if skip_init_failures:
init_errors.append(
{
"backend": str(raw_name),
"error_type": type(exc).__name__,
"error": str(exc),
"details": getattr(exc, "details", None),
}
)
continue
raise
if skip_init_failures is None:
return backends
return backends, init_errors
|