| 1 | # SPDX-License-Identifier: Apache-2.0 |
| 2 | # SPDX-FileCopyrightText: Copyright contributors to the vLLM project |
| 3 | """Attention layer with FlashAttention.""" |
| 4 | |
| 5 | import copy |
| 6 | import functools |
| 7 | from dataclasses import dataclass |
| 8 | from typing import Any, ClassVar |
| 9 | |
| 10 | import numpy as np |
| 11 | import torch |
| 12 | |
| 13 | from vllm.model_executor.layers.attention import Attention |
| 14 | from vllm.model_executor.warmup.jit_warmup import VllmJitKernel, WarmupChoices |
| 15 | from vllm.platforms import current_platform |
| 16 | from vllm.utils.torch_utils import ( |
| 17 | PIN_MEMORY, |
| 18 | canonicalize_singleton_dim_strides, |
| 19 | is_quantized_kv_cache, |
| 20 | ) |
| 21 | from vllm.v1.attention.backend import ( |
| 22 | AttentionBackend, |
| 23 | AttentionImpl, |
| 24 | AttentionType, |
| 25 | MultipleOf, |
| 26 | ) |
| 27 | from vllm.v1.attention.backends.fa_utils import ( |
| 28 | FA4_HD256_PAGE_SIZE, |
| 29 | flash_attn_supports_kv_cache_dtype, |
| 30 | flash_attn_supports_quant_query_input, |
| 31 | get_flash_attn_version, |
| 32 | is_fa_version_supported, |
| 33 | is_flash_attn_varlen_func_available, |
| 34 | uses_fa4_hd256_kernel, |
| 35 | ) |
| 36 | from vllm.v1.attention.backends.utils import ( |
| 37 | fill_mm_prefix_query_ranges, |
| 38 | get_dcp_local_seq_lens, |
| 39 | get_num_attention_heads_from_layers, |
| 40 | ) |
| 41 | from vllm.v1.attention.ops.dcp import ( |
| 42 | cp_lse_ag_out_rs, |
| 43 | dcp_a2a_lse_reduce, |
| 44 | ) |
| 45 | from vllm.v1.attention.ops.merge_attn_states import merge_attn_states |
| 46 | from vllm.v1.worker.workspace import current_workspace_manager |
| 47 | |
| 48 | if is_flash_attn_varlen_func_available(): |
| 49 | from vllm.v1.attention.backends.fa_utils import ( |
| 50 | compile_flash_attn_varlen_func_from_specs, |
| 51 | flash_attn_supports_sinks, |
| 52 | flash_attn_varlen_func, |
| 53 | get_scheduler_metadata, |
| 54 | reshape_and_cache_flash, |
| 55 | ) |
| 56 | import vllm.envs as envs |
| 57 | from vllm.config import ( |
| 58 | VllmConfig, |
| 59 | get_current_vllm_config_or_none, |
| 60 | get_layers_from_vllm_config, |
| 61 | ) |
| 62 | from vllm.config.cache import CacheDType |
| 63 | from vllm.distributed.parallel_state import ( |
| 64 | get_dcp_group, |
| 65 | get_dcp_world_size_and_rank, |
| 66 | ) |
| 67 | from vllm.logger import init_logger |
| 68 | from vllm.platforms.interface import DeviceCapability |
| 69 | from vllm.utils.math_utils import cdiv, round_up |
| 70 | from vllm.v1.attention.backend import ( |
| 71 | AttentionCGSupport, |
| 72 | AttentionMetadataBuilder, |
| 73 | CommonAttentionMetadata, |
| 74 | ) |
| 75 | from vllm.v1.kv_cache_interface import ( |
| 76 | AttentionSpec, |
| 77 | FullAttentionSpec, |
| 78 | KVCacheSpec, |
| 79 | KVQuantMode, |
| 80 | SlidingWindowSpec, |
| 81 | ) |
| 82 | from vllm.v1.worker.cp_utils import ( |
| 83 | run_split_fa2_dcp_context_attention, |
| 84 | should_skip_dcp_context_attention, |
| 85 | should_split_fa2_dcp_context_attention, |
| 86 | split_dcp_context_queries, |
| 87 | ) |
| 88 | |
| 89 | logger = init_logger(__name__) |
| 90 | |
| 91 | FA4_DENSE_FLOAT_DTYPES = (torch.bfloat16, torch.float16) |
| 92 | FA4_DENSE_Q_TILE = 128 |
| 93 | FA4_DENSE_NUM_BLOCKS = 256 |
| 94 | FA4_DENSE_MAX_SEQLEN_K = 8192 |
| 95 | |
| 96 | |
| 97 | class FA4DenseAttentionKernel(VllmJitKernel["FA4DenseAttentionKernel.CompileKey"]): |
| 98 | """Warm paged, causal FA4 on Blackwell and Hopper.""" |
| 99 | |
| 100 | @dataclass(frozen=True) |
| 101 | class CompileKey: |
| 102 | q_stage: int |
| 103 | dtype: torch.dtype |
| 104 | qhead_per_kvhead: int |
| 105 | head_dim: int |
| 106 | page_size: int |
| 107 | has_window_left: bool |
| 108 | has_window_right: bool |
| 109 | softcap: float |
| 110 | is_single_batch: bool |
| 111 | num_splits: int |
| 112 | |
| 113 | @staticmethod |
| 114 | def kernel(*args: Any, **kwargs: Any) -> Any: |
| 115 | assert flash_attn_varlen_func is not None |
| 116 | return flash_attn_varlen_func(*args, **kwargs) |
| 117 | |
| 118 | def dispatch( # type: ignore[override] |
| 119 | self, |
| 120 | *, |
| 121 | q_stage: int, |
| 122 | dtype: torch.dtype, |
| 123 | num_qo_heads: int, |
| 124 | num_kv_heads: int, |
| 125 | head_dim: int, |
| 126 | page_size: int, |
| 127 | window_size: tuple[int, int], |
| 128 | softcap: float, |
| 129 | is_single_batch: bool = True, |
| 130 | num_splits: int = 1, |
| 131 | ) -> CompileKey: |
| 132 | return self.CompileKey( |
| 133 | q_stage=q_stage, |
| 134 | dtype=dtype, |
| 135 | qhead_per_kvhead=num_qo_heads // num_kv_heads, |
| 136 | head_dim=head_dim, |
| 137 | page_size=page_size, |
| 138 | has_window_left=window_size[0] >= 0, |
| 139 | has_window_right=window_size[1] >= 0, |
| 140 | softcap=softcap, |
| 141 | is_single_batch=is_single_batch, |
| 142 | num_splits=num_splits, |
| 143 | ) |
| 144 | |
| 145 | def get_warmup_keys( |
| 146 | self, |
| 147 | *, |
| 148 | vllm_config: VllmConfig, |
| 149 | kv_cache_spec: AttentionSpec, |
| 150 | num_qo_heads: int, |
| 151 | ) -> list[CompileKey]: |
| 152 | capability = current_platform.get_device_capability() |
| 153 | major = capability.major if capability is not None else None |
| 154 | if ( |
| 155 | vllm_config.parallel_config.decode_context_parallel_size != 1 |
| 156 | or vllm_config.model_config.rswa_window is not None |
| 157 | or kv_cache_spec.kv_quant_mode != KVQuantMode.NONE |
| 158 | or kv_cache_spec.dtype not in FA4_DENSE_FLOAT_DTYPES |
| 159 | or not isinstance(kv_cache_spec, (FullAttentionSpec, SlidingWindowSpec)) |
| 160 | ): |
| 161 | return [] |
| 162 | head_dim = kv_cache_spec.head_size |
| 163 | page_size = kv_cache_spec.block_size |
| 164 | sliding_window = kv_cache_spec.sliding_window |
| 165 | window_size = ( |
| 166 | (sliding_window - 1, 0) if sliding_window is not None else (-1, -1) |
| 167 | ) |
| 168 | softcap = float( |
| 169 | getattr( |
| 170 | vllm_config.model_config.hf_text_config, "attn_logit_softcapping", 0 |
| 171 | ) |
| 172 | or 0 |
| 173 | ) |
| 174 | if ( |
| 175 | major in (10, 11) |
| 176 | and uses_fa4_hd256_kernel(head_dim) |
| 177 | and page_size == FA4_HD256_PAGE_SIZE |
| 178 | ): |
| 179 | q_stages = WarmupChoices(1, 2) |
| 180 | single_batches = WarmupChoices(True, False) |
| 181 | split_counts = WarmupChoices(1) |
| 182 | elif ( |
| 183 | major == 9 and head_dim == 512 and window_size == (-1, -1) and softcap == 0 |
| 184 | ): |
| 185 | # SM90 forward has split/non-split variants; its transitive |
| 186 | # combine specializes at 32/64/128/256 splits. |
| 187 | q_stages = WarmupChoices(1) |
| 188 | single_batches = WarmupChoices(False) |
| 189 | split_counts = WarmupChoices(1, 32, 64, 128, 256) |
| 190 | else: |
| 191 | return [] |
| 192 | |
| 193 | return self._trace_dispatch(self.dispatch)( |
| 194 | q_stage=q_stages, |
| 195 | is_single_batch=single_batches, |
| 196 | num_splits=split_counts, |
| 197 | dtype=vllm_config.model_config.dtype, |
| 198 | num_qo_heads=num_qo_heads, |
| 199 | num_kv_heads=kv_cache_spec.num_kv_heads, |
| 200 | head_dim=head_dim, |
| 201 | page_size=page_size, |
| 202 | window_size=WarmupChoices(window_size), |
| 203 | softcap=softcap, |
| 204 | ) |
| 205 | |
| 206 | def compile(self, compile_key: CompileKey) -> None: |
| 207 | assert compile_flash_attn_varlen_func_from_specs is not None |
| 208 | max_seqlen_q = FA4_DENSE_Q_TILE + 1 if compile_key.q_stage == 2 else 1 |
| 209 | batch_size = 1 if compile_key.is_single_batch else 2 |
| 210 | kv_shape = ( |
| 211 | FA4_DENSE_NUM_BLOCKS, |
| 212 | compile_key.page_size, |
| 213 | 1, |
| 214 | compile_key.head_dim, |
| 215 | ) |
| 216 | kv_stride = ( |
| 217 | 2 * compile_key.page_size * compile_key.head_dim, |
| 218 | 2 * compile_key.head_dim, |
| 219 | 2 * compile_key.head_dim, |
| 220 | 1, |
| 221 | ) |
| 222 | compile_flash_attn_varlen_func_from_specs( |
| 223 | q_shape=( |
| 224 | batch_size * max_seqlen_q, |
| 225 | compile_key.qhead_per_kvhead, |
| 226 | compile_key.head_dim, |
| 227 | ), |
| 228 | k_shape=kv_shape, |
| 229 | v_shape=kv_shape, |
| 230 | k_stride=kv_stride, |
| 231 | v_stride=kv_stride, |
| 232 | q_dtype=compile_key.dtype, |
| 233 | cu_seqlens_q_shape=(batch_size + 1,), |
| 234 | seqused_k_shape=(batch_size,), |
| 235 | page_table_shape=( |
| 236 | batch_size, |
| 237 | FA4_DENSE_MAX_SEQLEN_K // compile_key.page_size, |
| 238 | ), |
| 239 | max_seqlen_q=max_seqlen_q, |
| 240 | max_seqlen_k=FA4_DENSE_MAX_SEQLEN_K, |
| 241 | causal=True, |
| 242 | window_size=[ |
| 243 | 1 if compile_key.has_window_left else -1, |
| 244 | 1 if compile_key.has_window_right else -1, |
| 245 | ], |
| 246 | softcap=compile_key.softcap, |
| 247 | num_splits=2 if compile_key.num_splits > 1 else 1, |
| 248 | fa_version=4, |
| 249 | return_softmax_lse=False, |
| 250 | ) |
| 251 | if compile_key.head_dim == 512 and compile_key.num_splits > 1: |
| 252 | from torch._subclasses.fake_tensor import FakeTensorMode |
| 253 | |
| 254 | from vllm.vllm_flash_attn.cute.interface import _flash_attn_fwd_combine |
| 255 | |
| 256 | # FA4's forward compile-only entry returns before its transitive |
| 257 | # SplitKV combine. Call the native combine entry in FakeTensorMode: |
| 258 | # it populates the native cache without launching the kernel. |
| 259 | with FakeTensorMode(): |
| 260 | _flash_attn_fwd_combine( |
| 261 | torch.empty( |
| 262 | (compile_key.num_splits, 2, 1, 512), |
| 263 | dtype=torch.float32, |
| 264 | device="cuda", |
| 265 | ), |
| 266 | torch.empty( |
| 267 | (compile_key.num_splits, 2, 1), |
| 268 | dtype=torch.float32, |
| 269 | device="cuda", |
| 270 | ), |
| 271 | torch.empty((2, 1, 512), dtype=compile_key.dtype, device="cuda"), |
| 272 | cu_seqlens=torch.empty((3,), dtype=torch.int32, device="cuda"), |
| 273 | ) |
| 274 | |
| 275 | def __call__( |
| 276 | self, |
| 277 | *, |
| 278 | q: torch.Tensor, |
| 279 | k: torch.Tensor, |
| 280 | v: torch.Tensor, |
| 281 | **kwargs: Any, |
| 282 | ) -> Any: |
| 283 | return self.kernel(q=q, k=k, v=v, **kwargs) |
| 284 | |
| 285 | |
| 286 | class FlashAttentionBackend(AttentionBackend): |
| 287 | supported_dtypes: ClassVar[list[torch.dtype]] = [torch.float16, torch.bfloat16] |
| 288 | supported_kv_cache_dtypes: ClassVar[list[CacheDType]] = [ |
| 289 | "auto", |
| 290 | "float16", |
| 291 | "bfloat16", |
| 292 | "fp8", |
| 293 | "fp8_e4m3", |
| 294 | ] |
| 295 | head_size_v: int | None = None |
| 296 | |
| 297 | @staticmethod |
| 298 | def _get_sm90_fa4_fp8_kv_block_size( |
| 299 | kv_cache_spec: KVCacheSpec | None = None, |
| 300 | ) -> int | None: |
| 301 | if kv_cache_spec is None: |
| 302 | vllm_config = get_current_vllm_config_or_none() |
| 303 | if vllm_config is None or vllm_config.model_config is None: |
| 304 | return None |
| 305 | head_size = vllm_config.model_config.get_head_size() |
| 306 | uses_fp8_kv = vllm_config.cache_config.cache_dtype in ( |
| 307 | "fp8", |
| 308 | "fp8_e4m3", |
| 309 | ) |
| 310 | else: |
| 311 | if not isinstance(kv_cache_spec, AttentionSpec): |
| 312 | return None |
| 313 | head_size = kv_cache_spec.head_size |
| 314 | uses_fp8_kv = kv_cache_spec.kv_quant_mode == KVQuantMode.FP8_PER_TENSOR |
| 315 | |
| 316 | if ( |
| 317 | current_platform.is_device_capability_family(90) |
| 318 | and uses_fp8_kv |
| 319 | and head_size == 512 |
| 320 | and get_flash_attn_version(head_size=head_size) == 4 |
| 321 | ): |
| 322 | # The SM90 FP8-KV-dequant kernel uses a 64-token TMA tile/page. |
| 323 | return 64 |
| 324 | return None |
| 325 | |
| 326 | @classmethod |
| 327 | def _get_fa4_hd256_block_size( |
| 328 | cls, kv_cache_spec: KVCacheSpec | None = None |
| 329 | ) -> int | None: |
| 330 | if kv_cache_spec is None: |
| 331 | vllm_config = get_current_vllm_config_or_none() |
| 332 | if vllm_config is None or vllm_config.model_config is None: |
| 333 | return None |
| 334 | head_size = vllm_config.model_config.get_head_size() |
| 335 | head_size_v = cls.head_size_v |
| 336 | else: |
| 337 | if not isinstance(kv_cache_spec, AttentionSpec): |
| 338 | return None |
| 339 | head_size = kv_cache_spec.head_size |
| 340 | head_size_v = kv_cache_spec.head_size_v |
| 341 | |
| 342 | if ( |
| 343 | uses_fa4_hd256_kernel(head_size, head_size_v) |
| 344 | and get_flash_attn_version( |
| 345 | head_size=head_size, |
| 346 | head_size_v=head_size_v, |
| 347 | supports_fa4_hd256=True, |
| 348 | ) |
| 349 | == 4 |
| 350 | ): |
| 351 | return FA4_HD256_PAGE_SIZE |
| 352 | return None |
| 353 | |
| 354 | @classmethod |
| 355 | def get_supported_kernel_block_sizes( |
| 356 | cls, kv_cache_spec: KVCacheSpec | None = None |
| 357 | ) -> list[int | MultipleOf]: |
| 358 | if block_size := cls._get_sm90_fa4_fp8_kv_block_size(kv_cache_spec): |
| 359 | return [block_size] |
| 360 | if block_size := cls._get_fa4_hd256_block_size(kv_cache_spec): |
| 361 | # Sliding-window specs select the smallest advertised size. |
| 362 | return [block_size] |
| 363 | return [MultipleOf(16)] |
| 364 | |
| 365 | forward_includes_kv_cache_update: bool = False |
| 366 | |
| 367 | @classmethod |
| 368 | def get_preferred_block_size(cls, default_block_size: int) -> int: |
| 369 | if block_size := cls._get_sm90_fa4_fp8_kv_block_size(): |
| 370 | return max(default_block_size, block_size) |
| 371 | if block_size := cls._get_fa4_hd256_block_size(): |
| 372 | return max(default_block_size, block_size) |
| 373 | if current_platform.is_xpu(): |
| 374 | return max(default_block_size, 64) |
| 375 | return super().get_preferred_block_size(default_block_size) |
| 376 | |
| 377 | @staticmethod |
| 378 | def get_name() -> str: |
| 379 | return "FLASH_ATTN" |
| 380 | |
| 381 | @classmethod |
| 382 | def supports_rswa(cls) -> bool: |
| 383 | return True |
| 384 | |
| 385 | @classmethod |
| 386 | def supports_sliding_window(cls) -> bool: |
| 387 | return True |
| 388 | |
| 389 | @classmethod |
| 390 | def supports_batch_invariance(cls) -> bool: |
| 391 | return True |
| 392 | |
| 393 | @classmethod |
| 394 | def supports_non_causal(cls) -> bool: |
| 395 | return True |
| 396 | |
| 397 | @classmethod |
| 398 | def supports_attn_type(cls, attn_type: str) -> bool: |
| 399 | """FlashAttention supports all attention types.""" |
| 400 | return attn_type in ( |
| 401 | AttentionType.DECODER, |
| 402 | AttentionType.ENCODER, |
| 403 | AttentionType.ENCODER_ONLY, |
| 404 | AttentionType.ENCODER_DECODER, |
| 405 | ) |
| 406 | |
| 407 | @classmethod |
| 408 | def supports_per_head_quant_scales(cls) -> bool: |
| 409 | fa_version = get_flash_attn_version() |
| 410 | return fa_version is not None and fa_version >= 3 |
| 411 | |
| 412 | @staticmethod |
| 413 | def get_impl_cls() -> type["FlashAttentionImpl"]: |
| 414 | return FlashAttentionImpl |
| 415 | |
| 416 | @staticmethod |
| 417 | def get_builder_cls() -> type["FlashAttentionMetadataBuilder"]: |
| 418 | return FlashAttentionMetadataBuilder |
| 419 | |
| 420 | @classmethod |
| 421 | def supports_head_size(cls, head_size: int) -> bool: |
| 422 | if head_size % 8 != 0: |
| 423 | return False |
| 424 | if head_size <= 256: |
| 425 | return True |
| 426 | if is_fa_version_supported(4): |
| 427 | return head_size <= 512 |
| 428 | return False |
| 429 | |
| 430 | @classmethod |
| 431 | def supports_kv_cache_dtype(cls, kv_cache_dtype: CacheDType | None) -> bool: |
| 432 | if kv_cache_dtype is None: |
| 433 | return True |
| 434 | if kv_cache_dtype not in cls.supported_kv_cache_dtypes: |
| 435 | return False |
| 436 | if is_quantized_kv_cache(kv_cache_dtype): |
| 437 | return flash_attn_supports_kv_cache_dtype(kv_cache_dtype) |
| 438 | return True |
| 439 | |
| 440 | @classmethod |
| 441 | def supports_mm_prefix(cls) -> bool: |
| 442 | return is_fa_version_supported(4) |
| 443 | |
| 444 | @classmethod |
| 445 | def supports_sink(cls) -> bool: |
| 446 | if not is_flash_attn_varlen_func_available(): |
| 447 | return False |
| 448 | return flash_attn_supports_sinks() |
| 449 | |
| 450 | @classmethod |
| 451 | def supports_compute_capability(cls, capability: DeviceCapability) -> bool: |
| 452 | return capability >= DeviceCapability(8, 0) |
| 453 | |
| 454 | @classmethod |
| 455 | def supports_combination( |
| 456 | cls, |
| 457 | head_size: int, |
| 458 | dtype: torch.dtype, |
| 459 | kv_cache_dtype: CacheDType | None, |
| 460 | block_size: int | None, |
| 461 | use_mla: bool, |
| 462 | has_sink: bool, |
| 463 | use_sparse: bool, |
| 464 | use_mm_prefix: bool, |
| 465 | device_capability: DeviceCapability, |
| 466 | ) -> str | None: |
| 467 | if has_sink and device_capability < DeviceCapability(9, 0): |
| 468 | return "sink not supported on compute capability < 9.0" |
| 469 | if ( |
| 470 | use_mm_prefix |
| 471 | and kv_cache_dtype is not None |
| 472 | and is_quantized_kv_cache(kv_cache_dtype) |
| 473 | and device_capability == DeviceCapability(9, 0) |
| 474 | ): |
| 475 | return "SM90 FP8 KV with mm_prefix requires Triton" |
| 476 | if ( |
| 477 | kv_cache_dtype is not None |
| 478 | and is_quantized_kv_cache(kv_cache_dtype) |
| 479 | and not flash_attn_supports_kv_cache_dtype( |
| 480 | kv_cache_dtype, |
| 481 | head_size=head_size, |
| 482 | head_size_v=head_size, |
| 483 | has_sinks=has_sink, |
| 484 | kv_cache_block_size=block_size, |
| 485 | supports_fa4_hd256=True, |
| 486 | ) |
| 487 | ): |
| 488 | return ( |
| 489 | "FP8 KV cache requires FA3 on SM90, FA4 with head_size=512 " |
| 490 | "on SM90, or FA4 on SM100" |
| 491 | ) |
| 492 | if ( |
| 493 | use_mm_prefix |
| 494 | and get_flash_attn_version( |
| 495 | head_size=head_size, |
| 496 | has_sinks=has_sink, |
| 497 | kv_cache_block_size=block_size, |
| 498 | supports_fa4_hd256=True, |
| 499 | ) |
| 500 | != 4 |
| 501 | ): |
| 502 | return ( |
| 503 | "mm_prefix (PrefixLM bidirectional attention) requires " |
| 504 | "FlashAttention v4, which does not resolve for this " |
| 505 | "head_size" |
| 506 | ) |
| 507 | return None |
| 508 | |
| 509 | |
| 510 | @dataclass |
| 511 | class FlashAttentionMetadata: |
| 512 | # NOTE(sang): Definition of context_len, query_len, and seq_len. |
| 513 | # |---------- N-1 iteration --------| |
| 514 | # |---------------- N iteration ---------------------| |
| 515 | # |- tokenA -|......................|-- newTokens ---| |
| 516 | # |---------- context_len ----------| |
| 517 | # |-------------------- seq_len ---------------------| |
| 518 | # |-- query_len ---| |
| 519 | |
| 520 | num_actual_tokens: int # Number of tokens excluding padding. |
| 521 | max_query_len: int |
| 522 | query_start_loc: torch.Tensor |
| 523 | max_seq_len: int |
| 524 | seq_lens: torch.Tensor |
| 525 | block_table: torch.Tensor |
| 526 | slot_mapping: torch.Tensor |
| 527 | |
| 528 | # For cascade attention. |
| 529 | use_cascade: bool |
| 530 | common_prefix_len: int |
| 531 | cu_prefix_query_lens: torch.Tensor | None |
| 532 | prefix_kv_lens: torch.Tensor | None |
| 533 | suffix_kv_lens: torch.Tensor | None |
| 534 | |
| 535 | # For GQA DCP |
| 536 | max_dcp_context_kv_len: int | None = None |
| 537 | dcp_context_kv_lens: torch.Tensor | None = None |
| 538 | |
| 539 | # Split counts for FA2 DCP context attention. num_prefill_* tracks |
| 540 | # context-bearing extend rows; pure prefills do not attend to DCP context. |
| 541 | num_decode_reqs: int = 0 |
| 542 | num_prefill_reqs: int = 0 |
| 543 | num_decode_tokens: int = 0 |
| 544 | num_prefill_tokens: int = 0 |
| 545 | |
| 546 | # Optional aot scheduling |
| 547 | scheduler_metadata: torch.Tensor | None = None |
| 548 | prefix_scheduler_metadata: torch.Tensor | None = None |
| 549 | max_num_splits: int = 0 |
| 550 | |
| 551 | causal: bool | torch.Tensor = True |
| 552 | |
| 553 | # PrefixLM bidirectional range containing each scheduled query token. |
| 554 | # Shape: (num_actual_tokens, 2) int32, absolute [start, end] bounds; |
| 555 | # (-1, -1) for query tokens outside every multimodal range. |
| 556 | mm_prefix_query_range_tensor: torch.Tensor | None = None |
| 557 | |
| 558 | # Reference Sliding Window Attention (R-SWA) fields. |
| 559 | # rswa_prefix_lens: per-request prompt lengths [num_reqs], int32, CUDA. |
| 560 | # rswa_window: sliding window size (scalar int, for logic checks). |
| 561 | # rswa_window_tensor: [1] int32 CUDA tensor — pre-allocated in build() so |
| 562 | # no CPU→CUDA copy is needed inside forward() during CUDA graph capture. |
| 563 | # Only populated when the model uses R-SWA (Unlimited-OCR). |
| 564 | rswa_prefix_lens: torch.Tensor | None = None |
| 565 | rswa_window: int | None = None |
| 566 | rswa_window_tensor: torch.Tensor | None = None |
| 567 | |
| 568 | |
| 569 | def _get_sliding_window_configs( |
| 570 | vllm_config: VllmConfig, |
| 571 | ) -> set[tuple[int, int] | None]: |
| 572 | """Get the set of all sliding window configs used in the model. |
| 573 | |
| 574 | Only inspects FlashAttentionImpl layers, including implementations wrapped |
| 575 | by a composite backend. Other backends (e.g. TurboQuant, MLA) use their own |
| 576 | metadata builders and are skipped. |
| 577 | """ |
| 578 | sliding_window_configs: set[tuple[int, int] | None] = set() |
| 579 | layers = get_layers_from_vllm_config(vllm_config, Attention) |
| 580 | for layer in layers.values(): |
| 581 | for impl in layer.impl.get_impl_variants(): |
| 582 | if isinstance(impl, FlashAttentionImpl): |
| 583 | sliding_window_configs.add(impl.sliding_window) |
| 584 | return sliding_window_configs |
| 585 | |
| 586 | |
| 587 | def _maybe_symmetrize_window( |
| 588 | window: tuple[int, int] | None, |
| 589 | causal: bool | torch.Tensor, |
| 590 | ) -> tuple[int, int] | None: |
| 591 | """Make a causal sliding window ``(w, 0)`` symmetric ``(w, w)`` when attention |
| 592 | is non-causal, so bidirectional queries attend in both directions. Leaves |
| 593 | full-attention ``(-1, -1)`` and already-symmetric windows untouched. |
| 594 | """ |
| 595 | non_causal = isinstance(causal, torch.Tensor) or causal is False |
| 596 | if window is not None and window[0] >= 0 and window[1] == 0 and non_causal: |
| 597 | return (window[0], window[0]) |
| 598 | return window |
| 599 | |
| 600 | |
| 601 | class FlashAttentionMetadataBuilder(AttentionMetadataBuilder[FlashAttentionMetadata]): |
| 602 | # FA3: |
| 603 | # Supports full cudagraphs for all cases. |
| 604 | # |
| 605 | # FA2: |
| 606 | # For FA2, a graph is captured with max_query_len=1, (which is what we |
| 607 | # capture by default for num_tokens <= max_num_seqs when there is no |
| 608 | # spec-decode) then these graphs will not work for mixed prefill-decode |
| 609 | # (unlike FA3). This is due to special max_query_len=1 packed-GQA handling |
| 610 | # in FA2. |
| 611 | # In summary if we are running with spec decodes the graphs would |
| 612 | # work for mixed prefill-decode and uniform-decode. But for non-spec decodes |
| 613 | # the graphs would not work for mixed prefill-decode; sorta the inverse |
| 614 | # of UNIFORM_SINGLE_TOKEN_DECODE. |
| 615 | # There's probably a better way to describe this using `AttentionCGSupport` |
| 616 | # but for now just set it to `UNIFORM_BATCH` to get use to drop down |
| 617 | # to FULL_AND_PIECEWISE. |
| 618 | # TODO(luka, lucas): audit FA2 as part of: |
| 619 | # https://github.com/vllm-project/vllm/issues/22945 |
| 620 | _cudagraph_support = ( |
| 621 | AttentionCGSupport.ALWAYS |
| 622 | if get_flash_attn_version() == 3 |
| 623 | else AttentionCGSupport.UNIFORM_BATCH |
| 624 | ) |
| 625 | supports_update_block_table: bool = True |
| 626 | |
| 627 | @classmethod |
| 628 | def get_cudagraph_support( |
| 629 | cls, |
| 630 | vllm_config: "VllmConfig", |
| 631 | kv_cache_spec: "KVCacheSpec", |
| 632 | ) -> AttentionCGSupport: |
| 633 | return cls._cudagraph_support |
| 634 | |
| 635 | def _get_scheduler_metadata( |
| 636 | self, |
| 637 | *, |
| 638 | aot_schedule: bool, |
| 639 | batch_size: int, |
| 640 | cu_query_lens: torch.Tensor, |
| 641 | max_query_len: int, |
| 642 | seqlens: torch.Tensor, |
| 643 | max_seq_len: int, |
| 644 | causal: bool | torch.Tensor, |
| 645 | max_num_splits: int, |
| 646 | ) -> torch.Tensor | None: |
| 647 | if not aot_schedule: |
| 648 | return None |
| 649 | |
| 650 | cache_dtype = self.cache_config.cache_dtype |
| 651 | if is_quantized_kv_cache(cache_dtype): |
| 652 | qkv_dtype = current_platform.fp8_dtype() |
| 653 | else: |
| 654 | qkv_dtype = self.kv_cache_dtype |
| 655 | return get_scheduler_metadata( |
| 656 | batch_size=batch_size, |
| 657 | max_seqlen_q=max_query_len, |
| 658 | max_seqlen_k=max_seq_len, |
| 659 | num_heads_q=self.num_heads_q * self.dcp_world_size, |
| 660 | num_heads_kv=self.num_heads_kv, |
| 661 | headdim=self.headdim, |
| 662 | cache_seqlens=seqlens, |
| 663 | qkv_dtype=qkv_dtype, |
| 664 | cu_seqlens_q=cu_query_lens, |
| 665 | page_size=self.block_size, |
| 666 | causal=causal, |
| 667 | window_size=_maybe_symmetrize_window(self.aot_sliding_window, causal), |
| 668 | num_splits=max_num_splits, |
| 669 | ) |
| 670 | |
| 671 | def _store_scheduler_metadata( |
| 672 | self, scheduler_metadata: torch.Tensor | None |
| 673 | ) -> torch.Tensor | None: |
| 674 | if self.use_full_cuda_graph and scheduler_metadata is not None: |
| 675 | n = scheduler_metadata.shape[0] |
| 676 | assert self.scheduler_metadata is not None |
| 677 | self.scheduler_metadata[:n] = scheduler_metadata |
| 678 | # NOTE(woosuk): We should zero out the rest of the scheduler |
| 679 | # metadata to guarantee the correctness. Otherwise, some thread |
| 680 | # blocks may use the invalid scheduler metadata and overwrite the |
| 681 | # output buffer. |
| 682 | self.scheduler_metadata[n:] = 0 |
| 683 | return self.scheduler_metadata[:n] |
| 684 | return scheduler_metadata |
| 685 | |
| 686 | def __init__( |
| 687 | self, |
| 688 | kv_cache_spec: AttentionSpec, |
| 689 | layer_names: list[str], |
| 690 | vllm_config: VllmConfig, |
| 691 | device: torch.device, |
| 692 | ): |
| 693 | super().__init__(kv_cache_spec, layer_names, vllm_config, device) |
| 694 | self.model_config = vllm_config.model_config |
| 695 | self.parallel_config = vllm_config.parallel_config |
| 696 | self.cache_config = vllm_config.cache_config |
| 697 | self.compilation_config = vllm_config.compilation_config |
| 698 | self.attention_config = vllm_config.attention_config |
| 699 | |
| 700 | self.num_heads_q = get_num_attention_heads_from_layers( |
| 701 | vllm_config, layer_names |
| 702 | ) or self.model_config.get_num_attention_heads(self.parallel_config) |
| 703 | self.num_heads_kv = kv_cache_spec.num_kv_heads |
| 704 | self.kv_cache_dtype = kv_cache_spec.dtype |
| 705 | self.headdim = kv_cache_spec.head_size |
| 706 | self.block_size = kv_cache_spec.block_size |
| 707 | |
| 708 | self.max_num_splits = 0 # No upper bound on the number of splits. |
| 709 | self.aot_schedule = get_flash_attn_version() == 3 |
| 710 | head_size_v = getattr(kv_cache_spec, "head_size_v", None) |
| 711 | fa_version = get_flash_attn_version( |
| 712 | head_size=self.headdim, |
| 713 | head_size_v=head_size_v, |
| 714 | kv_cache_block_size=self.block_size, |
| 715 | supports_fa4_hd256=True, |
| 716 | ) |
| 717 | self.fa4_hd256 = fa_version == 4 and uses_fa4_hd256_kernel( |
| 718 | self.headdim, head_size_v |
| 719 | ) |
| 720 | |
| 721 | if ( |
| 722 | vllm_config.kernel_config.enable_jit_warmup |
| 723 | and self.model_config.hf_config.model_type in ("gemma4", "gemma4_unified") |
| 724 | and fa_version == 4 |
| 725 | ): |
| 726 | _FA4_DENSE_ATTENTION_KERNEL.register_warmup( |
| 727 | vllm_config=vllm_config, |
| 728 | kv_cache_spec=kv_cache_spec, |
| 729 | num_qo_heads=self.num_heads_q, |
| 730 | ) |
| 731 | |
| 732 | self.dcp_world_size, self.dcp_rank = get_dcp_world_size_and_rank( |
| 733 | kv_cache_spec.dcp_sharded |
| 734 | ) |
| 735 | |
| 736 | # Fused draft decode reuses the captured metadata object across draft |
| 737 | # steps. For DCP, build-time host-side decisions such as |
| 738 | # skip_dcp_context_attention() can change the metadata shape/control |
| 739 | # path (for example max_dcp_context_kv_len), and those Python-side |
| 740 | # fields are not refreshed in-place between graph replays. Keep the |
| 741 | # fused path disabled until DCP gets a full replay-safe refresh model. |
| 742 | self.supports_draft_decode_metadata_update = self.dcp_world_size == 1 |
| 743 | |
| 744 | self.cp_kv_cache_interleave_size = ( |
| 745 | self.parallel_config.cp_kv_cache_interleave_size |
| 746 | ) |
| 747 | |
| 748 | self.use_full_cuda_graph = ( |
| 749 | self.compilation_config.cudagraph_mode.has_full_cudagraphs() |
| 750 | ) |
| 751 | self.max_cudagraph_size = self.compilation_config.max_cudagraph_capture_size |
| 752 | |
| 753 | if self.use_full_cuda_graph and self.aot_schedule: |
| 754 | # FA3 scheduler_metadata size: 1 + round_up(batch_size, 4) * 4 |
| 755 | # The +1 is for the tile_count_semaphore (synchronization). |
| 756 | # The 4 slots per batch element (num_prepare_batch_vectors) are: |
| 757 | # prepare_varlen + dynamic_split + sort_batches + head_swizzle |
| 758 | # See: https://github.com/vllm-project/flash-attention/blob/5824e6e/hopper/flash_api.cpp#L664-L671 # noqa: E501 |
| 759 | max_batch_size = max( |
| 760 | vllm_config.scheduler_config.max_num_seqs, |
| 761 | self.max_cudagraph_size or 0, |
| 762 | ) |
| 763 | self.scheduler_metadata = torch.zeros( |
| 764 | 1 + round_up(max_batch_size, 4) * 4, |
| 765 | dtype=torch.int32, |
| 766 | device=self.device, |
| 767 | ) |
| 768 | # When using cuda graph, we need to set the upper bound of the |
| 769 | # number of splits so that large enough intermediate buffers are |
| 770 | # pre-allocated during capture. |
| 771 | self.max_num_splits = ( |
| 772 | self.attention_config.flash_attn_max_num_splits_for_cuda_graph |
| 773 | ) |
| 774 | |
| 775 | if self.dcp_world_size > 1: |
| 776 | max_num_reqs = vllm_config.scheduler_config.max_num_seqs |
| 777 | self._dcp_context_kv_lens = torch.zeros( |
| 778 | max_num_reqs, |
| 779 | dtype=torch.int32, |
| 780 | device=self.device, |
| 781 | ) |
| 782 | |
| 783 | # Sliding window size to be used with the AOT scheduler will be |
| 784 | # populated on first build() call. |
| 785 | self.aot_sliding_window: tuple[int, int] | None = None |
| 786 | |
| 787 | # R-SWA: persistent CUDA-graph-safe buffers owned by this builder. |
| 788 | self.rswa_window: int | None = self.model_config.rswa_window |
| 789 | self.persistent_rswa_prefix_lens: torch.Tensor | None = None |
| 790 | self.persistent_rswa_window_tensor: torch.Tensor | None = None |
| 791 | if self.rswa_window is not None: |
| 792 | max_num_reqs = vllm_config.scheduler_config.max_num_seqs |
| 793 | self.persistent_rswa_prefix_lens = torch.zeros( |
| 794 | max_num_reqs, dtype=torch.int32, device=self.device |
| 795 | ) |
| 796 | self.persistent_rswa_window_tensor = torch.tensor( |
| 797 | [self.rswa_window], dtype=torch.int32, device=self.device |
| 798 | ) |
| 799 | |
| 800 | # mm_prefix: persistent staging + device buffers owned by this builder, |
| 801 | # sized by scheduled query tokens so build() never allocates. |
| 802 | self.mm_prefix_query_ranges_cpu: torch.Tensor | None = None |
| 803 | self.mm_prefix_query_ranges_np: np.ndarray | None = None |
| 804 | self.mm_prefix_query_ranges_gpu: torch.Tensor | None = None |
| 805 | if self.model_config.is_mm_prefix_lm: |
| 806 | max_num_tokens = vllm_config.scheduler_config.max_num_batched_tokens |
| 807 | self.mm_prefix_query_ranges_cpu = torch.empty( |
| 808 | (max_num_tokens, 2), dtype=torch.int32, pin_memory=PIN_MEMORY |
| 809 | ) |
| 810 | self.mm_prefix_query_ranges_np = self.mm_prefix_query_ranges_cpu.numpy() |
| 811 | self.mm_prefix_query_ranges_gpu = torch.empty( |
| 812 | (max_num_tokens, 2), dtype=torch.int32, device=self.device |
| 813 | ) |
| 814 | |
| 815 | def build( |
| 816 | self, |
| 817 | common_prefix_len: int, |
| 818 | common_attn_metadata: CommonAttentionMetadata, |
| 819 | fast_build: bool = False, |
| 820 | ) -> FlashAttentionMetadata: |
| 821 | """fast_build disables AOT scheduling, used when there will be few |
| 822 | iterations i.e. spec-decode |
| 823 | """ |
| 824 | num_reqs = common_attn_metadata.num_reqs |
| 825 | num_actual_tokens = common_attn_metadata.num_actual_tokens |
| 826 | max_query_len = common_attn_metadata.max_query_len |
| 827 | max_seq_len = common_attn_metadata.max_seq_len |
| 828 | query_start_loc = common_attn_metadata.query_start_loc |
| 829 | seq_lens = common_attn_metadata.seq_lens |
| 830 | block_table_tensor = common_attn_metadata.block_table_tensor |
| 831 | slot_mapping = common_attn_metadata.slot_mapping |
| 832 | causal = common_attn_metadata.causal |
| 833 | |
| 834 | # Disable AOT schedule for spec-decode proposer (not worth the overhead) |
| 835 | # and for batch invariance (schedule varies with max_seqlen_q/k). |
| 836 | aot_schedule = ( |
| 837 | self.aot_schedule and not fast_build and not envs.VLLM_BATCH_INVARIANT |
| 838 | ) |
| 839 | |
| 840 | if self.aot_sliding_window is None: |
| 841 | self.aot_sliding_window = (-1, -1) |
| 842 | # For the AOT scheduler we need the sliding window value to be |
| 843 | # constant for all layers to. We have to populate this on the first |
| 844 | # build() call so the layers are constructed (cannot populate) |
| 845 | # in __init__. |
| 846 | if aot_schedule: |
| 847 | sliding_window_configs = _get_sliding_window_configs(self.vllm_config) |
| 848 | if len(sliding_window_configs) == 1: |
| 849 | sliding_window_config = sliding_window_configs.pop() |
| 850 | if sliding_window_config is not None: |
| 851 | self.aot_sliding_window = sliding_window_config |
| 852 | elif len(sliding_window_configs) > 1: |
| 853 | self.aot_schedule = False |
| 854 | aot_schedule = False |
| 855 | |
| 856 | max_num_splits = 0 # 0 means use FA3's heuristics, not CG compatible |
| 857 | if ( |
| 858 | self.use_full_cuda_graph |
| 859 | and self.max_cudagraph_size is not None |
| 860 | and num_actual_tokens <= self.max_cudagraph_size |
| 861 | ): |
| 862 | # NOTE(woosuk): Setting num_splits > 1 may increase the memory |
| 863 | # usage, because the intermediate buffers of size [num_splits, |
| 864 | # num_heads, num_tokens, head_size] are allocated. Therefore, |
| 865 | # we only set num_splits when using cuda graphs. |
| 866 | max_num_splits = self.max_num_splits |
| 867 | |
| 868 | if envs.VLLM_BATCH_INVARIANT: |
| 869 | max_num_splits = 1 |
| 870 | |
| 871 | use_cascade = common_prefix_len > 0 |
| 872 | max_dcp_context_kv_len = 0 |
| 873 | dcp_context_kv_lens = None |
| 874 | num_decode_reqs = 0 |
| 875 | num_prefill_reqs = 0 |
| 876 | num_decode_tokens = 0 |
| 877 | num_prefill_tokens = 0 |
| 878 | |
| 879 | cu_prefix_query_lens = None |
| 880 | prefix_kv_lens = None |
| 881 | suffix_kv_lens = None |
| 882 | prefix_scheduler_metadata = None |
| 883 | |
| 884 | if self.dcp_world_size > 1: |
| 885 | query_lens = query_start_loc[1:] - query_start_loc[:-1] |
| 886 | context_kv_lens = seq_lens - query_lens |
| 887 | local_context_kv_lens = get_dcp_local_seq_lens( |
| 888 | context_kv_lens, |
| 889 | self.dcp_world_size, |
| 890 | self.dcp_rank, |
| 891 | self.cp_kv_cache_interleave_size, |
| 892 | ) |
| 893 | self._dcp_context_kv_lens[:num_reqs] = local_context_kv_lens |
| 894 | self._dcp_context_kv_lens[num_reqs:] = 0 |
| 895 | dcp_context_kv_lens = self._dcp_context_kv_lens[:num_reqs] |
| 896 | |
| 897 | skip_dcp_context_attention = False |
| 898 | if common_attn_metadata.seq_lens_cpu_upper_bound is not None: |
| 899 | query_lens_cpu = ( |
| 900 | common_attn_metadata.query_start_loc_cpu[1 : num_reqs + 1] |
| 901 | - common_attn_metadata.query_start_loc_cpu[:num_reqs] |
| 902 | ) |
| 903 | context_kv_lens_cpu = ( |
| 904 | common_attn_metadata.seq_lens_cpu_upper_bound[:num_reqs] |
| 905 | - query_lens_cpu |
| 906 | ) |
| 907 | skip_dcp_context_attention = should_skip_dcp_context_attention( |
| 908 | context_kv_lens_cpu |
| 909 | ) |
| 910 | |
| 911 | if max_query_len > 1: |
| 912 | ( |
| 913 | num_decode_reqs, |
| 914 | num_prefill_reqs, |
| 915 | num_decode_tokens, |
| 916 | num_prefill_tokens, |
| 917 | ) = split_dcp_context_queries( |
| 918 | common_attn_metadata.query_start_loc_cpu, |
| 919 | common_attn_metadata.seq_lens_cpu_upper_bound, |
| 920 | max_query_len, |
| 921 | num_actual_tokens, |
| 922 | ) |
| 923 | |
| 924 | # After DCP distribution, the maximum number of tokens for any rank is |
| 925 | # ceil(L / (N * I)) * I, where L is max_seq_len, N is dcp_world_size, |
| 926 | # and I is cp_kv_cache_interleave_size. |
| 927 | # This eliminates GPU->CPU sync while minimizing workspace over-allocation. |
| 928 | if skip_dcp_context_attention: |
| 929 | max_dcp_context_kv_len = 0 |
| 930 | scheduler_metadata = None |
| 931 | else: |
| 932 | num_partitions = self.dcp_world_size * self.cp_kv_cache_interleave_size |
| 933 | max_dcp_context_kv_len = ( |
| 934 | (max_seq_len + num_partitions - 1) // num_partitions |
| 935 | ) * self.cp_kv_cache_interleave_size |
| 936 | |
| 937 | scheduler_metadata = self._get_scheduler_metadata( |
| 938 | aot_schedule=aot_schedule, |
| 939 | batch_size=num_reqs, |
| 940 | cu_query_lens=query_start_loc, |
| 941 | max_query_len=max_query_len, |
| 942 | seqlens=dcp_context_kv_lens, |
| 943 | max_seq_len=max_dcp_context_kv_len, |
| 944 | causal=False, |
| 945 | max_num_splits=max_num_splits, |
| 946 | ) |
| 947 | elif use_cascade: |
| 948 | cu_prefix_query_lens = torch.tensor( |
| 949 | [0, num_actual_tokens], dtype=torch.int32, device=self.device |
| 950 | ) |
| 951 | prefix_kv_lens = torch.tensor( |
| 952 | [common_prefix_len], dtype=torch.int32, device=self.device |
| 953 | ) |
| 954 | # Use GPU tensor directly - no CPU sync needed |
| 955 | suffix_kv_lens = seq_lens[:num_reqs] - common_prefix_len |
| 956 | prefix_scheduler_metadata = self._get_scheduler_metadata( |
| 957 | aot_schedule=aot_schedule, |
| 958 | batch_size=1, |
| 959 | cu_query_lens=cu_prefix_query_lens, |
| 960 | max_query_len=num_actual_tokens, |
| 961 | seqlens=prefix_kv_lens, |
| 962 | max_seq_len=common_prefix_len, |
| 963 | causal=False, |
| 964 | max_num_splits=max_num_splits, |
| 965 | ) |
| 966 | scheduler_metadata = self._get_scheduler_metadata( |
| 967 | aot_schedule=aot_schedule, |
| 968 | batch_size=num_reqs, |
| 969 | cu_query_lens=query_start_loc, |
| 970 | max_query_len=max_query_len, |
| 971 | seqlens=suffix_kv_lens, |
| 972 | max_seq_len=max_seq_len - common_prefix_len, |
| 973 | causal=True, |
| 974 | max_num_splits=max_num_splits, |
| 975 | ) |
| 976 | else: |
| 977 | scheduler_metadata = self._get_scheduler_metadata( |
| 978 | aot_schedule=aot_schedule, |
| 979 | batch_size=num_reqs, |
| 980 | cu_query_lens=query_start_loc, |
| 981 | max_query_len=max_query_len, |
| 982 | seqlens=seq_lens, |
| 983 | max_seq_len=max_seq_len, |
| 984 | causal=causal, |
| 985 | max_num_splits=max_num_splits, |
| 986 | ) |
| 987 | scheduler_metadata = self._store_scheduler_metadata(scheduler_metadata) |
| 988 | |
| 989 | if isinstance(causal, torch.Tensor) and causal.dtype != torch.int32: |
| 990 | raise ValueError( |
| 991 | f"Per-request causal tensor must be int32, got {causal.dtype}. " |
| 992 | "Casting here would allocate a fresh tensor each build() and " |
| 993 | "break FULL CUDA graph replay; allocate the buffer as int32." |
| 994 | ) |
| 995 | |
| 996 | attn_metadata = FlashAttentionMetadata( |
| 997 | num_actual_tokens=num_actual_tokens, |
| 998 | max_query_len=max_query_len, |
| 999 | query_start_loc=query_start_loc, |
| 1000 | max_seq_len=max_seq_len, |
| 1001 | seq_lens=seq_lens, |
| 1002 | block_table=block_table_tensor, |
| 1003 | slot_mapping=slot_mapping, |
| 1004 | max_dcp_context_kv_len=max_dcp_context_kv_len, |
| 1005 | dcp_context_kv_lens=dcp_context_kv_lens, |
| 1006 | num_decode_reqs=num_decode_reqs, |
| 1007 | num_prefill_reqs=num_prefill_reqs, |
| 1008 | num_decode_tokens=num_decode_tokens, |
| 1009 | num_prefill_tokens=num_prefill_tokens, |
| 1010 | use_cascade=use_cascade, |
| 1011 | common_prefix_len=common_prefix_len, |
| 1012 | scheduler_metadata=scheduler_metadata, |
| 1013 | cu_prefix_query_lens=cu_prefix_query_lens, |
| 1014 | prefix_kv_lens=prefix_kv_lens, |
| 1015 | suffix_kv_lens=suffix_kv_lens, |
| 1016 | prefix_scheduler_metadata=prefix_scheduler_metadata, |
| 1017 | max_num_splits=max_num_splits, |
| 1018 | causal=causal, |
| 1019 | ) |
| 1020 | |
| 1021 | # Compute mm_prefix range tensor if the batch contains |
| 1022 | # multimodal tokens with bidirectional ranges. Built for every FA |
| 1023 | # group; Gemma4 nulls the field for its non-sliding layers. |
| 1024 | mm_ranges = common_attn_metadata.mm_req_doc_ranges |
| 1025 | if mm_ranges is not None and self.mm_prefix_query_ranges_np is not None: |
| 1026 | # The upper bound is exact for prefill rows, which is where |
| 1027 | # mm_prefix ranges live; decode rows only ever get an optimistic |
| 1028 | # (larger) context, moving them further past every range. |
| 1029 | assert common_attn_metadata.seq_lens_cpu_upper_bound is not None, ( |
| 1030 | "mm_prefix requires seq_lens_cpu_upper_bound" |
| 1031 | ) |
| 1032 | num_mm_tokens = fill_mm_prefix_query_ranges( |
| 1033 | self.mm_prefix_query_ranges_np, |
| 1034 | mm_ranges, |
| 1035 | common_attn_metadata.query_start_loc_cpu, |
| 1036 | common_attn_metadata.seq_lens_cpu_upper_bound, |
| 1037 | ) |
| 1038 | if num_mm_tokens > 0: |
| 1039 | assert self.mm_prefix_query_ranges_cpu is not None |
| 1040 | assert self.mm_prefix_query_ranges_gpu is not None |
| 1041 | mm_query_ranges = self.mm_prefix_query_ranges_gpu[:num_mm_tokens] |
| 1042 | mm_query_ranges.copy_( |
| 1043 | self.mm_prefix_query_ranges_cpu[:num_mm_tokens], |
| 1044 | non_blocking=True, |
| 1045 | ) |
| 1046 | attn_metadata.mm_prefix_query_range_tensor = mm_query_ranges |
| 1047 | |
| 1048 | # R-SWA: copy prefix lengths into persistent buffers (outside the |
| 1049 | # compiled region) so forward() never allocates during CUDA graph |
| 1050 | # capture. rswa_window is a static model config scalar read here. |
| 1051 | if ( |
| 1052 | self.rswa_window is not None |
| 1053 | and common_attn_metadata.rswa_prefix_lens is not None |
| 1054 | ): |
| 1055 | assert self.persistent_rswa_prefix_lens is not None |
| 1056 | assert self.persistent_rswa_window_tensor is not None |
| 1057 | src = common_attn_metadata.rswa_prefix_lens |
| 1058 | rswa_prefix_lens = self.persistent_rswa_prefix_lens[:num_reqs] |
| 1059 | rswa_prefix_lens.copy_(src[:num_reqs], non_blocking=True) |
| 1060 | attn_metadata.rswa_prefix_lens = rswa_prefix_lens |
| 1061 | attn_metadata.rswa_window = self.rswa_window |
| 1062 | attn_metadata.rswa_window_tensor = self.persistent_rswa_window_tensor |
| 1063 | |
| 1064 | return attn_metadata |
| 1065 | |
| 1066 | def update_block_table( |
| 1067 | self, |
| 1068 | metadata: FlashAttentionMetadata, |
| 1069 | blk_table: torch.Tensor, |
| 1070 | slot_mapping: torch.Tensor, |
| 1071 | ) -> FlashAttentionMetadata: |
| 1072 | new_metadata = copy.copy(metadata) |
| 1073 | new_metadata.block_table = blk_table |
| 1074 | new_metadata.slot_mapping = slot_mapping |
| 1075 | return new_metadata |
| 1076 | |
| 1077 | def update_draft_decode_metadata(self, metadata: FlashAttentionMetadata) -> None: |
| 1078 | if metadata.scheduler_metadata is None: |
| 1079 | return |
| 1080 | |
| 1081 | num_reqs = metadata.num_decode_reqs or metadata.seq_lens.shape[0] |
| 1082 | |
| 1083 | assert self.dcp_world_size == 1 |
| 1084 | assert not metadata.use_cascade |
| 1085 | |
| 1086 | scheduler_metadata = self._get_scheduler_metadata( |
| 1087 | aot_schedule=True, |
| 1088 | batch_size=num_reqs, |
| 1089 | cu_query_lens=metadata.query_start_loc, |
| 1090 | max_query_len=metadata.max_query_len, |
| 1091 | seqlens=metadata.seq_lens, |
| 1092 | max_seq_len=metadata.max_seq_len, |
| 1093 | causal=metadata.causal, |
| 1094 | max_num_splits=metadata.max_num_splits, |
| 1095 | ) |
| 1096 | |
| 1097 | metadata.scheduler_metadata = self._store_scheduler_metadata(scheduler_metadata) |
| 1098 | |
| 1099 | def use_cascade_attention(self, *args, **kwargs) -> bool: |
| 1100 | if self.fa4_hd256: |
| 1101 | # Cascade may use a non-page-aligned prefix length. |
| 1102 | return False |
| 1103 | return use_cascade_attention(*args, **kwargs) |
| 1104 | |
| 1105 | |
| 1106 | class FlashAttentionImpl(AttentionImpl): |
| 1107 | can_return_lse_for_decode: bool = True |
| 1108 | supports_dcp: bool = True |
| 1109 | |
| 1110 | def __init__( |
| 1111 | self, |
| 1112 | num_heads: int, |
| 1113 | head_size: int, |
| 1114 | scale: float, |
| 1115 | num_kv_heads: int, |
| 1116 | alibi_slopes: list[float] | None, |
| 1117 | sliding_window: int | None, |
| 1118 | kv_cache_dtype: str, |
| 1119 | logits_soft_cap: float | None = None, |
| 1120 | attn_type: AttentionType = AttentionType.DECODER, |
| 1121 | kv_sharing_target_layer_name: str | None = None, |
| 1122 | sinks: torch.Tensor | None = None, |
| 1123 | ) -> None: |
| 1124 | self.num_heads = num_heads |
| 1125 | self.head_size = head_size |
| 1126 | self.scale = float(scale) |
| 1127 | self.num_kv_heads = num_kv_heads |
| 1128 | if alibi_slopes is not None: |
| 1129 | alibi_slopes = torch.tensor(alibi_slopes, dtype=torch.float32) |
| 1130 | self.alibi_slopes = alibi_slopes |
| 1131 | if sliding_window is None: |
| 1132 | self.sliding_window = (-1, -1) |
| 1133 | elif attn_type == AttentionType.ENCODER_ONLY: |
| 1134 | self.sliding_window = (sliding_window - 1, sliding_window - 1) |
| 1135 | else: |
| 1136 | self.sliding_window = (sliding_window - 1, 0) |
| 1137 | self.kv_cache_dtype = kv_cache_dtype |
| 1138 | if logits_soft_cap is None: |
| 1139 | # In flash-attn, setting logits_soft_cap as 0 means no soft cap. |
| 1140 | logits_soft_cap = 0 |
| 1141 | self.logits_soft_cap = logits_soft_cap |
| 1142 | self.kv_sharing_target_layer_name = kv_sharing_target_layer_name |
| 1143 | |
| 1144 | self.num_queries_per_kv = self.num_heads // self.num_kv_heads |
| 1145 | |
| 1146 | self.attn_type = attn_type |
| 1147 | vllm_config = get_current_vllm_config_or_none() |
| 1148 | uses_kv_cache = attn_type not in ( |
| 1149 | AttentionType.ENCODER, |
| 1150 | AttentionType.ENCODER_ONLY, |
| 1151 | ) |
| 1152 | # The final KV cache block size is unavailable during construction. |
| 1153 | self.vllm_flash_attn_version = get_flash_attn_version( |
| 1154 | requires_alibi=alibi_slopes is not None, |
| 1155 | head_size=head_size, |
| 1156 | has_sinks=sinks is not None, |
| 1157 | requires_softcap=bool(self.logits_soft_cap), |
| 1158 | supports_fa4_hd256=True, |
| 1159 | ) |
| 1160 | self.fa4_hd256 = self.vllm_flash_attn_version == 4 and uses_fa4_hd256_kernel( |
| 1161 | head_size |
| 1162 | ) |
| 1163 | if self.fa4_hd256 and not uses_kv_cache and sliding_window is not None: |
| 1164 | # The hd256 kernel requires seqused_k for local attention. |
| 1165 | logger.warning_once( |
| 1166 | "FA4's Blackwell head_size=256 kernel does not support local " |
| 1167 | "attention on encoder inputs, defaulting to FA version 2." |
| 1168 | ) |
| 1169 | self.vllm_flash_attn_version = 2 |
| 1170 | self.fa4_hd256 = False |
| 1171 | logger.info_once( |
| 1172 | "Using FlashAttention version %s", |
| 1173 | self.vllm_flash_attn_version, |
| 1174 | ) |
| 1175 | # Cache the batch invariant result for use in forward passes |
| 1176 | self.batch_invariant_enabled = envs.VLLM_BATCH_INVARIANT |
| 1177 | |
| 1178 | if is_quantized_kv_cache( |
| 1179 | self.kv_cache_dtype |
| 1180 | ) and not flash_attn_supports_kv_cache_dtype( |
| 1181 | self.kv_cache_dtype, |
| 1182 | requires_alibi=alibi_slopes is not None, |
| 1183 | head_size=head_size, |
| 1184 | head_size_v=head_size, |
| 1185 | has_sinks=sinks is not None, |
| 1186 | requires_softcap=bool(self.logits_soft_cap), |
| 1187 | supports_fa4_hd256=True, |
| 1188 | ): |
| 1189 | raise NotImplementedError( |
| 1190 | f"FlashAttention does not support {self.kv_cache_dtype}" |
| 1191 | " kv-cache on this device." |
| 1192 | ) |
| 1193 | |
| 1194 | self.sinks = sinks |
| 1195 | if self.sinks is not None: |
| 1196 | assert flash_attn_supports_sinks(), ( |
| 1197 | "Sinks are only supported in FlashAttention 3" |
| 1198 | ) |
| 1199 | assert self.sinks.shape[0] == num_heads, ( |
| 1200 | "Sinks must have the same number of heads as the number of " |
| 1201 | "heads in the layer" |
| 1202 | ) |
| 1203 | |
| 1204 | # FA4's SM90 FP8-KV path consumes native FP16/BF16 Q and dequantizes |
| 1205 | # FP8 K/V in-kernel. Other FA4 paths (notably SM100) still require Q, |
| 1206 | # K, and V to have the same FP8 dtype. |
| 1207 | uses_sm90_fa4_fp8_kv_dequant = ( |
| 1208 | self.vllm_flash_attn_version == 4 |
| 1209 | and current_platform.is_device_capability_family(90) |
| 1210 | and self.kv_cache_dtype in ("fp8", "fp8_e4m3") |
| 1211 | ) |
| 1212 | self.supports_quant_query_input = flash_attn_supports_quant_query_input() and ( |
| 1213 | not uses_sm90_fa4_fp8_kv_dequant |
| 1214 | ) |
| 1215 | |
| 1216 | dcp_a2a = ( |
| 1217 | vllm_config is not None |
| 1218 | and vllm_config.parallel_config.decode_context_parallel_size > 1 |
| 1219 | and vllm_config.parallel_config.dcp_comm_backend == "a2a" |
| 1220 | ) |
| 1221 | self.dcp_combine = dcp_a2a_lse_reduce if dcp_a2a else cp_lse_ag_out_rs |
| 1222 | |
| 1223 | self._dcp_dtype: torch.dtype | None = None |
| 1224 | self._dcp_max_num_tokens: int = 0 |
| 1225 | if vllm_config is not None and self.dcp_world_size > 1: |
| 1226 | self._dcp_dtype = vllm_config.model_config.dtype |
| 1227 | self._dcp_max_num_tokens = ( |
| 1228 | vllm_config.scheduler_config.max_num_batched_tokens |
| 1229 | ) |
| 1230 | |
| 1231 | def forward( |
| 1232 | self, |
| 1233 | layer: torch.nn.Module, |
| 1234 | query: torch.Tensor, |
| 1235 | key: torch.Tensor, |
| 1236 | value: torch.Tensor, |
| 1237 | kv_cache: torch.Tensor, |
| 1238 | attn_metadata: FlashAttentionMetadata, |
| 1239 | output: torch.Tensor, |
| 1240 | output_scale: torch.Tensor | None = None, |
| 1241 | output_block_scale: torch.Tensor | None = None, |
| 1242 | ) -> torch.Tensor: |
| 1243 | """Forward pass with FlashAttention. |
| 1244 | |
| 1245 | Args: |
| 1246 | layer: The attention layer, providing the q/k/v quantization scales. |
| 1247 | query: shape = [num_tokens, num_heads, head_size] |
| 1248 | key: shape = [num_tokens, num_kv_heads, head_size] |
| 1249 | value: shape = [num_tokens, num_kv_heads, head_size] |
| 1250 | kv_cache: shape = |
| 1251 | [num_blocks, num_kv_heads, block_size, 2 * head_size] |
| 1252 | attn_metadata: Metadata for attention. |
| 1253 | output: Tensor that the attention result is written into. |
| 1254 | output_scale: Scale for fused output quantization; not supported |
| 1255 | by this backend. |
| 1256 | output_block_scale: Block scale for fused output quantization; |
| 1257 | not supported by this backend. |
| 1258 | |
| 1259 | Returns: |
| 1260 | shape = [num_tokens, num_heads * head_size] |
| 1261 | NOTE: FP8 quantization, flash-attn expect the size of |
| 1262 | {q,k,v}_descale to be (num_sequences, num_kv_heads). |
| 1263 | We use torch's .expand() to avoid duplicating values |
| 1264 | |
| 1265 | """ |
| 1266 | assert self.vllm_flash_attn_version is not None, ( |
| 1267 | "FlashAttention version not detected." |
| 1268 | ) |
| 1269 | |
| 1270 | if output_scale is not None or output_block_scale is not None: |
| 1271 | raise NotImplementedError( |
| 1272 | "fused output quantization is not yet supported for FlashAttentionImpl" |
| 1273 | ) |
| 1274 | |
| 1275 | if attn_metadata is None: |
| 1276 | # Profiling run. |
| 1277 | return output.fill_(0) |
| 1278 | |
| 1279 | attn_type = self.attn_type |
| 1280 | |
| 1281 | # IMPORTANT! |
| 1282 | # NOTE(woosuk): With piece-wise CUDA graphs, this method is executed in |
| 1283 | # eager-mode PyTorch. Thus, we need to be careful about any CPU overhead |
| 1284 | # in this method. For example, `view` and `slice` (or `[:n]`) operations |
| 1285 | # are surprisingly slow even in the case they do not invoke any GPU ops. |
| 1286 | # Minimize the PyTorch ops in this method as much as possible. |
| 1287 | # Whenever making a change in this method, please benchmark the |
| 1288 | # performance to make sure it does not introduce any overhead. |
| 1289 | |
| 1290 | num_actual_tokens = attn_metadata.num_actual_tokens |
| 1291 | |
| 1292 | # Handle encoder attention differently - no KV cache needed |
| 1293 | if attn_type in (AttentionType.ENCODER_ONLY, AttentionType.ENCODER): |
| 1294 | # For encoder attention, |
| 1295 | # we use direct Q, K, V tensors without caching |
| 1296 | return self._forward_encoder_attention( |
| 1297 | query[:num_actual_tokens], |
| 1298 | key[:num_actual_tokens], |
| 1299 | value[:num_actual_tokens], |
| 1300 | output[:num_actual_tokens], |
| 1301 | attn_metadata, |
| 1302 | layer, |
| 1303 | ) |
| 1304 | |
| 1305 | # (B, H, N, 2*D) -> ((B, N, H, D), (B, N, H, D)) |
| 1306 | key_cache, value_cache = kv_cache.transpose(1, 2).split(self.head_size, dim=-1) |
| 1307 | # Fix degenerate strides on size-1 dims (e.g. num_kv_heads=1 with TP). |
| 1308 | # FA3/4 on H100+ uses TMA, which requires ≥16-byte stride alignment. |
| 1309 | # See vllm.utils.torch_utils.canonicalize_singleton_dim_strides. |
| 1310 | fixed_k = canonicalize_singleton_dim_strides(key_cache) |
| 1311 | fixed_v = canonicalize_singleton_dim_strides(value_cache) |
| 1312 | if fixed_k is not key_cache or fixed_v is not value_cache: |
| 1313 | logger.debug( |
| 1314 | "Canonicalized degenerate KV cache strides (FlashAttention): " |
| 1315 | "shape=%s, key strides before=%s after=%s, " |
| 1316 | "value strides before=%s after=%s", |
| 1317 | key_cache.shape, |
| 1318 | key_cache.stride(), |
| 1319 | fixed_k.stride(), |
| 1320 | value_cache.stride(), |
| 1321 | fixed_v.stride(), |
| 1322 | ) |
| 1323 | key_cache, value_cache = fixed_k, fixed_v |
| 1324 | |
| 1325 | if is_quantized_kv_cache(self.kv_cache_dtype): |
| 1326 | # queries are quantized in the attention layer |
| 1327 | key_cache = key_cache.view(current_platform.fp8_dtype()) |
| 1328 | value_cache = value_cache.view(current_platform.fp8_dtype()) |
| 1329 | |
| 1330 | if not attn_metadata.use_cascade: |
| 1331 | cu_seqlens_q = attn_metadata.query_start_loc |
| 1332 | seqused_k = attn_metadata.seq_lens |
| 1333 | max_seqlen_q = attn_metadata.max_query_len |
| 1334 | max_seqlen_k = attn_metadata.max_seq_len |
| 1335 | block_table = attn_metadata.block_table |
| 1336 | scheduler_metadata = attn_metadata.scheduler_metadata |
| 1337 | |
| 1338 | descale_shape = (cu_seqlens_q.shape[0] - 1, self.num_kv_heads) |
| 1339 | |
| 1340 | q_descale = ( |
| 1341 | layer._q_scale.expand(descale_shape) |
| 1342 | if self.supports_quant_query_input |
| 1343 | else None |
| 1344 | ) |
| 1345 | k_descale = layer._k_scale.expand(descale_shape) |
| 1346 | v_descale = layer._v_scale.expand(descale_shape) |
| 1347 | |
| 1348 | if self.dcp_world_size > 1: |
| 1349 | self._forward_with_dcp( |
| 1350 | query[:num_actual_tokens], |
| 1351 | key[:num_actual_tokens], |
| 1352 | value[:num_actual_tokens], |
| 1353 | key_cache, |
| 1354 | value_cache, |
| 1355 | output[:num_actual_tokens], |
| 1356 | attn_metadata, |
| 1357 | q_descale=q_descale, |
| 1358 | k_descale=k_descale, |
| 1359 | v_descale=v_descale, |
| 1360 | ) |
| 1361 | return output |
| 1362 | else: |
| 1363 | causal = attn_metadata.causal |
| 1364 | is_dynamic_causal = isinstance(causal, torch.Tensor) |
| 1365 | |
| 1366 | # The layer's own window wins over the group's: one KV cache |
| 1367 | # group can hold both windowed and global layers (e.g. Gemma-3 |
| 1368 | # with the hybrid KV cache manager disabled), and the group spec |
| 1369 | # cannot describe both. |
| 1370 | window = _maybe_symmetrize_window(self.sliding_window, causal) |
| 1371 | sliding_window_size: list[int] | None = ( |
| 1372 | list(window) if window is not None else None |
| 1373 | ) |
| 1374 | |
| 1375 | mm_prefix_query_ranges = attn_metadata.mm_prefix_query_range_tensor |
| 1376 | mm_mask_mod = None |
| 1377 | mm_aux = None |
| 1378 | if ( |
| 1379 | mm_prefix_query_ranges is not None |
| 1380 | and not is_dynamic_causal |
| 1381 | and causal is True |
| 1382 | and self.vllm_flash_attn_version == 4 |
| 1383 | ): |
| 1384 | # Triton convention: 1 + window_size[0]. Global layers store |
| 1385 | # (-1, -1) → sw stays None. |
| 1386 | layer_window = self.sliding_window |
| 1387 | sw_val = ( |
| 1388 | 1 + layer_window[0] |
| 1389 | if layer_window is not None and layer_window[0] >= 0 |
| 1390 | else None |
| 1391 | ) |
| 1392 | # Gemma4: also clamp the bidirectional block to the |
| 1393 | # sliding window when the layer opts in |
| 1394 | # (mm_prefix_clamp_sliding_window flag from PR #47217). |
| 1395 | mm_clamp_sw = 0 |
| 1396 | if ( |
| 1397 | getattr(layer, "mm_prefix_clamp_sliding_window", False) |
| 1398 | and sw_val is not None |
| 1399 | ): |
| 1400 | mm_clamp_sw = sw_val |
| 1401 | mm_mask_mod = _make_mm_prefix_mask_mod( |
| 1402 | sliding_window=mm_clamp_sw, |
| 1403 | sliding_window_left=sw_val, |
| 1404 | ) |
| 1405 | mm_aux = [mm_prefix_query_ranges, attn_metadata.query_start_loc] |
| 1406 | # mm_prefix is (causal ∧ window) ∨ bidirectional-range — |
| 1407 | # not ⊆ causal. FA #155 stopped auto-clearing causal/local |
| 1408 | # when mask_mod is set, so the caller must disable them or |
| 1409 | # the built-in causal path shorts out / clips the mask_mod. |
| 1410 | causal = False |
| 1411 | sliding_window_size = None |
| 1412 | |
| 1413 | # R-SWA: use CuTE-DSL mask_mod on FA4 for exact token-level |
| 1414 | # mask without block-size approximation. The mask_mod encodes |
| 1415 | # "causal AND (kv < prefix_len OR q - kv < rswa_window)", which |
| 1416 | # supersedes any FA-layer sliding_window_size parameter. |
| 1417 | rswa_mask_mod_fn = None |
| 1418 | rswa_aux = None |
| 1419 | if ( |
| 1420 | attn_metadata.rswa_prefix_lens is not None |
| 1421 | and self.vllm_flash_attn_version == 4 |
| 1422 | and not is_dynamic_causal |
| 1423 | ): |
| 1424 | rswa_mask_mod_fn = _make_rswa_mask_mod() |
| 1425 | rswa_aux = [ |
| 1426 | attn_metadata.rswa_prefix_lens.to(torch.int32), |
| 1427 | attn_metadata.rswa_window_tensor, # pre-allocated CUDA tensor |
| 1428 | ] |
| 1429 | # mask_mod fully expresses R-SWA; disable FA's own window. |
| 1430 | sliding_window_size = None |
| 1431 | |
| 1432 | dynamic_causal = None |
| 1433 | if isinstance(causal, torch.Tensor): |
| 1434 | if self.vllm_flash_attn_version != 4: |
| 1435 | raise NotImplementedError( |
| 1436 | "Per-sequence causal requires FA4. Current version: " |
| 1437 | f"FA{self.vllm_flash_attn_version}" |
| 1438 | ) |
| 1439 | dynamic_causal = causal |
| 1440 | has_window = ( |
| 1441 | sliding_window_size is not None and sliding_window_size[1] >= 0 |
| 1442 | ) |
| 1443 | causal = not has_window |
| 1444 | |
| 1445 | num_splits = attn_metadata.max_num_splits |
| 1446 | if self.fa4_hd256: |
| 1447 | # hd256 requires page-aligned lengths, exact-width block |
| 1448 | # tables, and no SplitKV. |
| 1449 | num_pages = cdiv(max_seqlen_k, FA4_HD256_PAGE_SIZE) |
| 1450 | max_seqlen_k = num_pages * FA4_HD256_PAGE_SIZE |
| 1451 | block_table = block_table[:, :num_pages] |
| 1452 | num_splits = 1 |
| 1453 | |
| 1454 | _FA4_DENSE_ATTENTION_KERNEL( |
| 1455 | q=query[:num_actual_tokens], |
| 1456 | k=key_cache, |
| 1457 | v=value_cache, |
| 1458 | out=output[:num_actual_tokens], |
| 1459 | cu_seqlens_q=cu_seqlens_q, |
| 1460 | max_seqlen_q=max_seqlen_q, |
| 1461 | seqused_k=seqused_k, |
| 1462 | max_seqlen_k=max_seqlen_k, |
| 1463 | softmax_scale=self.scale, |
| 1464 | causal=causal, |
| 1465 | alibi_slopes=self.alibi_slopes, |
| 1466 | window_size=sliding_window_size, |
| 1467 | block_table=block_table, |
| 1468 | softcap=self.logits_soft_cap, |
| 1469 | scheduler_metadata=scheduler_metadata, |
| 1470 | fa_version=self.vllm_flash_attn_version, |
| 1471 | q_descale=q_descale, |
| 1472 | k_descale=k_descale, |
| 1473 | v_descale=v_descale, |
| 1474 | dynamic_causal=dynamic_causal, |
| 1475 | num_splits=num_splits, |
| 1476 | s_aux=self.sinks, |
| 1477 | mask_mod=rswa_mask_mod_fn or mm_mask_mod, |
| 1478 | aux_tensors=rswa_aux or mm_aux, |
| 1479 | ) |
| 1480 | return output |
| 1481 | |
| 1482 | # Cascade attention (rare case). |
| 1483 | cascade_attention( |
| 1484 | output[:num_actual_tokens], |
| 1485 | query[:num_actual_tokens], |
| 1486 | key_cache, |
| 1487 | value_cache, |
| 1488 | cu_query_lens=attn_metadata.query_start_loc, |
| 1489 | max_query_len=attn_metadata.max_query_len, |
| 1490 | cu_prefix_query_lens=attn_metadata.cu_prefix_query_lens, |
| 1491 | prefix_kv_lens=attn_metadata.prefix_kv_lens, |
| 1492 | suffix_kv_lens=attn_metadata.suffix_kv_lens, |
| 1493 | max_kv_len=attn_metadata.max_seq_len, |
| 1494 | softmax_scale=self.scale, |
| 1495 | alibi_slopes=self.alibi_slopes, |
| 1496 | sliding_window=self.sliding_window, |
| 1497 | logits_soft_cap=self.logits_soft_cap, |
| 1498 | block_table=attn_metadata.block_table, |
| 1499 | common_prefix_len=attn_metadata.common_prefix_len, |
| 1500 | max_num_splits=attn_metadata.max_num_splits, |
| 1501 | fa_version=self.vllm_flash_attn_version, |
| 1502 | prefix_scheduler_metadata=attn_metadata.prefix_scheduler_metadata, |
| 1503 | suffix_scheduler_metadata=attn_metadata.scheduler_metadata, |
| 1504 | q_descale=layer._q_scale, |
| 1505 | k_descale=layer._k_scale, |
| 1506 | v_descale=layer._v_scale, |
| 1507 | s_aux=self.sinks, |
| 1508 | ) |
| 1509 | return output |
| 1510 | |
| 1511 | def do_kv_cache_update( |
| 1512 | self, |
| 1513 | layer: torch.nn.Module, |
| 1514 | key: torch.Tensor, |
| 1515 | value: torch.Tensor, |
| 1516 | kv_cache: torch.Tensor, |
| 1517 | slot_mapping: torch.Tensor, |
| 1518 | ) -> None: |
| 1519 | if self.attn_type in (AttentionType.ENCODER_ONLY, AttentionType.ENCODER): |
| 1520 | # For encoder attention, |
| 1521 | # we use direct Q, K, V tensors without caching |
| 1522 | return |
| 1523 | |
| 1524 | # Scatter write into the KV cache using slot_mapping indices. |
| 1525 | # No TMA kernel is invoked here, so stride canonicalization is not needed. |
| 1526 | # (B, H, N, 2*D) -> ((B, N, H, D), (B, N, H, D)) |
| 1527 | key_cache, value_cache = kv_cache.transpose(1, 2).split(self.head_size, dim=-1) |
| 1528 | |
| 1529 | # Reshape the input keys and values and store them in the cache. |
| 1530 | # Skip this if sharing KV cache with an earlier attention layer. |
| 1531 | # NOTE(woosuk): Here, key and value are padded while slot_mapping is |
| 1532 | # not padded. However, we don't need to do key[:num_actual_tokens] |
| 1533 | # and value[:num_actual_tokens] because the reshape_and_cache_flash |
| 1534 | # op uses the slot_mapping's shape to determine the number of |
| 1535 | # actual tokens. |
| 1536 | reshape_and_cache_flash( |
| 1537 | key, |
| 1538 | value, |
| 1539 | key_cache, |
| 1540 | value_cache, |
| 1541 | slot_mapping, |
| 1542 | self.kv_cache_dtype, |
| 1543 | layer._k_scale, |
| 1544 | layer._v_scale, |
| 1545 | ) |
| 1546 | |
| 1547 | def _forward_with_dcp( |
| 1548 | self, |
| 1549 | query: torch.Tensor, |
| 1550 | key: torch.Tensor, |
| 1551 | value: torch.Tensor, |
| 1552 | key_cache: torch.Tensor, |
| 1553 | value_cache: torch.Tensor, |
| 1554 | output: torch.Tensor, |
| 1555 | attn_metadata: FlashAttentionMetadata, |
| 1556 | q_descale: torch.Tensor | None = None, |
| 1557 | k_descale: torch.Tensor | None = None, |
| 1558 | v_descale: torch.Tensor | None = None, |
| 1559 | ) -> torch.Tensor: |
| 1560 | assert self.vllm_flash_attn_version is not None, ( |
| 1561 | "FlashAttention version not detected." |
| 1562 | ) |
| 1563 | |
| 1564 | cu_seqlens_q = attn_metadata.query_start_loc |
| 1565 | max_seqlen_q = attn_metadata.max_query_len |
| 1566 | block_table = attn_metadata.block_table |
| 1567 | |
| 1568 | query = query.contiguous() |
| 1569 | if attn_metadata.max_dcp_context_kv_len == 0: |
| 1570 | flash_attn_varlen_func( |
| 1571 | q=query, |
| 1572 | k=key, |
| 1573 | v=value, |
| 1574 | out=output, |
| 1575 | cu_seqlens_q=cu_seqlens_q, |
| 1576 | max_seqlen_q=max_seqlen_q, |
| 1577 | cu_seqlens_k=cu_seqlens_q, |
| 1578 | max_seqlen_k=max_seqlen_q, |
| 1579 | softmax_scale=self.scale, |
| 1580 | causal=attn_metadata.causal, |
| 1581 | alibi_slopes=self.alibi_slopes, |
| 1582 | window_size=list(self.sliding_window) |
| 1583 | if self.sliding_window is not None |
| 1584 | else None, |
| 1585 | softcap=self.logits_soft_cap, |
| 1586 | return_softmax_lse=True, |
| 1587 | fa_version=self.vllm_flash_attn_version, |
| 1588 | q_descale=q_descale, |
| 1589 | k_descale=k_descale, |
| 1590 | v_descale=v_descale, |
| 1591 | num_splits=attn_metadata.max_num_splits, |
| 1592 | ) |
| 1593 | return output |
| 1594 | |
| 1595 | query_across_dcp = get_dcp_group().all_gather(query, dim=1) |
| 1596 | sliding_window_size = ( |
| 1597 | list(self.sliding_window) if self.sliding_window is not None else None |
| 1598 | ) |
| 1599 | n = query_across_dcp.shape[0] |
| 1600 | num_reqs = cu_seqlens_q.shape[0] - 1 |
| 1601 | num_decodes = attn_metadata.num_decode_reqs |
| 1602 | num_context_prefills = attn_metadata.num_prefill_reqs |
| 1603 | num_decode_tokens = attn_metadata.num_decode_tokens |
| 1604 | num_context_prefill_tokens = attn_metadata.num_prefill_tokens |
| 1605 | split_dcp_context = should_split_fa2_dcp_context_attention( |
| 1606 | self.vllm_flash_attn_version, |
| 1607 | max_seqlen_q, |
| 1608 | num_reqs, |
| 1609 | num_decodes, |
| 1610 | num_context_prefills, |
| 1611 | ) |
| 1612 | dcp_context_out_tokens = max(n, self._dcp_max_num_tokens) |
| 1613 | dcp_context_out_spec = ( |
| 1614 | ( |
| 1615 | dcp_context_out_tokens, |
| 1616 | self.num_heads * self.dcp_world_size, |
| 1617 | self.head_size, |
| 1618 | ), |
| 1619 | self._dcp_dtype, |
| 1620 | ) |
| 1621 | (dcp_context_out_workspace,) = current_workspace_manager().get_simultaneous( |
| 1622 | dcp_context_out_spec, |
| 1623 | ) |
| 1624 | dcp_context_out = dcp_context_out_workspace[:n] |
| 1625 | |
| 1626 | if split_dcp_context: |
| 1627 | # TODO: Remove this DCP + FA2 mixed decode/prefill workaround once |
| 1628 | # FA4 supports this Qwen3.5 shape. |
| 1629 | assert attn_metadata.dcp_context_kv_lens is not None |
| 1630 | assert attn_metadata.max_dcp_context_kv_len is not None |
| 1631 | assert self.vllm_flash_attn_version is not None |
| 1632 | context_attn_out, context_lse = run_split_fa2_dcp_context_attention( |
| 1633 | flash_attn_varlen_func, |
| 1634 | query_across_dcp, |
| 1635 | key_cache, |
| 1636 | value_cache, |
| 1637 | dcp_context_out, |
| 1638 | cu_seqlens_q, |
| 1639 | max_seqlen_q, |
| 1640 | attn_metadata.dcp_context_kv_lens, |
| 1641 | attn_metadata.max_dcp_context_kv_len, |
| 1642 | self.scale, |
| 1643 | self.alibi_slopes, |
| 1644 | sliding_window_size, |
| 1645 | block_table, |
| 1646 | self.logits_soft_cap, |
| 1647 | self.vllm_flash_attn_version, |
| 1648 | q_descale, |
| 1649 | k_descale, |
| 1650 | v_descale, |
| 1651 | attn_metadata.max_num_splits, |
| 1652 | self.num_heads, |
| 1653 | self.dcp_world_size, |
| 1654 | num_decodes, |
| 1655 | num_context_prefills, |
| 1656 | num_decode_tokens, |
| 1657 | num_context_prefill_tokens, |
| 1658 | ) |
| 1659 | else: |
| 1660 | context_attn_out, context_lse = flash_attn_varlen_func( |
| 1661 | q=query_across_dcp, |
| 1662 | k=key_cache, |
| 1663 | v=value_cache, |
| 1664 | out=dcp_context_out, |
| 1665 | cu_seqlens_q=cu_seqlens_q, |
| 1666 | max_seqlen_q=max_seqlen_q, |
| 1667 | seqused_k=attn_metadata.dcp_context_kv_lens, |
| 1668 | max_seqlen_k=attn_metadata.max_dcp_context_kv_len, |
| 1669 | softmax_scale=self.scale, |
| 1670 | causal=False, |
| 1671 | alibi_slopes=self.alibi_slopes, |
| 1672 | window_size=sliding_window_size, |
| 1673 | block_table=block_table, |
| 1674 | softcap=self.logits_soft_cap, |
| 1675 | return_softmax_lse=True, |
| 1676 | scheduler_metadata=attn_metadata.scheduler_metadata, |
| 1677 | fa_version=self.vllm_flash_attn_version, |
| 1678 | q_descale=q_descale, |
| 1679 | k_descale=k_descale, |
| 1680 | v_descale=v_descale, |
| 1681 | num_splits=attn_metadata.max_num_splits, |
| 1682 | ) |
| 1683 | # FA returns LSE in shape [ H, B ] but DCP combine wants [ B, H ] |
| 1684 | context_attn_out_cor, context_lse_cor = self.dcp_combine( |
| 1685 | context_attn_out, |
| 1686 | context_lse.transpose(0, 1), |
| 1687 | get_dcp_group(), |
| 1688 | return_lse=True, |
| 1689 | ) |
| 1690 | context_lse_cor = context_lse_cor.transpose(0, 1).contiguous() |
| 1691 | |
| 1692 | query_attn_out, query_lse = flash_attn_varlen_func( |
| 1693 | q=query, |
| 1694 | k=key, |
| 1695 | v=value, |
| 1696 | out=output, |
| 1697 | cu_seqlens_q=cu_seqlens_q, |
| 1698 | max_seqlen_q=max_seqlen_q, |
| 1699 | cu_seqlens_k=cu_seqlens_q, |
| 1700 | max_seqlen_k=max_seqlen_q, |
| 1701 | softmax_scale=self.scale, |
| 1702 | causal=attn_metadata.causal, |
| 1703 | alibi_slopes=self.alibi_slopes, |
| 1704 | window_size=sliding_window_size, |
| 1705 | softcap=self.logits_soft_cap, |
| 1706 | return_softmax_lse=True, |
| 1707 | fa_version=self.vllm_flash_attn_version, |
| 1708 | q_descale=q_descale, |
| 1709 | k_descale=k_descale, |
| 1710 | v_descale=v_descale, |
| 1711 | num_splits=attn_metadata.max_num_splits, |
| 1712 | ) |
| 1713 | assert context_attn_out_cor.shape == query_attn_out.shape |
| 1714 | assert context_lse_cor.shape == query_lse.shape |
| 1715 | merge_attn_states( |
| 1716 | output, |
| 1717 | context_attn_out_cor, |
| 1718 | context_lse_cor, |
| 1719 | query_attn_out, |
| 1720 | query_lse, |
| 1721 | ) |
| 1722 | |
| 1723 | def _forward_encoder_attention( |
| 1724 | self, |
| 1725 | query: torch.Tensor, |
| 1726 | key: torch.Tensor, |
| 1727 | value: torch.Tensor, |
| 1728 | output: torch.Tensor, |
| 1729 | attn_metadata: FlashAttentionMetadata, |
| 1730 | layer: torch.nn.Module, |
| 1731 | ) -> torch.Tensor: |
| 1732 | """Forward pass for encoder attention without KV cache. |
| 1733 | |
| 1734 | Args: |
| 1735 | query: shape = [num_encoder_tokens, num_heads, head_size] |
| 1736 | key: shape = [num_encoder_tokens, num_kv_heads, head_size] |
| 1737 | value: shape = [num_encoder_tokens, num_kv_heads, head_size] |
| 1738 | output: shape = [num_encoder_tokens, num_heads, head_size] |
| 1739 | attn_metadata: Encoder attention metadata |
| 1740 | layer: The attention layer |
| 1741 | |
| 1742 | """ |
| 1743 | assert self.vllm_flash_attn_version is not None, ( |
| 1744 | "FlashAttention version not detected." |
| 1745 | ) |
| 1746 | |
| 1747 | # For encoder attention, process FP8 quantization if needed |
| 1748 | if is_quantized_kv_cache(self.kv_cache_dtype): |
| 1749 | raise NotImplementedError( |
| 1750 | "quantization is not supported for encoder attention" |
| 1751 | ) |
| 1752 | |
| 1753 | # Use encoder-specific metadata for sequence information |
| 1754 | cu_seqlens_q = attn_metadata.query_start_loc |
| 1755 | cu_seqlens_k = attn_metadata.query_start_loc |
| 1756 | max_seqlen_q = attn_metadata.max_query_len |
| 1757 | max_seqlen_k = attn_metadata.max_query_len |
| 1758 | |
| 1759 | descale_shape = ( |
| 1760 | cu_seqlens_q.shape[0] - 1, # type: ignore[union-attr] |
| 1761 | self.num_kv_heads, |
| 1762 | ) |
| 1763 | |
| 1764 | # Call flash attention directly on Q, K, V tensors |
| 1765 | sliding_window_size = ( |
| 1766 | list(self.sliding_window) if self.sliding_window is not None else None |
| 1767 | ) |
| 1768 | flash_attn_varlen_func( |
| 1769 | q=query, |
| 1770 | k=key, |
| 1771 | v=value, |
| 1772 | out=output, |
| 1773 | cu_seqlens_q=cu_seqlens_q, |
| 1774 | cu_seqlens_k=cu_seqlens_k, |
| 1775 | max_seqlen_q=max_seqlen_q, |
| 1776 | max_seqlen_k=max_seqlen_k, |
| 1777 | softmax_scale=self.scale, |
| 1778 | causal=False, # Encoder attention is bidirectional |
| 1779 | alibi_slopes=self.alibi_slopes, |
| 1780 | window_size=sliding_window_size, |
| 1781 | softcap=self.logits_soft_cap, |
| 1782 | fa_version=self.vllm_flash_attn_version, |
| 1783 | q_descale=layer._q_scale.expand(descale_shape) # type: ignore[operator] |
| 1784 | if self.supports_quant_query_input |
| 1785 | else None, |
| 1786 | k_descale=layer._k_scale.expand(descale_shape), # type: ignore[operator] |
| 1787 | v_descale=layer._v_scale.expand(descale_shape), # type: ignore[operator] |
| 1788 | # The hd256 kernel does not support SplitKV. |
| 1789 | num_splits=1 if self.batch_invariant_enabled or self.fa4_hd256 else 0, |
| 1790 | s_aux=self.sinks, |
| 1791 | ) |
| 1792 | |
| 1793 | return output |
| 1794 | |
| 1795 | |
| 1796 | @functools.cache |
| 1797 | def _make_mm_prefix_mask_mod( |
| 1798 | sliding_window: int = 0, |
| 1799 | sliding_window_left: int | None = None, |
| 1800 | ): |
| 1801 | """Build a CuTE-DSL mask_mod implementing |
| 1802 | ``(causal AND sliding_window) OR mm_prefix``. |
| 1803 | |
| 1804 | Cached so identical ``(sliding_window, sliding_window_left)`` reuse the |
| 1805 | same function object. FA4's ``hash_callable`` mixes ``repr()`` of closure |
| 1806 | cells into the compile key; the nested ``_load_q_range`` would otherwise |
| 1807 | get a new address each call and force a full JIT recompile every forward. |
| 1808 | |
| 1809 | The FA4 kernel passes *local* ``q_idx`` (0-based within the current |
| 1810 | prefill chunk) while ``kv_idx`` is absolute (0-based over the full |
| 1811 | KV cache). We recover the absolute Q position via |
| 1812 | ``q_abs = q_idx + seqlen_k - seqlen_q`` (the context-length offset) |
| 1813 | so that causal, sliding-window, and mm_prefix range comparisons all |
| 1814 | use consistent absolute positions. This matches the Triton |
| 1815 | reference path (``compute_kv_seq_mask``). |
| 1816 | |
| 1817 | ``aux_tensors[0]`` holds the absolute ``[start, end]`` bounds of the |
| 1818 | mm_prefix range containing each scheduled query token (``(-1, -1)`` when |
| 1819 | none), and ``aux_tensors[1]`` is ``cu_seqlens_q``, used to turn the local |
| 1820 | ``q_idx`` into a packed row index. Because mm_prefix ranges never overlap, |
| 1821 | ``r_start <= kv_idx <= r_end`` is exactly "query and key share a range", so |
| 1822 | no key-side lookup is needed and ``kv_idx`` is never used as an index. |
| 1823 | The ``(-1, -1)`` sentinel falls out for free: ``kv_idx <= -1`` is false for |
| 1824 | every valid key. |
| 1825 | |
| 1826 | ``sliding_window_left`` enforces the sliding window on the causal |
| 1827 | term (None = full causal, no window). ``sliding_window`` clamps the |
| 1828 | bidirectional block to the window (0 = unclamped; >0 = Gemma4 local |
| 1829 | layers via ``mm_prefix_clamp_sliding_window``). |
| 1830 | """ |
| 1831 | import cutlass |
| 1832 | import cutlass.cute as cute |
| 1833 | from cutlass import Int32 # type: ignore[attr-defined] |
| 1834 | |
| 1835 | from vllm.vllm_flash_attn.cute.utils import ( # type: ignore[import-untyped] |
| 1836 | scalar_to_ssa, |
| 1837 | ssa_to_scalar, |
| 1838 | ) |
| 1839 | |
| 1840 | @cute.jit |
| 1841 | def _load_q_range(q_idx, seqlen_info, aux_tensors, batch_idx): |
| 1842 | """Load the mm_prefix range bounds for this query row. |
| 1843 | |
| 1844 | Both loads depend only on the query index, so they hoist out of the |
| 1845 | unrolled per-element mask loop. ``ssa_to_scalar`` reads lane 0, so one |
| 1846 | call must not span query rows (hence the ``__vec_size__`` pin below). |
| 1847 | Clamping keeps ``token_idx`` in bounds for partial tiles and padded |
| 1848 | (``seqlen_q == 0``) rows; neither produces output. |
| 1849 | """ |
| 1850 | q_ranges = aux_tensors[0] |
| 1851 | cu_seqlens_q = aux_tensors[1] |
| 1852 | b = batch_idx[0] |
| 1853 | q_local = cutlass.min(ssa_to_scalar(q_idx), seqlen_info.seqlen_q - Int32(1)) |
| 1854 | token_idx = cutlass.max(cu_seqlens_q[b] + q_local, Int32(0)) |
| 1855 | return ( |
| 1856 | scalar_to_ssa(q_ranges[token_idx, 0], Int32), |
| 1857 | scalar_to_ssa(q_ranges[token_idx, 1], Int32), |
| 1858 | ) |
| 1859 | |
| 1860 | if sliding_window_left is not None: |
| 1861 | |
| 1862 | @cute.jit |
| 1863 | def mm_prefix_mask_mod( |
| 1864 | batch_idx: cute.TensorSSA, |
| 1865 | head_idx: cute.TensorSSA, |
| 1866 | q_idx: cute.TensorSSA, |
| 1867 | kv_idx: cute.TensorSSA, |
| 1868 | seqlen_info, |
| 1869 | aux_tensors, |
| 1870 | ): |
| 1871 | ctx_off = scalar_to_ssa(seqlen_info.seqlen_k - seqlen_info.seqlen_q, Int32) |
| 1872 | q_abs = q_idx + ctx_off |
| 1873 | sw = scalar_to_ssa(Int32(sliding_window_left), Int32) |
| 1874 | keep = (kv_idx <= q_abs) & ((q_abs - kv_idx) < sw) |
| 1875 | r_start, r_end = _load_q_range(q_idx, seqlen_info, aux_tensors, batch_idx) |
| 1876 | mm = (kv_idx >= r_start) & (kv_idx <= r_end) |
| 1877 | if sliding_window > 0: |
| 1878 | mm = mm & ((q_abs - kv_idx) < sw) |
| 1879 | keep = keep | mm |
| 1880 | return keep |
| 1881 | |
| 1882 | else: |
| 1883 | |
| 1884 | @cute.jit |
| 1885 | def mm_prefix_mask_mod( |
| 1886 | batch_idx: cute.TensorSSA, |
| 1887 | head_idx: cute.TensorSSA, |
| 1888 | q_idx: cute.TensorSSA, |
| 1889 | kv_idx: cute.TensorSSA, |
| 1890 | seqlen_info, |
| 1891 | aux_tensors, |
| 1892 | ): |
| 1893 | ctx_off = scalar_to_ssa(seqlen_info.seqlen_k - seqlen_info.seqlen_q, Int32) |
| 1894 | q_abs = q_idx + ctx_off |
| 1895 | keep = kv_idx <= q_abs |
| 1896 | r_start, r_end = _load_q_range(q_idx, seqlen_info, aux_tensors, batch_idx) |
| 1897 | keep = keep | ((kv_idx >= r_start) & (kv_idx <= r_end)) |
| 1898 | return keep |
| 1899 | |
| 1900 | mm_prefix_mask_mod.use_fast_sampling = True |
| 1901 | mm_prefix_mask_mod.__vec_size__ = 1 # _load_q_range takes lane 0 of q_idx |
| 1902 | return mm_prefix_mask_mod |
| 1903 | |
| 1904 | |
| 1905 | def _make_rswa_mask_mod(): |
| 1906 | """Build a CuTE-DSL mask_mod for Reference Sliding Window Attention (R-SWA). |
| 1907 | |
| 1908 | FA4 varlen + paged-KV convention (verified from cute/mask.py apply_mask): |
| 1909 | q_idx = LOCAL query-token offset (0 .. seqlen_q - 1) within this sequence. |
| 1910 | kv_idx = LOCAL KV-token position (0 .. seqlen_k - 1) within this sequence. |
| 1911 | |
| 1912 | To recover the ABSOLUTE token position (needed for causal and the sliding |
| 1913 | window distance), use the standard offset: |
| 1914 | abs_q = q_idx + (seqlen_k - seqlen_q) |
| 1915 | |
| 1916 | R-SWA keep condition: |
| 1917 | abs_q >= kv_idx (causal: KV at or before the query) |
| 1918 | AND (kv_idx < prefix_len (global prefix is always visible) |
| 1919 | OR abs_q - kv_idx < window) (generated tokens: sliding window) |
| 1920 | |
| 1921 | aux_tensors[0]: prefix_lens [num_reqs] int32 — per-request prefill length. |
| 1922 | aux_tensors[1]: rswa_window [1] int32 — decode sliding window size. |
| 1923 | |
| 1924 | use_fast_sampling=True lets FA4 skip fully-masked KV blocks (gap blocks) |
| 1925 | without loading their data. |
| 1926 | """ |
| 1927 | import cutlass.cute as cute |
| 1928 | from cutlass import Int32 # type: ignore[attr-defined] |
| 1929 | |
| 1930 | from vllm.vllm_flash_attn.cute.utils import ( # type: ignore[import-untyped] |
| 1931 | scalar_to_ssa, |
| 1932 | ) |
| 1933 | |
| 1934 | @cute.jit |
| 1935 | def rswa_mask_mod( |
| 1936 | batch_idx: cute.TensorSSA, |
| 1937 | head_idx: cute.TensorSSA, |
| 1938 | q_idx: cute.TensorSSA, |
| 1939 | kv_idx: cute.TensorSSA, |
| 1940 | seqlen_info, |
| 1941 | aux_tensors, |
| 1942 | ): |
| 1943 | b = batch_idx[0] |
| 1944 | prefix_len = scalar_to_ssa(aux_tensors[0][b], Int32) |
| 1945 | window = scalar_to_ssa(aux_tensors[1][0], Int32) |
| 1946 | # Convert local q offset to absolute token position. |
| 1947 | offset = scalar_to_ssa(seqlen_info.seqlen_k - seqlen_info.seqlen_q, Int32) |
| 1948 | abs_q = q_idx + offset |
| 1949 | causal = kv_idx <= abs_q |
| 1950 | in_prefix = kv_idx < prefix_len |
| 1951 | in_window = (abs_q - kv_idx) < window |
| 1952 | return causal & (in_prefix | in_window) |
| 1953 | |
| 1954 | rswa_mask_mod.use_fast_sampling = True |
| 1955 | return rswa_mask_mod |
| 1956 | |
| 1957 | |
| 1958 | def use_cascade_attention( |
| 1959 | common_prefix_len: int, |
| 1960 | query_lens: np.ndarray, |
| 1961 | num_query_heads: int, |
| 1962 | num_kv_heads: int, |
| 1963 | use_alibi: bool, |
| 1964 | use_sliding_window: bool, |
| 1965 | use_local_attention: bool, |
| 1966 | num_sms: int, |
| 1967 | dcp_world_size: int, |
| 1968 | ) -> bool: |
| 1969 | """Decide whether to use cascade attention. |
| 1970 | |
| 1971 | This function 1) checks whether cascade attention is supported with the |
| 1972 | given configuration, and 2) heuristically decides whether using cascade |
| 1973 | attention can improve performance. |
| 1974 | """ |
| 1975 | # Too short common prefix. Probably not worth using cascade attention. |
| 1976 | # We use an arbitrary threshold of 256 tokens. TODO: Tune this threshold. |
| 1977 | # NOTE(woosuk): This is the common case. We should return False as soon as |
| 1978 | # possible to avoid any unnecessary computation. |
| 1979 | if common_prefix_len < 256: |
| 1980 | return False |
| 1981 | # Cascade attention is currently not supported with these variants. |
| 1982 | if use_alibi or use_sliding_window or use_local_attention: |
| 1983 | return False |
| 1984 | # Too few queries. Probably not worth using cascade attention. |
| 1985 | # We use an arbitrary threshold of 8 queries. TODO: Tune this threshold. |
| 1986 | num_reqs = len(query_lens) |
| 1987 | if num_reqs < 8: |
| 1988 | return False |
| 1989 | # disable cascade attention for DCP |
| 1990 | if dcp_world_size > 1: |
| 1991 | return False |
| 1992 | |
| 1993 | # Heuristics to decide whether using cascade attention is beneficial. |
| 1994 | # 1. When FlashDecoding is not used for normal attention, cascade attention |
| 1995 | # is likely to be faster since it saves memory bandwidth. |
| 1996 | num_queries_per_kv = num_query_heads // num_kv_heads |
| 1997 | # The criteria for using FlashDecoding can be found in the following link: |
| 1998 | # https://github.com/vllm-project/flash-attention/blob/96266b1111111f3d11aabefaf3bacbab6a89d03c/csrc/flash_attn/flash_api.cpp#L535 |
| 1999 | use_flash_decoding = ( |
| 2000 | num_queries_per_kv > 1 |
| 2001 | and not use_sliding_window |
| 2002 | and not use_alibi |
| 2003 | and np.all(query_lens == 1) |
| 2004 | ) |
| 2005 | if not use_flash_decoding: |
| 2006 | # Use cascade attention. |
| 2007 | return True |
| 2008 | |
| 2009 | # 2. When FlashDecoding is used for normal attention, it is not clear |
| 2010 | # whether cascade attention is beneficial, because FlashDecoding can |
| 2011 | # launch more CTAs than cascade attention. |
| 2012 | # We use a simple performance model to compare the two methods. |
| 2013 | # NOTE(woosuk): The performance model is very rough and may not be |
| 2014 | # accurate. |
| 2015 | num_tokens = num_reqs |
| 2016 | # NOTE(woosuk): These are default tile sizes. flash-attn might use |
| 2017 | # different tile sizes (e.g., 64 or 256) depending on the configuration. |
| 2018 | q_tile_size = 128 |
| 2019 | kv_tile_size = 128 |
| 2020 | num_prefix_tiles = cdiv(common_prefix_len, kv_tile_size) |
| 2021 | |
| 2022 | cascade_ctas = num_query_heads * cdiv(num_tokens, q_tile_size) |
| 2023 | cascade_waves = cdiv(cascade_ctas, num_sms) |
| 2024 | cascade_time = cascade_waves * num_prefix_tiles |
| 2025 | |
| 2026 | flash_decoding_ctas = ( |
| 2027 | num_reqs * num_kv_heads * cdiv(num_queries_per_kv, q_tile_size) |
| 2028 | ) |
| 2029 | flash_decoding_ctas *= num_prefix_tiles |
| 2030 | flash_decoding_time = cdiv(flash_decoding_ctas, num_sms) |
| 2031 | |
| 2032 | # Use cascade attention if it is faster than FlashDecoding. |
| 2033 | return cascade_time < flash_decoding_time |
| 2034 | |
| 2035 | |
| 2036 | def cascade_attention( |
| 2037 | output: torch.Tensor, |
| 2038 | query: torch.Tensor, |
| 2039 | key_cache: torch.Tensor, |
| 2040 | value_cache: torch.Tensor, |
| 2041 | cu_query_lens: torch.Tensor, |
| 2042 | max_query_len: int, |
| 2043 | cu_prefix_query_lens: torch.Tensor, |
| 2044 | prefix_kv_lens: torch.Tensor, |
| 2045 | suffix_kv_lens: torch.Tensor, |
| 2046 | max_kv_len: int, |
| 2047 | softmax_scale: float, |
| 2048 | alibi_slopes: torch.Tensor | None, |
| 2049 | sliding_window: tuple[int, int], |
| 2050 | logits_soft_cap: float, |
| 2051 | block_table: torch.Tensor, |
| 2052 | common_prefix_len: int, |
| 2053 | max_num_splits: int, |
| 2054 | fa_version: int, |
| 2055 | prefix_scheduler_metadata: torch.Tensor | None = None, |
| 2056 | suffix_scheduler_metadata: torch.Tensor | None = None, |
| 2057 | q_descale: torch.Tensor | None = None, |
| 2058 | k_descale: torch.Tensor | None = None, |
| 2059 | v_descale: torch.Tensor | None = None, |
| 2060 | s_aux: torch.Tensor | None = None, |
| 2061 | ) -> torch.Tensor: |
| 2062 | assert alibi_slopes is None, "Cascade attention does not support ALiBi." |
| 2063 | # TODO: Support sliding window. |
| 2064 | assert sliding_window == (-1, -1), ( |
| 2065 | "Cascade attention does not support sliding window." |
| 2066 | ) |
| 2067 | |
| 2068 | num_tokens = query.shape[0] |
| 2069 | block_size = key_cache.shape[-3] |
| 2070 | assert common_prefix_len % block_size == 0 |
| 2071 | num_common_kv_blocks = common_prefix_len // block_size |
| 2072 | assert num_common_kv_blocks > 0 |
| 2073 | descale_shape = (cu_prefix_query_lens.shape[0] - 1, key_cache.shape[-2]) |
| 2074 | |
| 2075 | # Process shared prefix. |
| 2076 | prefix_output, prefix_lse = flash_attn_varlen_func( |
| 2077 | q=query, |
| 2078 | k=key_cache, |
| 2079 | v=value_cache, |
| 2080 | cu_seqlens_q=cu_prefix_query_lens, |
| 2081 | seqused_k=prefix_kv_lens, |
| 2082 | max_seqlen_q=num_tokens, |
| 2083 | max_seqlen_k=common_prefix_len, |
| 2084 | softmax_scale=softmax_scale, |
| 2085 | causal=False, |
| 2086 | window_size=list(sliding_window), |
| 2087 | block_table=block_table[:1], |
| 2088 | softcap=logits_soft_cap, |
| 2089 | return_softmax_lse=True, |
| 2090 | scheduler_metadata=prefix_scheduler_metadata, |
| 2091 | fa_version=fa_version, |
| 2092 | q_descale=q_descale.expand(descale_shape) if q_descale is not None else None, |
| 2093 | k_descale=k_descale.expand(descale_shape) if k_descale is not None else None, |
| 2094 | v_descale=v_descale.expand(descale_shape) if v_descale is not None else None, |
| 2095 | # s_aux is incorporated into prefix_lse inside the GPU kernel, |
| 2096 | # enabling its effect during the final attention merge. |
| 2097 | s_aux=s_aux, |
| 2098 | num_splits=1 if envs.VLLM_BATCH_INVARIANT else max_num_splits, |
| 2099 | ) |
| 2100 | |
| 2101 | descale_shape = (cu_query_lens.shape[0] - 1, key_cache.shape[-2]) |
| 2102 | |
| 2103 | # Process suffix per query. |
| 2104 | suffix_output, suffix_lse = flash_attn_varlen_func( |
| 2105 | q=query, |
| 2106 | k=key_cache, |
| 2107 | v=value_cache, |
| 2108 | cu_seqlens_q=cu_query_lens, |
| 2109 | seqused_k=suffix_kv_lens, |
| 2110 | max_seqlen_q=max_query_len, |
| 2111 | max_seqlen_k=max_kv_len - common_prefix_len, |
| 2112 | softmax_scale=softmax_scale, |
| 2113 | causal=True, |
| 2114 | window_size=list(sliding_window), |
| 2115 | block_table=block_table[:, num_common_kv_blocks:], |
| 2116 | softcap=logits_soft_cap, |
| 2117 | return_softmax_lse=True, |
| 2118 | scheduler_metadata=suffix_scheduler_metadata, |
| 2119 | fa_version=fa_version, |
| 2120 | q_descale=q_descale.expand(descale_shape) if q_descale is not None else None, |
| 2121 | k_descale=k_descale.expand(descale_shape) if k_descale is not None else None, |
| 2122 | v_descale=v_descale.expand(descale_shape) if v_descale is not None else None, |
| 2123 | num_splits=1 if envs.VLLM_BATCH_INVARIANT else max_num_splits, |
| 2124 | ) |
| 2125 | |
| 2126 | # Merge prefix and suffix outputs, and store the result in output. |
| 2127 | merge_attn_states(output, prefix_output, prefix_lse, suffix_output, suffix_lse) |
| 2128 | |
| 2129 | |
| 2130 | _FA4_DENSE_ATTENTION_KERNEL = FA4DenseAttentionKernel() |