| 1 | # SPDX-License-Identifier: Apache-2.0 |
| 2 | # SPDX-FileCopyrightText: Copyright contributors to the vLLM project |
| 3 | """FlashAttention backend for MLA prefill.""" |
| 4 | |
| 5 | import functools |
| 6 | from collections.abc import Callable |
| 7 | from dataclasses import dataclass |
| 8 | from typing import TYPE_CHECKING, Any |
| 9 | |
| 10 | import torch |
| 11 | |
| 12 | import vllm.envs as envs |
| 13 | from vllm.model_executor.layers.quantization.utils.quant_utils import ( |
| 14 | kFp8StaticTensorSym, |
| 15 | ) |
| 16 | from vllm.model_executor.warmup.jit_warmup import VllmJitKernel, zip_inputs |
| 17 | from vllm.platforms import current_platform |
| 18 | from vllm.v1.attention.backends.fa_utils import ( |
| 19 | compile_flash_attn_varlen_func_from_specs, |
| 20 | get_flash_attn_version, |
| 21 | is_flash_attn_varlen_func_available, |
| 22 | ) |
| 23 | from vllm.v1.attention.backends.mla.prefill.base import ( |
| 24 | MLADimensions, |
| 25 | MLAPrefillBackend, |
| 26 | ) |
| 27 | |
| 28 | if TYPE_CHECKING: |
| 29 | from vllm.config import VllmConfig |
| 30 | from vllm.model_executor.layers.attention.mla_attention import ( |
| 31 | MLACommonPrefillMetadata, |
| 32 | MLADims, |
| 33 | ) |
| 34 | from vllm.model_executor.layers.quantization.utils.quant_utils import QuantKey |
| 35 | |
| 36 | if is_flash_attn_varlen_func_available(): |
| 37 | from vllm.v1.attention.backends.fa_utils import flash_attn_varlen_func |
| 38 | else: |
| 39 | flash_attn_varlen_func = None # type: ignore[assignment] |
| 40 | |
| 41 | |
| 42 | FA4_STANDARD_DTYPES = (torch.bfloat16, torch.float16) |
| 43 | |
| 44 | # Current vLLM MLA prefill expands K/V to num_heads before FA4, so this plan |
| 45 | # covers qhead_per_kvhead=1. Batch is not a current FA4 MLA-prefill key field. |
| 46 | # Use b1 for compile-only specs because it is the conservative case for |
| 47 | # Split-KV shape heuristics. |
| 48 | # TODO(roberto): FA4 also has direct-GQA and qv/top-k absorbed-MLA paths, but |
| 49 | # vLLM does not use them in this backend yet; they need a separate |
| 50 | # num_kv_heads/qv/top-k-aware warmup plan if wired in later. |
| 51 | FA4_MLA_PREFILL_COMPILE_BATCH_SIZE = 1 |
| 52 | FA4_MLA_PREFILL_Q_TILE = 128 |
| 53 | FA4_MLA_PREFILL_K_TILE = 128 |
| 54 | FA4_MLA_PREFILL_LONG_K_BLOCKS = 32 |
| 55 | FA4_MLA_PREFILL_VERY_LONG_K_BLOCKS = 64 |
| 56 | FA4_MLA_PREFILL_CAUSAL_OPTIONS = (False, True) |
| 57 | FA4_MLA_PREFILL_LSE_OPTIONS = (False, True) |
| 58 | |
| 59 | |
| 60 | class FA4MLAPrefillKernel(VllmJitKernel["FA4MLAPrefillKernel.CompileKey"]): |
| 61 | """FA4 MLA prefill compile-key wrapper used by generic JIT warmup.""" |
| 62 | |
| 63 | @dataclass(frozen=True) |
| 64 | class CompileKey: |
| 65 | """High-level FA4 compile-only key used by vLLM warmup. |
| 66 | |
| 67 | This is not the CuTeDSL cache key. FA4 owns the selector that maps these |
| 68 | serving inputs to the actual compile-static fields: tile sizes, q_stage, |
| 69 | Split-KV, scheduler choice, layout-presence booleans, dtype/head dims, |
| 70 | arch, and related fields. |
| 71 | """ |
| 72 | |
| 73 | q_shape: tuple[int, ...] |
| 74 | k_shape: tuple[int, ...] |
| 75 | v_shape: tuple[int, ...] |
| 76 | q_dtype: torch.dtype |
| 77 | max_seqlen_q: int |
| 78 | max_seqlen_k: int |
| 79 | softmax_scale: float |
| 80 | causal: bool |
| 81 | fa_version: int |
| 82 | v_stride: tuple[int, ...] | None = None |
| 83 | cu_seqlens_q_shape: tuple[int, ...] | None = None |
| 84 | cu_seqlens_k_shape: tuple[int, ...] | None = None |
| 85 | window_size: tuple[int, int] | None = None |
| 86 | return_softmax_lse: bool = False |
| 87 | num_splits: int = 0 |
| 88 | |
| 89 | @staticmethod |
| 90 | def kernel( |
| 91 | *args: Any, |
| 92 | runtime_kernel: Callable[..., Any] | None = None, |
| 93 | **kwargs: Any, |
| 94 | ) -> Any: |
| 95 | kernel = flash_attn_varlen_func if runtime_kernel is None else runtime_kernel |
| 96 | assert kernel is not None |
| 97 | return kernel(*args, **kwargs) |
| 98 | |
| 99 | def dispatch( # type: ignore[override] |
| 100 | self, |
| 101 | *, |
| 102 | batch_size: int, |
| 103 | dtype: torch.dtype, |
| 104 | num_heads: int, |
| 105 | mla_dims: "MLADims", |
| 106 | max_seqlen_q: int, |
| 107 | max_seqlen_k: int, |
| 108 | requires_v_padding: bool, |
| 109 | causal: bool, |
| 110 | fa_version: int, |
| 111 | window_size: tuple[int, int] | None = None, |
| 112 | return_lse: bool = False, |
| 113 | num_splits: int = 0, |
| 114 | ) -> CompileKey: |
| 115 | return self.CompileKey( |
| 116 | q_shape=( |
| 117 | batch_size * max_seqlen_q, |
| 118 | num_heads, |
| 119 | mla_dims.qk_nope_head_dim + mla_dims.qk_rope_head_dim, |
| 120 | ), |
| 121 | k_shape=( |
| 122 | batch_size * max_seqlen_k, |
| 123 | num_heads, |
| 124 | mla_dims.qk_nope_head_dim + mla_dims.qk_rope_head_dim, |
| 125 | ), |
| 126 | v_shape=( |
| 127 | batch_size * max_seqlen_k, |
| 128 | num_heads, |
| 129 | ( |
| 130 | mla_dims.qk_nope_head_dim + mla_dims.qk_rope_head_dim |
| 131 | if requires_v_padding |
| 132 | else mla_dims.v_head_dim |
| 133 | ), |
| 134 | ), |
| 135 | q_dtype=dtype, |
| 136 | max_seqlen_q=max_seqlen_q, |
| 137 | max_seqlen_k=max_seqlen_k, |
| 138 | softmax_scale=(mla_dims.qk_nope_head_dim + mla_dims.qk_rope_head_dim) |
| 139 | ** -0.5, |
| 140 | causal=causal, |
| 141 | fa_version=fa_version, |
| 142 | v_stride=( |
| 143 | None |
| 144 | if requires_v_padding |
| 145 | else ( |
| 146 | num_heads * (mla_dims.qk_nope_head_dim + mla_dims.v_head_dim), |
| 147 | mla_dims.qk_nope_head_dim + mla_dims.v_head_dim, |
| 148 | 1, |
| 149 | ) |
| 150 | ), |
| 151 | cu_seqlens_q_shape=(batch_size + 1,), |
| 152 | cu_seqlens_k_shape=(batch_size + 1,), |
| 153 | window_size=window_size, |
| 154 | return_softmax_lse=return_lse, |
| 155 | num_splits=num_splits, |
| 156 | ) |
| 157 | |
| 158 | def _is_valid_warmup_shape_probe( |
| 159 | self, |
| 160 | *, |
| 161 | max_seqlen_k: int, |
| 162 | num_splits: int, |
| 163 | is_sm90: bool, |
| 164 | qk_head_dim: int, |
| 165 | effective_v_head_dim: int, |
| 166 | ) -> bool: |
| 167 | long_k = FA4_MLA_PREFILL_LONG_K_BLOCKS * FA4_MLA_PREFILL_K_TILE |
| 168 | very_long_k = FA4_MLA_PREFILL_VERY_LONG_K_BLOCKS * FA4_MLA_PREFILL_K_TILE |
| 169 | return ( |
| 170 | max_seqlen_k < long_k |
| 171 | or (num_splits != 1 and max_seqlen_k == long_k) |
| 172 | or ( |
| 173 | num_splits != 1 |
| 174 | and max_seqlen_k == very_long_k |
| 175 | and not is_sm90 |
| 176 | and qk_head_dim != effective_v_head_dim |
| 177 | ) |
| 178 | ) |
| 179 | |
| 180 | def get_warmup_keys(self, vllm_config: "VllmConfig") -> list[CompileKey]: |
| 181 | from vllm.model_executor.layers.attention.mla_attention import ( |
| 182 | get_mla_dims, |
| 183 | ) |
| 184 | |
| 185 | mla_dims = get_mla_dims(vllm_config.model_config) |
| 186 | qk_head_dim = mla_dims.qk_nope_head_dim + mla_dims.qk_rope_head_dim |
| 187 | fa_version = get_flash_attn_version(head_size=qk_head_dim) |
| 188 | if fa_version != 4: |
| 189 | return [] |
| 190 | |
| 191 | dtype = vllm_config.model_config.dtype |
| 192 | if dtype not in FA4_STANDARD_DTYPES: |
| 193 | dtype = torch.bfloat16 |
| 194 | |
| 195 | is_sm90 = current_platform.is_device_capability(90) |
| 196 | is_sm100_family = current_platform.is_device_capability_family(100) |
| 197 | is_sm120 = current_platform.is_device_capability_family(120) |
| 198 | if not (is_sm90 or is_sm100_family or is_sm120): |
| 199 | return [] |
| 200 | |
| 201 | num_splits = 1 if envs.VLLM_BATCH_INVARIANT else 0 |
| 202 | if is_sm120 and num_splits != 1: |
| 203 | return [] |
| 204 | |
| 205 | num_heads = vllm_config.model_config.get_num_attention_heads( |
| 206 | vllm_config.parallel_config |
| 207 | ) |
| 208 | if num_heads <= 0: |
| 209 | return [] |
| 210 | |
| 211 | requires_v_padding = False |
| 212 | effective_v_head_dim = ( |
| 213 | qk_head_dim if requires_v_padding else mla_dims.v_head_dim |
| 214 | ) |
| 215 | shape_probes = ( |
| 216 | dict(max_seqlen_q=1, max_seqlen_k=FA4_MLA_PREFILL_K_TILE), |
| 217 | dict( |
| 218 | max_seqlen_q=FA4_MLA_PREFILL_Q_TILE + 1, |
| 219 | max_seqlen_k=4 * FA4_MLA_PREFILL_K_TILE, |
| 220 | ), |
| 221 | dict( |
| 222 | max_seqlen_q=1, |
| 223 | max_seqlen_k=FA4_MLA_PREFILL_LONG_K_BLOCKS * FA4_MLA_PREFILL_K_TILE, |
| 224 | ), |
| 225 | dict( |
| 226 | max_seqlen_q=FA4_MLA_PREFILL_Q_TILE + 1, |
| 227 | max_seqlen_k=FA4_MLA_PREFILL_LONG_K_BLOCKS * FA4_MLA_PREFILL_K_TILE, |
| 228 | ), |
| 229 | dict( |
| 230 | max_seqlen_q=1, |
| 231 | max_seqlen_k=FA4_MLA_PREFILL_VERY_LONG_K_BLOCKS |
| 232 | * FA4_MLA_PREFILL_K_TILE, |
| 233 | ), |
| 234 | dict( |
| 235 | max_seqlen_q=FA4_MLA_PREFILL_Q_TILE + 1, |
| 236 | max_seqlen_k=FA4_MLA_PREFILL_VERY_LONG_K_BLOCKS |
| 237 | * FA4_MLA_PREFILL_K_TILE, |
| 238 | ), |
| 239 | ) |
| 240 | |
| 241 | return self._trace_dispatch(self.dispatch)( |
| 242 | zip_inputs(*shape_probes), |
| 243 | batch_size=FA4_MLA_PREFILL_COMPILE_BATCH_SIZE, |
| 244 | dtype=dtype, |
| 245 | num_heads=num_heads, |
| 246 | mla_dims=mla_dims, |
| 247 | requires_v_padding=requires_v_padding, |
| 248 | is_sm90=is_sm90, |
| 249 | qk_head_dim=qk_head_dim, |
| 250 | effective_v_head_dim=effective_v_head_dim, |
| 251 | causal=FA4_MLA_PREFILL_CAUSAL_OPTIONS, |
| 252 | return_lse=FA4_MLA_PREFILL_LSE_OPTIONS, |
| 253 | num_splits=num_splits, |
| 254 | fa_version=fa_version, |
| 255 | _when=self._is_valid_warmup_shape_probe, |
| 256 | ) |
| 257 | |
| 258 | def compile(self, compile_key: CompileKey) -> None: |
| 259 | assert compile_flash_attn_varlen_func_from_specs is not None |
| 260 | window_size = ( |
| 261 | list(compile_key.window_size) |
| 262 | if compile_key.window_size is not None |
| 263 | else None |
| 264 | ) |
| 265 | compile_flash_attn_varlen_func_from_specs( |
| 266 | q_shape=compile_key.q_shape, |
| 267 | k_shape=compile_key.k_shape, |
| 268 | v_shape=compile_key.v_shape, |
| 269 | q_dtype=compile_key.q_dtype, |
| 270 | v_stride=compile_key.v_stride, |
| 271 | cu_seqlens_q_shape=compile_key.cu_seqlens_q_shape, |
| 272 | cu_seqlens_k_shape=compile_key.cu_seqlens_k_shape, |
| 273 | max_seqlen_q=compile_key.max_seqlen_q, |
| 274 | max_seqlen_k=compile_key.max_seqlen_k, |
| 275 | softmax_scale=compile_key.softmax_scale, |
| 276 | causal=compile_key.causal, |
| 277 | window_size=window_size, |
| 278 | return_softmax_lse=compile_key.return_softmax_lse, |
| 279 | fa_version=compile_key.fa_version, |
| 280 | num_splits=compile_key.num_splits, |
| 281 | ) |
| 282 | |
| 283 | def __call__( |
| 284 | self, |
| 285 | *, |
| 286 | q: torch.Tensor, |
| 287 | k: torch.Tensor, |
| 288 | v: torch.Tensor, |
| 289 | runtime_kernel: Callable[..., Any] | None = None, |
| 290 | **kwargs: Any, |
| 291 | ) -> Any: |
| 292 | return self.kernel( |
| 293 | q=q, |
| 294 | k=k, |
| 295 | v=v, |
| 296 | runtime_kernel=runtime_kernel, |
| 297 | **kwargs, |
| 298 | ) |
| 299 | |
| 300 | |
| 301 | class FlashAttnPrefillBackend(MLAPrefillBackend): |
| 302 | """FlashAttention backend for MLA prefill.""" |
| 303 | |
| 304 | @staticmethod |
| 305 | def get_name() -> str: |
| 306 | return "FLASH_ATTN" |
| 307 | |
| 308 | @classmethod |
| 309 | def is_available(cls) -> bool: |
| 310 | return is_flash_attn_varlen_func_available() |
| 311 | |
| 312 | @classmethod |
| 313 | def supports_mla_dimensions(cls, mla_dimensions: MLADimensions) -> bool: |
| 314 | return mla_dimensions in [ |
| 315 | MLADimensions( |
| 316 | qk_nope_head_dim=128, |
| 317 | qk_rope_head_dim=64, |
| 318 | v_head_dim=128, |
| 319 | ), |
| 320 | MLADimensions( |
| 321 | qk_nope_head_dim=192, |
| 322 | qk_rope_head_dim=64, |
| 323 | v_head_dim=256, |
| 324 | ), |
| 325 | MLADimensions( |
| 326 | qk_nope_head_dim=64, |
| 327 | qk_rope_head_dim=64, |
| 328 | v_head_dim=128, |
| 329 | ), |
| 330 | # GLM5Next NoPE layout: qk_head_dim 256 + 0 and v_head_dim 256 run |
| 331 | # the same kernels as the (192, 64, 256) DeepSeek-V3.2 layout. |
| 332 | MLADimensions( |
| 333 | qk_nope_head_dim=256, |
| 334 | qk_rope_head_dim=0, |
| 335 | v_head_dim=256, |
| 336 | ), |
| 337 | ] |
| 338 | |
| 339 | def __init__( |
| 340 | self, |
| 341 | num_heads: int, |
| 342 | scale: float, |
| 343 | kv_lora_rank: int, |
| 344 | qk_nope_head_dim: int, |
| 345 | qk_rope_head_dim: int, |
| 346 | v_head_dim: int, |
| 347 | vllm_config: "VllmConfig", |
| 348 | ) -> None: |
| 349 | super().__init__( |
| 350 | num_heads=num_heads, |
| 351 | scale=scale, |
| 352 | kv_lora_rank=kv_lora_rank, |
| 353 | qk_nope_head_dim=qk_nope_head_dim, |
| 354 | qk_rope_head_dim=qk_rope_head_dim, |
| 355 | v_head_dim=v_head_dim, |
| 356 | vllm_config=vllm_config, |
| 357 | ) |
| 358 | |
| 359 | # Handle the differences between the flash_attn_varlen from |
| 360 | # flash_attn and the one from vllm_flash_attn |
| 361 | assert flash_attn_varlen_func is not None, ( |
| 362 | "FlashAttnPrefillBackend requires flash_attn_varlen_func. " |
| 363 | "Ensure FlashAttnPrefillBackend.is_available() is checked first." |
| 364 | ) |
| 365 | qk_head_dim = qk_nope_head_dim + qk_rope_head_dim |
| 366 | self.flash_attn_varlen_func = flash_attn_varlen_func |
| 367 | self.vllm_flash_attn_version = get_flash_attn_version( |
| 368 | head_size=qk_head_dim, head_size_v=v_head_dim |
| 369 | ) |
| 370 | if self.vllm_flash_attn_version is not None: |
| 371 | self.flash_attn_varlen_func = functools.partial( |
| 372 | flash_attn_varlen_func, fa_version=self.vllm_flash_attn_version |
| 373 | ) |
| 374 | |
| 375 | if ( |
| 376 | vllm_config.kernel_config.enable_jit_warmup |
| 377 | and self.vllm_flash_attn_version == 4 |
| 378 | ): |
| 379 | _FA4_MLA_PREFILL_KERNEL.register_warmup() |
| 380 | |
| 381 | # Determine if we need to pad V |
| 382 | # For MLA the v head dim is smaller than qk head dim so we pad out |
| 383 | # v with 0s to match the qk head dim for attention backends that do |
| 384 | # not support different headdims. |
| 385 | # FA3 on Hopper (SM90) and FA4 natively handle diff headdims. |
| 386 | device_capability = current_platform.get_device_capability() |
| 387 | self.requires_v_padding = self.vllm_flash_attn_version is None or not ( |
| 388 | ( |
| 389 | self.vllm_flash_attn_version == 3 |
| 390 | and device_capability is not None |
| 391 | and device_capability[0] == 9 |
| 392 | ) |
| 393 | or self.vllm_flash_attn_version == 4 |
| 394 | ) |
| 395 | |
| 396 | # Track whether we're using vllm's FA or upstream (for ROCm) |
| 397 | self._is_vllm_fa = current_platform.is_cuda() or current_platform.is_xpu() |
| 398 | |
| 399 | def supports_quant_output(self, quant_key: "QuantKey") -> bool: |
| 400 | device_capability = current_platform.get_device_capability() |
| 401 | return ( |
| 402 | self.vllm_flash_attn_version == 4 |
| 403 | and self._is_vllm_fa |
| 404 | and device_capability is not None |
| 405 | and device_capability[0] in (10, 11) |
| 406 | and quant_key == kFp8StaticTensorSym |
| 407 | ) |
| 408 | |
| 409 | def _flash_attn_varlen_diff_headdims( |
| 410 | self, |
| 411 | q: torch.Tensor, |
| 412 | k: torch.Tensor, |
| 413 | v: torch.Tensor, |
| 414 | return_softmax_lse: bool = False, |
| 415 | softmax_scale: float | None = None, |
| 416 | out: torch.Tensor | None = None, |
| 417 | output_scale: torch.Tensor | None = None, |
| 418 | **kwargs, |
| 419 | ) -> torch.Tensor | tuple[torch.Tensor, torch.Tensor]: |
| 420 | maybe_padded_v = v |
| 421 | if self.requires_v_padding: |
| 422 | maybe_padded_v = torch.nn.functional.pad( |
| 423 | v, [0, q.shape[-1] - v.shape[-1]], value=0 |
| 424 | ) |
| 425 | |
| 426 | if self._is_vllm_fa: |
| 427 | kwargs["return_softmax_lse"] = return_softmax_lse |
| 428 | kwargs["out"] = out |
| 429 | kwargs["output_scale"] = output_scale |
| 430 | else: |
| 431 | # ROCm leverages the upstream flash_attn, which takes a parameter |
| 432 | # called "return_attn_probs" instead of return_softmax_lse |
| 433 | kwargs["return_attn_probs"] = return_softmax_lse |
| 434 | assert out is None and output_scale is None |
| 435 | if envs.VLLM_BATCH_INVARIANT: |
| 436 | kwargs["num_splits"] = 1 |
| 437 | |
| 438 | attn_out = _FA4_MLA_PREFILL_KERNEL( |
| 439 | q=q, |
| 440 | k=k, |
| 441 | v=maybe_padded_v, |
| 442 | runtime_kernel=self.flash_attn_varlen_func, |
| 443 | softmax_scale=softmax_scale, |
| 444 | **kwargs, |
| 445 | ) |
| 446 | |
| 447 | # Unpack the output if there are multiple results |
| 448 | lse = None |
| 449 | if isinstance(attn_out, tuple): |
| 450 | attn_out, lse = attn_out[0], attn_out[1] |
| 451 | |
| 452 | # Remain consistent with old `flash_attn_varlen_func` where there |
| 453 | # is only one output tensor if `return_softmax_lse` is False. |
| 454 | if return_softmax_lse: |
| 455 | return attn_out, lse |
| 456 | return attn_out |
| 457 | |
| 458 | def supports_out(self) -> bool: |
| 459 | # A padded V produces a qk_head_dim output that cannot be written into |
| 460 | # a v_head_dim `out`; only the unpadded path honors `out`. |
| 461 | return not self.requires_v_padding |
| 462 | |
| 463 | def run_prefill_new_tokens( |
| 464 | self, |
| 465 | q: torch.Tensor, |
| 466 | k: torch.Tensor, |
| 467 | v: torch.Tensor, |
| 468 | return_softmax_lse: bool, |
| 469 | out: torch.Tensor | None = None, |
| 470 | output_scale: torch.Tensor | None = None, |
| 471 | ) -> torch.Tensor | tuple[torch.Tensor, torch.Tensor]: |
| 472 | return self._flash_attn_varlen_diff_headdims( |
| 473 | q=q, |
| 474 | k=k, |
| 475 | v=v, |
| 476 | cu_seqlens_q=self._prefill_metadata.query_start_loc, |
| 477 | cu_seqlens_k=self._prefill_metadata.query_start_loc, |
| 478 | max_seqlen_q=self._prefill_metadata.max_query_len, |
| 479 | max_seqlen_k=self._prefill_metadata.max_query_len, |
| 480 | softmax_scale=self.scale, |
| 481 | causal=True, |
| 482 | return_softmax_lse=return_softmax_lse, |
| 483 | out=out, |
| 484 | output_scale=output_scale, |
| 485 | ) |
| 486 | |
| 487 | def run_prefill_context_chunk( |
| 488 | self, |
| 489 | chunk: "MLACommonPrefillMetadata.ContextChunk", |
| 490 | q: torch.Tensor, |
| 491 | k: torch.Tensor, |
| 492 | v: torch.Tensor, |
| 493 | out: torch.Tensor | None = None, |
| 494 | ) -> tuple[torch.Tensor, torch.Tensor]: |
| 495 | return self._flash_attn_varlen_diff_headdims( |
| 496 | q=q, |
| 497 | k=k, |
| 498 | v=v, |
| 499 | cu_seqlens_q=chunk.query_start_loc, |
| 500 | cu_seqlens_k=chunk.cu_seq_lens, |
| 501 | max_seqlen_q=chunk.max_query_len, |
| 502 | max_seqlen_k=chunk.max_seq_len, |
| 503 | softmax_scale=self.scale, |
| 504 | causal=False, # Context is unmasked |
| 505 | return_softmax_lse=True, |
| 506 | out=out, |
| 507 | ) |
| 508 | |
| 509 | |
| 510 | _FA4_MLA_PREFILL_KERNEL = FA4MLAPrefillKernel() |