| 1 | # SPDX-License-Identifier: Apache-2.0 |
| 2 | # SPDX-FileCopyrightText: Copyright contributors to the vLLM project |
| 3 | |
| 4 | from dataclasses import dataclass |
| 5 | from typing import ClassVar |
| 6 | |
| 7 | import torch |
| 8 | |
| 9 | import vllm.envs as envs |
| 10 | from vllm.config import VllmConfig |
| 11 | from vllm.config.cache import CacheDType |
| 12 | from vllm.logger import init_logger |
| 13 | from vllm.model_executor.layers.attention.mla_attention import ( |
| 14 | MLACommonBackend, |
| 15 | MLACommonDecodeMetadata, |
| 16 | MLACommonImpl, |
| 17 | MLACommonMetadata, |
| 18 | MLACommonMetadataBuilder, |
| 19 | QueryLenSupport, |
| 20 | ) |
| 21 | from vllm.platforms.interface import DeviceCapability |
| 22 | from vllm.utils.math_utils import round_up |
| 23 | from vllm.utils.torch_utils import is_quantized_kv_cache |
| 24 | from vllm.v1.attention.backend import ( |
| 25 | AttentionCGSupport, |
| 26 | AttentionLayer, |
| 27 | AttentionType, |
| 28 | MultipleOf, |
| 29 | ) |
| 30 | from vllm.v1.attention.backends.fa_utils import ( |
| 31 | flash_attn_supports_mla, |
| 32 | get_flash_attn_version, |
| 33 | ) |
| 34 | from vllm.v1.kv_cache_interface import AttentionSpec |
| 35 | from vllm.vllm_flash_attn import ( # type: ignore[attr-defined] |
| 36 | flash_attn_varlen_func, |
| 37 | get_scheduler_metadata, |
| 38 | ) |
| 39 | |
| 40 | logger = init_logger(__name__) |
| 41 | |
| 42 | |
| 43 | class FlashAttnMLABackend(MLACommonBackend): |
| 44 | supported_dtypes: ClassVar[list[torch.dtype]] = [torch.float16, torch.bfloat16] |
| 45 | supported_kv_cache_dtypes: ClassVar[list[CacheDType]] = [ |
| 46 | "auto", |
| 47 | "float16", |
| 48 | "bfloat16", |
| 49 | ] |
| 50 | |
| 51 | @staticmethod |
| 52 | def get_supported_kernel_block_sizes(kv_cache_spec=None) -> list[int | MultipleOf]: |
| 53 | return [MultipleOf(16)] |
| 54 | |
| 55 | @staticmethod |
| 56 | def get_name() -> str: |
| 57 | return "FLASH_ATTN_MLA" |
| 58 | |
| 59 | @classmethod |
| 60 | def supports_batch_invariance(cls) -> bool: |
| 61 | return True |
| 62 | |
| 63 | @staticmethod |
| 64 | def get_builder_cls() -> type["FlashAttnMLAMetadataBuilder"]: |
| 65 | return FlashAttnMLAMetadataBuilder |
| 66 | |
| 67 | @staticmethod |
| 68 | def get_impl_cls() -> type["FlashAttnMLAImpl"]: |
| 69 | return FlashAttnMLAImpl |
| 70 | |
| 71 | @classmethod |
| 72 | def supports_compute_capability(cls, capability: DeviceCapability) -> bool: |
| 73 | return capability.major == 9 |
| 74 | |
| 75 | @classmethod |
| 76 | def supports_combination( |
| 77 | cls, |
| 78 | head_size: int, |
| 79 | dtype: torch.dtype, |
| 80 | kv_cache_dtype: CacheDType | None, |
| 81 | block_size: int | None, |
| 82 | use_mla: bool, |
| 83 | has_sink: bool, |
| 84 | use_sparse: bool, |
| 85 | use_mm_prefix: bool, |
| 86 | device_capability: DeviceCapability, |
| 87 | ) -> str | None: |
| 88 | if not flash_attn_supports_mla(): |
| 89 | return "FlashAttention MLA not supported on this device" |
| 90 | return None |
| 91 | |
| 92 | |
| 93 | @dataclass |
| 94 | class FlashAttnMLADecodeMetadata(MLACommonDecodeMetadata): |
| 95 | query_start_loc: torch.Tensor |
| 96 | max_query_len: int |
| 97 | max_seq_len: int |
| 98 | scheduler_metadata: torch.Tensor | None = None |
| 99 | max_num_splits: int = 0 |
| 100 | |
| 101 | |
| 102 | @dataclass |
| 103 | class FlashAttnMLAMetadata(MLACommonMetadata[FlashAttnMLADecodeMetadata]): |
| 104 | pass |
| 105 | |
| 106 | |
| 107 | class FlashAttnMLAMetadataBuilder(MLACommonMetadataBuilder[FlashAttnMLAMetadata]): |
| 108 | _cudagraph_support: ClassVar[AttentionCGSupport] = AttentionCGSupport.UNIFORM_BATCH |
| 109 | query_len_support: ClassVar[QueryLenSupport] = QueryLenSupport.VARLEN |
| 110 | reorder_batch_threshold: int = 512 # process small prefills with decode pathway |
| 111 | |
| 112 | def __init__( |
| 113 | self, |
| 114 | kv_cache_spec: AttentionSpec, |
| 115 | layer_names: list[str], |
| 116 | vllm_config: VllmConfig, |
| 117 | device: torch.device, |
| 118 | ): |
| 119 | interleave_size = vllm_config.parallel_config.cp_kv_cache_interleave_size |
| 120 | super().__init__( |
| 121 | kv_cache_spec, |
| 122 | layer_names, |
| 123 | vllm_config, |
| 124 | device, |
| 125 | FlashAttnMLAMetadata, |
| 126 | supports_dcp_with_varlen=(interleave_size == 1), |
| 127 | ) |
| 128 | self.max_num_splits = 0 # No upper bound on the number of splits. |
| 129 | self.fa_aot_schedule = get_flash_attn_version() == 3 |
| 130 | |
| 131 | self.use_full_cuda_graph = ( |
| 132 | self.compilation_config.cudagraph_mode.has_full_cudagraphs() |
| 133 | ) |
| 134 | self.max_cudagraph_size = self.compilation_config.max_cudagraph_capture_size |
| 135 | |
| 136 | if self.use_full_cuda_graph and self.fa_aot_schedule: |
| 137 | # FA3 scheduler_metadata size: 1 + round_up(batch_size, 4) * 4 |
| 138 | # The +1 is for the tile_count_semaphore (synchronization). |
| 139 | # The 4 slots per batch element (num_prepare_batch_vectors) are: |
| 140 | # prepare_varlen + dynamic_split + sort_batches + head_swizzle |
| 141 | # See: https://github.com/vllm-project/flash-attention/blob/5824e6e/hopper/flash_api.cpp#L664-L671 # noqa: E501 |
| 142 | max_batch_size = max( |
| 143 | vllm_config.scheduler_config.max_num_seqs, |
| 144 | self.max_cudagraph_size or 0, |
| 145 | ) |
| 146 | self.scheduler_metadata = torch.zeros( |
| 147 | 1 + round_up(max_batch_size, 4) * 4, |
| 148 | dtype=torch.int32, |
| 149 | device=self.device, |
| 150 | ) |
| 151 | # When using cuda graph, we need to set the upper bound of the |
| 152 | # number of splits so that large enough intermediate buffers are |
| 153 | # pre-allocated during capture. |
| 154 | self.max_num_splits = ( |
| 155 | vllm_config.attention_config.flash_attn_max_num_splits_for_cuda_graph |
| 156 | ) |
| 157 | |
| 158 | if envs.VLLM_BATCH_INVARIANT: |
| 159 | self.max_num_splits = 1 |
| 160 | |
| 161 | def _schedule_decode( |
| 162 | self, |
| 163 | num_reqs, |
| 164 | cu_query_lens, |
| 165 | max_query_len, |
| 166 | seqlens, |
| 167 | max_seq_len, |
| 168 | causal, |
| 169 | max_num_splits, |
| 170 | ): |
| 171 | if self.fa_aot_schedule: |
| 172 | return get_scheduler_metadata( |
| 173 | batch_size=num_reqs, |
| 174 | max_seqlen_q=max_query_len, |
| 175 | max_seqlen_k=max_seq_len, |
| 176 | num_heads_q=self.num_heads * self.dcp_world_size, |
| 177 | num_heads_kv=1, |
| 178 | headdim=self.mla_dims.qk_rope_head_dim, |
| 179 | cache_seqlens=seqlens, |
| 180 | qkv_dtype=self.kv_cache_spec.dtype, |
| 181 | headdim_v=self.mla_dims.kv_lora_rank, |
| 182 | page_size=self.page_size, |
| 183 | cu_seqlens_q=cu_query_lens, |
| 184 | causal=causal, |
| 185 | num_splits=max_num_splits, |
| 186 | ) |
| 187 | return None |
| 188 | |
| 189 | def _build_decode( |
| 190 | self, |
| 191 | block_table_tensor: torch.Tensor, |
| 192 | seq_lens_device: torch.Tensor, |
| 193 | max_seq_len: int, |
| 194 | query_start_loc_cpu: torch.Tensor, |
| 195 | query_start_loc_device: torch.Tensor, |
| 196 | num_decode_tokens: int, |
| 197 | max_query_len: int, |
| 198 | dcp_tot_seq_lens_device: torch.Tensor | None, |
| 199 | ) -> FlashAttnMLADecodeMetadata: |
| 200 | # Measured, not the promised bound: _schedule_decode asserts metadata size |
| 201 | # against a persistent buffer. |
| 202 | query_lens_cpu = query_start_loc_cpu[1:] - query_start_loc_cpu[:-1] |
| 203 | max_query_len = query_lens_cpu.max().item() |
| 204 | |
| 205 | # For Flash Attention MLA + full cudagraph |
| 206 | max_num_splits = 0 |
| 207 | if ( |
| 208 | self.use_full_cuda_graph |
| 209 | and self.max_cudagraph_size is not None |
| 210 | and num_decode_tokens <= self.max_cudagraph_size |
| 211 | ): |
| 212 | # NOTE(woosuk): Setting num_splits > 1 may increase the memory |
| 213 | # usage, because the intermediate buffers of size [num_splits, |
| 214 | # num_heads, num_tokens, head_size] are allocated. Therefore, |
| 215 | # we only set num_splits when using cuda graphs. |
| 216 | max_num_splits = self.max_num_splits |
| 217 | |
| 218 | if envs.VLLM_BATCH_INVARIANT: |
| 219 | max_num_splits = 1 |
| 220 | |
| 221 | scheduler_metadata = self._schedule_decode( |
| 222 | num_reqs=seq_lens_device.shape[0], |
| 223 | cu_query_lens=query_start_loc_device, |
| 224 | max_query_len=max_query_len, |
| 225 | seqlens=seq_lens_device, |
| 226 | max_seq_len=max_seq_len, |
| 227 | causal=True, |
| 228 | max_num_splits=max_num_splits, |
| 229 | ) |
| 230 | |
| 231 | if self.use_full_cuda_graph and scheduler_metadata is not None: |
| 232 | n = scheduler_metadata.shape[0] |
| 233 | # Ensure the persistent buffer is large enough |
| 234 | assert n <= self.scheduler_metadata.shape[0], ( |
| 235 | f"Scheduler metadata size {n} exceeds buffer size " |
| 236 | f"{self.scheduler_metadata.shape[0]}" |
| 237 | ) |
| 238 | self.scheduler_metadata[:n] = scheduler_metadata |
| 239 | # NOTE(woosuk): We should zero out the rest of the scheduler |
| 240 | # metadata to guarantee the correctness. Otherwise, some thread |
| 241 | # blocks may use the invalid scheduler metadata and overwrite the |
| 242 | # output buffer. |
| 243 | self.scheduler_metadata[n:] = 0 |
| 244 | scheduler_metadata = self.scheduler_metadata[:n] |
| 245 | |
| 246 | metadata = FlashAttnMLADecodeMetadata( |
| 247 | block_table=block_table_tensor, |
| 248 | seq_lens=seq_lens_device, |
| 249 | query_start_loc=query_start_loc_device, |
| 250 | max_query_len=max_query_len, |
| 251 | max_seq_len=max_seq_len, |
| 252 | scheduler_metadata=scheduler_metadata, |
| 253 | max_num_splits=max_num_splits, |
| 254 | dcp_tot_seq_lens=dcp_tot_seq_lens_device, |
| 255 | ) |
| 256 | return metadata |
| 257 | |
| 258 | |
| 259 | class FlashAttnMLAImpl(MLACommonImpl[FlashAttnMLAMetadata]): |
| 260 | can_return_lse_for_decode: bool = True |
| 261 | supports_dcp: bool = True |
| 262 | |
| 263 | def __init__( |
| 264 | self, |
| 265 | num_heads: int, |
| 266 | head_size: int, |
| 267 | scale: float, |
| 268 | num_kv_heads: int, |
| 269 | alibi_slopes: list[float] | None, |
| 270 | sliding_window: int | None, |
| 271 | kv_cache_dtype: str, |
| 272 | logits_soft_cap: float | None, |
| 273 | attn_type: str, |
| 274 | kv_sharing_target_layer_name: str | None, |
| 275 | # MLA Specific Arguments |
| 276 | **mla_args, |
| 277 | ) -> None: |
| 278 | super().__init__( |
| 279 | num_heads, |
| 280 | head_size, |
| 281 | scale, |
| 282 | num_kv_heads, |
| 283 | alibi_slopes, |
| 284 | sliding_window, |
| 285 | kv_cache_dtype, |
| 286 | logits_soft_cap, |
| 287 | attn_type, |
| 288 | kv_sharing_target_layer_name, |
| 289 | **mla_args, |
| 290 | ) |
| 291 | |
| 292 | assert flash_attn_supports_mla(), "FlashAttnMLA is not supported on this device" |
| 293 | |
| 294 | unsupported_features = [alibi_slopes, sliding_window, logits_soft_cap] |
| 295 | if any(unsupported_features): |
| 296 | raise NotImplementedError( |
| 297 | "FlashAttnMLAImpl does not support one of the following: " |
| 298 | "alibi_slopes, sliding_window, logits_soft_cap" |
| 299 | ) |
| 300 | |
| 301 | if attn_type != AttentionType.DECODER: |
| 302 | raise NotImplementedError( |
| 303 | "Encoder self-attention and " |
| 304 | "encoder/decoder cross-attention " |
| 305 | "are not implemented for " |
| 306 | "FlashAttnMLAImpl" |
| 307 | ) |
| 308 | |
| 309 | if is_quantized_kv_cache(self.kv_cache_dtype): |
| 310 | raise NotImplementedError( |
| 311 | "FlashAttnMLA V1 with FP8 KV cache not yet supported" |
| 312 | ) |
| 313 | |
| 314 | def forward_mqa( |
| 315 | self, |
| 316 | q: torch.Tensor | tuple[torch.Tensor, torch.Tensor], |
| 317 | kv_c_and_k_pe_cache: torch.Tensor, |
| 318 | attn_metadata: FlashAttnMLAMetadata, |
| 319 | layer: AttentionLayer, |
| 320 | ) -> tuple[torch.Tensor, torch.Tensor | None]: |
| 321 | assert kv_c_and_k_pe_cache.numel() > 0 |
| 322 | assert attn_metadata.decode is not None |
| 323 | |
| 324 | if type(q) is tuple: |
| 325 | q_nope, q_pe = q |
| 326 | else: |
| 327 | q_nope, q_pe = torch.split( |
| 328 | q, [self.kv_lora_rank, self.qk_rope_head_dim], dim=-1 |
| 329 | ) |
| 330 | |
| 331 | if is_quantized_kv_cache(self.kv_cache_dtype): |
| 332 | raise NotImplementedError("FP8 FlashAttention MLA not yet supported") |
| 333 | |
| 334 | kv_c_cache = kv_c_and_k_pe_cache[..., : self.kv_lora_rank] |
| 335 | k_pe_cache = kv_c_and_k_pe_cache[..., self.kv_lora_rank :] |
| 336 | |
| 337 | # NOTE(matt): During CUDA graph capture, max_query_len can be 0, but the |
| 338 | # kernel uses this to calculate grid dimensions. Ensure it's at least 1 |
| 339 | # to prevent invalid grid configuration during graph capture. |
| 340 | max_seqlen_q = max(attn_metadata.decode.max_query_len, 1) |
| 341 | |
| 342 | attn_out = flash_attn_varlen_func( |
| 343 | q=q_pe, |
| 344 | k=k_pe_cache.unsqueeze(-2), # Add head dim of 1 |
| 345 | v=kv_c_cache.unsqueeze(-2), # Add head dim of 1 |
| 346 | q_v=q_nope, |
| 347 | max_seqlen_q=max_seqlen_q, |
| 348 | cu_seqlens_q=attn_metadata.decode.query_start_loc, |
| 349 | max_seqlen_k=attn_metadata.decode.max_seq_len, |
| 350 | seqused_k=attn_metadata.decode.seq_lens, |
| 351 | block_table=attn_metadata.decode.block_table, |
| 352 | softmax_scale=self.scale, |
| 353 | causal=True, |
| 354 | return_softmax_lse=self.need_to_return_lse_for_decode, |
| 355 | fa_version=3, # only version 3 is supported |
| 356 | scheduler_metadata=attn_metadata.decode.scheduler_metadata, |
| 357 | num_splits=attn_metadata.decode.max_num_splits, |
| 358 | cp_world_size=self.dcp_world_size, |
| 359 | cp_rank=self.dcp_rank, |
| 360 | cp_tot_seqused_k=attn_metadata.decode.dcp_tot_seq_lens, |
| 361 | ) |
| 362 | |
| 363 | if self.need_to_return_lse_for_decode: |
| 364 | o, lse = attn_out |
| 365 | # FA returns LSE in shape [ H, B ] but DCP wants [ B, H ] |
| 366 | return o, lse.transpose(0, 1) # [ H, B ] -> [ B, H ] |
| 367 | else: |
| 368 | o = attn_out |
| 369 | return o, None |