| 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.platform_utils import num_compute_units |
| 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.utils import ( |
| 31 | reshape_attn_output_for_spec_decode, |
| 32 | reshape_query_for_spec_decode, |
| 33 | ) |
| 34 | from vllm.v1.attention.ops.flashmla import ( |
| 35 | FlashMLASchedMeta, |
| 36 | flash_mla_with_kvcache, |
| 37 | flash_mla_with_kvcache_fp8, |
| 38 | get_mla_metadata, |
| 39 | get_mla_metadata_dense_fp8, |
| 40 | is_flashmla_dense_supported, |
| 41 | ) |
| 42 | from vllm.v1.kv_cache_interface import AttentionSpec |
| 43 | |
| 44 | logger = init_logger(__name__) |
| 45 | |
| 46 | |
| 47 | class FlashMLABackend(MLACommonBackend): |
| 48 | supported_dtypes: ClassVar[list[torch.dtype]] = [torch.float16, torch.bfloat16] |
| 49 | supported_kv_cache_dtypes: ClassVar[list[CacheDType]] = [ |
| 50 | "auto", |
| 51 | "float16", |
| 52 | "bfloat16", |
| 53 | "fp8", |
| 54 | "fp8_e4m3", |
| 55 | ] |
| 56 | |
| 57 | @staticmethod |
| 58 | def get_supported_kernel_block_sizes(kv_cache_spec=None) -> list[int | MultipleOf]: |
| 59 | return [64] |
| 60 | |
| 61 | @staticmethod |
| 62 | def get_name() -> str: |
| 63 | return "FLASHMLA" |
| 64 | |
| 65 | @staticmethod |
| 66 | def get_builder_cls() -> type["FlashMLAMetadataBuilder"]: |
| 67 | return FlashMLAMetadataBuilder |
| 68 | |
| 69 | @staticmethod |
| 70 | def get_impl_cls() -> type["FlashMLAImpl"]: |
| 71 | return FlashMLAImpl |
| 72 | |
| 73 | @classmethod |
| 74 | def supports_compute_capability(cls, capability: DeviceCapability) -> bool: |
| 75 | return capability.major in [9, 10] |
| 76 | |
| 77 | @classmethod |
| 78 | def supports_combination( |
| 79 | cls, |
| 80 | head_size: int, |
| 81 | dtype: torch.dtype, |
| 82 | kv_cache_dtype: CacheDType | None, |
| 83 | block_size: int | None, |
| 84 | use_mla: bool, |
| 85 | has_sink: bool, |
| 86 | use_sparse: bool, |
| 87 | use_mm_prefix: bool, |
| 88 | device_capability: DeviceCapability, |
| 89 | ) -> str | None: |
| 90 | if use_sparse: |
| 91 | from vllm.v1.attention.ops.flashmla import is_flashmla_sparse_supported |
| 92 | |
| 93 | return is_flashmla_sparse_supported()[1] |
| 94 | else: |
| 95 | from vllm.v1.attention.ops.flashmla import is_flashmla_dense_supported |
| 96 | |
| 97 | return is_flashmla_dense_supported()[1] |
| 98 | |
| 99 | |
| 100 | @dataclass |
| 101 | class FlashMLADecodeMetadata(MLACommonDecodeMetadata): |
| 102 | scheduler_metadata: FlashMLASchedMeta |
| 103 | |
| 104 | |
| 105 | @dataclass |
| 106 | class FlashMLAMetadata(MLACommonMetadata[FlashMLADecodeMetadata]): |
| 107 | pass |
| 108 | |
| 109 | |
| 110 | class FlashMLAMetadataBuilder(MLACommonMetadataBuilder[FlashMLAMetadata]): |
| 111 | _cudagraph_support: ClassVar[AttentionCGSupport] = AttentionCGSupport.UNIFORM_BATCH |
| 112 | query_len_support: ClassVar[QueryLenSupport] = QueryLenSupport.UNIFORM |
| 113 | reorder_batch_threshold: int = 128 # process small prefills with decode pathway |
| 114 | # ^ TODO(matt): tune this |
| 115 | |
| 116 | def __init__( |
| 117 | self, |
| 118 | kv_cache_spec: AttentionSpec, |
| 119 | layer_names: list[str], |
| 120 | vllm_config: VllmConfig, |
| 121 | device: torch.device, |
| 122 | ): |
| 123 | super().__init__( |
| 124 | kv_cache_spec, layer_names, vllm_config, device, FlashMLAMetadata |
| 125 | ) |
| 126 | |
| 127 | self.num_q_heads = vllm_config.model_config.get_num_attention_heads( |
| 128 | vllm_config.parallel_config |
| 129 | ) |
| 130 | |
| 131 | self.cg_buf_tile_scheduler_metadata = None |
| 132 | self.cg_buf_num_splits = None |
| 133 | self.is_fp8_kvcache = is_quantized_kv_cache( |
| 134 | vllm_config.cache_config.cache_dtype |
| 135 | ) |
| 136 | |
| 137 | num_sms = num_compute_units(self.device.index) |
| 138 | |
| 139 | if self.compilation_config.cudagraph_mode.has_full_cudagraphs(): |
| 140 | self.cg_buf_tile_scheduler_metadata = torch.zeros( |
| 141 | # Upper bound on size (<= #SMs, TileSchedulerMetaDataSize) |
| 142 | # TileSchedulerMetaDataSize = 8 |
| 143 | (num_sms, 8), |
| 144 | device=self.device, |
| 145 | dtype=torch.int32, |
| 146 | ) |
| 147 | self.cg_buf_num_splits = torch.empty( |
| 148 | (vllm_config.scheduler_config.max_num_seqs + 1), |
| 149 | device=self.device, |
| 150 | dtype=torch.int32, |
| 151 | ) |
| 152 | |
| 153 | def _build_decode( |
| 154 | self, |
| 155 | block_table_tensor: torch.Tensor, |
| 156 | seq_lens_device: torch.Tensor, |
| 157 | max_seq_len: int, |
| 158 | query_start_loc_cpu: torch.Tensor, |
| 159 | query_start_loc_device: torch.Tensor, |
| 160 | num_decode_tokens: int, |
| 161 | max_query_len: int, |
| 162 | dcp_tot_seq_lens_device: torch.Tensor | None, |
| 163 | ) -> FlashMLADecodeMetadata: |
| 164 | query_lens_cpu = query_start_loc_cpu[1:] - query_start_loc_cpu[:-1] |
| 165 | # we use the max but all should be the same due to uniform length requirement |
| 166 | max_query_len = query_lens_cpu.max().item() |
| 167 | num_q_heads = self.num_q_heads |
| 168 | if self.dcp_world_size > 1: |
| 169 | num_q_heads *= self.dcp_world_size |
| 170 | num_q_tokens_per_head_k = max_query_len * num_q_heads // 1 |
| 171 | scheduler_metadata, _ = get_mla_metadata( |
| 172 | seq_lens_device, |
| 173 | num_q_tokens_per_head_k, |
| 174 | 1, # MQA for the decode path |
| 175 | is_fp8_kvcache=self.is_fp8_kvcache, |
| 176 | ) |
| 177 | if self.is_fp8_kvcache: |
| 178 | tile_scheduler_metadata, num_splits = get_mla_metadata_dense_fp8( |
| 179 | seq_lens_device, |
| 180 | num_q_tokens_per_head_k, |
| 181 | 1, # MQA for the decode path |
| 182 | ) |
| 183 | |
| 184 | # Copy FP8 metadata into persistent CUDA graph buffers |
| 185 | if self.compilation_config.cudagraph_mode.has_full_cudagraphs(): |
| 186 | assert self.cg_buf_tile_scheduler_metadata is not None |
| 187 | assert self.cg_buf_num_splits is not None |
| 188 | n = tile_scheduler_metadata.size(0) |
| 189 | assert n <= self.cg_buf_tile_scheduler_metadata.size(0) |
| 190 | self.cg_buf_tile_scheduler_metadata[:n].copy_(tile_scheduler_metadata) |
| 191 | tile_scheduler_metadata = self.cg_buf_tile_scheduler_metadata[:n] |
| 192 | |
| 193 | n = num_splits.size(0) |
| 194 | assert n <= self.cg_buf_num_splits.size(0) |
| 195 | self.cg_buf_num_splits[:n].copy_(num_splits) |
| 196 | num_splits = self.cg_buf_num_splits[:n] |
| 197 | |
| 198 | scheduler_metadata.tile_scheduler_metadata = tile_scheduler_metadata |
| 199 | scheduler_metadata.num_splits = num_splits |
| 200 | |
| 201 | return FlashMLADecodeMetadata( |
| 202 | block_table=block_table_tensor, |
| 203 | seq_lens=seq_lens_device, |
| 204 | scheduler_metadata=scheduler_metadata, |
| 205 | dcp_tot_seq_lens=dcp_tot_seq_lens_device, |
| 206 | ) |
| 207 | |
| 208 | |
| 209 | class FlashMLAImpl(MLACommonImpl[FlashMLAMetadata]): |
| 210 | can_return_lse_for_decode: bool = True |
| 211 | supports_dcp: bool = True |
| 212 | |
| 213 | def __init__( |
| 214 | self, |
| 215 | num_heads: int, |
| 216 | head_size: int, |
| 217 | scale: float, |
| 218 | num_kv_heads: int, |
| 219 | alibi_slopes: list[float] | None, |
| 220 | sliding_window: int | None, |
| 221 | kv_cache_dtype: str, |
| 222 | logits_soft_cap: float | None, |
| 223 | attn_type: str, |
| 224 | kv_sharing_target_layer_name: str | None, |
| 225 | # MLA Specific Arguments |
| 226 | **mla_args, |
| 227 | ) -> None: |
| 228 | super().__init__( |
| 229 | num_heads, |
| 230 | head_size, |
| 231 | scale, |
| 232 | num_kv_heads, |
| 233 | alibi_slopes, |
| 234 | sliding_window, |
| 235 | kv_cache_dtype, |
| 236 | logits_soft_cap, |
| 237 | attn_type, |
| 238 | kv_sharing_target_layer_name, |
| 239 | **mla_args, |
| 240 | ) |
| 241 | |
| 242 | is_supported, reason = is_flashmla_dense_supported() |
| 243 | assert is_supported, reason |
| 244 | |
| 245 | unsupported_features = [alibi_slopes, sliding_window, logits_soft_cap] |
| 246 | if any(unsupported_features): |
| 247 | raise NotImplementedError( |
| 248 | "FlashMLAImpl does not support one of the following: " |
| 249 | "alibi_slopes, sliding_window, logits_soft_cap" |
| 250 | ) |
| 251 | |
| 252 | if attn_type != AttentionType.DECODER: |
| 253 | raise NotImplementedError( |
| 254 | "Encoder self-attention and " |
| 255 | "encoder/decoder cross-attention " |
| 256 | "are not implemented for " |
| 257 | "FlashMLAImpl" |
| 258 | ) |
| 259 | |
| 260 | def forward_mqa( |
| 261 | self, |
| 262 | q: torch.Tensor | tuple[torch.Tensor, torch.Tensor], |
| 263 | kv_c_and_k_pe_cache: torch.Tensor, |
| 264 | attn_metadata: FlashMLAMetadata, |
| 265 | layer: AttentionLayer, |
| 266 | ) -> tuple[torch.Tensor, torch.Tensor | None]: |
| 267 | # TODO: (zyongye) decode function for mla here |
| 268 | assert kv_c_and_k_pe_cache.numel() > 0 |
| 269 | assert attn_metadata.decode is not None |
| 270 | |
| 271 | if type(q) is tuple: |
| 272 | q = torch.cat(q, dim=-1) |
| 273 | |
| 274 | # mypy assertion: q is now always a tensor |
| 275 | assert isinstance(q, torch.Tensor) |
| 276 | |
| 277 | num_decodes = attn_metadata.num_decodes |
| 278 | q = reshape_query_for_spec_decode(q, num_decodes) |
| 279 | |
| 280 | scheduler_metadata = attn_metadata.decode.scheduler_metadata |
| 281 | if envs.VLLM_BATCH_INVARIANT and not is_quantized_kv_cache(self.kv_cache_dtype): |
| 282 | device = q.device |
| 283 | dtype = torch.int32 |
| 284 | |
| 285 | B = q.shape[0] |
| 286 | # block_table shape: [batch_size, max_num_blocks_per_seq] |
| 287 | # The number of blocks per sequence is in the second dimension |
| 288 | topk = attn_metadata.decode.block_table.shape[-1] |
| 289 | B_TOPK = 64 |
| 290 | assert topk % B_TOPK == 0, f"topk ({topk}) must be divisible by {B_TOPK}" |
| 291 | end_block_idx = topk // B_TOPK |
| 292 | |
| 293 | # Single partition => num_sm_parts = 1 |
| 294 | # TileSchedulerMetaDataSize = 8, layout: |
| 295 | # [begin_idx, begin_block_idx, end_idx, end_block_idx, |
| 296 | # begin_n_split_idx, _, _, _] |
| 297 | tile_scheduler_metadata = torch.zeros((1, 8), dtype=dtype, device=device) |
| 298 | tile_scheduler_metadata[0, 0] = 0 # begin_idx |
| 299 | tile_scheduler_metadata[0, 1] = 0 # sched_begin_block_idx |
| 300 | tile_scheduler_metadata[0, 2] = B - 1 # end_idx |
| 301 | tile_scheduler_metadata[0, 3] = end_block_idx |
| 302 | tile_scheduler_metadata[0, 4] = 0 # begin_n_split_idx |
| 303 | # fields [5..7] stay 0 |
| 304 | |
| 305 | # Non-split path ignores num_splits, but the API requires it: |
| 306 | # zeros of length B+1 |
| 307 | num_splits = torch.zeros((B + 1,), dtype=dtype, device=device) |
| 308 | scheduler_metadata.tile_scheduler_metadata = tile_scheduler_metadata |
| 309 | scheduler_metadata.num_splits = num_splits |
| 310 | |
| 311 | if is_quantized_kv_cache(self.kv_cache_dtype): |
| 312 | o, lse = flash_mla_with_kvcache_fp8( |
| 313 | q=q, |
| 314 | k_cache=kv_c_and_k_pe_cache.unsqueeze(-2), # Add head dim of 1 |
| 315 | block_table=attn_metadata.decode.block_table, |
| 316 | cache_seqlens=attn_metadata.decode.seq_lens, |
| 317 | head_dim_v=self.kv_lora_rank, |
| 318 | tile_scheduler_metadata=scheduler_metadata.tile_scheduler_metadata, |
| 319 | num_splits=scheduler_metadata.num_splits, |
| 320 | softmax_scale=self.scale, |
| 321 | causal=True, |
| 322 | descale_q=layer._q_scale.reshape(1), |
| 323 | descale_k=layer._k_scale.reshape(1), |
| 324 | ) |
| 325 | else: |
| 326 | o, lse = flash_mla_with_kvcache( |
| 327 | q=q, |
| 328 | k_cache=kv_c_and_k_pe_cache.unsqueeze(-2), # Add head dim of 1 |
| 329 | block_table=attn_metadata.decode.block_table, |
| 330 | cache_seqlens=attn_metadata.decode.seq_lens, |
| 331 | head_dim_v=self.kv_lora_rank, |
| 332 | tile_scheduler_metadata=scheduler_metadata, |
| 333 | softmax_scale=self.scale, |
| 334 | causal=True, |
| 335 | is_fp8_kvcache=False, |
| 336 | ) |
| 337 | |
| 338 | o = reshape_attn_output_for_spec_decode(o) |
| 339 | |
| 340 | if self.need_to_return_lse_for_decode: |
| 341 | # FlashMLA returns LSE as [batch, heads, seq_len]; the DCP reducer |
| 342 | # consumes [tokens, heads]. Flattening matters under spec-decode, |
| 343 | # where seq_len > 1. Only DCP consumes lse, so skip the copy |
| 344 | # otherwise. |
| 345 | num_decodes, q_num_heads, seq_len = lse.shape |
| 346 | lse = ( |
| 347 | lse.permute(0, 2, 1) |
| 348 | .reshape(num_decodes * seq_len, q_num_heads) |
| 349 | .contiguous() |
| 350 | ) |
| 351 | |
| 352 | return o, lse |