| 1 | # SPDX-License-Identifier: Apache-2.0 |
| 2 | # SPDX-FileCopyrightText: Copyright contributors to the vLLM project |
| 3 | """Shared MHA implementation and metadata builder for sparse MLA backends.""" |
| 4 | |
| 5 | import math |
| 6 | from dataclasses import dataclass |
| 7 | from shutil import which |
| 8 | from typing import TYPE_CHECKING, Any, ClassVar, Generic, TypeVar |
| 9 | |
| 10 | import numpy as np |
| 11 | import torch |
| 12 | |
| 13 | from vllm import _custom_ops as ops |
| 14 | from vllm.config import get_current_vllm_config |
| 15 | from vllm.distributed import ( |
| 16 | get_dcp_group, |
| 17 | get_tensor_model_parallel_world_size, |
| 18 | ) |
| 19 | from vllm.logger import init_logger |
| 20 | from vllm.model_executor.layers.attention.mla_attention import ( |
| 21 | MLACommonBaseImpl, |
| 22 | MLACommonDecodeMetadata, |
| 23 | MLACommonMetadata, |
| 24 | MLACommonPrefillMetadata, |
| 25 | accumulate_mla_context_chunk, |
| 26 | align_mla_chunked_context_workspace_size, |
| 27 | build_mla_chunked_context_metadata, |
| 28 | get_mla_dims, |
| 29 | init_mla_context_partial, |
| 30 | ) |
| 31 | from vllm.platforms import current_platform |
| 32 | from vllm.triton_utils import tl, triton |
| 33 | from vllm.utils.flashinfer import has_flashinfer |
| 34 | from vllm.utils.torch_utils import ( |
| 35 | PIN_MEMORY, |
| 36 | is_quantized_kv_cache, |
| 37 | np_to_pinned_tensor, |
| 38 | ) |
| 39 | from vllm.v1.attention.backend import AttentionMetadataBuilder |
| 40 | from vllm.v1.attention.backends.fa_utils import get_flash_attn_version |
| 41 | from vllm.v1.attention.backends.mla.index_group import ( |
| 42 | SparseMLAIndexGroup, |
| 43 | SparseMLAIndexGroupBuilder, |
| 44 | ) |
| 45 | from vllm.v1.attention.backends.utils import split_decodes_and_prefills |
| 46 | from vllm.v1.attention.ops.dcp import MLADCPManager |
| 47 | from vllm.v1.attention.ops.merge_attn_states import merge_attn_states |
| 48 | from vllm.v1.hisparse.runtime import ( |
| 49 | HiSparsePrefillStagingPlan, |
| 50 | build_hisparse_prefill_staging_plan, |
| 51 | ) |
| 52 | |
| 53 | if TYPE_CHECKING: |
| 54 | from vllm.config import VllmConfig |
| 55 | from vllm.model_executor.layers.linear import ColumnParallelLinear |
| 56 | from vllm.v1.attention.backend import CommonAttentionMetadata |
| 57 | from vllm.v1.kv_cache_interface import AttentionSpec |
| 58 | |
| 59 | logger = init_logger(__name__) |
| 60 | |
| 61 | GLOBAL_TOPK_MASK_MAX_BYTES = 128 * 1024 * 1024 # 128 MiB |
| 62 | |
| 63 | |
| 64 | def _topk_mask_shape( |
| 65 | batch_size: int, |
| 66 | max_query_len: int, |
| 67 | max_key_len: int, |
| 68 | reserve_key_starts_word: bool = False, |
| 69 | ) -> tuple[int, int, int]: |
| 70 | """Shape of a bit-packed top-k mask, shared by every site that builds one.""" |
| 71 | tile_m = 128 if max_query_len <= 128 else 256 |
| 72 | padded_q_len = triton.cdiv(max_query_len, tile_m) * tile_m |
| 73 | num_words = triton.cdiv(max_key_len, 32) + int(reserve_key_starts_word) |
| 74 | num_words = triton.cdiv(num_words, 4) * 4 |
| 75 | return batch_size, padded_q_len, num_words |
| 76 | |
| 77 | |
| 78 | def _masked_mha_workspace_fits( |
| 79 | batch_size: int, |
| 80 | max_query_len: int, |
| 81 | max_context_chunk_seq_len: int, |
| 82 | workspace_numel: int, |
| 83 | ) -> bool: |
| 84 | """Return whether the suffix and per-context-chunk masks fit the workspace. |
| 85 | |
| 86 | The global mask is excluded: it always needs more, and has its own check. |
| 87 | """ |
| 88 | max_key_len = max(max_query_len, max_context_chunk_seq_len) |
| 89 | needed = math.prod(_topk_mask_shape(batch_size, max_query_len, max_key_len)) |
| 90 | return needed <= workspace_numel |
| 91 | |
| 92 | |
| 93 | def _is_masked_mha_available( |
| 94 | num_heads_total: int, |
| 95 | kv_lora_rank: int, |
| 96 | qk_nope_head_dim: int, |
| 97 | qk_rope_head_dim: int, |
| 98 | v_head_dim: int, |
| 99 | kv_cache_dtype: str, |
| 100 | ) -> bool: |
| 101 | """Check if masked MHA can ever fire for this model configuration.""" |
| 102 | if not current_platform.is_device_capability_family(100): |
| 103 | return False |
| 104 | model_dims = ( |
| 105 | num_heads_total, |
| 106 | kv_lora_rank, |
| 107 | qk_nope_head_dim, |
| 108 | qk_rope_head_dim, |
| 109 | v_head_dim, |
| 110 | ) |
| 111 | if model_dims not in ( |
| 112 | (128, 512, 128, 64, 128), |
| 113 | (64, 512, 192, 64, 256), |
| 114 | # GLM-5.3-Flash: NoPE, qk_head_dim 256 == the (192, 64, 256) kernel. |
| 115 | (64, 512, 256, 0, 256), |
| 116 | ): |
| 117 | return False |
| 118 | qk_head_dim = qk_nope_head_dim + qk_rope_head_dim |
| 119 | if qk_head_dim == 256 and v_head_dim == 256: |
| 120 | # This path uses contiguous K/V, so it does not need the paged-KV |
| 121 | # features that keep head-dim 256 disabled in the general FA selector. |
| 122 | fa_version = get_flash_attn_version() |
| 123 | else: |
| 124 | fa_version = get_flash_attn_version( |
| 125 | head_size=qk_head_dim, head_size_v=v_head_dim |
| 126 | ) |
| 127 | return fa_version == 4 and not is_quantized_kv_cache(kv_cache_dtype) |
| 128 | |
| 129 | |
| 130 | def _use_dense_mha_prefill( |
| 131 | vllm_config: "VllmConfig", prefill_max_seq_len: int, topk_tokens: int |
| 132 | ) -> bool: |
| 133 | attention_config = vllm_config.attention_config |
| 134 | return ( |
| 135 | attention_config.hisparse_config is None |
| 136 | and not attention_config.sparse_mla_force_mqa |
| 137 | and prefill_max_seq_len <= topk_tokens |
| 138 | ) |
| 139 | |
| 140 | |
| 141 | @dataclass |
| 142 | class SparseMLAPrefillMetadata(MLACommonPrefillMetadata): |
| 143 | host_staging_plan: HiSparsePrefillStagingPlan | None = None |
| 144 | |
| 145 | |
| 146 | @dataclass(kw_only=True) |
| 147 | class SparseMLACommonMetadata(MLACommonMetadata[MLACommonDecodeMetadata]): |
| 148 | block_table: torch.Tensor |
| 149 | req_id_per_token: torch.Tensor |
| 150 | seq_lens: torch.Tensor | None = None |
| 151 | block_size: int = 64 |
| 152 | topk_tokens: int = 2048 |
| 153 | num_decodes: int = 0 |
| 154 | num_prefills: int = 0 |
| 155 | num_decode_tokens: int = 0 |
| 156 | decode_max_query_len: int = 0 |
| 157 | prefill_max_seq_len: int = 0 |
| 158 | prefill: SparseMLAPrefillMetadata | None = None |
| 159 | cp_kv_cache_interleave_size: int = 1 |
| 160 | |
| 161 | |
| 162 | T = TypeVar("T", bound=SparseMLACommonMetadata) |
| 163 | |
| 164 | |
| 165 | class SparseMLACommonMetadataBuilder(AttentionMetadataBuilder[T]): |
| 166 | metadata_cls: type[T] |
| 167 | require_uniform_decodes: ClassVar[bool] = False |
| 168 | |
| 169 | def __init__( |
| 170 | self, |
| 171 | kv_cache_spec: "AttentionSpec", |
| 172 | layer_names: list[str], |
| 173 | vllm_config: "VllmConfig", |
| 174 | device: torch.device, |
| 175 | ) -> None: |
| 176 | super().__init__(kv_cache_spec, layer_names, vllm_config, device) |
| 177 | self.vllm_config = vllm_config |
| 178 | self.device = device |
| 179 | self.model_config = vllm_config.model_config |
| 180 | self.mla_dims = get_mla_dims(self.model_config) |
| 181 | self.topk_tokens: int = vllm_config.model_config.hf_text_config.index_topk |
| 182 | self.req_id_per_token_buffer = torch.empty( |
| 183 | (vllm_config.scheduler_config.max_num_batched_tokens,), |
| 184 | dtype=torch.int32, |
| 185 | device=device, |
| 186 | ) |
| 187 | self._context_lens_cpu = torch.empty( |
| 188 | vllm_config.scheduler_config.max_num_seqs, |
| 189 | dtype=torch.int32, |
| 190 | device="cpu", |
| 191 | pin_memory=PIN_MEMORY, |
| 192 | ) |
| 193 | parallel_config = vllm_config.parallel_config |
| 194 | self.pcp_world_size = parallel_config.prefill_context_parallel_size |
| 195 | self.use_pcp = self.pcp_world_size > 1 |
| 196 | try: |
| 197 | self.dcp_world_size = get_dcp_group().world_size |
| 198 | except AssertionError: |
| 199 | # DCP might not be initialized in testing |
| 200 | self.dcp_world_size = 1 |
| 201 | self.cp_kv_cache_interleave_size = parallel_config.cp_kv_cache_interleave_size |
| 202 | self.dcp_local_block_size = self.cp_kv_cache_interleave_size |
| 203 | self.dcp_virtual_block_size = self.dcp_local_block_size * self.dcp_world_size |
| 204 | |
| 205 | self.chunked_prefill_workspace_size = ( |
| 206 | self.determine_chunked_prefill_workspace_size(vllm_config) |
| 207 | ) |
| 208 | workspace_head_size = ( |
| 209 | self.mla_dims.kv_lora_rank + self.mla_dims.qk_rope_head_dim |
| 210 | ) |
| 211 | workspace_rows = self.chunked_prefill_workspace_size |
| 212 | if self.dcp_world_size > 1: |
| 213 | # DCP gathers each rank's local KV shard into the workspace, so it |
| 214 | # needs an extra 1/DCP rows beyond the TP allocation. |
| 215 | assert self.chunked_prefill_workspace_size % self.dcp_world_size == 0 |
| 216 | workspace_rows += self.chunked_prefill_workspace_size // self.dcp_world_size |
| 217 | self.chunked_prefill_workspace = torch.empty( |
| 218 | (workspace_rows, workspace_head_size), |
| 219 | dtype=self.model_config.dtype, |
| 220 | device=device, |
| 221 | ) |
| 222 | self.topk_mask_workspace: torch.Tensor | None = None |
| 223 | if _is_masked_mha_available( |
| 224 | self.model_config.model_arch_config.total_num_attention_heads, |
| 225 | self.mla_dims.kv_lora_rank, |
| 226 | self.mla_dims.qk_nope_head_dim, |
| 227 | self.mla_dims.qk_rope_head_dim, |
| 228 | self.mla_dims.v_head_dim, |
| 229 | vllm_config.cache_config.cache_dtype, |
| 230 | ): |
| 231 | self.topk_mask_workspace = torch.zeros( |
| 232 | GLOBAL_TOPK_MASK_MAX_BYTES // torch.int32.itemsize, |
| 233 | dtype=torch.int32, |
| 234 | device=device, |
| 235 | ) |
| 236 | attention_layer = vllm_config.compilation_config.static_forward_context[ |
| 237 | layer_names[0] |
| 238 | ] |
| 239 | layer_prefill_backend = attention_layer.prefill_backend |
| 240 | self.dcp_manager: MLADCPManager | None = None |
| 241 | if self.dcp_world_size > 1: |
| 242 | self.dcp_manager = getattr(attention_layer, "dcp_manager", None) |
| 243 | assert isinstance(self.dcp_manager, MLADCPManager) |
| 244 | if layer_prefill_backend is not None: |
| 245 | self.dcp_manager.init_kv_gather( |
| 246 | self.chunked_prefill_workspace, |
| 247 | self.chunked_prefill_workspace_size, |
| 248 | ) |
| 249 | self._prefill_backend = ( |
| 250 | layer_prefill_backend.clone() if layer_prefill_backend is not None else None |
| 251 | ) |
| 252 | |
| 253 | def _init_reorder_batch_threshold( |
| 254 | self, |
| 255 | reorder_batch_threshold: int | None = 1, |
| 256 | supports_spec_as_decode: bool = False, |
| 257 | supports_dcp_with_varlen: bool = False, |
| 258 | ) -> None: |
| 259 | if self.vllm_config.attention_config.hisparse_config is not None: |
| 260 | reorder_batch_threshold = 1 |
| 261 | super()._init_reorder_batch_threshold( |
| 262 | reorder_batch_threshold, |
| 263 | supports_spec_as_decode, |
| 264 | supports_dcp_with_varlen, |
| 265 | ) |
| 266 | |
| 267 | @staticmethod |
| 268 | def determine_chunked_prefill_workspace_size(vllm_config: "VllmConfig") -> int: |
| 269 | scheduler_config = vllm_config.scheduler_config |
| 270 | cache_config = vllm_config.cache_config |
| 271 | model_config = vllm_config.model_config |
| 272 | topk_tokens = model_config.hf_text_config.index_topk |
| 273 | |
| 274 | workspace_size = min( |
| 275 | max( |
| 276 | 8 * model_config.max_model_len, |
| 277 | 4 * scheduler_config.max_num_seqs * cache_config.block_size, |
| 278 | ), |
| 279 | 64 * 1024, |
| 280 | scheduler_config.max_num_seqs * topk_tokens, |
| 281 | ) |
| 282 | workspace_size = max(workspace_size, cache_config.block_size) |
| 283 | if vllm_config.parallel_config.decode_context_parallel_size > 1: |
| 284 | return align_mla_chunked_context_workspace_size(vllm_config, workspace_size) |
| 285 | return workspace_size |
| 286 | |
| 287 | def _build_req_id_per_token( |
| 288 | self, |
| 289 | common_attn_metadata: "CommonAttentionMetadata", |
| 290 | ) -> torch.Tensor: |
| 291 | return common_attn_metadata.token_to_req_indices(self.req_id_per_token_buffer) |
| 292 | |
| 293 | def _build_chunked_context_fields( |
| 294 | self, |
| 295 | common_attn_metadata: "CommonAttentionMetadata", |
| 296 | num_decodes: int, |
| 297 | num_prefills: int, |
| 298 | prefill_query_lens_cpu: torch.Tensor | None, |
| 299 | ) -> "MLACommonPrefillMetadata.ChunkedContextMetadata | None": |
| 300 | if num_prefills == 0 or prefill_query_lens_cpu is None: |
| 301 | return None |
| 302 | |
| 303 | seq_lens_cpu = common_attn_metadata.seq_lens_cpu_upper_bound |
| 304 | assert seq_lens_cpu is not None |
| 305 | context_lens_cpu = self._context_lens_cpu[:num_prefills] |
| 306 | torch.subtract( |
| 307 | seq_lens_cpu[num_decodes : num_decodes + num_prefills], |
| 308 | prefill_query_lens_cpu, |
| 309 | out=context_lens_cpu, |
| 310 | ) |
| 311 | qsl_cpu = common_attn_metadata.query_start_loc_cpu |
| 312 | prefill_query_start_loc_cpu = qsl_cpu[num_decodes:] - qsl_cpu[num_decodes] |
| 313 | |
| 314 | return build_mla_chunked_context_metadata( |
| 315 | context_lens_cpu=context_lens_cpu, |
| 316 | prefill_query_start_loc_cpu=prefill_query_start_loc_cpu, |
| 317 | chunked_prefill_workspace=self.chunked_prefill_workspace, |
| 318 | chunked_prefill_workspace_size=self.chunked_prefill_workspace_size, |
| 319 | block_size=self.kv_cache_spec.block_size, |
| 320 | align_chunk_to_block=current_platform.is_cuda(), |
| 321 | device=self.device, |
| 322 | dcp_world_size=self.dcp_world_size, |
| 323 | dcp_local_block_size=self.dcp_local_block_size, |
| 324 | dcp_virtual_block_size=self.dcp_virtual_block_size, |
| 325 | dcp_manager=self.dcp_manager, |
| 326 | ) |
| 327 | |
| 328 | def build( |
| 329 | self, |
| 330 | common_prefix_len: int, |
| 331 | common_attn_metadata: "CommonAttentionMetadata", |
| 332 | fast_build: bool = False, |
| 333 | ) -> T: |
| 334 | req_id_per_token = self._build_req_id_per_token(common_attn_metadata) |
| 335 | |
| 336 | num_decodes, num_prefills, num_decode_tokens, _ = split_decodes_and_prefills( |
| 337 | common_attn_metadata, |
| 338 | decode_threshold=self.reorder_batch_threshold or 1, |
| 339 | treat_short_extends_as_decodes=not self.use_pcp, |
| 340 | require_uniform=self.require_uniform_decodes, |
| 341 | ) |
| 342 | ( |
| 343 | prefill_query_start_loc, |
| 344 | prefill_max_query_len, |
| 345 | prefill_query_lens_cpu, |
| 346 | ) = self._build_prefill_fields(common_attn_metadata, num_decodes, num_prefills) |
| 347 | decode_max_query_len = 0 |
| 348 | if num_decodes: |
| 349 | if common_attn_metadata.max_logits_per_req is not None: |
| 350 | decode_max_query_len = common_attn_metadata.max_logits_per_req |
| 351 | elif num_prefills == 0: |
| 352 | decode_max_query_len = common_attn_metadata.max_query_len |
| 353 | else: |
| 354 | decode_query_lens_cpu = torch.diff( |
| 355 | common_attn_metadata.query_start_loc_cpu[: num_decodes + 1] |
| 356 | ) |
| 357 | decode_max_query_len = int(decode_query_lens_cpu.max().item()) |
| 358 | |
| 359 | prefill_max_seq_len = 0 |
| 360 | prefill: SparseMLAPrefillMetadata | None = None |
| 361 | if num_prefills > 0 and self._prefill_backend is not None: |
| 362 | seq_lens_cpu = common_attn_metadata.seq_lens_cpu_upper_bound |
| 363 | assert seq_lens_cpu is not None |
| 364 | prefill_max_seq_len = int( |
| 365 | seq_lens_cpu[num_decodes : num_decodes + num_prefills].max().item() |
| 366 | ) |
| 367 | block_table = common_attn_metadata.block_table_tensor[num_decodes:, ...] |
| 368 | chunked_context = self._build_chunked_context_fields( |
| 369 | common_attn_metadata, |
| 370 | num_decodes, |
| 371 | num_prefills, |
| 372 | prefill_query_lens_cpu, |
| 373 | ) |
| 374 | staging_plan = None |
| 375 | if self.vllm_config.attention_config.hisparse_config is not None: |
| 376 | prefill_seq_lens_cpu = seq_lens_cpu[ |
| 377 | num_decodes : num_decodes + num_prefills |
| 378 | ] |
| 379 | staging_block_capacity = int( |
| 380 | ( |
| 381 | (prefill_seq_lens_cpu + self.kv_cache_spec.block_size - 1) |
| 382 | // self.kv_cache_spec.block_size |
| 383 | ).sum() |
| 384 | ) |
| 385 | staging_plan = build_hisparse_prefill_staging_plan( |
| 386 | block_table, |
| 387 | common_attn_metadata.seq_lens[num_decodes:], |
| 388 | self.kv_cache_spec.block_size, |
| 389 | staging_block_capacity, |
| 390 | ) |
| 391 | if chunked_context is not None: |
| 392 | block_table = staging_plan.block_table |
| 393 | prefill = SparseMLAPrefillMetadata( |
| 394 | block_table=block_table, |
| 395 | query_start_loc=prefill_query_start_loc, |
| 396 | max_query_len=prefill_max_query_len, |
| 397 | query_lens_cpu=prefill_query_lens_cpu, |
| 398 | chunked_context=chunked_context, |
| 399 | q_data_type=self.model_config.dtype, |
| 400 | output_dtype=self.model_config.dtype, |
| 401 | prefill_backend=self._prefill_backend, |
| 402 | use_dense_mha=( |
| 403 | _use_dense_mha_prefill( |
| 404 | self.vllm_config, prefill_max_seq_len, self.topk_tokens |
| 405 | ) |
| 406 | # Dense MHA cannot be used under PCP+DCP. |
| 407 | and not (self.use_pcp and self.dcp_world_size > 1) |
| 408 | ), |
| 409 | topk_mask_workspace=self.topk_mask_workspace, |
| 410 | host_staging_plan=staging_plan, |
| 411 | ) |
| 412 | self._prefill_backend.prepare_metadata(prefill) |
| 413 | |
| 414 | return self.metadata_cls( # type: ignore[call-arg] |
| 415 | num_reqs=common_attn_metadata.num_reqs, |
| 416 | max_query_len=common_attn_metadata.max_query_len, |
| 417 | max_seq_len=common_attn_metadata.max_seq_len, |
| 418 | num_actual_tokens=common_attn_metadata.num_actual_tokens, |
| 419 | query_start_loc=common_attn_metadata.query_start_loc, |
| 420 | slot_mapping=common_attn_metadata.slot_mapping, |
| 421 | block_table=common_attn_metadata.block_table_tensor, |
| 422 | req_id_per_token=req_id_per_token, |
| 423 | seq_lens=common_attn_metadata.seq_lens, |
| 424 | block_size=self.kv_cache_spec.block_size, |
| 425 | topk_tokens=self.topk_tokens, |
| 426 | num_decodes=num_decodes, |
| 427 | num_prefills=num_prefills, |
| 428 | num_decode_tokens=num_decode_tokens, |
| 429 | decode_max_query_len=decode_max_query_len, |
| 430 | prefill_max_seq_len=prefill_max_seq_len, |
| 431 | prefill=prefill, |
| 432 | cp_kv_cache_interleave_size=self.cp_kv_cache_interleave_size, |
| 433 | ) |
| 434 | |
| 435 | @staticmethod |
| 436 | def _build_prefill_fields( |
| 437 | common_attn_metadata: "CommonAttentionMetadata", |
| 438 | num_decodes: int, |
| 439 | num_prefills: int, |
| 440 | ) -> tuple[ |
| 441 | torch.Tensor | None, # prefill_query_start_loc |
| 442 | int, # prefill_max_query_len |
| 443 | torch.Tensor | None, # prefill_query_lens_cpu |
| 444 | ]: |
| 445 | if num_prefills == 0: |
| 446 | return None, 0, None |
| 447 | |
| 448 | offset = common_attn_metadata.query_start_loc[num_decodes] |
| 449 | prefill_query_start_loc = ( |
| 450 | common_attn_metadata.query_start_loc[num_decodes:] - offset |
| 451 | ) |
| 452 | |
| 453 | qsl_cpu = common_attn_metadata.query_start_loc_cpu |
| 454 | prefill_qsl_cpu = qsl_cpu[num_decodes:] - qsl_cpu[num_decodes] |
| 455 | prefill_query_lens = prefill_qsl_cpu[1:] - prefill_qsl_cpu[:-1] |
| 456 | prefill_max_query_len = int(prefill_query_lens.max().item()) |
| 457 | |
| 458 | return ( |
| 459 | prefill_query_start_loc, |
| 460 | prefill_max_query_len, |
| 461 | prefill_query_lens, |
| 462 | ) |
| 463 | |
| 464 | |
| 465 | @triton.jit |
| 466 | def _scatter_topk_kernel( |
| 467 | mask_ptr, |
| 468 | topk_ptr, |
| 469 | cu_q_lens_ptr, |
| 470 | num_words: tl.constexpr, |
| 471 | mask_row_stride: tl.constexpr, |
| 472 | num_topk: tl.constexpr, |
| 473 | topk_stride: tl.constexpr, |
| 474 | max_q_len: tl.constexpr, |
| 475 | BLOCK_TOPK: tl.constexpr, |
| 476 | BLOCK_WORDS: tl.constexpr, |
| 477 | ): |
| 478 | row_idx = tl.program_id(0) |
| 479 | row_ptr = mask_ptr + row_idx * mask_row_stride |
| 480 | word_offsets = tl.arange(0, BLOCK_WORDS) |
| 481 | tl.store( |
| 482 | row_ptr + word_offsets, |
| 483 | tl.zeros([BLOCK_WORDS], dtype=tl.int32), |
| 484 | mask=word_offsets < num_words, |
| 485 | ) |
| 486 | |
| 487 | req_idx = row_idx // max_q_len |
| 488 | q_local = row_idx % max_q_len |
| 489 | q_start = tl.load(cu_q_lens_ptr + req_idx) |
| 490 | q_len = tl.load(cu_q_lens_ptr + req_idx + 1) - q_start |
| 491 | if q_local < q_len: |
| 492 | src_row = q_start + q_local |
| 493 | offsets = tl.arange(0, BLOCK_TOPK) |
| 494 | in_range = offsets < num_topk |
| 495 | indices = tl.load( |
| 496 | topk_ptr + src_row * topk_stride + offsets, |
| 497 | mask=in_range, |
| 498 | other=-1, |
| 499 | ) |
| 500 | valid = in_range & (indices >= 0) |
| 501 | word_indices = indices >> 5 |
| 502 | bits = (1 << (indices & 31)).to(tl.int32) |
| 503 | tl.atomic_or(row_ptr + word_indices, bits, mask=valid) |
| 504 | |
| 505 | |
| 506 | @triton.jit |
| 507 | def _scatter_topk_single_req_kernel( |
| 508 | mask_ptr, |
| 509 | topk_ptr, |
| 510 | num_words: tl.constexpr, |
| 511 | mask_row_stride: tl.constexpr, |
| 512 | num_topk: tl.constexpr, |
| 513 | topk_stride: tl.constexpr, |
| 514 | total_q: tl.constexpr, |
| 515 | BLOCK_TOPK: tl.constexpr, |
| 516 | BLOCK_WORDS: tl.constexpr, |
| 517 | ): |
| 518 | row_idx = tl.program_id(0) |
| 519 | row_ptr = mask_ptr + row_idx * mask_row_stride |
| 520 | word_offsets = tl.arange(0, BLOCK_WORDS) |
| 521 | tl.store( |
| 522 | row_ptr + word_offsets, |
| 523 | tl.zeros([BLOCK_WORDS], dtype=tl.int32), |
| 524 | mask=word_offsets < num_words, |
| 525 | ) |
| 526 | if row_idx < total_q: |
| 527 | offsets = tl.arange(0, BLOCK_TOPK) |
| 528 | in_range = offsets < num_topk |
| 529 | indices = tl.load( |
| 530 | topk_ptr + row_idx * topk_stride + offsets, |
| 531 | mask=in_range, |
| 532 | other=-1, |
| 533 | ) |
| 534 | valid = in_range & (indices >= 0) |
| 535 | word_indices = indices >> 5 |
| 536 | bits = (1 << (indices & 31)).to(tl.int32) |
| 537 | tl.atomic_or(row_ptr + word_indices, bits, mask=valid) |
| 538 | |
| 539 | |
| 540 | def _build_topk_mask( |
| 541 | topk_indices_per_req: list[torch.Tensor], |
| 542 | q_lens: list[int], |
| 543 | max_q_len: int, |
| 544 | max_seq_len: int, |
| 545 | out: torch.Tensor, |
| 546 | ) -> torch.Tensor: |
| 547 | """Build a bit-packed top-k mask while preserving padded row storage.""" |
| 548 | batch_size = len(q_lens) |
| 549 | num_words = (max_seq_len + 31) // 32 |
| 550 | total_rows = batch_size * max_q_len |
| 551 | if total_rows == 0: |
| 552 | return out[:batch_size, :max_q_len] |
| 553 | |
| 554 | total_q = sum(q_lens) |
| 555 | mask_row_stride = out.stride(-2) |
| 556 | block_words = triton.next_power_of_2(num_words) |
| 557 | |
| 558 | if batch_size == 1: |
| 559 | topk_packed = topk_indices_per_req[0] |
| 560 | num_topk = topk_packed.shape[1] |
| 561 | _scatter_topk_single_req_kernel[(max_q_len,)]( |
| 562 | out, |
| 563 | topk_packed, |
| 564 | num_words=num_words, |
| 565 | mask_row_stride=mask_row_stride, |
| 566 | num_topk=num_topk, |
| 567 | topk_stride=topk_packed.stride(0), |
| 568 | total_q=total_q, |
| 569 | BLOCK_TOPK=triton.next_power_of_2(num_topk), |
| 570 | BLOCK_WORDS=block_words, |
| 571 | ) |
| 572 | return out[:1, :max_q_len] |
| 573 | |
| 574 | topk_packed = torch.cat(topk_indices_per_req, dim=0) |
| 575 | num_topk = topk_packed.shape[1] |
| 576 | q_lens_tensor = np_to_pinned_tensor(np.asarray(q_lens, dtype=np.int32)).to( |
| 577 | out.device, non_blocking=True |
| 578 | ) |
| 579 | cu_q_lens = out.new_zeros(batch_size + 1) |
| 580 | torch.cumsum(q_lens_tensor, dim=0, out=cu_q_lens[1:]) |
| 581 | _scatter_topk_kernel[(total_rows,)]( |
| 582 | out, |
| 583 | topk_packed, |
| 584 | cu_q_lens, |
| 585 | num_words=num_words, |
| 586 | mask_row_stride=mask_row_stride, |
| 587 | num_topk=num_topk, |
| 588 | topk_stride=topk_packed.stride(0), |
| 589 | max_q_len=max_q_len, |
| 590 | BLOCK_TOPK=triton.next_power_of_2(num_topk), |
| 591 | BLOCK_WORDS=block_words, |
| 592 | ) |
| 593 | return out[:batch_size, :max_q_len] |
| 594 | |
| 595 | |
| 596 | class SharedTopkIndicesBuffer: |
| 597 | """Resolves the shared top-k index buffer for sparse MLA implementations. |
| 598 | |
| 599 | The indexer owns the buffer, but `LLMBaseProposer.load_model` repoints the |
| 600 | draft's indexers at the target's buffer after the impls are constructed, so |
| 601 | it must be resolved per read rather than snapshotted. Backbone skip-topk |
| 602 | layers have no indexer and pass the buffer explicitly. |
| 603 | """ |
| 604 | |
| 605 | _indexer: object | None = None |
| 606 | _topk_indices_buffer: torch.Tensor | None = None |
| 607 | |
| 608 | def init_topk_indices_buffer( |
| 609 | self, |
| 610 | indexer: object | None, |
| 611 | topk_indices_buffer: torch.Tensor | None, |
| 612 | ) -> None: |
| 613 | self._indexer = indexer |
| 614 | self._topk_indices_buffer = topk_indices_buffer |
| 615 | |
| 616 | @property |
| 617 | def topk_indices_buffer(self) -> torch.Tensor | None: |
| 618 | if self._indexer is not None: |
| 619 | return self._indexer.topk_indices_buffer # type: ignore[attr-defined] |
| 620 | return self._topk_indices_buffer |
| 621 | |
| 622 | @topk_indices_buffer.setter |
| 623 | def topk_indices_buffer(self, buffer: torch.Tensor | None) -> None: |
| 624 | # An explicit assignment supersedes the indexer. |
| 625 | self._indexer = None |
| 626 | self._topk_indices_buffer = buffer |
| 627 | |
| 628 | |
| 629 | class SparseMLACommonImpl(MLACommonBaseImpl[T], SharedTopkIndicesBuffer, Generic[T]): |
| 630 | """Sparse MLA base with dense and masked-MHA prefill paths.""" |
| 631 | |
| 632 | is_sparse = True |
| 633 | |
| 634 | def __init__( |
| 635 | self, |
| 636 | num_heads: int, |
| 637 | head_size: int, |
| 638 | scale: float, |
| 639 | num_kv_heads: int, |
| 640 | alibi_slopes: list[float] | None, |
| 641 | sliding_window: int | None, |
| 642 | kv_cache_dtype: str, |
| 643 | logits_soft_cap: float | None, |
| 644 | attn_type: str, |
| 645 | kv_sharing_target_layer_name: str | None, |
| 646 | q_lora_rank: int | None, |
| 647 | kv_lora_rank: int, |
| 648 | qk_nope_head_dim: int, |
| 649 | qk_rope_head_dim: int, |
| 650 | qk_head_dim: int, |
| 651 | v_head_dim: int, |
| 652 | kv_b_proj: "ColumnParallelLinear", |
| 653 | indexer: object | None = None, |
| 654 | topk_indices_buffer: torch.Tensor | None = None, |
| 655 | index_group_builder: SparseMLAIndexGroupBuilder | None = None, |
| 656 | q_pad_num_heads: int | None = None, |
| 657 | ) -> None: |
| 658 | super().__init__( |
| 659 | num_heads, |
| 660 | head_size, |
| 661 | scale, |
| 662 | num_kv_heads, |
| 663 | kv_cache_dtype, |
| 664 | kv_lora_rank, |
| 665 | qk_nope_head_dim, |
| 666 | qk_rope_head_dim, |
| 667 | qk_head_dim, |
| 668 | v_head_dim, |
| 669 | kv_b_proj, |
| 670 | ) |
| 671 | |
| 672 | self.init_topk_indices_buffer(indexer, topk_indices_buffer) |
| 673 | self.index_group: SparseMLAIndexGroup | None = None |
| 674 | self.index_group_index = 0 |
| 675 | if index_group_builder is None and self.topk_indices_buffer is not None: |
| 676 | index_group_builder = SparseMLAIndexGroupBuilder(self.topk_indices_buffer) |
| 677 | if index_group_builder is not None: |
| 678 | vllm_config = get_current_vllm_config() |
| 679 | self.index_group, self.index_group_index = ( |
| 680 | index_group_builder.register_layer( |
| 681 | indexer is not None, |
| 682 | vllm_config, |
| 683 | head_size=head_size, |
| 684 | kv_cache_dtype=kv_cache_dtype, |
| 685 | ) |
| 686 | ) |
| 687 | |
| 688 | self._use_flashinfer_concat_mla_k = ( |
| 689 | has_flashinfer() |
| 690 | and which("ninja") is not None |
| 691 | and (self.num_heads == 128) |
| 692 | and (self.qk_nope_head_dim == 128) |
| 693 | and (self.qk_rope_head_dim == 64) |
| 694 | ) |
| 695 | self.masked_mha_available = _is_masked_mha_available( |
| 696 | num_heads_total=num_heads * get_tensor_model_parallel_world_size(), |
| 697 | kv_lora_rank=kv_lora_rank, |
| 698 | qk_nope_head_dim=qk_nope_head_dim, |
| 699 | qk_rope_head_dim=qk_rope_head_dim, |
| 700 | v_head_dim=v_head_dim, |
| 701 | kv_cache_dtype=kv_cache_dtype, |
| 702 | ) |
| 703 | |
| 704 | def record_logical_topk_ready(self) -> None: |
| 705 | if self.index_group is not None: |
| 706 | self.index_group.set_logical_topk_ready(self.index_group_index) |
| 707 | |
| 708 | def prepare_for_batch(self, attn_metadata: T | None) -> None: |
| 709 | if self.index_group is not None: |
| 710 | self.index_group.prepare_for_batch(self.index_group_index, attn_metadata) |
| 711 | |
| 712 | def _convert_logical_to_physical_topk( |
| 713 | self, |
| 714 | logical_topk_indices: torch.Tensor, |
| 715 | attn_metadata: Any, |
| 716 | *, |
| 717 | block_stride_rows: int | None, |
| 718 | return_valid_counts: bool, |
| 719 | ) -> torch.Tensor | tuple[torch.Tensor, torch.Tensor]: |
| 720 | group = self.index_group |
| 721 | assert group is not None |
| 722 | assert self.dcp_world_size == 1 |
| 723 | return group.convert_logical_to_physical_topk( |
| 724 | self.index_group_index, |
| 725 | logical_topk_indices, |
| 726 | attn_metadata, |
| 727 | block_stride_rows=block_stride_rows, |
| 728 | return_valid_counts=return_valid_counts, |
| 729 | ) |
| 730 | |
| 731 | @staticmethod |
| 732 | def masked_mha_workspace_fits(prefill: MLACommonPrefillMetadata) -> bool: |
| 733 | """Whether this prefill batch's top-k masks fit the workspace.""" |
| 734 | workspace = prefill.topk_mask_workspace |
| 735 | if workspace is None or prefill.query_lens_cpu is None: |
| 736 | return False |
| 737 | max_context_chunk_seq_len = 0 |
| 738 | if prefill.chunked_context is not None: |
| 739 | max_context_chunk_seq_len = max( |
| 740 | chunk.max_seq_len for chunk in prefill.chunked_context.chunks |
| 741 | ) |
| 742 | fits = _masked_mha_workspace_fits( |
| 743 | batch_size=len(prefill.query_lens_cpu), |
| 744 | max_query_len=prefill.max_query_len, |
| 745 | max_context_chunk_seq_len=max_context_chunk_seq_len, |
| 746 | workspace_numel=workspace.numel(), |
| 747 | ) |
| 748 | if not fits: |
| 749 | logger.warning_once( |
| 750 | "Sparse MLA top-k mask workspace (%d MiB) is too small for some " |
| 751 | "prefill batches; those fall back to slower sparse MQA.", |
| 752 | workspace.numel() * torch.int32.itemsize // (1024 * 1024), |
| 753 | ) |
| 754 | return fits |
| 755 | |
| 756 | @staticmethod |
| 757 | def _slice_topk_per_req( |
| 758 | topk_all: torch.Tensor, |
| 759 | q_lens: list[int], |
| 760 | ) -> list[torch.Tensor]: |
| 761 | topk_per_req = [] |
| 762 | offset = 0 |
| 763 | for q_len in q_lens: |
| 764 | topk_per_req.append(topk_all[offset : offset + q_len]) |
| 765 | offset += q_len |
| 766 | return topk_per_req |
| 767 | |
| 768 | @staticmethod |
| 769 | def _remap_topk_to_ranges( |
| 770 | topk_per_req: list[torch.Tensor], |
| 771 | range_starts: list[int] | torch.Tensor, |
| 772 | range_lens: list[int], |
| 773 | ) -> list[torch.Tensor]: |
| 774 | remapped = [] |
| 775 | for topk, start, length in zip(topk_per_req, range_starts, range_lens): |
| 776 | valid = (topk >= start) & (topk < start + length) |
| 777 | remapped.append(torch.where(valid, topk - start, -1)) |
| 778 | return remapped |
| 779 | |
| 780 | def _project_kv( |
| 781 | self, kv_c_normed: torch.Tensor, k_pe: torch.Tensor |
| 782 | ) -> tuple[torch.Tensor, torch.Tensor]: |
| 783 | kv_nope = self.kv_b_proj(kv_c_normed)[0].view( |
| 784 | -1, |
| 785 | self.num_heads, |
| 786 | self.qk_nope_head_dim + self.v_head_dim, |
| 787 | ) |
| 788 | k_nope, v = kv_nope.split([self.qk_nope_head_dim, self.v_head_dim], dim=-1) |
| 789 | return self._concat_k_nope_k_pe(k_nope, k_pe), v |
| 790 | |
| 791 | @staticmethod |
| 792 | def _try_build_global_mask( |
| 793 | topk_per_req: list[torch.Tensor], |
| 794 | q_lens: list[int], |
| 795 | max_query_len: int, |
| 796 | max_seq_len: int, |
| 797 | topk_mask_workspace: torch.Tensor, |
| 798 | ) -> torch.Tensor | None: |
| 799 | """Build a full-sequence top-k mask if it fits within the budget. |
| 800 | |
| 801 | When the mask fits, it is reused across the suffix and all context |
| 802 | chunks, avoiding per-chunk mask rebuilds. Returns None when the |
| 803 | mask is too large, signalling the caller to fall back to per-chunk |
| 804 | index remapping. |
| 805 | """ |
| 806 | batch_size, padded_q_len, num_words_padded = _topk_mask_shape( |
| 807 | len(q_lens), |
| 808 | max_query_len, |
| 809 | max_seq_len, |
| 810 | reserve_key_starts_word=True, |
| 811 | ) |
| 812 | needed = batch_size * padded_q_len * num_words_padded |
| 813 | if needed > topk_mask_workspace.numel(): |
| 814 | return None |
| 815 | |
| 816 | mask = topk_mask_workspace[:needed].view( |
| 817 | batch_size, padded_q_len, num_words_padded |
| 818 | ) |
| 819 | _build_topk_mask( |
| 820 | topk_per_req, |
| 821 | q_lens, |
| 822 | padded_q_len, |
| 823 | max_seq_len, |
| 824 | mask, |
| 825 | ) |
| 826 | return mask |
| 827 | |
| 828 | def _run_masked_mha( |
| 829 | self, |
| 830 | q: torch.Tensor, |
| 831 | k: torch.Tensor, |
| 832 | v: torch.Tensor, |
| 833 | cu_seqlens_q: torch.Tensor, |
| 834 | cu_seqlens_k: torch.Tensor, |
| 835 | max_seqlen_q: int, |
| 836 | max_seqlen_k: int, |
| 837 | topk_per_req: list[torch.Tensor], |
| 838 | q_lens: list[int], |
| 839 | causal: bool, |
| 840 | return_softmax_lse: bool = False, |
| 841 | dense_mask: torch.Tensor | None = None, |
| 842 | key_starts: torch.Tensor | None = None, |
| 843 | topk_mask_workspace: torch.Tensor | None = None, |
| 844 | ) -> torch.Tensor | tuple[torch.Tensor, torch.Tensor]: |
| 845 | from vllm.model_executor.layers.attention.sparse_mla_mask import ( |
| 846 | dense_mask_mod, |
| 847 | offset_dense_mask_mod, |
| 848 | ) |
| 849 | from vllm.vllm_flash_attn import flash_attn_varlen_func |
| 850 | |
| 851 | if dense_mask is None: |
| 852 | assert topk_mask_workspace is not None |
| 853 | batch_size, padded_q_len, num_words = _topk_mask_shape( |
| 854 | len(q_lens), max_seqlen_q, max_seqlen_k |
| 855 | ) |
| 856 | words_needed = batch_size * padded_q_len * num_words |
| 857 | if words_needed > topk_mask_workspace.numel(): |
| 858 | raise ValueError( |
| 859 | f"Sparse MLA top-k mask needs {words_needed} int32 words (batch=" |
| 860 | f"{len(q_lens)}, q={max_seqlen_q}, k={max_seqlen_k}) but the " |
| 861 | f"workspace holds {topk_mask_workspace.numel()}." |
| 862 | ) |
| 863 | workspace_3d = topk_mask_workspace[:words_needed].view( |
| 864 | batch_size, padded_q_len, num_words |
| 865 | ) |
| 866 | dense_mask = _build_topk_mask( |
| 867 | topk_per_req, |
| 868 | q_lens, |
| 869 | padded_q_len, |
| 870 | max_seqlen_k, |
| 871 | workspace_3d, |
| 872 | ) |
| 873 | if key_starts is not None: |
| 874 | dense_mask[:, 0, -1].copy_(key_starts) |
| 875 | kwargs = { |
| 876 | "q": q, |
| 877 | "k": k, |
| 878 | "v": v, |
| 879 | "cu_seqlens_q": cu_seqlens_q, |
| 880 | "cu_seqlens_k": cu_seqlens_k, |
| 881 | "max_seqlen_q": max_seqlen_q, |
| 882 | "max_seqlen_k": max_seqlen_k, |
| 883 | "softmax_scale": self.scale, |
| 884 | "return_softmax_lse": return_softmax_lse, |
| 885 | "fa_version": 4, |
| 886 | "mask_mod": dense_mask_mod if key_starts is None else offset_dense_mask_mod, |
| 887 | "aux_tensors": [dense_mask], |
| 888 | "aux_tensor_leading_dims": [2], |
| 889 | "causal": causal, |
| 890 | } |
| 891 | |
| 892 | return flash_attn_varlen_func(**kwargs) |
| 893 | |
| 894 | def _compute_context_mha( |
| 895 | self, |
| 896 | q: torch.Tensor, |
| 897 | kv_c_and_k_pe_cache: torch.Tensor, |
| 898 | prefill_metadata: MLACommonPrefillMetadata, |
| 899 | k_scale: torch.Tensor, |
| 900 | q_lens: list[int], |
| 901 | topk_per_req: list[torch.Tensor], |
| 902 | dense_mask: torch.Tensor | None = None, |
| 903 | ) -> tuple[torch.Tensor, torch.Tensor]: |
| 904 | if self.dcp_world_size > 1: |
| 905 | raise NotImplementedError( |
| 906 | "Masked MHA with context does not yet support decode context " |
| 907 | "parallelism" |
| 908 | ) |
| 909 | |
| 910 | chunked_context = prefill_metadata.chunked_context |
| 911 | assert chunked_context is not None |
| 912 | output: torch.Tensor | None = None |
| 913 | output_lse: torch.Tensor | None = None |
| 914 | workspace = chunked_context.workspace |
| 915 | |
| 916 | for chunk in chunked_context.chunks: |
| 917 | toks = chunk.num_context_tokens |
| 918 | requests = chunk.request_slice |
| 919 | ops.gather_and_maybe_dequant_cache( |
| 920 | src_cache=kv_c_and_k_pe_cache, |
| 921 | dst=workspace, |
| 922 | block_table=prefill_metadata.block_table[requests], |
| 923 | cu_seq_lens=chunk.cu_seq_lens, |
| 924 | token_to_seq=chunk.token_to_seq, |
| 925 | num_tokens=toks, |
| 926 | kv_cache_dtype=self.kv_cache_dtype, |
| 927 | scale=k_scale, |
| 928 | seq_starts=chunk.starts, |
| 929 | ) |
| 930 | |
| 931 | chunk_kv_c = workspace[:toks, : self.kv_lora_rank] |
| 932 | chunk_k_pe = workspace[:toks, self.kv_lora_rank :].unsqueeze(1) |
| 933 | k, v = self._project_kv(chunk_kv_c, chunk_k_pe) |
| 934 | if dense_mask is not None: |
| 935 | chunk_mask: torch.Tensor | None = dense_mask[requests] |
| 936 | chunk_topk = topk_per_req[requests] |
| 937 | key_starts: torch.Tensor | None = chunk.starts |
| 938 | else: |
| 939 | chunk_mask = None |
| 940 | chunk_topk = self._remap_topk_to_ranges( |
| 941 | topk_per_req[requests], |
| 942 | chunk.starts, |
| 943 | chunk.seq_lens.tolist(), |
| 944 | ) |
| 945 | key_starts = None |
| 946 | attn_out, lse = self._run_masked_mha( |
| 947 | q=q[chunk.token_slice], |
| 948 | k=k, |
| 949 | v=v, |
| 950 | cu_seqlens_q=chunk.query_start_loc, |
| 951 | cu_seqlens_k=chunk.cu_seq_lens, |
| 952 | max_seqlen_q=chunk.max_query_len, |
| 953 | max_seqlen_k=chunk.max_seq_len, |
| 954 | topk_per_req=chunk_topk, |
| 955 | q_lens=q_lens[requests], |
| 956 | causal=False, |
| 957 | return_softmax_lse=True, |
| 958 | dense_mask=chunk_mask, |
| 959 | key_starts=key_starts, |
| 960 | topk_mask_workspace=prefill_metadata.topk_mask_workspace, |
| 961 | ) |
| 962 | |
| 963 | if output is None: |
| 964 | if ( |
| 965 | len(chunked_context.chunks) == 1 |
| 966 | and not chunked_context.empty_token_slices |
| 967 | ): |
| 968 | return attn_out, lse |
| 969 | output, output_lse = init_mla_context_partial( |
| 970 | chunked_context, |
| 971 | attn_out, |
| 972 | lse, |
| 973 | num_tokens=q.shape[0], |
| 974 | ) |
| 975 | accumulate_mla_context_chunk(chunk, attn_out, lse, output, output_lse) |
| 976 | |
| 977 | assert output is not None and output_lse is not None |
| 978 | return output, output_lse |
| 979 | |
| 980 | def forward_mha( # type: ignore[override] |
| 981 | self, |
| 982 | q: torch.Tensor, |
| 983 | kv_c_normed: torch.Tensor, |
| 984 | k_pe: torch.Tensor, |
| 985 | kv_c_and_k_pe_cache: torch.Tensor, |
| 986 | attn_metadata: T, |
| 987 | k_scale: torch.Tensor, |
| 988 | output: torch.Tensor, |
| 989 | output_scale: torch.Tensor | None = None, |
| 990 | ) -> None: |
| 991 | prefill_metadata = attn_metadata.prefill |
| 992 | assert prefill_metadata is not None |
| 993 | prefill_max_seq_len = attn_metadata.prefill_max_seq_len |
| 994 | topk_tokens = attn_metadata.topk_tokens |
| 995 | force_dense = getattr(self, "_sparse_mla_force_dense_mha", False) |
| 996 | force_masked = getattr(self, "_sparse_mla_force_masked_mha", False) |
| 997 | if force_dense or (prefill_max_seq_len <= topk_tokens and not force_masked): |
| 998 | return super().forward_mha( |
| 999 | q, |
| 1000 | kv_c_normed, |
| 1001 | k_pe, |
| 1002 | kv_c_and_k_pe_cache, |
| 1003 | attn_metadata, |
| 1004 | k_scale, |
| 1005 | output, |
| 1006 | output_scale, |
| 1007 | ) |
| 1008 | |
| 1009 | assert output_scale is None |
| 1010 | assert self.masked_mha_available |
| 1011 | assert prefill_metadata.query_lens_cpu is not None |
| 1012 | assert self.topk_indices_buffer is not None |
| 1013 | |
| 1014 | q_lens = prefill_metadata.query_lens_cpu.tolist() |
| 1015 | num_decode_tokens = attn_metadata.num_decode_tokens |
| 1016 | topk_all = self.topk_indices_buffer[ |
| 1017 | num_decode_tokens : num_decode_tokens + q.shape[0] |
| 1018 | ] |
| 1019 | topk_per_req = self._slice_topk_per_req(topk_all, q_lens) |
| 1020 | |
| 1021 | k, v = self._project_kv(kv_c_normed, k_pe) |
| 1022 | chunked_context = prefill_metadata.chunked_context |
| 1023 | if chunked_context is None: |
| 1024 | attn_out = self._run_masked_mha( |
| 1025 | q=q, |
| 1026 | k=k, |
| 1027 | v=v, |
| 1028 | cu_seqlens_q=prefill_metadata.query_start_loc, |
| 1029 | cu_seqlens_k=prefill_metadata.query_start_loc, |
| 1030 | max_seqlen_q=prefill_metadata.max_query_len, |
| 1031 | max_seqlen_k=prefill_metadata.max_query_len, |
| 1032 | topk_per_req=topk_per_req, |
| 1033 | q_lens=q_lens, |
| 1034 | causal=True, |
| 1035 | topk_mask_workspace=prefill_metadata.topk_mask_workspace, |
| 1036 | ) |
| 1037 | assert isinstance(attn_out, torch.Tensor) |
| 1038 | output.copy_(attn_out[..., : self.v_head_dim].flatten(start_dim=-2)) |
| 1039 | return |
| 1040 | |
| 1041 | context_lens = chunked_context.context_lens_list |
| 1042 | dense_mask = self._try_build_global_mask( |
| 1043 | topk_per_req, |
| 1044 | q_lens, |
| 1045 | prefill_metadata.max_query_len, |
| 1046 | prefill_max_seq_len, |
| 1047 | prefill_metadata.topk_mask_workspace, |
| 1048 | ) |
| 1049 | if dense_mask is not None: |
| 1050 | suffix_topk = topk_per_req |
| 1051 | else: |
| 1052 | suffix_topk = self._remap_topk_to_ranges(topk_per_req, context_lens, q_lens) |
| 1053 | suffix_output, suffix_lse = self._run_masked_mha( |
| 1054 | q=q, |
| 1055 | k=k, |
| 1056 | v=v, |
| 1057 | cu_seqlens_q=prefill_metadata.query_start_loc, |
| 1058 | cu_seqlens_k=prefill_metadata.query_start_loc, |
| 1059 | max_seqlen_q=prefill_metadata.max_query_len, |
| 1060 | max_seqlen_k=prefill_metadata.max_query_len, |
| 1061 | topk_per_req=suffix_topk, |
| 1062 | q_lens=q_lens, |
| 1063 | causal=True, |
| 1064 | return_softmax_lse=True, |
| 1065 | dense_mask=dense_mask, |
| 1066 | key_starts=( |
| 1067 | chunked_context.context_lens if dense_mask is not None else None |
| 1068 | ), |
| 1069 | topk_mask_workspace=prefill_metadata.topk_mask_workspace, |
| 1070 | ) |
| 1071 | context_output, context_lse = self._compute_context_mha( |
| 1072 | q=q, |
| 1073 | kv_c_and_k_pe_cache=kv_c_and_k_pe_cache, |
| 1074 | prefill_metadata=prefill_metadata, |
| 1075 | k_scale=k_scale, |
| 1076 | q_lens=q_lens, |
| 1077 | topk_per_req=topk_per_req, |
| 1078 | dense_mask=dense_mask, |
| 1079 | ) |
| 1080 | merge_attn_states( |
| 1081 | output=output.view(-1, self.num_heads, self.v_head_dim), |
| 1082 | prefix_output=context_output[..., : self.v_head_dim], |
| 1083 | prefix_lse=context_lse, |
| 1084 | suffix_output=suffix_output[..., : self.v_head_dim], |
| 1085 | suffix_lse=suffix_lse, |
| 1086 | ) |