| 1 | # SPDX-License-Identifier: Apache-2.0 |
| 2 | # SPDX-FileCopyrightText: Copyright contributors to the vLLM project |
| 3 | from collections.abc import Iterator |
| 4 | from dataclasses import dataclass, replace |
| 5 | from typing import TYPE_CHECKING |
| 6 | |
| 7 | import numpy as np |
| 8 | import torch |
| 9 | |
| 10 | from vllm.config import CUDAGraphMode, VllmConfig |
| 11 | from vllm.distributed.parallel_state import get_dcp_group, get_pcp_group |
| 12 | from vllm.logger import init_logger |
| 13 | from vllm.utils.torch_utils import async_tensor_h2d |
| 14 | from vllm.v1.attention.backends.utils import PAD_SLOT_ID, get_dcp_local_seq_lens |
| 15 | from vllm.v1.worker.gpu.block_table import BlockTables |
| 16 | from vllm.v1.worker.gpu.input_batch import ( |
| 17 | InputBatch, |
| 18 | InputBuffers, |
| 19 | ) |
| 20 | |
| 21 | if TYPE_CHECKING: |
| 22 | from vllm.v1.worker.gpu.cudagraph_utils import BatchExecutionDescriptor |
| 23 | |
| 24 | logger = init_logger(__name__) |
| 25 | |
| 26 | |
| 27 | @dataclass(frozen=True) |
| 28 | class RankSegment: |
| 29 | global_batch_req_idx: int |
| 30 | global_batch_slice: slice |
| 31 | rank_local_batch_slice: slice |
| 32 | |
| 33 | @property |
| 34 | def num_tokens(self) -> int: |
| 35 | return self.global_batch_slice.stop - self.global_batch_slice.start |
| 36 | |
| 37 | |
| 38 | class PCPManager: |
| 39 | """MRV2 PC batch manager. |
| 40 | |
| 41 | The model runner keeps the global scheduled batch. This manager rewrites only |
| 42 | the per-step InputBatch into rank-local DualChunkSwap rows and keeps the |
| 43 | global-batch view private to restore to the global batch shape before |
| 44 | sampling/postprocess. |
| 45 | """ |
| 46 | |
| 47 | def __init__( |
| 48 | self, |
| 49 | pcp_world_size: int, |
| 50 | pcp_rank: int, |
| 51 | device: torch.device, |
| 52 | shard_decode_requests: bool, |
| 53 | max_num_reqs: int | None = None, |
| 54 | max_num_tokens: int | None = None, |
| 55 | block_tables: BlockTables | None = None, |
| 56 | dcp_world_size: int = 1, |
| 57 | dcp_rank: int = 0, |
| 58 | cp_interleave: int = 1, |
| 59 | ) -> None: |
| 60 | self.pcp_world_size = pcp_world_size |
| 61 | self.pcp_rank = pcp_rank |
| 62 | self.device = device |
| 63 | self.dcp_world_size = dcp_world_size |
| 64 | self.dcp_rank = dcp_rank |
| 65 | self.cp_interleave = cp_interleave |
| 66 | self.shard_decode_requests = shard_decode_requests |
| 67 | |
| 68 | self._global_batch: InputBatch | None = None |
| 69 | self._local_batch: InputBatch | None = None |
| 70 | self._local_gather_idx: torch.Tensor | None = None |
| 71 | self.draft_prefill_batch: InputBatch | None = None |
| 72 | self._block_tables = block_tables |
| 73 | self._hidden_restore_idx: torch.Tensor | None = None |
| 74 | self._padded_gather_idx: torch.Tensor | None = None |
| 75 | self._gathered_kv_write_mask: torch.Tensor | None = None |
| 76 | self._pad_slot_id = torch.tensor(PAD_SLOT_ID, dtype=torch.int64, device=device) |
| 77 | |
| 78 | max_num_local_reqs = 2 * max_num_reqs if max_num_reqs is not None else None |
| 79 | self._input_buffers = ( |
| 80 | InputBuffers(max_num_local_reqs, max_num_tokens, device) |
| 81 | if max_num_local_reqs is not None and max_num_tokens is not None |
| 82 | else None |
| 83 | ) |
| 84 | self._local_block_tables: tuple[torch.Tensor, ...] | None |
| 85 | self._local_block_table_ptrs: torch.Tensor | None |
| 86 | if block_tables is not None and max_num_local_reqs is not None: |
| 87 | self._local_block_tables = tuple( |
| 88 | table.new_zeros((max_num_local_reqs, table.shape[1])) |
| 89 | for table in block_tables.input_block_tables |
| 90 | ) |
| 91 | self._local_block_table_ptrs = torch.tensor( |
| 92 | [table.data_ptr() for table in self._local_block_tables], |
| 93 | dtype=torch.uint64, |
| 94 | device=device, |
| 95 | ) |
| 96 | else: |
| 97 | self._local_block_tables = None |
| 98 | self._local_block_table_ptrs = None |
| 99 | num_kv_cache_groups = ( |
| 100 | block_tables.num_kv_cache_groups if block_tables is not None else 0 |
| 101 | ) |
| 102 | self._global_batch_slot_mappings = ( |
| 103 | torch.empty( |
| 104 | num_kv_cache_groups, |
| 105 | max_num_tokens, |
| 106 | dtype=torch.int64, |
| 107 | device=device, |
| 108 | ) |
| 109 | if max_num_tokens is not None and num_kv_cache_groups > 0 |
| 110 | else None |
| 111 | ) |
| 112 | self._gathered_kv_slot_mappings = ( |
| 113 | torch.empty( |
| 114 | num_kv_cache_groups, |
| 115 | max_num_tokens * pcp_world_size, |
| 116 | dtype=torch.int64, |
| 117 | device=device, |
| 118 | ) |
| 119 | if max_num_tokens is not None and num_kv_cache_groups > 0 |
| 120 | else None |
| 121 | ) |
| 122 | |
| 123 | @staticmethod |
| 124 | def validate_config( |
| 125 | vllm_config: VllmConfig, |
| 126 | supports_mm_inputs: bool, |
| 127 | ) -> None: |
| 128 | parallel_config = vllm_config.parallel_config |
| 129 | model_config = vllm_config.model_config |
| 130 | pcp_size = parallel_config.prefill_context_parallel_size |
| 131 | if pcp_size <= 1: |
| 132 | return |
| 133 | |
| 134 | if not model_config.use_mla: |
| 135 | raise NotImplementedError("MRV2 PCP currently supports MLA models only.") |
| 136 | if model_config.is_encoder_decoder: |
| 137 | raise NotImplementedError( |
| 138 | "MRV2 PCP does not support encoder-decoder models yet." |
| 139 | ) |
| 140 | if supports_mm_inputs: |
| 141 | raise NotImplementedError("MRV2 PCP does not support MM inputs yet.") |
| 142 | if vllm_config.lora_config is not None: |
| 143 | raise NotImplementedError("MRV2 PCP does not support LoRA yet.") |
| 144 | speculative_config = vllm_config.speculative_config |
| 145 | if speculative_config is not None: |
| 146 | if speculative_config.use_dspark(): |
| 147 | dcp_size = parallel_config.decode_context_parallel_size |
| 148 | if ( |
| 149 | dcp_size not in (1, pcp_size) |
| 150 | and speculative_config.draft_model_config.use_mla |
| 151 | ): |
| 152 | raise NotImplementedError( |
| 153 | "MRV2 PCP DSpark requires DCP=1 or DCP=PCP; got " |
| 154 | f"DCP={dcp_size}, PCP={pcp_size}." |
| 155 | ) |
| 156 | elif ( |
| 157 | speculative_config.method != "mtp" |
| 158 | or speculative_config.use_multi_module_mtp() |
| 159 | ): |
| 160 | raise NotImplementedError( |
| 161 | "MRV2 PCP only supports DSpark or single-module MTP " |
| 162 | "speculative decoding." |
| 163 | ) |
| 164 | cudagraph_mode = vllm_config.compilation_config.cudagraph_mode |
| 165 | is_sparse_mla = hasattr(model_config.hf_text_config, "index_topk") |
| 166 | if parallel_config.decode_context_parallel_size > 1 and not is_sparse_mla: |
| 167 | # Dense MLA prefill sizes its DCP KV gather from each rank's own |
| 168 | # chunk rows, so the ranks' collectives diverge (#53573). |
| 169 | raise NotImplementedError("MRV2 PCP + DCP supports sparse MLA models only.") |
| 170 | if ( |
| 171 | is_sparse_mla |
| 172 | and parallel_config.decode_context_parallel_size == 1 |
| 173 | and cudagraph_mode != CUDAGraphMode.NONE |
| 174 | ): |
| 175 | raise NotImplementedError( |
| 176 | "MRV2 sparse MLA PCP does not support CUDA graphs yet. " |
| 177 | "Set -cc.cudagraph_mode=NONE." |
| 178 | ) |
| 179 | if ( |
| 180 | cudagraph_mode.has_full_cudagraphs() |
| 181 | and not cudagraph_mode.separate_routine() |
| 182 | ): |
| 183 | raise NotImplementedError( |
| 184 | "MRV2 PCP supports full CUDA graphs for decode-only routines. " |
| 185 | "Use FULL_DECODE_ONLY, FULL_AND_PIECEWISE, PIECEWISE, or NONE." |
| 186 | ) |
| 187 | if ( |
| 188 | parallel_config.decode_context_parallel_size > 1 |
| 189 | and parallel_config.dcp_comm_backend != "ag_rs" |
| 190 | ): |
| 191 | raise NotImplementedError( |
| 192 | "MRV2 PCP + DCP requires dcp_comm_backend='ag_rs'; got " |
| 193 | f"'{parallel_config.dcp_comm_backend}'." |
| 194 | ) |
| 195 | |
| 196 | @staticmethod |
| 197 | def _reorder_segments( |
| 198 | segments: list[RankSegment], is_prefilling: np.ndarray |
| 199 | ) -> list[RankSegment]: |
| 200 | """Order this rank's rows decodes-first, then prefills, canonically.""" |
| 201 | |
| 202 | def sort_key(segment: RankSegment) -> tuple[bool, int, int]: |
| 203 | req_idx = segment.global_batch_req_idx |
| 204 | return ( |
| 205 | bool(is_prefilling[req_idx]), |
| 206 | req_idx, |
| 207 | segment.global_batch_slice.start, |
| 208 | ) |
| 209 | |
| 210 | segments.sort(key=sort_key) |
| 211 | rank_offset = 0 |
| 212 | for index, segment in enumerate(segments): |
| 213 | segments[index] = replace( |
| 214 | segment, |
| 215 | rank_local_batch_slice=slice( |
| 216 | rank_offset, rank_offset + segment.num_tokens |
| 217 | ), |
| 218 | ) |
| 219 | rank_offset += segment.num_tokens |
| 220 | return segments |
| 221 | |
| 222 | def replicated_requests( |
| 223 | self, num_scheduled_tokens: np.ndarray, is_prefilling: np.ndarray |
| 224 | ) -> np.ndarray: |
| 225 | """Per global request, whether every PCP rank gets the whole query.""" |
| 226 | num_chunks = 2 * self.pcp_world_size |
| 227 | query_lens = np.asarray(num_scheduled_tokens, dtype=np.int64) |
| 228 | replicated = ~np.asarray(is_prefilling, dtype=np.bool_) |
| 229 | if self.dcp_world_size > 1: |
| 230 | chunk_sizes = (query_lens + num_chunks - 1) // num_chunks |
| 231 | drops_a_chunk = (num_chunks - 1) * chunk_sizes >= query_lens |
| 232 | replicated |= drops_a_chunk |
| 233 | return replicated |
| 234 | |
| 235 | def _iter_rank_chunks( |
| 236 | self, |
| 237 | rank: int, |
| 238 | num_scheduled_tokens: np.ndarray, |
| 239 | is_prefilling: np.ndarray, |
| 240 | ) -> Iterator[tuple[int, int, int]]: |
| 241 | """Yield ``(request index, query offset, length)`` for one PCP rank. |
| 242 | |
| 243 | PCP=4 partitions each prefill into eight chunks: |
| 244 | |
| 245 | full: | 0 | 1 | 2 | 3 | 4 | 5 | 6 | 7 | |
| 246 | rank 0: 0 7 |
| 247 | rank 1: 1 6 |
| 248 | rank 2: 2 5 |
| 249 | rank 3: 3 4 |
| 250 | |
| 251 | Decodes, and prefills too short to fill all eight, are replicated instead. |
| 252 | """ |
| 253 | decode_ordinal = 0 |
| 254 | num_chunks = 2 * self.pcp_world_size |
| 255 | replicated = self.replicated_requests(num_scheduled_tokens, is_prefilling) |
| 256 | for global_batch_req_idx, num_tokens in enumerate(num_scheduled_tokens): |
| 257 | query_len = int(num_tokens) |
| 258 | if query_len == 0: |
| 259 | continue |
| 260 | chunk_indices: tuple[int, ...] |
| 261 | if not replicated[global_batch_req_idx]: |
| 262 | chunk_size = (query_len + num_chunks - 1) // num_chunks |
| 263 | chunk_indices = (rank, num_chunks - 1 - rank) |
| 264 | elif self.shard_decode_requests: |
| 265 | chunk_size = query_len |
| 266 | # KV and hidden states are gathered back to every PCP rank, so |
| 267 | # decode ownership does not need to persist across steps. Use a |
| 268 | # compact decode-only ordinal to keep each step exactly |
| 269 | # balanced even when prefills and zero-token rows are present. |
| 270 | owner_rank = decode_ordinal % self.pcp_world_size |
| 271 | decode_ordinal += 1 |
| 272 | chunk_indices = (0,) if rank == owner_rank else () |
| 273 | else: # DCP requires decode queries on every participating rank. |
| 274 | chunk_size = query_len |
| 275 | chunk_indices = (0,) |
| 276 | |
| 277 | for chunk_idx in chunk_indices: |
| 278 | chunk_offset = chunk_idx * chunk_size |
| 279 | chunk_len = min(chunk_size, query_len - chunk_offset) |
| 280 | if chunk_len <= 0: |
| 281 | continue |
| 282 | yield global_batch_req_idx, chunk_offset, chunk_len |
| 283 | |
| 284 | def _get_rank_segments( |
| 285 | self, |
| 286 | rank: int, |
| 287 | num_scheduled_tokens: np.ndarray, |
| 288 | is_prefilling: np.ndarray, |
| 289 | query_start_loc_np: np.ndarray, |
| 290 | ) -> list[RankSegment]: |
| 291 | rank_segments = [] |
| 292 | rank_offset = 0 |
| 293 | for global_batch_req_idx, chunk_offset, chunk_len in self._iter_rank_chunks( |
| 294 | rank, num_scheduled_tokens, is_prefilling |
| 295 | ): |
| 296 | global_batch_start = int(query_start_loc_np[global_batch_req_idx]) |
| 297 | chunk_start = global_batch_start + chunk_offset |
| 298 | rank_segments.append( |
| 299 | RankSegment( |
| 300 | global_batch_req_idx=global_batch_req_idx, |
| 301 | global_batch_slice=slice(chunk_start, chunk_start + chunk_len), |
| 302 | rank_local_batch_slice=slice(rank_offset, rank_offset + chunk_len), |
| 303 | ) |
| 304 | ) |
| 305 | rank_offset += chunk_len |
| 306 | return self._reorder_segments(rank_segments, is_prefilling) |
| 307 | |
| 308 | def _build_batch_layout( |
| 309 | self, |
| 310 | num_scheduled_tokens: np.ndarray, |
| 311 | num_computed_tokens: np.ndarray, |
| 312 | is_prefilling: np.ndarray, |
| 313 | query_start_loc_np: np.ndarray, |
| 314 | padded_num_tokens: int | None = None, |
| 315 | ) -> tuple[list[list[RankSegment]], list[int]]: |
| 316 | replicated = self.replicated_requests(num_scheduled_tokens, is_prefilling) |
| 317 | segments_by_rank = [] |
| 318 | per_rank_num_tokens = [] |
| 319 | for rank in range(self.pcp_world_size): |
| 320 | segments = self._get_rank_segments( |
| 321 | rank, |
| 322 | num_scheduled_tokens, |
| 323 | is_prefilling, |
| 324 | query_start_loc_np, |
| 325 | ) |
| 326 | num_rank_tokens = sum(segment.num_tokens for segment in segments) |
| 327 | segments_by_rank.append(segments) |
| 328 | per_rank_num_tokens.append(num_rank_tokens) |
| 329 | |
| 330 | # PCP=2 example: |
| 331 | # global batch: [A B C D E F G] |
| 332 | # rank 0 / rank 1: [A B G] / [C D E F] |
| 333 | # padded gathered: [A B G _ | C D E F] |
| 334 | # hidden_restore_idx: [0, 1, 4, 5, 6, 7, 2] |
| 335 | # padded_gather_idx: [0, 1, 6, 0, 2, 3, 4, 5] |
| 336 | # Therefore global = gathered[hidden_restore_idx] and |
| 337 | # padded_gathered = global[padded_gather_idx]. |
| 338 | hidden_restore_idx = np.empty(int(query_start_loc_np[-1]), dtype=np.int64) |
| 339 | if padded_num_tokens is None: |
| 340 | padded_num_tokens = max(per_rank_num_tokens) |
| 341 | elif padded_num_tokens < max(per_rank_num_tokens): |
| 342 | raise ValueError( |
| 343 | "PCP padded token count is smaller than the largest rank-local " |
| 344 | f"batch: {padded_num_tokens} < {max(per_rank_num_tokens)}." |
| 345 | ) |
| 346 | num_expanded_tokens = padded_num_tokens * self.pcp_world_size |
| 347 | padded_gather_idx = np.zeros(num_expanded_tokens, dtype=np.int64) |
| 348 | gathered_kv_write_mask = np.zeros(num_expanded_tokens, dtype=np.bool_) |
| 349 | for rank, segments in enumerate(segments_by_rank): |
| 350 | expanded_rank_offset = rank * padded_num_tokens |
| 351 | for segment in segments: |
| 352 | padded_gathered_slice = slice( |
| 353 | expanded_rank_offset + segment.rank_local_batch_slice.start, |
| 354 | expanded_rank_offset + segment.rank_local_batch_slice.stop, |
| 355 | ) |
| 356 | padded_gather_idx[padded_gathered_slice] = np.arange( |
| 357 | segment.global_batch_slice.start, |
| 358 | segment.global_batch_slice.stop, |
| 359 | dtype=np.int64, |
| 360 | ) |
| 361 | # Replicated rows contain identical KV, so rank 0 is the |
| 362 | # canonical writer. A sharded decode has exactly one owner and |
| 363 | # therefore needs no de-duplication here. |
| 364 | is_sharded_decode = ( |
| 365 | not bool(is_prefilling[segment.global_batch_req_idx]) |
| 366 | and self.shard_decode_requests |
| 367 | ) |
| 368 | if ( |
| 369 | replicated[segment.global_batch_req_idx] |
| 370 | and not is_sharded_decode |
| 371 | and rank != 0 |
| 372 | ): |
| 373 | continue |
| 374 | gathered_kv_write_mask[padded_gathered_slice] = True |
| 375 | hidden_restore_idx[segment.global_batch_slice] = np.arange( |
| 376 | padded_gathered_slice.start, |
| 377 | padded_gathered_slice.stop, |
| 378 | dtype=np.int64, |
| 379 | ) |
| 380 | |
| 381 | self._hidden_restore_idx = async_tensor_h2d( |
| 382 | hidden_restore_idx, device=self.device |
| 383 | ) |
| 384 | self._padded_gather_idx = async_tensor_h2d( |
| 385 | padded_gather_idx, device=self.device |
| 386 | ) |
| 387 | self._gathered_kv_write_mask = async_tensor_h2d( |
| 388 | gathered_kv_write_mask, device=self.device |
| 389 | ) |
| 390 | return segments_by_rank, per_rank_num_tokens |
| 391 | |
| 392 | def get_num_tokens_for_dispatch( |
| 393 | self, |
| 394 | num_scheduled_tokens: np.ndarray, |
| 395 | is_prefilling: np.ndarray, |
| 396 | ) -> int: |
| 397 | """Return the largest real rank-local batch before graph padding.""" |
| 398 | return max( |
| 399 | sum( |
| 400 | chunk_len |
| 401 | for _, _, chunk_len in self._iter_rank_chunks( |
| 402 | rank, num_scheduled_tokens, is_prefilling |
| 403 | ) |
| 404 | ) |
| 405 | for rank in range(self.pcp_world_size) |
| 406 | ) |
| 407 | |
| 408 | @staticmethod |
| 409 | def _resolve_num_reqs_after_padding( |
| 410 | input_batch: InputBatch, |
| 411 | padded_num_reqs: int | None, |
| 412 | num_local_reqs: int, |
| 413 | ) -> int: |
| 414 | if padded_num_reqs is None: |
| 415 | return num_local_reqs |
| 416 | if input_batch.has_prefill: |
| 417 | raise RuntimeError("PCP FULL graphs require a decode-only batch.") |
| 418 | assert padded_num_reqs >= num_local_reqs, ( |
| 419 | "PCP graph request capacity must cover the rank-local batch: " |
| 420 | f"{padded_num_reqs} < {num_local_reqs}." |
| 421 | ) |
| 422 | return padded_num_reqs |
| 423 | |
| 424 | @property |
| 425 | def input_buffers(self) -> InputBuffers: |
| 426 | assert self._input_buffers is not None |
| 427 | return self._input_buffers |
| 428 | |
| 429 | @property |
| 430 | def global_batch(self) -> InputBatch: |
| 431 | assert self._global_batch is not None |
| 432 | return self._global_batch |
| 433 | |
| 434 | def partition_batch( |
| 435 | self, input_batch: InputBatch, batch_desc: "BatchExecutionDescriptor" |
| 436 | ) -> InputBatch: |
| 437 | assert self._input_buffers is not None |
| 438 | input_buffers = self._input_buffers |
| 439 | |
| 440 | global_batch = input_batch |
| 441 | self._global_batch = global_batch |
| 442 | |
| 443 | num_scheduled_tokens = global_batch.num_scheduled_tokens |
| 444 | num_computed_tokens = global_batch.num_computed_tokens_np |
| 445 | is_prefilling = global_batch.is_prefilling_np |
| 446 | |
| 447 | padded_num_tokens = None |
| 448 | padded_num_reqs = None |
| 449 | if batch_desc.cg_mode != CUDAGraphMode.NONE: |
| 450 | padded_num_tokens = batch_desc.num_tokens |
| 451 | if batch_desc.cg_mode == CUDAGraphMode.FULL: |
| 452 | padded_num_reqs = batch_desc.num_reqs |
| 453 | |
| 454 | segments_by_rank, per_rank_num_tokens = self._build_batch_layout( |
| 455 | num_scheduled_tokens, |
| 456 | num_computed_tokens, |
| 457 | is_prefilling, |
| 458 | global_batch.query_start_loc_np, |
| 459 | padded_num_tokens=padded_num_tokens, |
| 460 | ) |
| 461 | |
| 462 | local_segments = segments_by_rank[self.pcp_rank] |
| 463 | if not local_segments: |
| 464 | local_segments = [ |
| 465 | RankSegment( |
| 466 | global_batch_req_idx=0, |
| 467 | global_batch_slice=slice(0, 0), |
| 468 | rank_local_batch_slice=slice(0, 0), |
| 469 | ) |
| 470 | ] |
| 471 | |
| 472 | num_local_reqs = len(local_segments) |
| 473 | num_reqs_after_padding = self._resolve_num_reqs_after_padding( |
| 474 | global_batch, padded_num_reqs, num_local_reqs |
| 475 | ) |
| 476 | if num_reqs_after_padding > input_buffers.max_num_reqs: |
| 477 | raise RuntimeError( |
| 478 | "PCP padded local request count exceeds the MRV2 input buffer size: " |
| 479 | f"{num_reqs_after_padding} > {input_buffers.max_num_reqs}." |
| 480 | ) |
| 481 | |
| 482 | local_to_global_batch_req_idx_np = np.fromiter( |
| 483 | (segment.global_batch_req_idx for segment in local_segments), |
| 484 | dtype=np.intp, |
| 485 | count=num_local_reqs, |
| 486 | ) |
| 487 | local_start_pos_np = np.fromiter( |
| 488 | ( |
| 489 | num_computed_tokens[segment.global_batch_req_idx] |
| 490 | + segment.global_batch_slice.start |
| 491 | - global_batch.query_start_loc_np[segment.global_batch_req_idx] |
| 492 | for segment in local_segments |
| 493 | ), |
| 494 | dtype=np.int32, |
| 495 | count=num_local_reqs, |
| 496 | ) |
| 497 | local_num_scheduled_tokens = np.fromiter( |
| 498 | (segment.num_tokens for segment in local_segments), |
| 499 | dtype=np.int32, |
| 500 | count=num_local_reqs, |
| 501 | ) |
| 502 | local_to_global_req_idx_np = global_batch.idx_mapping_np[ |
| 503 | local_to_global_batch_req_idx_np |
| 504 | ] |
| 505 | local_req_ids = [ |
| 506 | global_batch.req_ids[global_batch_req_idx] |
| 507 | for global_batch_req_idx in local_to_global_batch_req_idx_np |
| 508 | ] |
| 509 | |
| 510 | num_local_tokens = int(local_num_scheduled_tokens.sum()) |
| 511 | num_local_tokens_padded = ( |
| 512 | max(per_rank_num_tokens) if padded_num_tokens is None else padded_num_tokens |
| 513 | ) |
| 514 | fresh_prefills = int( |
| 515 | np.count_nonzero(is_prefilling & (num_computed_tokens == 0)) |
| 516 | ) |
| 517 | continued_prefills = int( |
| 518 | np.count_nonzero(is_prefilling & (num_computed_tokens > 0)) |
| 519 | ) |
| 520 | logger.debug( |
| 521 | "PCP batch: rank=%d global_batch_reqs=%d fresh_prefills=%d " |
| 522 | "continued_prefills=%d decodes=%d local_reqs=%d " |
| 523 | "local_tokens=%d per_rank_tokens=%s", |
| 524 | self.pcp_rank, |
| 525 | global_batch.num_reqs, |
| 526 | fresh_prefills, |
| 527 | continued_prefills, |
| 528 | global_batch.num_reqs - fresh_prefills - continued_prefills, |
| 529 | num_local_reqs, |
| 530 | num_local_tokens, |
| 531 | per_rank_num_tokens, |
| 532 | ) |
| 533 | if num_local_tokens_padded > input_buffers.max_num_tokens: |
| 534 | raise RuntimeError( |
| 535 | "PCP local token count exceeds the MRV2 input buffer size: " |
| 536 | f"{num_local_tokens_padded} > {input_buffers.max_num_tokens}." |
| 537 | ) |
| 538 | rank_token_start = self.pcp_rank * num_local_tokens_padded |
| 539 | assert self._padded_gather_idx is not None |
| 540 | local_gather_idx = self._padded_gather_idx[ |
| 541 | rank_token_start : rank_token_start + num_local_tokens_padded |
| 542 | ] |
| 543 | self._local_gather_idx = local_gather_idx |
| 544 | torch.index_select( |
| 545 | global_batch.input_ids, |
| 546 | 0, |
| 547 | local_gather_idx, |
| 548 | out=input_buffers.input_ids[:num_local_tokens_padded], |
| 549 | ) |
| 550 | # Keep the GPU request-state cursor materialized by prepare_inputs(). |
| 551 | # The CPU cursor can lag after speculative rejection. |
| 552 | torch.index_select( |
| 553 | global_batch.positions, |
| 554 | 0, |
| 555 | local_gather_idx, |
| 556 | out=input_buffers.positions[:num_local_tokens_padded], |
| 557 | ) |
| 558 | |
| 559 | local_query_start_loc_np = np.empty( |
| 560 | input_buffers.max_num_reqs + 1, dtype=np.int32 |
| 561 | ) |
| 562 | local_query_start_loc_np[0] = 0 |
| 563 | local_query_start_loc_out = local_query_start_loc_np[1 : num_local_reqs + 1] |
| 564 | np.cumsum(local_num_scheduled_tokens, out=local_query_start_loc_out) |
| 565 | local_query_start_loc_np[num_local_reqs + 1 :] = num_local_tokens |
| 566 | async_tensor_h2d(local_query_start_loc_np, out=input_buffers.query_start_loc) |
| 567 | local_query_start_loc = input_buffers.query_start_loc[ |
| 568 | : num_reqs_after_padding + 1 |
| 569 | ] |
| 570 | |
| 571 | local_to_global_req_idx = async_tensor_h2d( |
| 572 | local_to_global_req_idx_np, device=self.device |
| 573 | ) |
| 574 | seq_lens = input_buffers.seq_lens[:num_reqs_after_padding] |
| 575 | real_seq_lens = seq_lens[:num_local_reqs] |
| 576 | if num_local_tokens > 0: |
| 577 | local_end_positions = torch.index_select( |
| 578 | input_buffers.positions, |
| 579 | 0, |
| 580 | local_query_start_loc[1 : num_local_reqs + 1] - 1, |
| 581 | ) |
| 582 | real_seq_lens.copy_(local_end_positions + 1) |
| 583 | else: |
| 584 | real_seq_lens.zero_() |
| 585 | seq_lens[num_local_reqs:].zero_() |
| 586 | is_padding = input_buffers.is_padding[:num_local_tokens_padded] |
| 587 | is_padding[:num_local_tokens].fill_(False) |
| 588 | is_padding[num_local_tokens:].fill_(True) |
| 589 | if num_local_tokens_padded > num_local_tokens: |
| 590 | input_buffers.input_ids[:num_local_tokens_padded].masked_fill_( |
| 591 | is_padding, 0 |
| 592 | ) |
| 593 | input_buffers.positions[:num_local_tokens_padded].masked_fill_( |
| 594 | is_padding, 0 |
| 595 | ) |
| 596 | |
| 597 | total_num_logits = num_local_reqs if num_local_tokens > 0 else 0 |
| 598 | if total_num_logits > 0: |
| 599 | cu_num_logits_np = np.arange(num_local_reqs + 1, dtype=np.int32) |
| 600 | cu_num_logits = torch.arange( |
| 601 | num_local_reqs + 1, device=self.device, dtype=torch.int32 |
| 602 | ) |
| 603 | else: |
| 604 | cu_num_logits_np = np.zeros(num_local_reqs + 1, dtype=np.int32) |
| 605 | cu_num_logits = torch.zeros( |
| 606 | num_local_reqs + 1, device=self.device, dtype=torch.int32 |
| 607 | ) |
| 608 | # Local logits are never sampled. The complete hidden-state tensor is |
| 609 | # restored first and sampled with the untouched global InputBatch. |
| 610 | logits_indices = local_query_start_loc[1:] - 1 |
| 611 | |
| 612 | local_prefill_len_np = global_batch.prefill_len_np[ |
| 613 | local_to_global_batch_req_idx_np |
| 614 | ] |
| 615 | local_num_computed_prefill_tokens_np = np.minimum( |
| 616 | local_start_pos_np, local_prefill_len_np |
| 617 | ) |
| 618 | real_local_is_prefilling_np = ( |
| 619 | local_num_computed_prefill_tokens_np < local_prefill_len_np |
| 620 | ) |
| 621 | local_is_prefilling_np = np.zeros(num_reqs_after_padding, dtype=np.bool_) |
| 622 | local_is_prefilling_np[:num_local_reqs] = real_local_is_prefilling_np |
| 623 | local_has_prefill = bool(local_is_prefilling_np.any()) |
| 624 | seq_lens_cpu_upper_bound_np = np.zeros(num_reqs_after_padding, dtype=np.int32) |
| 625 | seq_lens_cpu_upper_bound_np[:num_local_reqs] = ( |
| 626 | local_start_pos_np + local_num_scheduled_tokens |
| 627 | ) |
| 628 | dcp_local_seq_lens_cpu_upper_bound = None |
| 629 | if self.dcp_world_size > 1: |
| 630 | # The largest DCP shard of each row's whole request, identical on |
| 631 | # every PCP rank: the sparse backends pad their KV gather to it. |
| 632 | request_seq_lens = (num_computed_tokens + num_scheduled_tokens)[ |
| 633 | local_to_global_batch_req_idx_np |
| 634 | ] |
| 635 | real_dcp_local_seq_lens_cpu_upper_bound = get_dcp_local_seq_lens( |
| 636 | torch.from_numpy(request_seq_lens.astype(np.int32)), |
| 637 | self.dcp_world_size, |
| 638 | 0, |
| 639 | self.cp_interleave, |
| 640 | ) |
| 641 | dcp_local_seq_lens_cpu_upper_bound = torch.zeros( |
| 642 | num_reqs_after_padding, dtype=torch.int32 |
| 643 | ) |
| 644 | dcp_local_seq_lens_cpu_upper_bound[:num_local_reqs].copy_( |
| 645 | real_dcp_local_seq_lens_cpu_upper_bound |
| 646 | ) |
| 647 | |
| 648 | self._local_batch = replace( |
| 649 | input_batch, |
| 650 | req_ids=local_req_ids, |
| 651 | num_reqs=num_local_reqs, |
| 652 | num_reqs_after_padding=num_reqs_after_padding, |
| 653 | idx_mapping=local_to_global_req_idx, |
| 654 | idx_mapping_np=local_to_global_req_idx_np, |
| 655 | expanded_idx_mapping=local_to_global_req_idx, |
| 656 | expanded_local_pos=torch.zeros( |
| 657 | num_local_reqs, dtype=torch.int32, device=self.device |
| 658 | ), |
| 659 | num_scheduled_tokens=local_num_scheduled_tokens, |
| 660 | num_tokens=num_local_tokens, |
| 661 | num_tokens_after_padding=num_local_tokens_padded, |
| 662 | num_draft_tokens=0, |
| 663 | num_draft_tokens_per_req=None, |
| 664 | query_start_loc=local_query_start_loc, |
| 665 | query_start_loc_np=local_query_start_loc_np[: num_reqs_after_padding + 1], |
| 666 | seq_lens=seq_lens, |
| 667 | seq_lens_cpu_upper_bound=torch.from_numpy(seq_lens_cpu_upper_bound_np), |
| 668 | dcp_local_seq_lens=None, |
| 669 | dcp_local_seq_lens_cpu_upper_bound=dcp_local_seq_lens_cpu_upper_bound, |
| 670 | num_computed_tokens_np=local_start_pos_np, |
| 671 | prefill_len_np=local_prefill_len_np, |
| 672 | num_computed_prefill_tokens_np=local_num_computed_prefill_tokens_np, |
| 673 | is_prefilling_np=local_is_prefilling_np, |
| 674 | max_seq_len_np=None, |
| 675 | has_prefill=local_has_prefill, |
| 676 | decode_graph_eligible=not local_has_prefill, |
| 677 | prefill_runs_as_decode_np=None, |
| 678 | input_ids=input_buffers.input_ids[:num_local_tokens_padded], |
| 679 | positions=input_buffers.positions[:num_local_tokens_padded], |
| 680 | is_padding=is_padding, |
| 681 | logits_indices=logits_indices, |
| 682 | cu_num_logits=cu_num_logits, |
| 683 | cu_num_logits_np=cu_num_logits_np, |
| 684 | prompt_lens=None, |
| 685 | ) |
| 686 | return self._local_batch |
| 687 | |
| 688 | def prepare_inputs_to_capture(self, input_batch: InputBatch) -> InputBatch: |
| 689 | """Stage a capture or dummy batch in persistent PCP input buffers.""" |
| 690 | input_buffers = self.input_buffers |
| 691 | num_reqs = input_batch.num_reqs_after_padding |
| 692 | num_tokens = input_batch.num_tokens_after_padding |
| 693 | input_batch = replace( |
| 694 | input_batch, |
| 695 | input_ids=input_buffers.input_ids[:num_tokens].copy_(input_batch.input_ids), |
| 696 | positions=input_buffers.positions[:num_tokens].copy_(input_batch.positions), |
| 697 | is_padding=input_buffers.is_padding[:num_tokens].copy_( |
| 698 | input_batch.is_padding |
| 699 | ), |
| 700 | query_start_loc=input_buffers.query_start_loc[: num_reqs + 1].copy_( |
| 701 | input_batch.query_start_loc |
| 702 | ), |
| 703 | seq_lens=input_buffers.seq_lens[:num_reqs].copy_(input_batch.seq_lens), |
| 704 | ) |
| 705 | return input_batch |
| 706 | |
| 707 | def get_dummy_block_tables(self, num_reqs: int) -> tuple[torch.Tensor, ...]: |
| 708 | assert self._local_block_tables is not None |
| 709 | return tuple( |
| 710 | block_table[:num_reqs].zero_() for block_table in self._local_block_tables |
| 711 | ) |
| 712 | |
| 713 | def prepare_attn( |
| 714 | self, input_batch: InputBatch |
| 715 | ) -> tuple[tuple[torch.Tensor, ...], torch.Tensor]: |
| 716 | assert self._block_tables is not None |
| 717 | assert self._local_block_tables is not None |
| 718 | assert self._local_block_table_ptrs is not None |
| 719 | block_tables = self._block_tables.gather_block_tables( |
| 720 | input_batch.idx_mapping, |
| 721 | input_batch.num_reqs_after_padding, |
| 722 | out=self._local_block_tables, |
| 723 | out_ptrs=self._local_block_table_ptrs, |
| 724 | ) |
| 725 | slot_mappings = self.prepare_slot_mappings() |
| 726 | return block_tables, slot_mappings |
| 727 | |
| 728 | def prepare_slot_mappings(self) -> torch.Tensor: |
| 729 | assert self._block_tables is not None |
| 730 | assert self._global_batch_slot_mappings is not None |
| 731 | assert self._global_batch is not None |
| 732 | global_batch = self._global_batch |
| 733 | global_batch_slot_mappings = self._block_tables.compute_slot_mappings( |
| 734 | global_batch.idx_mapping, |
| 735 | global_batch.query_start_loc, |
| 736 | global_batch.positions, |
| 737 | global_batch.num_tokens, |
| 738 | out=self._global_batch_slot_mappings, |
| 739 | ) |
| 740 | return self._convert_to_gathered_slot_mappings(global_batch_slot_mappings) |
| 741 | |
| 742 | def get_dummy_slot_mappings(self, num_tokens: int) -> torch.Tensor: |
| 743 | assert self._gathered_kv_slot_mappings is not None |
| 744 | self._gathered_kv_slot_mappings.fill_(PAD_SLOT_ID) |
| 745 | return self._gathered_kv_slot_mappings[:, : num_tokens * self.pcp_world_size] |
| 746 | |
| 747 | def _convert_to_gathered_slot_mappings( |
| 748 | self, global_batch_slot_mappings: torch.Tensor |
| 749 | ) -> torch.Tensor: |
| 750 | assert self._padded_gather_idx is not None |
| 751 | assert self._gathered_kv_write_mask is not None |
| 752 | padded_gather_idx = self._padded_gather_idx |
| 753 | num_expanded_tokens = padded_gather_idx.shape[0] |
| 754 | if self._gathered_kv_slot_mappings is None: |
| 755 | self._gathered_kv_slot_mappings = global_batch_slot_mappings.new_empty( |
| 756 | global_batch_slot_mappings.shape[0], num_expanded_tokens |
| 757 | ) |
| 758 | gathered_kv_slot_mappings = self._gathered_kv_slot_mappings[ |
| 759 | :, :num_expanded_tokens |
| 760 | ] |
| 761 | torch.index_select( |
| 762 | global_batch_slot_mappings, |
| 763 | 1, |
| 764 | padded_gather_idx, |
| 765 | out=gathered_kv_slot_mappings, |
| 766 | ) |
| 767 | torch.where( |
| 768 | self._gathered_kv_write_mask.unsqueeze(0), |
| 769 | gathered_kv_slot_mappings, |
| 770 | self._pad_slot_id, |
| 771 | out=gathered_kv_slot_mappings, |
| 772 | ) |
| 773 | return gathered_kv_slot_mappings |
| 774 | |
| 775 | def restore_hidden_states(self, hidden_states: torch.Tensor) -> torch.Tensor: |
| 776 | if self._hidden_restore_idx is None: |
| 777 | return hidden_states |
| 778 | gathered = get_pcp_group().all_gather(hidden_states, dim=0) |
| 779 | return gathered[self._hidden_restore_idx] |
| 780 | |
| 781 | def get_draft_input_buffers( |
| 782 | self, input_buffers: InputBuffers |
| 783 | ) -> InputBatch | InputBuffers: |
| 784 | return self.draft_prefill_batch or input_buffers |
| 785 | |
| 786 | def prepare_draft_prefill( |
| 787 | self, input_batch: InputBatch, input_ids: torch.Tensor |
| 788 | ) -> None: |
| 789 | self.draft_prefill_batch = None |
| 790 | if input_batch is not self._global_batch or self._local_batch is None: |
| 791 | return |
| 792 | local_batch = self._local_batch |
| 793 | assert self._local_gather_idx is not None |
| 794 | num_local_tokens = self._local_gather_idx.shape[0] |
| 795 | torch.index_select( |
| 796 | input_ids, |
| 797 | 0, |
| 798 | self._local_gather_idx, |
| 799 | out=local_batch.input_ids[:num_local_tokens], |
| 800 | ) |
| 801 | self.draft_prefill_batch = local_batch |
| 802 | |
| 803 | def restore_draft_prefill( |
| 804 | self, |
| 805 | last_hidden_states: torch.Tensor, |
| 806 | hidden_states: torch.Tensor, |
| 807 | ) -> tuple[torch.Tensor, torch.Tensor]: |
| 808 | if self.draft_prefill_batch is None: |
| 809 | return last_hidden_states, hidden_states |
| 810 | local_last_hidden_states = last_hidden_states |
| 811 | last_hidden_states = self.restore_hidden_states(local_last_hidden_states) |
| 812 | hidden_states = ( |
| 813 | last_hidden_states |
| 814 | if local_last_hidden_states is hidden_states |
| 815 | else self.restore_hidden_states(hidden_states) |
| 816 | ) |
| 817 | self.draft_prefill_batch = None |
| 818 | return last_hidden_states, hidden_states |
| 819 | |
| 820 | def restore_for_sampling( |
| 821 | self, hidden_states: torch.Tensor, aux_hidden_states: list[torch.Tensor] | None |
| 822 | ) -> tuple[torch.Tensor, list[torch.Tensor] | None, InputBatch]: |
| 823 | assert self._global_batch is not None |
| 824 | hidden_states = self.restore_hidden_states(hidden_states) |
| 825 | if aux_hidden_states is not None: |
| 826 | aux_hidden_states = [ |
| 827 | self.restore_hidden_states(states) for states in aux_hidden_states |
| 828 | ] |
| 829 | return hidden_states, aux_hidden_states, self._global_batch |
| 830 | |
| 831 | |
| 832 | def maybe_build_pcp_manager( |
| 833 | vllm_config: VllmConfig, |
| 834 | device: torch.device, |
| 835 | supports_mm_inputs: bool, |
| 836 | block_tables: BlockTables, |
| 837 | cls: type[PCPManager] = PCPManager, |
| 838 | ) -> PCPManager | None: |
| 839 | parallel_config = vllm_config.parallel_config |
| 840 | pcp_size = parallel_config.prefill_context_parallel_size |
| 841 | if pcp_size <= 1: |
| 842 | return None |
| 843 | |
| 844 | cls.validate_config(vllm_config, supports_mm_inputs) |
| 845 | |
| 846 | pcp_rank = get_pcp_group().rank_in_group |
| 847 | dcp_size = parallel_config.decode_context_parallel_size |
| 848 | dcp_rank = get_dcp_group().rank_in_group if dcp_size > 1 else 0 |
| 849 | |
| 850 | return cls( |
| 851 | pcp_world_size=pcp_size, |
| 852 | pcp_rank=pcp_rank, |
| 853 | device=device, |
| 854 | shard_decode_requests=parallel_config.pcp_shard_decode_requests, |
| 855 | max_num_reqs=vllm_config.scheduler_config.max_num_seqs, |
| 856 | max_num_tokens=vllm_config.scheduler_config.max_num_batched_tokens, |
| 857 | block_tables=block_tables, |
| 858 | dcp_world_size=dcp_size, |
| 859 | dcp_rank=dcp_rank, |
| 860 | cp_interleave=parallel_config.cp_kv_cache_interleave_size, |
| 861 | ) |