| 1 | # SPDX-License-Identifier: Apache-2.0 |
| 2 | # SPDX-FileCopyrightText: Copyright contributors to the vLLM project |
| 3 | """Microbatching (DBO) helpers for the V2 GPU model runner.""" |
| 4 | |
| 5 | import threading |
| 6 | from collections.abc import Callable |
| 7 | from contextlib import ExitStack |
| 8 | from dataclasses import replace |
| 9 | from typing import Any, NamedTuple |
| 10 | |
| 11 | import numpy as np |
| 12 | import torch |
| 13 | |
| 14 | from vllm.config import CUDAGraphMode, VllmConfig |
| 15 | from vllm.distributed import get_dcp_group |
| 16 | from vllm.forward_context import ( |
| 17 | DPMetadata, |
| 18 | ForwardContext, |
| 19 | create_forward_context, |
| 20 | override_forward_context, |
| 21 | ) |
| 22 | from vllm.logger import init_logger |
| 23 | from vllm.sequence import IntermediateTensors |
| 24 | from vllm.utils.torch_utils import current_stream |
| 25 | from vllm.v1.kv_cache_interface import KVCacheConfig |
| 26 | from vllm.v1.worker.gpu.attn_utils import build_slot_mappings_by_layer |
| 27 | from vllm.v1.worker.gpu.cp_utils import prepare_dcp_local_seq_lens |
| 28 | from vllm.v1.worker.gpu.input_batch import InputBatch |
| 29 | from vllm.v1.worker.gpu.model_states.interface import ModelState |
| 30 | from vllm.v1.worker.ubatch_utils import ( |
| 31 | UBatchSlice, |
| 32 | UBatchSlices, |
| 33 | create_sm_control_context, |
| 34 | maybe_create_ubatch_slices, |
| 35 | ) |
| 36 | from vllm.v1.worker.ubatching import make_ubatch_contexts |
| 37 | from vllm.v1.worker.utils import AttentionGroup |
| 38 | |
| 39 | logger = init_logger(__name__) |
| 40 | |
| 41 | |
| 42 | class UBatchState(NamedTuple): |
| 43 | """Per-microbatch inputs for one dual-batch-overlap step.""" |
| 44 | |
| 45 | slices: UBatchSlices |
| 46 | forward_contexts: list[ForwardContext] |
| 47 | |
| 48 | |
| 49 | def create_ubatch_slices(input_batch: InputBatch, num_ubatches: int) -> UBatchSlices: |
| 50 | """Split a DP-padded batch into the slices its microbatches run on. |
| 51 | |
| 52 | Splitting at the midpoint of the DP-padded token count leaves the trailing |
| 53 | microbatch anywhere from full to entirely inside the padding. Clamping its |
| 54 | request slice onto the last request leaves it holding that request with |
| 55 | zero query tokens -- the shape a straddling request already takes -- so it |
| 56 | stays well-formed with no work to do, like a dummy run. |
| 57 | """ |
| 58 | _, ubatch_slices = maybe_create_ubatch_slices( |
| 59 | True, |
| 60 | input_batch.num_scheduled_tokens, |
| 61 | input_batch.num_tokens_after_padding, |
| 62 | input_batch.num_reqs_after_padding, |
| 63 | num_ubatches, |
| 64 | ) |
| 65 | assert ubatch_slices is not None |
| 66 | return [ |
| 67 | UBatchSlice( |
| 68 | slice( |
| 69 | min(s.request_slice.start, input_batch.num_reqs - 1), |
| 70 | min(s.request_slice.stop, input_batch.num_reqs_after_padding), |
| 71 | ), |
| 72 | s.token_slice, |
| 73 | ) |
| 74 | for s in ubatch_slices |
| 75 | ] |
| 76 | |
| 77 | |
| 78 | def _slice_input_batch( |
| 79 | input_batch: InputBatch, |
| 80 | ubatch_slice: UBatchSlice, |
| 81 | query_start_loc_buf: torch.Tensor, |
| 82 | seq_lens_buf: torch.Tensor, |
| 83 | dcp_local_seq_lens_buf: torch.Tensor | None = None, |
| 84 | *, |
| 85 | dcp_size: int = 1, |
| 86 | dcp_rank: int = 0, |
| 87 | cp_interleave: int = 1, |
| 88 | ) -> InputBatch: |
| 89 | """Build the sub-`InputBatch` a single microbatch runs on. |
| 90 | |
| 91 | A request straddling the boundary appears in both microbatches, its query |
| 92 | truncated to the tokens each one owns. |
| 93 | |
| 94 | Query offsets and sequence lengths (including DCP-local lengths) are |
| 95 | written to the caller's persistent buffers after truncation. Other GPU |
| 96 | inputs are views of the full batch's buffers. |
| 97 | |
| 98 | The sub-batch describes the forward pass only. Sampling runs once over the |
| 99 | merged batch, so `logits_indices`, `cu_num_logits` and the draft-token |
| 100 | fields are carried over unchanged and must not be read off a sub-batch. |
| 101 | """ |
| 102 | assert not ubatch_slice.is_empty(), f"Ubatch slice {ubatch_slice} is empty" |
| 103 | |
| 104 | req_start = ubatch_slice.request_slice.start |
| 105 | req_stop = ubatch_slice.request_slice.stop |
| 106 | tok_start = ubatch_slice.token_slice.start |
| 107 | tok_stop = ubatch_slice.token_slice.stop |
| 108 | |
| 109 | num_reqs_padded = req_stop - req_start |
| 110 | num_tokens_padded = tok_stop - tok_start |
| 111 | # The trailing microbatch may extend past the real batch into CUDA graph |
| 112 | # padding; the unpadded counts stop at the last real request/token. |
| 113 | num_reqs = max(0, min(req_stop, input_batch.num_reqs) - req_start) |
| 114 | num_tokens = max(0, min(tok_stop, input_batch.num_tokens) - tok_start) |
| 115 | |
| 116 | query_start_loc = _slice_query_start_loc( |
| 117 | input_batch.query_start_loc, |
| 118 | query_start_loc_buf[: num_reqs_padded + 1], |
| 119 | req_start, |
| 120 | req_stop, |
| 121 | tok_start, |
| 122 | num_tokens_padded, |
| 123 | ) |
| 124 | query_start_loc_np = np.clip( |
| 125 | input_batch.query_start_loc_np[req_start : req_stop + 1] - tok_start, |
| 126 | 0, |
| 127 | num_tokens_padded, |
| 128 | ).astype(np.int32) |
| 129 | |
| 130 | seq_lens = _slice_seq_lens( |
| 131 | input_batch.seq_lens, |
| 132 | seq_lens_buf[:num_reqs_padded], |
| 133 | input_batch.query_start_loc, |
| 134 | req_start, |
| 135 | req_stop, |
| 136 | tok_stop, |
| 137 | ) |
| 138 | # Same truncation as `_slice_seq_lens`, on the host-side upper bound. |
| 139 | seq_lens_cpu_upper_bound = input_batch.seq_lens_cpu_upper_bound[ |
| 140 | req_start:req_stop |
| 141 | ].clone() |
| 142 | tokens_truncated = max(0, int(input_batch.query_start_loc_np[req_stop]) - tok_stop) |
| 143 | if tokens_truncated: |
| 144 | seq_lens_cpu_upper_bound[-1] -= tokens_truncated |
| 145 | |
| 146 | # Query lengths of the truncated requests, so consumers that derive |
| 147 | # max_query_len from this array see the microbatch's own lengths. |
| 148 | num_scheduled_tokens = np.diff(query_start_loc_np)[:num_reqs] |
| 149 | |
| 150 | dcp_local_seq_lens = None |
| 151 | if dcp_size > 1: |
| 152 | assert dcp_local_seq_lens_buf is not None |
| 153 | dcp_local_seq_lens = prepare_dcp_local_seq_lens( |
| 154 | dcp_local_seq_lens_buf, |
| 155 | seq_lens, |
| 156 | num_reqs, |
| 157 | dcp_size, |
| 158 | dcp_rank, |
| 159 | cp_interleave, |
| 160 | num_reqs_padded=num_reqs_padded, |
| 161 | ) |
| 162 | |
| 163 | return replace( |
| 164 | input_batch, |
| 165 | req_ids=input_batch.req_ids[req_start : min(req_stop, input_batch.num_reqs)], |
| 166 | num_reqs=num_reqs, |
| 167 | num_reqs_after_padding=num_reqs_padded, |
| 168 | idx_mapping=input_batch.idx_mapping[req_start:req_stop], |
| 169 | idx_mapping_np=input_batch.idx_mapping_np[req_start:req_stop], |
| 170 | num_scheduled_tokens=num_scheduled_tokens, |
| 171 | num_tokens=num_tokens, |
| 172 | num_tokens_after_padding=num_tokens_padded, |
| 173 | query_start_loc=query_start_loc, |
| 174 | query_start_loc_np=query_start_loc_np, |
| 175 | seq_lens=seq_lens, |
| 176 | seq_lens_cpu_upper_bound=seq_lens_cpu_upper_bound, |
| 177 | dcp_local_seq_lens=dcp_local_seq_lens, |
| 178 | num_computed_tokens_np=input_batch.num_computed_tokens_np[req_start:req_stop], |
| 179 | prefill_len_np=input_batch.prefill_len_np[req_start:req_stop], |
| 180 | num_computed_prefill_tokens_np=input_batch.num_computed_prefill_tokens_np[ |
| 181 | req_start:req_stop |
| 182 | ], |
| 183 | is_prefilling_np=input_batch.is_prefilling_np[req_start:req_stop], |
| 184 | prefill_runs_as_decode_np=( |
| 185 | None |
| 186 | if input_batch.prefill_runs_as_decode_np is None |
| 187 | else input_batch.prefill_runs_as_decode_np[req_start:req_stop] |
| 188 | ), |
| 189 | input_ids=input_batch.input_ids[tok_start:tok_stop], |
| 190 | positions=input_batch.positions[tok_start:tok_stop], |
| 191 | is_padding=input_batch.is_padding[tok_start:tok_stop], |
| 192 | prompt_lens=( |
| 193 | None |
| 194 | if input_batch.prompt_lens is None |
| 195 | else input_batch.prompt_lens[req_start:req_stop] |
| 196 | ), |
| 197 | max_seq_len_np=None, |
| 198 | fast_prefill=None, |
| 199 | ) |
| 200 | |
| 201 | |
| 202 | def _slice_query_start_loc( |
| 203 | query_start_loc: torch.Tensor, |
| 204 | out: torch.Tensor, |
| 205 | req_start: int, |
| 206 | req_stop: int, |
| 207 | tok_start: int, |
| 208 | num_tokens_padded: int, |
| 209 | ) -> torch.Tensor: |
| 210 | """Rebase query_start_loc onto the microbatch's token range. |
| 211 | |
| 212 | The clamp truncates the requests straddling either boundary to the tokens |
| 213 | this microbatch owns. |
| 214 | """ |
| 215 | torch.sub(query_start_loc[req_start : req_stop + 1], tok_start, out=out) |
| 216 | return out.clamp_(0, num_tokens_padded) |
| 217 | |
| 218 | |
| 219 | def _slice_seq_lens( |
| 220 | seq_lens: torch.Tensor, |
| 221 | out: torch.Tensor, |
| 222 | query_start_loc: torch.Tensor, |
| 223 | req_start: int, |
| 224 | req_stop: int, |
| 225 | tok_stop: int, |
| 226 | ) -> torch.Tensor: |
| 227 | """Slice seq_lens, shortening the request truncated at the boundary. |
| 228 | |
| 229 | Only tokens past the end of this microbatch are missing; tokens it owns in |
| 230 | an *earlier* microbatch are computed before this one reads them, so they |
| 231 | still count. Deriving the truncation from this microbatch's own query |
| 232 | lengths would subtract them twice. It is computed from tensors, not Python |
| 233 | ints, so this stays valid inside a CUDA graph. |
| 234 | """ |
| 235 | out.copy_(seq_lens[req_start:req_stop]) |
| 236 | last = req_stop - req_start - 1 |
| 237 | out[last] -= (query_start_loc[req_stop] - tok_stop).clamp_(min=0) |
| 238 | return out |
| 239 | |
| 240 | |
| 241 | def slice_model_inputs( |
| 242 | model_inputs: dict[str, Any], token_slice: slice |
| 243 | ) -> dict[str, Any]: |
| 244 | """Narrow the model's per-token inputs to one microbatch.""" |
| 245 | sliced = dict(model_inputs) |
| 246 | for key in ("input_ids", "inputs_embeds"): |
| 247 | value = model_inputs.get(key) |
| 248 | if value is not None: |
| 249 | sliced[key] = value[token_slice] |
| 250 | |
| 251 | positions = model_inputs["positions"] |
| 252 | # M-RoPE carries a leading section dim. |
| 253 | sliced["positions"] = ( |
| 254 | positions[:, token_slice] if positions.ndim == 2 else positions[token_slice] |
| 255 | ) |
| 256 | |
| 257 | intermediate_tensors = model_inputs.get("intermediate_tensors") |
| 258 | if intermediate_tensors is not None: |
| 259 | sliced["intermediate_tensors"] = intermediate_tensors[token_slice] |
| 260 | return sliced |
| 261 | |
| 262 | |
| 263 | def merge_ubatch_outputs(outputs: list[Any]) -> Any: |
| 264 | """Reassemble the full-batch model output from the per-microbatch ones. |
| 265 | |
| 266 | Models return hidden states, a tuple of those plus auxiliary hidden states |
| 267 | (EAGLE3), or IntermediateTensors on non-final pipeline ranks. |
| 268 | """ |
| 269 | first = outputs[0] |
| 270 | if isinstance(first, IntermediateTensors): |
| 271 | return IntermediateTensors( |
| 272 | { |
| 273 | key: torch.cat([out.tensors[key] for out in outputs], dim=0) |
| 274 | for key in first.tensors |
| 275 | } |
| 276 | ) |
| 277 | if isinstance(first, tuple): |
| 278 | return tuple(torch.cat(parts, dim=0) for parts in zip(*outputs)) |
| 279 | return torch.cat(outputs, dim=0) |
| 280 | |
| 281 | |
| 282 | class UBatchRunner: |
| 283 | """Runs the model on two or more microbatches that overlap each other. |
| 284 | |
| 285 | Each microbatch gets a thread and its own forward context. Only one thread |
| 286 | holds the GPU at a time; they hand off to each other at the communication |
| 287 | points inside the model (see `vllm.v1.worker.ubatching`), so one |
| 288 | microbatch's expert all-to-all overlaps the other's compute. |
| 289 | """ |
| 290 | |
| 291 | def __init__( |
| 292 | self, |
| 293 | vllm_config: VllmConfig, |
| 294 | device: torch.device, |
| 295 | model_state: ModelState, |
| 296 | attn_groups: list[list[AttentionGroup]], |
| 297 | kv_cache_config: KVCacheConfig, |
| 298 | max_num_reqs: int, |
| 299 | ): |
| 300 | self.vllm_config = vllm_config |
| 301 | self.parallel_config = vllm_config.parallel_config |
| 302 | self.num_ubatches = self.parallel_config.num_ubatches |
| 303 | self.dcp_size = self.parallel_config.decode_context_parallel_size |
| 304 | self.dcp_rank = get_dcp_group().rank_in_group if self.dcp_size > 1 else 0 |
| 305 | self.cp_interleave = self.parallel_config.cp_kv_cache_interleave_size |
| 306 | self.device = device |
| 307 | self.model_state = model_state |
| 308 | self.attn_groups = attn_groups |
| 309 | self.kv_cache_config = kv_cache_config |
| 310 | # `query_start_loc` and `seq_lens` are rebased onto each microbatch's |
| 311 | # own token range, so they cannot be views of the full batch's buffers. |
| 312 | # Allocating up front keeps their addresses stable across replays. |
| 313 | self.ubatch_query_start_loc = [ |
| 314 | torch.zeros(max_num_reqs + 1, dtype=torch.int32, device=device) |
| 315 | for _ in range(self.num_ubatches) |
| 316 | ] |
| 317 | self.ubatch_seq_lens = [ |
| 318 | torch.zeros(max_num_reqs, dtype=torch.int32, device=device) |
| 319 | for _ in range(self.num_ubatches) |
| 320 | ] |
| 321 | self.ubatch_dcp_local_seq_lens = [ |
| 322 | torch.zeros(max_num_reqs, dtype=torch.int32, device=device) |
| 323 | if self.dcp_size > 1 |
| 324 | else None |
| 325 | for _ in range(self.num_ubatches) |
| 326 | ] |
| 327 | self.comm_stream = torch.cuda.Stream(device=device) |
| 328 | # Threads and graph capture must use the same compute stream. |
| 329 | self.capture_stream = torch.cuda.Stream(device=device) |
| 330 | # The microbatch threads plus the thread that starts them. |
| 331 | self.ready_barrier = threading.Barrier(self.num_ubatches + 1) |
| 332 | self._pending_finish: Callable[[], Any] | None = None |
| 333 | self.sm_control = create_sm_control_context(self.parallel_config) |
| 334 | |
| 335 | def prepare( |
| 336 | self, |
| 337 | input_batch: InputBatch, |
| 338 | block_tables: tuple[torch.Tensor, ...], |
| 339 | slot_mappings: torch.Tensor, |
| 340 | cg_mode: CUDAGraphMode = CUDAGraphMode.NONE, |
| 341 | for_capture: bool = False, |
| 342 | ) -> UBatchState: |
| 343 | """Split the batch into the microbatches the step will run on. |
| 344 | |
| 345 | Attention metadata is built per microbatch and carried in the forward |
| 346 | contexts the threads install, not in the caller's. |
| 347 | |
| 348 | FULL uses padded sizes; `for_capture` refreshes metadata from dummy |
| 349 | block tables. |
| 350 | """ |
| 351 | ubatch_slices = create_ubatch_slices(input_batch, self.num_ubatches) |
| 352 | |
| 353 | attn_metadata = [] |
| 354 | slot_mappings_by_layer = [] |
| 355 | is_padding = [] |
| 356 | for i, ubatch_slice in enumerate(ubatch_slices): |
| 357 | ubatch = _slice_input_batch( |
| 358 | input_batch, |
| 359 | ubatch_slice, |
| 360 | self.ubatch_query_start_loc[i], |
| 361 | self.ubatch_seq_lens[i], |
| 362 | self.ubatch_dcp_local_seq_lens[i], |
| 363 | dcp_size=self.dcp_size, |
| 364 | dcp_rank=self.dcp_rank, |
| 365 | cp_interleave=self.cp_interleave, |
| 366 | ) |
| 367 | ubatch_slot_mappings = slot_mappings[:, ubatch_slice.token_slice] |
| 368 | ubatch_block_tables = tuple( |
| 369 | block_table[ubatch_slice.request_slice] for block_table in block_tables |
| 370 | ) |
| 371 | attn_metadata.append( |
| 372 | self.model_state.prepare_attn( |
| 373 | ubatch, |
| 374 | cg_mode, |
| 375 | ubatch_block_tables, |
| 376 | ubatch_slot_mappings, |
| 377 | self.attn_groups, |
| 378 | self.kv_cache_config, |
| 379 | ubatch_idx=i, |
| 380 | for_capture=for_capture, |
| 381 | ) |
| 382 | ) |
| 383 | slot_mappings_by_layer.append( |
| 384 | build_slot_mappings_by_layer(ubatch_slot_mappings, self.kv_cache_config) |
| 385 | ) |
| 386 | is_padding.append(ubatch.is_padding) |
| 387 | |
| 388 | return UBatchState( |
| 389 | slices=ubatch_slices, |
| 390 | forward_contexts=self._make_forward_contexts( |
| 391 | ubatch_slices, attn_metadata, slot_mappings_by_layer, is_padding |
| 392 | ), |
| 393 | ) |
| 394 | |
| 395 | def _make_forward_contexts( |
| 396 | self, |
| 397 | ubatch_slices: UBatchSlices, |
| 398 | attn_metadata: list[dict[str, Any]], |
| 399 | slot_mappings_by_layer: list[dict[str, torch.Tensor]], |
| 400 | is_padding: list[torch.Tensor | None], |
| 401 | ) -> list[ForwardContext]: |
| 402 | """Build one forward context per microbatch. |
| 403 | |
| 404 | The expert all-to-all runs once per microbatch, so each context's DP |
| 405 | metadata carries that microbatch's token count. Microbatching pads every |
| 406 | DP rank alike, so each rank can assume that count for the others too. |
| 407 | """ |
| 408 | dp_size = self.parallel_config.data_parallel_size |
| 409 | forward_contexts = [] |
| 410 | for i, ubatch_slice in enumerate(ubatch_slices): |
| 411 | num_tokens = ubatch_slice.num_tokens |
| 412 | num_tokens_across_dp = torch.full( |
| 413 | (dp_size,), num_tokens, dtype=torch.int32, device="cpu" |
| 414 | ) |
| 415 | forward_contexts.append( |
| 416 | create_forward_context( |
| 417 | attn_metadata[i], |
| 418 | self.vllm_config, |
| 419 | dp_metadata=DPMetadata.make( |
| 420 | self.parallel_config, num_tokens, num_tokens_across_dp |
| 421 | ), |
| 422 | cudagraph_runtime_mode=CUDAGraphMode.NONE, |
| 423 | slot_mapping=slot_mappings_by_layer[i], |
| 424 | is_padding=is_padding[i], |
| 425 | ) |
| 426 | ) |
| 427 | return forward_contexts |
| 428 | |
| 429 | def run( |
| 430 | self, |
| 431 | model: Any, |
| 432 | model_inputs: dict[str, Any], |
| 433 | ubatch_state: UBatchState, |
| 434 | ) -> Any: |
| 435 | finish = self.begin_capturable_run(model, model_inputs, ubatch_state) |
| 436 | return finish() |
| 437 | |
| 438 | def begin_capturable_run( |
| 439 | self, |
| 440 | model: Any, |
| 441 | model_inputs: dict[str, Any], |
| 442 | ubatch_state: UBatchState, |
| 443 | for_capture: bool = False, |
| 444 | ) -> Callable[[], Any]: |
| 445 | """Start microbatch threads and return a callback to run and join them. |
| 446 | |
| 447 | For capture, start threads outside the graph and call the callback inside. |
| 448 | `for_capture` selects the capture stream and initializes cuBLAS workspaces |
| 449 | before capture. |
| 450 | |
| 451 | Call the callback exactly once on this thread, or call `abort_pending_run` |
| 452 | on failure to release parked threads. |
| 453 | """ |
| 454 | ubatch_slices = ubatch_state.slices |
| 455 | assert len(ubatch_slices) == len(ubatch_state.forward_contexts) |
| 456 | assert len(ubatch_slices) == self.num_ubatches |
| 457 | |
| 458 | compute_stream = self.capture_stream if for_capture else current_stream() |
| 459 | ubatch_contexts = make_ubatch_contexts( |
| 460 | num_micro_batches=self.num_ubatches, |
| 461 | comm_stream=self.comm_stream, |
| 462 | compute_stream=compute_stream, |
| 463 | forward_contexts=ubatch_state.forward_contexts, |
| 464 | ready_barrier=self.ready_barrier, |
| 465 | ) |
| 466 | |
| 467 | outputs: dict[int, Any] = {} |
| 468 | errors: dict[int, BaseException] = {} |
| 469 | |
| 470 | @torch.inference_mode() |
| 471 | def run_ubatch(ubatch_context, inputs: dict[str, Any]) -> None: |
| 472 | try: |
| 473 | # New threads must select the worker's device. |
| 474 | torch.accelerator.set_device_index(self.device.index) |
| 475 | if for_capture: |
| 476 | # Allocate cuBLAS workspaces before capture. |
| 477 | for stream in (compute_stream, self.comm_stream): |
| 478 | with torch.cuda.stream(stream): |
| 479 | torch.cuda.current_blas_handle() |
| 480 | with ubatch_context: |
| 481 | outputs[ubatch_context.id] = model(**inputs) |
| 482 | except BaseException as e: # noqa: BLE001 |
| 483 | # Recorded so the join below can report which microbatch broke. |
| 484 | # Note the siblings are not told to unwind: microbatches hand |
| 485 | # control to each other, so one that dies mid-forward leaves |
| 486 | # the others parked on a handoff that never comes and the step |
| 487 | # hangs. The V1 runner behaves the same way; fixing it means |
| 488 | # changing the shared handoff protocol in ubatching.py, which |
| 489 | # is deliberately out of scope here. |
| 490 | errors[ubatch_context.id] = e |
| 491 | |
| 492 | # Threads own the forward context; keep it cleared and SMs reserved |
| 493 | # until finish joins them. |
| 494 | stack = ExitStack() |
| 495 | stack.enter_context(override_forward_context(None)) |
| 496 | stack.enter_context(self.sm_control) |
| 497 | |
| 498 | threads = [] |
| 499 | for ubatch_context, ubatch_slice in zip(ubatch_contexts, ubatch_slices): |
| 500 | thread = threading.Thread( |
| 501 | target=run_ubatch, |
| 502 | args=( |
| 503 | ubatch_context, |
| 504 | slice_model_inputs(model_inputs, ubatch_slice.token_slice), |
| 505 | ), |
| 506 | ) |
| 507 | threads.append(thread) |
| 508 | thread.start() |
| 509 | |
| 510 | # Park all threads before the caller starts capture. |
| 511 | self.ready_barrier.wait() |
| 512 | |
| 513 | def finish() -> Any: |
| 514 | try: |
| 515 | ubatch_contexts[0].cpu_wait_event.set() |
| 516 | for thread in threads: |
| 517 | thread.join() |
| 518 | finally: |
| 519 | self._pending_finish = None |
| 520 | stack.close() |
| 521 | |
| 522 | if errors: |
| 523 | failed = min(errors) |
| 524 | raise RuntimeError( |
| 525 | f"Microbatch {failed} of {self.num_ubatches} failed" |
| 526 | ) from errors[failed] |
| 527 | return merge_ubatch_outputs([outputs[i] for i in range(self.num_ubatches)]) |
| 528 | |
| 529 | self._pending_finish = finish |
| 530 | return finish |
| 531 | |
| 532 | def abort_pending_run(self) -> None: |
| 533 | """Release and join parked threads after a failed capture.""" |
| 534 | if self._pending_finish is None: |
| 535 | return |
| 536 | try: |
| 537 | self._pending_finish() |
| 538 | except Exception: |
| 539 | logger.exception("Failed to join microbatch threads after an abort") |
| 540 | finally: |
| 541 | self._pending_finish = None |
| 542 | |
| 543 | |
| 544 | def maybe_build_ubatch_runner( |
| 545 | vllm_config: VllmConfig, |
| 546 | device: torch.device, |
| 547 | model_state: ModelState, |
| 548 | attn_groups: list[list[AttentionGroup]], |
| 549 | kv_cache_config: KVCacheConfig, |
| 550 | max_num_reqs: int, |
| 551 | ) -> UBatchRunner | None: |
| 552 | """Build the microbatch runner, or None when DBO is not in use. |
| 553 | |
| 554 | Microbatching needs the DP handshake to agree on it, so it is only |
| 555 | available with more than one DP rank (as in the V1 runner). |
| 556 | """ |
| 557 | parallel_config = vllm_config.parallel_config |
| 558 | if not parallel_config.use_ubatching or parallel_config.data_parallel_size <= 1: |
| 559 | return None |
| 560 | |
| 561 | logger.info_once("Dual batch overlap is enabled.") |
| 562 | return UBatchRunner( |
| 563 | vllm_config, device, model_state, attn_groups, kv_cache_config, max_num_reqs |
| 564 | ) |