| 1 | # SPDX-License-Identifier: Apache-2.0 |
| 2 | # SPDX-FileCopyrightText: Copyright contributors to the vLLM project |
| 3 | |
| 4 | import functools |
| 5 | import gc |
| 6 | import itertools |
| 7 | import threading |
| 8 | import time |
| 9 | from collections import defaultdict |
| 10 | from collections.abc import Callable, Iterable, Iterator, Sequence |
| 11 | from contextlib import AbstractContextManager, contextmanager, nullcontext |
| 12 | from copy import copy, deepcopy |
| 13 | from dataclasses import replace |
| 14 | from functools import reduce |
| 15 | from typing import TYPE_CHECKING, Any, NamedTuple, TypeAlias, cast |
| 16 | |
| 17 | import numpy as np |
| 18 | import torch |
| 19 | import torch.distributed |
| 20 | import torch.nn as nn |
| 21 | from tqdm import tqdm |
| 22 | |
| 23 | import vllm.envs as envs |
| 24 | from vllm.compilation.breakable_cudagraph import ( |
| 25 | BreakableCUDAGraphWrapper, |
| 26 | is_breakable_cudagraph_enabled, |
| 27 | ) |
| 28 | from vllm.compilation.counter import compilation_counter |
| 29 | from vllm.compilation.cuda_graph import CUDAGraphStat, CUDAGraphWrapper |
| 30 | from vllm.compilation.monitor import set_cudagraph_capturing_enabled |
| 31 | from vllm.config import ( |
| 32 | CompilationMode, |
| 33 | CUDAGraphMode, |
| 34 | VllmConfig, |
| 35 | get_layers_from_vllm_config, |
| 36 | set_current_vllm_config, |
| 37 | update_config, |
| 38 | ) |
| 39 | from vllm.config.cache import CacheConfig |
| 40 | from vllm.config.ec_manager_config import EncoderCacheManagerMetadata |
| 41 | from vllm.config.model import PROCESSED_LOGPROBS_MODES |
| 42 | from vllm.distributed.ec_transfer import get_ec_transfer, has_ec_transfer |
| 43 | from vllm.distributed.eplb.eplb_state import EplbState |
| 44 | from vllm.distributed.kv_transfer import get_kv_transfer_group, has_kv_transfer_group |
| 45 | from vllm.distributed.kv_transfer.kv_connector.utils import ( |
| 46 | copy_kv_blocks, |
| 47 | ) |
| 48 | from vllm.distributed.parallel_state import ( |
| 49 | get_dcp_group, |
| 50 | get_pp_group, |
| 51 | get_tp_group, |
| 52 | graph_capture, |
| 53 | is_global_first_rank, |
| 54 | ) |
| 55 | from vllm.forward_context import ( |
| 56 | BatchDescriptor, |
| 57 | set_forward_context, |
| 58 | ) |
| 59 | from vllm.logger import init_logger |
| 60 | from vllm.lora.layers import BaseLayerWithLoRA, LoRAMapping, LoRAMappingType |
| 61 | from vllm.model_executor.layers.attention import Attention |
| 62 | from vllm.model_executor.layers.attention_layer_base import AttentionLayerBase |
| 63 | from vllm.model_executor.layers.fused_moe.all2all_utils import get_ep_all2all_manager |
| 64 | from vllm.model_executor.layers.mamba.mamba_utils import MambaStateCopyFuncsByType |
| 65 | from vllm.model_executor.layers.mamba.ops.ssu_dispatch import ( |
| 66 | initialize_mamba_ssu_backend, |
| 67 | ) |
| 68 | from vllm.model_executor.layers.rotary_embedding import MRotaryEmbedding |
| 69 | from vllm.model_executor.model_loader import get_model_loader |
| 70 | from vllm.model_executor.model_loader.reload import ( |
| 71 | finalize_layerwise_reload, |
| 72 | initialize_layerwise_reload, |
| 73 | ) |
| 74 | from vllm.model_executor.models.interfaces import ( |
| 75 | MixtureOfExperts, |
| 76 | MultiModalEmbeddings, |
| 77 | SupportsMRoPE, |
| 78 | SupportsMultiModal, |
| 79 | get_mixture_of_experts_model, |
| 80 | supports_eagle3, |
| 81 | supports_mrope, |
| 82 | supports_multimodal_pruning, |
| 83 | supports_realtime, |
| 84 | supports_transcription, |
| 85 | ) |
| 86 | from vllm.model_executor.models.interfaces_base import ( |
| 87 | VllmModelForPooling, |
| 88 | is_pooling_model, |
| 89 | is_text_generation_model, |
| 90 | ) |
| 91 | from vllm.model_executor.offloader import ( |
| 92 | create_offloader, |
| 93 | get_offloader, |
| 94 | set_offloader, |
| 95 | ) |
| 96 | from vllm.model_executor.warmup.jit_warmup import JitWarmupRegistry |
| 97 | from vllm.multimodal import MULTIMODAL_REGISTRY |
| 98 | from vllm.multimodal.encoder_budget import MultiModalBudget |
| 99 | from vllm.multimodal.inputs import ( |
| 100 | BatchedTensorInputs, |
| 101 | MultiModalKwargsItem, |
| 102 | PlaceholderRange, |
| 103 | ) |
| 104 | from vllm.multimodal.utils import ( |
| 105 | copy_mm_embedding_modality, |
| 106 | get_mm_features_in_window, |
| 107 | group_and_batch_mm_kwargs, |
| 108 | set_mm_embedding_modality, |
| 109 | ) |
| 110 | from vllm.platforms import current_platform |
| 111 | from vllm.pooling_params import PoolingParams |
| 112 | from vllm.sampling_params import SamplingType |
| 113 | from vllm.sequence import IntermediateTensors |
| 114 | from vllm.tasks import GenerationTask, PoolingTask, SupportedTask |
| 115 | from vllm.tracing import instrument |
| 116 | from vllm.utils import length_from_prompt_token_ids_or_embeds |
| 117 | from vllm.utils.gc_utils import freeze_gc_for_cudagraph_capture |
| 118 | from vllm.utils.gpu_sync_debug import gpu_sync_allowed |
| 119 | from vllm.utils.math_utils import cdiv, round_up |
| 120 | from vllm.utils.mem_utils import DeviceMemoryProfiler, format_gib |
| 121 | from vllm.utils.nvtx_pytorch_hooks import PytHooks |
| 122 | from vllm.utils.platform_utils import num_compute_units |
| 123 | from vllm.utils.torch_utils import ( |
| 124 | PIN_MEMORY, |
| 125 | async_tensor_h2d, |
| 126 | kv_cache_dtype_str_to_dtype, |
| 127 | ) |
| 128 | from vllm.v1.attention.backend import ( |
| 129 | AttentionBackend, |
| 130 | AttentionCGSupport, |
| 131 | AttentionMetadata, |
| 132 | AttentionMetadataBuilder, |
| 133 | AttentionType, |
| 134 | CommonAttentionMetadata, |
| 135 | ) |
| 136 | from vllm.v1.attention.backends.gdn_attn import GDNAttentionMetadataBuilder |
| 137 | from vllm.v1.attention.backends.linear_attn import ( |
| 138 | BailingLinearAttentionMetadataBuilder, |
| 139 | ) |
| 140 | from vllm.v1.attention.backends.mamba2_attn import Mamba2AttentionMetadataBuilder |
| 141 | from vllm.v1.attention.backends.short_conv_attn import ShortConvAttentionMetadataBuilder |
| 142 | from vllm.v1.attention.backends.utils import ( |
| 143 | NULL_BLOCK_ID, |
| 144 | create_fast_prefill_custom_backend, |
| 145 | get_dcp_local_seq_lens, |
| 146 | reorder_batch_to_split_decodes_and_prefills, |
| 147 | ) |
| 148 | from vllm.v1.core.sched.output import NewRequestData |
| 149 | from vllm.v1.cudagraph_dispatcher import CudagraphDispatcher |
| 150 | from vllm.v1.kv_cache_interface import ( |
| 151 | AttentionSpec, |
| 152 | ChunkedLocalAttentionSpec, |
| 153 | CrossAttentionSpec, |
| 154 | EncoderOnlyAttentionSpec, |
| 155 | FullAttentionSpec, |
| 156 | KVCacheConfig, |
| 157 | KVCacheGroupSpec, |
| 158 | KVCacheSpec, |
| 159 | KVCacheSpecKind, |
| 160 | SlidingWindowSpec, |
| 161 | UniformTypeKVCacheSpecs, |
| 162 | get_kv_cache_spec_kind, |
| 163 | ) |
| 164 | from vllm.v1.kv_cache_spec_registry import KVCacheSpecRegistry |
| 165 | from vllm.v1.outputs import ( |
| 166 | EMPTY_MODEL_RUNNER_OUTPUT, |
| 167 | AsyncModelRunnerOutput, |
| 168 | DraftTokenIds, |
| 169 | ECConnectorOutput, |
| 170 | KVConnectorOutput, |
| 171 | LogprobsLists, |
| 172 | LogprobsTensors, |
| 173 | ModelRunnerOutput, |
| 174 | PoolerOutput, |
| 175 | SamplerOutput, |
| 176 | make_empty_encoder_model_runner_output, |
| 177 | ) |
| 178 | from vllm.v1.pool.late_interaction_runner import LateInteractionRunner |
| 179 | from vllm.v1.pool.metadata import PoolingMetadata, PoolingStates |
| 180 | from vllm.v1.sample.logits_processor import LogitsProcessors, build_logitsprocs |
| 181 | from vllm.v1.sample.logits_processor.interface import LogitsProcessor |
| 182 | from vllm.v1.sample.metadata import SamplingMetadata |
| 183 | from vllm.v1.sample.rejection_sampler import RejectionSampler |
| 184 | from vllm.v1.sample.sampler import Sampler |
| 185 | from vllm.v1.spec_decode.custom_class_proposer import create_custom_proposer |
| 186 | from vllm.v1.spec_decode.dflash import DFlashProposer |
| 187 | from vllm.v1.spec_decode.draft_model import DraftModelProposer |
| 188 | from vllm.v1.spec_decode.eagle import EagleProposer |
| 189 | from vllm.v1.spec_decode.extract_hidden_states import ExtractHiddenStatesProposer |
| 190 | from vllm.v1.spec_decode.gemma4 import Gemma4Proposer |
| 191 | from vllm.v1.spec_decode.medusa import MedusaProposer |
| 192 | from vllm.v1.spec_decode.metadata import SpecDecodeMetadata |
| 193 | from vllm.v1.spec_decode.ngram_proposer_gpu import ( |
| 194 | NgramProposerGPU, |
| 195 | copy_num_valid_draft_tokens, |
| 196 | update_ngram_gpu_tensors_incremental, |
| 197 | update_scheduler_for_invalid_drafts, |
| 198 | ) |
| 199 | from vllm.v1.spec_decode.step3p5 import Step3p5MTPProposer |
| 200 | from vllm.v1.spec_decode.suffix_decoding import SuffixDecodingProposer |
| 201 | from vllm.v1.spec_decode.utils import update_num_computed_tokens_for_batch_change |
| 202 | from vllm.v1.structured_output.utils import apply_grammar_bitmask |
| 203 | from vllm.v1.utils import CpuGpuBuffer, record_function_or_nullcontext |
| 204 | from vllm.v1.worker import mamba_utils |
| 205 | from vllm.v1.worker.block_table import SlotMappingMode |
| 206 | from vllm.v1.worker.cp_utils import ( |
| 207 | check_attention_cp_compatibility, |
| 208 | get_dcp_dummy_context_len, |
| 209 | prepare_dcp_dummy_context_metadata, |
| 210 | ) |
| 211 | from vllm.v1.worker.dp_utils import coordinate_batch_across_dp |
| 212 | from vllm.v1.worker.ec_connector_model_runner_mixin import ECConnectorModelRunnerMixin |
| 213 | from vllm.v1.worker.gpu_input_batch import CachedRequestState, InputBatch |
| 214 | from vllm.v1.worker.gpu_ubatch_wrapper import UBatchWrapper |
| 215 | from vllm.v1.worker.kv_connector_model_runner_mixin import KVConnectorModelRunnerMixin |
| 216 | from vllm.v1.worker.lora_model_runner_mixin import LoRAModelRunnerMixin |
| 217 | from vllm.v1.worker.ubatch_utils import ( |
| 218 | UBatchSlices, |
| 219 | check_ubatch_thresholds, |
| 220 | maybe_create_ubatch_slices, |
| 221 | split_attn_metadata, |
| 222 | ) |
| 223 | from vllm.v1.worker.utils import ( |
| 224 | EncoderTimingStats, |
| 225 | is_residual_scattered_for_sp, |
| 226 | raise_if_nan_logits, |
| 227 | ) |
| 228 | from vllm.v1.worker.workspace import lock_workspace |
| 229 | |
| 230 | from .utils import ( |
| 231 | AttentionGroup, |
| 232 | KVBlockZeroer, |
| 233 | add_kv_sharing_layers_to_kv_cache_groups, |
| 234 | allocate_kv_cache, |
| 235 | bind_kv_cache, |
| 236 | copy_kv_cache_blocks_inplace, |
| 237 | prepare_kernel_block_sizes, |
| 238 | sanity_check_mm_encoder_outputs, |
| 239 | ) |
| 240 | |
| 241 | if TYPE_CHECKING: |
| 242 | from vllm.v1.core.sched.output import GrammarOutput, SchedulerOutput |
| 243 | from vllm.v1.spec_decode.ngram_proposer import NgramProposer |
| 244 | from vllm.v1.worker.encoder_cudagraph import EncoderCudaGraphManager |
| 245 | |
| 246 | logger = init_logger(__name__) |
| 247 | |
| 248 | |
| 249 | def _get_parameter_for_reload(model: nn.Module, name: str) -> nn.Parameter: |
| 250 | """Resolve checkpoint names without changing the model's module tree.""" |
| 251 | module_name, _, parameter_name = name.rpartition(".") |
| 252 | module = model.get_submodule(module_name) |
| 253 | if isinstance(module, BaseLayerWithLoRA): |
| 254 | module = module.base_layer |
| 255 | return module.get_parameter(parameter_name) |
| 256 | |
| 257 | |
| 258 | AttnMetadataDict: TypeAlias = dict[str, AttentionMetadata] |
| 259 | # list when ubatching is enabled |
| 260 | PerLayerAttnMetadata: TypeAlias = list[AttnMetadataDict] | AttnMetadataDict |
| 261 | |
| 262 | |
| 263 | def count_nans_per_row(logits: torch.Tensor) -> torch.Tensor: |
| 264 | """Per-row NaN counts, left on device.""" |
| 265 | return logits.isnan().sum(dim=-1, dtype=torch.int32) |
| 266 | |
| 267 | |
| 268 | def nans_to_dict(counts: list[int], req_id_to_index: dict[str, int]) -> dict[str, int]: |
| 269 | """Map per-row NaN counts onto request ids. |
| 270 | |
| 271 | Rows and requests need not correspond one-to-one, so requests without a |
| 272 | row are reported as 0. |
| 273 | """ |
| 274 | return { |
| 275 | req_id: int(counts[i]) if i < len(counts) else 0 |
| 276 | for req_id, i in req_id_to_index.items() |
| 277 | } |
| 278 | |
| 279 | |
| 280 | # Wrapper for ModelRunnerOutput to support overlapped execution. |
| 281 | class AsyncGPUModelRunnerOutput(AsyncModelRunnerOutput): |
| 282 | def __init__( |
| 283 | self, |
| 284 | model_runner_output: ModelRunnerOutput, |
| 285 | sampled_token_ids: torch.Tensor, |
| 286 | logprobs_tensors: LogprobsTensors | None, |
| 287 | invalid_req_indices: list[int], |
| 288 | async_output_copy_stream: torch.cuda.Stream, |
| 289 | vocab_size: int, |
| 290 | check_ep_fault: bool = False, |
| 291 | num_nans: torch.Tensor | None = None, |
| 292 | ): |
| 293 | self._model_runner_output = model_runner_output |
| 294 | self._invalid_req_indices = invalid_req_indices |
| 295 | |
| 296 | # Event on the copy stream so we can synchronize the non-blocking copy. |
| 297 | # Blocking (sleep) event to avoid busy-polling the CUDA driver lock. |
| 298 | self.async_copy_ready_event = torch.cuda.Event(blocking=True) |
| 299 | |
| 300 | # Keep a reference to the device tensor to avoid it being |
| 301 | # deallocated until we finish copying it to the host. |
| 302 | self._sampled_token_ids = sampled_token_ids |
| 303 | self.vocab_size = vocab_size |
| 304 | self._logprobs_tensors = logprobs_tensors |
| 305 | self._num_nans = num_nans |
| 306 | self._has_fault: torch.Tensor | None = None |
| 307 | |
| 308 | # Initiate the copy on a separate stream, but do not synchronize it. |
| 309 | default_stream = torch.cuda.current_stream() |
| 310 | with torch.cuda.stream(async_output_copy_stream): |
| 311 | async_output_copy_stream.wait_stream(default_stream) |
| 312 | self.sampled_token_ids_cpu = self._sampled_token_ids.to( |
| 313 | "cpu", non_blocking=True |
| 314 | ) |
| 315 | self._logprobs_tensors_cpu = ( |
| 316 | self._logprobs_tensors.to_cpu_nonblocking() |
| 317 | if self._logprobs_tensors |
| 318 | else None |
| 319 | ) |
| 320 | self._num_nans_cpu = ( |
| 321 | self._num_nans.to("cpu", non_blocking=True) |
| 322 | if self._num_nans is not None |
| 323 | else None |
| 324 | ) |
| 325 | if check_ep_fault: |
| 326 | has_fault = get_ep_all2all_manager().query_fault() |
| 327 | self._has_fault = has_fault.to("cpu", non_blocking=True) |
| 328 | self.async_copy_ready_event.record() |
| 329 | |
| 330 | def get_output(self) -> ModelRunnerOutput: |
| 331 | """Copy the device tensors to the host and return a ModelRunnerOutput. |
| 332 | |
| 333 | This function blocks until the copy is finished. |
| 334 | """ |
| 335 | max_gen_len = self.sampled_token_ids_cpu.shape[-1] |
| 336 | self.async_copy_ready_event.synchronize() |
| 337 | |
| 338 | # Release the device tensors once the copy has completed. |
| 339 | del self._logprobs_tensors |
| 340 | del self._sampled_token_ids |
| 341 | if max_gen_len == 1: |
| 342 | valid_sampled_token_ids = self.sampled_token_ids_cpu.tolist() |
| 343 | for i in self._invalid_req_indices: |
| 344 | valid_sampled_token_ids[i].clear() |
| 345 | logprobs_lists = None |
| 346 | if self._logprobs_tensors_cpu is not None: |
| 347 | logprobs_lists = self._logprobs_tensors_cpu.tolists() |
| 348 | else: |
| 349 | valid_sampled_token_ids, logprobs_lists = RejectionSampler.parse_output( |
| 350 | self.sampled_token_ids_cpu, |
| 351 | self.vocab_size, |
| 352 | self._invalid_req_indices, |
| 353 | logprobs_tensors=self._logprobs_tensors_cpu, |
| 354 | ) |
| 355 | |
| 356 | output = self._model_runner_output |
| 357 | output.sampled_token_ids = valid_sampled_token_ids |
| 358 | output.logprobs = logprobs_lists |
| 359 | |
| 360 | if self._num_nans_cpu is not None: |
| 361 | output.num_nans_in_logits = nans_to_dict( |
| 362 | self._num_nans_cpu.tolist(), output.req_id_to_index |
| 363 | ) |
| 364 | if envs.VLLM_RAISE_ON_LOGIT_NANS: |
| 365 | raise_if_nan_logits(output.num_nans_in_logits) |
| 366 | del self._num_nans |
| 367 | |
| 368 | if self._has_fault is not None and self._has_fault.item(): |
| 369 | mask = get_ep_all2all_manager().query_active_mask() |
| 370 | raise RuntimeError( |
| 371 | "Fault detected in EP all2all communication: " |
| 372 | "one or more ranks timed out during dispatch/combine. " |
| 373 | f"Mask: {mask.cpu().tolist()}" |
| 374 | ) |
| 375 | |
| 376 | return output |
| 377 | |
| 378 | |
| 379 | def _copy_pooler_output_to_cpu( |
| 380 | raw_pooler_output: PoolerOutput, finished_mask: list[bool] |
| 381 | ) -> list[torch.Tensor | None]: |
| 382 | num_reqs = len(finished_mask) |
| 383 | |
| 384 | if isinstance(raw_pooler_output, torch.Tensor): |
| 385 | if raw_pooler_output.shape[0] != num_reqs: |
| 386 | raise ValueError( |
| 387 | "Pooler output batch size does not match finished mask size: " |
| 388 | f"{raw_pooler_output.shape[0]} != {num_reqs}." |
| 389 | ) |
| 390 | |
| 391 | num_finished = sum(finished_mask) |
| 392 | if num_finished == 0: |
| 393 | return [None] * num_reqs |
| 394 | if num_finished == num_reqs: |
| 395 | return list(raw_pooler_output.to("cpu", non_blocking=True)) |
| 396 | |
| 397 | # partial finished |
| 398 | finished_indices = [i for i, include in enumerate(finished_mask) if include] |
| 399 | index_tensor = torch.tensor( |
| 400 | finished_indices, device=raw_pooler_output.device, dtype=torch.long |
| 401 | ) |
| 402 | finished_outputs = raw_pooler_output.index_select(0, index_tensor).to( |
| 403 | "cpu", non_blocking=True |
| 404 | ) |
| 405 | partial_pooler_output: list[torch.Tensor | None] = [None] * num_reqs |
| 406 | for i, out in zip(finished_indices, finished_outputs): |
| 407 | partial_pooler_output[i] = out |
| 408 | return partial_pooler_output |
| 409 | |
| 410 | assert isinstance(raw_pooler_output, list) |
| 411 | if len(raw_pooler_output) != num_reqs: |
| 412 | raise ValueError( |
| 413 | "Pooler output batch size does not match finished mask size: " |
| 414 | f"{len(raw_pooler_output)} != {num_reqs}." |
| 415 | ) |
| 416 | |
| 417 | pooler_output: list[torch.Tensor | None] = [None] * num_reqs |
| 418 | for i, (out, include) in enumerate(zip(raw_pooler_output, finished_mask)): |
| 419 | if include and out is not None: |
| 420 | pooler_output[i] = out.to("cpu", non_blocking=True) |
| 421 | return pooler_output |
| 422 | |
| 423 | |
| 424 | class AsyncGPUPoolingModelRunnerOutput(AsyncModelRunnerOutput): |
| 425 | def __init__( |
| 426 | self, |
| 427 | model_runner_output: ModelRunnerOutput, |
| 428 | raw_pooler_output: PoolerOutput, |
| 429 | finished_mask: list[bool], |
| 430 | async_output_copy_stream: torch.cuda.Stream, |
| 431 | ): |
| 432 | self._model_runner_output = model_runner_output |
| 433 | |
| 434 | # Event on the copy stream so we can synchronize the non-blocking copy. |
| 435 | # Blocking (sleep) event to avoid busy-polling the CUDA driver lock. |
| 436 | self.async_copy_ready_event = torch.cuda.Event(blocking=True) |
| 437 | |
| 438 | # Keep a reference to the device tensors to avoid them being |
| 439 | # deallocated until we finish copying it to the host. |
| 440 | self._raw_pooler_output = raw_pooler_output |
| 441 | |
| 442 | # Initiate the copy on a separate stream, but do not synchronize it. |
| 443 | default_stream = torch.cuda.current_stream() |
| 444 | with torch.cuda.stream(async_output_copy_stream): |
| 445 | async_output_copy_stream.wait_stream(default_stream) |
| 446 | self._model_runner_output.pooler_output = _copy_pooler_output_to_cpu( |
| 447 | raw_pooler_output=self._raw_pooler_output, |
| 448 | finished_mask=finished_mask, |
| 449 | ) |
| 450 | self.async_copy_ready_event.record() |
| 451 | |
| 452 | def get_output(self) -> ModelRunnerOutput: |
| 453 | """Copy the device tensors to the host and return a ModelRunnerOutput. |
| 454 | This function blocks until the copy is finished. |
| 455 | """ |
| 456 | self.async_copy_ready_event.synchronize() |
| 457 | |
| 458 | # Release the device tensors once the copy has completed. |
| 459 | del self._raw_pooler_output |
| 460 | return self._model_runner_output |
| 461 | |
| 462 | |
| 463 | class ExecuteModelState(NamedTuple): |
| 464 | """Ephemeral cached state transferred between execute_model() and |
| 465 | sample_tokens(), after execute_model() returns None.""" |
| 466 | |
| 467 | scheduler_output: "SchedulerOutput" |
| 468 | logits: torch.Tensor |
| 469 | spec_decode_metadata: SpecDecodeMetadata | None |
| 470 | spec_decode_common_attn_metadata: CommonAttentionMetadata | None |
| 471 | hidden_states: torch.Tensor |
| 472 | sample_hidden_states: torch.Tensor |
| 473 | aux_hidden_states: list[torch.Tensor] | None |
| 474 | ec_connector_output: ECConnectorOutput | None |
| 475 | cudagraph_stats: CUDAGraphStat | None |
| 476 | slot_mappings: dict[str, torch.Tensor] | list[dict[str, torch.Tensor]] | None |
| 477 | |
| 478 | |
| 479 | class GPUModelRunner( |
| 480 | LoRAModelRunnerMixin, KVConnectorModelRunnerMixin, ECConnectorModelRunnerMixin |
| 481 | ): |
| 482 | @JitWarmupRegistry.capture |
| 483 | def __init__( |
| 484 | self, |
| 485 | vllm_config: VllmConfig, |
| 486 | device: torch.device, |
| 487 | ): |
| 488 | self.vllm_config = vllm_config |
| 489 | self.model_config = vllm_config.model_config |
| 490 | self.cache_config = vllm_config.cache_config |
| 491 | self.offload_config = vllm_config.offload_config |
| 492 | self.compilation_config = vllm_config.compilation_config |
| 493 | self.lora_config = vllm_config.lora_config |
| 494 | self.load_config = vllm_config.load_config |
| 495 | self.parallel_config = vllm_config.parallel_config |
| 496 | self.scheduler_config = vllm_config.scheduler_config |
| 497 | self.speculative_config = vllm_config.speculative_config |
| 498 | self.observability_config = vllm_config.observability_config |
| 499 | |
| 500 | model_config = self.model_config |
| 501 | cache_config = self.cache_config |
| 502 | scheduler_config = self.scheduler_config |
| 503 | parallel_config = self.parallel_config |
| 504 | self.device = device |
| 505 | self.dtype = self.model_config.dtype |
| 506 | |
| 507 | self.check_ep_fault = False |
| 508 | if parallel_config.data_parallel_size > 1 and self.model_config.is_moe: |
| 509 | self.check_ep_fault = get_ep_all2all_manager().support_fault_tolerance |
| 510 | |
| 511 | self.kv_cache_dtype = kv_cache_dtype_str_to_dtype( |
| 512 | cache_config.cache_dtype, self.model_config |
| 513 | ) |
| 514 | |
| 515 | self.is_pooling_model = model_config.runner_type == "pooling" |
| 516 | self.enable_prompt_embeds = model_config.enable_prompt_embeds |
| 517 | self.is_multimodal_raw_input_only_model = ( |
| 518 | model_config.is_multimodal_raw_input_only_model |
| 519 | ) |
| 520 | # These will be overridden in load_model() |
| 521 | self.is_multimodal_pruning_enabled = False |
| 522 | self.requires_sequential_video_encoding = False |
| 523 | self.max_model_len = model_config.max_model_len |
| 524 | |
| 525 | # Always set to false after the first forward pass |
| 526 | self.dcp_world_size = self.parallel_config.decode_context_parallel_size |
| 527 | self.dcp_rank = 0 if self.dcp_world_size <= 1 else get_dcp_group().rank_in_group |
| 528 | self.max_num_tokens = scheduler_config.max_num_batched_tokens |
| 529 | self.max_num_reqs = scheduler_config.max_num_seqs |
| 530 | |
| 531 | # Broadcast PP output for external_launcher (torchrun) |
| 532 | # to make sure we are synced across pp ranks |
| 533 | # TODO: Support overlapping micro-batches |
| 534 | # https://github.com/vllm-project/vllm/issues/18019 |
| 535 | self.broadcast_pp_output = ( |
| 536 | self.parallel_config.distributed_executor_backend == "external_launcher" |
| 537 | and len(get_pp_group().ranks) > 1 |
| 538 | ) |
| 539 | |
| 540 | # Model-related. |
| 541 | self.num_query_heads = model_config.get_num_attention_heads(parallel_config) |
| 542 | self.inputs_embeds_size = model_config.get_inputs_embeds_size() |
| 543 | # Only relevant for models using ALiBi (e.g, MPT) |
| 544 | self.use_alibi = model_config.uses_alibi |
| 545 | |
| 546 | self.cascade_attn_enabled = not self.model_config.disable_cascade_attn |
| 547 | self.is_mm_prefix_lm = self.model_config.is_mm_prefix_lm |
| 548 | |
| 549 | # Multi-modal data support |
| 550 | self.mm_registry = MULTIMODAL_REGISTRY |
| 551 | self.uses_mrope = model_config.uses_mrope |
| 552 | self.mrope_num_dims = model_config.mrope_num_dims |
| 553 | self.supports_mm_inputs = model_config.supports_multimodal_inputs |
| 554 | |
| 555 | if self.model_config.is_encoder_decoder: |
| 556 | # Maximum length of the encoder input, only for encoder-decoder |
| 557 | # models. |
| 558 | self.max_encoder_len = scheduler_config.max_num_encoder_input_tokens |
| 559 | else: |
| 560 | self.max_encoder_len = 0 |
| 561 | |
| 562 | # Async scheduling |
| 563 | self.use_async_scheduling = self.scheduler_config.async_scheduling |
| 564 | |
| 565 | # Async PP broadcast of sampled token ids, waited on in _prepare_input_ids. |
| 566 | self._pp_recv_work: torch.distributed.Work | None = None |
| 567 | |
| 568 | # Sampler |
| 569 | self.sampler = Sampler( |
| 570 | logprobs_mode=self.model_config.logprobs_mode, |
| 571 | use_fp64_gumbel=self.model_config.use_fp64_gumbel, |
| 572 | ) |
| 573 | |
| 574 | self.eplb_state: EplbState | None = None |
| 575 | self._moe_model: MixtureOfExperts | None = None |
| 576 | # NOTE(yongji): flag to temporarily disable EPLB during scaling up/down |
| 577 | self.eep_eplb_suppressed = False |
| 578 | """ |
| 579 | State of the expert parallelism load balancer. |
| 580 | |
| 581 | Will be lazily initialized when the model is loaded. |
| 582 | """ |
| 583 | |
| 584 | # Lazy initializations |
| 585 | # self.model: nn.Module # Set after load_model |
| 586 | # Initialize in initialize_kv_cache |
| 587 | self.kv_caches: list[torch.Tensor] = [] |
| 588 | # indexes: [kv_cache_group_id][attn_group] |
| 589 | self.attn_groups: list[list[AttentionGroup]] = [] |
| 590 | # self.kv_cache_config: KVCacheConfig |
| 591 | |
| 592 | # mm_hash -> encoder_output |
| 593 | self.encoder_cache: dict[str, torch.Tensor] = {} |
| 594 | self.late_interaction_runner = LateInteractionRunner() |
| 595 | |
| 596 | # Encoder CUDA graph manager (initialized after model load if enabled) |
| 597 | self.encoder_cudagraph_manager: EncoderCudaGraphManager | None = None |
| 598 | |
| 599 | self.use_aux_hidden_state_outputs = False |
| 600 | # Set up speculative decoding. |
| 601 | # NOTE(Jiayi): currently we put the entire draft model on |
| 602 | # the last PP rank. This is not ideal if there are many |
| 603 | # layers in the draft model. |
| 604 | if self.speculative_config and get_pp_group().is_last_rank: |
| 605 | self.drafter: ( |
| 606 | NgramProposer # noqa: F823 |
| 607 | | NgramProposerGPU |
| 608 | | SuffixDecodingProposer |
| 609 | | EagleProposer |
| 610 | | DFlashProposer |
| 611 | | DraftModelProposer |
| 612 | | MedusaProposer |
| 613 | | ExtractHiddenStatesProposer |
| 614 | | Gemma4Proposer |
| 615 | | Step3p5MTPProposer |
| 616 | ) |
| 617 | if self.speculative_config.method == "custom_class": |
| 618 | self.drafter = create_custom_proposer( # type: ignore[assignment] |
| 619 | self.vllm_config |
| 620 | ) |
| 621 | elif self.speculative_config.method == "ngram": |
| 622 | from vllm.v1.spec_decode.ngram_proposer import NgramProposer |
| 623 | |
| 624 | self.drafter = NgramProposer(self.vllm_config) |
| 625 | elif self.speculative_config.uses_draft_model(): |
| 626 | self.drafter = DraftModelProposer( |
| 627 | vllm_config=self.vllm_config, |
| 628 | device=self.device, |
| 629 | runner=self, |
| 630 | ) |
| 631 | elif self.speculative_config.use_ngram_gpu(): |
| 632 | self.drafter = NgramProposerGPU(self.vllm_config, self.device, self) |
| 633 | self.num_tokens_no_spec_gpu = torch.zeros( |
| 634 | self.max_num_reqs, dtype=torch.int32, device=device |
| 635 | ) |
| 636 | self.token_ids_gpu_tensor = torch.zeros( |
| 637 | self.max_num_reqs, |
| 638 | self.max_model_len, |
| 639 | dtype=torch.int32, |
| 640 | device=device, |
| 641 | ) |
| 642 | self._ngram_pinned_idx_buf = torch.zeros( |
| 643 | self.max_num_reqs, dtype=torch.long, pin_memory=True |
| 644 | ) |
| 645 | self._ngram_pinned_val_buf = torch.zeros( |
| 646 | self.max_num_reqs, dtype=torch.int32, pin_memory=True |
| 647 | ) |
| 648 | elif self.speculative_config.use_gemma4_mtp(): |
| 649 | self.drafter = Gemma4Proposer(self.vllm_config, self.device, self) |
| 650 | elif self.speculative_config.use_step3p5_mtp(): |
| 651 | self.drafter = Step3p5MTPProposer(self.vllm_config, self.device, self) |
| 652 | elif self.speculative_config.use_dflash(): |
| 653 | self.drafter = DFlashProposer(self.vllm_config, self.device, self) |
| 654 | self.use_aux_hidden_state_outputs = True |
| 655 | elif self.speculative_config.method == "suffix": |
| 656 | self.drafter = SuffixDecodingProposer(self.vllm_config) |
| 657 | elif self.speculative_config.use_eagle(): |
| 658 | self.drafter = EagleProposer(self.vllm_config, self.device, self) |
| 659 | if self.speculative_config.method == "eagle3": |
| 660 | self.use_aux_hidden_state_outputs = ( |
| 661 | self.drafter.eagle3_use_aux_hidden_state |
| 662 | ) |
| 663 | elif self.speculative_config.method == "medusa": |
| 664 | self.drafter = MedusaProposer( |
| 665 | vllm_config=self.vllm_config, device=self.device |
| 666 | ) |
| 667 | elif self.speculative_config.method == "extract_hidden_states": |
| 668 | self.drafter = ExtractHiddenStatesProposer( |
| 669 | vllm_config=self.vllm_config, device=self.device |
| 670 | ) |
| 671 | self.use_aux_hidden_state_outputs = True |
| 672 | else: |
| 673 | raise ValueError( |
| 674 | "Unknown speculative decoding method: " |
| 675 | f"{self.speculative_config.method}" |
| 676 | ) |
| 677 | self.rejection_sampler = RejectionSampler( |
| 678 | self.sampler, self.speculative_config, self.device |
| 679 | ) |
| 680 | |
| 681 | self.num_spec_tokens = 0 |
| 682 | self.prev_num_spec_tokens = 0 |
| 683 | self.valid_sampled_token_count_gpu: torch.Tensor | None = None |
| 684 | if self.speculative_config: |
| 685 | self.num_spec_tokens = self.speculative_config.num_speculative_tokens |
| 686 | self.prev_num_spec_tokens = self.num_spec_tokens |
| 687 | draft_config = self.speculative_config.draft_model_config |
| 688 | if draft_config is not None and draft_config.max_model_len is not None: |
| 689 | self.effective_drafter_max_model_len = draft_config.max_model_len |
| 690 | else: |
| 691 | self.effective_drafter_max_model_len = self.max_model_len |
| 692 | self.use_async_spec_decode = ( |
| 693 | self.use_async_scheduling and self.num_spec_tokens > 0 |
| 694 | ) |
| 695 | |
| 696 | # Request states. |
| 697 | self.requests: dict[str, CachedRequestState] = {} |
| 698 | # NOTE(rob): num_prompt_logprobs only includes reqs |
| 699 | # that are currently in the prefill phase. |
| 700 | self.num_prompt_logprobs: dict[str, int] = {} |
| 701 | |
| 702 | # Input Batch |
| 703 | # NOTE(Chen): Ideally, we should initialize the input batch inside |
| 704 | # `initialize_kv_cache` based on the kv cache config. However, as in |
| 705 | # https://github.com/vllm-project/vllm/pull/18298, due to some unknown |
| 706 | # reasons, we have to initialize the input batch before `load_model`, |
| 707 | # quantization + weight offloading will fail otherwise. As a temporary |
| 708 | # solution, we initialize the input batch here, and re-initialize it |
| 709 | # in `initialize_kv_cache` if the block_sizes here is different from |
| 710 | # the block_sizes in the kv cache config. |
| 711 | logits_processors = model_config.logits_processors |
| 712 | custom_logitsprocs: Sequence[str | type[LogitsProcessor]] = ( |
| 713 | tuple(logits_processors) if logits_processors is not None else () |
| 714 | ) |
| 715 | placeholder_block_size = ( |
| 716 | self.cache_config.block_size or CacheConfig.DEFAULT_BLOCK_SIZE |
| 717 | ) |
| 718 | placeholder_max_num_blocks = cdiv( |
| 719 | max(self.max_model_len, self.max_encoder_len), placeholder_block_size |
| 720 | ) |
| 721 | self._init_block_sizes = [placeholder_block_size] |
| 722 | self._init_kernel_block_sizes = [placeholder_block_size] |
| 723 | self._init_max_num_blocks = [placeholder_max_num_blocks] |
| 724 | self._init_slot_mapping_modes = [SlotMappingMode.TOKEN_TO_KV_SLOT] |
| 725 | self.cp_kv_cache_interleave_size = ( |
| 726 | self.parallel_config.cp_kv_cache_interleave_size |
| 727 | ) |
| 728 | # Capture warmup providers registered by the initial placeholder InputBatch |
| 729 | self.input_batch = InputBatch( |
| 730 | max_num_reqs=self.max_num_reqs, |
| 731 | # We need to use the encoder length for encoder-decoder |
| 732 | # because of KV cache for cross-attention. |
| 733 | max_model_len=max(self.max_model_len, self.max_encoder_len), |
| 734 | max_num_batched_tokens=self.max_num_tokens, |
| 735 | device=self.device, |
| 736 | vocab_size=self.model_config.get_vocab_size(), |
| 737 | block_sizes=[placeholder_block_size], |
| 738 | kernel_block_sizes=[placeholder_block_size], |
| 739 | max_num_blocks_per_req=[placeholder_max_num_blocks], |
| 740 | num_spec_tokens=self.num_spec_tokens, |
| 741 | logitsprocs=build_logitsprocs( |
| 742 | self.vllm_config, |
| 743 | self.device, |
| 744 | PIN_MEMORY, |
| 745 | self.is_pooling_model, |
| 746 | custom_logitsprocs, |
| 747 | ), |
| 748 | # We currently don't know whether a particular custom logits processor |
| 749 | # uses output token ids so we set this conservatively. Thinking-budget |
| 750 | # tracking is requested dynamically when a budgeted request is in the |
| 751 | # batch. |
| 752 | logitsprocs_need_output_token_ids=bool(custom_logitsprocs), |
| 753 | is_pooling_model=self.is_pooling_model, |
| 754 | cp_kv_cache_interleave_size=self.parallel_config.cp_kv_cache_interleave_size, |
| 755 | reasoning_config=self.vllm_config.reasoning_config, |
| 756 | use_replayssm=self.cache_config.use_replayssm, |
| 757 | ) |
| 758 | |
| 759 | # Separate cuda stream for overlapping transfer of sampled token ids from |
| 760 | # GPU to CPU when async scheduling is enabled. |
| 761 | self.async_output_copy_stream: torch.cuda.Stream | None = None |
| 762 | # cuda event to synchronize use of reused CPU tensors between steps |
| 763 | # when async scheduling is enabled. |
| 764 | self.prepare_inputs_event: torch.Event | None = None |
| 765 | if self.use_async_scheduling: |
| 766 | self.async_output_copy_stream = torch.cuda.Stream() |
| 767 | # Blocking (sleep) event to avoid busy-polling the CUDA driver lock; |
| 768 | # under TP contention that spin can balloon and make the rank a straggler. |
| 769 | self.prepare_inputs_event = torch.cuda.Event(blocking=True) |
| 770 | |
| 771 | # self.cudagraph_batch_sizes sorts in ascending order. |
| 772 | if ( |
| 773 | self.compilation_config.cudagraph_capture_sizes |
| 774 | and self.compilation_config.cudagraph_mode != CUDAGraphMode.NONE |
| 775 | ): |
| 776 | self.cudagraph_batch_sizes = sorted( |
| 777 | self.compilation_config.cudagraph_capture_sizes |
| 778 | ) |
| 779 | else: |
| 780 | self.cudagraph_batch_sizes = [] |
| 781 | |
| 782 | # Cache the device properties. |
| 783 | self._init_device_properties() |
| 784 | |
| 785 | # Encoder timing registry for observability |
| 786 | self.encoder_timing_registry: dict[str, EncoderTimingStats] = {} |
| 787 | self._encoder_timing_lock = threading.Lock() |
| 788 | |
| 789 | # Persistent buffers for CUDA graphs. |
| 790 | self.input_ids = self._make_buffer(self.max_num_tokens, dtype=torch.int32) |
| 791 | # [max_num_reqs, depth] prompt token ids preceding each request's chunk |
| 792 | # start, for models exposing `token_lookback_depth`. Sized after load. |
| 793 | self.lookback_token_ids: CpuGpuBuffer | None = None |
| 794 | self.positions = torch.zeros( |
| 795 | self.max_num_tokens, dtype=torch.int64, device=self.device |
| 796 | ) |
| 797 | self.is_padding = torch.zeros( |
| 798 | self.max_num_tokens, dtype=torch.bool, device=self.device |
| 799 | ) |
| 800 | self.query_start_loc = self._make_buffer( |
| 801 | self.max_num_reqs + 1, dtype=torch.int32 |
| 802 | ) |
| 803 | self.seq_lens = torch.zeros( |
| 804 | self.max_num_reqs, dtype=torch.int32, device=self.device |
| 805 | ) |
| 806 | self.optimistic_seq_lens_cpu = torch.zeros( |
| 807 | self.max_num_reqs, dtype=torch.int32, pin_memory=PIN_MEMORY |
| 808 | ) |
| 809 | self.num_computed_tokens = torch.zeros( |
| 810 | self.max_num_reqs, dtype=torch.int32, device=self.device |
| 811 | ) |
| 812 | self.prev_num_draft_tokens = self._make_buffer( |
| 813 | self.max_num_reqs, dtype=torch.int32 |
| 814 | ) |
| 815 | self.req_indices = self._make_buffer(self.max_num_tokens, dtype=torch.int64) |
| 816 | # Maps current batch position -> previous batch position (-1 for new reqs) |
| 817 | self.prev_positions = self._make_buffer(self.max_num_reqs, dtype=torch.int64) |
| 818 | self.num_scheduled_tokens = self._make_buffer( |
| 819 | self.max_num_reqs, dtype=torch.int32 |
| 820 | ) |
| 821 | |
| 822 | self.encoder_seq_lens = self._make_buffer(self.max_num_reqs, dtype=torch.int32) |
| 823 | if self.dcp_world_size > 1: |
| 824 | self.dcp_local_seq_lens = self._make_buffer( |
| 825 | self.max_num_reqs, dtype=torch.int32 |
| 826 | ) |
| 827 | # Because inputs_embeds may be bfloat16 and we don't need a numpy |
| 828 | # version of this tensor, avoid a RuntimeError by not creating a |
| 829 | # numpy buffer. |
| 830 | self.inputs_embeds = self._make_buffer( |
| 831 | self.max_num_tokens, self.inputs_embeds_size, dtype=self.dtype, numpy=False |
| 832 | ) |
| 833 | self.is_token_ids = self._make_buffer(self.max_num_tokens, dtype=torch.bool) |
| 834 | self.discard_request_mask = self._make_buffer( |
| 835 | self.max_num_reqs, dtype=torch.bool |
| 836 | ) |
| 837 | self.num_decode_draft_tokens = self._make_buffer( |
| 838 | self.max_num_reqs, dtype=torch.int32 |
| 839 | ) |
| 840 | self.num_accepted_tokens = self._make_buffer( |
| 841 | self.max_num_reqs, dtype=torch.int32 |
| 842 | ) |
| 843 | |
| 844 | # Only relevant for models using M-RoPE (e.g, Qwen2-VL) |
| 845 | if self.uses_mrope: |
| 846 | # NOTE: `mrope_positions` is implemented with one additional dummy |
| 847 | # position on purpose to make it non-contiguous so that it can work |
| 848 | # with torch compile. |
| 849 | # See detailed explanation in https://github.com/vllm-project/vllm/pull/12128#discussion_r1926431923 |
| 850 | |
| 851 | # NOTE: For text-only inputs, each dimension has identical position |
| 852 | # IDs, making M-RoPE functionally equivalent to 1D-RoPE. |
| 853 | # See page 5 of https://arxiv.org/abs/2409.12191 |
| 854 | self.mrope_positions = self._make_buffer( |
| 855 | (self.mrope_num_dims, self.max_num_tokens + 1), dtype=torch.int64 |
| 856 | ) |
| 857 | |
| 858 | # None in the first PP rank. The rest are set after load_model. |
| 859 | self.intermediate_tensors: IntermediateTensors | None = None |
| 860 | |
| 861 | # OPTIMIZATION: Cache the arange tensors rather than creating them |
| 862 | # every step. Keep in int64 to avoid overflow with long context. |
| 863 | # - arange_np: immutable [0, 1, 2, ...] used as source for batched computation |
| 864 | # - query_pos: CpuGpuBuffer for the computed batched arange result |
| 865 | arange_size = max(self.max_num_reqs + 1, self.max_num_tokens) |
| 866 | self.arange_np = np.arange(arange_size, dtype=np.int64) |
| 867 | self.query_pos = self._make_buffer(arange_size, dtype=torch.int64) |
| 868 | self._arange_scratch = np.empty(arange_size, dtype=np.int64) |
| 869 | |
| 870 | # Layer pairings for cross-layer KV sharing. |
| 871 | # If an Attention layer `layer_name` is in the keys of this dict, it |
| 872 | # means this layer will perform attention using the keys and values |
| 873 | # from the KV cache of `shared_kv_cache_layers[layer_name]`. |
| 874 | self.shared_kv_cache_layers: dict[str, str] = {} |
| 875 | self.kv_sharing_fast_prefill_eligible_layers: set[str] = set() |
| 876 | |
| 877 | self.kv_sharing_fast_prefill_logits_indices = None |
| 878 | if self.cache_config.kv_sharing_fast_prefill: |
| 879 | self.kv_sharing_fast_prefill_logits_indices = torch.zeros( |
| 880 | self.max_num_tokens, dtype=torch.int32, device=self.device |
| 881 | ) |
| 882 | |
| 883 | self.uniform_decode_query_len = 1 + self.num_spec_tokens |
| 884 | |
| 885 | # Cudagraph dispatcher for runtime cudagraph dispatching. |
| 886 | self.cudagraph_dispatcher = CudagraphDispatcher(self.vllm_config) |
| 887 | |
| 888 | self.mm_budget = ( |
| 889 | MultiModalBudget(self.vllm_config, self.mm_registry) |
| 890 | if self.supports_mm_inputs |
| 891 | else None |
| 892 | ) |
| 893 | |
| 894 | self.reorder_batch_threshold: int | None = None |
| 895 | |
| 896 | # Attention layers that are only in the KVCacheConfig of the runner |
| 897 | # (e.g., KV sharing, encoder-only attention), but not in the |
| 898 | # KVCacheConfig of the scheduler. |
| 899 | self.runner_only_attn_layers: set[str] = set() |
| 900 | |
| 901 | # Cached outputs. |
| 902 | self._draft_token_ids: list[list[int]] | torch.Tensor | None = None |
| 903 | self._draft_probs: torch.Tensor | None = None |
| 904 | self._draft_prob_req_ids: list[str] | None = None |
| 905 | # N-gram GPU path: async D2H buffer/event for per-request valid draft counts. |
| 906 | self._num_valid_draft_tokens: torch.Tensor | None = None |
| 907 | self._num_valid_draft_tokens_cpu: torch.Tensor | None = None |
| 908 | self._num_valid_draft_tokens_event: torch.cuda.Event | None = None |
| 909 | self._num_valid_draft_tokens_copy_stream: torch.cuda.Stream | None = None |
| 910 | if ( |
| 911 | self.speculative_config is not None |
| 912 | and self.speculative_config.use_ngram_gpu() |
| 913 | ): |
| 914 | self._num_valid_draft_tokens_cpu = torch.empty( |
| 915 | self.max_num_reqs, dtype=torch.int32, pin_memory=PIN_MEMORY |
| 916 | ) |
| 917 | self._num_valid_draft_tokens_event = torch.cuda.Event() |
| 918 | self._num_valid_draft_tokens_copy_stream = torch.cuda.Stream() |
| 919 | |
| 920 | self._draft_token_req_ids: list[str] | None = None |
| 921 | self.transfer_event = torch.Event() |
| 922 | self.sampled_token_ids_pinned_cpu = torch.empty( |
| 923 | (self.max_num_reqs, 1), |
| 924 | dtype=torch.int64, |
| 925 | device="cpu", |
| 926 | pin_memory=PIN_MEMORY, |
| 927 | ) |
| 928 | |
| 929 | # Pre-allocated tensor for copying valid sampled token counts to CPU, |
| 930 | # with dedicated stream for overlapping and event for coordination. |
| 931 | self.valid_sampled_token_count_event: torch.Event | None = None |
| 932 | self.valid_sampled_token_count_copy_stream: torch.cuda.Stream | None = None |
| 933 | # We also copy the drafted tokens to the CPU asynchronously, |
| 934 | # in case we need them for structured outputs. |
| 935 | self.draft_token_ids_event: torch.Event | None = None |
| 936 | self.draft_token_ids_copy_stream: torch.cuda.Stream | None = None |
| 937 | self.valid_sampled_token_count_cpu: torch.Tensor | None = None |
| 938 | self.draft_token_ids_cpu: torch.Tensor | None = None |
| 939 | self.num_accepted_tokens_event: torch.Event | None = None |
| 940 | if self.num_spec_tokens: |
| 941 | self.draft_token_ids_event = torch.Event() |
| 942 | self.num_accepted_tokens_event = torch.Event() |
| 943 | self.draft_token_ids_copy_stream = torch.cuda.Stream() |
| 944 | self.draft_token_ids_cpu = torch.empty( |
| 945 | (self.max_num_reqs, self.num_spec_tokens), |
| 946 | dtype=torch.int64, |
| 947 | device="cpu", |
| 948 | pin_memory=PIN_MEMORY, |
| 949 | ) |
| 950 | if self.use_async_scheduling: |
| 951 | self.valid_sampled_token_count_event = torch.Event() |
| 952 | self.valid_sampled_token_count_copy_stream = torch.cuda.Stream() |
| 953 | self.valid_sampled_token_count_cpu = torch.empty( |
| 954 | self.max_num_reqs, |
| 955 | dtype=torch.int32, |
| 956 | device="cpu", |
| 957 | pin_memory=PIN_MEMORY, |
| 958 | ) |
| 959 | |
| 960 | # Model weight offloader |
| 961 | # Make sure this is called before any get_offloader call |
| 962 | set_offloader(create_offloader(self.offload_config)) |
| 963 | |
| 964 | # Ephemeral state transferred between execute_model() and sample_tokens(). |
| 965 | self.execute_model_state: ExecuteModelState | None = None |
| 966 | self.kv_connector_output: KVConnectorOutput | None = None |
| 967 | self.mamba_state_idx: dict[str, int] = {} |
| 968 | self._mamba_bufs: mamba_utils.MambaBuffers | None = None |
| 969 | self._mamba_state_copy_funcs: MambaStateCopyFuncsByType | None = None |
| 970 | self.layerwise_nvtx_hooks_registered = False |
| 971 | |
| 972 | def update_max_model_len(self, max_model_len: int) -> None: |
| 973 | self.max_model_len = max_model_len |
| 974 | if self.speculative_config: |
| 975 | draft_config = self.speculative_config.draft_model_config |
| 976 | if draft_config is None or draft_config.max_model_len is None: |
| 977 | self.effective_drafter_max_model_len = self.max_model_len |
| 978 | |
| 979 | def reset_mm_cache(self) -> None: |
| 980 | """Clear the multi-modal cache that was used during profiling, |
| 981 | but no longer needed during inference. |
| 982 | """ |
| 983 | if self.mm_budget: |
| 984 | self.mm_budget.reset_cache() |
| 985 | self.late_interaction_runner.clear() |
| 986 | |
| 987 | def reset_encoder_cache(self) -> None: |
| 988 | """Clear the GPU-side encoder cache storing vision embeddings. |
| 989 | |
| 990 | This should be called when model weights are updated to ensure |
| 991 | stale embeddings computed with old weights are not reused. |
| 992 | """ |
| 993 | self.encoder_cache.clear() |
| 994 | self.late_interaction_runner.clear() |
| 995 | |
| 996 | def release_late_interaction_query_cache(self, query_keys: list[str]) -> None: |
| 997 | self.late_interaction_runner.release_queries(query_keys) |
| 998 | |
| 999 | def _get_positions(self, num_tokens: Any): |
| 1000 | if isinstance(num_tokens, int): |
| 1001 | if self.uses_mrope: |
| 1002 | return self.mrope_positions.gpu[:, :num_tokens] |
| 1003 | return self.positions[:num_tokens] |
| 1004 | else: |
| 1005 | if self.uses_mrope: |
| 1006 | return self.mrope_positions.gpu[:, num_tokens] |
| 1007 | return self.positions[num_tokens] |
| 1008 | |
| 1009 | def _make_buffer( |
| 1010 | self, *size: int | torch.SymInt, dtype: torch.dtype, numpy: bool = True |
| 1011 | ) -> CpuGpuBuffer: |
| 1012 | return CpuGpuBuffer( |
| 1013 | *size, |
| 1014 | dtype=dtype, |
| 1015 | device=self.device, |
| 1016 | with_numpy=numpy, |
| 1017 | ) |
| 1018 | |
| 1019 | def _get_mamba_state_copy_funcs(self) -> MambaStateCopyFuncsByType: |
| 1020 | if self._mamba_state_copy_funcs is None: |
| 1021 | mamba_groups = mamba_utils.get_mamba_groups(self.kv_cache_config) |
| 1022 | mamba_types = {spec.mamba_type for spec in mamba_groups} |
| 1023 | copy_funcs = self.model.get_mamba_state_copy_funcs(mamba_types) |
| 1024 | mamba_utils.validate_mamba_state_copy_funcs(mamba_groups, copy_funcs) |
| 1025 | self._mamba_state_copy_funcs = copy_funcs |
| 1026 | return self._mamba_state_copy_funcs |
| 1027 | |
| 1028 | def _get_mamba_bufs(self) -> mamba_utils.MambaBuffers: |
| 1029 | # Only reachable on the ``mamba_cache_mode == "align"`` path. |
| 1030 | # The postprocess sub-object is additionally gated on spec |
| 1031 | # decode + hybrid model. |
| 1032 | assert self.cache_config.mamba_cache_mode == "align" |
| 1033 | if self._mamba_bufs is None: |
| 1034 | self._mamba_bufs = mamba_utils.MambaBuffers.create( |
| 1035 | max_num_reqs=self.max_num_reqs, |
| 1036 | kv_cache_config=self.kv_cache_config, |
| 1037 | copy_funcs=self._get_mamba_state_copy_funcs(), |
| 1038 | make_buffer=self._make_buffer, |
| 1039 | device=self.device, |
| 1040 | with_postprocess_align=( |
| 1041 | self.speculative_config is not None and self.model_config.is_hybrid |
| 1042 | ), |
| 1043 | ) |
| 1044 | return self._mamba_bufs |
| 1045 | |
| 1046 | def _prepare_lookback_token_ids(self, num_reqs: int) -> torch.Tensor: |
| 1047 | """Gather, per request, the `depth` prompt token ids preceding its |
| 1048 | first scheduled token (column j is position start - 1 - j); -1 where |
| 1049 | the position is before the prompt or already past it. Generated |
| 1050 | positions are left to the model: under async scheduling the CPU token |
| 1051 | table holds placeholders for them.""" |
| 1052 | buf = self.lookback_token_ids |
| 1053 | assert buf is not None |
| 1054 | buf.np.fill(-1) |
| 1055 | if num_reqs > 0: |
| 1056 | depth = buf.np.shape[1] |
| 1057 | starts = self.input_batch.num_computed_tokens_cpu[:num_reqs, None] |
| 1058 | pos = starts - np.arange(1, depth + 1) |
| 1059 | num_prompt = self.input_batch.num_prompt_tokens[:num_reqs, None] |
| 1060 | valid = (pos >= 0) & (pos < num_prompt) |
| 1061 | rows = np.arange(num_reqs)[:, None] |
| 1062 | ids = self.input_batch.token_ids_cpu[rows, np.clip(pos, 0, None)] |
| 1063 | buf.np[:num_reqs] = np.where(valid, ids, -1) |
| 1064 | return buf.copy_to_gpu() |
| 1065 | |
| 1066 | def _init_model_kwargs(self, num_reqs: int | None = None): |
| 1067 | model_kwargs = dict[str, Any]() |
| 1068 | |
| 1069 | if self.lookback_token_ids is not None: |
| 1070 | if num_reqs is None: |
| 1071 | num_reqs = self.input_batch.num_reqs |
| 1072 | model_kwargs["lookback_token_ids"] = self._prepare_lookback_token_ids( |
| 1073 | num_reqs |
| 1074 | ) |
| 1075 | |
| 1076 | if not self.is_pooling_model: |
| 1077 | return model_kwargs |
| 1078 | |
| 1079 | num_reqs = self.input_batch.num_reqs |
| 1080 | pooling_params = self.input_batch.get_pooling_params() |
| 1081 | |
| 1082 | token_type_id_requests = dict[int, Any]() |
| 1083 | for i, param in enumerate(pooling_params): |
| 1084 | if ( |
| 1085 | param.extra_kwargs is not None |
| 1086 | and (token_types := param.extra_kwargs.get("compressed_token_type_ids")) |
| 1087 | is not None |
| 1088 | ): |
| 1089 | token_type_id_requests[i] = token_types |
| 1090 | |
| 1091 | if len(token_type_id_requests) == 0: |
| 1092 | return model_kwargs |
| 1093 | |
| 1094 | # Build ids on CPU using the CPU-resident upper bound for seq_lens; |
| 1095 | # `torch.arange(seq_lens[i])` with a GPU scalar would force a sync. |
| 1096 | seq_lens_cpu = self.optimistic_seq_lens_cpu[:num_reqs].tolist() |
| 1097 | token_type_ids = [] |
| 1098 | |
| 1099 | for i in range(num_reqs): |
| 1100 | seq_len_i = seq_lens_cpu[i] |
| 1101 | pos = token_type_id_requests.get(i, seq_len_i) |
| 1102 | ids = (torch.arange(seq_len_i) >= pos).int() |
| 1103 | token_type_ids.append(ids) |
| 1104 | |
| 1105 | token_type_ids_cpu = torch.empty( |
| 1106 | sum(seq_lens_cpu), dtype=torch.int32, pin_memory=PIN_MEMORY |
| 1107 | ) |
| 1108 | torch.cat(token_type_ids, out=token_type_ids_cpu) |
| 1109 | model_kwargs["token_type_ids"] = token_type_ids_cpu.to( |
| 1110 | device=self.device, non_blocking=True |
| 1111 | ) |
| 1112 | return model_kwargs |
| 1113 | |
| 1114 | def _may_reorder_batch(self, scheduler_output: "SchedulerOutput") -> None: |
| 1115 | """Update the order of requests in the batch based on the attention |
| 1116 | backend's needs. For example, some attention backends (namely MLA) may |
| 1117 | want to separate requests based on if the attention computation will be |
| 1118 | compute-bound or memory-bound. |
| 1119 | |
| 1120 | Args: |
| 1121 | scheduler_output: The scheduler output. |
| 1122 | |
| 1123 | """ |
| 1124 | # Attention free models have zero kv_cache_groups, however models |
| 1125 | # like Mamba are also attention free but use the kv_cache for |
| 1126 | # keeping its internal state. This is why we check the number |
| 1127 | # of kv_cache groups instead of solely checking |
| 1128 | # for self.model_config.is_attention_free. |
| 1129 | if len(self.kv_cache_config.kv_cache_groups) == 0: |
| 1130 | return |
| 1131 | |
| 1132 | if self.reorder_batch_threshold is not None: |
| 1133 | reorder_batch_to_split_decodes_and_prefills( |
| 1134 | self.input_batch, |
| 1135 | scheduler_output, |
| 1136 | decode_threshold=self.reorder_batch_threshold, |
| 1137 | ) |
| 1138 | |
| 1139 | def _init_kv_zero_meta(self) -> None: |
| 1140 | """One-time precomputation for _zero_block_ids. |
| 1141 | |
| 1142 | Called from gpu_worker.py outside the CuMem pool context. |
| 1143 | """ |
| 1144 | self._kv_block_zeroer = KVBlockZeroer( |
| 1145 | self.device, |
| 1146 | attn_groups_iter=self._kv_cache_spec_attn_group_iterator(), |
| 1147 | kernel_block_sizes=self._kernel_block_sizes, |
| 1148 | runner_only_attn_layers=self.runner_only_attn_layers, |
| 1149 | static_forward_context=self.compilation_config.static_forward_context, |
| 1150 | num_blocks=self.kv_cache_config.num_blocks, |
| 1151 | ) |
| 1152 | |
| 1153 | def _zero_block_ids(self, block_ids: list[int]) -> None: |
| 1154 | """Zero the KV cache memory for the given block IDs.""" |
| 1155 | if hasattr(self, "_kv_block_zeroer"): |
| 1156 | self._kv_block_zeroer.zero_block_ids(block_ids) |
| 1157 | |
| 1158 | # Note: used for model runner override. |
| 1159 | def _init_device_properties(self) -> None: |
| 1160 | """Initialize attributes from torch.cuda.get_device_properties""" |
| 1161 | self.num_sms = num_compute_units(self.device.index) |
| 1162 | |
| 1163 | # Note: used for model runner override. |
| 1164 | def _sync_device(self) -> None: |
| 1165 | torch.accelerator.synchronize() |
| 1166 | |
| 1167 | def _get_or_create_async_output_copy_stream(self) -> torch.cuda.Stream: |
| 1168 | stream = self.async_output_copy_stream |
| 1169 | if stream is None: |
| 1170 | stream = torch.cuda.Stream() |
| 1171 | self.async_output_copy_stream = stream |
| 1172 | return stream |
| 1173 | |
| 1174 | def _on_request_state_removed( |
| 1175 | self, |
| 1176 | req_id: str, |
| 1177 | req_state: CachedRequestState | None, |
| 1178 | ) -> None: |
| 1179 | """Hook for platform runners to clean request-scoped side caches.""" |
| 1180 | del req_id, req_state |
| 1181 | |
| 1182 | def _process_encoder_cache_scheduler_output( |
| 1183 | self, |
| 1184 | scheduler_output: "SchedulerOutput", |
| 1185 | ) -> None: |
| 1186 | """Apply scheduler-side encoder cache lifecycle updates.""" |
| 1187 | for mm_hash in scheduler_output.free_encoder_mm_hashes: |
| 1188 | self.encoder_cache.pop(mm_hash, None) |
| 1189 | |
| 1190 | def _update_states(self, scheduler_output: "SchedulerOutput") -> Callable | None: |
| 1191 | """Update the cached states and the persistent batch with the scheduler |
| 1192 | output. |
| 1193 | |
| 1194 | The updated states are used by the `_prepare_inputs` function to create |
| 1195 | the input GPU tensors for the model. |
| 1196 | |
| 1197 | The SamplingMetadata is updated and copied to the GPU if there is a |
| 1198 | new/resumed/paused/finished request in the batch. |
| 1199 | """ |
| 1200 | # Remove finished requests from the cached states. |
| 1201 | for req_id in scheduler_output.finished_req_ids: |
| 1202 | req_state = self.requests.pop(req_id, None) |
| 1203 | self._on_request_state_removed(req_id, req_state) |
| 1204 | self.num_prompt_logprobs.pop(req_id, None) |
| 1205 | self.late_interaction_runner.on_requests_finished( |
| 1206 | scheduler_output.finished_req_ids |
| 1207 | ) |
| 1208 | # Remove the finished requests from the persistent batch. |
| 1209 | # NOTE(woosuk): There could be an edge case where finished_req_ids and |
| 1210 | # scheduled_req_ids overlap. This happens when a request is aborted and |
| 1211 | # then resubmitted with the same ID. In this case, we treat them as two |
| 1212 | # distinct requests - clearing the cached states for the first request |
| 1213 | # and handling the second as a new request. |
| 1214 | for req_id in scheduler_output.finished_req_ids: |
| 1215 | self.input_batch.remove_request(req_id) |
| 1216 | |
| 1217 | # Zero GPU memory for freshly allocated cache blocks to prevent |
| 1218 | # stale NaN/data from corrupting attention or SSM computation. |
| 1219 | if scheduler_output.new_block_ids_to_zero: |
| 1220 | self._zero_block_ids(scheduler_output.new_block_ids_to_zero) |
| 1221 | if scheduler_output.kv_cache_block_copies: |
| 1222 | copy_kv_cache_blocks_inplace( |
| 1223 | self.kv_caches, |
| 1224 | self.kv_cache_config.num_blocks, |
| 1225 | scheduler_output.kv_cache_block_copies, |
| 1226 | ) |
| 1227 | |
| 1228 | # Free the cached encoder outputs. |
| 1229 | self._process_encoder_cache_scheduler_output(scheduler_output) |
| 1230 | |
| 1231 | # Remove the unscheduled requests from the persistent batch. |
| 1232 | # NOTE(woosuk): The unscheduled requests are either preempted requests |
| 1233 | # or running requests that are not scheduled in this step. We remove |
| 1234 | # them from the persistent batch but keep their cached states since |
| 1235 | # they will be scheduled again sometime in the future. |
| 1236 | scheduled_req_ids = scheduler_output.num_scheduled_tokens.keys() |
| 1237 | cached_req_ids = self.input_batch.req_id_to_index.keys() |
| 1238 | resumed_req_ids = scheduler_output.scheduled_cached_reqs.resumed_req_ids |
| 1239 | # NOTE(zhuohan): cached_req_ids and resumed_req_ids are usually disjoint, |
| 1240 | # so `(scheduled_req_ids - resumed_req_ids) == scheduled_req_ids` holds |
| 1241 | # apart from the forced-preemption case in reset_prefix_cache. And in |
| 1242 | # that case we include the resumed_req_ids in the unscheduled set so |
| 1243 | # that they get cleared from the persistent batch before being re-scheduled |
| 1244 | # in the normal resumed request path. |
| 1245 | unscheduled_req_ids = cached_req_ids - (scheduled_req_ids - resumed_req_ids) |
| 1246 | # NOTE(woosuk): The persistent batch optimization assumes that |
| 1247 | # consecutive batches contain mostly the same requests. If batches |
| 1248 | # have low request overlap (e.g., alternating between two distinct |
| 1249 | # sets of requests), this optimization becomes very inefficient. |
| 1250 | for req_id in unscheduled_req_ids: |
| 1251 | self.input_batch.remove_request(req_id) |
| 1252 | |
| 1253 | is_ngram_gpu = ( |
| 1254 | self.speculative_config is not None |
| 1255 | and self.speculative_config.use_ngram_gpu() |
| 1256 | ) |
| 1257 | if is_ngram_gpu: |
| 1258 | ngram_gpu_new_reqs: list[CachedRequestState] = [] |
| 1259 | |
| 1260 | reqs_to_add: list[CachedRequestState] = [] |
| 1261 | deferred_spec_decode_corrections = [] |
| 1262 | |
| 1263 | # Add new requests to the cached states. |
| 1264 | for new_req_data in scheduler_output.scheduled_new_reqs: |
| 1265 | req_id = new_req_data.req_id |
| 1266 | if req_id in self.requests: |
| 1267 | # For streaming case only. |
| 1268 | req_state = self._update_streaming_request(req_id, new_req_data) |
| 1269 | reqs_to_add.append(req_state) |
| 1270 | continue |
| 1271 | |
| 1272 | sampling_params = new_req_data.sampling_params |
| 1273 | pooling_params = new_req_data.pooling_params |
| 1274 | |
| 1275 | if ( |
| 1276 | sampling_params |
| 1277 | and sampling_params.sampling_type == SamplingType.RANDOM_SEED |
| 1278 | ): |
| 1279 | generator = torch.Generator(device=self.device) |
| 1280 | generator.manual_seed(sampling_params.seed) |
| 1281 | else: |
| 1282 | generator = None |
| 1283 | |
| 1284 | if self.is_pooling_model: |
| 1285 | assert pooling_params is not None |
| 1286 | task = pooling_params.task |
| 1287 | assert task is not None, "You did not set `task` in the API" |
| 1288 | |
| 1289 | model = cast(VllmModelForPooling, self.get_model()) |
| 1290 | to_update = model.pooler.get_pooling_updates(task) |
| 1291 | to_update.apply(pooling_params) |
| 1292 | |
| 1293 | req_state = CachedRequestState( |
| 1294 | req_id=req_id, |
| 1295 | prompt_token_ids=new_req_data.prompt_token_ids, |
| 1296 | prompt_embeds=new_req_data.prompt_embeds, |
| 1297 | prompt_is_token_ids=new_req_data.prompt_is_token_ids, |
| 1298 | mm_features=new_req_data.mm_features, |
| 1299 | sampling_params=sampling_params, |
| 1300 | pooling_params=pooling_params, |
| 1301 | generator=generator, |
| 1302 | block_ids=new_req_data.block_ids, |
| 1303 | num_computed_tokens=new_req_data.num_computed_tokens, |
| 1304 | output_token_ids=[], |
| 1305 | lora_request=new_req_data.lora_request, |
| 1306 | ) |
| 1307 | self.requests[req_id] = req_state |
| 1308 | self.late_interaction_runner.register_request(req_id, pooling_params) |
| 1309 | |
| 1310 | if sampling_params and sampling_params.prompt_logprobs is not None: |
| 1311 | self.num_prompt_logprobs[req_id] = ( |
| 1312 | self.input_batch.vocab_size |
| 1313 | if sampling_params.prompt_logprobs == -1 |
| 1314 | else sampling_params.prompt_logprobs |
| 1315 | ) |
| 1316 | |
| 1317 | # Only relevant for models using M-RoPE (e.g, Qwen2-VL) |
| 1318 | if self.uses_mrope: |
| 1319 | self._init_mrope_positions(req_state) |
| 1320 | |
| 1321 | reqs_to_add.append(req_state) |
| 1322 | # Track new requests for ngram_gpu full tensor copy |
| 1323 | if is_ngram_gpu: |
| 1324 | ngram_gpu_new_reqs.append(req_state) |
| 1325 | |
| 1326 | # Update the states of the running/resumed requests. |
| 1327 | is_last_rank = get_pp_group().is_last_rank |
| 1328 | req_data = scheduler_output.scheduled_cached_reqs |
| 1329 | scheduled_spec_tokens = scheduler_output.scheduled_spec_decode_tokens |
| 1330 | |
| 1331 | # Save scheduler-allocated spec lengths before trimming so |
| 1332 | # prev_num_draft_len keeps the optimistic count for rejection correction. |
| 1333 | original_num_spec_per_req: dict[str, int] = {} |
| 1334 | if ( |
| 1335 | self.speculative_config is not None |
| 1336 | and self.speculative_config.use_ngram_gpu() |
| 1337 | ): |
| 1338 | for req_id, toks in scheduled_spec_tokens.items(): |
| 1339 | original_num_spec_per_req[req_id] = len(toks) |
| 1340 | update_scheduler_for_invalid_drafts( |
| 1341 | self._num_valid_draft_tokens_event, |
| 1342 | self._num_valid_draft_tokens_cpu, |
| 1343 | scheduler_output, |
| 1344 | self.input_batch.req_id_to_index, |
| 1345 | ) |
| 1346 | if self.use_async_spec_decode: |
| 1347 | self.prev_num_draft_tokens.np.fill(0) |
| 1348 | |
| 1349 | for i, req_id in enumerate(req_data.req_ids): |
| 1350 | req_state = self.requests[req_id] |
| 1351 | num_computed_tokens = req_data.num_computed_tokens[i] |
| 1352 | new_block_ids = req_data.new_block_ids[i] |
| 1353 | resumed_from_preemption = req_id in req_data.resumed_req_ids |
| 1354 | num_output_tokens = req_data.num_output_tokens[i] |
| 1355 | req_index = self.input_batch.req_id_to_index.get(req_id) |
| 1356 | |
| 1357 | if req_state.prev_num_draft_len and self.use_async_scheduling: |
| 1358 | # prev_num_draft_len is used in async scheduling mode with |
| 1359 | # spec decode. it indicates if need to update num_computed_tokens |
| 1360 | # of the request. for example: |
| 1361 | # first step: num_computed_tokens = 0, spec_tokens = [], |
| 1362 | # prev_num_draft_len = 0. |
| 1363 | # second step: num_computed_tokens = 100(prompt length), |
| 1364 | # spec_tokens = [a,b], prev_num_draft_len = 0. |
| 1365 | # third step: num_computed_tokens = 100 + 2, spec_tokens = [c,d], |
| 1366 | # prev_num_draft_len = 2. |
| 1367 | # num_computed_tokens in first step and second step doesn't contain |
| 1368 | # the spec tokens length, but in third step it contains the |
| 1369 | # spec tokens length. we only need to update num_computed_tokens |
| 1370 | # when prev_num_draft_len > 0. |
| 1371 | if req_index is None: |
| 1372 | req_state.prev_num_draft_len = 0 |
| 1373 | else: |
| 1374 | # Optimistically assume all accepted; queue up a correction |
| 1375 | # to be called after the model forward to preserve async |
| 1376 | # scheduling. Corrected on GPU in _prepare_inputs. |
| 1377 | optimistic_num_accepted = req_state.prev_num_draft_len |
| 1378 | req_state.output_token_ids.extend([-1] * optimistic_num_accepted) |
| 1379 | |
| 1380 | deferred_spec_decode_corrections.append( |
| 1381 | (req_id, optimistic_num_accepted, req_state) |
| 1382 | ) |
| 1383 | |
| 1384 | prev_req_index = ( |
| 1385 | self.input_batch.prev_req_id_to_index.get(req_id) |
| 1386 | if self.input_batch.prev_req_id_to_index |
| 1387 | else None |
| 1388 | ) |
| 1389 | if prev_req_index is not None: |
| 1390 | self.prev_num_draft_tokens.np[prev_req_index] = ( |
| 1391 | optimistic_num_accepted |
| 1392 | ) |
| 1393 | |
| 1394 | if is_ngram_gpu and optimistic_num_accepted > 0: |
| 1395 | self.input_batch.num_tokens_no_spec[req_index] += ( |
| 1396 | optimistic_num_accepted |
| 1397 | ) |
| 1398 | |
| 1399 | # Update the cached states. |
| 1400 | req_state.num_computed_tokens = num_computed_tokens |
| 1401 | |
| 1402 | if not is_last_rank: |
| 1403 | if not req_data.new_token_ids: |
| 1404 | # Async scheduled PP: Sampled tokens propagated via GPU broadcast. |
| 1405 | new_token_ids: list[int] = [] |
| 1406 | else: |
| 1407 | # Non-async scheduling with PP: The scheduler sends |
| 1408 | # sampled token ids back because there's no direct communication |
| 1409 | # between the first-stage worker and the last-stage worker. |
| 1410 | new_token_ids = req_data.new_token_ids[i] |
| 1411 | # Add the sampled token(s) from the previous step (if any). |
| 1412 | # This doesn't include "unverified" tokens like spec tokens. |
| 1413 | num_new_tokens = ( |
| 1414 | num_computed_tokens + len(new_token_ids) - req_state.num_tokens |
| 1415 | ) |
| 1416 | if num_new_tokens == 1: |
| 1417 | # Avoid slicing list in most common case. |
| 1418 | req_state.output_token_ids.append(new_token_ids[-1]) |
| 1419 | elif num_new_tokens > 0: |
| 1420 | req_state.output_token_ids.extend( |
| 1421 | new_token_ids[-num_new_tokens:] |
| 1422 | ) |
| 1423 | elif num_output_tokens < len(req_state.output_token_ids): |
| 1424 | # Some output tokens were discarded due to a sync-KV-load |
| 1425 | # failure, or output_token_ids was inflated by the optimistic |
| 1426 | # extend above (async spec decode). Align the cached state. |
| 1427 | del req_state.output_token_ids[num_output_tokens:] |
| 1428 | if req_index is not None: |
| 1429 | end_idx = ( |
| 1430 | self.input_batch.num_prompt_tokens[req_index] |
| 1431 | + num_output_tokens |
| 1432 | ) |
| 1433 | self.input_batch.num_tokens_no_spec[req_index] = end_idx |
| 1434 | |
| 1435 | # Update the block IDs. |
| 1436 | if not resumed_from_preemption: |
| 1437 | if new_block_ids is not None: |
| 1438 | # Append the new blocks to the existing block IDs. |
| 1439 | for block_ids, new_ids in zip(req_state.block_ids, new_block_ids): |
| 1440 | block_ids.extend(new_ids) |
| 1441 | else: |
| 1442 | assert req_index is None |
| 1443 | assert new_block_ids is not None |
| 1444 | # The request is resumed from preemption. |
| 1445 | # Replace the existing block IDs with the new ones. |
| 1446 | req_state.block_ids = new_block_ids |
| 1447 | |
| 1448 | if req_index is None: |
| 1449 | # The request is not in the persistent batch. |
| 1450 | # The request was either preempted and resumed later, or was not |
| 1451 | # scheduled in the previous step and needs to be added again. |
| 1452 | |
| 1453 | if self.use_async_scheduling and num_output_tokens > 0: |
| 1454 | # We must recover the output token ids for resumed requests in the |
| 1455 | # async scheduling case, so that correct input_ids are obtained. |
| 1456 | resumed_token_ids = req_data.all_token_ids[req_id] |
| 1457 | req_state.output_token_ids = resumed_token_ids[-num_output_tokens:] |
| 1458 | |
| 1459 | reqs_to_add.append(req_state) |
| 1460 | # Track resumed requests for ngram_gpu full tensor copy |
| 1461 | if is_ngram_gpu: |
| 1462 | ngram_gpu_new_reqs.append(req_state) |
| 1463 | continue |
| 1464 | |
| 1465 | # Update the persistent batch. |
| 1466 | self.input_batch.num_computed_tokens_cpu[req_index] = num_computed_tokens |
| 1467 | if new_block_ids is not None: |
| 1468 | self.input_batch.block_table.append_row(new_block_ids, req_index) |
| 1469 | |
| 1470 | # For the last rank, we don't need to update the token_ids_cpu |
| 1471 | # because the sampled tokens are already cached. |
| 1472 | if not is_last_rank: |
| 1473 | start_token_index = self.input_batch.num_tokens_no_spec[req_index] |
| 1474 | # For chunked prefill, num_computed_tokens may less |
| 1475 | # than num_tokens_no_spec. |
| 1476 | # Async scheduled PP: no new_token_ids, advance num_tokens_no_spec |
| 1477 | # according to num_computed_tokens. |
| 1478 | end_token_index = max( |
| 1479 | start_token_index, |
| 1480 | num_computed_tokens + len(new_token_ids), |
| 1481 | ) |
| 1482 | if end_token_index > start_token_index: |
| 1483 | if new_token_ids: |
| 1484 | # Add new_token_ids to token_ids_cpu. |
| 1485 | num_new_tokens = end_token_index - start_token_index |
| 1486 | tokens_to_append = new_token_ids[-num_new_tokens:] |
| 1487 | self.input_batch.token_ids_cpu[ |
| 1488 | req_index, start_token_index:end_token_index |
| 1489 | ] = tokens_to_append |
| 1490 | self.input_batch.is_token_ids[ |
| 1491 | req_index, start_token_index:end_token_index |
| 1492 | ] = True |
| 1493 | self.input_batch.num_tokens_no_spec[req_index] = end_token_index |
| 1494 | |
| 1495 | # Add spec_token_ids to token_ids_cpu. |
| 1496 | self.input_batch.update_req_spec_token_ids(req_state, scheduled_spec_tokens) |
| 1497 | # Restore scheduler-side draft count after ngram trimming. |
| 1498 | if original_num_spec_per_req: |
| 1499 | orig = original_num_spec_per_req.get(req_id, 0) |
| 1500 | if orig != req_state.prev_num_draft_len: |
| 1501 | req_state.prev_num_draft_len = orig |
| 1502 | |
| 1503 | # Add the new or resumed requests to the persistent batch. |
| 1504 | # The smaller empty indices are filled first. |
| 1505 | for request in reqs_to_add: |
| 1506 | self.input_batch.add_request(request) |
| 1507 | self.input_batch.update_req_spec_token_ids(request, scheduled_spec_tokens) |
| 1508 | |
| 1509 | # Condense the batched states if there are gaps left by removed requests |
| 1510 | self.input_batch.condense() |
| 1511 | # Allow attention backend to reorder the batch, potentially |
| 1512 | self._may_reorder_batch(scheduler_output) |
| 1513 | # Refresh batch metadata with any pending updates. |
| 1514 | self.input_batch.refresh_metadata() |
| 1515 | |
| 1516 | # Incrementally update ngram_gpu tensors after batch is stable |
| 1517 | if is_ngram_gpu: |
| 1518 | update_ngram_gpu_tensors_incremental( |
| 1519 | self.input_batch, |
| 1520 | self.token_ids_gpu_tensor, |
| 1521 | self.num_tokens_no_spec_gpu, |
| 1522 | ngram_gpu_new_reqs, |
| 1523 | self.device, |
| 1524 | _pinned_idx_buf=self._ngram_pinned_idx_buf, |
| 1525 | _pinned_val_buf=self._ngram_pinned_val_buf, |
| 1526 | ) |
| 1527 | |
| 1528 | if deferred_spec_decode_corrections: |
| 1529 | |
| 1530 | def correct_spec_decode_token_counts(): |
| 1531 | valid_sampled_token_count = self._get_valid_sampled_token_count() |
| 1532 | if not valid_sampled_token_count: |
| 1533 | return |
| 1534 | prev_req_id_to_index = self.input_batch.prev_req_id_to_index |
| 1535 | if not prev_req_id_to_index: |
| 1536 | return |
| 1537 | for ( |
| 1538 | req_id, |
| 1539 | optimistic_num_accepted, |
| 1540 | req_state, |
| 1541 | ) in deferred_spec_decode_corrections: |
| 1542 | prev_req_index = prev_req_id_to_index.get(req_id) |
| 1543 | if prev_req_index is None: |
| 1544 | continue |
| 1545 | num_accepted = valid_sampled_token_count[prev_req_index] - 1 |
| 1546 | correction = optimistic_num_accepted - num_accepted |
| 1547 | req_state.num_computed_tokens -= correction |
| 1548 | cur_req_index = self.input_batch.req_id_to_index.get(req_id) |
| 1549 | if cur_req_index is None: |
| 1550 | continue |
| 1551 | self.input_batch.num_computed_tokens_cpu[cur_req_index] -= ( |
| 1552 | correction |
| 1553 | ) |
| 1554 | if is_ngram_gpu and correction > 0: |
| 1555 | self.input_batch.num_tokens_no_spec[cur_req_index] -= correction |
| 1556 | self.num_tokens_no_spec_gpu[cur_req_index] -= correction |
| 1557 | |
| 1558 | return correct_spec_decode_token_counts |
| 1559 | else: |
| 1560 | return None |
| 1561 | |
| 1562 | def _update_states_after_model_execute( |
| 1563 | self, output_token_ids: torch.Tensor, scheduler_output: "SchedulerOutput" |
| 1564 | ) -> None: |
| 1565 | """Update the cached states after model execution. |
| 1566 | |
| 1567 | This is used for MTP/EAGLE for hybrid models, as in linear attention, |
| 1568 | only the last token's state is kept. In MTP/EAGLE, for draft tokens |
| 1569 | the state are kept util we decide how many tokens are accepted for |
| 1570 | each sequence, and a shifting is done during the next iteration |
| 1571 | based on the number of accepted tokens. |
| 1572 | """ |
| 1573 | if not self.speculative_config or not self.model_config.is_hybrid: |
| 1574 | return |
| 1575 | |
| 1576 | # Count the number of accepted tokens for each sequence. |
| 1577 | # Valid tokens are contiguous from position 0, so counting non-(-1) |
| 1578 | # tokens gives us the first -1 position (i.e., number of accepted). |
| 1579 | num_reqs = output_token_ids.size(0) |
| 1580 | self.num_accepted_tokens.gpu[:num_reqs] = (output_token_ids != -1).sum(dim=1) |
| 1581 | |
| 1582 | if self.cache_config.mamba_cache_mode == "align": |
| 1583 | # Fused GPU postprocess: state copies + per-request accepted-token |
| 1584 | # update without CPU-GPU sync. The metadata |
| 1585 | # (num_scheduled_tokens, num_draft_tokens, num_computed_tokens) is |
| 1586 | # pre-staged to GPU buffers in _prepare_inputs. |
| 1587 | mamba_utils.postprocess_mamba_align_gpu( |
| 1588 | bufs=self._get_mamba_bufs(), |
| 1589 | num_reqs=num_reqs, |
| 1590 | num_accepted_tokens_gpu=self.num_accepted_tokens.gpu, |
| 1591 | num_accepted_tokens_cpu_tensor=( |
| 1592 | self.input_batch.num_accepted_tokens_cpu_tensor |
| 1593 | ), |
| 1594 | input_batch=self.input_batch, |
| 1595 | kv_cache_config=self.kv_cache_config, |
| 1596 | forward_context=self.compilation_config.static_forward_context, |
| 1597 | mamba_state_copy_funcs=self._get_mamba_state_copy_funcs(), |
| 1598 | ) |
| 1599 | |
| 1600 | assert self.num_accepted_tokens_event is not None |
| 1601 | self.num_accepted_tokens_event.record() |
| 1602 | else: |
| 1603 | self.input_batch.num_accepted_tokens_cpu_tensor[:num_reqs].copy_( |
| 1604 | self.num_accepted_tokens.gpu[:num_reqs], non_blocking=True |
| 1605 | ) |
| 1606 | assert self.num_accepted_tokens_event is not None |
| 1607 | self.num_accepted_tokens_event.record() |
| 1608 | |
| 1609 | def _update_streaming_request( |
| 1610 | self, req_id: str, new_req_data: NewRequestData |
| 1611 | ) -> CachedRequestState: |
| 1612 | """Updates streaming session request from `scheduled_new_reqs`. |
| 1613 | |
| 1614 | Removes the request from InputBatch (if present), updates the cached |
| 1615 | state, and prepares it for re-addition to the batch. |
| 1616 | |
| 1617 | NOTE: prompt_token_ids includes intermediate output tokens - tokens |
| 1618 | previously generated but now are input context (part of the prompt). |
| 1619 | """ |
| 1620 | self.input_batch.remove_request(req_id) |
| 1621 | req_state = self.requests[req_id] |
| 1622 | |
| 1623 | req_state.prompt_token_ids = new_req_data.prompt_token_ids |
| 1624 | req_state.mm_features = new_req_data.mm_features |
| 1625 | req_state.prompt_embeds = new_req_data.prompt_embeds |
| 1626 | req_state.sampling_params = new_req_data.sampling_params |
| 1627 | req_state.pooling_params = new_req_data.pooling_params |
| 1628 | self.late_interaction_runner.register_request(req_id, req_state.pooling_params) |
| 1629 | req_state.block_ids = new_req_data.block_ids |
| 1630 | req_state.num_computed_tokens = new_req_data.num_computed_tokens |
| 1631 | req_state.num_prompt_tokens = length_from_prompt_token_ids_or_embeds( |
| 1632 | req_state.prompt_token_ids, req_state.prompt_embeds |
| 1633 | ) |
| 1634 | |
| 1635 | # Clear `output_token_ids` as previous output tokens are now part of |
| 1636 | # `prompt_token_ids`. |
| 1637 | req_state.output_token_ids.clear() |
| 1638 | |
| 1639 | if self.uses_mrope: |
| 1640 | self._init_mrope_positions(req_state) |
| 1641 | |
| 1642 | return req_state |
| 1643 | |
| 1644 | def _init_mrope_positions(self, req_state: CachedRequestState): |
| 1645 | model = self.get_model() |
| 1646 | assert supports_mrope(model), "M-RoPE support is not implemented." |
| 1647 | mrope_model = cast(SupportsMRoPE, model) |
| 1648 | |
| 1649 | # `prompt_embeds` is a passthrough modality (no grid_thw), models' |
| 1650 | # M-RoPE code assumes per-feature grid info, so filter it out. The |
| 1651 | # prompt_embeds positions are treated as text positions for M-RoPE. |
| 1652 | mrope_features = [ |
| 1653 | f for f in req_state.mm_features if f.modality != "prompt_embeds" |
| 1654 | ] |
| 1655 | |
| 1656 | if req_state.prompt_token_ids is not None: |
| 1657 | input_tokens = req_state.prompt_token_ids |
| 1658 | elif req_state.prompt_embeds is not None: |
| 1659 | # For embeddings-only inputs, get_mrope_input_positions only |
| 1660 | # needs the sequence length when mm_features is empty (which is |
| 1661 | # the case here since prompt_embeds are filtered out above). |
| 1662 | seq_len = req_state.prompt_embeds.shape[0] |
| 1663 | input_tokens = list(range(seq_len)) |
| 1664 | else: |
| 1665 | raise ValueError( |
| 1666 | "M-RoPE requires either prompt_token_ids or prompt_embeds." |
| 1667 | ) |
| 1668 | |
| 1669 | req_state.mrope_positions, req_state.mrope_position_delta = ( |
| 1670 | mrope_model.get_mrope_input_positions( |
| 1671 | input_tokens, |
| 1672 | mrope_features, |
| 1673 | ) |
| 1674 | ) |
| 1675 | |
| 1676 | def _extract_mm_kwargs( |
| 1677 | self, |
| 1678 | scheduler_output: "SchedulerOutput", |
| 1679 | ) -> BatchedTensorInputs: |
| 1680 | if not scheduler_output or not self.is_multimodal_raw_input_only_model: |
| 1681 | return {} |
| 1682 | |
| 1683 | mm_kwargs = list[tuple[str, MultiModalKwargsItem]]() |
| 1684 | for req in scheduler_output.scheduled_new_reqs: |
| 1685 | for feature in req.mm_features: |
| 1686 | if feature.data is not None: |
| 1687 | mm_kwargs.append((feature.modality, feature.data)) |
| 1688 | |
| 1689 | # Input all modalities at once |
| 1690 | mm_kwargs_combined: BatchedTensorInputs = {} |
| 1691 | for _, _, mm_kwargs_batch in group_and_batch_mm_kwargs( |
| 1692 | mm_kwargs, |
| 1693 | device=self.device, |
| 1694 | pin_memory=PIN_MEMORY, |
| 1695 | ): |
| 1696 | mm_kwargs_combined.update(mm_kwargs_batch) |
| 1697 | |
| 1698 | return mm_kwargs_combined |
| 1699 | |
| 1700 | def _dummy_mm_kwargs(self, num_seqs: int) -> BatchedTensorInputs: |
| 1701 | if not self.is_multimodal_raw_input_only_model: |
| 1702 | return {} |
| 1703 | |
| 1704 | mm_budget = self.mm_budget |
| 1705 | assert mm_budget is not None |
| 1706 | |
| 1707 | if not mm_budget.mm_max_toks_per_item: |
| 1708 | return {} # No tower modalities (embed-only mode) |
| 1709 | |
| 1710 | dummy_modality = mm_budget.get_modality_with_max_tokens() |
| 1711 | return self._get_mm_dummy_batch(dummy_modality, num_seqs) |
| 1712 | |
| 1713 | def _get_cumsum_and_arange( |
| 1714 | self, |
| 1715 | num_tokens: np.ndarray, |
| 1716 | arange_out: np.ndarray, |
| 1717 | cumsum_dtype: np.dtype | None = None, |
| 1718 | ) -> np.ndarray: |
| 1719 | """Get the cumulative sum and batched arange of the given array. |
| 1720 | E.g., [2, 5, 3] -> [2, 7, 10], arange written to |
| 1721 | arange_out[:10] as [0, 1, 0, 1, 2, 3, 4, 0, 1, 2]. |
| 1722 | Equivalent to but faster than: |
| 1723 | np.concatenate([np.arange(n) for n in num_tokens]) |
| 1724 | """ |
| 1725 | # Step 1. [2, 5, 3] -> [2, 7, 10] |
| 1726 | cu_num_tokens = np.cumsum(num_tokens, dtype=cumsum_dtype) |
| 1727 | total_num_tokens = cu_num_tokens[-1] |
| 1728 | # Step 2. [2, 7, 10] -> [0, 0, 2, 2, 2, 2, 2, 7, 7, 7] |
| 1729 | cumsums_offsets = np.repeat(cu_num_tokens - num_tokens, num_tokens) |
| 1730 | # Step 3. [0, 1, 0, 1, 2, 3, 4, 0, 1, 2] |
| 1731 | np.subtract( |
| 1732 | self.arange_np[:total_num_tokens], |
| 1733 | cumsums_offsets, |
| 1734 | out=arange_out[:total_num_tokens], |
| 1735 | ) |
| 1736 | |
| 1737 | return cu_num_tokens |
| 1738 | |
| 1739 | def _compute_prev_positions(self, num_reqs: int) -> None: |
| 1740 | """Build prev_positions mapping: current pos -> previous pos (-1 if new). |
| 1741 | |
| 1742 | Populates self.prev_positions.np[:num_reqs] with the mapping. |
| 1743 | """ |
| 1744 | prev_req_id_to_index = self.input_batch.prev_req_id_to_index |
| 1745 | prev_positions = self.prev_positions.np[:num_reqs] |
| 1746 | |
| 1747 | if not prev_req_id_to_index: |
| 1748 | prev_positions.fill(-1) |
| 1749 | return |
| 1750 | |
| 1751 | for i, req_id in enumerate(self.input_batch.req_ids[:num_reqs]): |
| 1752 | prev_positions[i] = prev_req_id_to_index.get(req_id, -1) |
| 1753 | |
| 1754 | def _prepare_input_ids( |
| 1755 | self, |
| 1756 | scheduler_output: "SchedulerOutput", |
| 1757 | num_reqs: int, |
| 1758 | total_num_scheduled_tokens: int, |
| 1759 | cu_num_tokens: np.ndarray, |
| 1760 | ) -> None: |
| 1761 | """Prepare the input IDs for the current batch. |
| 1762 | |
| 1763 | Carefully handles the `prev_sampled_token_ids` which can be cached |
| 1764 | from the previous engine iteration, in which case those tokens on the |
| 1765 | GPU need to be copied into the corresponding slots into input_ids. |
| 1766 | |
| 1767 | Uses self.prev_positions[:num_reqs] which maps current pos -> prev pos |
| 1768 | (-1 for new requests). |
| 1769 | """ |
| 1770 | # Sync the async PP broadcast before reading sampled tokens. |
| 1771 | if self._pp_recv_work is not None: |
| 1772 | self._pp_recv_work.wait() |
| 1773 | self._pp_recv_work = None |
| 1774 | |
| 1775 | if self.input_batch.prev_sampled_token_ids is None: |
| 1776 | # Normal scheduling case |
| 1777 | self.input_ids.copy_to_gpu(total_num_scheduled_tokens) |
| 1778 | if self.enable_prompt_embeds: |
| 1779 | self.inputs_embeds.copy_to_gpu(total_num_scheduled_tokens) |
| 1780 | self.is_token_ids.copy_to_gpu(total_num_scheduled_tokens) |
| 1781 | return |
| 1782 | |
| 1783 | # Async scheduling case, where some decode requests from the previous |
| 1784 | # iteration won't have entries in input_ids_cpu and need to be copied |
| 1785 | # on the GPU from prev_sampled_token_ids. |
| 1786 | prev_positions = self.prev_positions.np[:num_reqs] |
| 1787 | scheduled_spec_tokens = scheduler_output.scheduled_spec_decode_tokens |
| 1788 | sample_flattened_indices: list[int] = [] |
| 1789 | spec_flattened_indices: list[int] = [] |
| 1790 | prev_draft_token_indices: list[int] = [] |
| 1791 | prev_indices: list[int] = [] |
| 1792 | common_indices_match = True |
| 1793 | max_flattened_index = -1 |
| 1794 | total_num_spec_tokens = 0 |
| 1795 | |
| 1796 | for cur_index in range(num_reqs): |
| 1797 | prev_index = prev_positions[cur_index] |
| 1798 | if prev_index < 0: |
| 1799 | continue |
| 1800 | prev_indices.append(prev_index) |
| 1801 | req_id = self.input_batch.req_ids[cur_index] |
| 1802 | # We need to compute the flattened input_ids index of the |
| 1803 | # last token in each common request. |
| 1804 | draft_len = len(scheduled_spec_tokens.get(req_id, ())) |
| 1805 | total_num_spec_tokens += draft_len |
| 1806 | flattened_index = cu_num_tokens[cur_index].item() - 1 |
| 1807 | # example: cu_num_tokens = [2, 5, 8], draft_tokens = [1, 2, 2] |
| 1808 | # sample_flattened_indices = [0, 2, 5] |
| 1809 | # spec_flattened_indices = [1, 3, 4, 6, 7] |
| 1810 | sample_flattened_indices.append(flattened_index - draft_len) |
| 1811 | spec_flattened_indices.extend( |
| 1812 | range(flattened_index - draft_len + 1, flattened_index + 1) |
| 1813 | ) |
| 1814 | start = prev_index * self.prev_num_spec_tokens |
| 1815 | # prev_draft_token_indices is used to find which draft_tokens_id |
| 1816 | # should be copied to input_ids |
| 1817 | # example: prev draft_tokens_id [[1,2], [3,4], [5, 6]] |
| 1818 | # flatten draft_tokens_id [1,2,3,4,5,6] |
| 1819 | # draft_len of each request [1, 2, 1] |
| 1820 | # then prev_draft_token_indices is [0, 2, 3, 4] |
| 1821 | prev_draft_token_indices.extend(range(start, start + draft_len)) |
| 1822 | common_indices_match &= prev_index == flattened_index |
| 1823 | max_flattened_index = max(max_flattened_index, flattened_index) |
| 1824 | |
| 1825 | num_common_tokens = len(sample_flattened_indices) |
| 1826 | total_without_spec = total_num_scheduled_tokens - total_num_spec_tokens |
| 1827 | if self.enable_prompt_embeds: |
| 1828 | # The multimodal embed path reads is_token_ids.gpu; its .cpu copy is |
| 1829 | # refreshed every step but the async fast path below replaces draft |
| 1830 | # tokens directly on the GPU. Mark those positions as token IDs before |
| 1831 | # refreshing the GPU mask so their embeddings are rebuilt. Sampled |
| 1832 | # tokens are already marked by bookkeeping. |
| 1833 | if spec_flattened_indices: |
| 1834 | self.is_token_ids.np[spec_flattened_indices] = True |
| 1835 | self.is_token_ids.copy_to_gpu(total_num_scheduled_tokens) |
| 1836 | if num_common_tokens < total_without_spec: |
| 1837 | # If not all requests are decodes from the last iteration, |
| 1838 | # we need to copy the input_ids_cpu to the GPU first. |
| 1839 | self.input_ids.copy_to_gpu(total_num_scheduled_tokens) |
| 1840 | if self.enable_prompt_embeds: |
| 1841 | self.inputs_embeds.copy_to_gpu(total_num_scheduled_tokens) |
| 1842 | if num_common_tokens == 0: |
| 1843 | # No requests in common with the previous iteration |
| 1844 | # So input_ids.cpu will have all the input ids. |
| 1845 | return |
| 1846 | if common_indices_match and max_flattened_index == (num_common_tokens - 1): |
| 1847 | # Common-case optimization: the batch is unchanged |
| 1848 | # and no reordering happened. |
| 1849 | # The indices are both the same permutation of 0..N-1 so |
| 1850 | # we can copy directly using a single slice. |
| 1851 | self.input_ids.gpu[:num_common_tokens].copy_( |
| 1852 | self.input_batch.prev_sampled_token_ids[:num_common_tokens, 0], |
| 1853 | non_blocking=True, |
| 1854 | ) |
| 1855 | return |
| 1856 | # Upload the index tensors asynchronously so the scatter can be non-blocking. |
| 1857 | sampled_tokens_index_tensor = torch.tensor( |
| 1858 | sample_flattened_indices, dtype=torch.int64, pin_memory=PIN_MEMORY |
| 1859 | ).to(self.device, non_blocking=True) |
| 1860 | prev_common_req_indices_tensor = torch.tensor( |
| 1861 | prev_indices, dtype=torch.int64, pin_memory=PIN_MEMORY |
| 1862 | ).to(self.device, non_blocking=True) |
| 1863 | self.input_ids.gpu.scatter_( |
| 1864 | dim=0, |
| 1865 | index=sampled_tokens_index_tensor, |
| 1866 | src=self.input_batch.prev_sampled_token_ids[ |
| 1867 | prev_common_req_indices_tensor, 0 |
| 1868 | ], |
| 1869 | ) |
| 1870 | |
| 1871 | # Scatter the draft tokens after the sampled tokens are scattered. |
| 1872 | if self._draft_token_ids is None or not spec_flattened_indices: |
| 1873 | return |
| 1874 | |
| 1875 | assert isinstance(self._draft_token_ids, torch.Tensor) |
| 1876 | draft_tokens_index_tensor = torch.tensor( |
| 1877 | spec_flattened_indices, dtype=torch.int64, pin_memory=PIN_MEMORY |
| 1878 | ).to(self.device, non_blocking=True) |
| 1879 | prev_draft_token_indices_tensor = torch.tensor( |
| 1880 | prev_draft_token_indices, dtype=torch.int64, pin_memory=PIN_MEMORY |
| 1881 | ).to(self.device, non_blocking=True) |
| 1882 | |
| 1883 | # because input_ids dtype is torch.int32, |
| 1884 | # so convert draft_token_ids to torch.int32 here. |
| 1885 | draft_token_ids = self._draft_token_ids.to(dtype=torch.int32) |
| 1886 | |
| 1887 | self.input_ids.gpu.scatter_( |
| 1888 | dim=0, |
| 1889 | index=draft_tokens_index_tensor, |
| 1890 | src=draft_token_ids.flatten()[prev_draft_token_indices_tensor], |
| 1891 | ) |
| 1892 | |
| 1893 | def _get_encoder_seq_lens( |
| 1894 | self, |
| 1895 | num_scheduled_tokens: dict[str, int], |
| 1896 | kv_cache_spec: KVCacheSpec, |
| 1897 | num_reqs: int, |
| 1898 | for_cudagraph_capture: bool = False, |
| 1899 | ) -> tuple[torch.Tensor | None, np.ndarray | None]: |
| 1900 | if not isinstance(kv_cache_spec, CrossAttentionSpec): |
| 1901 | return None, None |
| 1902 | |
| 1903 | # Zero out buffer for padding requests that are not actually scheduled (CGs) |
| 1904 | self.encoder_seq_lens.np[:num_reqs] = 0 |
| 1905 | |
| 1906 | # Build encoder_seq_lens array mapping request indices to |
| 1907 | # encoder lengths for inputs scheduled in this batch |
| 1908 | for req_id in num_scheduled_tokens: |
| 1909 | req_index = self.input_batch.req_id_to_index[req_id] |
| 1910 | req_state = self.requests[req_id] |
| 1911 | if req_state.mm_features is None: |
| 1912 | self.encoder_seq_lens.np[req_index] = 0 |
| 1913 | continue |
| 1914 | |
| 1915 | # Get the total number of encoder input tokens for running encoder requests |
| 1916 | # whether encoding is finished or not so that cross-attention knows how |
| 1917 | # many encoder tokens to attend to. |
| 1918 | encoder_input_tokens = sum( |
| 1919 | feature.mm_position.length for feature in req_state.mm_features |
| 1920 | ) |
| 1921 | self.encoder_seq_lens.np[req_index] = encoder_input_tokens |
| 1922 | if for_cudagraph_capture: |
| 1923 | # During CUDA graph capture, we need to use realistic encoder lengths |
| 1924 | # so that max_seqlen_k is captured with the correct value. |
| 1925 | max_encoder_len = getattr( |
| 1926 | self.model_config.hf_config, |
| 1927 | "max_source_positions", |
| 1928 | self.max_encoder_len, |
| 1929 | ) |
| 1930 | self.encoder_seq_lens.np[:num_reqs] = max_encoder_len |
| 1931 | |
| 1932 | self.encoder_seq_lens.copy_to_gpu(num_reqs) |
| 1933 | encoder_seq_lens = self.encoder_seq_lens.gpu[:num_reqs] |
| 1934 | encoder_seq_lens_cpu = self.encoder_seq_lens.np[:num_reqs] |
| 1935 | |
| 1936 | return encoder_seq_lens, encoder_seq_lens_cpu |
| 1937 | |
| 1938 | def _prepare_inputs( |
| 1939 | self, |
| 1940 | scheduler_output: "SchedulerOutput", |
| 1941 | num_scheduled_tokens: np.ndarray, |
| 1942 | ) -> tuple[ |
| 1943 | torch.Tensor, |
| 1944 | SpecDecodeMetadata | None, |
| 1945 | int, |
| 1946 | ]: |
| 1947 | """Returns: |
| 1948 | tuple[logits_indices, spec_decode_metadata, max_num_sampled_tokens] |
| 1949 | |
| 1950 | """ |
| 1951 | total_num_scheduled_tokens = scheduler_output.total_num_scheduled_tokens |
| 1952 | assert total_num_scheduled_tokens > 0 |
| 1953 | num_reqs = self.input_batch.num_reqs |
| 1954 | assert num_reqs > 0 |
| 1955 | |
| 1956 | # OPTIMIZATION: Start copying the block table first. |
| 1957 | # This way, we can overlap the copy with the following CPU operations. |
| 1958 | self.input_batch.block_table.commit_block_table(num_reqs) |
| 1959 | |
| 1960 | # Get request indices. |
| 1961 | # E.g., [2, 5, 3] -> [0, 0, 1, 1, 1, 1, 1, 2, 2, 2] |
| 1962 | req_indices = np.repeat(self.arange_np[:num_reqs], num_scheduled_tokens) |
| 1963 | |
| 1964 | # cu_num_tokens: [2, 5, 3] -> [2, 7, 10] |
| 1965 | # self.query_pos.np[:10]: [0, 1, 0, 1, 2, 3, 4, 0, 1, 2] |
| 1966 | cu_num_tokens = self._get_cumsum_and_arange( |
| 1967 | num_scheduled_tokens, self.query_pos.np |
| 1968 | ) |
| 1969 | |
| 1970 | # Get positions. |
| 1971 | positions_np = ( |
| 1972 | self.input_batch.num_computed_tokens_cpu[req_indices] |
| 1973 | + self.query_pos.np[: cu_num_tokens[-1]] |
| 1974 | ) |
| 1975 | |
| 1976 | # Calculate M-RoPE positions. |
| 1977 | # Only relevant for models using M-RoPE (e.g, Qwen2-VL) |
| 1978 | if self.uses_mrope: |
| 1979 | self._calc_mrope_positions(scheduler_output) |
| 1980 | |
| 1981 | # Get token indices. |
| 1982 | # E.g., [0, 1, 0, 1, 2, 3, 4, 0, 1, 2] |
| 1983 | # -> [0, 1, M, M + 1, M + 2, M + 3, M + 4, 2 * M, 2 * M + 1, 2 * M + 2] |
| 1984 | # where M is the max_model_len. |
| 1985 | token_indices = ( |
| 1986 | positions_np + req_indices * self.input_batch.token_ids_cpu.shape[1] |
| 1987 | ) |
| 1988 | token_indices_tensor = torch.from_numpy(token_indices) |
| 1989 | |
| 1990 | # NOTE(woosuk): We use torch.index_select instead of np.take here |
| 1991 | # because torch.index_select is much faster than np.take for large |
| 1992 | # tensors. |
| 1993 | torch.index_select( |
| 1994 | self.input_batch.token_ids_cpu_tensor.flatten(), |
| 1995 | 0, |
| 1996 | token_indices_tensor, |
| 1997 | out=self.input_ids.cpu[:total_num_scheduled_tokens], |
| 1998 | ) |
| 1999 | if self.enable_prompt_embeds: |
| 2000 | is_token_ids = self.input_batch.is_token_ids_tensor.flatten() |
| 2001 | torch.index_select( |
| 2002 | is_token_ids, |
| 2003 | 0, |
| 2004 | token_indices_tensor, |
| 2005 | out=self.is_token_ids.cpu[:total_num_scheduled_tokens], |
| 2006 | ) |
| 2007 | |
| 2008 | # Because we did not pre-allocate a massive prompt_embeds CPU tensor on |
| 2009 | # the InputBatch, we need to fill in the prompt embeds into the expected |
| 2010 | # spots in the GpuModelRunner's pre-allocated prompt_embeds tensor. |
| 2011 | if self.input_batch.req_prompt_embeds: |
| 2012 | output_idx = 0 |
| 2013 | for req_idx in range(num_reqs): |
| 2014 | num_sched = num_scheduled_tokens[req_idx] |
| 2015 | |
| 2016 | # Skip if this request doesn't have embeddings |
| 2017 | if req_idx not in self.input_batch.req_prompt_embeds: |
| 2018 | output_idx += num_sched |
| 2019 | continue |
| 2020 | |
| 2021 | # Skip if no tokens scheduled |
| 2022 | if num_sched <= 0: |
| 2023 | output_idx += num_sched |
| 2024 | continue |
| 2025 | |
| 2026 | req_embeds = self.input_batch.req_prompt_embeds[req_idx] |
| 2027 | start_pos = self.input_batch.num_computed_tokens_cpu[req_idx] |
| 2028 | |
| 2029 | # Skip if trying to read beyond available embeddings |
| 2030 | if start_pos >= req_embeds.shape[0]: |
| 2031 | output_idx += num_sched |
| 2032 | continue |
| 2033 | |
| 2034 | # Copy available embeddings |
| 2035 | end_pos = start_pos + num_sched |
| 2036 | actual_end = min(end_pos, req_embeds.shape[0]) |
| 2037 | actual_num_sched = actual_end - start_pos |
| 2038 | |
| 2039 | if actual_num_sched > 0: |
| 2040 | self.inputs_embeds.cpu[ |
| 2041 | output_idx : output_idx + actual_num_sched |
| 2042 | ].copy_(req_embeds[start_pos:actual_end]) |
| 2043 | |
| 2044 | output_idx += num_sched |
| 2045 | |
| 2046 | # Prepare the attention metadata. |
| 2047 | self.query_start_loc.np[0] = 0 |
| 2048 | self.query_start_loc.np[1 : num_reqs + 1] = cu_num_tokens |
| 2049 | # Note: pad query_start_loc to be non-decreasing, as kernels |
| 2050 | # like FlashAttention requires that |
| 2051 | self.query_start_loc.np[num_reqs + 1 :].fill(cu_num_tokens[-1]) |
| 2052 | self.query_start_loc.copy_to_gpu() |
| 2053 | query_start_loc = self.query_start_loc.gpu[: num_reqs + 1] |
| 2054 | |
| 2055 | # Compute optimistic seq_lens (assumes all draft tokens from previous |
| 2056 | # iteration accepted). Store in optimistic_seq_lens_cpu for use by |
| 2057 | # _build_attention_metadata (max_seq_len) and discard_request_mask. |
| 2058 | # seq_lens (GPU) will be computed later using the same optimistic values. |
| 2059 | torch.add( |
| 2060 | self.input_batch.num_computed_tokens_cpu_tensor[:num_reqs], |
| 2061 | torch.from_numpy(num_scheduled_tokens), |
| 2062 | out=self.optimistic_seq_lens_cpu[:num_reqs], |
| 2063 | ) |
| 2064 | self.optimistic_seq_lens_cpu[num_reqs:].fill_(0) |
| 2065 | |
| 2066 | # Build prev_positions mapping: current pos -> prev pos (-1 if new). |
| 2067 | # Used for gathering from previous iteration's GPU tensors. |
| 2068 | prev_req_id_to_index = self.input_batch.prev_req_id_to_index |
| 2069 | self._compute_prev_positions(num_reqs) |
| 2070 | |
| 2071 | num_tokens = [self.requests[r].num_tokens for r in self.input_batch.req_ids] |
| 2072 | num_tokens_np = np.array(num_tokens, dtype=np.int32) |
| 2073 | |
| 2074 | # Record which requests should not be sampled, |
| 2075 | # so that we could clear the sampled tokens before returning |
| 2076 | self.discard_request_mask.np[:num_reqs] = ( |
| 2077 | self.optimistic_seq_lens_cpu[:num_reqs].numpy() < num_tokens_np |
| 2078 | ) |
| 2079 | self.discard_request_mask.copy_to_gpu(num_reqs) |
| 2080 | |
| 2081 | # Sync num_accepted_tokens from CPU (set by |
| 2082 | # _update_states_after_model_execute for hybrid models). |
| 2083 | # Skipped under async scheduling (non-align): the CPU copy races with |
| 2084 | # the in-flight D2H copy and with input-batch row moves. |
| 2085 | needs_cpu_accepted_counts = self.num_accepted_tokens_event is not None and not ( |
| 2086 | self.use_async_scheduling and self.cache_config.mamba_cache_mode != "align" |
| 2087 | ) |
| 2088 | if needs_cpu_accepted_counts: |
| 2089 | assert self.num_accepted_tokens_event is not None |
| 2090 | self.num_accepted_tokens_event.synchronize() |
| 2091 | # Async mode: condense() reordered indices, use prev_positions mapping |
| 2092 | if self.use_async_scheduling and prev_req_id_to_index: |
| 2093 | prev_idx = self.prev_positions.np[:num_reqs] |
| 2094 | new_mask = prev_idx < 0 |
| 2095 | self.num_accepted_tokens.np[:num_reqs] = ( |
| 2096 | self.input_batch.num_accepted_tokens_cpu[ |
| 2097 | np.where(new_mask, 0, prev_idx) |
| 2098 | ] |
| 2099 | ) |
| 2100 | self.num_accepted_tokens.np[:num_reqs][new_mask] = 1 |
| 2101 | self.input_batch.num_accepted_tokens_cpu[:num_reqs] = ( |
| 2102 | self.num_accepted_tokens.np[:num_reqs] |
| 2103 | ) |
| 2104 | else: |
| 2105 | # Non-async mode: use values directly |
| 2106 | self.num_accepted_tokens.np[:num_reqs] = ( |
| 2107 | self.input_batch.num_accepted_tokens_cpu[:num_reqs] |
| 2108 | ) |
| 2109 | self.num_accepted_tokens.np[num_reqs:].fill(1) |
| 2110 | self.num_accepted_tokens.copy_to_gpu() |
| 2111 | else: |
| 2112 | # Default to 1; update_num_computed_tokens_for_batch_change below |
| 2113 | # corrects rows that had drafts from valid_sampled_token_count. |
| 2114 | self.num_accepted_tokens.np.fill(1) |
| 2115 | self.num_accepted_tokens.gpu.fill_(1) |
| 2116 | |
| 2117 | # Update num_computed_tokens on GPU. In async spec decode, |
| 2118 | # CPU values are optimistic (all drafts accepted). The kernel |
| 2119 | # corrects on GPU using the previous step's |
| 2120 | # valid_sampled_token_count_gpu. Otherwise, just copy from CPU. |
| 2121 | if ( |
| 2122 | self.use_async_spec_decode |
| 2123 | and self.valid_sampled_token_count_gpu is not None |
| 2124 | and prev_req_id_to_index |
| 2125 | ): |
| 2126 | self.prev_positions.copy_to_gpu(num_reqs) |
| 2127 | self.prev_num_draft_tokens.copy_to_gpu() |
| 2128 | cpu_values = self.input_batch.num_computed_tokens_cpu_tensor[:num_reqs].to( |
| 2129 | device=self.device, non_blocking=True |
| 2130 | ) |
| 2131 | update_num_computed_tokens_for_batch_change( |
| 2132 | self.num_computed_tokens, |
| 2133 | self.num_accepted_tokens.gpu[:num_reqs], |
| 2134 | self.prev_positions.gpu[:num_reqs], |
| 2135 | self.valid_sampled_token_count_gpu, |
| 2136 | self.prev_num_draft_tokens.gpu, |
| 2137 | cpu_values, |
| 2138 | ) |
| 2139 | else: |
| 2140 | self.num_computed_tokens[:num_reqs].copy_( |
| 2141 | self.input_batch.num_computed_tokens_cpu_tensor[:num_reqs], |
| 2142 | non_blocking=True, |
| 2143 | ) |
| 2144 | |
| 2145 | self.req_indices.np[:total_num_scheduled_tokens] = req_indices |
| 2146 | self.req_indices.copy_to_gpu(total_num_scheduled_tokens) |
| 2147 | req_indices_gpu = self.req_indices.gpu[:total_num_scheduled_tokens] |
| 2148 | |
| 2149 | self.query_pos.copy_to_gpu(total_num_scheduled_tokens) |
| 2150 | self.num_scheduled_tokens.np[:num_reqs] = num_scheduled_tokens |
| 2151 | self.num_scheduled_tokens.copy_to_gpu(num_reqs) |
| 2152 | num_scheduled_tokens_gpu = self.num_scheduled_tokens.gpu[:num_reqs] |
| 2153 | self.positions[:total_num_scheduled_tokens] = ( |
| 2154 | self.num_computed_tokens[req_indices_gpu].to(torch.int64) |
| 2155 | + self.query_pos.gpu[:total_num_scheduled_tokens] |
| 2156 | ) |
| 2157 | self.seq_lens[:num_reqs] = ( |
| 2158 | self.num_computed_tokens[:num_reqs] + num_scheduled_tokens_gpu |
| 2159 | ) |
| 2160 | self.seq_lens[num_reqs:].fill_(0) |
| 2161 | |
| 2162 | self.input_batch.block_table.compute_slot_mapping( |
| 2163 | num_reqs, |
| 2164 | self.query_start_loc.gpu[: num_reqs + 1], |
| 2165 | self.positions[:total_num_scheduled_tokens], |
| 2166 | ) |
| 2167 | |
| 2168 | # Copy the tensors to the GPU. |
| 2169 | self._prepare_input_ids( |
| 2170 | scheduler_output, |
| 2171 | num_reqs, |
| 2172 | total_num_scheduled_tokens, |
| 2173 | cu_num_tokens, |
| 2174 | ) |
| 2175 | |
| 2176 | if self.uses_mrope: |
| 2177 | # Only relevant for models using M-RoPE (e.g, Qwen2-VL) |
| 2178 | # Copy one row at a time. mrope_positions is allocated as |
| 2179 | # [num_dims, max_num_tokens + 1] with a dummy trailing column to keep it |
| 2180 | # non-contiguous for torch.compile, so cpu[:, :N] is a strided view. |
| 2181 | # copy_() cannot express a strided source as a single |
| 2182 | # cudaMemcpyAsync, so it first gathers into a contiguous *pageable* |
| 2183 | # temporary, and a pageable H2D ignores non_blocking=True and |
| 2184 | # synchronizes the stream before the transfer starts. Each row is |
| 2185 | # contiguous within the pinned allocation, so per-row copies stay on |
| 2186 | # the pinned path and are genuinely asynchronous. |
| 2187 | for row in range(self.mrope_positions.gpu.shape[0]): |
| 2188 | self.mrope_positions.gpu[row, :total_num_scheduled_tokens].copy_( |
| 2189 | self.mrope_positions.cpu[row, :total_num_scheduled_tokens], |
| 2190 | non_blocking=True, |
| 2191 | ) |
| 2192 | if self.use_async_spec_decode and self.uses_mrope: |
| 2193 | drift = self.num_computed_tokens[req_indices_gpu].to( |
| 2194 | torch.int64 |
| 2195 | ) - async_tensor_h2d( |
| 2196 | self.input_batch.num_computed_tokens_cpu_tensor[req_indices], |
| 2197 | device=self.device, |
| 2198 | dtype=torch.int64, |
| 2199 | ) |
| 2200 | self.mrope_positions.gpu[:, :total_num_scheduled_tokens] += drift |
| 2201 | |
| 2202 | use_spec_decode = len(scheduler_output.scheduled_spec_decode_tokens) > 0 |
| 2203 | if not use_spec_decode: |
| 2204 | # NOTE(woosuk): Due to chunked prefills, the batch may contain |
| 2205 | # partial requests. While we should not sample any token |
| 2206 | # from these partial requests, we do so for simplicity. |
| 2207 | # We will ignore the sampled tokens from the partial requests. |
| 2208 | # TODO: Support prompt logprobs. |
| 2209 | logits_indices = query_start_loc[1:] - 1 |
| 2210 | spec_decode_metadata = None |
| 2211 | num_sampled_tokens = np.ones(num_reqs, dtype=np.int32) |
| 2212 | else: |
| 2213 | # Get the number of draft tokens for each request. |
| 2214 | # Iterate over the dictionary rather than all requests since not all |
| 2215 | # requests have draft tokens. |
| 2216 | num_draft_tokens = np.zeros(num_reqs, dtype=np.int32) |
| 2217 | # For chunked prefills, use -1 as mask rather than 0, as guided |
| 2218 | # decoding may rollback speculative tokens. |
| 2219 | num_decode_draft_tokens = np.full(num_reqs, -1, dtype=np.int32) |
| 2220 | for ( |
| 2221 | req_id, |
| 2222 | draft_token_ids, |
| 2223 | ) in scheduler_output.scheduled_spec_decode_tokens.items(): |
| 2224 | req_idx = self.input_batch.req_id_to_index[req_id] |
| 2225 | draft_len = len(draft_token_ids) |
| 2226 | num_draft_tokens[req_idx] = draft_len |
| 2227 | if num_scheduled_tokens[req_idx] == draft_len + 1: |
| 2228 | num_decode_draft_tokens[req_idx] = draft_len |
| 2229 | spec_decode_metadata = self._calc_spec_decode_metadata( |
| 2230 | num_draft_tokens, cu_num_tokens |
| 2231 | ) |
| 2232 | logits_indices = spec_decode_metadata.logits_indices |
| 2233 | num_sampled_tokens = num_draft_tokens + 1 |
| 2234 | # For DECODE only cuda graph of some attention backends (e.g., GDN). |
| 2235 | self.num_decode_draft_tokens.np[:num_reqs] = num_decode_draft_tokens |
| 2236 | self.num_decode_draft_tokens.np[num_reqs:].fill(-1) |
| 2237 | self.num_decode_draft_tokens.copy_to_gpu() |
| 2238 | |
| 2239 | # Hot-Swap lora model |
| 2240 | if self.lora_config: |
| 2241 | assert ( |
| 2242 | np.sum(num_sampled_tokens) |
| 2243 | <= self.vllm_config.scheduler_config.max_num_batched_tokens |
| 2244 | ) |
| 2245 | self.set_active_loras( |
| 2246 | self.input_batch, num_scheduled_tokens, num_sampled_tokens |
| 2247 | ) |
| 2248 | |
| 2249 | return ( |
| 2250 | logits_indices, |
| 2251 | spec_decode_metadata, |
| 2252 | int(num_sampled_tokens.max()), |
| 2253 | ) |
| 2254 | |
| 2255 | def _build_attention_metadata( |
| 2256 | self, |
| 2257 | num_tokens: int, |
| 2258 | num_reqs: int, |
| 2259 | max_query_len: int, |
| 2260 | num_tokens_padded: int | None = None, |
| 2261 | num_reqs_padded: int | None = None, |
| 2262 | ubatch_slices: UBatchSlices | None = None, |
| 2263 | logits_indices: torch.Tensor | None = None, |
| 2264 | max_num_sampled_tokens: int | None = None, |
| 2265 | use_spec_decode: bool = False, |
| 2266 | for_cudagraph_capture: bool = False, |
| 2267 | num_scheduled_tokens: dict[str, int] | None = None, |
| 2268 | cascade_attn_prefix_lens: list[list[int]] | None = None, |
| 2269 | slot_mappings: dict[int, torch.Tensor] | None = None, |
| 2270 | ) -> tuple[PerLayerAttnMetadata, CommonAttentionMetadata | None]: |
| 2271 | """Returns: |
| 2272 | tuple[attn_metadata, spec_decode_common_attn_metadata] |
| 2273 | |
| 2274 | """ |
| 2275 | # Attention metadata is not needed for attention free models |
| 2276 | if len(self.kv_cache_config.kv_cache_groups) == 0: |
| 2277 | return {}, None |
| 2278 | |
| 2279 | num_tokens_padded = num_tokens_padded or num_tokens |
| 2280 | num_reqs_padded = num_reqs_padded or num_reqs |
| 2281 | assert num_reqs_padded is not None and num_tokens_padded is not None |
| 2282 | |
| 2283 | attn_metadata: PerLayerAttnMetadata = {} |
| 2284 | if ubatch_slices is not None: |
| 2285 | attn_metadata = [dict() for _ in range(len(ubatch_slices))] |
| 2286 | |
| 2287 | if for_cudagraph_capture: |
| 2288 | # For some attention backends (e.g. FA) with sliding window models we need |
| 2289 | # to make sure the backend see a max_seq_len that is larger to the sliding |
| 2290 | # window size when capturing to make sure the correct kernel is selected. |
| 2291 | max_seq_len = self.max_model_len |
| 2292 | else: |
| 2293 | max_seq_len = self.optimistic_seq_lens_cpu.numpy()[:num_reqs].max().item() |
| 2294 | |
| 2295 | kv_cache_groups = self.kv_cache_config.kv_cache_groups |
| 2296 | |
| 2297 | def _get_block_table(kv_cache_gid: int): |
| 2298 | assert num_reqs_padded is not None and num_tokens_padded is not None |
| 2299 | kv_cache_spec = kv_cache_groups[kv_cache_gid].kv_cache_spec |
| 2300 | if isinstance(kv_cache_spec, EncoderOnlyAttentionSpec): |
| 2301 | blk_table_tensor = torch.zeros( |
| 2302 | (num_reqs_padded, 1), |
| 2303 | dtype=torch.int32, |
| 2304 | device=self.device, |
| 2305 | ) |
| 2306 | else: |
| 2307 | blk_table = self.input_batch.block_table[kv_cache_gid] |
| 2308 | blk_table_tensor = blk_table.get_device_tensor(num_reqs_padded) |
| 2309 | |
| 2310 | # Fill unused block table entries with NULL_BLOCK_ID (null block) |
| 2311 | # for CUDAGraph padding. Block 0 is reserved for padding. |
| 2312 | blk_table_tensor[num_reqs:num_reqs_padded].fill_(NULL_BLOCK_ID) |
| 2313 | return blk_table_tensor |
| 2314 | |
| 2315 | assert slot_mappings is not None |
| 2316 | block_table_gid_0 = _get_block_table(0) |
| 2317 | slot_mapping_gid_0 = slot_mappings[0] |
| 2318 | |
| 2319 | num_computed_tokens_cpu = self.input_batch.num_computed_tokens_cpu_tensor[ |
| 2320 | :num_reqs_padded |
| 2321 | ] |
| 2322 | num_prompt_tokens_cpu = self.input_batch.num_prompt_tokens_cpu_tensor[ |
| 2323 | :num_reqs_padded |
| 2324 | ] |
| 2325 | seq_lens_cpu_upper_bound = self.optimistic_seq_lens_cpu[:num_reqs_padded] |
| 2326 | |
| 2327 | # is_prefilling: True if request is still in prefill phase. |
| 2328 | # Used by mamba backends to distinguish actual decodes from |
| 2329 | # short extends. |
| 2330 | is_prefilling = num_computed_tokens_cpu < num_prompt_tokens_cpu |
| 2331 | # Zero out padded rows so stale data from condense() doesn't |
| 2332 | # misclassify padding as prefill in CUDA graph mode. |
| 2333 | is_prefilling[num_reqs:] = False |
| 2334 | |
| 2335 | # Compute mm_prefix bidirectional ranges before building |
| 2336 | # attention metadata so builders handle them during build(). |
| 2337 | # By default, ranges exceeding sliding_window are skipped to prevent |
| 2338 | # early tokens from attending across the entire image span. Models that |
| 2339 | # clamp mm_prefix to the sliding window *in-kernel* (e.g. Gemma4, which |
| 2340 | # needs HF's (causal OR blockwise) AND sliding_window on sliding layers) |
| 2341 | # opt out of the skip so the bidirectional range survives for images |
| 2342 | # larger than the window; the kernel then bounds it per-query. |
| 2343 | req_doc_ranges: dict[int, list[tuple[int, int]]] | None = None |
| 2344 | if self.is_mm_prefix_lm: |
| 2345 | req_doc_ranges = {} |
| 2346 | hf_text_config = self.model_config.hf_text_config |
| 2347 | _bidi_sw = getattr(hf_text_config, "sliding_window", None) |
| 2348 | _clamps_in_kernel = getattr( |
| 2349 | self.model, "mm_prefix_clamp_sliding_window", False |
| 2350 | ) or getattr(hf_text_config, "mm_prefix_clamp_sliding_window", False) |
| 2351 | # Some models (DeepSeek-V4 vision) define the bidirectional span |
| 2352 | # over the whole sentinel block ([IMAGE_START, IMAGE_END]) rather |
| 2353 | # than the embed tokens, and prepend a position-dependent |
| 2354 | # alignment pad before the first sentinel. For those, derive the |
| 2355 | # span from the full placeholder range and strip the pad. |
| 2356 | # TODO(Isotr0py): Refactor mm_prefix_lm implementation |
| 2357 | # for better readability and maintainability. |
| 2358 | _span_pad_modulus = getattr( |
| 2359 | hf_text_config, "mm_prefix_span_leading_pad_modulus", 0 |
| 2360 | ) |
| 2361 | for req_id in self.input_batch.req_ids: |
| 2362 | image_doc_ranges = [] |
| 2363 | req_state = self.requests[req_id] |
| 2364 | for mm_feature in req_state.mm_features: |
| 2365 | if mm_feature.modality == "audio": |
| 2366 | continue |
| 2367 | pos_info = mm_feature.mm_position |
| 2368 | if _span_pad_modulus: |
| 2369 | pad = ( |
| 2370 | _span_pad_modulus - 1 - pos_info.offset % _span_pad_modulus |
| 2371 | ) |
| 2372 | img_doc_range = [ |
| 2373 | ( |
| 2374 | pos_info.offset + pad, |
| 2375 | pos_info.offset + pos_info.length - 1, |
| 2376 | ) |
| 2377 | ] |
| 2378 | else: |
| 2379 | img_doc_range = pos_info.extract_embeds_range() |
| 2380 | for r in img_doc_range: |
| 2381 | if ( |
| 2382 | not _clamps_in_kernel |
| 2383 | and _bidi_sw is not None |
| 2384 | and (r[1] - r[0] + 1) > _bidi_sw |
| 2385 | ): |
| 2386 | continue |
| 2387 | image_doc_ranges.append(r) |
| 2388 | req_idx = self.input_batch.req_id_to_index[req_id] |
| 2389 | req_doc_ranges[req_idx] = image_doc_ranges |
| 2390 | |
| 2391 | # Reference Sliding Window Attention (R-SWA): pass per-request prompt |
| 2392 | # lengths so the attention backend can keep the prefix globally visible. |
| 2393 | # The backend owns the persistent CUDA-graph-safe GPU buffer. |
| 2394 | rswa_prefix_lens = None |
| 2395 | if self.model_config.rswa_window is not None: |
| 2396 | rswa_prefix_lens = num_prompt_tokens_cpu |
| 2397 | |
| 2398 | replayssm_decode_base_cpu = None |
| 2399 | if self.cache_config.use_replayssm: |
| 2400 | replayssm_decode_base_cpu = ( |
| 2401 | self.input_batch.replayssm_decode_base_cpu_tensor[:num_reqs_padded] |
| 2402 | ) |
| 2403 | |
| 2404 | cm_base = CommonAttentionMetadata( |
| 2405 | query_start_loc=self.query_start_loc.gpu[: num_reqs_padded + 1], |
| 2406 | query_start_loc_cpu=self.query_start_loc.cpu[: num_reqs_padded + 1], |
| 2407 | seq_lens=self.seq_lens[:num_reqs_padded], |
| 2408 | seq_lens_cpu_upper_bound=seq_lens_cpu_upper_bound, |
| 2409 | replayssm_decode_base_cpu=replayssm_decode_base_cpu, |
| 2410 | num_reqs=num_reqs_padded, |
| 2411 | num_actual_tokens=num_tokens_padded, |
| 2412 | max_query_len=max_query_len, |
| 2413 | max_seq_len=max_seq_len, |
| 2414 | block_table_tensor=block_table_gid_0, |
| 2415 | slot_mapping=slot_mapping_gid_0, |
| 2416 | causal=True, |
| 2417 | is_prefilling=is_prefilling, |
| 2418 | positions=self.positions[:num_tokens_padded], |
| 2419 | mm_req_doc_ranges=req_doc_ranges, |
| 2420 | rswa_prefix_lens=rswa_prefix_lens, |
| 2421 | ) |
| 2422 | |
| 2423 | if self.dcp_world_size > 1: |
| 2424 | self.dcp_local_seq_lens.cpu[:num_reqs] = get_dcp_local_seq_lens( |
| 2425 | self.optimistic_seq_lens_cpu[:num_reqs], |
| 2426 | self.dcp_world_size, |
| 2427 | self.dcp_rank, |
| 2428 | self.parallel_config.cp_kv_cache_interleave_size, |
| 2429 | ) |
| 2430 | self.dcp_local_seq_lens.cpu[num_reqs:].fill_(0) |
| 2431 | self.dcp_local_seq_lens.copy_to_gpu(num_reqs_padded) |
| 2432 | |
| 2433 | cm_base.dcp_local_seq_lens = self.dcp_local_seq_lens.gpu[:num_reqs_padded] |
| 2434 | cm_base.dcp_local_seq_lens_cpu_upper_bound = self.dcp_local_seq_lens.cpu[ |
| 2435 | :num_reqs_padded |
| 2436 | ] |
| 2437 | |
| 2438 | if logits_indices is not None and self.cache_config.kv_sharing_fast_prefill: |
| 2439 | cm_base.num_logits_indices = logits_indices.size(0) |
| 2440 | cm_base.max_logits_per_req = max_num_sampled_tokens |
| 2441 | cm_base.logits_indices_padded = self._prepare_kv_sharing_fast_prefill( |
| 2442 | logits_indices |
| 2443 | ) |
| 2444 | |
| 2445 | # Cache attention metadata builds across hybrid KV-cache groups |
| 2446 | # The only thing that changes between different hybrid KV-cache groups when the |
| 2447 | # same metadata builder and KVCacheSpec is the same is the block table, so we |
| 2448 | # can cache the attention metadata builds and just update the block table using |
| 2449 | # `builder.update_block_table` if the builder supports it. |
| 2450 | cached_attn_metadata: dict[ |
| 2451 | tuple[KVCacheSpec, type[AttentionMetadataBuilder]], AttentionMetadata |
| 2452 | ] = {} |
| 2453 | |
| 2454 | def _build_attn_group_metadata( |
| 2455 | kv_cache_gid: int, |
| 2456 | attn_gid: int, |
| 2457 | common_attn_metadata: CommonAttentionMetadata, |
| 2458 | ubid: int | None = None, |
| 2459 | ) -> None: |
| 2460 | attn_group = self.attn_groups[kv_cache_gid][attn_gid] |
| 2461 | builder = attn_group.get_metadata_builder(ubid or 0) |
| 2462 | kv_cache_spec = kv_cache_groups[kv_cache_gid].kv_cache_spec |
| 2463 | if isinstance(kv_cache_spec, UniformTypeKVCacheSpecs): |
| 2464 | kv_cache_spec = kv_cache_spec.kv_cache_specs[attn_group.layer_names[0]] |
| 2465 | cache_key = (kv_cache_spec, type(builder)) |
| 2466 | |
| 2467 | cascade_attn_prefix_len = ( |
| 2468 | cascade_attn_prefix_lens[kv_cache_gid][attn_gid] |
| 2469 | if cascade_attn_prefix_lens |
| 2470 | else 0 |
| 2471 | ) |
| 2472 | |
| 2473 | extra_attn_metadata_args = {} |
| 2474 | if use_spec_decode and isinstance( |
| 2475 | builder, |
| 2476 | ( |
| 2477 | Mamba2AttentionMetadataBuilder, |
| 2478 | GDNAttentionMetadataBuilder, |
| 2479 | BailingLinearAttentionMetadataBuilder, |
| 2480 | ShortConvAttentionMetadataBuilder, |
| 2481 | ), |
| 2482 | ): |
| 2483 | assert ubid is None, ( |
| 2484 | "UBatching not supported with GDN or linear attn yet" |
| 2485 | ) |
| 2486 | extra_attn_metadata_args = dict( |
| 2487 | num_accepted_tokens=self.num_accepted_tokens.gpu[:num_reqs_padded], |
| 2488 | num_decode_draft_tokens_cpu=self.num_decode_draft_tokens.cpu[ |
| 2489 | :num_reqs_padded |
| 2490 | ], |
| 2491 | ) |
| 2492 | |
| 2493 | if for_cudagraph_capture: |
| 2494 | attn_metadata_i = builder.build_for_cudagraph_capture( |
| 2495 | common_attn_metadata |
| 2496 | ) |
| 2497 | elif ( |
| 2498 | cache_key in cached_attn_metadata |
| 2499 | and builder.supports_update_block_table |
| 2500 | ): |
| 2501 | attn_metadata_i = builder.update_block_table( |
| 2502 | cached_attn_metadata[cache_key], |
| 2503 | common_attn_metadata.block_table_tensor, |
| 2504 | common_attn_metadata.slot_mapping, |
| 2505 | ) |
| 2506 | else: |
| 2507 | attn_metadata_i = builder.build( |
| 2508 | common_prefix_len=cascade_attn_prefix_len, |
| 2509 | common_attn_metadata=common_attn_metadata, |
| 2510 | **extra_attn_metadata_args, |
| 2511 | ) |
| 2512 | if builder.supports_update_block_table: |
| 2513 | cached_attn_metadata[cache_key] = attn_metadata_i |
| 2514 | |
| 2515 | if ubid is None: |
| 2516 | assert isinstance(attn_metadata, dict) |
| 2517 | attn_metadata_dict = attn_metadata |
| 2518 | else: |
| 2519 | assert isinstance(attn_metadata, list) |
| 2520 | attn_metadata_dict = attn_metadata[ubid] |
| 2521 | |
| 2522 | for layer_name in attn_group.layer_names: |
| 2523 | attn_metadata_dict[layer_name] = attn_metadata_i |
| 2524 | |
| 2525 | # Prepare the attention metadata for each KV cache group and make layers |
| 2526 | # in the same group share the same metadata. |
| 2527 | spec_decode_common_attn_metadata = None |
| 2528 | for kv_cache_gid, kv_cache_group in enumerate(kv_cache_groups): |
| 2529 | cm = copy(cm_base) # shallow copy |
| 2530 | |
| 2531 | # Basically only the encoder seq_lens, block_table and slot_mapping change |
| 2532 | # for each kv_cache_group. |
| 2533 | cm.encoder_seq_lens, cm.encoder_seq_lens_cpu = self._get_encoder_seq_lens( |
| 2534 | num_scheduled_tokens or {}, |
| 2535 | kv_cache_group.kv_cache_spec, |
| 2536 | num_reqs_padded, |
| 2537 | for_cudagraph_capture=for_cudagraph_capture, |
| 2538 | ) |
| 2539 | if kv_cache_gid > 0: |
| 2540 | cm.block_table_tensor = _get_block_table(kv_cache_gid) |
| 2541 | cm.slot_mapping = slot_mappings[kv_cache_gid] |
| 2542 | |
| 2543 | if self.speculative_config and spec_decode_common_attn_metadata is None: |
| 2544 | if isinstance( |
| 2545 | self.drafter, |
| 2546 | ( |
| 2547 | EagleProposer, |
| 2548 | DFlashProposer, |
| 2549 | Gemma4Proposer, |
| 2550 | ExtractHiddenStatesProposer, |
| 2551 | ), |
| 2552 | ): |
| 2553 | if self.drafter.kv_cache_gid == kv_cache_gid: |
| 2554 | spec_decode_common_attn_metadata = cm |
| 2555 | else: |
| 2556 | spec_decode_common_attn_metadata = cm |
| 2557 | # Capture per-group block tables for multi-group proposers. |
| 2558 | if self.speculative_config and isinstance(self.drafter, Step3p5MTPProposer): |
| 2559 | self.drafter.set_per_group_attn_metadata( |
| 2560 | kv_cache_gid, cm.block_table_tensor, cm.slot_mapping |
| 2561 | ) |
| 2562 | elif self.speculative_config and isinstance(self.drafter, Gemma4Proposer): |
| 2563 | self.drafter.set_per_group_block_table( |
| 2564 | kv_cache_gid, cm.block_table_tensor |
| 2565 | ) |
| 2566 | |
| 2567 | for attn_gid in range(len(self.attn_groups[kv_cache_gid])): |
| 2568 | if ubatch_slices is not None: |
| 2569 | for ubid, _cm in enumerate(split_attn_metadata(ubatch_slices, cm)): |
| 2570 | _build_attn_group_metadata(kv_cache_gid, attn_gid, _cm, ubid) |
| 2571 | |
| 2572 | else: |
| 2573 | _build_attn_group_metadata(kv_cache_gid, attn_gid, cm) |
| 2574 | |
| 2575 | if spec_decode_common_attn_metadata is not None and ( |
| 2576 | num_reqs != num_reqs_padded or num_tokens != num_tokens_padded |
| 2577 | ): |
| 2578 | # Currently the drafter still only uses piecewise cudagraphs (and modifies |
| 2579 | # the attention metadata in directly), and therefore does not want to use |
| 2580 | # padded attention metadata. |
| 2581 | spec_decode_common_attn_metadata = ( |
| 2582 | spec_decode_common_attn_metadata.unpadded(num_tokens, num_reqs) |
| 2583 | ) |
| 2584 | |
| 2585 | return attn_metadata, spec_decode_common_attn_metadata |
| 2586 | |
| 2587 | def _compute_cascade_attn_prefix_lens( |
| 2588 | self, |
| 2589 | num_scheduled_tokens: np.ndarray, |
| 2590 | num_computed_tokens: np.ndarray, |
| 2591 | num_common_prefix_blocks: list[int], |
| 2592 | ) -> list[list[int]] | None: |
| 2593 | """Returns: |
| 2594 | Optional[cascade_attn_prefix_lens] |
| 2595 | cascade_attn_prefix_lens is 2D: |
| 2596 | ``[kv_cache_group_id][attn_group_idx]``, |
| 2597 | None if we should not use cascade attention |
| 2598 | |
| 2599 | """ |
| 2600 | use_cascade_attn = False |
| 2601 | num_kv_cache_groups = len(self.kv_cache_config.kv_cache_groups) |
| 2602 | cascade_attn_prefix_lens: list[list[int]] = [ |
| 2603 | [] for _ in range(num_kv_cache_groups) |
| 2604 | ] |
| 2605 | |
| 2606 | for kv_cache_gid in range(num_kv_cache_groups): |
| 2607 | for attn_group in self.attn_groups[kv_cache_gid]: |
| 2608 | if isinstance(attn_group.kv_cache_spec, EncoderOnlyAttentionSpec): |
| 2609 | cascade_attn_prefix_len = 0 |
| 2610 | else: |
| 2611 | # 0 if cascade attention should not be used |
| 2612 | cascade_attn_prefix_len = self._compute_cascade_attn_prefix_len( |
| 2613 | num_scheduled_tokens, |
| 2614 | num_computed_tokens, |
| 2615 | num_common_prefix_blocks[kv_cache_gid], |
| 2616 | attn_group.kv_cache_spec, |
| 2617 | attn_group.get_metadata_builder(), |
| 2618 | ) |
| 2619 | cascade_attn_prefix_lens[kv_cache_gid].append(cascade_attn_prefix_len) |
| 2620 | use_cascade_attn |= cascade_attn_prefix_len > 0 |
| 2621 | |
| 2622 | return cascade_attn_prefix_lens if use_cascade_attn else None |
| 2623 | |
| 2624 | def _compute_cascade_attn_prefix_len( |
| 2625 | self, |
| 2626 | num_scheduled_tokens: np.ndarray, |
| 2627 | num_computed_tokens: np.ndarray, |
| 2628 | num_common_prefix_blocks: int, |
| 2629 | kv_cache_spec: KVCacheSpec, |
| 2630 | attn_metadata_builder: AttentionMetadataBuilder, |
| 2631 | ) -> int: |
| 2632 | """Compute the length of the common prefix for cascade attention. |
| 2633 | |
| 2634 | NOTE(woosuk): The common prefix length returned by this function |
| 2635 | represents the length used specifically for cascade attention, not the |
| 2636 | actual number of tokens shared between requests. When cascade attention |
| 2637 | is disabled (use_cascade=False), this function returns 0 even if |
| 2638 | requests share common tokens. Additionally, the common prefix length is |
| 2639 | truncated to a multiple of the block size and may be further truncated |
| 2640 | due to implementation details explained below. |
| 2641 | |
| 2642 | Args: |
| 2643 | num_scheduled_tokens: Number of tokens scheduled per request. |
| 2644 | num_computed_tokens: Number of tokens already computed per request. |
| 2645 | num_common_prefix_blocks: Number of shared KV cache blocks. |
| 2646 | kv_cache_spec: The KV cache spec of the group being considered. |
| 2647 | attn_metadata_builder: Builder used to decide whether cascade |
| 2648 | attention can be used. |
| 2649 | |
| 2650 | Returns: |
| 2651 | int: Length of common prefix in tokens. |
| 2652 | |
| 2653 | """ |
| 2654 | common_prefix_len = num_common_prefix_blocks * kv_cache_spec.block_size |
| 2655 | if common_prefix_len == 0: |
| 2656 | # Common case. |
| 2657 | return 0 |
| 2658 | |
| 2659 | # NOTE(woosuk): Cascade attention uses two attention kernels: one |
| 2660 | # for the common prefix and the other for the rest. For the first |
| 2661 | # kernel, we concatenate all the query tokens (possibly from |
| 2662 | # different requests) and treat them as if they are from the same |
| 2663 | # request. Then, we use bi-directional attention to process the |
| 2664 | # common prefix in the KV cache. Importantly, this means that the |
| 2665 | # first kernel does not do any masking. |
| 2666 | |
| 2667 | # Consider the following example: |
| 2668 | # Request 1's input query: [D, E, X] |
| 2669 | # Request 1's kv cache: [A, B, C, D, E, X] |
| 2670 | # Request 1's num_computed_tokens: 3 (i.e., [A, B, C]) |
| 2671 | # Request 2's input query: [E, Y] |
| 2672 | # Request 2's kv cache: [A, B, C, D, E, Y] |
| 2673 | # Request 2's num_computed_tokens: 4 (i.e., [A, B, C, D]) |
| 2674 | |
| 2675 | # If we use [A, B, C, D, E] as the common prefix, then the |
| 2676 | # first kernel will compute the bi-directional attention between |
| 2677 | # input query [D, E, X, E, Y] and common prefix [A, B, C, D, E]. |
| 2678 | # However, this is wrong because D in Request 1 should not attend to |
| 2679 | # E in the common prefix (i.e., we need masking). |
| 2680 | # To avoid this, [A, B, C, D] should be the common prefix. |
| 2681 | # That is, the common prefix should be capped by the minimum |
| 2682 | # num_computed_tokens among the requests, and plus one to include |
| 2683 | # the first token of the query. |
| 2684 | |
| 2685 | # In practice, we use [A, B, C] as the common prefix, instead of |
| 2686 | # [A, B, C, D] (i.e., the common prefix is capped by the minimum |
| 2687 | # num_computed_tokens, without plus one). |
| 2688 | # This is because of an implementation detail: We want to always |
| 2689 | # use two kernels for cascade attention. Let's imagine: |
| 2690 | # Request 3's input query: [D] |
| 2691 | # Request 3's kv cache: [A, B, C, D] |
| 2692 | # Request 3's num_computed_tokens: 3 (i.e., [A, B, C]) |
| 2693 | # If we use [A, B, C, D] as the common prefix for Request 1-3, |
| 2694 | # then Request 3 will be processed only by the first kernel, |
| 2695 | # and the second kernel will get an empty input. While this is not |
| 2696 | # a fundamental problem, our current implementation does not support |
| 2697 | # this case. |
| 2698 | common_prefix_len = min(common_prefix_len, num_computed_tokens.min()) |
| 2699 | # common_prefix_len should be a multiple of the block size. |
| 2700 | common_prefix_len = ( |
| 2701 | common_prefix_len // kv_cache_spec.block_size * kv_cache_spec.block_size |
| 2702 | ) |
| 2703 | use_sliding_window = isinstance(kv_cache_spec, SlidingWindowSpec) or ( |
| 2704 | isinstance(kv_cache_spec, FullAttentionSpec) |
| 2705 | and kv_cache_spec.sliding_window is not None |
| 2706 | ) |
| 2707 | use_local_attention = isinstance(kv_cache_spec, ChunkedLocalAttentionSpec) or ( |
| 2708 | isinstance(kv_cache_spec, FullAttentionSpec) |
| 2709 | and kv_cache_spec.attention_chunk_size is not None |
| 2710 | ) |
| 2711 | assert isinstance(kv_cache_spec, AttentionSpec) |
| 2712 | use_cascade = attn_metadata_builder.use_cascade_attention( |
| 2713 | common_prefix_len=common_prefix_len, |
| 2714 | query_lens=num_scheduled_tokens, |
| 2715 | num_query_heads=self.num_query_heads, |
| 2716 | num_kv_heads=kv_cache_spec.num_kv_heads, |
| 2717 | use_alibi=self.use_alibi, |
| 2718 | use_sliding_window=use_sliding_window, |
| 2719 | use_local_attention=use_local_attention, |
| 2720 | num_sms=self.num_sms, |
| 2721 | dcp_world_size=self.dcp_world_size, |
| 2722 | ) |
| 2723 | return common_prefix_len if use_cascade else 0 |
| 2724 | |
| 2725 | def _calc_mrope_positions(self, scheduler_output: "SchedulerOutput"): |
| 2726 | mrope_pos_ptr = 0 |
| 2727 | for index, req_id in enumerate(self.input_batch.req_ids): |
| 2728 | req = self.requests[req_id] |
| 2729 | assert req.mrope_positions is not None |
| 2730 | |
| 2731 | num_computed_tokens = self.input_batch.num_computed_tokens_cpu[index] |
| 2732 | num_scheduled_tokens = scheduler_output.num_scheduled_tokens[req_id] |
| 2733 | num_prompt_tokens = length_from_prompt_token_ids_or_embeds( |
| 2734 | req.prompt_token_ids, req.prompt_embeds |
| 2735 | ) |
| 2736 | |
| 2737 | if num_computed_tokens + num_scheduled_tokens > num_prompt_tokens: |
| 2738 | prompt_part_len = max(0, num_prompt_tokens - num_computed_tokens) |
| 2739 | completion_part_len = max(0, num_scheduled_tokens - prompt_part_len) |
| 2740 | else: |
| 2741 | prompt_part_len = num_scheduled_tokens |
| 2742 | completion_part_len = 0 |
| 2743 | |
| 2744 | assert num_scheduled_tokens == prompt_part_len + completion_part_len |
| 2745 | |
| 2746 | if prompt_part_len > 0: |
| 2747 | # prompt's mrope_positions are pre-computed |
| 2748 | dst_start = mrope_pos_ptr |
| 2749 | dst_end = mrope_pos_ptr + prompt_part_len |
| 2750 | src_start = num_computed_tokens |
| 2751 | src_end = num_computed_tokens + prompt_part_len |
| 2752 | |
| 2753 | self.mrope_positions.cpu[:, dst_start:dst_end] = req.mrope_positions[ |
| 2754 | :, src_start:src_end |
| 2755 | ] |
| 2756 | mrope_pos_ptr += prompt_part_len |
| 2757 | |
| 2758 | if completion_part_len > 0: |
| 2759 | # compute completion's mrope_positions on-the-fly |
| 2760 | dst_start = mrope_pos_ptr |
| 2761 | dst_end = mrope_pos_ptr + completion_part_len |
| 2762 | |
| 2763 | assert req.mrope_position_delta is not None |
| 2764 | MRotaryEmbedding.get_next_input_positions_tensor( |
| 2765 | out=self.mrope_positions.np, |
| 2766 | out_offset=dst_start, |
| 2767 | mrope_position_delta=req.mrope_position_delta, |
| 2768 | context_len=num_computed_tokens + prompt_part_len, |
| 2769 | num_new_tokens=completion_part_len, |
| 2770 | ) |
| 2771 | |
| 2772 | mrope_pos_ptr += completion_part_len |
| 2773 | |
| 2774 | def _calc_spec_decode_metadata( |
| 2775 | self, |
| 2776 | num_draft_tokens: np.ndarray, |
| 2777 | cu_num_scheduled_tokens: np.ndarray, |
| 2778 | ) -> SpecDecodeMetadata: |
| 2779 | # Inputs: |
| 2780 | # cu_num_scheduled_tokens: [ 4, 104, 107, 207, 209] |
| 2781 | # num_draft_tokens: [ 3, 0, 2, 0, 1] |
| 2782 | # Outputs: |
| 2783 | # cu_num_draft_tokens: [ 3, 3, 5, 5, 6] |
| 2784 | # logits_indices: [ 0, 1, 2, 3, 103, 104, 105, 106, |
| 2785 | # 206, 207, 208] |
| 2786 | # target_logits_indices: [ 0, 1, 2, 5, 6, 9] |
| 2787 | # bonus_logits_indices: [ 3, 4, 7, 8, 10] |
| 2788 | |
| 2789 | # Compute the logits indices. |
| 2790 | # [4, 1, 3, 1, 2] |
| 2791 | num_sampled_tokens = num_draft_tokens + 1 |
| 2792 | |
| 2793 | # Step 1. |
| 2794 | # cu_num_sampled_tokens: [4, 5, 8, 9, 11] |
| 2795 | # _arange_scratch[:11]: [0, 1, 2, 3, 0, 0, 1, 2, 0, 0, 1] |
| 2796 | cu_num_sampled_tokens = self._get_cumsum_and_arange( |
| 2797 | num_sampled_tokens, self._arange_scratch, cumsum_dtype=np.int32 |
| 2798 | ) |
| 2799 | # Step 2. [0, 0, 0, 0, 103, 104, 104, 104, 206, 207, 207] |
| 2800 | logits_indices = np.repeat( |
| 2801 | cu_num_scheduled_tokens - num_sampled_tokens, num_sampled_tokens |
| 2802 | ) |
| 2803 | # Step 3. [0, 1, 2, 3, 103, 104, 105, 106, 206, 207, 208] |
| 2804 | logits_indices += self._arange_scratch[: cu_num_sampled_tokens[-1]] |
| 2805 | |
| 2806 | # Compute the bonus logits indices. |
| 2807 | bonus_logits_indices = cu_num_sampled_tokens - 1 |
| 2808 | |
| 2809 | # Compute the draft logits indices. |
| 2810 | # cu_num_draft_tokens: [3, 3, 5, 5, 6] |
| 2811 | # _arange_scratch[:6]: [0, 1, 2, 0, 1, 0] |
| 2812 | cu_num_draft_tokens = self._get_cumsum_and_arange( |
| 2813 | num_draft_tokens, self._arange_scratch, cumsum_dtype=np.int32 |
| 2814 | ) |
| 2815 | # [0, 0, 0, 5, 5, 9] |
| 2816 | target_logits_indices = np.repeat( |
| 2817 | cu_num_sampled_tokens - num_sampled_tokens, num_draft_tokens |
| 2818 | ) |
| 2819 | # [0, 1, 2, 5, 6, 9] |
| 2820 | target_logits_indices += self._arange_scratch[: cu_num_draft_tokens[-1]] |
| 2821 | |
| 2822 | cu_num_draft_tokens = async_tensor_h2d(cu_num_draft_tokens, device=self.device) |
| 2823 | cu_num_sampled_tokens = async_tensor_h2d( |
| 2824 | cu_num_sampled_tokens, device=self.device |
| 2825 | ) |
| 2826 | logits_indices = async_tensor_h2d(logits_indices, device=self.device) |
| 2827 | target_logits_indices = async_tensor_h2d( |
| 2828 | target_logits_indices, device=self.device |
| 2829 | ) |
| 2830 | bonus_logits_indices = async_tensor_h2d( |
| 2831 | bonus_logits_indices, device=self.device |
| 2832 | ) |
| 2833 | |
| 2834 | # Compute the draft token ids. |
| 2835 | # draft_token_indices: [ 1, 2, 3, 105, 106, 208] |
| 2836 | draft_token_ids = self.input_ids.gpu[logits_indices] |
| 2837 | draft_token_ids = draft_token_ids[target_logits_indices + 1] |
| 2838 | |
| 2839 | return SpecDecodeMetadata( |
| 2840 | draft_token_ids=draft_token_ids, |
| 2841 | num_draft_tokens=num_draft_tokens.tolist(), |
| 2842 | cu_num_draft_tokens=cu_num_draft_tokens, |
| 2843 | cu_num_sampled_tokens=cu_num_sampled_tokens, |
| 2844 | target_logits_indices=target_logits_indices, |
| 2845 | bonus_logits_indices=bonus_logits_indices, |
| 2846 | logits_indices=logits_indices, |
| 2847 | ) |
| 2848 | |
| 2849 | def _prepare_kv_sharing_fast_prefill( |
| 2850 | self, |
| 2851 | logits_indices: torch.Tensor, |
| 2852 | ) -> torch.Tensor: |
| 2853 | assert self.kv_sharing_fast_prefill_logits_indices is not None |
| 2854 | num_logits = logits_indices.shape[0] |
| 2855 | assert num_logits > 0 |
| 2856 | self.kv_sharing_fast_prefill_logits_indices[:num_logits].copy_(logits_indices) |
| 2857 | # There might have leftover indices in logits_indices[num_logits:] |
| 2858 | # from previous iterations, whose values may be greater than the |
| 2859 | # batch size in the current iteration. To ensure indices are always |
| 2860 | # valid, fill the padded indices with the last index. Broadcast the |
| 2861 | # scalar GPU-side to avoid a D2H sync on `.item()`. |
| 2862 | self.kv_sharing_fast_prefill_logits_indices[num_logits:] = logits_indices[-1] |
| 2863 | # Dispatch for the decoder portion of the model. |
| 2864 | _, batch_desc = self.cudagraph_dispatcher.dispatch( |
| 2865 | num_logits, invalid_modes={CUDAGraphMode.FULL} |
| 2866 | ) |
| 2867 | num_logits_padded = batch_desc.num_tokens |
| 2868 | logits_indices_padded = self.kv_sharing_fast_prefill_logits_indices[ |
| 2869 | :num_logits_padded |
| 2870 | ] |
| 2871 | return logits_indices_padded |
| 2872 | |
| 2873 | def _batch_mm_inputs_from_scheduler( |
| 2874 | self, |
| 2875 | scheduler_output: "SchedulerOutput", |
| 2876 | ) -> tuple[ |
| 2877 | list[str], |
| 2878 | list[tuple[str, MultiModalKwargsItem]], |
| 2879 | list[tuple[str, PlaceholderRange]], |
| 2880 | ]: |
| 2881 | """Batch multimodal inputs from scheduled encoder inputs. |
| 2882 | |
| 2883 | Args: |
| 2884 | scheduler_output: The scheduler output containing scheduled encoder |
| 2885 | inputs. |
| 2886 | |
| 2887 | Returns: |
| 2888 | A tuple of (mm_hashes, mm_kwargs, mm_lora_refs) where: |
| 2889 | - mm_hashes: List of multimodal hashes for each item |
| 2890 | - mm_kwargs: List of multimodal kwargs for each item |
| 2891 | - mm_lora_refs: List of (req_id, placeholder_range) for each item |
| 2892 | |
| 2893 | """ |
| 2894 | scheduled_encoder_inputs = scheduler_output.scheduled_encoder_inputs |
| 2895 | if not scheduled_encoder_inputs: |
| 2896 | return [], [], [] |
| 2897 | |
| 2898 | mm_hashes = list[str]() |
| 2899 | mm_kwargs = list[tuple[str, MultiModalKwargsItem]]() |
| 2900 | # Multimodal LoRA reference info to map each multimodal item |
| 2901 | # back to its request & position |
| 2902 | mm_lora_refs = list[tuple[str, PlaceholderRange]]() |
| 2903 | for req_id, encoder_input_ids in scheduled_encoder_inputs.items(): |
| 2904 | req_state = self.requests[req_id] |
| 2905 | |
| 2906 | for mm_input_id in encoder_input_ids: |
| 2907 | mm_feature = req_state.mm_features[mm_input_id] |
| 2908 | if mm_feature.data is None: |
| 2909 | continue |
| 2910 | |
| 2911 | mm_hashes.append(mm_feature.identifier) |
| 2912 | mm_kwargs.append((mm_feature.modality, mm_feature.data)) |
| 2913 | mm_lora_refs.append((req_id, mm_feature.mm_position)) |
| 2914 | |
| 2915 | return mm_hashes, mm_kwargs, mm_lora_refs |
| 2916 | |
| 2917 | def _cache_encoder_output( |
| 2918 | self, |
| 2919 | mm_hash: str, |
| 2920 | output: torch.Tensor, |
| 2921 | ec_manager_metadata: "EncoderCacheManagerMetadata | None", |
| 2922 | free_encoder_mm_hashes: list[str], |
| 2923 | ) -> None: |
| 2924 | """Store an encoder output for later multimodal embedding gather.""" |
| 2925 | del ec_manager_metadata, free_encoder_mm_hashes |
| 2926 | self.encoder_cache[mm_hash] = output |
| 2927 | self.maybe_save_ec_to_connector(self.encoder_cache, mm_hash) |
| 2928 | |
| 2929 | def _execute_mm_encoder( |
| 2930 | self, scheduler_output: "SchedulerOutput" |
| 2931 | ) -> list[torch.Tensor]: |
| 2932 | mm_hashes, mm_kwargs, mm_lora_refs = self._batch_mm_inputs_from_scheduler( |
| 2933 | scheduler_output |
| 2934 | ) |
| 2935 | |
| 2936 | if not mm_kwargs: |
| 2937 | return [] |
| 2938 | |
| 2939 | # `prompt_embeds` is a passthrough modality, the tensor is already in |
| 2940 | # the model embedding space, so no encoder runs. Inject each |
| 2941 | # `prompt_embeds` tensor directly into the encoder cache here so that |
| 2942 | # `_gather_mm_embeddings` can splice it via the standard `is_mm_embed` |
| 2943 | # path. |
| 2944 | pe_indices = [ |
| 2945 | i |
| 2946 | for i, (modality, _) in enumerate(mm_kwargs) |
| 2947 | if modality == "prompt_embeds" |
| 2948 | ] |
| 2949 | if pe_indices: |
| 2950 | for i in pe_indices: |
| 2951 | pe_tensor = mm_kwargs[i][1]["embedding"].data |
| 2952 | assert isinstance(pe_tensor, torch.Tensor) |
| 2953 | |
| 2954 | self._cache_encoder_output( |
| 2955 | mm_hashes[i], |
| 2956 | async_tensor_h2d(pe_tensor, device=self.device), |
| 2957 | scheduler_output.ec_manager_metadata, |
| 2958 | scheduler_output.free_encoder_mm_hashes, |
| 2959 | ) |
| 2960 | # Filter out `prompt_embeds` items from mm_kwargs/mm_hashes/mm_lora_refs |
| 2961 | # since they don't require further encoder processing. |
| 2962 | mm_hashes = [h for i, h in enumerate(mm_hashes) if i not in pe_indices] |
| 2963 | mm_kwargs = [k for i, k in enumerate(mm_kwargs) if i not in pe_indices] |
| 2964 | mm_lora_refs = [ |
| 2965 | r for i, r in enumerate(mm_lora_refs) if i not in pe_indices |
| 2966 | ] |
| 2967 | if not mm_kwargs: |
| 2968 | return [] # nothing left to encode after filtering out `prompt_embeds` |
| 2969 | |
| 2970 | should_time = bool( |
| 2971 | self.observability_config |
| 2972 | and self.observability_config.enable_mm_processor_stats |
| 2973 | and scheduler_output.scheduled_encoder_inputs |
| 2974 | ) |
| 2975 | |
| 2976 | # Batch mm inputs as much as we can: if a request in the batch has |
| 2977 | # multiple modalities or a different modality than the previous one, |
| 2978 | # we process it separately to preserve item order. |
| 2979 | # FIXME(ywang96): This is a hacky way to deal with multiple modalities |
| 2980 | # in the same batch while still being able to benefit from batching |
| 2981 | # multimodal inputs. The proper solution should be reordering the |
| 2982 | # encoder outputs. |
| 2983 | model = cast(SupportsMultiModal, self.model) |
| 2984 | |
| 2985 | if self.lora_config and self.lora_manager.supports_tower_connector_lora(): |
| 2986 | # Build LoRA mappings independently for encoder inputs |
| 2987 | # (encoder batch structure is different from main batch) |
| 2988 | prompt_lora_mapping = [] |
| 2989 | token_lora_mapping = [] |
| 2990 | lora_requests = set() |
| 2991 | encoder_token_counts = [] |
| 2992 | connector_token_counts = [] |
| 2993 | |
| 2994 | for (req_id, pos_info), (modality, mm_item) in zip( |
| 2995 | mm_lora_refs, |
| 2996 | mm_kwargs, |
| 2997 | ): |
| 2998 | req_idx = self.input_batch.req_id_to_index[req_id] |
| 2999 | lora_id = int(self.input_batch.request_lora_mapping[req_idx]) |
| 3000 | |
| 3001 | tower_tokens, connector_tokens = self.model.get_mm_lora_token_counts( # type: ignore[attr-defined] |
| 3002 | modality=modality, |
| 3003 | mm_kwargs=mm_item, |
| 3004 | num_mm_embeds=pos_info.get_num_embeds(), |
| 3005 | ) |
| 3006 | prompt_lora_mapping.append(lora_id) |
| 3007 | token_lora_mapping.extend([lora_id] * tower_tokens) |
| 3008 | encoder_token_counts.append(tower_tokens) |
| 3009 | connector_token_counts.append(connector_tokens) |
| 3010 | |
| 3011 | if lora_id > 0: |
| 3012 | lora_request = self.input_batch.lora_id_to_lora_request.get(lora_id) |
| 3013 | if lora_request is not None: |
| 3014 | lora_requests.add(lora_request) |
| 3015 | |
| 3016 | # Set tower adapter mapping |
| 3017 | tower_mapping = LoRAMapping( |
| 3018 | tuple(token_lora_mapping), |
| 3019 | tuple(prompt_lora_mapping), |
| 3020 | is_prefill=True, |
| 3021 | type=LoRAMappingType.TOWER, |
| 3022 | ) |
| 3023 | self.lora_manager.set_active_adapters(lora_requests, tower_mapping) |
| 3024 | |
| 3025 | # Only set connector mapping if the model actually has a connector. |
| 3026 | # Some multimodal models inherit a stub `get_num_mm_connector_tokens` |
| 3027 | # from `SupportsMultiModal`, which returns None and should not be |
| 3028 | # treated as a signal that connector LoRA is supported. |
| 3029 | mm_mapping = ( |
| 3030 | self.model.get_mm_mapping() # type: ignore[attr-defined] |
| 3031 | if hasattr(self.model, "get_mm_mapping") |
| 3032 | else None |
| 3033 | ) |
| 3034 | if ( |
| 3035 | mm_mapping is not None |
| 3036 | and mm_mapping.connector |
| 3037 | and all(count is not None for count in connector_token_counts) |
| 3038 | ): |
| 3039 | connector_token_mapping = np.repeat( |
| 3040 | np.array(prompt_lora_mapping, dtype=np.int32), |
| 3041 | np.array(connector_token_counts, dtype=np.int32), |
| 3042 | ) |
| 3043 | connector_mapping = LoRAMapping( |
| 3044 | index_mapping=tuple(connector_token_mapping.tolist()), |
| 3045 | prompt_mapping=tuple(prompt_lora_mapping), |
| 3046 | is_prefill=True, |
| 3047 | type=LoRAMappingType.CONNECTOR, |
| 3048 | ) |
| 3049 | |
| 3050 | self.lora_manager.set_active_adapters( |
| 3051 | lora_requests, |
| 3052 | connector_mapping, |
| 3053 | ) |
| 3054 | |
| 3055 | encoder_outputs: list[torch.Tensor] = [] |
| 3056 | # Track the current index in mm_kwargs/mm_lora_refs to map groups to request IDs |
| 3057 | current_item_idx = 0 |
| 3058 | for modality, num_items, mm_kwargs_batch in group_and_batch_mm_kwargs( |
| 3059 | mm_kwargs, device=self.device, pin_memory=PIN_MEMORY |
| 3060 | ): |
| 3061 | batch_outputs: MultiModalEmbeddings |
| 3062 | |
| 3063 | # EVS and dynamic res video related change. |
| 3064 | # (ekhvedchenia): Temporary hack to limit peak memory usage when |
| 3065 | # processing multimodal data. This solves the issue with scheduler |
| 3066 | # putting too many video samples into a single batch. Scheduler |
| 3067 | # uses pruned vision tokens count to compare it versus compute |
| 3068 | # budget which is incorrect (Either input media size or non-pruned |
| 3069 | # output vision tokens count should be considered) |
| 3070 | # dynamic res video for nemotron temporarily uses this hack via |
| 3071 | # requires_sequential_video_encoding |
| 3072 | # because it doesn't yet support video batching. |
| 3073 | # TODO(ywang96): Fix memory profiling to take EVS into account and |
| 3074 | # remove this hack. |
| 3075 | if ( |
| 3076 | ( |
| 3077 | self.is_multimodal_pruning_enabled |
| 3078 | or self.requires_sequential_video_encoding |
| 3079 | ) |
| 3080 | and modality == "video" |
| 3081 | and num_items > 1 |
| 3082 | ): |
| 3083 | batch_outputs_lst = list[torch.Tensor]() |
| 3084 | for video_idx in range(num_items): |
| 3085 | video_mm_kwargs_item = mm_kwargs[current_item_idx + video_idx] |
| 3086 | with self.timed_encoder_operation( |
| 3087 | should_time, mm_lora_refs, current_item_idx + video_idx, 1 |
| 3088 | ): |
| 3089 | _, _, micro_batch_mm_inputs = next( |
| 3090 | group_and_batch_mm_kwargs( |
| 3091 | [video_mm_kwargs_item], |
| 3092 | device=self.device, |
| 3093 | pin_memory=PIN_MEMORY, |
| 3094 | ) |
| 3095 | ) |
| 3096 | |
| 3097 | micro_batch_outputs = model.embed_multimodal( |
| 3098 | **micro_batch_mm_inputs |
| 3099 | ) |
| 3100 | |
| 3101 | batch_outputs_lst.extend(micro_batch_outputs) |
| 3102 | |
| 3103 | batch_outputs = batch_outputs_lst |
| 3104 | else: |
| 3105 | # Run the encoder. |
| 3106 | # `batch_outputs` is either of the following: |
| 3107 | # 1. A tensor of shape (num_items, feature_size, hidden_size) |
| 3108 | # in case feature_size is fixed across all multimodal items. |
| 3109 | # 2. A list or tuple (length: num_items) of tensors, |
| 3110 | # each of shape (feature_size, hidden_size) in case the feature |
| 3111 | # size is dynamic depending on the input multimodal items. |
| 3112 | |
| 3113 | with self.timed_encoder_operation( |
| 3114 | should_time, mm_lora_refs, current_item_idx, num_items |
| 3115 | ): |
| 3116 | cudagraph_output = None |
| 3117 | if ( |
| 3118 | self.encoder_cudagraph_manager is not None |
| 3119 | and self.encoder_cudagraph_manager.supports_modality(modality) |
| 3120 | ): |
| 3121 | cudagraph_output = self.encoder_cudagraph_manager.execute( |
| 3122 | mm_kwargs_batch, |
| 3123 | ) |
| 3124 | |
| 3125 | if cudagraph_output is not None: |
| 3126 | batch_outputs = cudagraph_output |
| 3127 | else: |
| 3128 | batch_outputs = model.embed_multimodal(**mm_kwargs_batch) |
| 3129 | |
| 3130 | sanity_check_mm_encoder_outputs(batch_outputs, expected_num_items=num_items) |
| 3131 | encoder_outputs.extend(batch_outputs) |
| 3132 | |
| 3133 | current_item_idx += num_items |
| 3134 | |
| 3135 | # Cache the encoder outputs by mm_hash |
| 3136 | for mm_hash, output in zip(mm_hashes, encoder_outputs): |
| 3137 | self._cache_encoder_output( |
| 3138 | mm_hash, |
| 3139 | output, |
| 3140 | scheduler_output.ec_manager_metadata, |
| 3141 | scheduler_output.free_encoder_mm_hashes, |
| 3142 | ) |
| 3143 | logger.debug("Finish execute for mm hash %s", mm_hash) |
| 3144 | |
| 3145 | return encoder_outputs |
| 3146 | |
| 3147 | def _get_encoder_output_from_cache(self, mm_hash: str) -> torch.Tensor | None: |
| 3148 | """Return a cached encoder output for multimodal |
| 3149 | embedding gather.""" |
| 3150 | return self.encoder_cache.get(mm_hash, None) |
| 3151 | |
| 3152 | def _gather_mm_embeddings( |
| 3153 | self, |
| 3154 | scheduler_output: "SchedulerOutput", |
| 3155 | shift_computed_tokens: int = 0, |
| 3156 | ) -> tuple[list[torch.Tensor], torch.Tensor]: |
| 3157 | total_num_scheduled_tokens = scheduler_output.total_num_scheduled_tokens |
| 3158 | |
| 3159 | mm_embeds = list[torch.Tensor]() |
| 3160 | is_mm_embed = torch.zeros( |
| 3161 | total_num_scheduled_tokens, |
| 3162 | dtype=torch.bool, |
| 3163 | device="cpu", |
| 3164 | pin_memory=PIN_MEMORY, |
| 3165 | ) |
| 3166 | |
| 3167 | req_start_idx = 0 |
| 3168 | should_sync_mrope_positions = False |
| 3169 | |
| 3170 | for req_id in self.input_batch.req_ids: |
| 3171 | mm_embeds_req: list[torch.Tensor] = [] |
| 3172 | |
| 3173 | num_scheduled_tokens = scheduler_output.num_scheduled_tokens[req_id] |
| 3174 | req_state = self.requests[req_id] |
| 3175 | num_computed_tokens = req_state.num_computed_tokens + shift_computed_tokens |
| 3176 | |
| 3177 | mm_features = req_state.mm_features |
| 3178 | lo, hi = get_mm_features_in_window( |
| 3179 | mm_features, |
| 3180 | start=num_computed_tokens, |
| 3181 | end=num_computed_tokens + num_scheduled_tokens, |
| 3182 | ) |
| 3183 | for i in range(lo, hi): |
| 3184 | mm_feature = mm_features[i] |
| 3185 | pos_info = mm_feature.mm_position |
| 3186 | start_pos = pos_info.offset |
| 3187 | num_encoder_tokens = pos_info.length |
| 3188 | |
| 3189 | start_idx = max(num_computed_tokens - start_pos, 0) |
| 3190 | end_idx = min( |
| 3191 | num_computed_tokens - start_pos + num_scheduled_tokens, |
| 3192 | num_encoder_tokens, |
| 3193 | ) |
| 3194 | assert start_idx < end_idx |
| 3195 | curr_embeds_start, curr_embeds_end = ( |
| 3196 | pos_info.get_embeds_indices_in_range(start_idx, end_idx) |
| 3197 | ) |
| 3198 | # If there are no embeddings in the current range, we skip |
| 3199 | # gathering the embeddings. |
| 3200 | if curr_embeds_start == curr_embeds_end: |
| 3201 | continue |
| 3202 | |
| 3203 | mm_hash = mm_feature.identifier |
| 3204 | encoder_output = self._get_encoder_output_from_cache(mm_hash) |
| 3205 | if encoder_output is None: |
| 3206 | # A feature starting at/after the processed boundary is only |
| 3207 | # reached via the drafter's +1 look-ahead and might not be |
| 3208 | # encoded yet; fall back to the token embedding for drafting. |
| 3209 | if ( |
| 3210 | start_pos |
| 3211 | >= req_state.num_computed_tokens + num_scheduled_tokens |
| 3212 | ): |
| 3213 | continue |
| 3214 | raise RuntimeError(f"Encoder cache miss for {mm_hash}.") |
| 3215 | |
| 3216 | if (is_embed := pos_info.is_embed) is not None: |
| 3217 | is_embed = is_embed[start_idx:end_idx] |
| 3218 | mm_embeds_item = encoder_output[curr_embeds_start:curr_embeds_end] |
| 3219 | else: |
| 3220 | mm_embeds_item = encoder_output[start_idx:end_idx] |
| 3221 | |
| 3222 | req_start_pos = req_start_idx + start_pos - num_computed_tokens |
| 3223 | # OR mask for overlapping mm_features (use_audio_in_video) |
| 3224 | if is_embed is None: |
| 3225 | is_mm_embed[req_start_pos + start_idx : req_start_pos + end_idx] = ( |
| 3226 | True |
| 3227 | ) |
| 3228 | else: |
| 3229 | is_mm_embed[ |
| 3230 | req_start_pos + start_idx : req_start_pos + end_idx |
| 3231 | ] |= is_embed |
| 3232 | set_mm_embedding_modality(mm_embeds_item, mm_feature.modality) |
| 3233 | mm_embeds_req.append(mm_embeds_item) |
| 3234 | |
| 3235 | if self.is_multimodal_pruning_enabled and self.uses_mrope: |
| 3236 | assert req_state.mrope_positions is not None |
| 3237 | should_sync_mrope_positions = True |
| 3238 | old_mm_embeds_req = mm_embeds_req |
| 3239 | mm_embeds_req, new_mrope_positions, new_delta = ( |
| 3240 | self.model.recompute_mrope_positions( |
| 3241 | input_ids=req_state.prompt_token_ids, |
| 3242 | multimodal_embeddings=mm_embeds_req, |
| 3243 | mrope_positions=req_state.mrope_positions, |
| 3244 | num_computed_tokens=req_state.num_computed_tokens, |
| 3245 | ) |
| 3246 | ) |
| 3247 | mm_embeds_req = [ |
| 3248 | copy_mm_embedding_modality(src, dst) |
| 3249 | for src, dst in zip(old_mm_embeds_req, mm_embeds_req) |
| 3250 | ] |
| 3251 | req_state.mrope_positions.copy_(new_mrope_positions) |
| 3252 | req_state.mrope_position_delta = new_delta |
| 3253 | |
| 3254 | mm_embeds.extend(mm_embeds_req) |
| 3255 | req_start_idx += num_scheduled_tokens |
| 3256 | |
| 3257 | if should_sync_mrope_positions: |
| 3258 | self._calc_mrope_positions(scheduler_output) |
| 3259 | self.mrope_positions.copy_to_gpu(total_num_scheduled_tokens) |
| 3260 | |
| 3261 | return mm_embeds, is_mm_embed |
| 3262 | |
| 3263 | def get_model(self) -> nn.Module: |
| 3264 | if not hasattr(self, "model"): |
| 3265 | raise ValueError("Cannot get model before model has been initialized") |
| 3266 | if isinstance( |
| 3267 | self.model, (CUDAGraphWrapper, UBatchWrapper, BreakableCUDAGraphWrapper) |
| 3268 | ): |
| 3269 | # get raw model out of the cudagraph wrapper. |
| 3270 | return self.model.unwrap() |
| 3271 | return self.model |
| 3272 | |
| 3273 | def get_draft_model(self) -> nn.Module | None: |
| 3274 | drafter = getattr(self, "drafter", None) |
| 3275 | if drafter is None: |
| 3276 | return None |
| 3277 | model = getattr(drafter, "model", None) |
| 3278 | if isinstance( |
| 3279 | model, (CUDAGraphWrapper, UBatchWrapper, BreakableCUDAGraphWrapper) |
| 3280 | ): |
| 3281 | return cast(nn.Module, model.unwrap()) |
| 3282 | return cast(nn.Module | None, model) |
| 3283 | |
| 3284 | def get_supported_generation_tasks(self) -> list[GenerationTask]: |
| 3285 | model = self.get_model() |
| 3286 | supported_tasks = list[GenerationTask]() |
| 3287 | |
| 3288 | if is_text_generation_model(model): |
| 3289 | supported_tasks.append("generate") |
| 3290 | |
| 3291 | if supports_transcription(model): |
| 3292 | if model.supports_transcription_only: |
| 3293 | return ["transcription"] |
| 3294 | |
| 3295 | supported_tasks.append("transcription") |
| 3296 | |
| 3297 | if supports_realtime(model): |
| 3298 | supported_tasks.append("realtime") |
| 3299 | |
| 3300 | return supported_tasks |
| 3301 | |
| 3302 | def get_supported_pooling_tasks(self) -> list[PoolingTask]: |
| 3303 | model = self.get_model() |
| 3304 | if not is_pooling_model(model): |
| 3305 | return [] |
| 3306 | |
| 3307 | return list(model.pooler.get_supported_tasks()) |
| 3308 | |
| 3309 | def get_supported_tasks(self) -> tuple[SupportedTask, ...]: |
| 3310 | tasks = list[SupportedTask]() |
| 3311 | |
| 3312 | if self.model_config.runner_type == "generate": |
| 3313 | tasks.extend(self.get_supported_generation_tasks()) |
| 3314 | if self.model_config.runner_type == "pooling": |
| 3315 | tasks.extend(self.get_supported_pooling_tasks()) |
| 3316 | |
| 3317 | return tuple(tasks) |
| 3318 | |
| 3319 | def sync_and_gather_intermediate_tensors( |
| 3320 | self, |
| 3321 | num_tokens: int, |
| 3322 | intermediate_tensors: IntermediateTensors | None, |
| 3323 | sync_self: bool, |
| 3324 | ) -> IntermediateTensors: |
| 3325 | assert self.intermediate_tensors is not None |
| 3326 | |
| 3327 | tp = self.vllm_config.parallel_config.tensor_parallel_size |
| 3328 | is_rs = is_residual_scattered_for_sp(self.vllm_config, num_tokens) |
| 3329 | |
| 3330 | # When sequence parallelism is enabled, the "residual" tensor is |
| 3331 | # sharded across TP ranks. All-gather it here because downstream |
| 3332 | # QKV + Attention needs the full residual before the SP split point. |
| 3333 | if sync_self: |
| 3334 | assert intermediate_tensors is not None |
| 3335 | for k, v in intermediate_tensors.items(): |
| 3336 | is_scattered = k == "residual" and is_rs |
| 3337 | if is_scattered: |
| 3338 | local_len = num_tokens // tp |
| 3339 | v = get_tp_group().all_gather(v[:local_len], dim=0) |
| 3340 | |
| 3341 | self.intermediate_tensors[k][:num_tokens].copy_( |
| 3342 | v[:num_tokens], non_blocking=True |
| 3343 | ) |
| 3344 | |
| 3345 | return IntermediateTensors( |
| 3346 | {k: v[:num_tokens] for k, v in self.intermediate_tensors.items()} |
| 3347 | ) |
| 3348 | |
| 3349 | def eplb_step(self, is_dummy: bool = False, is_profile: bool = False) -> None: |
| 3350 | """Step for the EPLB (Expert Parallelism Load Balancing) state.""" |
| 3351 | if not self.parallel_config.enable_eplb or self.eep_eplb_suppressed: |
| 3352 | return |
| 3353 | |
| 3354 | assert self.eplb_state is not None |
| 3355 | assert self._moe_model is not None |
| 3356 | self.eplb_state.step( |
| 3357 | is_dummy, |
| 3358 | is_profile, |
| 3359 | log_stats=self.parallel_config.eplb_config.log_balancedness, |
| 3360 | ) |
| 3361 | |
| 3362 | def setup_eplb_from_mapping( |
| 3363 | self, |
| 3364 | expanded_physical_to_logical: torch.Tensor, |
| 3365 | ) -> None: |
| 3366 | assert self.eplb_state is not None |
| 3367 | self.eplb_state.update_mapping( |
| 3368 | self.model_config, |
| 3369 | expanded_physical_to_logical, |
| 3370 | ) |
| 3371 | |
| 3372 | def _pool( |
| 3373 | self, |
| 3374 | hidden_states: torch.Tensor, |
| 3375 | num_scheduled_tokens: int, |
| 3376 | num_scheduled_tokens_np: np.ndarray, |
| 3377 | kv_connector_output: KVConnectorOutput | None, |
| 3378 | ) -> ModelRunnerOutput | AsyncModelRunnerOutput: |
| 3379 | num_reqs = self.input_batch.num_reqs |
| 3380 | assert num_reqs == len(self.input_batch.pooling_params), ( |
| 3381 | "Either all or none of the requests in a batch must be pooling request" |
| 3382 | ) |
| 3383 | |
| 3384 | hidden_states = hidden_states[:num_scheduled_tokens] |
| 3385 | seq_lens_cpu = self.optimistic_seq_lens_cpu[:num_reqs] |
| 3386 | |
| 3387 | pooling_metadata = self.input_batch.get_pooling_metadata() |
| 3388 | pooling_metadata.build_pooling_cursor( |
| 3389 | num_scheduled_tokens_np, |
| 3390 | seq_lens_cpu, |
| 3391 | device=hidden_states.device, |
| 3392 | query_start_loc_gpu=self.query_start_loc.gpu[: num_reqs + 1], |
| 3393 | ) |
| 3394 | |
| 3395 | model = cast(VllmModelForPooling, self.model) |
| 3396 | raw_pooler_output: PoolerOutput = model.pooler( |
| 3397 | hidden_states=hidden_states, pooling_metadata=pooling_metadata |
| 3398 | ) |
| 3399 | |
| 3400 | finished_mask = pooling_metadata.get_pooling_cursor().get_finished_mask() |
| 3401 | raw_pooler_output = self.late_interaction_runner.postprocess_pooler_output( |
| 3402 | raw_pooler_output=raw_pooler_output, |
| 3403 | pooling_params=pooling_metadata.pooling_params, |
| 3404 | req_ids=self.input_batch.req_ids, |
| 3405 | finished_mask=finished_mask, |
| 3406 | ) |
| 3407 | |
| 3408 | model_runner_output = ModelRunnerOutput( |
| 3409 | req_ids=self.input_batch.req_ids.copy(), |
| 3410 | req_id_to_index=self.input_batch.req_id_to_index.copy(), |
| 3411 | kv_connector_output=kv_connector_output, |
| 3412 | ) |
| 3413 | |
| 3414 | if raw_pooler_output is None or not any(finished_mask): |
| 3415 | self._sync_device() |
| 3416 | model_runner_output.pooler_output = [None] * num_reqs |
| 3417 | return model_runner_output |
| 3418 | |
| 3419 | if not current_platform.is_cuda_alike(): |
| 3420 | # cpu/xpu runners cannot use the CUDA stream/event-based wrapper. |
| 3421 | model_runner_output.pooler_output = _copy_pooler_output_to_cpu( |
| 3422 | raw_pooler_output=raw_pooler_output, |
| 3423 | finished_mask=finished_mask, |
| 3424 | ) |
| 3425 | self._sync_device() |
| 3426 | return model_runner_output |
| 3427 | |
| 3428 | return AsyncGPUPoolingModelRunnerOutput( |
| 3429 | model_runner_output=model_runner_output, |
| 3430 | raw_pooler_output=raw_pooler_output, |
| 3431 | finished_mask=finished_mask, |
| 3432 | async_output_copy_stream=self._get_or_create_async_output_copy_stream(), |
| 3433 | ) |
| 3434 | |
| 3435 | def _pad_for_sequence_parallelism(self, num_scheduled_tokens: int) -> int: |
| 3436 | # Pad tokens to multiple of tensor_parallel_size when |
| 3437 | # enabled collective fusion for SP |
| 3438 | tp_size = self.vllm_config.parallel_config.tensor_parallel_size |
| 3439 | if self.compilation_config.pass_config.enable_sp and tp_size > 1: |
| 3440 | return round_up(num_scheduled_tokens, tp_size) |
| 3441 | return num_scheduled_tokens |
| 3442 | |
| 3443 | def _prepare_padding_mask( |
| 3444 | self, num_tokens_unpadded: int, num_tokens_padded: int |
| 3445 | ) -> torch.Tensor: |
| 3446 | padding_mask = self.is_padding[:num_tokens_padded] |
| 3447 | padding_mask[:num_tokens_unpadded].fill_(False) |
| 3448 | padding_mask[num_tokens_unpadded:].fill_(True) |
| 3449 | return padding_mask |
| 3450 | |
| 3451 | def _prepare_mm_inputs( |
| 3452 | self, num_tokens: int |
| 3453 | ) -> tuple[torch.Tensor | None, torch.Tensor]: |
| 3454 | if self.model.requires_raw_input_tokens: |
| 3455 | input_ids = self.input_ids.gpu[:num_tokens] |
| 3456 | else: |
| 3457 | input_ids = None |
| 3458 | |
| 3459 | inputs_embeds = self.inputs_embeds.gpu[:num_tokens] |
| 3460 | return input_ids, inputs_embeds |
| 3461 | |
| 3462 | def _preprocess( |
| 3463 | self, |
| 3464 | scheduler_output: "SchedulerOutput", |
| 3465 | num_input_tokens: int, # Padded |
| 3466 | intermediate_tensors: IntermediateTensors | None = None, |
| 3467 | ) -> tuple[ |
| 3468 | torch.Tensor | None, |
| 3469 | torch.Tensor | None, |
| 3470 | torch.Tensor, |
| 3471 | IntermediateTensors | None, |
| 3472 | dict[str, Any], |
| 3473 | ECConnectorOutput | None, |
| 3474 | ]: |
| 3475 | num_scheduled_tokens = scheduler_output.total_num_scheduled_tokens |
| 3476 | is_first_rank = get_pp_group().is_first_rank |
| 3477 | is_encoder_decoder = self.model_config.is_encoder_decoder |
| 3478 | |
| 3479 | # Clamp speculative scheduler placeholders (-1) before embedding lookup. |
| 3480 | if self.speculative_config is not None: |
| 3481 | self.input_ids.gpu[:num_input_tokens].clamp_(min=0) |
| 3482 | |
| 3483 | # _prepare_inputs may reorder the batch, so we must gather multi |
| 3484 | # modal outputs after that to ensure the correct order |
| 3485 | ec_connector_output = None |
| 3486 | |
| 3487 | if self.supports_mm_inputs and is_first_rank and not is_encoder_decoder: |
| 3488 | # Run the multimodal encoder if any. |
| 3489 | with self.maybe_get_ec_connector_output( |
| 3490 | scheduler_output, |
| 3491 | encoder_cache=self.encoder_cache, |
| 3492 | ) as ec_connector_output: |
| 3493 | self._execute_mm_encoder(scheduler_output) |
| 3494 | mm_embeds, is_mm_embed = self._gather_mm_embeddings(scheduler_output) |
| 3495 | |
| 3496 | # NOTE(woosuk): To unify token ids and soft tokens (vision |
| 3497 | # embeddings), we always use embeddings (rather than token ids) |
| 3498 | # as input to the multimodal model, even when the input is text. |
| 3499 | if self.enable_prompt_embeds and self.input_batch.req_prompt_embeds: |
| 3500 | # Some positions carry precomputed prompt_embeds: they are |
| 3501 | # already in self.inputs_embeds and marked is_token_ids=False. |
| 3502 | # Embed only the token-id positions (zeroing the placeholder ids |
| 3503 | # at prompt_embeds positions so the embedding gather cannot read |
| 3504 | # out-of-range ids), and write them back without clobbering the |
| 3505 | # prompt_embeds positions. |
| 3506 | is_token_ids = self.is_token_ids.gpu[:num_scheduled_tokens] |
| 3507 | safe_input_ids = torch.where( |
| 3508 | is_token_ids, |
| 3509 | self.input_ids.gpu[:num_scheduled_tokens], |
| 3510 | 0, |
| 3511 | ) |
| 3512 | inputs_embeds_scheduled = self.model.embed_input_ids( |
| 3513 | safe_input_ids, |
| 3514 | multimodal_embeddings=mm_embeds, |
| 3515 | is_multimodal=is_mm_embed, |
| 3516 | ) |
| 3517 | target = self.inputs_embeds.gpu[:num_scheduled_tokens] |
| 3518 | self.inputs_embeds.gpu[:num_scheduled_tokens] = torch.where( |
| 3519 | is_token_ids.unsqueeze(-1), |
| 3520 | inputs_embeds_scheduled, |
| 3521 | target, |
| 3522 | ) |
| 3523 | else: |
| 3524 | inputs_embeds_scheduled = self.model.embed_input_ids( |
| 3525 | self.input_ids.gpu[:num_scheduled_tokens], |
| 3526 | multimodal_embeddings=mm_embeds, |
| 3527 | is_multimodal=is_mm_embed, |
| 3528 | ) |
| 3529 | |
| 3530 | # TODO(woosuk): Avoid the copy. Optimize. |
| 3531 | self.inputs_embeds.gpu[:num_scheduled_tokens].copy_( |
| 3532 | inputs_embeds_scheduled |
| 3533 | ) |
| 3534 | |
| 3535 | input_ids, inputs_embeds = self._prepare_mm_inputs(num_input_tokens) |
| 3536 | model_kwargs = { |
| 3537 | **self._init_model_kwargs(), |
| 3538 | **self._extract_mm_kwargs(scheduler_output), |
| 3539 | } |
| 3540 | elif self.enable_prompt_embeds and is_first_rank: |
| 3541 | # Get the input embeddings for the tokens that are not input embeds, |
| 3542 | # then put them into the appropriate positions. |
| 3543 | # TODO(qthequartermasterman): Since even when prompt embeds are |
| 3544 | # enabled, (a) not all requests will use prompt embeds, and (b) |
| 3545 | # after the initial prompt is processed, the rest of the generated |
| 3546 | # tokens will be token ids, it is not desirable to have the |
| 3547 | # embedding layer outside of the CUDA graph all the time. The v0 |
| 3548 | # engine avoids this by "double compiling" the CUDA graph, once |
| 3549 | # with input_ids and again with inputs_embeds, for all num_tokens. |
| 3550 | # If a batch only has token ids, then including the embedding layer |
| 3551 | # in the CUDA graph will be more performant (like in the else case |
| 3552 | # below). |
| 3553 | is_token_ids = self.is_token_ids.np[:num_scheduled_tokens] |
| 3554 | token_ids_idx_np = np.nonzero(is_token_ids)[0] |
| 3555 | # Some tokens ids may need to become embeds |
| 3556 | if token_ids_idx_np.size > 0: |
| 3557 | token_ids_idx = async_tensor_h2d(token_ids_idx_np, device=self.device) |
| 3558 | token_ids = self.input_ids.gpu[token_ids_idx] |
| 3559 | tokens_to_embeds = self.model.embed_input_ids(input_ids=token_ids) |
| 3560 | self.inputs_embeds.gpu[token_ids_idx] = tokens_to_embeds |
| 3561 | |
| 3562 | inputs_embeds = self.inputs_embeds.gpu[:num_input_tokens] |
| 3563 | model_kwargs = self._init_model_kwargs() |
| 3564 | input_ids = None |
| 3565 | else: |
| 3566 | # For text-only models, we use token ids as input. |
| 3567 | # While it is possible to use embeddings as input just like the |
| 3568 | # multimodal models, it is not desirable for performance since |
| 3569 | # then the embedding layer is not included in the CUDA graph. |
| 3570 | input_ids = self.input_ids.gpu[:num_input_tokens] |
| 3571 | inputs_embeds = None |
| 3572 | model_kwargs = self._init_model_kwargs() |
| 3573 | |
| 3574 | if self.uses_mrope: |
| 3575 | positions = self.mrope_positions.gpu[:, :num_input_tokens] |
| 3576 | else: |
| 3577 | positions = self.positions[:num_input_tokens] |
| 3578 | if num_input_tokens > num_scheduled_tokens: |
| 3579 | self.positions[num_scheduled_tokens:num_input_tokens].zero_() |
| 3580 | |
| 3581 | if is_first_rank: |
| 3582 | intermediate_tensors = None |
| 3583 | else: |
| 3584 | assert intermediate_tensors is not None |
| 3585 | intermediate_tensors = self.sync_and_gather_intermediate_tensors( |
| 3586 | num_input_tokens, intermediate_tensors, True |
| 3587 | ) |
| 3588 | |
| 3589 | if is_encoder_decoder and scheduler_output.scheduled_encoder_inputs: |
| 3590 | # Run the encoder, just like we do with other multimodal inputs. |
| 3591 | # For an encoder-decoder model, our processing here is a bit |
| 3592 | # simpler, because the outputs are just passed to the decoder. |
| 3593 | # We are not doing any prompt replacement. We also will only |
| 3594 | # ever have a single encoder input. |
| 3595 | encoder_outputs = self._execute_mm_encoder(scheduler_output) |
| 3596 | model_kwargs.update({"encoder_outputs": encoder_outputs}) |
| 3597 | |
| 3598 | return ( |
| 3599 | input_ids, |
| 3600 | inputs_embeds, |
| 3601 | positions, |
| 3602 | intermediate_tensors, |
| 3603 | model_kwargs, |
| 3604 | ec_connector_output, |
| 3605 | ) |
| 3606 | |
| 3607 | def _sample( |
| 3608 | self, |
| 3609 | logits: torch.Tensor | None, |
| 3610 | spec_decode_metadata: SpecDecodeMetadata | None, |
| 3611 | ) -> SamplerOutput: |
| 3612 | # Sample the next token and get logprobs if needed. |
| 3613 | sampling_metadata = self.input_batch.sampling_metadata |
| 3614 | # Update output token ids with tokens sampled in last step |
| 3615 | # if async scheduling and required by current sampling params. |
| 3616 | self.input_batch.update_async_output_token_ids() |
| 3617 | if spec_decode_metadata is None: |
| 3618 | return self.sampler( |
| 3619 | logits=logits, |
| 3620 | sampling_metadata=sampling_metadata, |
| 3621 | ) |
| 3622 | |
| 3623 | # Update spec_token_ids with real draft tokens from pre step only when |
| 3624 | # output_token_ids is needed (penalties or bad_words are in use). |
| 3625 | if self.use_async_scheduling and self._draft_token_req_ids is not None: |
| 3626 | draft_token_ids_cpu, _ = self._get_draft_token_ids_cpu() |
| 3627 | self.input_batch.update_async_spec_token_ids(draft_token_ids_cpu) |
| 3628 | |
| 3629 | draft_probs = self._get_spec_decode_draft_probs(spec_decode_metadata) |
| 3630 | sampler_output = self.rejection_sampler( |
| 3631 | spec_decode_metadata, |
| 3632 | draft_probs, |
| 3633 | logits, |
| 3634 | sampling_metadata, |
| 3635 | ) |
| 3636 | return sampler_output |
| 3637 | |
| 3638 | def _bookkeeping_sync( |
| 3639 | self, |
| 3640 | scheduler_output: "SchedulerOutput", |
| 3641 | sampler_output: SamplerOutput, |
| 3642 | logits: torch.Tensor | None, |
| 3643 | hidden_states: torch.Tensor, |
| 3644 | num_scheduled_tokens: int, |
| 3645 | ) -> tuple[ |
| 3646 | dict[str, int], |
| 3647 | torch.Tensor | None, |
| 3648 | LogprobsLists | None, |
| 3649 | list[list[int]], |
| 3650 | dict[str, LogprobsTensors | None], |
| 3651 | list[str], |
| 3652 | dict[str, int], |
| 3653 | list[int], |
| 3654 | ]: |
| 3655 | num_nans: torch.Tensor | None = None |
| 3656 | num_nans_in_logits: dict[str, int] = {} |
| 3657 | if envs.VLLM_COMPUTE_NANS_IN_LOGITS: |
| 3658 | if self.use_async_scheduling: |
| 3659 | # Keep the counts on device; they ride the async output copy |
| 3660 | # stream rather than blocking here. |
| 3661 | num_nans = None if logits is None else count_nans_per_row(logits) |
| 3662 | else: |
| 3663 | num_nans_in_logits = self._get_nans_in_logits(logits) |
| 3664 | |
| 3665 | num_reqs = self.input_batch.num_reqs |
| 3666 | discard_sampled_tokens_req_indices = np.nonzero( |
| 3667 | self.discard_request_mask.np[:num_reqs] |
| 3668 | )[0] |
| 3669 | for i in discard_sampled_tokens_req_indices: |
| 3670 | gen = self.input_batch.generators.get(int(i)) |
| 3671 | if gen is not None: |
| 3672 | gen.set_offset(gen.get_offset() - 4) |
| 3673 | |
| 3674 | # Copy some objects so they don't get modified after returning. |
| 3675 | # This is important when using async scheduling. |
| 3676 | req_ids_output_copy = self.input_batch.req_ids.copy() |
| 3677 | req_id_to_index_output_copy = self.input_batch.req_id_to_index.copy() |
| 3678 | |
| 3679 | num_sampled_tokens = sampler_output.sampled_token_ids.shape[0] |
| 3680 | sampled_token_ids = sampler_output.sampled_token_ids |
| 3681 | logprobs_tensors = sampler_output.logprobs_tensors |
| 3682 | invalid_req_indices = [] |
| 3683 | logprobs_lists = None |
| 3684 | if not self.use_async_scheduling: |
| 3685 | with gpu_sync_allowed(): |
| 3686 | # Get the valid generated tokens. |
| 3687 | max_gen_len = sampled_token_ids.shape[-1] |
| 3688 | if max_gen_len == 1: |
| 3689 | # No spec decode tokens. |
| 3690 | valid_sampled_token_ids = self._to_list(sampled_token_ids) |
| 3691 | # Mask out the sampled tokens that should not be sampled. |
| 3692 | for i in discard_sampled_tokens_req_indices: |
| 3693 | valid_sampled_token_ids[int(i)].clear() |
| 3694 | |
| 3695 | if logprobs_tensors is not None: |
| 3696 | logprobs_lists = logprobs_tensors.tolists() |
| 3697 | else: |
| 3698 | # Includes spec decode tokens. |
| 3699 | valid_sampled_token_ids, logprobs_lists = ( |
| 3700 | RejectionSampler.parse_output( |
| 3701 | sampled_token_ids, |
| 3702 | self.input_batch.vocab_size, |
| 3703 | discard_sampled_tokens_req_indices, |
| 3704 | logprobs_tensors=logprobs_tensors, |
| 3705 | ) |
| 3706 | ) |
| 3707 | else: |
| 3708 | valid_sampled_token_ids = [] |
| 3709 | invalid_req_indices = discard_sampled_tokens_req_indices.tolist() |
| 3710 | invalid_req_indices_set = set(invalid_req_indices) |
| 3711 | |
| 3712 | # Cache the sampled tokens on the GPU and avoid CPU sync. |
| 3713 | # These will be copied into input_ids in the next step |
| 3714 | # when preparing inputs. |
| 3715 | # With spec decoding, this is done in propose_draft_token_ids(). |
| 3716 | if self.input_batch.prev_sampled_token_ids is None: |
| 3717 | assert sampled_token_ids.shape[-1] == 1 |
| 3718 | self.input_batch.prev_sampled_token_ids = sampled_token_ids |
| 3719 | self.input_batch.prev_req_id_to_index = { |
| 3720 | req_id: i |
| 3721 | for i, req_id in enumerate(self.input_batch.req_ids) |
| 3722 | if i not in invalid_req_indices_set |
| 3723 | } |
| 3724 | |
| 3725 | # Cache the sampled tokens in the model runner, so that the scheduler |
| 3726 | # doesn't need to send them back. |
| 3727 | # NOTE(woosuk): As an exception, when using PP, the scheduler sends |
| 3728 | # the sampled tokens back, because there's no direct communication |
| 3729 | # between the first-stage worker and the last-stage worker. |
| 3730 | req_ids = self.input_batch.req_ids |
| 3731 | for req_idx in range(num_sampled_tokens): |
| 3732 | if self.use_async_scheduling: |
| 3733 | sampled_ids = [-1] if req_idx not in invalid_req_indices_set else None |
| 3734 | else: |
| 3735 | sampled_ids = valid_sampled_token_ids[req_idx] |
| 3736 | |
| 3737 | num_sampled_ids: int = len(sampled_ids) if sampled_ids else 0 |
| 3738 | |
| 3739 | if not sampled_ids: |
| 3740 | continue |
| 3741 | |
| 3742 | start_idx = self.input_batch.num_tokens_no_spec[req_idx] |
| 3743 | end_idx = start_idx + num_sampled_ids |
| 3744 | assert end_idx <= self.max_model_len, ( |
| 3745 | "Sampled token IDs exceed the max model length. " |
| 3746 | f"Total number of tokens: {end_idx} > max_model_len: " |
| 3747 | f"{self.max_model_len}" |
| 3748 | ) |
| 3749 | |
| 3750 | self.input_batch.token_ids_cpu[req_idx, start_idx:end_idx] = sampled_ids |
| 3751 | self.input_batch.is_token_ids[req_idx, start_idx:end_idx] = True |
| 3752 | self.input_batch.num_tokens_no_spec[req_idx] = end_idx |
| 3753 | |
| 3754 | req_id = req_ids[req_idx] |
| 3755 | req_state = self.requests[req_id] |
| 3756 | req_state.output_token_ids.extend(sampled_ids) |
| 3757 | |
| 3758 | # Compute prompt logprobs if needed. |
| 3759 | prompt_logprobs_dict = self._get_prompt_logprobs_dict( |
| 3760 | hidden_states[:num_scheduled_tokens], |
| 3761 | scheduler_output.num_scheduled_tokens, |
| 3762 | ) |
| 3763 | |
| 3764 | return ( |
| 3765 | num_nans_in_logits, |
| 3766 | num_nans, |
| 3767 | logprobs_lists, |
| 3768 | valid_sampled_token_ids, |
| 3769 | prompt_logprobs_dict, |
| 3770 | req_ids_output_copy, |
| 3771 | req_id_to_index_output_copy, |
| 3772 | invalid_req_indices, |
| 3773 | ) |
| 3774 | |
| 3775 | @contextmanager |
| 3776 | def synchronize_input_prep(self): |
| 3777 | if self.prepare_inputs_event is None: |
| 3778 | yield |
| 3779 | return |
| 3780 | |
| 3781 | # Ensure prior step has finished with reused CPU tensors. |
| 3782 | # This is required in the async scheduling case because |
| 3783 | # the CPU->GPU transfer happens async. |
| 3784 | self.prepare_inputs_event.synchronize() |
| 3785 | try: |
| 3786 | yield |
| 3787 | finally: |
| 3788 | self.prepare_inputs_event.record() |
| 3789 | |
| 3790 | def _model_forward( |
| 3791 | self, |
| 3792 | input_ids: torch.Tensor | None = None, |
| 3793 | positions: torch.Tensor | None = None, |
| 3794 | intermediate_tensors: IntermediateTensors | None = None, |
| 3795 | inputs_embeds: torch.Tensor | None = None, |
| 3796 | **model_kwargs: dict[str, Any], |
| 3797 | ) -> Any: |
| 3798 | """Helper method to call the model forward pass. |
| 3799 | |
| 3800 | This method can be overridden by subclasses for model execution. |
| 3801 | Motivation: We can inspect only this method versus |
| 3802 | the whole execute_model, which has additional logic. |
| 3803 | |
| 3804 | Args: |
| 3805 | input_ids: Input token IDs |
| 3806 | positions: Token positions |
| 3807 | intermediate_tensors: Tensors from previous pipeline stages |
| 3808 | inputs_embeds: Input embeddings (alternative to input_ids) |
| 3809 | **model_kwargs: Additional model arguments |
| 3810 | |
| 3811 | Returns: |
| 3812 | Model output tensor |
| 3813 | |
| 3814 | """ |
| 3815 | return self.model( |
| 3816 | input_ids=input_ids, |
| 3817 | positions=positions, |
| 3818 | intermediate_tensors=intermediate_tensors, |
| 3819 | inputs_embeds=inputs_embeds, |
| 3820 | **model_kwargs, |
| 3821 | ) |
| 3822 | |
| 3823 | @staticmethod |
| 3824 | def _is_uniform_decode( |
| 3825 | max_num_scheduled_tokens: int, |
| 3826 | uniform_decode_query_len: int, |
| 3827 | num_tokens: int, |
| 3828 | num_reqs: int, |
| 3829 | force_uniform_decode: bool | None = None, |
| 3830 | ) -> bool: |
| 3831 | """Checks if it's a decode batch with same amount scheduled tokens |
| 3832 | across all requests. |
| 3833 | """ |
| 3834 | return ( |
| 3835 | ( |
| 3836 | (max_num_scheduled_tokens == uniform_decode_query_len) |
| 3837 | and (num_tokens == max_num_scheduled_tokens * num_reqs) |
| 3838 | ) |
| 3839 | if force_uniform_decode is None |
| 3840 | else force_uniform_decode |
| 3841 | ) |
| 3842 | |
| 3843 | def _allow_microbatching( |
| 3844 | self, num_reqs: int, num_scheduled_tokens_np: np.ndarray |
| 3845 | ) -> bool: |
| 3846 | """Refuse to microbatch a step that splits a prefix from its writer. |
| 3847 | |
| 3848 | A request can be admitted on a prefix cache hit against blocks another |
| 3849 | request in the same batch is only computing now. Run whole, the step |
| 3850 | issues every KV cache write before any attention read and the hit |
| 3851 | holds; split, a reader in the first half would attend over blocks the |
| 3852 | writer in the second half has not filled in yet. Vetoing on one rank |
| 3853 | settles it for all, since ranks agree on microbatching collectively. |
| 3854 | """ |
| 3855 | if not self.parallel_config.use_ubatching or num_reqs < 2: |
| 3856 | return True |
| 3857 | computed = self.input_batch.num_computed_tokens_cpu[:num_reqs] |
| 3858 | query_lens = num_scheduled_tokens_np[:num_reqs] |
| 3859 | # A decode reads nothing it was not already given in an earlier step. |
| 3860 | readers = np.flatnonzero( |
| 3861 | (query_lens > (self.reorder_batch_threshold or 1)) & (computed > 0) |
| 3862 | ) |
| 3863 | if readers.size == 0: |
| 3864 | return True |
| 3865 | |
| 3866 | for block_table in self.input_batch.block_table.block_tables: |
| 3867 | block_size = block_table.block_size |
| 3868 | table = block_table.get_numpy_array() |
| 3869 | start = computed // block_size |
| 3870 | stop = (computed + query_lens + block_size - 1) // block_size |
| 3871 | span = int((stop - start).max()) |
| 3872 | columns = start[:, None] + np.arange(span)[None, :] |
| 3873 | written = np.unique( |
| 3874 | np.take_along_axis( |
| 3875 | table[:num_reqs], |
| 3876 | np.minimum(columns, table.shape[1] - 1), |
| 3877 | axis=1, |
| 3878 | )[columns < stop[:, None]] |
| 3879 | ) |
| 3880 | for reader in readers: |
| 3881 | # Whole blocks only: a reader ends inside a partly filled one, |
| 3882 | # which is private to it and which it fills before reading. |
| 3883 | whole = computed[reader] // block_size |
| 3884 | if np.isin(table[reader, :whole], written).any(): |
| 3885 | return False |
| 3886 | return True |
| 3887 | |
| 3888 | def _determine_batch_execution_and_padding( |
| 3889 | self, |
| 3890 | num_tokens: int, |
| 3891 | num_reqs: int, |
| 3892 | num_scheduled_tokens_np: np.ndarray, |
| 3893 | max_num_scheduled_tokens: int, |
| 3894 | use_cascade_attn: bool, |
| 3895 | allow_microbatching: bool = True, |
| 3896 | force_eager: bool = False, |
| 3897 | # For cudagraph capture TODO(lucas): Refactor how we capture cudagraphs (will |
| 3898 | # be improved in model runner v2) |
| 3899 | force_uniform_decode: bool | None = None, |
| 3900 | force_has_lora: bool | None = None, |
| 3901 | force_num_active_loras: int | None = None, |
| 3902 | num_encoder_reqs: int = 0, |
| 3903 | ) -> tuple[ |
| 3904 | CUDAGraphMode, |
| 3905 | BatchDescriptor, |
| 3906 | bool, |
| 3907 | torch.Tensor | None, |
| 3908 | CUDAGraphStat | None, |
| 3909 | ]: |
| 3910 | uniform_decode = self._is_uniform_decode( |
| 3911 | max_num_scheduled_tokens=max_num_scheduled_tokens, |
| 3912 | uniform_decode_query_len=self.uniform_decode_query_len, |
| 3913 | num_tokens=num_tokens, |
| 3914 | num_reqs=num_reqs, |
| 3915 | force_uniform_decode=force_uniform_decode, |
| 3916 | ) |
| 3917 | # Encoder-decoder models only support CG for decoder_step > 0 (no enc_output |
| 3918 | # is present). Also, chunked-prefill is disabled, so batch are uniform. |
| 3919 | has_encoder_output = ( |
| 3920 | self.model_config.is_encoder_decoder and num_encoder_reqs > 0 |
| 3921 | ) |
| 3922 | |
| 3923 | # Compute LoRA state for cudagraph dispatch |
| 3924 | num_active_loras = ( |
| 3925 | force_num_active_loras |
| 3926 | if force_num_active_loras is not None |
| 3927 | else len(self.input_batch.lora_id_to_lora_request) |
| 3928 | ) |
| 3929 | has_lora = num_active_loras > 0 if force_has_lora is None else force_has_lora |
| 3930 | |
| 3931 | num_tokens_padded = self._pad_for_sequence_parallelism(num_tokens) |
| 3932 | |
| 3933 | def dispatch_cudagraph(num_tokens, disable_full=False, valid_modes=None): |
| 3934 | return self.cudagraph_dispatcher.dispatch( |
| 3935 | num_tokens=num_tokens, |
| 3936 | has_lora=has_lora, |
| 3937 | uniform_decode=uniform_decode, |
| 3938 | num_active_loras=num_active_loras, |
| 3939 | valid_modes={CUDAGraphMode.NONE} if force_eager else valid_modes, |
| 3940 | invalid_modes={CUDAGraphMode.FULL} if disable_full else None, |
| 3941 | ) |
| 3942 | |
| 3943 | cudagraph_mode, batch_descriptor = dispatch_cudagraph( |
| 3944 | num_tokens_padded, disable_full=use_cascade_attn or has_encoder_output |
| 3945 | ) |
| 3946 | num_tokens_padded = batch_descriptor.num_tokens |
| 3947 | if self.compilation_config.pass_config.enable_sp: |
| 3948 | assert ( |
| 3949 | batch_descriptor.num_tokens |
| 3950 | % self.vllm_config.parallel_config.tensor_parallel_size |
| 3951 | == 0 |
| 3952 | ), ( |
| 3953 | "Sequence parallelism requires num_tokens to be " |
| 3954 | "a multiple of tensor parallel size" |
| 3955 | ) |
| 3956 | |
| 3957 | # Extra coordination when running data-parallel since we need to coordinate |
| 3958 | # across ranks |
| 3959 | should_ubatch, num_tokens_across_dp = False, None |
| 3960 | if self.vllm_config.parallel_config.data_parallel_size > 1: |
| 3961 | should_ubatch, num_tokens_across_dp, synced_cudagraph_mode = ( |
| 3962 | coordinate_batch_across_dp( |
| 3963 | num_tokens_unpadded=num_tokens, |
| 3964 | parallel_config=self.parallel_config, |
| 3965 | allow_microbatching=allow_microbatching, |
| 3966 | num_tokens_padded=num_tokens_padded, |
| 3967 | uniform_decode=uniform_decode, |
| 3968 | cudagraph_mode=cudagraph_mode.value, |
| 3969 | ) |
| 3970 | ) |
| 3971 | |
| 3972 | # Extract DP-synced values |
| 3973 | if num_tokens_across_dp is not None: |
| 3974 | dp_rank = self.parallel_config.data_parallel_rank |
| 3975 | num_tokens_padded = int(num_tokens_across_dp[dp_rank].item()) |
| 3976 | # Re-dispatch with DP padding so we have the correct batch_descriptor |
| 3977 | cudagraph_mode, batch_descriptor = dispatch_cudagraph( |
| 3978 | num_tokens_padded, |
| 3979 | valid_modes={CUDAGraphMode(synced_cudagraph_mode)}, |
| 3980 | ) |
| 3981 | # Assert to make sure the agreed upon token count is correct otherwise |
| 3982 | # num_tokens_across_dp will no-longer be valid |
| 3983 | assert batch_descriptor.num_tokens == num_tokens_padded |
| 3984 | |
| 3985 | cudagraph_stats = None |
| 3986 | if self.vllm_config.observability_config.cudagraph_metrics: |
| 3987 | cudagraph_stats = CUDAGraphStat( |
| 3988 | num_unpadded_tokens=num_tokens, |
| 3989 | num_padded_tokens=batch_descriptor.num_tokens, |
| 3990 | num_paddings=batch_descriptor.num_tokens - num_tokens, |
| 3991 | runtime_mode=str(cudagraph_mode), |
| 3992 | ) |
| 3993 | |
| 3994 | return ( |
| 3995 | cudagraph_mode, |
| 3996 | batch_descriptor, |
| 3997 | should_ubatch, |
| 3998 | num_tokens_across_dp, |
| 3999 | cudagraph_stats, |
| 4000 | ) |
| 4001 | |
| 4002 | def _register_layerwise_nvtx_hooks(self) -> None: |
| 4003 | """Register layerwise NVTX hooks if --enable-layerwise-nvtx-tracing is enabled |
| 4004 | to trace detailed information of each layer or module in the model. |
| 4005 | """ |
| 4006 | if ( |
| 4007 | self.vllm_config.observability_config.enable_layerwise_nvtx_tracing |
| 4008 | and not self.layerwise_nvtx_hooks_registered |
| 4009 | ): |
| 4010 | if self.compilation_config.cudagraph_mode != CUDAGraphMode.NONE: |
| 4011 | logger.debug_once( |
| 4012 | "layerwise NVTX tracing is not supported when CUDA graph is " |
| 4013 | "turned off; you may observe part or all of the model " |
| 4014 | "missing NVTX markers" |
| 4015 | ) |
| 4016 | |
| 4017 | # In STOCK_TORCH_COMPILE mode, after registering hooks here, |
| 4018 | # the __call__ function of nn.module will be recompiled with |
| 4019 | # fullgraph=True. Since nvtx.range_push/pop are not traceable |
| 4020 | # by torch dynamo, we can't register hook functions here |
| 4021 | # because hook functions will also be traced by torch dynamo. |
| 4022 | if ( |
| 4023 | self.vllm_config.compilation_config.mode |
| 4024 | == CompilationMode.STOCK_TORCH_COMPILE |
| 4025 | ): |
| 4026 | logger.debug_once( |
| 4027 | "layerwise NVTX tracing is not supported when " |
| 4028 | "CompilationMode is STOCK_TORCH_COMPILE, skipping " |
| 4029 | "function hooks registration" |
| 4030 | ) |
| 4031 | else: |
| 4032 | pyt_hooks = PytHooks() |
| 4033 | pyt_hooks.register_hooks(self.model, self.model.__class__.__name__) |
| 4034 | self.layerwise_nvtx_hooks_registered = True |
| 4035 | |
| 4036 | def _get_slot_mappings( |
| 4037 | self, |
| 4038 | num_tokens_padded: int, |
| 4039 | num_reqs_padded: int, |
| 4040 | num_tokens_unpadded: int, |
| 4041 | ubatch_slices: "UBatchSlices | None" = None, |
| 4042 | ) -> tuple[ |
| 4043 | dict[int, torch.Tensor] | None, |
| 4044 | dict[str, torch.Tensor] | list[dict[str, torch.Tensor]] | None, |
| 4045 | ]: |
| 4046 | """Build slot mappings in both formats needed by the system. |
| 4047 | |
| 4048 | Args: |
| 4049 | num_tokens_padded: Total number of tokens (padded) |
| 4050 | num_reqs_padded: Total number of requests (padded) |
| 4051 | num_tokens_unpadded: Actual number of tokens (unpadded) |
| 4052 | ubatch_slices: Optional ubatch slicing info for DBO |
| 4053 | |
| 4054 | Returns: |
| 4055 | A tuple of: |
| 4056 | - slot_mappings_by_gid: dict[int, torch.Tensor] for attention metadata |
| 4057 | - slot_mappings_by_layer: dict[str, torch.Tensor] or list for ForwardContext |
| 4058 | |
| 4059 | """ |
| 4060 | if not ( |
| 4061 | hasattr(self, "kv_cache_config") |
| 4062 | and self.kv_cache_config is not None |
| 4063 | and len(self.kv_cache_config.kv_cache_groups) > 0 |
| 4064 | ): |
| 4065 | return None, None |
| 4066 | |
| 4067 | def _get_slot_mapping(kv_cache_gid: int): |
| 4068 | assert num_reqs_padded is not None and num_tokens_padded is not None |
| 4069 | kv_cache_spec = self.kv_cache_config.kv_cache_groups[ |
| 4070 | kv_cache_gid |
| 4071 | ].kv_cache_spec |
| 4072 | if isinstance(kv_cache_spec, EncoderOnlyAttentionSpec): |
| 4073 | slot_mapping = torch.zeros( |
| 4074 | (num_tokens_padded,), |
| 4075 | dtype=torch.int64, |
| 4076 | device=self.device, |
| 4077 | ) |
| 4078 | else: |
| 4079 | blk_table = self.input_batch.block_table[kv_cache_gid] |
| 4080 | slot_mapping = blk_table.slot_mapping.gpu[:num_tokens_padded] |
| 4081 | |
| 4082 | # Fill unused with -1. Needed for reshape_and_cache in full cuda |
| 4083 | # graph mode. `blk_table_tensor` -1 to match mamba PAD_SLOT_ID |
| 4084 | slot_mapping[num_tokens_unpadded:num_tokens_padded].fill_(-1) |
| 4085 | |
| 4086 | return slot_mapping |
| 4087 | |
| 4088 | slot_mappings_by_gid = { |
| 4089 | gid: _get_slot_mapping(gid) |
| 4090 | for gid, _ in enumerate(self.kv_cache_config.kv_cache_groups) |
| 4091 | } |
| 4092 | |
| 4093 | slot_mappings_by_layer: dict[str, torch.Tensor] = {} |
| 4094 | for gid, kv_cache_group in enumerate(self.kv_cache_config.kv_cache_groups): |
| 4095 | slot_mapping = slot_mappings_by_gid[gid] |
| 4096 | for layer_name in kv_cache_group.layer_names: |
| 4097 | slot_mappings_by_layer[layer_name] = slot_mapping |
| 4098 | |
| 4099 | if ubatch_slices is not None: |
| 4100 | result: list[dict[str, torch.Tensor]] = [] |
| 4101 | for ubatch in ubatch_slices: |
| 4102 | sliced_mappings: dict[str, torch.Tensor] = {} |
| 4103 | for layer_name, slot_mapping in slot_mappings_by_layer.items(): |
| 4104 | sliced_mappings[layer_name] = slot_mapping[ubatch.token_slice] |
| 4105 | result.append(sliced_mappings) |
| 4106 | return slot_mappings_by_gid, result |
| 4107 | |
| 4108 | return slot_mappings_by_gid, slot_mappings_by_layer |
| 4109 | |
| 4110 | def _is_all_reqs_chunked_prefill(self) -> bool: |
| 4111 | """Check if all scheduled requests are marked to discard sampled tokens. |
| 4112 | |
| 4113 | This is true when `discard_request_mask` is set for every scheduled |
| 4114 | request (e.g., for chunked prefill requests that are not the last |
| 4115 | prefill chunk).""" |
| 4116 | num_reqs = self.input_batch.num_reqs |
| 4117 | return bool(self.discard_request_mask.np[:num_reqs].all()) |
| 4118 | |
| 4119 | @torch.inference_mode() |
| 4120 | def execute_model( |
| 4121 | self, |
| 4122 | scheduler_output: "SchedulerOutput", |
| 4123 | intermediate_tensors: IntermediateTensors | None = None, |
| 4124 | ) -> ModelRunnerOutput | AsyncModelRunnerOutput | IntermediateTensors | None: |
| 4125 | if self.execute_model_state is not None: |
| 4126 | raise RuntimeError( |
| 4127 | "State error: sample_tokens() must be called " |
| 4128 | "after execute_model() returns None." |
| 4129 | ) |
| 4130 | |
| 4131 | # If ngram_gpu is used, we need to copy the scheduler_output to avoid |
| 4132 | # the modification has influence on the scheduler_output in engine core process. |
| 4133 | # The replace is much faster than deepcopy. |
| 4134 | if ( |
| 4135 | self.speculative_config is not None |
| 4136 | and self.speculative_config.use_ngram_gpu() |
| 4137 | ): |
| 4138 | num_scheduled_tokens_copy = scheduler_output.num_scheduled_tokens.copy() |
| 4139 | spec_decode_tokens_copy = ( |
| 4140 | scheduler_output.scheduled_spec_decode_tokens.copy() |
| 4141 | ) |
| 4142 | scheduler_output = replace( |
| 4143 | scheduler_output, |
| 4144 | num_scheduled_tokens=num_scheduled_tokens_copy, |
| 4145 | scheduled_spec_decode_tokens=spec_decode_tokens_copy, |
| 4146 | ) |
| 4147 | |
| 4148 | if has_kv_transfer_group(): |
| 4149 | kv_connector_metadata = scheduler_output.kv_connector_metadata |
| 4150 | assert kv_connector_metadata is not None |
| 4151 | get_kv_transfer_group().handle_preemptions(kv_connector_metadata) |
| 4152 | |
| 4153 | num_scheduled_tokens = scheduler_output.total_num_scheduled_tokens |
| 4154 | with ( |
| 4155 | record_function_or_nullcontext("gpu_model_runner: preprocess"), |
| 4156 | self.synchronize_input_prep(), |
| 4157 | ): |
| 4158 | # Update persistent batch states. |
| 4159 | deferred_state_corrections_fn = self._update_states(scheduler_output) |
| 4160 | |
| 4161 | if has_ec_transfer() and not get_ec_transfer().is_consumer: |
| 4162 | with self.maybe_get_ec_connector_output( |
| 4163 | scheduler_output, |
| 4164 | encoder_cache=self.encoder_cache, |
| 4165 | ) as ec_connector_output: |
| 4166 | self._execute_mm_encoder(scheduler_output) |
| 4167 | return make_empty_encoder_model_runner_output(scheduler_output) |
| 4168 | |
| 4169 | if not num_scheduled_tokens: |
| 4170 | if ( |
| 4171 | self.parallel_config.distributed_executor_backend |
| 4172 | == "external_launcher" |
| 4173 | and self.parallel_config.data_parallel_size > 1 |
| 4174 | ): |
| 4175 | # this is a corner case when both external launcher |
| 4176 | # and DP are enabled, num_scheduled_tokens could be |
| 4177 | # 0, and has_unfinished_requests in the outer loop |
| 4178 | # returns True. before returning early here we call |
| 4179 | # dummy run to ensure coordinate_batch_across_dp |
| 4180 | # is called into to avoid out of sync issues. |
| 4181 | self._dummy_run(1) |
| 4182 | if not has_kv_transfer_group(): |
| 4183 | # Return empty ModelRunnerOutput if no work to do. |
| 4184 | return EMPTY_MODEL_RUNNER_OUTPUT |
| 4185 | return self.kv_connector_no_forward(scheduler_output, self.vllm_config) |
| 4186 | |
| 4187 | if self.cache_config.kv_sharing_fast_prefill: |
| 4188 | assert not self.num_prompt_logprobs, ( |
| 4189 | "--kv-sharing-fast-prefill produces incorrect " |
| 4190 | "logprobs for prompt tokens, tokens, please disable " |
| 4191 | "it when the requests need prompt logprobs" |
| 4192 | ) |
| 4193 | |
| 4194 | num_reqs = self.input_batch.num_reqs |
| 4195 | req_ids = self.input_batch.req_ids |
| 4196 | tokens = [scheduler_output.num_scheduled_tokens[i] for i in req_ids] |
| 4197 | num_scheduled_tokens_np = np.array(tokens, dtype=np.int32) |
| 4198 | max_num_scheduled_tokens = int(num_scheduled_tokens_np.max()) |
| 4199 | num_tokens_unpadded = scheduler_output.total_num_scheduled_tokens |
| 4200 | |
| 4201 | logits_indices, spec_decode_metadata, max_num_sampled_tokens = ( |
| 4202 | self._prepare_inputs(scheduler_output, num_scheduled_tokens_np) |
| 4203 | ) |
| 4204 | |
| 4205 | cascade_attn_prefix_lens = None |
| 4206 | # Disable cascade attention when using microbatching (DBO) |
| 4207 | if self.cascade_attn_enabled and not self.parallel_config.use_ubatching: |
| 4208 | # Pre-compute cascade attention prefix lengths |
| 4209 | cascade_attn_prefix_lens = self._compute_cascade_attn_prefix_lens( |
| 4210 | num_scheduled_tokens_np, |
| 4211 | self.input_batch.num_computed_tokens_cpu[:num_reqs], |
| 4212 | scheduler_output.num_common_prefix_blocks, |
| 4213 | ) |
| 4214 | |
| 4215 | ( |
| 4216 | cudagraph_mode, |
| 4217 | batch_desc, |
| 4218 | should_ubatch, |
| 4219 | num_tokens_across_dp, |
| 4220 | cudagraph_stats, |
| 4221 | ) = self._determine_batch_execution_and_padding( |
| 4222 | num_tokens=num_tokens_unpadded, |
| 4223 | num_reqs=num_reqs, |
| 4224 | num_scheduled_tokens_np=num_scheduled_tokens_np, |
| 4225 | max_num_scheduled_tokens=max_num_scheduled_tokens, |
| 4226 | use_cascade_attn=cascade_attn_prefix_lens is not None, |
| 4227 | num_encoder_reqs=len(scheduler_output.scheduled_encoder_inputs), |
| 4228 | allow_microbatching=self._allow_microbatching( |
| 4229 | num_reqs, num_scheduled_tokens_np |
| 4230 | ), |
| 4231 | ) |
| 4232 | |
| 4233 | logger.debug( |
| 4234 | "Running batch with cudagraph_mode: %s, batch_descriptor: %s, " |
| 4235 | "should_ubatch: %s, num_tokens_across_dp: %s", |
| 4236 | cudagraph_mode, |
| 4237 | batch_desc, |
| 4238 | should_ubatch, |
| 4239 | num_tokens_across_dp, |
| 4240 | ) |
| 4241 | |
| 4242 | num_tokens_padded = batch_desc.num_tokens |
| 4243 | num_reqs_padded = ( |
| 4244 | batch_desc.num_reqs if batch_desc.num_reqs is not None else num_reqs |
| 4245 | ) |
| 4246 | ubatch_slices, ubatch_slices_padded = maybe_create_ubatch_slices( |
| 4247 | should_ubatch, |
| 4248 | num_scheduled_tokens_np, |
| 4249 | num_tokens_padded, |
| 4250 | num_reqs_padded, |
| 4251 | self.parallel_config.num_ubatches, |
| 4252 | ) |
| 4253 | |
| 4254 | logger.debug( |
| 4255 | "ubatch_slices: %s, ubatch_slices_padded: %s", |
| 4256 | ubatch_slices, |
| 4257 | ubatch_slices_padded, |
| 4258 | ) |
| 4259 | |
| 4260 | # True if any attention backend handles KV cache update separately |
| 4261 | # from forward() (i.e., forward_includes_kv_cache_update=False). When true, |
| 4262 | # slot_mappings must use padded dimensions to match the key/value tensors. |
| 4263 | has_separate_kv_update = not all( |
| 4264 | all( |
| 4265 | g.backend.forward_includes_kv_cache_update |
| 4266 | for g in self.attn_groups[id] |
| 4267 | ) |
| 4268 | for id, spec in enumerate(self.kv_cache_config.kv_cache_groups) |
| 4269 | if not isinstance(spec.kv_cache_spec, EncoderOnlyAttentionSpec) |
| 4270 | ) |
| 4271 | pad_attn = cudagraph_mode == CUDAGraphMode.FULL |
| 4272 | |
| 4273 | if self.cache_config.mamba_cache_mode == "align": |
| 4274 | # preprocess_mamba reads req_state.num_computed_tokens (CPU) |
| 4275 | # to decide copy operations, so we must apply deferred |
| 4276 | # corrections before it runs. |
| 4277 | if deferred_state_corrections_fn: |
| 4278 | deferred_state_corrections_fn() |
| 4279 | deferred_state_corrections_fn = None |
| 4280 | mamba_bufs = self._get_mamba_bufs() |
| 4281 | mamba_utils.preprocess_mamba( |
| 4282 | scheduler_output, |
| 4283 | self.kv_cache_config, |
| 4284 | self.cache_config, |
| 4285 | self.mamba_state_idx, |
| 4286 | self.input_batch, |
| 4287 | self.requests, |
| 4288 | self.compilation_config.static_forward_context, |
| 4289 | self._get_mamba_state_copy_funcs(), |
| 4290 | mamba_bufs.preprocess, |
| 4291 | align_ctx=mamba_bufs.postprocess_align, |
| 4292 | ) |
| 4293 | # preprocess_mamba resets num_accepted_tokens_cpu to 1 |
| 4294 | # for requests whose state was copied to a new block. |
| 4295 | # Re-sync to GPU so the mamba kernel reads from the |
| 4296 | # correct initial state slot (init_token_idx = 0). |
| 4297 | self.num_accepted_tokens.np[:num_reqs] = ( |
| 4298 | self.input_batch.num_accepted_tokens_cpu[:num_reqs] |
| 4299 | ) |
| 4300 | self.num_accepted_tokens.copy_to_gpu(num_reqs) |
| 4301 | |
| 4302 | # Stage per-request inputs for the fused postprocess kernel |
| 4303 | # only when that kernel will actually run. The kernel is |
| 4304 | # gated on spec-decode + hybrid (see MambaBuffers.create); |
| 4305 | # without it, ``mamba_bufs.postprocess_align`` is None and |
| 4306 | # the staging buffers don't exist. |
| 4307 | if mamba_bufs.postprocess_align is not None: |
| 4308 | mamba_utils.stage_postprocess_inputs_to_gpu( |
| 4309 | mamba_bufs.postprocess_align, |
| 4310 | scheduler_output, |
| 4311 | self.input_batch.req_ids, |
| 4312 | num_reqs, |
| 4313 | self.requests, |
| 4314 | self.mamba_state_idx, |
| 4315 | ) |
| 4316 | |
| 4317 | use_spec_decode = len(scheduler_output.scheduled_spec_decode_tokens) > 0 |
| 4318 | ubatch_slices_attn = ubatch_slices_padded if pad_attn else ubatch_slices |
| 4319 | |
| 4320 | slot_mappings_by_group, slot_mappings = self._get_slot_mappings( |
| 4321 | num_tokens_padded=num_tokens_padded |
| 4322 | if pad_attn or has_separate_kv_update |
| 4323 | else num_tokens_unpadded, |
| 4324 | num_reqs_padded=( |
| 4325 | num_reqs_padded if pad_attn or has_separate_kv_update else num_reqs |
| 4326 | ), |
| 4327 | num_tokens_unpadded=num_tokens_unpadded, |
| 4328 | ubatch_slices=ubatch_slices_padded, |
| 4329 | ) |
| 4330 | |
| 4331 | attn_metadata, spec_decode_common_attn_metadata = ( |
| 4332 | self._build_attention_metadata( |
| 4333 | num_tokens=num_tokens_unpadded, |
| 4334 | num_tokens_padded=num_tokens_padded if pad_attn else None, |
| 4335 | num_reqs=num_reqs, |
| 4336 | num_reqs_padded=num_reqs_padded if pad_attn else None, |
| 4337 | max_query_len=max_num_scheduled_tokens, |
| 4338 | ubatch_slices=ubatch_slices_attn, |
| 4339 | logits_indices=logits_indices, |
| 4340 | max_num_sampled_tokens=max_num_sampled_tokens, |
| 4341 | use_spec_decode=use_spec_decode, |
| 4342 | num_scheduled_tokens=scheduler_output.num_scheduled_tokens, |
| 4343 | cascade_attn_prefix_lens=cascade_attn_prefix_lens, |
| 4344 | slot_mappings=slot_mappings_by_group, |
| 4345 | ) |
| 4346 | ) |
| 4347 | |
| 4348 | ( |
| 4349 | input_ids, |
| 4350 | inputs_embeds, |
| 4351 | positions, |
| 4352 | intermediate_tensors, |
| 4353 | model_kwargs, |
| 4354 | ec_connector_output, |
| 4355 | ) = self._preprocess( |
| 4356 | scheduler_output, num_tokens_padded, intermediate_tensors |
| 4357 | ) |
| 4358 | |
| 4359 | # Encoder-decoder models can only compile the pure decode steps where no |
| 4360 | # encoder inputs are present. Use eager for the first pass. |
| 4361 | num_encoder_reqs = len(scheduler_output.scheduled_encoder_inputs) |
| 4362 | has_encoder_input = ( |
| 4363 | self.model_config.is_encoder_decoder and num_encoder_reqs > 0 |
| 4364 | ) |
| 4365 | |
| 4366 | # Run the model. |
| 4367 | # Use persistent buffers for CUDA graphs. |
| 4368 | # When spec decode is enabled, defer connector finalization |
| 4369 | # (wait_for_save + clear metadata) until after draft model runs. |
| 4370 | defer_kv_connector_finalize = self.speculative_config is not None |
| 4371 | # Update the EPLB meta. |
| 4372 | if self.eplb_state is not None: |
| 4373 | self.eplb_state.prepare_forward( |
| 4374 | self.model_config, |
| 4375 | num_tokens_unpadded, |
| 4376 | ubatch_slices_padded, |
| 4377 | ) |
| 4378 | is_padding = self._prepare_padding_mask(num_tokens_unpadded, num_tokens_padded) |
| 4379 | with ( |
| 4380 | set_forward_context( |
| 4381 | attn_metadata, |
| 4382 | self.vllm_config, |
| 4383 | num_tokens=num_tokens_padded, |
| 4384 | num_tokens_across_dp=num_tokens_across_dp, |
| 4385 | cudagraph_runtime_mode=cudagraph_mode, |
| 4386 | batch_descriptor=batch_desc, |
| 4387 | ubatch_slices=ubatch_slices_padded, |
| 4388 | slot_mapping=slot_mappings, |
| 4389 | skip_compiled=has_encoder_input, |
| 4390 | is_padding=is_padding, |
| 4391 | ), |
| 4392 | record_function_or_nullcontext("gpu_model_runner: forward"), |
| 4393 | self.maybe_get_kv_connector_output( |
| 4394 | scheduler_output, |
| 4395 | defer_finalize=defer_kv_connector_finalize, |
| 4396 | ) as kv_connector_output, |
| 4397 | ): |
| 4398 | model_output = self._model_forward( |
| 4399 | input_ids=input_ids, |
| 4400 | positions=positions, |
| 4401 | intermediate_tensors=intermediate_tensors, |
| 4402 | inputs_embeds=inputs_embeds, |
| 4403 | **model_kwargs, |
| 4404 | ) |
| 4405 | |
| 4406 | with record_function_or_nullcontext("gpu_model_runner: postprocess"): |
| 4407 | if self.use_aux_hidden_state_outputs: |
| 4408 | # True when EAGLE 3 is used. |
| 4409 | hidden_states, aux_hidden_states = model_output |
| 4410 | else: |
| 4411 | # Common case. |
| 4412 | hidden_states = model_output |
| 4413 | aux_hidden_states = None |
| 4414 | |
| 4415 | if not self.broadcast_pp_output: |
| 4416 | # Common case. |
| 4417 | if not get_pp_group().is_last_rank: |
| 4418 | # Return the intermediate tensors. |
| 4419 | assert isinstance(hidden_states, IntermediateTensors) |
| 4420 | self.kv_connector_output = kv_connector_output |
| 4421 | return hidden_states |
| 4422 | |
| 4423 | if self.is_pooling_model: |
| 4424 | # Return the pooling output. |
| 4425 | return self._pool( |
| 4426 | hidden_states, |
| 4427 | num_scheduled_tokens, |
| 4428 | num_scheduled_tokens_np, |
| 4429 | kv_connector_output, |
| 4430 | ) |
| 4431 | |
| 4432 | sample_hidden_states = hidden_states[logits_indices] |
| 4433 | logits = self.model.compute_logits(sample_hidden_states) |
| 4434 | else: |
| 4435 | # Rare case. |
| 4436 | assert not self.is_pooling_model |
| 4437 | |
| 4438 | sample_hidden_states = hidden_states[logits_indices] |
| 4439 | if not get_pp_group().is_last_rank: |
| 4440 | all_gather_tensors = { |
| 4441 | "residual": not is_residual_scattered_for_sp( |
| 4442 | self.vllm_config, num_tokens_padded |
| 4443 | ) |
| 4444 | } |
| 4445 | get_pp_group().send_tensor_dict( |
| 4446 | hidden_states.tensors, |
| 4447 | all_gather_group=get_tp_group(), |
| 4448 | all_gather_tensors=all_gather_tensors, |
| 4449 | ) |
| 4450 | logits = None |
| 4451 | else: |
| 4452 | logits = self.model.compute_logits(sample_hidden_states) |
| 4453 | |
| 4454 | model_output_broadcast_data: dict[str, Any] = {} |
| 4455 | if logits is not None: |
| 4456 | model_output_broadcast_data["logits"] = logits.contiguous() |
| 4457 | |
| 4458 | broadcasted = get_pp_group().broadcast_tensor_dict( |
| 4459 | model_output_broadcast_data, src=len(get_pp_group().ranks) - 1 |
| 4460 | ) |
| 4461 | assert broadcasted is not None |
| 4462 | logits = broadcasted["logits"] |
| 4463 | |
| 4464 | self.execute_model_state = ExecuteModelState( |
| 4465 | scheduler_output, |
| 4466 | logits, |
| 4467 | spec_decode_metadata, |
| 4468 | spec_decode_common_attn_metadata, |
| 4469 | hidden_states, |
| 4470 | sample_hidden_states, |
| 4471 | aux_hidden_states, |
| 4472 | ec_connector_output, |
| 4473 | cudagraph_stats, |
| 4474 | slot_mappings, |
| 4475 | ) |
| 4476 | self.kv_connector_output = kv_connector_output |
| 4477 | |
| 4478 | # Now the batch has been launched we can wait for corrections from the |
| 4479 | # previous model forward without breaking async scheduling. |
| 4480 | if deferred_state_corrections_fn: |
| 4481 | deferred_state_corrections_fn() |
| 4482 | |
| 4483 | return None |
| 4484 | |
| 4485 | def _input_fits_in_drafter( |
| 4486 | self, common_attn_metadata: CommonAttentionMetadata | None |
| 4487 | ) -> bool: |
| 4488 | if common_attn_metadata is None: |
| 4489 | return False |
| 4490 | assert self.speculative_config is not None |
| 4491 | # DFlash queries one extra token (the bonus token) beyond num_spec_tokens |
| 4492 | num_drafter_query_tokens = self.num_spec_tokens + ( |
| 4493 | 1 if self.speculative_config.use_dflash() else 0 |
| 4494 | ) |
| 4495 | return ( |
| 4496 | common_attn_metadata.max_seq_len + num_drafter_query_tokens |
| 4497 | <= self.effective_drafter_max_model_len |
| 4498 | ) |
| 4499 | |
| 4500 | @torch.inference_mode |
| 4501 | def sample_tokens( |
| 4502 | self, grammar_output: "GrammarOutput | None" |
| 4503 | ) -> ModelRunnerOutput | AsyncModelRunnerOutput | IntermediateTensors: |
| 4504 | if self.execute_model_state is None: |
| 4505 | kv_connector_output = self.kv_connector_output |
| 4506 | self.kv_connector_output = None |
| 4507 | # receive sampled token ids from the last PP rank. |
| 4508 | if self.use_async_scheduling and not get_pp_group().is_last_rank: |
| 4509 | self._pp_receive_prev_sampled_token_ids_to_input_batch() |
| 4510 | # In case of PP with kv transfer, we need to pass through the |
| 4511 | # kv_connector_output |
| 4512 | return ModelRunnerOutput.with_kv_conn_output_only(kv_connector_output) |
| 4513 | |
| 4514 | # Unpack ephemeral state. |
| 4515 | ( |
| 4516 | scheduler_output, |
| 4517 | logits, |
| 4518 | spec_decode_metadata, |
| 4519 | spec_decode_common_attn_metadata, |
| 4520 | hidden_states, |
| 4521 | sample_hidden_states, |
| 4522 | aux_hidden_states, |
| 4523 | ec_connector_output, |
| 4524 | cudagraph_stats, |
| 4525 | slot_mappings, |
| 4526 | ) = self.execute_model_state |
| 4527 | # Clear ephemeral state. |
| 4528 | self.execute_model_state = None |
| 4529 | |
| 4530 | # Apply structured output bitmasks if present. |
| 4531 | if grammar_output is not None: |
| 4532 | apply_grammar_bitmask( |
| 4533 | scheduler_output, grammar_output, self.input_batch, logits |
| 4534 | ) |
| 4535 | |
| 4536 | with record_function_or_nullcontext("gpu_model_runner: sample"): |
| 4537 | sampler_output = self._sample(logits, spec_decode_metadata) |
| 4538 | |
| 4539 | self._update_states_after_model_execute( |
| 4540 | sampler_output.sampled_token_ids, scheduler_output |
| 4541 | ) |
| 4542 | if self.use_async_scheduling: |
| 4543 | pp = get_pp_group() |
| 4544 | # For torchrun external_launcher PP mode with broadcast_pp_output=True, |
| 4545 | # PP outputs have been broadcasted to all ranks at logits computation. |
| 4546 | # Therefore, here is no need to send sampled token ids again in this case. |
| 4547 | if not self.broadcast_pp_output and pp.world_size > 1 and pp.is_last_rank: |
| 4548 | self._pp_broadcast_prev_sampled_token_ids( |
| 4549 | sampler_output.sampled_token_ids |
| 4550 | ) |
| 4551 | |
| 4552 | self._draft_token_ids = None |
| 4553 | self._draft_probs = None |
| 4554 | self._draft_prob_req_ids = None |
| 4555 | self._draft_token_req_ids = None |
| 4556 | self.valid_sampled_token_count_gpu = None |
| 4557 | self.input_batch.prev_sampled_token_ids = None |
| 4558 | |
| 4559 | def propose_draft_token_ids(sampled_token_ids): |
| 4560 | assert spec_decode_common_attn_metadata is not None |
| 4561 | with record_function_or_nullcontext("gpu_model_runner: draft"): |
| 4562 | self._draft_token_ids = self.propose_draft_token_ids( |
| 4563 | scheduler_output, |
| 4564 | sampled_token_ids, |
| 4565 | self.input_batch.sampling_metadata, |
| 4566 | hidden_states, |
| 4567 | sample_hidden_states, |
| 4568 | aux_hidden_states, |
| 4569 | spec_decode_metadata, |
| 4570 | spec_decode_common_attn_metadata, |
| 4571 | slot_mappings, |
| 4572 | ) |
| 4573 | self._copy_draft_token_ids_to_cpu(scheduler_output) |
| 4574 | |
| 4575 | spec_config = self.speculative_config |
| 4576 | draft_after_bookkeeping = False |
| 4577 | if spec_config is not None: |
| 4578 | # Decide whether to run the drafter or zero out draft tokens. |
| 4579 | input_fits_in_drafter = self._input_fits_in_drafter( |
| 4580 | spec_decode_common_attn_metadata |
| 4581 | ) |
| 4582 | # Whether the drafter runs a GPU model forward (and thus carries |
| 4583 | # TP/EP/DP collectives), independent of padded-batch timing. |
| 4584 | drafter_runs_model_forward = ( |
| 4585 | spec_config.use_eagle() |
| 4586 | or spec_config.uses_draft_model() |
| 4587 | or spec_config.uses_extract_hidden_states() |
| 4588 | ) |
| 4589 | use_gpu_toks = ( |
| 4590 | drafter_runs_model_forward |
| 4591 | and not spec_config.disable_padded_drafter_batch |
| 4592 | ) |
| 4593 | if use_gpu_toks: |
| 4594 | # EAGLE/DraftModel speculative decoding can use the GPU sampled tokens |
| 4595 | # as inputs, and does not need to wait for bookkeeping to finish. |
| 4596 | assert isinstance( |
| 4597 | self.drafter, |
| 4598 | EagleProposer |
| 4599 | | DFlashProposer |
| 4600 | | DraftModelProposer |
| 4601 | | ExtractHiddenStatesProposer |
| 4602 | | Gemma4Proposer, |
| 4603 | ) |
| 4604 | sampled_token_ids = sampler_output.sampled_token_ids |
| 4605 | if input_fits_in_drafter: |
| 4606 | propose_draft_token_ids(sampled_token_ids) |
| 4607 | else: |
| 4608 | if self.valid_sampled_token_count_event is not None: |
| 4609 | assert spec_decode_common_attn_metadata is not None |
| 4610 | next_token_ids, valid_sampled_tokens_count = ( |
| 4611 | self.drafter.prepare_next_token_ids_padded( |
| 4612 | sampled_token_ids, |
| 4613 | self.requests, |
| 4614 | self.input_batch, |
| 4615 | self.discard_request_mask.gpu, |
| 4616 | ) |
| 4617 | ) |
| 4618 | self._copy_valid_sampled_token_count( |
| 4619 | next_token_ids, valid_sampled_tokens_count |
| 4620 | ) |
| 4621 | if self.parallel_config.data_parallel_size > 1: |
| 4622 | # Prevent hang when DP ranks disagree on input_fits_in_drafter |
| 4623 | self.drafter.dummy_run(num_tokens=1) |
| 4624 | elif ( |
| 4625 | spec_config.use_ngram_gpu() |
| 4626 | and not spec_config.disable_padded_drafter_batch |
| 4627 | ): |
| 4628 | assert isinstance(self.drafter, NgramProposerGPU) |
| 4629 | sampled_token_ids = sampler_output.sampled_token_ids |
| 4630 | if input_fits_in_drafter: |
| 4631 | propose_draft_token_ids(sampled_token_ids) |
| 4632 | elif self.valid_sampled_token_count_event is not None: |
| 4633 | assert spec_decode_common_attn_metadata is not None |
| 4634 | next_token_ids, valid_sampled_tokens_count, _ = ( |
| 4635 | self.drafter.update_token_ids_ngram( |
| 4636 | sampled_token_ids, |
| 4637 | self.input_batch, |
| 4638 | self.token_ids_gpu_tensor, |
| 4639 | self.num_tokens_no_spec_gpu, |
| 4640 | self.discard_request_mask.gpu, |
| 4641 | ) |
| 4642 | ) |
| 4643 | self._copy_valid_sampled_token_count( |
| 4644 | next_token_ids, valid_sampled_tokens_count |
| 4645 | ) |
| 4646 | else: |
| 4647 | # These drafters consume CPU sampled tokens, so they run |
| 4648 | # after bookkeeping. |
| 4649 | draft_after_bookkeeping = True |
| 4650 | |
| 4651 | if not input_fits_in_drafter: |
| 4652 | # Zero out draft tokens so the scheduler doesn't schedule |
| 4653 | # stale drafts from the previous step. |
| 4654 | # For Nemotron-H: it is necessary to zero out the draft tokens, |
| 4655 | # otherwise the stale tokens will corrupt Mamba recurrent |
| 4656 | # state and logprobs for sequences near max_model_len. |
| 4657 | self._draft_token_ids = torch.zeros( |
| 4658 | 1, device=self.device, dtype=torch.int32 |
| 4659 | ).expand(len(self.input_batch.req_ids), self.num_spec_tokens) |
| 4660 | self._draft_probs = None |
| 4661 | self._draft_prob_req_ids = None |
| 4662 | self._copy_draft_token_ids_to_cpu(scheduler_output, zeros_only=True) |
| 4663 | |
| 4664 | with record_function_or_nullcontext("gpu_model_runner: bookkeep"): |
| 4665 | ( |
| 4666 | num_nans_in_logits, |
| 4667 | num_nans_device, |
| 4668 | logprobs_lists, |
| 4669 | valid_sampled_token_ids, |
| 4670 | prompt_logprobs_dict, |
| 4671 | req_ids_output_copy, |
| 4672 | req_id_to_index_output_copy, |
| 4673 | invalid_req_indices, |
| 4674 | ) = self._bookkeeping_sync( |
| 4675 | scheduler_output, |
| 4676 | sampler_output, |
| 4677 | logits, |
| 4678 | hidden_states, |
| 4679 | scheduler_output.total_num_scheduled_tokens, |
| 4680 | ) |
| 4681 | |
| 4682 | if draft_after_bookkeeping: |
| 4683 | # ngram and other speculative decoding methods use the sampled |
| 4684 | # tokens on the CPU, so they are run after bookkeeping. |
| 4685 | if input_fits_in_drafter: |
| 4686 | propose_draft_token_ids(valid_sampled_token_ids) |
| 4687 | elif ( |
| 4688 | drafter_runs_model_forward |
| 4689 | and self.parallel_config.data_parallel_size > 1 |
| 4690 | ): |
| 4691 | # Prevent hang when DP ranks disagree on input_fits_in_drafter |
| 4692 | assert isinstance( |
| 4693 | self.drafter, |
| 4694 | EagleProposer |
| 4695 | | DFlashProposer |
| 4696 | | DraftModelProposer |
| 4697 | | ExtractHiddenStatesProposer |
| 4698 | | Gemma4Proposer, |
| 4699 | ) |
| 4700 | self.drafter.dummy_run(num_tokens=1) |
| 4701 | |
| 4702 | # Finalize KV connector (wait_for_save + clear metadata) after |
| 4703 | # draft model runs. Deferred from target model forward to allow |
| 4704 | # draft model to also save its KV cache. |
| 4705 | if spec_config is not None: |
| 4706 | self.finalize_kv_connector() |
| 4707 | |
| 4708 | with record_function_or_nullcontext("gpu_model_runner: eplb"): |
| 4709 | self.eplb_step() |
| 4710 | |
| 4711 | # self.kv_connector_output may be modified during drafting |
| 4712 | kv_connector_output = self.kv_connector_output |
| 4713 | self.kv_connector_output = None |
| 4714 | |
| 4715 | with record_function_or_nullcontext("gpu_model_runner: ModelRunnerOutput"): |
| 4716 | output = ModelRunnerOutput( |
| 4717 | req_ids=req_ids_output_copy, |
| 4718 | req_id_to_index=req_id_to_index_output_copy, |
| 4719 | sampled_token_ids=valid_sampled_token_ids, |
| 4720 | logprobs=logprobs_lists, |
| 4721 | prompt_logprobs_dict=prompt_logprobs_dict, |
| 4722 | kv_connector_output=kv_connector_output, |
| 4723 | ec_connector_output=ec_connector_output |
| 4724 | if self.supports_mm_inputs |
| 4725 | else None, |
| 4726 | num_nans_in_logits=num_nans_in_logits, |
| 4727 | cudagraph_stats=cudagraph_stats, |
| 4728 | ) |
| 4729 | |
| 4730 | if not self.use_async_scheduling: |
| 4731 | return output |
| 4732 | |
| 4733 | with record_function_or_nullcontext( |
| 4734 | "gpu_model_runner: AsyncGPUModelRunnerOutput" |
| 4735 | ): |
| 4736 | async_output = AsyncGPUModelRunnerOutput( |
| 4737 | model_runner_output=output, |
| 4738 | sampled_token_ids=sampler_output.sampled_token_ids, |
| 4739 | logprobs_tensors=sampler_output.logprobs_tensors, |
| 4740 | invalid_req_indices=invalid_req_indices, |
| 4741 | async_output_copy_stream=self._get_or_create_async_output_copy_stream(), |
| 4742 | vocab_size=self.input_batch.vocab_size, |
| 4743 | check_ep_fault=self.check_ep_fault, |
| 4744 | num_nans=num_nans_device, |
| 4745 | ) |
| 4746 | with record_function_or_nullcontext( |
| 4747 | "gpu_model_runner: set_async_sampled_token_ids" |
| 4748 | ): |
| 4749 | # Save ref of sampled_token_ids CPU tensor if the batch contains |
| 4750 | # any requests with sampling params that require output ids. |
| 4751 | self.input_batch.set_async_sampled_token_ids( |
| 4752 | async_output.sampled_token_ids_cpu, |
| 4753 | async_output.async_copy_ready_event, |
| 4754 | ) |
| 4755 | |
| 4756 | return async_output |
| 4757 | |
| 4758 | def _pp_broadcast_prev_sampled_token_ids( |
| 4759 | self, sampled_token_ids: torch.Tensor |
| 4760 | ) -> None: |
| 4761 | """Broadcast sampled token ids (GPU) from last PP stage""" |
| 4762 | pp = get_pp_group() |
| 4763 | assert pp.is_last_rank |
| 4764 | # `prev_sampled_token_ids` is expected to have shape [num_reqs, 1]. |
| 4765 | assert sampled_token_ids.dim() == 2 and sampled_token_ids.shape[-1] == 1, ( |
| 4766 | "PP+async expects sampled_token_ids to have shape [num_reqs, 1]" |
| 4767 | ) |
| 4768 | # Skip for chunked prefill: sampled tokens are dummy |
| 4769 | # and will be discarded, no need to broadcast. |
| 4770 | if not self._is_all_reqs_chunked_prefill(): |
| 4771 | torch.distributed.broadcast( |
| 4772 | sampled_token_ids, src=pp.rank, group=pp.device_group |
| 4773 | ) |
| 4774 | |
| 4775 | def _pp_receive_prev_sampled_token_ids_to_input_batch(self) -> None: |
| 4776 | """Receive sampled token ids broadcast from last PP stage""" |
| 4777 | pp = get_pp_group() |
| 4778 | assert not pp.is_last_rank |
| 4779 | num_reqs = self.input_batch.num_reqs |
| 4780 | # `prev_sampled_token_ids` is expected to have shape [num_reqs, 1]. |
| 4781 | recv = torch.empty((num_reqs, 1), dtype=torch.int32, device=self.device) |
| 4782 | # skip for chunked prefill. |
| 4783 | if not self._is_all_reqs_chunked_prefill(): |
| 4784 | self._pp_recv_work = torch.distributed.broadcast( |
| 4785 | recv, src=pp.last_rank, group=pp.device_group, async_op=True |
| 4786 | ) |
| 4787 | self.input_batch.prev_sampled_token_ids = recv |
| 4788 | |
| 4789 | # construct `prev_req_id_to_index` here so `_prepare_input_ids` |
| 4790 | # can map req_id -> previous batch row |
| 4791 | discard_req_indices = np.nonzero(self.discard_request_mask.np[:num_reqs])[0] |
| 4792 | discard_req_indices_set = set(discard_req_indices) |
| 4793 | prev_req_id_to_index: dict[str, int] = {} |
| 4794 | for i, req_id in enumerate(self.input_batch.req_ids): |
| 4795 | if i in discard_req_indices_set: |
| 4796 | continue |
| 4797 | prev_req_id_to_index[req_id] = i |
| 4798 | # PP+async scheduling: advance per-request local cached output length by |
| 4799 | # appending a placeholder (-1) token id. |
| 4800 | if (req_state := self.requests.get(req_id)) is not None: |
| 4801 | req_state.output_token_ids.append(-1) |
| 4802 | pos = self.input_batch.num_tokens_no_spec[i] |
| 4803 | self.input_batch.is_token_ids[i, pos] = True |
| 4804 | self.input_batch.num_tokens_no_spec[i] = pos + 1 |
| 4805 | self.input_batch.prev_req_id_to_index = prev_req_id_to_index |
| 4806 | |
| 4807 | @contextmanager |
| 4808 | def preserve_serving_state(self): |
| 4809 | multi_block_table = self.input_batch.block_table |
| 4810 | saved = [ |
| 4811 | (bt.block_table.gpu.clone(), bt.block_table.cpu.clone()) |
| 4812 | for bt in multi_block_table.block_tables |
| 4813 | ] |
| 4814 | multi_block_table.clear() |
| 4815 | try: |
| 4816 | yield |
| 4817 | finally: |
| 4818 | for bt, (saved_gpu, saved_cpu) in zip( |
| 4819 | multi_block_table.block_tables, saved |
| 4820 | ): |
| 4821 | bt.block_table.gpu.copy_(saved_gpu) |
| 4822 | bt.block_table.cpu.copy_(saved_cpu) |
| 4823 | |
| 4824 | def take_draft_token_ids(self) -> DraftTokenIds | None: |
| 4825 | if not self.num_spec_tokens or not self._draft_token_req_ids: |
| 4826 | return None |
| 4827 | draft_token_ids, req_ids = self._get_draft_token_ids_cpu() |
| 4828 | return DraftTokenIds(req_ids, draft_token_ids) |
| 4829 | |
| 4830 | def _copy_draft_token_ids_to_cpu( |
| 4831 | self, scheduler_output: "SchedulerOutput", zeros_only: bool = False |
| 4832 | ) -> None: |
| 4833 | if torch.is_tensor(self._draft_token_ids): |
| 4834 | assert isinstance(self._draft_token_ids, torch.Tensor) |
| 4835 | self.prev_num_spec_tokens = self._draft_token_ids.shape[1] |
| 4836 | # Check if we need to copy draft tokens to CPU. In async scheduling, |
| 4837 | # we only copy when needed for structured output, penalties or bad_words. |
| 4838 | if self.use_async_scheduling and not ( |
| 4839 | scheduler_output.has_structured_output_requests |
| 4840 | or self.input_batch.sampling_metadata.output_token_ids |
| 4841 | ): |
| 4842 | return |
| 4843 | # We must also set the corresponding request ids. |
| 4844 | self._draft_token_req_ids = self.input_batch.req_ids.copy() |
| 4845 | |
| 4846 | draft_token_ids: torch.Tensor = self._draft_token_ids |
| 4847 | if not torch.is_tensor(draft_token_ids): |
| 4848 | return |
| 4849 | assert self.draft_token_ids_event is not None |
| 4850 | assert self.draft_token_ids_copy_stream is not None |
| 4851 | assert self.draft_token_ids_cpu is not None |
| 4852 | default_stream = torch.cuda.current_stream() |
| 4853 | num_reqs = draft_token_ids.shape[0] |
| 4854 | num_spec_tokens = draft_token_ids.shape[1] |
| 4855 | with torch.cuda.stream(self.draft_token_ids_copy_stream): |
| 4856 | if not zeros_only: |
| 4857 | # Trigger async copy of draft token ids to cpu. |
| 4858 | self.draft_token_ids_copy_stream.wait_stream(default_stream) |
| 4859 | self.draft_token_ids_cpu[:num_reqs, :num_spec_tokens].copy_( |
| 4860 | draft_token_ids, non_blocking=True |
| 4861 | ) |
| 4862 | else: |
| 4863 | # No copy needed, just zero-out cpu tensor. |
| 4864 | self.draft_token_ids_cpu[:num_reqs, :num_spec_tokens] = 0 |
| 4865 | self.draft_token_ids_event.record() |
| 4866 | |
| 4867 | def _get_draft_token_ids_cpu(self) -> tuple[list[list[int]], list[str]]: |
| 4868 | if isinstance(self._draft_token_ids, list): |
| 4869 | return self._draft_token_ids, self.input_batch.req_ids |
| 4870 | req_ids = self._draft_token_req_ids |
| 4871 | if req_ids is None: |
| 4872 | return [], [] |
| 4873 | assert self.draft_token_ids_event is not None |
| 4874 | assert self.draft_token_ids_cpu is not None |
| 4875 | self.draft_token_ids_event.synchronize() |
| 4876 | assert isinstance(self._draft_token_ids, torch.Tensor) |
| 4877 | num_spec_tokens = self._draft_token_ids.shape[1] |
| 4878 | return self.draft_token_ids_cpu[ |
| 4879 | : len(req_ids), :num_spec_tokens |
| 4880 | ].tolist(), req_ids |
| 4881 | |
| 4882 | def _copy_valid_sampled_token_count( |
| 4883 | self, next_token_ids: torch.Tensor, valid_sampled_tokens_count: torch.Tensor |
| 4884 | ) -> None: |
| 4885 | if self.valid_sampled_token_count_event is None: |
| 4886 | return |
| 4887 | |
| 4888 | default_stream = torch.cuda.current_stream() |
| 4889 | # Initialize a new stream to overlap the copy operation with |
| 4890 | # prepare_input of draft model. |
| 4891 | with torch.cuda.stream(self.valid_sampled_token_count_copy_stream): |
| 4892 | self.valid_sampled_token_count_copy_stream.wait_stream(default_stream) # type: ignore |
| 4893 | counts = valid_sampled_tokens_count |
| 4894 | counts_cpu = self.valid_sampled_token_count_cpu |
| 4895 | assert counts_cpu is not None |
| 4896 | counts_cpu[: counts.shape[0]].copy_(counts, non_blocking=True) |
| 4897 | self.valid_sampled_token_count_event.record() |
| 4898 | |
| 4899 | if self.use_async_spec_decode: |
| 4900 | # Stash for GPU-side correction in _prepare_inputs. |
| 4901 | self.valid_sampled_token_count_gpu = valid_sampled_tokens_count |
| 4902 | self.input_batch.prev_sampled_token_ids = next_token_ids.unsqueeze(1) |
| 4903 | |
| 4904 | def _get_valid_sampled_token_count(self) -> list[int]: |
| 4905 | # Wait until valid_sampled_tokens_count is copied to cpu, |
| 4906 | prev_sampled_token_ids = self.input_batch.prev_sampled_token_ids |
| 4907 | sampled_count_event = self.valid_sampled_token_count_event |
| 4908 | if sampled_count_event is None or prev_sampled_token_ids is None: |
| 4909 | return [] |
| 4910 | |
| 4911 | counts_cpu = self.valid_sampled_token_count_cpu |
| 4912 | assert counts_cpu is not None |
| 4913 | sampled_count_event.synchronize() |
| 4914 | return counts_cpu[: prev_sampled_token_ids.shape[0]].tolist() |
| 4915 | |
| 4916 | def _get_spec_decode_draft_probs( |
| 4917 | self, spec_decode_metadata: SpecDecodeMetadata |
| 4918 | ) -> torch.Tensor | None: |
| 4919 | if self._draft_probs is None or self._draft_prob_req_ids is None: |
| 4920 | return None |
| 4921 | |
| 4922 | row_by_req_id = { |
| 4923 | req_id: idx for idx, req_id in enumerate(self._draft_prob_req_ids) |
| 4924 | } |
| 4925 | draft_probs_rows: list[torch.Tensor] = [] |
| 4926 | for req_id, num_draft in zip( |
| 4927 | self.input_batch.req_ids, spec_decode_metadata.num_draft_tokens |
| 4928 | ): |
| 4929 | if num_draft == 0: |
| 4930 | continue |
| 4931 | row_idx = row_by_req_id.get(req_id) |
| 4932 | if row_idx is None: |
| 4933 | logger.warning( |
| 4934 | "Missing cached draft probabilities for request %s; " |
| 4935 | "falling back to legacy speculative rejection behavior.", |
| 4936 | req_id, |
| 4937 | ) |
| 4938 | return None |
| 4939 | draft_probs_rows.append(self._draft_probs[row_idx, :num_draft]) |
| 4940 | |
| 4941 | if not draft_probs_rows: |
| 4942 | return None |
| 4943 | return torch.cat(draft_probs_rows, dim=0).contiguous() |
| 4944 | |
| 4945 | def propose_draft_token_ids( |
| 4946 | self, |
| 4947 | scheduler_output: "SchedulerOutput", |
| 4948 | sampled_token_ids: torch.Tensor | list[list[int]], |
| 4949 | sampling_metadata: SamplingMetadata, |
| 4950 | hidden_states: torch.Tensor, |
| 4951 | sample_hidden_states: torch.Tensor, |
| 4952 | aux_hidden_states: list[torch.Tensor] | None, |
| 4953 | spec_decode_metadata: SpecDecodeMetadata | None, |
| 4954 | common_attn_metadata: CommonAttentionMetadata, |
| 4955 | slot_mappings: dict[str, torch.Tensor] | list[dict[str, torch.Tensor]] | None, |
| 4956 | ) -> list[list[int]] | torch.Tensor: |
| 4957 | num_scheduled_tokens = scheduler_output.total_num_scheduled_tokens |
| 4958 | spec_config = self.speculative_config |
| 4959 | assert spec_config is not None |
| 4960 | num_spec_tokens_to_schedule = scheduler_output.num_spec_tokens_to_schedule |
| 4961 | self._draft_probs = None |
| 4962 | self._draft_prob_req_ids = None |
| 4963 | if spec_config.method == "ngram": |
| 4964 | from vllm.v1.spec_decode.ngram_proposer import NgramProposer |
| 4965 | |
| 4966 | assert isinstance(sampled_token_ids, list) |
| 4967 | assert isinstance(self.drafter, NgramProposer) |
| 4968 | draft_token_ids = self.drafter.propose( |
| 4969 | num_spec_tokens_to_schedule, |
| 4970 | sampled_token_ids, |
| 4971 | self.input_batch.num_tokens_no_spec, |
| 4972 | self.input_batch.token_ids_cpu, |
| 4973 | slot_mappings=slot_mappings, |
| 4974 | ) |
| 4975 | elif spec_config.method == "custom_class": |
| 4976 | assert isinstance(sampled_token_ids, list) |
| 4977 | draft_token_ids = cast(Any, self.drafter).propose( |
| 4978 | sampled_token_ids, |
| 4979 | self.input_batch.num_tokens_no_spec, |
| 4980 | self.input_batch.token_ids_cpu, |
| 4981 | slot_mappings=slot_mappings, |
| 4982 | ) |
| 4983 | elif spec_config.use_ngram_gpu(): |
| 4984 | assert isinstance(self.drafter, NgramProposerGPU) |
| 4985 | ( |
| 4986 | next_token_ids, |
| 4987 | valid_sampled_tokens_count, |
| 4988 | valid_sampled_token_ids_gpu, |
| 4989 | ) = self.drafter.update_token_ids_ngram( |
| 4990 | sampled_token_ids, |
| 4991 | self.input_batch, |
| 4992 | self.token_ids_gpu_tensor, |
| 4993 | self.num_tokens_no_spec_gpu, |
| 4994 | self.discard_request_mask.gpu, |
| 4995 | ) |
| 4996 | self._copy_valid_sampled_token_count( |
| 4997 | next_token_ids, valid_sampled_tokens_count |
| 4998 | ) |
| 4999 | |
| 5000 | batch_size = next_token_ids.shape[0] |
Showing first 5,000 of 7,522 lines.
View raw