gpu_model_runner.py 304 KB
Newer Older
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 contextmanager
12
from copy import copy, deepcopy
13
from dataclasses import dataclass, replace
14
from functools import reduce
15
from typing import TYPE_CHECKING, Any, NamedTuple, TypeAlias, cast
16
17
18
19
20

import numpy as np
import torch
import torch.distributed
import torch.nn as nn
21
from tqdm import tqdm
22

23
import vllm.envs as envs
24
from vllm.compilation.counter import compilation_counter
25
from vllm.compilation.cuda_graph import CUDAGraphStat, CUDAGraphWrapper
26
from vllm.compilation.monitor import set_cudagraph_capturing_enabled
27
from vllm.config import (
28
    CompilationMode,
29
30
31
    CUDAGraphMode,
    VllmConfig,
    get_layers_from_vllm_config,
32
    set_current_vllm_config,
33
34
    update_config,
)
35
from vllm.config.cache import CacheConfig
36
from vllm.distributed.ec_transfer import get_ec_transfer, has_ec_transfer
37
from vllm.distributed.eplb.eplb_state import EplbState
38
from vllm.distributed.kv_transfer import get_kv_transfer_group, has_kv_transfer_group
39
from vllm.distributed.kv_transfer.kv_connector.utils import copy_kv_blocks
40
from vllm.distributed.parallel_state import (
41
    get_dcp_group,
42
43
44
45
46
47
    get_pp_group,
    get_tp_group,
    graph_capture,
    is_global_first_rank,
    prepare_communication_buffer_for_model,
)
48
49
50
51
from vllm.forward_context import (
    BatchDescriptor,
    set_forward_context,
)
52
from vllm.logger import init_logger
53
from vllm.lora.layers import LoRAMapping, LoRAMappingType
54
from vllm.model_executor.layers.attention import Attention, MLAAttention
55
from vllm.model_executor.layers.attention_layer_base import AttentionLayerBase
56
57
58
from vllm.model_executor.layers.fused_moe.routed_experts_capturer import (
    RoutedExpertsCapturer,
)
59
60
61
62
from vllm.model_executor.layers.rotary_embedding import (
    MRotaryEmbedding,
    XDRotaryEmbedding,
)
63
from vllm.model_executor.model_loader import get_model_loader
64
65
66
67
from vllm.model_executor.model_loader.reload import (
    finalize_layerwise_reload,
    initialize_layerwise_reload,
)
68
from vllm.model_executor.models.interfaces import (
69
    MultiModalEmbeddings,
70
    SupportsMRoPE,
71
    SupportsMultiModal,
72
    SupportsXDRoPE,
73
74
75
76
    is_mixture_of_experts,
    supports_eagle3,
    supports_mrope,
    supports_multimodal_pruning,
77
    supports_realtime,
78
    supports_transcription,
79
    supports_xdrope,
80
)
81
from vllm.model_executor.models.interfaces_base import (
82
83
84
85
    VllmModelForPooling,
    is_pooling_model,
    is_text_generation_model,
)
86
87
88
89
90
from vllm.model_executor.offloader import (
    create_offloader,
    get_offloader,
    set_offloader,
)
91
from vllm.multimodal import MULTIMODAL_REGISTRY
92
from vllm.multimodal.encoder_budget import MultiModalBudget
93
94
95
96
97
from vllm.multimodal.inputs import (
    BatchedTensorInputs,
    MultiModalKwargsItem,
    PlaceholderRange,
)
98
from vllm.multimodal.utils import group_and_batch_mm_kwargs
99
from vllm.platforms import current_platform
100
from vllm.pooling_params import PoolingParams
101
from vllm.sampling_params import SamplingType
102
from vllm.sequence import IntermediateTensors
103
from vllm.tasks import GenerationTask, PoolingTask, SupportedTask
104
from vllm.tracing import instrument
105
from vllm.utils import length_from_prompt_token_ids_or_embeds
106
from vllm.utils.math_utils import cdiv, round_up
107
from vllm.utils.mem_utils import DeviceMemoryProfiler, format_gib
108
from vllm.utils.nvtx_pytorch_hooks import PytHooks
109
from vllm.utils.platform_utils import is_pin_memory_available, num_compute_units
110
111
from vllm.utils.torch_utils import (
    get_dtype_size,
112
    is_quantized_kv_cache,
113
114
    kv_cache_dtype_str_to_dtype,
)
115
116
from vllm.v1.attention.backend import (
    AttentionBackend,
117
    AttentionCGSupport,
118
    AttentionMetadata,
119
    AttentionMetadataBuilder,
120
    AttentionType,
121
    CommonAttentionMetadata,
122
)
123
from vllm.v1.attention.backends.gdn_attn import GDNAttentionMetadataBuilder
124
from vllm.v1.attention.backends.mamba2_attn import Mamba2AttentionMetadataBuilder
125
from vllm.v1.attention.backends.utils import (
126
    NULL_BLOCK_ID,
127
    create_fast_prefill_custom_backend,
128
    get_dcp_local_seq_lens,
129
130
    reorder_batch_to_split_decodes_and_prefills,
)
131
from vllm.v1.core.sched.output import NewRequestData
132
from vllm.v1.cudagraph_dispatcher import CudagraphDispatcher
133
134
135
136
137
138
139
140
141
142
143
144
145
146
147
148
149
from vllm.v1.kv_cache_interface import (
    AttentionSpec,
    ChunkedLocalAttentionSpec,
    CrossAttentionSpec,
    EncoderOnlyAttentionSpec,
    FullAttentionSpec,
    KVCacheConfig,
    KVCacheGroupSpec,
    KVCacheSpec,
    MambaSpec,
    SlidingWindowSpec,
    UniformTypeKVCacheSpecs,
)
from vllm.v1.outputs import (
    EMPTY_MODEL_RUNNER_OUTPUT,
    AsyncModelRunnerOutput,
    DraftTokenIds,
150
    ECConnectorOutput,
151
    KVConnectorOutput,
152
153
154
155
156
    LogprobsLists,
    LogprobsTensors,
    ModelRunnerOutput,
    PoolerOutput,
    SamplerOutput,
157
    make_empty_encoder_model_runner_output,
158
)
159
from vllm.v1.pool.metadata import PoolingMetadata, PoolingStates
160
from vllm.v1.sample.logits_processor import LogitsProcessors, build_logitsprocs
161
from vllm.v1.sample.logits_processor.interface import LogitsProcessor
162
from vllm.v1.sample.metadata import SamplingMetadata
163
from vllm.v1.sample.rejection_sampler import RejectionSampler
164
from vllm.v1.sample.sampler import Sampler
165
from vllm.v1.spec_decode.dflash import DFlashProposer
166
from vllm.v1.spec_decode.draft_model import DraftModelProposer
167
from vllm.v1.spec_decode.eagle import EagleProposer
168
from vllm.v1.spec_decode.extract_hidden_states import ExtractHiddenStatesProposer
169
from vllm.v1.spec_decode.medusa import MedusaProposer
170
from vllm.v1.spec_decode.metadata import SpecDecodeMetadata
171
172
173
174
175
176
from vllm.v1.spec_decode.ngram_proposer_gpu import (
    NgramProposerGPU,
    copy_num_valid_draft_tokens,
    update_ngram_gpu_tensors_incremental,
    update_scheduler_for_invalid_drafts,
)
177
from vllm.v1.spec_decode.suffix_decoding import SuffixDecodingProposer
178
from vllm.v1.spec_decode.utils import update_num_computed_tokens_for_batch_change
179
from vllm.v1.structured_output.utils import apply_grammar_bitmask
180
from vllm.v1.utils import CpuGpuBuffer, record_function_or_nullcontext
181
182
183
184
185
from vllm.v1.worker import mamba_utils
from vllm.v1.worker.cp_utils import (
    check_attention_cp_compatibility,
    get_total_cp_world_size,
)
186
from vllm.v1.worker.dp_utils import coordinate_batch_across_dp
187
from vllm.v1.worker.ec_connector_model_runner_mixin import ECConnectorModelRunnerMixin
188
from vllm.v1.worker.gpu.pool.late_interaction_runner import LateInteractionRunner
189
from vllm.v1.worker.gpu_input_batch import CachedRequestState, InputBatch
190
from vllm.v1.worker.gpu_ubatch_wrapper import UBatchWrapper
191
from vllm.v1.worker.kv_connector_model_runner_mixin import KVConnectorModelRunnerMixin
192
from vllm.v1.worker.lora_model_runner_mixin import LoRAModelRunnerMixin
193
194
195
from vllm.v1.worker.ubatch_utils import (
    UBatchSlices,
    check_ubatch_thresholds,
196
    maybe_create_ubatch_slices,
197
    split_attn_metadata,
198
)
199
from vllm.v1.worker.utils import is_residual_scattered_for_sp
200
from vllm.v1.worker.workspace import lock_workspace
201

202
203
from .utils import (
    AttentionGroup,
204
    KVBlockZeroer,
205
206
    add_kv_sharing_layers_to_kv_cache_groups,
    bind_kv_cache,
207
    prepare_kernel_block_sizes,
208
209
    sanity_check_mm_encoder_outputs,
)
210

211
if TYPE_CHECKING:
212
    from vllm.v1.core.sched.output import GrammarOutput, SchedulerOutput
213
    from vllm.v1.spec_decode.ngram_proposer import NgramProposer
214
    from vllm.v1.worker.gpu.mm.encoder_cudagraph import EncoderCudaGraphManager
215
216
217

logger = init_logger(__name__)

218
219
AttnMetadataDict: TypeAlias = dict[str, AttentionMetadata]
# list when ubatching is enabled
220
PerLayerAttnMetadata: TypeAlias = list[AttnMetadataDict] | AttnMetadataDict
221

222

223
224
225
226
227
228
# Wrapper for ModelRunnerOutput to support overlapped execution.
class AsyncGPUModelRunnerOutput(AsyncModelRunnerOutput):
    def __init__(
        self,
        model_runner_output: ModelRunnerOutput,
        sampled_token_ids: torch.Tensor,
229
        logprobs_tensors: LogprobsTensors | None,
230
231
        invalid_req_indices: list[int],
        async_output_copy_stream: torch.cuda.Stream,
232
        vocab_size: int,
233
234
235
236
237
    ):
        self._model_runner_output = model_runner_output
        self._invalid_req_indices = invalid_req_indices

        # Event on the copy stream so we can synchronize the non-blocking copy.
238
        self.async_copy_ready_event = torch.Event()
239
240
241
242

        # Keep a reference to the device tensor to avoid it being
        # deallocated until we finish copying it to the host.
        self._sampled_token_ids = sampled_token_ids
243
        self.vocab_size = vocab_size
244
        self._logprobs_tensors = logprobs_tensors
245
246
247
248
249

        # Initiate the copy on a separate stream, but do not synchronize it.
        default_stream = torch.cuda.current_stream()
        with torch.cuda.stream(async_output_copy_stream):
            async_output_copy_stream.wait_stream(default_stream)
250
            self.sampled_token_ids_cpu = self._sampled_token_ids.to(
251
252
                "cpu", non_blocking=True
            )
253
254
255
256
257
            self._logprobs_tensors_cpu = (
                self._logprobs_tensors.to_cpu_nonblocking()
                if self._logprobs_tensors
                else None
            )
258
            self.async_copy_ready_event.record()
259
260
261

    def get_output(self) -> ModelRunnerOutput:
        """Copy the device tensors to the host and return a ModelRunnerOutput.
262

263
264
        This function blocks until the copy is finished.
        """
265
        max_gen_len = self.sampled_token_ids_cpu.shape[-1]
266
        self.async_copy_ready_event.synchronize()
267

268
269
        # Release the device tensors once the copy has completed.
        del self._logprobs_tensors
270
        del self._sampled_token_ids
271
        if max_gen_len == 1:
272
            valid_sampled_token_ids = self.sampled_token_ids_cpu.tolist()
273
274
            for i in self._invalid_req_indices:
                valid_sampled_token_ids[i].clear()
275
276
277
            logprobs_lists = None
            if self._logprobs_tensors_cpu is not None:
                logprobs_lists = self._logprobs_tensors_cpu.tolists()
278
        else:
279
            valid_sampled_token_ids, logprobs_lists = RejectionSampler.parse_output(
280
281
                self.sampled_token_ids_cpu,
                self.vocab_size,
282
                self._invalid_req_indices,
283
                logprobs_tensors=self._logprobs_tensors_cpu,
284
            )
285
286
287

        output = self._model_runner_output
        output.sampled_token_ids = valid_sampled_token_ids
288
        output.logprobs = logprobs_lists
289
290
291
        return output


292
293
294
295
296
297
298
299
300
301
302
303
304
305
306
307
308
309
310
311
312
313
314
315
316
317
318
319
320
321
322
323
324
325
326
327
328
329
330
331
332
333
334
335
336
def _copy_pooler_output_to_cpu(
    raw_pooler_output: PoolerOutput, finished_mask: list[bool]
) -> list[torch.Tensor | None]:
    num_reqs = len(finished_mask)

    if isinstance(raw_pooler_output, torch.Tensor):
        if raw_pooler_output.shape[0] != num_reqs:
            raise ValueError(
                "Pooler output batch size does not match finished mask size: "
                f"{raw_pooler_output.shape[0]} != {num_reqs}."
            )

        num_finished = sum(finished_mask)
        if num_finished == 0:
            return [None] * num_reqs
        if num_finished == num_reqs:
            return list(raw_pooler_output.to("cpu", non_blocking=True))

        # partial finished
        finished_indices = [i for i, include in enumerate(finished_mask) if include]
        index_tensor = torch.tensor(
            finished_indices, device=raw_pooler_output.device, dtype=torch.long
        )
        finished_outputs = raw_pooler_output.index_select(0, index_tensor).to(
            "cpu", non_blocking=True
        )
        partial_pooler_output: list[torch.Tensor | None] = [None] * num_reqs
        for i, out in zip(finished_indices, finished_outputs):
            partial_pooler_output[i] = out
        return partial_pooler_output

    assert isinstance(raw_pooler_output, list)
    if len(raw_pooler_output) != num_reqs:
        raise ValueError(
            "Pooler output batch size does not match finished mask size: "
            f"{len(raw_pooler_output)} != {num_reqs}."
        )

    pooler_output: list[torch.Tensor | None] = [None] * num_reqs
    for i, (out, include) in enumerate(zip(raw_pooler_output, finished_mask)):
        if include and out is not None:
            pooler_output[i] = out.to("cpu", non_blocking=True)
    return pooler_output


337
338
339
340
341
342
343
344
345
346
347
348
349
350
351
352
353
354
355
356
357
class AsyncGPUPoolingModelRunnerOutput(AsyncModelRunnerOutput):
    def __init__(
        self,
        model_runner_output: ModelRunnerOutput,
        raw_pooler_output: PoolerOutput,
        finished_mask: list[bool],
        async_output_copy_stream: torch.cuda.Stream,
    ):
        self._model_runner_output = model_runner_output

        # Event on the copy stream so we can synchronize the non-blocking copy.
        self.async_copy_ready_event = torch.Event()

        # Keep a reference to the device tensors to avoid them being
        # deallocated until we finish copying it to the host.
        self._raw_pooler_output = raw_pooler_output

        # Initiate the copy on a separate stream, but do not synchronize it.
        default_stream = torch.cuda.current_stream()
        with torch.cuda.stream(async_output_copy_stream):
            async_output_copy_stream.wait_stream(default_stream)
358
359
360
            self._model_runner_output.pooler_output = _copy_pooler_output_to_cpu(
                raw_pooler_output=self._raw_pooler_output,
                finished_mask=finished_mask,
361
362
363
364
365
366
367
368
369
370
371
372
373
374
            )
            self.async_copy_ready_event.record()

    def get_output(self) -> ModelRunnerOutput:
        """Copy the device tensors to the host and return a ModelRunnerOutput.
        This function blocks until the copy is finished.
        """
        self.async_copy_ready_event.synchronize()

        # Release the device tensors once the copy has completed.
        del self._raw_pooler_output
        return self._model_runner_output


375
376
377
378
379
380
381
382
383
384
385
class ExecuteModelState(NamedTuple):
    """Ephemeral cached state transferred between execute_model() and
    sample_tokens(), after execute_model() returns None."""

    scheduler_output: "SchedulerOutput"
    logits: torch.Tensor
    spec_decode_metadata: SpecDecodeMetadata | None
    spec_decode_common_attn_metadata: CommonAttentionMetadata | None
    hidden_states: torch.Tensor
    sample_hidden_states: torch.Tensor
    aux_hidden_states: list[torch.Tensor] | None
386
    ec_connector_output: ECConnectorOutput | None
387
    cudagraph_stats: CUDAGraphStat | None
388
    slot_mappings: dict[str, torch.Tensor] | list[dict[str, torch.Tensor]] | None
389
390


391
392
393
class GPUModelRunner(
    LoRAModelRunnerMixin, KVConnectorModelRunnerMixin, ECConnectorModelRunnerMixin
):
394
395
    def __init__(
        self,
396
        vllm_config: VllmConfig,
397
        device: torch.device,
398
    ):
399
400
401
        self.vllm_config = vllm_config
        self.model_config = vllm_config.model_config
        self.cache_config = vllm_config.cache_config
402
        self.offload_config = vllm_config.offload_config
403
        self.compilation_config = vllm_config.compilation_config
404
405
406
407
408
409
        self.lora_config = vllm_config.lora_config
        self.load_config = vllm_config.load_config
        self.parallel_config = vllm_config.parallel_config
        self.scheduler_config = vllm_config.scheduler_config
        self.speculative_config = vllm_config.speculative_config
        self.observability_config = vllm_config.observability_config
410

411
412
413
414
        model_config = self.model_config
        cache_config = self.cache_config
        scheduler_config = self.scheduler_config
        parallel_config = self.parallel_config
415
        self.device = device
416
417
        self.pin_memory = is_pin_memory_available()
        self.dtype = self.model_config.dtype
418

419
420
421
        self.kv_cache_dtype = kv_cache_dtype_str_to_dtype(
            cache_config.cache_dtype, self.model_config
        )
422

423
        self.is_pooling_model = model_config.runner_type == "pooling"
424
        self.enable_prompt_embeds = model_config.enable_prompt_embeds
425
        self.is_multimodal_raw_input_only_model = (
426
427
            model_config.is_multimodal_raw_input_only_model
        )
428
        # These will be overridden in load_model()
429
        self.is_multimodal_pruning_enabled = False
430
        self.requires_sequential_video_encoding = False
431
432
433
        # Set to True after init_routed_experts_capturer() completes.
        # Prevents routed experts code from running during profiling/dummy run.
        self.routed_experts_initialized = False
434
        self.max_model_len = model_config.max_model_len
435
436
437

        # Always set to false after the first forward pass
        self.calculate_kv_scales = self.cache_config.calculate_kv_scales
438
        self.dcp_world_size = self.parallel_config.decode_context_parallel_size
439
        self.dcp_rank = 0 if self.dcp_world_size <= 1 else get_dcp_group().rank_in_group
440
        self.max_num_tokens = scheduler_config.max_num_batched_tokens
441
        self.max_num_reqs = scheduler_config.max_num_seqs
442

443
444
        # Broadcast PP output for external_launcher (torchrun)
        # to make sure we are synced across pp ranks
445
        # TODO: Support overlapping micro-batches
446
447
        # https://github.com/vllm-project/vllm/issues/18019
        self.broadcast_pp_output = (
448
            self.parallel_config.distributed_executor_backend == "external_launcher"
449
            and len(get_pp_group().ranks) > 1
450
        )
451

452
        # Model-related.
453
        self.num_query_heads = model_config.get_num_attention_heads(parallel_config)
454
        self.inputs_embeds_size = model_config.get_inputs_embeds_size()
455
        # Only relevant for models using ALiBi (e.g, MPT)
456
        self.use_alibi = model_config.uses_alibi
457

458
        self.cascade_attn_enabled = not self.model_config.disable_cascade_attn
459
        self.is_mm_prefix_lm = self.model_config.is_mm_prefix_lm
460

461
        # Multi-modal data support
462
        self.mm_registry = MULTIMODAL_REGISTRY
463
        self.uses_mrope = model_config.uses_mrope
464
        self.uses_xdrope_dim = model_config.uses_xdrope_dim
465
        self.supports_mm_inputs = self.mm_registry.supports_multimodal_inputs(
466
            model_config
467
        )
468

469
470
471
        if self.model_config.is_encoder_decoder:
            # Maximum length of the encoder input, only for encoder-decoder
            # models.
472
            self.max_encoder_len = scheduler_config.max_num_encoder_input_tokens
473
474
475
        else:
            self.max_encoder_len = 0

476
477
478
        # Async scheduling
        self.use_async_scheduling = self.scheduler_config.async_scheduling

479
        # Sampler
480
        self.sampler = Sampler(logprobs_mode=self.model_config.logprobs_mode)
481

482
        self.eplb_state: EplbState | None = None
483
484
        # NOTE(yongji): flag to temporarily disable EPLB during scaling up/down
        self.eep_eplb_suppressed = False
485
486
487
488
489
490
        """
        State of the expert parallelism load balancer.

        Will be lazily initialized when the model is loaded.
        """

491
        # Lazy initializations
492
        # self.model: nn.Module  # Set after load_model
493
        # Initialize in initialize_kv_cache
494
        self.kv_caches: list[torch.Tensor] = []
495
496
497
        # Initialize in initialize_kv_cache_tensors
        self.cross_layers_kv_cache: torch.Tensor | None = None
        self.cross_layers_attn_backend: type[AttentionBackend] | None = None
498
499
        # indexes: [kv_cache_group_id][attn_group]
        self.attn_groups: list[list[AttentionGroup]] = []
500
501
        # self.kv_cache_config: KVCacheConfig

502
503
        # mm_hash ->  encoder_output
        self.encoder_cache: dict[str, torch.Tensor] = {}
504
        self.late_interaction_runner = LateInteractionRunner()
505

506
507
508
        # Encoder CUDA graph manager (initialized after model load if enabled)
        self.encoder_cudagraph_manager: EncoderCudaGraphManager | None = None

509
        self.use_aux_hidden_state_outputs = False
510
511
512
513
514
        # Set up speculative decoding.
        # NOTE(Jiayi): currently we put the entire draft model on
        # the last PP rank. This is not ideal if there are many
        # layers in the draft model.
        if self.speculative_config and get_pp_group().is_last_rank:
515
            self.drafter: (
516
                NgramProposer  # noqa: F823
517
                | NgramProposerGPU
518
519
                | SuffixDecodingProposer
                | EagleProposer
520
                | DFlashProposer
521
522
                | DraftModelProposer
                | MedusaProposer
523
                | ExtractHiddenStatesProposer
524
            )
525
            if self.speculative_config.method == "ngram":
526
527
                from vllm.v1.spec_decode.ngram_proposer import NgramProposer

528
                self.drafter = NgramProposer(self.vllm_config)
529
530
531
532
533
534
            elif self.speculative_config.uses_draft_model():
                self.drafter = DraftModelProposer(
                    vllm_config=self.vllm_config,
                    device=self.device,
                    runner=self,
                )
535
536
537
538
539
540
541
542
543
544
545
546
547
548
549
550
551
            elif self.speculative_config.use_ngram_gpu():
                self.drafter = NgramProposerGPU(self.vllm_config, self.device, self)
                self.num_tokens_no_spec_gpu = torch.zeros(
                    self.max_num_reqs, dtype=torch.int32, device=device
                )
                self.token_ids_gpu_tensor = torch.zeros(
                    self.max_num_reqs,
                    self.max_model_len,
                    dtype=torch.int32,
                    device=device,
                )
                self._ngram_pinned_idx_buf = torch.zeros(
                    self.max_num_reqs, dtype=torch.long, pin_memory=True
                )
                self._ngram_pinned_val_buf = torch.zeros(
                    self.max_num_reqs, dtype=torch.int32, pin_memory=True
                )
552
553
554
            elif self.speculative_config.use_dflash():
                self.drafter = DFlashProposer(self.vllm_config, self.device, self)
                self.use_aux_hidden_state_outputs = True
555
556
            elif self.speculative_config.method == "suffix":
                self.drafter = SuffixDecodingProposer(self.vllm_config)
557
            elif self.speculative_config.use_eagle():
558
                self.drafter = EagleProposer(self.vllm_config, self.device, self)
559
                if self.speculative_config.method == "eagle3":
560
561
562
                    self.use_aux_hidden_state_outputs = (
                        self.drafter.eagle3_use_aux_hidden_state
                    )
563
564
            elif self.speculative_config.method == "medusa":
                self.drafter = MedusaProposer(
565
                    vllm_config=self.vllm_config, device=self.device
566
                )
567
568
569
570
571
            elif self.speculative_config.method == "extract_hidden_states":
                self.drafter = ExtractHiddenStatesProposer(
                    vllm_config=self.vllm_config, device=self.device
                )
                self.use_aux_hidden_state_outputs = True
572
            else:
573
574
575
576
                raise ValueError(
                    "Unknown speculative decoding method: "
                    f"{self.speculative_config.method}"
                )
577
            self.rejection_sampler = RejectionSampler(self.sampler)
578

579
        self.num_spec_tokens = 0
580
        self.valid_sampled_token_count_gpu: torch.Tensor | None = None
581
582
        if self.speculative_config:
            self.num_spec_tokens = self.speculative_config.num_speculative_tokens
583
584
585
586
587
            draft_config = self.speculative_config.draft_model_config
            if draft_config is not None and draft_config.max_model_len is not None:
                self.effective_drafter_max_model_len = draft_config.max_model_len
            else:
                self.effective_drafter_max_model_len = self.max_model_len
588
589
590
        self.use_async_spec_decode = (
            self.use_async_scheduling and self.num_spec_tokens > 0
        )
591

592
        # Request states.
593
        self.requests: dict[str, CachedRequestState] = {}
594
595
596
        # NOTE(rob): num_prompt_logprobs only includes reqs
        # that are currently in the prefill phase.
        self.num_prompt_logprobs: dict[str, int] = {}
597

598
599
600
601
602
603
604
605
606
        # Input Batch
        # NOTE(Chen): Ideally, we should initialize the input batch inside
        # `initialize_kv_cache` based on the kv cache config. However, as in
        # https://github.com/vllm-project/vllm/pull/18298, due to some unknown
        # reasons, we have to initialize the input batch before `load_model`,
        # quantization + weight offloading will fail otherwise. As a temporary
        # solution, we initialize the input batch here, and re-initialize it
        # in `initialize_kv_cache` if the block_sizes here is different from
        # the block_sizes in the kv cache config.
607
608
609
610
        logits_processors = model_config.logits_processors
        custom_logitsprocs: Sequence[str | type[LogitsProcessor]] = (
            tuple(logits_processors) if logits_processors is not None else ()
        )
611
612
613
614
615
        placeholder_block_size = (
            self.cache_config.block_size or CacheConfig.DEFAULT_BLOCK_SIZE
        )
        self._init_block_sizes = [placeholder_block_size]
        self._init_kernel_block_sizes = [placeholder_block_size]
616
617
        self.input_batch = InputBatch(
            max_num_reqs=self.max_num_reqs,
618
            # We need to use the encoder length for encoder-decoder
619
620
            # because of KV cache for cross-attention.
            max_model_len=max(self.max_model_len, self.max_encoder_len),
621
622
623
            max_num_batched_tokens=self.max_num_tokens,
            device=self.device,
            pin_memory=self.pin_memory,
624
            vocab_size=self.model_config.get_vocab_size(),
625
626
            block_sizes=[placeholder_block_size],
            kernel_block_sizes=[placeholder_block_size],
627
            is_spec_decode=bool(self.vllm_config.speculative_config),
628
            logitsprocs=build_logitsprocs(
629
630
631
                self.vllm_config,
                self.device,
                self.pin_memory,
632
                self.is_pooling_model,
633
                custom_logitsprocs,
634
            ),
635
636
            # We currently don't know whether a particular custom logits processor
            # uses output token ids so we set this conservatively.
637
638
639
640
            # ThinkingTokenBudgetLogitsProcessor also needs output token ids to
            # correctly track think start/end token sequences in async scheduling.
            logitsprocs_need_output_token_ids=bool(custom_logitsprocs)
            or self.vllm_config.reasoning_config is not None,
641
            is_pooling_model=self.is_pooling_model,
642
            cp_kv_cache_interleave_size=self.parallel_config.cp_kv_cache_interleave_size,
643
        )
644

645
646
647
648
649
        # Separate cuda stream for overlapping transfer of sampled token ids from
        # GPU to CPU when async scheduling is enabled.
        self.async_output_copy_stream: torch.cuda.Stream | None = None
        # cuda event to synchronize use of reused CPU tensors between steps
        # when async scheduling is enabled.
650
        self.prepare_inputs_event: torch.Event | None = None
651
652
        if self.use_async_scheduling:
            self.async_output_copy_stream = torch.cuda.Stream()
653
            self.prepare_inputs_event = torch.Event()
654

655
656
657
658
659
660
661
662
663
664
665
        # self.cudagraph_batch_sizes sorts in ascending order.
        if (
            self.compilation_config.cudagraph_capture_sizes
            and self.compilation_config.cudagraph_mode != CUDAGraphMode.NONE
        ):
            self.cudagraph_batch_sizes = sorted(
                self.compilation_config.cudagraph_capture_sizes
            )
        else:
            self.cudagraph_batch_sizes = []

666
        # Cache the device properties.
667
        self._init_device_properties()
668

669
670
671
672
        # Encoder timing registry for observability
        self.encoder_timing_registry: dict[str, EncoderTimingStats] = {}
        self._encoder_timing_lock = threading.Lock()

673
        # Persistent buffers for CUDA graphs.
674
        self.input_ids = self._make_buffer(self.max_num_tokens, dtype=torch.int32)
675
676
677
        self.positions = torch.zeros(
            self.max_num_tokens, dtype=torch.int64, device=self.device
        )
678
679
680
        self.query_start_loc = self._make_buffer(
            self.max_num_reqs + 1, dtype=torch.int32
        )
681
682
683
684
685
686
687
688
689
690
691
692
693
694
695
696
697
698
699
        self.seq_lens = torch.zeros(
            self.max_num_reqs, dtype=torch.int32, device=self.device
        )
        self.optimistic_seq_lens_cpu = torch.zeros(
            self.max_num_reqs, dtype=torch.int32, pin_memory=self.pin_memory
        )
        self.num_computed_tokens = torch.zeros(
            self.max_num_reqs, dtype=torch.int32, device=self.device
        )
        self.prev_num_draft_tokens = self._make_buffer(
            self.max_num_reqs, dtype=torch.int32
        )
        self.req_indices = self._make_buffer(self.max_num_tokens, dtype=torch.int64)
        # Maps current batch position -> previous batch position (-1 for new reqs)
        self.prev_positions = self._make_buffer(self.max_num_reqs, dtype=torch.int64)
        self.num_scheduled_tokens = self._make_buffer(
            self.max_num_reqs, dtype=torch.int32
        )

700
        self.encoder_seq_lens = self._make_buffer(self.max_num_reqs, dtype=torch.int32)
701
702
703
704
        if self.dcp_world_size > 1:
            self.dcp_local_seq_lens = self._make_buffer(
                self.max_num_reqs, dtype=torch.int32
            )
705
706
707
        # Because inputs_embeds may be bfloat16 and we don't need a numpy
        # version of this tensor, avoid a RuntimeError by not creating a
        # numpy buffer.
708
        self.inputs_embeds = self._make_buffer(
709
            self.max_num_tokens, self.inputs_embeds_size, dtype=self.dtype, numpy=False
710
711
        )
        self.is_token_ids = self._make_buffer(self.max_num_tokens, dtype=torch.bool)
712
713
        self.discard_request_mask = self._make_buffer(
            self.max_num_reqs, dtype=torch.bool
714
715
716
717
718
        )
        self.num_decode_draft_tokens = self._make_buffer(
            self.max_num_reqs, dtype=torch.int32
        )
        self.num_accepted_tokens = self._make_buffer(
719
            self.max_num_reqs, dtype=torch.int32
720
        )
721

722
723
        # Only relevant for multimodal models
        if self.supports_mm_inputs:
724
725
726
727
728
729
730
            # Double buffer to avoid race condition: previous iteration's async
            # copy may still be reading from CPU while current iteration writes.
            self.is_mm_embed_buffers = [
                self._make_buffer(self.max_num_tokens, dtype=torch.bool),
                self._make_buffer(self.max_num_tokens, dtype=torch.bool),
            ]
            self.is_mm_embed_idx = 0
731

732
        # Only relevant for models using M-RoPE (e.g, Qwen2-VL)
733
        if self.uses_mrope:
Roger Wang's avatar
Roger Wang committed
734
735
736
737
            # NOTE: `mrope_positions` is implemented with one additional dummy
            # position on purpose to make it non-contiguous so that it can work
            # with torch compile.
            # See detailed explanation in https://github.com/vllm-project/vllm/pull/12128#discussion_r1926431923
738
739
740
741
742
743

            # NOTE: When M-RoPE is enabled, position ids are 3D regardless of
            # the modality of inputs. For text-only inputs, each dimension has
            # identical position IDs, making M-RoPE functionally equivalent to
            # 1D-RoPE.
            # See page 5 of https://arxiv.org/abs/2409.12191
744
            self.mrope_positions = self._make_buffer(
745
746
                (3, self.max_num_tokens + 1), dtype=torch.int64
            )
747

748
749
750
751
752
753
754
        # Only relevant for models using XD-RoPE (e.g, HunYuan-VL)
        if self.uses_xdrope_dim > 0:
            # Similar to mrope but use assigned dimension number for RoPE, 4 as default.
            self.xdrope_positions = self._make_buffer(
                (self.uses_xdrope_dim, self.max_num_tokens + 1), dtype=torch.int64
            )

755
        # None in the first PP rank. The rest are set after load_model.
756
        self.intermediate_tensors: IntermediateTensors | None = None
757

758
759
760
761
762
763
764
765
        # OPTIMIZATION: Cache the arange tensors rather than creating them
        # every step. Keep in int64 to avoid overflow with long context.
        # - arange_np: immutable [0, 1, 2, ...] used as source for batched computation
        # - query_pos: CpuGpuBuffer for the computed batched arange result
        arange_size = max(self.max_num_reqs + 1, self.max_num_tokens)
        self.arange_np = np.arange(arange_size, dtype=np.int64)
        self.query_pos = self._make_buffer(arange_size, dtype=torch.int64)
        self._arange_scratch = np.empty(arange_size, dtype=np.int64)
766

767
768
769
770
771
        # Layer pairings for cross-layer KV sharing.
        # If an Attention layer `layer_name` is in the keys of this dict, it
        # means this layer will perform attention using the keys and values
        # from the KV cache of `shared_kv_cache_layers[layer_name]`.
        self.shared_kv_cache_layers: dict[str, str] = {}
772
773
774
775
776
        self.kv_sharing_fast_prefill_eligible_layers: set[str] = set()

        self.kv_sharing_fast_prefill_logits_indices = None
        if self.cache_config.kv_sharing_fast_prefill:
            self.kv_sharing_fast_prefill_logits_indices = torch.zeros(
777
778
                self.max_num_tokens, dtype=torch.int32, device=self.device
            )
779

780
        self.uniform_decode_query_len = 1 + self.num_spec_tokens
781
782
783
784

        # Cudagraph dispatcher for runtime cudagraph dispatching.
        self.cudagraph_dispatcher = CudagraphDispatcher(self.vllm_config)

785
        self.mm_budget = (
786
            MultiModalBudget(self.vllm_config, self.mm_registry)
787
788
789
            if self.supports_mm_inputs
            else None
        )
790

791
        self.reorder_batch_threshold: int | None = None
792

793
794
795
796
797
        # Attention layers that are only in the KVCacheConfig of the runner
        # (e.g., KV sharing, encoder-only attention), but not in the
        # KVCacheConfig of the scheduler.
        self.runner_only_attn_layers: set[str] = set()

798
        # Cached outputs.
799
        self._draft_token_ids: list[list[int]] | torch.Tensor | None = None
800
801
802
803
804
805
806
807
808
809
810
811
812
813
814
        # N-gram GPU path: async D2H buffer/event for per-request valid draft counts.
        self._num_valid_draft_tokens: torch.Tensor | None = None
        self._num_valid_draft_tokens_cpu: torch.Tensor | None = None
        self._num_valid_draft_tokens_event: torch.cuda.Event | None = None
        self._num_valid_draft_tokens_copy_stream: torch.cuda.Stream | None = None
        if (
            self.speculative_config is not None
            and self.speculative_config.use_ngram_gpu()
        ):
            self._num_valid_draft_tokens_cpu = torch.empty(
                self.max_num_reqs, dtype=torch.int32, pin_memory=self.pin_memory
            )
            self._num_valid_draft_tokens_event = torch.cuda.Event()
            self._num_valid_draft_tokens_copy_stream = torch.cuda.Stream()

815
        self._draft_token_req_ids: list[str] | None = None
816
        self.transfer_event = torch.Event()
817
        self.sampled_token_ids_pinned_cpu = torch.empty(
818
            (self.max_num_reqs, 1),
819
820
            dtype=torch.int64,
            device="cpu",
821
822
            pin_memory=self.pin_memory,
        )
823

824
825
        # Pre-allocated tensor for copying valid sampled token counts to CPU,
        # with dedicated stream for overlapping and event for coordination.
826
        self.valid_sampled_token_count_event: torch.Event | None = None
827
        self.valid_sampled_token_count_copy_stream: torch.cuda.Stream | None = None
828
829
830
831
832
833
        # We also copy the drafted tokens to the CPU asynchronously,
        # in case we need them for structured outputs.
        self.draft_token_ids_event: torch.Event | None = None
        self.draft_token_ids_copy_stream: torch.cuda.Stream | None = None
        self.valid_sampled_token_count_cpu: torch.Tensor | None = None
        self.draft_token_ids_cpu: torch.Tensor | None = None
834
        self.num_accepted_tokens_event: torch.Event | None = None
835
836
        if self.num_spec_tokens:
            self.draft_token_ids_event = torch.Event()
837
            self.num_accepted_tokens_event = torch.Event()
838
839
840
841
842
843
844
845
846
847
848
849
            self.draft_token_ids_copy_stream = torch.cuda.Stream()
            self.draft_token_ids_cpu = torch.empty(
                (self.max_num_reqs, self.num_spec_tokens),
                dtype=torch.int64,
                device="cpu",
                pin_memory=self.pin_memory,
            )
            if self.use_async_scheduling:
                self.valid_sampled_token_count_event = torch.Event()
                self.valid_sampled_token_count_copy_stream = torch.cuda.Stream()
                self.valid_sampled_token_count_cpu = torch.empty(
                    self.max_num_reqs,
850
                    dtype=torch.int32,
851
852
853
                    device="cpu",
                    pin_memory=self.pin_memory,
                )
854

855
856
857
858
        # Model weight offloader
        # Make sure this is called before any get_offloader call
        set_offloader(create_offloader(self.offload_config))

859
860
        # Ephemeral state transferred between execute_model() and sample_tokens().
        self.execute_model_state: ExecuteModelState | None = None
861
        self.kv_connector_output: KVConnectorOutput | None = None
862
        self.mamba_state_idx: dict[str, int] = {}
863
        self._mamba_copy_bufs: mamba_utils.MambaCopyBuffers | None = None
864
        self.layerwise_nvtx_hooks_registered = False
865

866
867
868
869
870
871
872
    def update_max_model_len(self, max_model_len: int) -> None:
        self.max_model_len = max_model_len
        if self.speculative_config:
            draft_config = self.speculative_config.draft_model_config
            if draft_config is None or draft_config.max_model_len is None:
                self.effective_drafter_max_model_len = self.max_model_len

873
    def reset_mm_cache(self) -> None:
874
875
876
877
        """
        Clear the multi-modal cache that was used during profiling,
        but no longer needed during inference.
        """
878
879
        if self.mm_budget:
            self.mm_budget.reset_cache()
880
        self.late_interaction_runner.clear()
881

882
883
884
885
886
887
888
    def reset_encoder_cache(self) -> None:
        """Clear the GPU-side encoder cache storing vision embeddings.

        This should be called when model weights are updated to ensure
        stale embeddings computed with old weights are not reused.
        """
        self.encoder_cache.clear()
889
        self.late_interaction_runner.clear()
890

891
892
893
894
895
896
897
898
899
    @torch.inference_mode()
    def init_fp8_kv_scales(self) -> None:
        """
        Re-initialize the KV cache and FP8 scales after waking from sleep.
        1. Zero out the KV cache tensors to remove garbage data from re-allocation.
        2. Reset Attention layer scaling factors (_k_scale, _v_scale) to 1.0.
          If these are left at 0.0 (default after wake_up), all KV cache values
          become effectively zero, causing gibberish output.
        """
900
        if not is_quantized_kv_cache(self.cache_config.cache_dtype):
901
902
903
904
905
906
907
908
909
910
911
912
913
914
915
916
917
918
919
920
921
922
923
924
925
926
927
928
929
930
931
932
933
934
            return

        kv_caches = getattr(self, "kv_caches", [])
        for cache_tensor in kv_caches:
            if cache_tensor is not None:
                cache_tensor.zero_()

        k_attr_names = ("_k_scale", "k_scale")
        v_attr_names = ("_v_scale", "v_scale")

        attn_layers = self.compilation_config.static_forward_context
        for name, module in attn_layers.items():
            if isinstance(module, (Attention, MLAAttention)):
                # TODO: Generally, scale is 1.0 if user uses on-the-fly fp8
                # kvcache quant. However, to get better accuracy, compression
                # frameworks like llm-compressors allow users to tune the
                # scale. We may need to restore the specific calibrated scales
                # here in the future.
                k_scale_val, v_scale_val = 1.0, 1.0

                # Processing K Scale
                for attr in k_attr_names:
                    if hasattr(module, attr):
                        param = getattr(module, attr)
                        if isinstance(param, torch.Tensor):
                            param.fill_(k_scale_val)

                # Processing V Scale
                for attr in v_attr_names:
                    if hasattr(module, attr):
                        param = getattr(module, attr)
                        if isinstance(param, torch.Tensor):
                            param.fill_(v_scale_val)

935
936
937
938
    def _get_positions(self, num_tokens: Any):
        if isinstance(num_tokens, int):
            if self.uses_mrope:
                return self.mrope_positions.gpu[:, :num_tokens]
939
940
            if self.uses_xdrope_dim > 0:
                return self.xdrope_positions.gpu[:, :num_tokens]
941
            return self.positions[:num_tokens]
942
943
944
        else:
            if self.uses_mrope:
                return self.mrope_positions.gpu[:, num_tokens]
945
946
            if self.uses_xdrope_dim > 0:
                return self.xdrope_positions.gpu[:, num_tokens]
947
            return self.positions[num_tokens]
948

949
    def _make_buffer(
950
        self, *size: int | torch.SymInt, dtype: torch.dtype, numpy: bool = True
951
952
953
954
955
956
957
958
    ) -> CpuGpuBuffer:
        return CpuGpuBuffer(
            *size,
            dtype=dtype,
            device=self.device,
            pin_memory=self.pin_memory,
            with_numpy=numpy,
        )
959

960
961
962
963
964
965
966
967
968
969
    def _get_mamba_copy_bufs(self) -> mamba_utils.MambaCopyBuffers:
        if self._mamba_copy_bufs is None:
            self._mamba_copy_bufs = mamba_utils.MambaCopyBuffers.create(
                self.max_num_reqs,
                self.kv_cache_config,
                self.model.get_mamba_state_copy_func(),
                self._make_buffer,
            )
        return self._mamba_copy_bufs

970
    def _init_model_kwargs(self):
971
972
        model_kwargs = dict[str, Any]()

973
        if not self.is_pooling_model:
974
975
            return model_kwargs

976
977
        num_reqs = self.input_batch.num_reqs
        pooling_params = self.input_batch.get_pooling_params()
978
979
980

        token_type_id_requests = dict[int, Any]()
        for i, param in enumerate(pooling_params):
981
982
983
984
985
            if (
                param.extra_kwargs is not None
                and (token_types := param.extra_kwargs.get("compressed_token_type_ids"))
                is not None
            ):
986
987
988
989
990
                token_type_id_requests[i] = token_types

        if len(token_type_id_requests) == 0:
            return model_kwargs

991
        seq_lens = self.seq_lens[:num_reqs]
992
993
994
995
996
997
998
999
        token_type_ids = []

        for i in range(num_reqs):
            pos = token_type_id_requests.get(i, seq_lens[i])
            ids = (torch.arange(seq_lens[i]) >= pos).int()
            token_type_ids.append(ids)

        model_kwargs["token_type_ids"] = torch.concat(token_type_ids).to(
1000
1001
            device=self.device
        )
1002
1003
        return model_kwargs

1004
    def _may_reorder_batch(self, scheduler_output: "SchedulerOutput") -> None:
1005
1006
        """
        Update the order of requests in the batch based on the attention
1007
        backend's needs. For example, some attention backends (namely MLA) may
1008
1009
1010
1011
1012
1013
        want to separate requests based on if the attention computation will be
        compute-bound or memory-bound.

        Args:
            scheduler_output: The scheduler output.
        """
Jiayi Yan's avatar
Jiayi Yan committed
1014
        # Attention free models have zero kv_cache_groups, however models
1015
1016
1017
1018
1019
1020
1021
        # like Mamba are also attention free but use the kv_cache for
        # keeping its internal state. This is why we check the number
        # of kv_cache groups instead of solely checking
        # for self.model_config.is_attention_free.
        if len(self.kv_cache_config.kv_cache_groups) == 0:
            return

1022
1023
1024
1025
        if self.reorder_batch_threshold is not None:
            reorder_batch_to_split_decodes_and_prefills(
                self.input_batch,
                scheduler_output,
1026
1027
                decode_threshold=self.reorder_batch_threshold,
            )
1028

1029
1030
1031
1032
1033
1034
1035
1036
1037
1038
1039
1040
1041
1042
1043
1044
1045
1046
1047
1048
    def _init_kv_zero_meta(self) -> None:
        """One-time precomputation for _zero_block_ids.

        Delegates to KVBlockZeroer.init_meta with the runner's state.
        Called from gpu_worker.py outside the CuMem pool context.
        """
        self._kv_block_zeroer = KVBlockZeroer(self.device, self.pin_memory)
        self._kv_block_zeroer.init_meta(
            attn_groups_iter=self._kv_cache_spec_attn_group_iterator(),
            kernel_block_sizes=self._kernel_block_sizes,
            cache_dtype=self.cache_config.cache_dtype,
            runner_only_attn_layers=self.runner_only_attn_layers,
            static_forward_context=(self.compilation_config.static_forward_context),
        )

    def _zero_block_ids(self, block_ids: list[int]) -> None:
        """Zero the KV cache memory for the given block IDs."""
        if hasattr(self, "_kv_block_zeroer"):
            self._kv_block_zeroer.zero_block_ids(block_ids)

1049
1050
    # Note: used for model runner override.
    def _init_device_properties(self) -> None:
1051
        """Initialize attributes from torch.cuda.get_device_properties"""
1052
1053

        self.num_sms = num_compute_units(self.device.index)
1054
1055
1056

    # Note: used for model runner override.
    def _sync_device(self) -> None:
1057
        torch.accelerator.synchronize()
1058

1059
    def _update_states(self, scheduler_output: "SchedulerOutput") -> Callable | None:
1060
1061
1062
1063
1064
1065
        """Update the cached states and the persistent batch with the scheduler
        output.

        The updated states are used by the `_prepare_inputs` function to create
        the input GPU tensors for the model.

1066
1067
        The SamplingMetadata is updated and copied to the GPU if there is a
        new/resumed/paused/finished request in the batch.
1068
1069
        """
        # Remove finished requests from the cached states.
1070
1071
        for req_id in scheduler_output.finished_req_ids:
            self.requests.pop(req_id, None)
1072
            self.num_prompt_logprobs.pop(req_id, None)
1073
1074
1075
        self.late_interaction_runner.on_requests_finished(
            scheduler_output.finished_req_ids
        )
1076
1077
1078
1079
1080
1081
1082
        # Remove the finished requests from the persistent batch.
        # NOTE(woosuk): There could be an edge case where finished_req_ids and
        # scheduled_req_ids overlap. This happens when a request is aborted and
        # then resubmitted with the same ID. In this case, we treat them as two
        # distinct requests - clearing the cached states for the first request
        # and handling the second as a new request.
        for req_id in scheduler_output.finished_req_ids:
1083
            self.input_batch.remove_request(req_id)
1084

1085
1086
1087
1088
1089
        # Zero GPU memory for freshly allocated cache blocks to prevent
        # stale NaN/data from corrupting attention or SSM computation.
        if scheduler_output.new_block_ids_to_zero:
            self._zero_block_ids(scheduler_output.new_block_ids_to_zero)

1090
        # Free the cached encoder outputs.
1091
1092
        for mm_hash in scheduler_output.free_encoder_mm_hashes:
            self.encoder_cache.pop(mm_hash, None)
1093

1094
1095
1096
1097
1098
1099
1100
        # Remove the unscheduled requests from the persistent batch.
        # NOTE(woosuk): The unscheduled requests are either preempted requests
        # or running requests that are not scheduled in this step. We remove
        # them from the persistent batch but keep their cached states since
        # they will be scheduled again sometime in the future.
        scheduled_req_ids = scheduler_output.num_scheduled_tokens.keys()
        cached_req_ids = self.input_batch.req_id_to_index.keys()
1101
1102
1103
1104
1105
1106
1107
1108
        resumed_req_ids = scheduler_output.scheduled_cached_reqs.resumed_req_ids
        # NOTE(zhuohan): cached_req_ids and resumed_req_ids are usually disjoint,
        # so `(scheduled_req_ids - resumed_req_ids) == scheduled_req_ids` holds
        # apart from the forced-preemption case in reset_prefix_cache. And in
        # that case we include the resumed_req_ids in the unscheduled set so
        # that they get cleared from the persistent batch before being re-scheduled
        # in the normal resumed request path.
        unscheduled_req_ids = cached_req_ids - (scheduled_req_ids - resumed_req_ids)
1109
1110
1111
1112
1113
        # NOTE(woosuk): The persistent batch optimization assumes that
        # consecutive batches contain mostly the same requests. If batches
        # have low request overlap (e.g., alternating between two distinct
        # sets of requests), this optimization becomes very inefficient.
        for req_id in unscheduled_req_ids:
1114
            self.input_batch.remove_request(req_id)
1115

1116
1117
1118
1119
1120
1121
1122
        is_ngram_gpu = (
            self.speculative_config is not None
            and self.speculative_config.use_ngram_gpu()
        )
        if is_ngram_gpu:
            ngram_gpu_new_reqs: list[CachedRequestState] = []

1123
        reqs_to_add: list[CachedRequestState] = []
1124
1125
        deferred_spec_decode_corrections = []

1126
        # Add new requests to the cached states.
1127
1128
        for new_req_data in scheduler_output.scheduled_new_reqs:
            req_id = new_req_data.req_id
1129
1130
1131
1132
1133
1134
            if req_id in self.requests:
                # For streaming case only.
                req_state = self._update_streaming_request(req_id, new_req_data)
                reqs_to_add.append(req_state)
                continue

1135
            sampling_params = new_req_data.sampling_params
1136
            pooling_params = new_req_data.pooling_params
1137

1138
1139
1140
1141
            if (
                sampling_params
                and sampling_params.sampling_type == SamplingType.RANDOM_SEED
            ):
1142
1143
1144
1145
1146
                generator = torch.Generator(device=self.device)
                generator.manual_seed(sampling_params.seed)
            else:
                generator = None

1147
1148
            if self.is_pooling_model:
                assert pooling_params is not None
1149
1150
                task = pooling_params.task
                assert task is not None, "You did not set `task` in the API"
1151

1152
                model = cast(VllmModelForPooling, self.get_model())
1153
                to_update = model.pooler.get_pooling_updates(task)
1154
1155
                to_update.apply(pooling_params)

1156
            req_state = CachedRequestState(
1157
                req_id=req_id,
1158
                prompt_token_ids=new_req_data.prompt_token_ids,
1159
                prompt_embeds=new_req_data.prompt_embeds,
1160
                mm_features=new_req_data.mm_features,
1161
                sampling_params=sampling_params,
1162
                pooling_params=pooling_params,
1163
                generator=generator,
1164
1165
                block_ids=new_req_data.block_ids,
                num_computed_tokens=new_req_data.num_computed_tokens,
1166
                output_token_ids=[],
1167
                lora_request=new_req_data.lora_request,
1168
            )
1169
            self.requests[req_id] = req_state
1170
            self.late_interaction_runner.register_request(req_id, pooling_params)
1171

1172
1173
1174
1175
1176
1177
1178
            if sampling_params and sampling_params.prompt_logprobs is not None:
                self.num_prompt_logprobs[req_id] = (
                    self.input_batch.vocab_size
                    if sampling_params.prompt_logprobs == -1
                    else sampling_params.prompt_logprobs
                )

1179
            # Only relevant for models using M-RoPE (e.g, Qwen2-VL)
1180
            if self.uses_mrope:
1181
                self._init_mrope_positions(req_state)
1182

1183
1184
1185
1186
            # Only relevant for models using XD-RoPE (e.g, HunYuan-VL)
            if self.uses_xdrope_dim > 0:
                self._init_xdrope_positions(req_state)

1187
            reqs_to_add.append(req_state)
1188
1189
1190
            # Track new requests for ngram_gpu full tensor copy
            if is_ngram_gpu:
                ngram_gpu_new_reqs.append(req_state)
1191

1192
        # Update the states of the running/resumed requests.
1193
        is_last_rank = get_pp_group().is_last_rank
1194
        req_data = scheduler_output.scheduled_cached_reqs
1195
        scheduled_spec_tokens = scheduler_output.scheduled_spec_decode_tokens
1196

1197
1198
1199
1200
1201
1202
1203
1204
1205
1206
1207
1208
1209
1210
1211
        # Save scheduler-allocated spec lengths before trimming so
        # prev_num_draft_len keeps the optimistic count for rejection correction.
        original_num_spec_per_req: dict[str, int] = {}
        if (
            self.speculative_config is not None
            and self.speculative_config.use_ngram_gpu()
        ):
            for req_id, toks in scheduled_spec_tokens.items():
                original_num_spec_per_req[req_id] = len(toks)
            update_scheduler_for_invalid_drafts(
                self._num_valid_draft_tokens_event,
                self._num_valid_draft_tokens_cpu,
                scheduler_output,
                self.input_batch.req_id_to_index,
            )
1212
1213
        if self.use_async_spec_decode:
            self.prev_num_draft_tokens.np.fill(0)
1214

1215
        for i, req_id in enumerate(req_data.req_ids):
1216
            req_state = self.requests[req_id]
1217
1218
            num_computed_tokens = req_data.num_computed_tokens[i]
            new_block_ids = req_data.new_block_ids[i]
1219
            resumed_from_preemption = req_id in req_data.resumed_req_ids
1220
            num_output_tokens = req_data.num_output_tokens[i]
1221
            req_index = self.input_batch.req_id_to_index.get(req_id)
1222

1223
1224
1225
1226
            if req_state.prev_num_draft_len and self.use_async_scheduling:
                # prev_num_draft_len is used in async scheduling mode with
                # spec decode. it indicates if need to update num_computed_tokens
                # of the request. for example:
1227
                # first step: num_computed_tokens = 0, spec_tokens = [],
1228
                # prev_num_draft_len = 0.
Jiayi Yan's avatar
Jiayi Yan committed
1229
                # second step: num_computed_tokens = 100(prompt length),
1230
1231
1232
                # spec_tokens = [a,b], prev_num_draft_len = 0.
                # third step: num_computed_tokens = 100 + 2, spec_tokens = [c,d],
                # prev_num_draft_len = 2.
1233
                # num_computed_tokens in first step and second step doesn't contain
1234
1235
1236
                # the spec tokens length, but in third step it contains the
                # spec tokens length. we only need to update num_computed_tokens
                # when prev_num_draft_len > 0.
1237
1238
1239
                if req_index is None:
                    req_state.prev_num_draft_len = 0
                else:
1240
1241
1242
1243
1244
1245
1246
1247
1248
1249
1250
1251
1252
1253
1254
1255
1256
1257
1258
                    # Optimistically assume all accepted; queue up a correction
                    # to be called after the model forward to preserve async
                    # scheduling. Corrected on GPU in _prepare_inputs.
                    optimistic_num_accepted = req_state.prev_num_draft_len
                    req_state.output_token_ids.extend([-1] * optimistic_num_accepted)

                    deferred_spec_decode_corrections.append(
                        (req_id, optimistic_num_accepted, req_state)
                    )

                    prev_req_index = (
                        self.input_batch.prev_req_id_to_index.get(req_id)
                        if self.input_batch.prev_req_id_to_index
                        else None
                    )
                    if prev_req_index is not None:
                        self.prev_num_draft_tokens.np[prev_req_index] = (
                            optimistic_num_accepted
                        )
1259

1260
1261
1262
1263
                    if is_ngram_gpu and optimistic_num_accepted > 0:
                        self.input_batch.num_tokens_no_spec[req_index] += (
                            optimistic_num_accepted
                        )
1264

1265
            # Update the cached states.
1266
            req_state.num_computed_tokens = num_computed_tokens
1267
1268

            if not is_last_rank:
1269
1270
1271
1272
1273
1274
1275
1276
1277
1278
1279
1280
1281
1282
1283
1284
1285
1286
1287
1288
                if not req_data.new_token_ids:
                    # Async scheduled PP: Sampled tokens propagated via GPU broadcast.
                    new_token_ids: list[int] = []
                else:
                    # Non-async scheduling with PP: The scheduler sends
                    # sampled token ids back because there's no direct communication
                    # between the first-stage worker and the last-stage worker.
                    new_token_ids = req_data.new_token_ids[i]
                    # Add the sampled token(s) from the previous step (if any).
                    # This doesn't include "unverified" tokens like spec tokens.
                    num_new_tokens = (
                        num_computed_tokens + len(new_token_ids) - req_state.num_tokens
                    )
                    if num_new_tokens == 1:
                        # Avoid slicing list in most common case.
                        req_state.output_token_ids.append(new_token_ids[-1])
                    elif num_new_tokens > 0:
                        req_state.output_token_ids.extend(
                            new_token_ids[-num_new_tokens:]
                        )
1289
1290
            elif num_output_tokens < len(req_state.output_token_ids):
                # Some output tokens were discarded due to a sync-KV-load
1291
1292
                # failure, or output_token_ids was inflated by the optimistic
                # extend above (async spec decode). Align the cached state.
1293
1294
                del req_state.output_token_ids[num_output_tokens:]
                if req_index is not None:
1295
1296
1297
1298
                    end_idx = (
                        self.input_batch.num_prompt_tokens[req_index]
                        + num_output_tokens
                    )
1299
                    self.input_batch.num_tokens_no_spec[req_index] = end_idx
1300

1301
            # Update the block IDs.
1302
            if not resumed_from_preemption:
1303
1304
                if new_block_ids is not None:
                    # Append the new blocks to the existing block IDs.
1305
                    for block_ids, new_ids in zip(req_state.block_ids, new_block_ids):
1306
                        block_ids.extend(new_ids)
1307
            else:
1308
                assert req_index is None
1309
                assert new_block_ids is not None
1310
1311
                # The request is resumed from preemption.
                # Replace the existing block IDs with the new ones.
1312
                req_state.block_ids = new_block_ids
1313
1314
1315
1316
1317

            if req_index is None:
                # The request is not in the persistent batch.
                # The request was either preempted and resumed later, or was not
                # scheduled in the previous step and needs to be added again.
1318
1319
1320
1321
1322
1323
1324

                if self.use_async_scheduling and num_output_tokens > 0:
                    # We must recover the output token ids for resumed requests in the
                    # async scheduling case, so that correct input_ids are obtained.
                    resumed_token_ids = req_data.all_token_ids[req_id]
                    req_state.output_token_ids = resumed_token_ids[-num_output_tokens:]

1325
                reqs_to_add.append(req_state)
1326
1327
1328
                # Track resumed requests for ngram_gpu full tensor copy
                if is_ngram_gpu:
                    ngram_gpu_new_reqs.append(req_state)
1329
1330
1331
                continue

            # Update the persistent batch.
1332
            self.input_batch.num_computed_tokens_cpu[req_index] = num_computed_tokens
1333
            if new_block_ids is not None:
1334
                self.input_batch.block_table.append_row(new_block_ids, req_index)
1335
1336
1337
1338
1339
1340
1341

            # For the last rank, we don't need to update the token_ids_cpu
            # because the sampled tokens are already cached.
            if not is_last_rank:
                # Add new_token_ids to token_ids_cpu.
                start_token_index = num_computed_tokens
                end_token_index = num_computed_tokens + len(new_token_ids)
1342
                self.input_batch.token_ids_cpu[
1343
1344
1345
                    req_index, start_token_index:end_token_index
                ] = new_token_ids
                self.input_batch.num_tokens_no_spec[req_index] = end_token_index
1346

1347
            # Add spec_token_ids to token_ids_cpu.
1348
            self.input_batch.update_req_spec_token_ids(req_state, scheduled_spec_tokens)
1349
1350
1351
1352
1353
            # Restore scheduler-side draft count after ngram trimming.
            if original_num_spec_per_req:
                orig = original_num_spec_per_req.get(req_id, 0)
                if orig != req_state.prev_num_draft_len:
                    req_state.prev_num_draft_len = orig
1354

1355
1356
        # Add the new or resumed requests to the persistent batch.
        # The smaller empty indices are filled first.
1357
1358
        for request in reqs_to_add:
            self.input_batch.add_request(request)
1359
            self.input_batch.update_req_spec_token_ids(request, scheduled_spec_tokens)
1360

1361
1362
1363
1364
1365
1366
        # Condense the batched states if there are gaps left by removed requests
        self.input_batch.condense()
        # Allow attention backend to reorder the batch, potentially
        self._may_reorder_batch(scheduler_output)
        # Refresh batch metadata with any pending updates.
        self.input_batch.refresh_metadata()
1367

1368
1369
1370
1371
1372
1373
1374
1375
1376
1377
1378
1379
        # Incrementally update ngram_gpu tensors after batch is stable
        if is_ngram_gpu:
            update_ngram_gpu_tensors_incremental(
                self.input_batch,
                self.token_ids_gpu_tensor,
                self.num_tokens_no_spec_gpu,
                ngram_gpu_new_reqs,
                self.device,
                _pinned_idx_buf=self._ngram_pinned_idx_buf,
                _pinned_val_buf=self._ngram_pinned_val_buf,
            )

1380
1381
1382
1383
1384
1385
1386
1387
1388
1389
1390
1391
1392
1393
1394
1395
1396
1397
1398
1399
1400
1401
1402
1403
1404
1405
1406
1407
1408
1409
1410
1411
1412
1413
        if deferred_spec_decode_corrections:

            def correct_spec_decode_token_counts():
                valid_sampled_token_count = self._get_valid_sampled_token_count()
                if not valid_sampled_token_count:
                    return
                prev_req_id_to_index = self.input_batch.prev_req_id_to_index
                if not prev_req_id_to_index:
                    return
                for (
                    req_id,
                    optimistic_num_accepted,
                    req_state,
                ) in deferred_spec_decode_corrections:
                    prev_req_index = prev_req_id_to_index.get(req_id)
                    if prev_req_index is None:
                        continue
                    num_accepted = valid_sampled_token_count[prev_req_index] - 1
                    correction = optimistic_num_accepted - num_accepted
                    req_state.num_computed_tokens -= correction
                    cur_req_index = self.input_batch.req_id_to_index.get(req_id)
                    if cur_req_index is None:
                        continue
                    self.input_batch.num_computed_tokens_cpu[cur_req_index] -= (
                        correction
                    )
                    if is_ngram_gpu and correction > 0:
                        self.input_batch.num_tokens_no_spec[cur_req_index] -= correction
                        self.num_tokens_no_spec_gpu[cur_req_index] -= correction

            return correct_spec_decode_token_counts
        else:
            return None

1414
    def _update_states_after_model_execute(
1415
        self, output_token_ids: torch.Tensor, scheduler_output: "SchedulerOutput"
1416
    ) -> None:
1417
1418
1419
1420
1421
1422
1423
1424
        """Update the cached states after model execution.

        This is used for MTP/EAGLE for hybrid models, as in linear attention,
        only the last token's state is kept. In MTP/EAGLE, for draft tokens
        the state are kept util we decide how many tokens are accepted for
        each sequence, and a shifting is done during the next iteration
        based on the number of accepted tokens.
        """
1425
        if not self.speculative_config or not self.model_config.is_hybrid:
1426
1427
            return

1428
1429
1430
        # TODO: Remove .cpu() sync to enable fully async for hybrid model;
        # Use num_computed_tokens.gpu instead of req.num_computed_tokens to
        # support aligned mamba cache mode.
1431
        # Find the number of accepted tokens for each sequence.
1432
1433
        num_reqs = output_token_ids.size(0)
        self.num_accepted_tokens.gpu[:num_reqs] = (
1434
1435
1436
1437
1438
            (
                torch.cat(
                    [
                        output_token_ids,
                        torch.full(
1439
                            (num_reqs, 1),
1440
1441
1442
1443
1444
1445
1446
1447
1448
1449
1450
                            -1,
                            device=output_token_ids.device,
                        ),
                    ],
                    dim=1,
                )
                == -1
            )
            .int()
            .argmax(-1)
        )
1451

1452
        if self.cache_config.mamba_cache_mode == "align":
1453
1454
1455
1456
            for i, num_tokens in enumerate(
                self.num_accepted_tokens.gpu[:num_reqs].cpu().numpy()
            ):
                self.input_batch.num_accepted_tokens_cpu[i] = num_tokens
1457
1458
1459
1460
1461
1462
1463
1464
            mamba_utils.postprocess_mamba(
                scheduler_output,
                self.kv_cache_config,
                self.input_batch,
                self.requests,
                self.mamba_state_idx,
                self.compilation_config.static_forward_context,
                self.model.get_mamba_state_copy_func(),
1465
                self._get_mamba_copy_bufs(),
1466
            )
1467
1468
1469
1470
        else:
            self.input_batch.num_accepted_tokens_cpu_tensor[:num_reqs].copy_(
                self.num_accepted_tokens.gpu[:num_reqs], non_blocking=True
            )
1471
1472
            assert self.num_accepted_tokens_event is not None
            self.num_accepted_tokens_event.record()
1473

1474
1475
1476
1477
1478
1479
1480
1481
1482
1483
1484
1485
1486
1487
1488
1489
1490
1491
1492
    def _update_streaming_request(
        self, req_id: str, new_req_data: NewRequestData
    ) -> CachedRequestState:
        """Updates streaming session request from `scheduled_new_reqs`.

        Removes the request from InputBatch (if present), updates the cached
        state, and prepares it for re-addition to the batch.

        NOTE: prompt_token_ids includes intermediate output tokens - tokens
        previously generated but now are input context (part of the prompt).
        """
        self.input_batch.remove_request(req_id)
        req_state = self.requests[req_id]

        req_state.prompt_token_ids = new_req_data.prompt_token_ids
        req_state.mm_features = new_req_data.mm_features
        req_state.prompt_embeds = new_req_data.prompt_embeds
        req_state.sampling_params = new_req_data.sampling_params
        req_state.pooling_params = new_req_data.pooling_params
1493
        self.late_interaction_runner.register_request(req_id, req_state.pooling_params)
1494
1495
1496
1497
1498
1499
1500
1501
1502
1503
1504
1505
1506
1507
1508
        req_state.block_ids = new_req_data.block_ids
        req_state.num_computed_tokens = new_req_data.num_computed_tokens
        req_state.num_prompt_tokens = length_from_prompt_token_ids_or_embeds(
            req_state.prompt_token_ids, req_state.prompt_embeds
        )

        # Clear `output_token_ids` as previous output tokens are now part of
        # `prompt_token_ids`.
        req_state.output_token_ids.clear()

        if self.uses_mrope:
            self._init_mrope_positions(req_state)

        return req_state

1509
    def _init_mrope_positions(self, req_state: CachedRequestState):
1510
1511
        model = self.get_model()
        assert supports_mrope(model), "M-RoPE support is not implemented."
1512
1513
1514
1515
        assert req_state.prompt_token_ids is not None, (
            "M-RoPE requires prompt_token_ids to be available."
        )
        mrope_model = cast(SupportsMRoPE, model)
1516
1517

        req_state.mrope_positions, req_state.mrope_position_delta = (
1518
            mrope_model.get_mrope_input_positions(
1519
                req_state.prompt_token_ids,
1520
                req_state.mm_features,
1521
            )
1522
        )
1523

1524
1525
1526
1527
1528
1529
1530
1531
1532
1533
1534
1535
1536
    def _init_xdrope_positions(self, req_state: CachedRequestState):
        model = self.get_model()
        xdrope_model = cast(SupportsXDRoPE, model)
        assert req_state.prompt_token_ids is not None, (
            "XD-RoPE requires prompt_token_ids to be available."
        )
        assert supports_xdrope(model), "XD-RoPE support is not implemented."

        req_state.xdrope_positions = xdrope_model.get_xdrope_input_positions(
            req_state.prompt_token_ids,
            req_state.mm_features,
        )

1537
    def _extract_mm_kwargs(
1538
        self,
1539
1540
        scheduler_output: "SchedulerOutput",
    ) -> BatchedTensorInputs:
1541
        if not scheduler_output or not self.is_multimodal_raw_input_only_model:
1542
            return {}
1543

1544
        mm_kwargs = list[tuple[str, MultiModalKwargsItem]]()
1545
        for req in scheduler_output.scheduled_new_reqs:
1546
1547
            for feature in req.mm_features:
                if feature.data is not None:
1548
                    mm_kwargs.append((feature.modality, feature.data))
1549

1550
1551
        # Input all modalities at once
        mm_kwargs_combined: BatchedTensorInputs = {}
1552
        for _, _, mm_kwargs_batch in group_and_batch_mm_kwargs(
1553
1554
1555
            mm_kwargs,
            device=self.device,
            pin_memory=self.pin_memory,
1556
        ):
1557
            mm_kwargs_combined.update(mm_kwargs_batch)
1558

1559
        return mm_kwargs_combined
1560

1561
    def _dummy_mm_kwargs(self, num_seqs: int) -> BatchedTensorInputs:
1562
        if not self.is_multimodal_raw_input_only_model:
1563
            return {}
1564

1565
1566
1567
        mm_budget = self.mm_budget
        assert mm_budget is not None

1568
1569
1570
        if not mm_budget.mm_max_toks_per_item:
            return {}  # No tower modalities (embed-only mode)

1571
1572
        dummy_modality = mm_budget.get_modality_with_max_tokens()
        return self._get_mm_dummy_batch(dummy_modality, num_seqs)
1573

1574
1575
1576
    def _get_cumsum_and_arange(
        self,
        num_tokens: np.ndarray,
1577
        arange_out: np.ndarray,
1578
        cumsum_dtype: np.dtype | None = None,
1579
    ) -> np.ndarray:
1580
        """Get the cumulative sum and batched arange of the given array.
1581
1582
1583
1584
        E.g., [2, 5, 3] -> [2, 7, 10], arange written to
        arange_out[:10] as [0, 1, 0, 1, 2, 3, 4, 0, 1, 2].
        Equivalent to but faster than:
        np.concatenate([np.arange(n) for n in num_tokens])
1585
1586
1587
1588
1589
1590
1591
        """
        # Step 1. [2, 5, 3] -> [2, 7, 10]
        cu_num_tokens = np.cumsum(num_tokens, dtype=cumsum_dtype)
        total_num_tokens = cu_num_tokens[-1]
        # Step 2. [2, 7, 10] -> [0, 0, 2, 2, 2, 2, 2, 7, 7, 7]
        cumsums_offsets = np.repeat(cu_num_tokens - num_tokens, num_tokens)
        # Step 3. [0, 1, 0, 1, 2, 3, 4, 0, 1, 2]
1592
1593
1594
1595
1596
1597
1598
1599
1600
1601
1602
1603
1604
1605
1606
1607
1608
1609
1610
        np.subtract(
            self.arange_np[:total_num_tokens],
            cumsums_offsets,
            out=arange_out[:total_num_tokens],
        )

        return cu_num_tokens

    def _compute_prev_positions(self, num_reqs: int) -> None:
        """Build prev_positions mapping: current pos -> previous pos (-1 if new).

        Populates self.prev_positions.np[:num_reqs] with the mapping.
        """
        prev_req_id_to_index = self.input_batch.prev_req_id_to_index
        prev_positions = self.prev_positions.np[:num_reqs]

        if not prev_req_id_to_index:
            prev_positions.fill(-1)
            return
1611

1612
1613
        for i, req_id in enumerate(self.input_batch.req_ids[:num_reqs]):
            prev_positions[i] = prev_req_id_to_index.get(req_id, -1)
1614

1615
    def _prepare_input_ids(
1616
1617
        self,
        scheduler_output: "SchedulerOutput",
1618
        num_reqs: int,
1619
1620
        total_num_scheduled_tokens: int,
        cu_num_tokens: np.ndarray,
1621
    ) -> None:
1622
        """Prepare the input IDs for the current batch.
1623

1624
1625
        Carefully handles the `prev_sampled_token_ids` which can be cached
        from the previous engine iteration, in which case those tokens on the
1626
1627
1628
1629
1630
        GPU need to be copied into the corresponding slots into input_ids.

        Uses self.prev_positions[:num_reqs] which maps current pos -> prev pos
        (-1 for new requests).
        """
1631
1632
1633
1634

        if self.input_batch.prev_sampled_token_ids is None:
            # Normal scheduling case
            self.input_ids.copy_to_gpu(total_num_scheduled_tokens)
1635
1636
1637
            if self.enable_prompt_embeds:
                self.inputs_embeds.copy_to_gpu(total_num_scheduled_tokens)
                self.is_token_ids.copy_to_gpu(total_num_scheduled_tokens)
1638
1639
1640
1641
1642
            return

        # Async scheduling case, where some decode requests from the previous
        # iteration won't have entries in input_ids_cpu and need to be copied
        # on the GPU from prev_sampled_token_ids.
1643
1644
        prev_positions = self.prev_positions.np[:num_reqs]
        scheduled_spec_tokens = scheduler_output.scheduled_spec_decode_tokens
1645
1646
1647
        sample_flattened_indices: list[int] = []
        spec_flattened_indices: list[int] = []
        prev_draft_token_indices: list[int] = []
1648
1649
        prev_indices: list[int] = []
        common_indices_match = True
1650
        max_flattened_index = -1
1651
1652
        total_num_spec_tokens = 0

1653
1654
1655
1656
1657
1658
1659
1660
1661
1662
1663
1664
1665
1666
1667
1668
1669
1670
1671
1672
1673
1674
1675
1676
1677
1678
1679
1680
1681
        for cur_index in range(num_reqs):
            prev_index = prev_positions[cur_index]
            if prev_index < 0:
                continue
            prev_indices.append(prev_index)
            req_id = self.input_batch.req_ids[cur_index]
            # We need to compute the flattened input_ids index of the
            # last token in each common request.
            draft_len = len(scheduled_spec_tokens.get(req_id, ()))
            total_num_spec_tokens += draft_len
            flattened_index = cu_num_tokens[cur_index].item() - 1
            # example: cu_num_tokens = [2, 5, 8], draft_tokens = [1, 2, 2]
            # sample_flattened_indices = [0, 2, 5]
            # spec_flattened_indices = [1,   3, 4,    6, 7]
            sample_flattened_indices.append(flattened_index - draft_len)
            spec_flattened_indices.extend(
                range(flattened_index - draft_len + 1, flattened_index + 1)
            )
            start = prev_index * self.num_spec_tokens
            # prev_draft_token_indices is used to find which draft_tokens_id
            # should be copied to input_ids
            # example: prev draft_tokens_id [[1,2], [3,4], [5, 6]]
            # flatten draft_tokens_id [1,2,3,4,5,6]
            # draft_len of each request [1, 2, 1]
            # then prev_draft_token_indices is [0,   2, 3,   4]
            prev_draft_token_indices.extend(range(start, start + draft_len))
            common_indices_match &= prev_index == flattened_index
            max_flattened_index = max(max_flattened_index, flattened_index)

Jiayi Yan's avatar
Jiayi Yan committed
1682
        num_common_tokens = len(sample_flattened_indices)
1683
        total_without_spec = total_num_scheduled_tokens - total_num_spec_tokens
Jiayi Yan's avatar
Jiayi Yan committed
1684
        if num_common_tokens < total_without_spec:
1685
            # If not all requests are decodes from the last iteration,
1686
            # we need to copy the input_ids_cpu to the GPU first.
1687
            self.input_ids.copy_to_gpu(total_num_scheduled_tokens)
1688
1689
1690
            if self.enable_prompt_embeds:
                self.inputs_embeds.copy_to_gpu(total_num_scheduled_tokens)
                self.is_token_ids.copy_to_gpu(total_num_scheduled_tokens)
Jiayi Yan's avatar
Jiayi Yan committed
1691
        if num_common_tokens == 0:
1692
            # No requests in common with the previous iteration
1693
            # So input_ids.cpu will have all the input ids.
1694
            return
1695
        if common_indices_match and max_flattened_index == (num_common_tokens - 1):
1696
1697
1698
1699
            # Common-case optimization: the batch is unchanged
            # and no reordering happened.
            # The indices are both the same permutation of 0..N-1 so
            # we can copy directly using a single slice.
Jiayi Yan's avatar
Jiayi Yan committed
1700
1701
            self.input_ids.gpu[:num_common_tokens].copy_(
                self.input_batch.prev_sampled_token_ids[:num_common_tokens, 0],
1702
1703
                non_blocking=True,
            )
1704
            if self.enable_prompt_embeds:
Jiayi Yan's avatar
Jiayi Yan committed
1705
                self.is_token_ids.gpu[:num_common_tokens] = True
1706
            return
1707
        # Upload the index tensors asynchronously so the scatter can be non-blocking.
1708
1709
        sampled_tokens_index_tensor = torch.tensor(
            sample_flattened_indices, dtype=torch.int64, pin_memory=self.pin_memory
1710
        ).to(self.device, non_blocking=True)
1711
        prev_common_req_indices_tensor = torch.tensor(
1712
            prev_indices, dtype=torch.int64, pin_memory=self.pin_memory
1713
        ).to(self.device, non_blocking=True)
1714
1715
        self.input_ids.gpu.scatter_(
            dim=0,
1716
            index=sampled_tokens_index_tensor,
1717
            src=self.input_batch.prev_sampled_token_ids[
1718
1719
1720
                prev_common_req_indices_tensor, 0
            ],
        )
1721

1722
1723
1724
1725
1726
1727
1728
1729
1730
1731
1732
1733
1734
1735
1736
1737
1738
1739
1740
1741
1742
1743
        # Scatter the draft tokens after the sampled tokens are scattered.
        if self._draft_token_ids is None or not spec_flattened_indices:
            return

        assert isinstance(self._draft_token_ids, torch.Tensor)
        draft_tokens_index_tensor = torch.tensor(
            spec_flattened_indices, dtype=torch.int64, pin_memory=self.pin_memory
        ).to(self.device, non_blocking=True)
        prev_draft_token_indices_tensor = torch.tensor(
            prev_draft_token_indices, dtype=torch.int64, pin_memory=self.pin_memory
        ).to(self.device, non_blocking=True)

        # because input_ids dtype is torch.int32,
        # so convert draft_token_ids to torch.int32 here.
        draft_token_ids = self._draft_token_ids.to(dtype=torch.int32)

        self.input_ids.gpu.scatter_(
            dim=0,
            index=draft_tokens_index_tensor,
            src=draft_token_ids.flatten()[prev_draft_token_indices_tensor],
        )

1744
1745
    def _get_encoder_seq_lens(
        self,
1746
        num_scheduled_tokens: dict[str, int],
1747
1748
        kv_cache_spec: KVCacheSpec,
        num_reqs: int,
1749
        for_cudagraph_capture: bool = False,
1750
    ) -> tuple[torch.Tensor | None, np.ndarray | None]:
1751
        if not isinstance(kv_cache_spec, CrossAttentionSpec):
1752
            return None, None
1753

1754
1755
        # Zero out buffer for padding requests that are not actually scheduled (CGs)
        self.encoder_seq_lens.np[:num_reqs] = 0
1756

1757
1758
        # Build encoder_seq_lens array mapping request indices to
        # encoder lengths for inputs scheduled in this batch
1759
        for req_id in num_scheduled_tokens:
1760
            req_index = self.input_batch.req_id_to_index[req_id]
1761
1762
1763
1764
1765
1766
1767
1768
1769
1770
1771
1772
            req_state = self.requests[req_id]
            if req_state.mm_features is None:
                self.encoder_seq_lens.np[req_index] = 0
                continue

            # Get the total number of encoder input tokens for running encoder requests
            # whether encoding is finished or not so that cross-attention knows how
            # many encoder tokens to attend to.
            encoder_input_tokens = sum(
                feature.mm_position.length for feature in req_state.mm_features
            )
            self.encoder_seq_lens.np[req_index] = encoder_input_tokens
1773
1774
1775
1776
1777
1778
1779
1780
1781
        if for_cudagraph_capture:
            # During CUDA graph capture, we need to use realistic encoder lengths
            # so that max_seqlen_k is captured with the correct value.
            max_encoder_len = getattr(
                self.model_config.hf_config,
                "max_source_positions",
                self.max_encoder_len,
            )
            self.encoder_seq_lens.np[:num_reqs] = max_encoder_len
1782
1783
1784
1785

        self.encoder_seq_lens.copy_to_gpu(num_reqs)
        encoder_seq_lens = self.encoder_seq_lens.gpu[:num_reqs]
        encoder_seq_lens_cpu = self.encoder_seq_lens.np[:num_reqs]
1786

1787
        return encoder_seq_lens, encoder_seq_lens_cpu
1788

1789
    def _prepare_inputs(
1790
1791
1792
        self,
        scheduler_output: "SchedulerOutput",
        num_scheduled_tokens: np.ndarray,
1793
1794
    ) -> tuple[
        torch.Tensor,
1795
        SpecDecodeMetadata | None,
1796
    ]:
1797
1798
        """
        :return: tuple[
1799
            logits_indices, spec_decode_metadata,
1800
1801
        ]
        """
1802
1803
1804
1805
1806
1807
1808
        total_num_scheduled_tokens = scheduler_output.total_num_scheduled_tokens
        assert total_num_scheduled_tokens > 0
        num_reqs = self.input_batch.num_reqs
        assert num_reqs > 0

        # OPTIMIZATION: Start copying the block table first.
        # This way, we can overlap the copy with the following CPU operations.
1809
        self.input_batch.block_table.commit_block_table(num_reqs)
1810
1811
1812

        # Get request indices.
        # E.g., [2, 5, 3] -> [0, 0, 1, 1, 1, 1, 1, 2, 2, 2]
1813
        req_indices = np.repeat(self.arange_np[:num_reqs], num_scheduled_tokens)
1814

1815
        # cu_num_tokens: [2, 5, 3] -> [2, 7, 10]
1816
1817
1818
1819
        # self.query_pos.np[:10]: [0, 1, 0, 1, 2, 3, 4, 0, 1, 2]
        cu_num_tokens = self._get_cumsum_and_arange(
            num_scheduled_tokens, self.query_pos.np
        )
1820
1821

        # Get positions.
1822
1823
1824
        positions_np = (
            self.input_batch.num_computed_tokens_cpu[req_indices]
            + self.query_pos.np[: cu_num_tokens[-1]]
1825
        )
1826

1827
1828
        # Calculate M-RoPE positions.
        # Only relevant for models using M-RoPE (e.g, Qwen2-VL)
1829
        if self.uses_mrope:
1830
1831
            self._calc_mrope_positions(scheduler_output)

1832
1833
1834
1835
1836
        # Calculate XD-RoPE positions.
        # Only relevant for models using XD-RoPE (e.g, HunYuan-VL)
        if self.uses_xdrope_dim > 0:
            self._calc_xdrope_positions(scheduler_output)

1837
1838
1839
1840
        # Get token indices.
        # E.g., [0, 1, 0, 1, 2, 3, 4, 0, 1, 2]
        # -> [0, 1, M, M + 1, M + 2, M + 3, M + 4, 2 * M, 2 * M + 1, 2 * M + 2]
        # where M is the max_model_len.
1841
1842
1843
        token_indices = (
            positions_np + req_indices * self.input_batch.token_ids_cpu.shape[1]
        )
1844
        token_indices_tensor = torch.from_numpy(token_indices)
1845

1846
1847
1848
        # NOTE(woosuk): We use torch.index_select instead of np.take here
        # because torch.index_select is much faster than np.take for large
        # tensors.
1849
1850
1851
1852
1853
1854
        torch.index_select(
            self.input_batch.token_ids_cpu_tensor.flatten(),
            0,
            token_indices_tensor,
            out=self.input_ids.cpu[:total_num_scheduled_tokens],
        )
1855
        if self.enable_prompt_embeds:
1856
            is_token_ids = self.input_batch.is_token_ids_tensor.flatten()
1857
1858
1859
1860
            torch.index_select(
                is_token_ids,
                0,
                token_indices_tensor,
1861
1862
                out=self.is_token_ids.cpu[:total_num_scheduled_tokens],
            )
1863
1864
1865
1866
1867
1868
1869
1870
1871
1872
1873
1874
1875
1876
1877
1878
1879
1880
1881
1882
1883
1884
1885
1886
1887
1888
1889
1890
1891
1892
1893
1894
1895

        # Because we did not pre-allocate a massive prompt_embeds CPU tensor on
        # the InputBatch, we need to fill in the prompt embeds into the expected
        # spots in the GpuModelRunner's pre-allocated prompt_embeds tensor.
        if self.input_batch.req_prompt_embeds:
            output_idx = 0
            for req_idx in range(num_reqs):
                num_sched = num_scheduled_tokens[req_idx]

                # Skip if this request doesn't have embeddings
                if req_idx not in self.input_batch.req_prompt_embeds:
                    output_idx += num_sched
                    continue

                # Skip if no tokens scheduled
                if num_sched <= 0:
                    output_idx += num_sched
                    continue

                req_embeds = self.input_batch.req_prompt_embeds[req_idx]
                start_pos = self.input_batch.num_computed_tokens_cpu[req_idx]

                # Skip if trying to read beyond available embeddings
                if start_pos >= req_embeds.shape[0]:
                    output_idx += num_sched
                    continue

                # Copy available embeddings
                end_pos = start_pos + num_sched
                actual_end = min(end_pos, req_embeds.shape[0])
                actual_num_sched = actual_end - start_pos

                if actual_num_sched > 0:
1896
1897
1898
                    self.inputs_embeds.cpu[
                        output_idx : output_idx + actual_num_sched
                    ].copy_(req_embeds[start_pos:actual_end])
1899
1900

                output_idx += num_sched
1901
1902

        # Prepare the attention metadata.
1903
        self.query_start_loc.np[0] = 0
1904
        self.query_start_loc.np[1 : num_reqs + 1] = cu_num_tokens
1905
1906
        # Note: pad query_start_loc to be non-decreasing, as kernels
        # like FlashAttention requires that
1907
        self.query_start_loc.np[num_reqs + 1 :].fill(cu_num_tokens[-1])
1908
        self.query_start_loc.copy_to_gpu()
1909
        query_start_loc = self.query_start_loc.gpu[: num_reqs + 1]
1910

1911
1912
1913
1914
1915
1916
1917
1918
        # Compute optimistic seq_lens (assumes all draft tokens from previous
        # iteration accepted). Store in optimistic_seq_lens_cpu for use by
        # _build_attention_metadata (max_seq_len) and discard_request_mask.
        # seq_lens (GPU) will be computed later using the same optimistic values.
        torch.add(
            self.input_batch.num_computed_tokens_cpu_tensor[:num_reqs],
            torch.from_numpy(num_scheduled_tokens),
            out=self.optimistic_seq_lens_cpu[:num_reqs],
1919
        )
1920
1921
1922
1923
1924
1925
        self.optimistic_seq_lens_cpu[num_reqs:].fill_(0)

        # Build prev_positions mapping: current pos -> prev pos (-1 if new).
        # Used for gathering from previous iteration's GPU tensors.
        prev_req_id_to_index = self.input_batch.prev_req_id_to_index
        self._compute_prev_positions(num_reqs)
1926

1927
        num_tokens = [self.requests[r].num_tokens for r in self.input_batch.req_ids]
1928
1929
        num_tokens_np = np.array(num_tokens, dtype=np.int32)

1930
        # Record which requests should not be sampled,
1931
        # so that we could clear the sampled tokens before returning
1932
        self.discard_request_mask.np[:num_reqs] = (
1933
            self.optimistic_seq_lens_cpu[:num_reqs].numpy() < num_tokens_np
1934
        )
1935
        self.discard_request_mask.copy_to_gpu(num_reqs)
1936

1937
1938
1939
1940
        # Sync num_accepted_tokens from CPU (set by
        # _update_states_after_model_execute for hybrid models).
        if self.num_accepted_tokens_event is not None:
            self.num_accepted_tokens_event.synchronize()
1941
1942
1943
1944
1945
1946
1947
1948
1949
1950
1951
1952
1953
1954
1955
1956
1957
1958
            # Async mode: condense() reordered indices, use prev_positions mapping
            if self.use_async_scheduling and prev_req_id_to_index:
                prev_idx = self.prev_positions.np[:num_reqs]
                new_mask = prev_idx < 0
                self.num_accepted_tokens.np[:num_reqs] = (
                    self.input_batch.num_accepted_tokens_cpu[
                        np.where(new_mask, 0, prev_idx)
                    ]
                )
                self.num_accepted_tokens.np[:num_reqs][new_mask] = 1
                self.input_batch.num_accepted_tokens_cpu[:num_reqs] = (
                    self.num_accepted_tokens.np[:num_reqs]
                )
            else:
                # Non-async mode: use values directly
                self.num_accepted_tokens.np[:num_reqs] = (
                    self.input_batch.num_accepted_tokens_cpu[:num_reqs]
                )
1959
1960
1961
1962
1963
1964
1965
1966
1967
1968
1969
1970
1971
1972
1973
1974
1975
1976
1977
1978
1979
1980
1981
1982
1983
1984
1985
1986
1987
1988
1989
1990
1991
1992
1993
1994
1995
1996
1997
1998
1999
2000
2001
2002
2003
2004
2005
2006
2007
2008
2009
2010
2011
2012
2013
2014
2015
            self.num_accepted_tokens.np[num_reqs:].fill(1)
            self.num_accepted_tokens.copy_to_gpu()
        else:
            self.num_accepted_tokens.np.fill(1)
            self.num_accepted_tokens.gpu.fill_(1)

        # Update num_computed_tokens on GPU. In async spec decode,
        # CPU values are optimistic (all drafts accepted). The kernel
        # corrects on GPU using the previous step's
        # valid_sampled_token_count_gpu. Otherwise, just copy from CPU.
        if (
            self.use_async_spec_decode
            and self.valid_sampled_token_count_gpu is not None
            and prev_req_id_to_index
        ):
            self.prev_positions.copy_to_gpu(num_reqs)
            self.prev_num_draft_tokens.copy_to_gpu()
            cpu_values = self.input_batch.num_computed_tokens_cpu_tensor[:num_reqs].to(
                device=self.device, non_blocking=True
            )
            update_num_computed_tokens_for_batch_change(
                self.num_computed_tokens,
                self.num_accepted_tokens.gpu[:num_reqs],
                self.prev_positions.gpu[:num_reqs],
                self.valid_sampled_token_count_gpu,
                self.prev_num_draft_tokens.gpu,
                cpu_values,
            )
        else:
            self.num_computed_tokens[:num_reqs].copy_(
                self.input_batch.num_computed_tokens_cpu_tensor[:num_reqs],
                non_blocking=True,
            )

        self.req_indices.np[:total_num_scheduled_tokens] = req_indices
        self.req_indices.copy_to_gpu(total_num_scheduled_tokens)
        req_indices_gpu = self.req_indices.gpu[:total_num_scheduled_tokens]

        self.query_pos.copy_to_gpu(total_num_scheduled_tokens)
        self.num_scheduled_tokens.np[:num_reqs] = num_scheduled_tokens
        self.num_scheduled_tokens.copy_to_gpu(num_reqs)
        num_scheduled_tokens_gpu = self.num_scheduled_tokens.gpu[:num_reqs]
        self.positions[:total_num_scheduled_tokens] = (
            self.num_computed_tokens[req_indices_gpu].to(torch.int64)
            + self.query_pos.gpu[:total_num_scheduled_tokens]
        )
        self.seq_lens[:num_reqs] = (
            self.num_computed_tokens[:num_reqs] + num_scheduled_tokens_gpu
        )
        self.seq_lens[num_reqs:].fill_(0)

        self.input_batch.block_table.compute_slot_mapping(
            num_reqs,
            self.query_start_loc.gpu[: num_reqs + 1],
            self.positions[:total_num_scheduled_tokens],
        )

2016
        # Copy the tensors to the GPU.
2017
2018
        self._prepare_input_ids(
            scheduler_output,
2019
            num_reqs,
2020
2021
2022
            total_num_scheduled_tokens,
            cu_num_tokens,
        )
2023

2024
        if self.uses_mrope:
2025
            # Only relevant for models using M-RoPE (e.g, Qwen2-VL)
2026
2027
            self.mrope_positions.gpu[:, :total_num_scheduled_tokens].copy_(
                self.mrope_positions.cpu[:, :total_num_scheduled_tokens],
2028
2029
                non_blocking=True,
            )
2030
2031
2032
2033
2034
2035
        elif self.uses_xdrope_dim > 0:
            # Only relevant for models using XD-RoPE (e.g, HunYuan-VL)
            self.xdrope_positions.gpu[:, :total_num_scheduled_tokens].copy_(
                self.xdrope_positions.cpu[:, :total_num_scheduled_tokens],
                non_blocking=True,
            )
2036
2037
2038
2039
2040
2041
2042
2043
        if self.use_async_spec_decode and (self.uses_mrope or self.uses_xdrope_dim > 0):
            drift = self.num_computed_tokens[req_indices_gpu].to(
                torch.int64
            ) - self.input_batch.num_computed_tokens_cpu_tensor[req_indices].to(
                device=self.device, dtype=torch.int64, non_blocking=True
            )
            target = self.mrope_positions if self.uses_mrope else self.xdrope_positions
            target.gpu[:, :total_num_scheduled_tokens] += drift
2044

2045
        use_spec_decode = len(scheduler_output.scheduled_spec_decode_tokens) > 0
2046
2047
2048
2049
2050
2051
2052
2053
        if not use_spec_decode:
            # NOTE(woosuk): Due to chunked prefills, the batch may contain
            # partial requests. While we should not sample any token
            # from these partial requests, we do so for simplicity.
            # We will ignore the sampled tokens from the partial requests.
            # TODO: Support prompt logprobs.
            logits_indices = query_start_loc[1:] - 1
            spec_decode_metadata = None
2054
            num_sampled_tokens = np.ones(num_reqs, dtype=np.int32)
2055
2056
2057
2058
2059
        else:
            # Get the number of draft tokens for each request.
            # Iterate over the dictionary rather than all requests since not all
            # requests have draft tokens.
            num_draft_tokens = np.zeros(num_reqs, dtype=np.int32)
2060
2061
2062
            # For chunked prefills, use -1 as mask rather than 0, as guided
            # decoding may rollback speculative tokens.
            num_decode_draft_tokens = np.full(num_reqs, -1, dtype=np.int32)
2063
2064
2065
2066
            for (
                req_id,
                draft_token_ids,
            ) in scheduler_output.scheduled_spec_decode_tokens.items():
2067
                req_idx = self.input_batch.req_id_to_index[req_id]
2068
2069
                draft_len = len(draft_token_ids)
                num_draft_tokens[req_idx] = draft_len
2070
2071
2072
2073
                if (
                    self.input_batch.num_computed_tokens_cpu[req_idx]
                    >= self.input_batch.num_prompt_tokens[req_idx]
                ):
2074
                    num_decode_draft_tokens[req_idx] = draft_len
2075
            spec_decode_metadata = self._calc_spec_decode_metadata(
2076
2077
                num_draft_tokens, cu_num_tokens
            )
2078
            logits_indices = spec_decode_metadata.logits_indices
2079
            num_sampled_tokens = num_draft_tokens + 1
2080
            # For DECODE only cuda graph of some attention backends (e.g., GDN).
2081
            self.num_decode_draft_tokens.np[:num_reqs] = num_decode_draft_tokens
2082
2083
            self.num_decode_draft_tokens.np[num_reqs:].fill(-1)
            self.num_decode_draft_tokens.copy_to_gpu()
2084

2085
2086
2087
2088
2089
        # Hot-Swap lora model
        if self.lora_config:
            assert (
                np.sum(num_sampled_tokens)
                <= self.vllm_config.scheduler_config.max_num_batched_tokens
2090
            )
2091
2092
2093
2094
2095
2096
2097
2098
2099
2100
2101
            self.set_active_loras(
                self.input_batch, num_scheduled_tokens, num_sampled_tokens
            )

        return (
            logits_indices,
            spec_decode_metadata,
        )

    def _build_attention_metadata(
        self,
2102
        num_tokens: int,
2103
        num_reqs: int,
2104
2105
2106
        max_query_len: int,
        num_tokens_padded: int | None = None,
        num_reqs_padded: int | None = None,
2107
2108
2109
2110
        ubatch_slices: UBatchSlices | None = None,
        logits_indices: torch.Tensor | None = None,
        use_spec_decode: bool = False,
        for_cudagraph_capture: bool = False,
2111
        num_scheduled_tokens: dict[str, int] | None = None,
2112
        cascade_attn_prefix_lens: list[list[int]] | None = None,
2113
        slot_mappings: dict[int, torch.Tensor] | None = None,
2114
2115
2116
2117
    ) -> tuple[PerLayerAttnMetadata, CommonAttentionMetadata | None]:
        """
        :return: tuple[attn_metadata, spec_decode_common_attn_metadata]
        """
2118
2119
2120
2121
        # Attention metadata is not needed for attention free models
        if len(self.kv_cache_config.kv_cache_groups) == 0:
            return {}, None

2122
2123
        num_tokens_padded = num_tokens_padded or num_tokens
        num_reqs_padded = num_reqs_padded or num_reqs
2124
        assert num_reqs_padded is not None and num_tokens_padded is not None
2125

2126
2127
2128
        attn_metadata: PerLayerAttnMetadata = {}
        if ubatch_slices is not None:
            attn_metadata = [dict() for _ in range(len(ubatch_slices))]
2129

2130
2131
2132
2133
2134
2135
        if for_cudagraph_capture:
            # For some attention backends (e.g. FA) with sliding window models we need
            # to make sure the backend see a max_seq_len that is larger to the sliding
            # window size when capturing to make sure the correct kernel is selected.
            max_seq_len = self.max_model_len
        else:
2136
            max_seq_len = self.optimistic_seq_lens_cpu.numpy()[:num_reqs].max().item()
2137

2138
        kv_cache_groups = self.kv_cache_config.kv_cache_groups
2139

2140
        def _get_block_table(kv_cache_gid: int):
2141
2142
2143
            assert num_reqs_padded is not None and num_tokens_padded is not None
            kv_cache_spec = kv_cache_groups[kv_cache_gid].kv_cache_spec
            if isinstance(kv_cache_spec, EncoderOnlyAttentionSpec):
2144
                blk_table_tensor = torch.zeros(
2145
                    (num_reqs_padded, 1),
2146
                    dtype=torch.int32,
2147
2148
                    device=self.device,
                )
2149
            else:
2150
                blk_table = self.input_batch.block_table[kv_cache_gid]
2151
                blk_table_tensor = blk_table.get_device_tensor(num_reqs_padded)
2152

2153
2154
2155
            # Fill unused block table entries with NULL_BLOCK_ID (null block)
            # for CUDAGraph padding. Block 0 is reserved for padding.
            blk_table_tensor[num_reqs:num_reqs_padded].fill_(NULL_BLOCK_ID)
2156
            return blk_table_tensor
2157

2158
2159
2160
        assert slot_mappings is not None
        block_table_gid_0 = _get_block_table(0)
        slot_mapping_gid_0 = slot_mappings[0]
2161

2162
2163
2164
2165
        if self.routed_experts_initialized:
            attn_gid = self.routed_experts_attn_gid
            slot_mapping_attn = slot_mappings[attn_gid]
            self.slot_mapping = slot_mapping_attn[:num_tokens].cpu().numpy()
2166
2167
2168
2169
2170
2171
        num_computed_tokens_cpu = self.input_batch.num_computed_tokens_cpu_tensor[
            :num_reqs_padded
        ]
        num_prompt_tokens_cpu = self.input_batch.num_prompt_tokens_cpu_tensor[
            :num_reqs_padded
        ]
2172
2173
2174
2175
2176
        seq_lens_cpu = self.optimistic_seq_lens_cpu[:num_reqs_padded]

        # is_prefilling: True if request is still in prefill phase.
        # Used by mamba backends to distinguish actual decodes from
        # short extends.
2177
2178
        is_prefilling = num_computed_tokens_cpu < num_prompt_tokens_cpu

2179
2180
2181
2182
2183
        if self.use_async_spec_decode:
            # GPU tensors are authoritative in async mode.
            seq_lens_cpu = None
            num_computed_tokens_cpu = None

2184
2185
2186
        cm_base = CommonAttentionMetadata(
            query_start_loc=self.query_start_loc.gpu[: num_reqs_padded + 1],
            query_start_loc_cpu=self.query_start_loc.cpu[: num_reqs_padded + 1],
2187
2188
            seq_lens=self.seq_lens[:num_reqs_padded],
            _seq_lens_cpu=seq_lens_cpu,
2189
            _num_computed_tokens_cpu=num_computed_tokens_cpu,
2190
2191
2192
2193
2194
2195
2196
            num_reqs=num_reqs_padded,
            num_actual_tokens=num_tokens_padded,
            max_query_len=max_query_len,
            max_seq_len=max_seq_len,
            block_table_tensor=block_table_gid_0,
            slot_mapping=slot_mapping_gid_0,
            causal=True,
2197
            is_prefilling=is_prefilling,
2198
2199
2200
2201
        )

        if self.dcp_world_size > 1:
            self.dcp_local_seq_lens.cpu[:num_reqs] = get_dcp_local_seq_lens(
2202
                self.optimistic_seq_lens_cpu[:num_reqs],
2203
2204
2205
2206
2207
2208
2209
2210
2211
2212
2213
2214
2215
2216
2217
2218
2219
2220
                self.dcp_world_size,
                self.dcp_rank,
                self.parallel_config.cp_kv_cache_interleave_size,
            )
            self.dcp_local_seq_lens.cpu[num_reqs:].fill_(0)
            self.dcp_local_seq_lens.copy_to_gpu(num_reqs_padded)

            cm_base.dcp_local_seq_lens = self.dcp_local_seq_lens.gpu[:num_reqs_padded]
            cm_base.dcp_local_seq_lens_cpu = self.dcp_local_seq_lens.cpu[
                :num_reqs_padded
            ]

        if logits_indices is not None and self.cache_config.kv_sharing_fast_prefill:
            cm_base.num_logits_indices = logits_indices.size(0)
            cm_base.logits_indices_padded = self._prepare_kv_sharing_fast_prefill(
                logits_indices
            )

2221
2222
2223
2224
2225
2226
2227
2228
2229
        # Cache attention metadata builds across hybrid KV-cache groups
        # The only thing that changes between different hybrid KV-cache groups when the
        # same metadata builder and KVCacheSpec is the same is the block table, so we
        # can cache the attention metadata builds and just update the block table using
        # `builder.update_block_table` if the builder supports it.
        cached_attn_metadata: dict[
            tuple[KVCacheSpec, type[AttentionMetadataBuilder]], AttentionMetadata
        ] = {}

2230
2231
2232
2233
2234
2235
2236
        def _build_attn_group_metadata(
            kv_cache_gid: int,
            attn_gid: int,
            common_attn_metadata: CommonAttentionMetadata,
            ubid: int | None = None,
        ) -> None:
            attn_group = self.attn_groups[kv_cache_gid][attn_gid]
2237
            builder = attn_group.get_metadata_builder(ubid or 0)
2238
2239
2240
2241
            kv_cache_spec = kv_cache_groups[kv_cache_gid].kv_cache_spec
            if isinstance(kv_cache_spec, UniformTypeKVCacheSpecs):
                kv_cache_spec = kv_cache_spec.kv_cache_specs[attn_group.layer_names[0]]
            cache_key = (kv_cache_spec, type(builder))
2242

2243
2244
2245
2246
2247
2248
2249
            cascade_attn_prefix_len = (
                cascade_attn_prefix_lens[kv_cache_gid][attn_gid]
                if cascade_attn_prefix_lens
                else 0
            )

            extra_attn_metadata_args = {}
2250
2251
2252
            if use_spec_decode and isinstance(
                builder, (Mamba2AttentionMetadataBuilder, GDNAttentionMetadataBuilder)
            ):
2253
2254
2255
2256
2257
2258
2259
2260
2261
2262
2263
2264
                assert ubid is None, "UBatching not supported with GDN yet"
                extra_attn_metadata_args = dict(
                    num_accepted_tokens=self.num_accepted_tokens.gpu[:num_reqs_padded],
                    num_decode_draft_tokens_cpu=self.num_decode_draft_tokens.cpu[
                        :num_reqs_padded
                    ],
                )

            if for_cudagraph_capture:
                attn_metadata_i = builder.build_for_cudagraph_capture(
                    common_attn_metadata
                )
2265
2266
2267
2268
2269
2270
2271
2272
2273
            elif (
                cache_key in cached_attn_metadata
                and builder.supports_update_block_table
            ):
                attn_metadata_i = builder.update_block_table(
                    cached_attn_metadata[cache_key],
                    common_attn_metadata.block_table_tensor,
                    common_attn_metadata.slot_mapping,
                )
2274
2275
2276
2277
2278
2279
            else:
                attn_metadata_i = builder.build(
                    common_prefix_len=cascade_attn_prefix_len,
                    common_attn_metadata=common_attn_metadata,
                    **extra_attn_metadata_args,
                )
2280
2281
                if builder.supports_update_block_table:
                    cached_attn_metadata[cache_key] = attn_metadata_i
2282
2283
2284
2285
2286
2287
2288
2289
2290
2291
2292
2293
2294
2295
2296
2297
2298
2299
2300
2301
2302
2303
2304

            if ubid is None:
                assert isinstance(attn_metadata, dict)
                attn_metadata_dict = attn_metadata
            else:
                assert isinstance(attn_metadata, list)
                attn_metadata_dict = attn_metadata[ubid]

            for layer_name in attn_group.layer_names:
                attn_metadata_dict[layer_name] = attn_metadata_i

        # Prepare the attention metadata for each KV cache group and make layers
        # in the same group share the same metadata.
        spec_decode_common_attn_metadata = None
        for kv_cache_gid, kv_cache_group in enumerate(kv_cache_groups):
            cm = copy(cm_base)  # shallow copy

            # Basically only the encoder seq_lens, block_table and slot_mapping change
            # for each kv_cache_group.
            cm.encoder_seq_lens, cm.encoder_seq_lens_cpu = self._get_encoder_seq_lens(
                num_scheduled_tokens or {},
                kv_cache_group.kv_cache_spec,
                num_reqs_padded,
2305
                for_cudagraph_capture=for_cudagraph_capture,
2306
            )
2307
            if kv_cache_gid > 0:
2308
2309
                cm.block_table_tensor = _get_block_table(kv_cache_gid)
                cm.slot_mapping = slot_mappings[kv_cache_gid]
2310

2311
            if self.speculative_config and spec_decode_common_attn_metadata is None:
2312
                if isinstance(self.drafter, (EagleProposer, DFlashProposer)):
2313
                    if self.drafter.kv_cache_gid == kv_cache_gid:
2314
                        spec_decode_common_attn_metadata = cm
2315
                else:
2316
                    spec_decode_common_attn_metadata = cm
2317

2318
            for attn_gid in range(len(self.attn_groups[kv_cache_gid])):
2319
                if ubatch_slices is not None:
2320
2321
2322
                    for ubid, _cm in enumerate(split_attn_metadata(ubatch_slices, cm)):
                        _build_attn_group_metadata(kv_cache_gid, attn_gid, _cm, ubid)

2323
                else:
2324
                    _build_attn_group_metadata(kv_cache_gid, attn_gid, cm)
2325

2326
2327
2328
2329
2330
2331
2332
2333
2334
2335
2336
2337
2338
2339
2340
2341
2342
2343
2344
2345
        if self.is_mm_prefix_lm:
            req_doc_ranges = {}
            for req_id in self.input_batch.req_ids:
                image_doc_ranges = []
                req_state = self.requests[req_id]
                for mm_feature in req_state.mm_features:
                    pos_info = mm_feature.mm_position
                    img_doc_range = pos_info.extract_embeds_range()
                    image_doc_ranges.extend(img_doc_range)
                req_idx = self.input_batch.req_id_to_index[req_id]
                req_doc_ranges[req_idx] = image_doc_ranges

            if isinstance(attn_metadata, list):
                for ub_metadata in attn_metadata:
                    for _metadata in ub_metadata.values():
                        _metadata.mm_prefix_range = req_doc_ranges  # type: ignore[attr-defined]
            else:
                for _metadata in attn_metadata.values():
                    _metadata.mm_prefix_range = req_doc_ranges  # type: ignore[attr-defined]

2346
2347
2348
2349
2350
2351
2352
2353
2354
2355
        if spec_decode_common_attn_metadata is not None and (
            num_reqs != num_reqs_padded or num_tokens != num_tokens_padded
        ):
            # Currently the drafter still only uses piecewise cudagraphs (and modifies
            # the attention metadata in directly), and therefore does not want to use
            # padded attention metadata.
            spec_decode_common_attn_metadata = (
                spec_decode_common_attn_metadata.unpadded(num_tokens, num_reqs)
            )

2356
        return attn_metadata, spec_decode_common_attn_metadata
2357

2358
2359
2360
    def _compute_cascade_attn_prefix_lens(
        self,
        num_scheduled_tokens: np.ndarray,
2361
        num_computed_tokens: np.ndarray,
2362
2363
2364
2365
2366
2367
2368
        num_common_prefix_blocks: list[int],
    ) -> list[list[int]] | None:
        """
        :return: Optional[cascade_attn_prefix_lens]
            cascade_attn_prefix_lens is 2D: ``[kv_cache_group_id][attn_group_idx]``,
            None if we should not use cascade attention
        """
2369

2370
2371
2372
2373
2374
2375
2376
2377
2378
2379
2380
2381
2382
2383
        use_cascade_attn = False
        num_kv_cache_groups = len(self.kv_cache_config.kv_cache_groups)
        cascade_attn_prefix_lens: list[list[int]] = [
            [] for _ in range(num_kv_cache_groups)
        ]

        for kv_cache_gid in range(num_kv_cache_groups):
            for attn_group in self.attn_groups[kv_cache_gid]:
                if isinstance(attn_group.kv_cache_spec, EncoderOnlyAttentionSpec):
                    cascade_attn_prefix_len = 0
                else:
                    # 0 if cascade attention should not be used
                    cascade_attn_prefix_len = self._compute_cascade_attn_prefix_len(
                        num_scheduled_tokens,
2384
                        num_computed_tokens,
2385
2386
2387
2388
2389
2390
2391
2392
                        num_common_prefix_blocks[kv_cache_gid],
                        attn_group.kv_cache_spec,
                        attn_group.get_metadata_builder(),
                    )
                cascade_attn_prefix_lens[kv_cache_gid].append(cascade_attn_prefix_len)
                use_cascade_attn |= cascade_attn_prefix_len > 0

        return cascade_attn_prefix_lens if use_cascade_attn else None
2393

2394
2395
2396
    def _compute_cascade_attn_prefix_len(
        self,
        num_scheduled_tokens: np.ndarray,
2397
        num_computed_tokens: np.ndarray,
2398
        num_common_prefix_blocks: int,
2399
2400
        kv_cache_spec: KVCacheSpec,
        attn_metadata_builder: AttentionMetadataBuilder,
2401
2402
2403
2404
2405
2406
2407
2408
2409
2410
2411
2412
2413
2414
2415
2416
2417
2418
    ) -> int:
        """Compute the length of the common prefix for cascade attention.

        NOTE(woosuk): The common prefix length returned by this function
        represents the length used specifically for cascade attention, not the
        actual number of tokens shared between requests. When cascade attention
        is disabled (use_cascade=False), this function returns 0 even if
        requests share common tokens. Additionally, the common prefix length is
        truncated to a multiple of the block size and may be further truncated
        due to implementation details explained below.

        Args:
            num_scheduled_tokens: Number of tokens scheduled per request.
            num_common_prefix_blocks: Number of shared KV cache blocks.

        Returns:
            int: Length of common prefix in tokens.
        """
2419

2420
        common_prefix_len = num_common_prefix_blocks * kv_cache_spec.block_size
2421
2422
2423
2424
2425
2426
2427
2428
2429
2430
2431
2432
2433
2434
2435
2436
2437
2438
2439
2440
2441
2442
2443
2444
2445
2446
2447
2448
2449
2450
2451
2452
2453
2454
2455
2456
2457
        if common_prefix_len == 0:
            # Common case.
            return 0

        # NOTE(woosuk): Cascade attention uses two attention kernels: one
        # for the common prefix and the other for the rest. For the first
        # kernel, we concatenate all the query tokens (possibly from
        # different requests) and treat them as if they are from the same
        # request. Then, we use bi-directional attention to process the
        # common prefix in the KV cache. Importantly, this means that the
        # first kernel does not do any masking.

        # Consider the following example:
        # Request 1's input query: [D, E, X]
        # Request 1's kv cache: [A, B, C, D, E, X]
        # Request 1's num_computed_tokens: 3 (i.e., [A, B, C])
        # Request 2's input query: [E, Y]
        # Request 2's kv cache: [A, B, C, D, E, Y]
        # Request 2's num_computed_tokens: 4 (i.e., [A, B, C, D])

        # If we use [A, B, C, D, E] as the common prefix, then the
        # first kernel will compute the bi-directional attention between
        # input query [D, E, X, E, Y] and common prefix [A, B, C, D, E].
        # However, this is wrong because D in Request 1 should not attend to
        # E in the common prefix (i.e., we need masking).
        # To avoid this, [A, B, C, D] should be the common prefix.
        # That is, the common prefix should be capped by the minimum
        # num_computed_tokens among the requests, and plus one to include
        # the first token of the query.

        # In practice, we use [A, B, C] as the common prefix, instead of
        # [A, B, C, D] (i.e., the common prefix is capped by the minimum
        # num_computed_tokens, without plus one).
        # This is because of an implementation detail: We want to always
        # use two kernels for cascade attention. Let's imagine:
        # Request 3's input query: [D]
        # Request 3's kv cache: [A, B, C, D]
2458
        # Request 3's num_computed_tokens: 3 (i.e., [A, B, C])
2459
2460
2461
2462
2463
        # If we use [A, B, C, D] as the common prefix for Request 1-3,
        # then Request 3 will be processed only by the first kernel,
        # and the second kernel will get an empty input. While this is not
        # a fundamental problem, our current implementation does not support
        # this case.
2464
        common_prefix_len = min(common_prefix_len, num_computed_tokens.min())
2465
        # common_prefix_len should be a multiple of the block size.
2466
2467
2468
2469
2470
2471
2472
2473
2474
2475
2476
        common_prefix_len = (
            common_prefix_len // kv_cache_spec.block_size * kv_cache_spec.block_size
        )
        use_sliding_window = isinstance(kv_cache_spec, SlidingWindowSpec) or (
            isinstance(kv_cache_spec, FullAttentionSpec)
            and kv_cache_spec.sliding_window is not None
        )
        use_local_attention = isinstance(kv_cache_spec, ChunkedLocalAttentionSpec) or (
            isinstance(kv_cache_spec, FullAttentionSpec)
            and kv_cache_spec.attention_chunk_size is not None
        )
2477
2478
        assert isinstance(kv_cache_spec, AttentionSpec)
        use_cascade = attn_metadata_builder.use_cascade_attention(
2479
2480
2481
            common_prefix_len=common_prefix_len,
            query_lens=num_scheduled_tokens,
            num_query_heads=self.num_query_heads,
2482
            num_kv_heads=kv_cache_spec.num_kv_heads,
2483
            use_alibi=self.use_alibi,
2484
            use_sliding_window=use_sliding_window,
2485
            use_local_attention=use_local_attention,
2486
            num_sms=self.num_sms,
2487
            dcp_world_size=self.dcp_world_size,
2488
2489
2490
        )
        return common_prefix_len if use_cascade else 0

2491
2492
    def _calc_mrope_positions(self, scheduler_output: "SchedulerOutput"):
        mrope_pos_ptr = 0
2493
        for index, req_id in enumerate(self.input_batch.req_ids):
2494
2495
2496
            req = self.requests[req_id]
            assert req.mrope_positions is not None

2497
2498
            num_computed_tokens = self.input_batch.num_computed_tokens_cpu[index]
            num_scheduled_tokens = scheduler_output.num_scheduled_tokens[req_id]
2499
            num_prompt_tokens = length_from_prompt_token_ids_or_embeds(
2500
2501
                req.prompt_token_ids, req.prompt_embeds
            )
2502
2503

            if num_computed_tokens + num_scheduled_tokens > num_prompt_tokens:
2504
2505
                prompt_part_len = max(0, num_prompt_tokens - num_computed_tokens)
                completion_part_len = max(0, num_scheduled_tokens - prompt_part_len)
2506
2507
2508
2509
2510
2511
2512
2513
2514
2515
2516
2517
2518
            else:
                prompt_part_len = num_scheduled_tokens
                completion_part_len = 0

            assert num_scheduled_tokens == prompt_part_len + completion_part_len

            if prompt_part_len > 0:
                # prompt's mrope_positions are pre-computed
                dst_start = mrope_pos_ptr
                dst_end = mrope_pos_ptr + prompt_part_len
                src_start = num_computed_tokens
                src_end = num_computed_tokens + prompt_part_len

2519
2520
2521
                self.mrope_positions.cpu[:, dst_start:dst_end] = req.mrope_positions[
                    :, src_start:src_end
                ]
2522
2523
2524
2525
2526
2527
2528
                mrope_pos_ptr += prompt_part_len

            if completion_part_len > 0:
                # compute completion's mrope_positions on-the-fly
                dst_start = mrope_pos_ptr
                dst_end = mrope_pos_ptr + completion_part_len

2529
                assert req.mrope_position_delta is not None
2530
                MRotaryEmbedding.get_next_input_positions_tensor(
2531
                    out=self.mrope_positions.np,
2532
2533
2534
2535
2536
                    out_offset=dst_start,
                    mrope_position_delta=req.mrope_position_delta,
                    context_len=num_computed_tokens + prompt_part_len,
                    num_new_tokens=completion_part_len,
                )
2537
2538
2539

                mrope_pos_ptr += completion_part_len

2540
2541
2542
2543
2544
2545
2546
2547
2548
2549
2550
2551
2552
2553
2554
2555
2556
2557
2558
2559
2560
2561
2562
2563
2564
2565
2566
2567
2568
2569
2570
2571
2572
2573
2574
2575
2576
2577
2578
2579
2580
2581
2582
2583
2584
2585
2586
    def _calc_xdrope_positions(self, scheduler_output: "SchedulerOutput"):
        xdrope_pos_ptr = 0
        for index, req_id in enumerate(self.input_batch.req_ids):
            req = self.requests[req_id]
            assert req.xdrope_positions is not None

            num_computed_tokens = self.input_batch.num_computed_tokens_cpu[index]
            num_scheduled_tokens = scheduler_output.num_scheduled_tokens[req_id]
            num_prompt_tokens = length_from_prompt_token_ids_or_embeds(
                req.prompt_token_ids, req.prompt_embeds
            )

            if num_computed_tokens + num_scheduled_tokens > num_prompt_tokens:
                prompt_part_len = max(0, num_prompt_tokens - num_computed_tokens)
                completion_part_len = max(0, num_scheduled_tokens - prompt_part_len)
            else:
                prompt_part_len = num_scheduled_tokens
                completion_part_len = 0

            assert num_scheduled_tokens == prompt_part_len + completion_part_len

            if prompt_part_len > 0:
                # prompt's xdrope_positions are pre-computed
                dst_start = xdrope_pos_ptr
                dst_end = xdrope_pos_ptr + prompt_part_len
                src_start = num_computed_tokens
                src_end = num_computed_tokens + prompt_part_len

                self.xdrope_positions.cpu[:, dst_start:dst_end] = req.xdrope_positions[
                    :, src_start:src_end
                ]
                xdrope_pos_ptr += prompt_part_len

            if completion_part_len > 0:
                # compute completion's xdrope_positions on-the-fly
                dst_start = xdrope_pos_ptr
                dst_end = xdrope_pos_ptr + completion_part_len

                XDRotaryEmbedding.get_next_input_positions_tensor(
                    out=self.xdrope_positions.np,
                    out_offset=dst_start,
                    context_len=num_computed_tokens + prompt_part_len,
                    num_new_tokens=completion_part_len,
                )

                xdrope_pos_ptr += completion_part_len

2587
2588
    def _calc_spec_decode_metadata(
        self,
2589
2590
2591
2592
2593
2594
2595
2596
2597
2598
2599
2600
2601
2602
2603
2604
        num_draft_tokens: np.ndarray,
        cu_num_scheduled_tokens: np.ndarray,
    ) -> SpecDecodeMetadata:
        # Inputs:
        # cu_num_scheduled_tokens:  [  4, 104, 107, 207, 209]
        # num_draft_tokens:         [  3,   0,   2,   0,   1]
        # Outputs:
        # cu_num_draft_tokens:      [  3,   3,   5,   5,   6]
        # logits_indices:           [  0,   1,   2,   3, 103, 104, 105, 106,
        #                            206, 207, 208]
        # target_logits_indices:    [  0,   1,   2,   5,   6,   9]
        # bonus_logits_indices:     [  3,   4,   7,   8,  10]

        # Compute the logits indices.
        # [4, 1, 3, 1, 2]
        num_sampled_tokens = num_draft_tokens + 1
2605

2606
2607
2608
2609
2610
        # Step 1.
        # cu_num_sampled_tokens: [4, 5, 8, 9, 11]
        # _arange_scratch[:11]: [0, 1, 2, 3, 0, 0, 1, 2, 0, 0, 1]
        cu_num_sampled_tokens = self._get_cumsum_and_arange(
            num_sampled_tokens, self._arange_scratch, cumsum_dtype=np.int32
2611
        )
2612
        # Step 2. [0, 0, 0, 0, 103, 104, 104, 104, 206, 207, 207]
2613
        logits_indices = np.repeat(
2614
2615
            cu_num_scheduled_tokens - num_sampled_tokens, num_sampled_tokens
        )
2616
        # Step 3. [0, 1, 2, 3, 103, 104, 105, 106, 206, 207, 208]
2617
        logits_indices += self._arange_scratch[: cu_num_sampled_tokens[-1]]
2618
2619
2620
2621
2622

        # Compute the bonus logits indices.
        bonus_logits_indices = cu_num_sampled_tokens - 1

        # Compute the draft logits indices.
2623
        # cu_num_draft_tokens: [3, 3, 5, 5, 6]
2624
2625
2626
        # _arange_scratch[:6]: [0, 1, 2, 0, 1, 0]
        cu_num_draft_tokens = self._get_cumsum_and_arange(
            num_draft_tokens, self._arange_scratch, cumsum_dtype=np.int32
2627
        )
2628
2629
        # [0, 0, 0, 5, 5, 9]
        target_logits_indices = np.repeat(
2630
2631
            cu_num_sampled_tokens - num_sampled_tokens, num_draft_tokens
        )
2632
        # [0, 1, 2, 5, 6, 9]
2633
        target_logits_indices += self._arange_scratch[: cu_num_draft_tokens[-1]]
2634
2635
2636

        # TODO: Optimize the CPU -> GPU copy.
        cu_num_draft_tokens = torch.from_numpy(cu_num_draft_tokens).to(
2637
2638
            self.device, non_blocking=True
        )
2639
2640
2641
        cu_num_sampled_tokens = torch.from_numpy(cu_num_sampled_tokens).to(
            self.device, non_blocking=True
        )
2642
2643
2644
        logits_indices = torch.from_numpy(logits_indices).to(
            self.device, non_blocking=True
        )
2645
        target_logits_indices = torch.from_numpy(target_logits_indices).to(
2646
2647
            self.device, non_blocking=True
        )
2648
        bonus_logits_indices = torch.from_numpy(bonus_logits_indices).to(
2649
2650
            self.device, non_blocking=True
        )
2651

2652
2653
        # Compute the draft token ids.
        # draft_token_indices:      [  1,   2,   3, 105, 106, 208]
2654
        draft_token_ids = self.input_ids.gpu[logits_indices]
2655
2656
        draft_token_ids = draft_token_ids[target_logits_indices + 1]

2657
        return SpecDecodeMetadata(
2658
2659
2660
            draft_token_ids=draft_token_ids,
            num_draft_tokens=num_draft_tokens.tolist(),
            cu_num_draft_tokens=cu_num_draft_tokens,
2661
            cu_num_sampled_tokens=cu_num_sampled_tokens,
2662
2663
2664
2665
2666
            target_logits_indices=target_logits_indices,
            bonus_logits_indices=bonus_logits_indices,
            logits_indices=logits_indices,
        )

2667
2668
2669
2670
2671
2672
2673
    def _prepare_kv_sharing_fast_prefill(
        self,
        logits_indices: torch.Tensor,
    ) -> torch.Tensor:
        assert self.kv_sharing_fast_prefill_logits_indices is not None
        num_logits = logits_indices.shape[0]
        assert num_logits > 0
2674
        self.kv_sharing_fast_prefill_logits_indices[:num_logits].copy_(logits_indices)
2675
2676
2677
2678
2679
        # There might have leftover indices in logits_indices[num_logits:]
        # from previous iterations, whose values may be greater than the
        # batch size in the current iteration. To ensure indices are always
        # valid, we fill the padded indices with the last index.
        self.kv_sharing_fast_prefill_logits_indices[num_logits:].fill_(
2680
2681
            logits_indices[-1].item()
        )
2682
2683
        # Dispatch for the decoder portion of the model.
        _, batch_desc = self.cudagraph_dispatcher.dispatch(
2684
            num_logits, invalid_modes={CUDAGraphMode.FULL}
2685
2686
        )
        num_logits_padded = batch_desc.num_tokens
2687
2688
2689
        logits_indices_padded = self.kv_sharing_fast_prefill_logits_indices[
            :num_logits_padded
        ]
2690
2691
        return logits_indices_padded

2692
    def _batch_mm_inputs_from_scheduler(
2693
2694
        self,
        scheduler_output: "SchedulerOutput",
2695
2696
    ) -> tuple[
        list[str],
2697
        list[tuple[str, MultiModalKwargsItem]],
2698
2699
        list[tuple[str, PlaceholderRange]],
    ]:
2700
        """Batch multimodal inputs from scheduled encoder inputs.
2701
2702
2703

        Args:
            scheduler_output: The scheduler output containing scheduled encoder
2704
                inputs.
2705
2706

        Returns:
2707
            A tuple of (mm_hashes, mm_kwargs, mm_lora_refs) where:
2708
2709
            - mm_hashes: List of multimodal hashes for each item
            - mm_kwargs: List of multimodal kwargs for each item
2710
            - mm_lora_refs: List of (req_id, placeholder_range) for each item
2711
        """
2712
2713
        scheduled_encoder_inputs = scheduler_output.scheduled_encoder_inputs
        if not scheduled_encoder_inputs:
2714
            return [], [], []
2715
2716

        mm_hashes = list[str]()
2717
        mm_kwargs = list[tuple[str, MultiModalKwargsItem]]()
2718
2719
2720
        # Multimodal LoRA reference info to map each multimodal item
        # back to its request & position
        mm_lora_refs = list[tuple[str, PlaceholderRange]]()
2721
2722
        for req_id, encoder_input_ids in scheduled_encoder_inputs.items():
            req_state = self.requests[req_id]
2723
2724

            for mm_input_id in encoder_input_ids:
2725
                mm_feature = req_state.mm_features[mm_input_id]
2726
2727
                if mm_feature.data is None:
                    continue
2728
2729

                mm_hashes.append(mm_feature.identifier)
2730
                mm_kwargs.append((mm_feature.modality, mm_feature.data))
2731
                mm_lora_refs.append((req_id, mm_feature.mm_position))
2732

2733
        return mm_hashes, mm_kwargs, mm_lora_refs
2734

2735
2736
2737
    def _execute_mm_encoder(
        self, scheduler_output: "SchedulerOutput"
    ) -> list[torch.Tensor]:
2738
2739
2740
        mm_hashes, mm_kwargs, mm_lora_refs = self._batch_mm_inputs_from_scheduler(
            scheduler_output
        )
2741
2742

        if not mm_kwargs:
2743
            return []
2744

2745
2746
2747
2748
2749
2750
        should_time = bool(
            self.observability_config
            and self.observability_config.enable_mm_processor_stats
            and scheduler_output.scheduled_encoder_inputs
        )

2751
2752
2753
2754
2755
2756
2757
        # Batch mm inputs as much as we can: if a request in the batch has
        # multiple modalities or a different modality than the previous one,
        # we process it separately to preserve item order.
        # FIXME(ywang96): This is a hacky way to deal with multiple modalities
        # in the same batch while still being able to benefit from batching
        # multimodal inputs. The proper solution should be reordering the
        # encoder outputs.
2758
        model = cast(SupportsMultiModal, self.model)
2759
2760
2761
2762
2763
2764
2765
2766
2767
2768
2769
2770
2771
2772
2773

        if self.lora_config and self.lora_manager.supports_tower_connector_lora():
            # Build LoRA mappings independently for encoder inputs
            # (encoder batch structure is different from main batch)
            prompt_lora_mapping = []
            token_lora_mapping = []
            lora_requests = set()
            encoder_token_counts = []

            for req_id, pos_info in mm_lora_refs:
                req_idx = self.input_batch.req_id_to_index[req_id]
                lora_id = int(self.input_batch.request_lora_mapping[req_idx])

                # Prefer pos_info.get_num_embeds to count precise MM embedding tokens.
                num_tokens = self.model.get_num_mm_encoder_tokens(  # type: ignore[attr-defined]
2774
                    pos_info.get_num_embeds()
2775
2776
2777
2778
2779
2780
2781
2782
2783
2784
2785
2786
2787
2788
2789
2790
2791
2792
2793
2794
2795
2796
2797
2798
2799
2800
2801
2802
2803
2804
2805
2806
2807
2808
2809
2810
2811
2812
2813
2814
2815
                )
                prompt_lora_mapping.append(lora_id)
                token_lora_mapping.extend([lora_id] * num_tokens)
                encoder_token_counts.append(num_tokens)

                if lora_id > 0:
                    lora_request = self.input_batch.lora_id_to_lora_request.get(lora_id)
                    if lora_request is not None:
                        lora_requests.add(lora_request)

            # Set tower adapter mapping
            tower_mapping = LoRAMapping(
                tuple(token_lora_mapping),
                tuple(prompt_lora_mapping),
                is_prefill=True,
                type=LoRAMappingType.TOWER,
            )
            self.lora_manager.set_active_adapters(lora_requests, tower_mapping)

            if hasattr(self.model, "get_num_mm_connector_tokens"):
                post_op_counts = [
                    self.model.get_num_mm_connector_tokens(num_tokens)  # type: ignore[attr-defined]
                    for num_tokens in encoder_token_counts
                ]

                connector_token_mapping = np.repeat(
                    np.array(prompt_lora_mapping, dtype=np.int32),
                    np.array(post_op_counts, dtype=np.int32),
                )
                connector_mapping = LoRAMapping(
                    index_mapping=tuple(connector_token_mapping.tolist()),
                    prompt_mapping=tuple(prompt_lora_mapping),
                    is_prefill=True,
                    type=LoRAMappingType.CONNECTOR,
                )

                self.lora_manager.set_active_adapters(
                    lora_requests,
                    connector_mapping,
                )

2816
        encoder_outputs: list[torch.Tensor] = []
2817
2818
        # Track the current index in mm_kwargs/mm_lora_refs to map groups to request IDs
        current_item_idx = 0
2819
        for modality, num_items, mm_kwargs_batch in group_and_batch_mm_kwargs(
2820
2821
2822
            mm_kwargs,
            device=self.device,
            pin_memory=self.pin_memory,
2823
        ):
2824
            batch_outputs: MultiModalEmbeddings
2825

2826
            # EVS and dynamic res video related change.
2827
            # (ekhvedchenia): Temporary hack to limit peak memory usage when
2828
            # processing multimodal data. This solves the issue with scheduler
2829
2830
2831
2832
            # putting too many video samples into a single batch. Scheduler
            # uses pruned vision tokens count to compare it versus compute
            # budget which is incorrect (Either input media size or non-pruned
            # output vision tokens count should be considered)
2833
2834
2835
            # dynamic res video for nemotron temporarily uses this hack via
            # requires_sequential_video_encoding
            # because it doesn't yet support video batching.
2836
2837
2838
            # TODO(ywang96): Fix memory profiling to take EVS into account and
            # remove this hack.
            if (
2839
2840
2841
2842
                (
                    self.is_multimodal_pruning_enabled
                    or self.requires_sequential_video_encoding
                )
2843
2844
2845
                and modality == "video"
                and num_items > 1
            ):
2846
                batch_outputs_lst = list[torch.Tensor]()
2847
2848
2849
2850
2851
2852
                for video_idx in range(num_items):
                    video_mm_kwargs_item = mm_kwargs[current_item_idx + video_idx]
                    with self.timed_encoder_operation(
                        should_time, mm_lora_refs, current_item_idx + video_idx, 1
                    ):
                        _, _, micro_batch_mm_inputs = next(
2853
                            group_and_batch_mm_kwargs(
2854
2855
2856
2857
                                [video_mm_kwargs_item],
                                device=self.device,
                                pin_memory=self.pin_memory,
                            )
2858
                        )
2859

2860
2861
2862
                        micro_batch_outputs = model.embed_multimodal(
                            **micro_batch_mm_inputs
                        )
2863

2864
                        batch_outputs_lst.extend(micro_batch_outputs)
2865

2866
                batch_outputs = batch_outputs_lst
2867
2868
            else:
                # Run the encoder.
2869
                # `batch_outputs` is either of the following:
2870
2871
2872
2873
2874
                # 1. A tensor of shape (num_items, feature_size, hidden_size)
                # in case feature_size is fixed across all multimodal items.
                # 2. A list or tuple (length: num_items) of tensors,
                # each of shape (feature_size, hidden_size) in case the feature
                # size is dynamic depending on the input multimodal items.
2875
2876
2877
2878

                with self.timed_encoder_operation(
                    should_time, mm_lora_refs, current_item_idx, num_items
                ):
2879
2880
2881
2882
2883
2884
2885
2886
2887
2888
2889
2890
2891
                    cudagraph_output = None
                    if (
                        self.encoder_cudagraph_manager is not None
                        and self.encoder_cudagraph_manager.supports_modality(modality)
                    ):
                        cudagraph_output = self.encoder_cudagraph_manager.execute(
                            mm_kwargs_batch,
                        )

                    if cudagraph_output is not None:
                        batch_outputs = cudagraph_output
                    else:
                        batch_outputs = model.embed_multimodal(**mm_kwargs_batch)
2892

2893
2894
            sanity_check_mm_encoder_outputs(batch_outputs, expected_num_items=num_items)
            encoder_outputs.extend(batch_outputs)
2895

2896
2897
            current_item_idx += num_items

2898
        # Cache the encoder outputs by mm_hash
2899
        for mm_hash, output in zip(mm_hashes, encoder_outputs):
2900
            self.encoder_cache[mm_hash] = output
2901
2902
            logger.debug("Finish execute for mm hash %s", mm_hash)
            self.maybe_save_ec_to_connector(self.encoder_cache, mm_hash)
2903

2904
2905
        return encoder_outputs

2906
    def _gather_mm_embeddings(
2907
2908
        self,
        scheduler_output: "SchedulerOutput",
2909
        shift_computed_tokens: int = 0,
2910
2911
2912
    ) -> tuple[list[torch.Tensor], torch.Tensor]:
        total_num_scheduled_tokens = scheduler_output.total_num_scheduled_tokens

2913
2914
2915
2916
2917
        # Swap to the other buffer to avoid race condition with previous
        # iteration's async copy that may still be reading from CPU.
        self.is_mm_embed_idx = 1 - self.is_mm_embed_idx
        is_mm_embed_buf = self.is_mm_embed_buffers[self.is_mm_embed_idx]

2918
        mm_embeds = list[torch.Tensor]()
2919
        is_mm_embed = is_mm_embed_buf.cpu
2920
2921
2922
        is_mm_embed[:total_num_scheduled_tokens] = False

        req_start_idx = 0
2923
        should_sync_mrope_positions = False
2924
        should_sync_xdrope_positions = False
2925

2926
        for req_id in self.input_batch.req_ids:
2927
2928
            mm_embeds_req: list[torch.Tensor] = []

2929
            num_scheduled_tokens = scheduler_output.num_scheduled_tokens[req_id]
2930
            req_state = self.requests[req_id]
2931
            num_computed_tokens = req_state.num_computed_tokens + shift_computed_tokens
2932

2933
2934
            for mm_feature in req_state.mm_features:
                pos_info = mm_feature.mm_position
2935
2936
                start_pos = pos_info.offset
                num_encoder_tokens = pos_info.length
2937
2938
2939
2940
2941
2942
2943
2944
2945
2946
2947
2948
2949
2950
2951
2952

                # The encoder output is needed if the two ranges overlap:
                # [num_computed_tokens,
                #  num_computed_tokens + num_scheduled_tokens) and
                # [start_pos, start_pos + num_encoder_tokens)
                if start_pos >= num_computed_tokens + num_scheduled_tokens:
                    # The encoder output is not needed in this step.
                    break
                if start_pos + num_encoder_tokens <= num_computed_tokens:
                    # The encoder output is already processed and stored
                    # in the decoder's KV cache.
                    continue

                start_idx = max(num_computed_tokens - start_pos, 0)
                end_idx = min(
                    num_computed_tokens - start_pos + num_scheduled_tokens,
2953
2954
                    num_encoder_tokens,
                )
2955
                assert start_idx < end_idx
2956
2957
2958
2959
2960
2961
2962
                curr_embeds_start, curr_embeds_end = (
                    pos_info.get_embeds_indices_in_range(start_idx, end_idx)
                )
                # If there are no embeddings in the current range, we skip
                # gathering the embeddings.
                if curr_embeds_start == curr_embeds_end:
                    continue
2963

2964
                mm_hash = mm_feature.identifier
2965
                encoder_output = self.encoder_cache.get(mm_hash, None)
2966
                assert encoder_output is not None, f"Encoder cache miss for {mm_hash}."
2967
2968
2969

                if (is_embed := pos_info.is_embed) is not None:
                    is_embed = is_embed[start_idx:end_idx]
2970
2971
2972
                    mm_embeds_item = encoder_output[curr_embeds_start:curr_embeds_end]
                else:
                    mm_embeds_item = encoder_output[start_idx:end_idx]
2973

2974
                req_start_pos = req_start_idx + start_pos - num_computed_tokens
2975
2976
2977
2978
2979
2980
2981
2982
2983
                # OR mask for overlapping mm_features (use_audio_in_video)
                if is_embed is None:
                    is_mm_embed[req_start_pos + start_idx : req_start_pos + end_idx] = (
                        True
                    )
                else:
                    is_mm_embed[
                        req_start_pos + start_idx : req_start_pos + end_idx
                    ] |= is_embed
2984
2985
2986
                mm_embeds_req.append(mm_embeds_item)

            if self.is_multimodal_pruning_enabled and self.uses_mrope:
2987
                assert req_state.mrope_positions is not None
2988
2989
2990
2991
2992
2993
2994
                should_sync_mrope_positions = True
                mm_embeds_req, new_mrope_positions, new_delta = (
                    self.model.recompute_mrope_positions(
                        input_ids=req_state.prompt_token_ids,
                        multimodal_embeddings=mm_embeds_req,
                        mrope_positions=req_state.mrope_positions,
                        num_computed_tokens=req_state.num_computed_tokens,
2995
2996
                    )
                )
2997
2998
2999
3000
                req_state.mrope_positions.copy_(new_mrope_positions)
                req_state.mrope_position_delta = new_delta

            mm_embeds.extend(mm_embeds_req)
3001
3002
            req_start_idx += num_scheduled_tokens

3003
        is_mm_embed = is_mm_embed_buf.copy_to_gpu(total_num_scheduled_tokens)
3004
3005
3006

        if should_sync_mrope_positions:
            self._calc_mrope_positions(scheduler_output)
3007
            self.mrope_positions.copy_to_gpu(total_num_scheduled_tokens)
3008

3009
3010
3011
3012
        if should_sync_xdrope_positions:
            self._calc_xdrope_positions(scheduler_output)
            self.xdrope_positions.copy_to_gpu(total_num_scheduled_tokens)

3013
        return mm_embeds, is_mm_embed
3014

3015
    def get_model(self) -> nn.Module:
3016
3017
        if not hasattr(self, "model"):
            raise ValueError("Cannot get model before model has been initialized")
3018
        if isinstance(self.model, (CUDAGraphWrapper, UBatchWrapper)):
3019
            # get raw model out of the cudagraph wrapper.
3020
            return self.model.unwrap()
3021
3022
        return self.model

3023
3024
3025
3026
3027
3028
3029
3030
3031
3032
3033
3034
3035
    def get_supported_generation_tasks(self) -> list[GenerationTask]:
        model = self.get_model()
        supported_tasks = list[GenerationTask]()

        if is_text_generation_model(model):
            supported_tasks.append("generate")

        if supports_transcription(model):
            if model.supports_transcription_only:
                return ["transcription"]

            supported_tasks.append("transcription")

3036
3037
3038
        if supports_realtime(model):
            supported_tasks.append("realtime")

3039
3040
        return supported_tasks

3041
3042
3043
3044
3045
    def get_supported_pooling_tasks(self) -> list[PoolingTask]:
        model = self.get_model()
        if not is_pooling_model(model):
            return []

3046
        return list(model.pooler.get_supported_tasks())
3047

3048
3049
3050
3051
3052
3053
3054
3055
3056
3057
    def get_supported_tasks(self) -> tuple[SupportedTask, ...]:
        tasks = list[SupportedTask]()

        if self.model_config.runner_type == "generate":
            tasks.extend(self.get_supported_generation_tasks())
        if self.model_config.runner_type == "pooling":
            tasks.extend(self.get_supported_pooling_tasks())

        return tuple(tasks)

3058
    def sync_and_slice_intermediate_tensors(
3059
3060
        self,
        num_tokens: int,
3061
        intermediate_tensors: IntermediateTensors | None,
3062
3063
        sync_self: bool,
    ) -> IntermediateTensors:
3064
3065
3066
        assert self.intermediate_tensors is not None

        tp = self.vllm_config.parallel_config.tensor_parallel_size
3067
        is_rs = is_residual_scattered_for_sp(self.vllm_config, num_tokens)
3068
3069
3070
3071
3072
3073

        # When sequence parallelism is enabled, the "residual" tensor is sharded
        # across tensor parallel ranks, so each rank only needs its own slice.
        if sync_self:
            assert intermediate_tensors is not None
            for k, v in intermediate_tensors.items():
3074
                is_scattered = k == "residual" and is_rs
3075
                copy_len = num_tokens // tp if is_scattered else num_tokens
3076
                self.intermediate_tensors[k][:copy_len].copy_(
3077
3078
3079
3080
3081
3082
3083
3084
3085
3086
3087
3088
3089
                    v[:copy_len], non_blocking=True
                )

        return IntermediateTensors(
            {
                k: v[: num_tokens // tp]
                if k == "residual" and is_rs
                else v[:num_tokens]
                for k, v in self.intermediate_tensors.items()
            }
        )

    def eplb_step(self, is_dummy: bool = False, is_profile: bool = False) -> None:
3090
3091
3092
        """
        Step for the EPLB (Expert Parallelism Load Balancing) state.
        """
3093
        if not self.parallel_config.enable_eplb or self.eep_eplb_suppressed:
3094
3095
3096
            return

        assert self.eplb_state is not None
3097
3098
        model = self.get_model()
        assert is_mixture_of_experts(model)
3099
3100
3101
        self.eplb_state.step(
            is_dummy,
            is_profile,
3102
            log_stats=self.parallel_config.eplb_config.log_balancedness,
3103
3104
        )

3105
3106
3107
3108
3109
3110
3111
3112
3113
3114
3115
3116
3117
3118
3119
3120
3121
    def setup_eplb_from_mapping(
        self,
        expanded_physical_to_logical: torch.Tensor,
        old_num_physical_experts: int,
    ) -> None:
        model = self.get_model()
        assert is_mixture_of_experts(model)

        self.eplb_state = EplbState.from_mapping(
            model=model,
            model_config=self.model_config,
            device=self.device,
            parallel_config=self.parallel_config,
            expanded_physical_to_logical=expanded_physical_to_logical,
            num_valid_physical_experts=old_num_physical_experts,
        )

3122
3123
3124
3125
3126
    def _pool(
        self,
        hidden_states: torch.Tensor,
        num_scheduled_tokens: int,
        num_scheduled_tokens_np: np.ndarray,
3127
3128
3129
3130
        kv_connector_output: KVConnectorOutput | None,
    ) -> ModelRunnerOutput | AsyncModelRunnerOutput:
        num_reqs = self.input_batch.num_reqs
        assert num_reqs == len(self.input_batch.pooling_params), (
3131
3132
            "Either all or none of the requests in a batch must be pooling request"
        )
3133

3134
        hidden_states = hidden_states[:num_scheduled_tokens]
3135
        seq_lens_cpu = self.optimistic_seq_lens_cpu[:num_reqs]
3136

3137
        pooling_metadata = self.input_batch.get_pooling_metadata()
3138
        pooling_metadata.build_pooling_cursor(
3139
3140
3141
3142
            num_scheduled_tokens_np,
            seq_lens_cpu,
            device=hidden_states.device,
            query_start_loc_gpu=self.query_start_loc.gpu[: num_reqs + 1],
3143
        )
3144

3145
3146
        model = cast(VllmModelForPooling, self.model)
        raw_pooler_output: PoolerOutput = model.pooler(
3147
            hidden_states=hidden_states, pooling_metadata=pooling_metadata
3148
        )
3149
3150
3151
3152
3153

        finished_mask = [
            seq_len == prompt_len
            for seq_len, prompt_len in zip(seq_lens_cpu, pooling_metadata.prompt_lens)
        ]
3154
3155
3156
3157
3158
3159
        raw_pooler_output = self.late_interaction_runner.postprocess_pooler_output(
            raw_pooler_output=raw_pooler_output,
            pooling_params=pooling_metadata.pooling_params,
            req_ids=self.input_batch.req_ids,
            finished_mask=finished_mask,
        )
3160
3161
3162
3163
3164
3165
3166
3167
3168
3169
3170
3171
3172
3173
3174
3175
3176
3177
3178

        model_runner_output = ModelRunnerOutput(
            req_ids=self.input_batch.req_ids.copy(),
            req_id_to_index=self.input_batch.req_id_to_index.copy(),
            kv_connector_output=kv_connector_output,
        )

        if raw_pooler_output is None or not any(finished_mask):
            model_runner_output.pooler_output = [None] * num_reqs
            return model_runner_output

        if self.use_async_scheduling:
            return AsyncGPUPoolingModelRunnerOutput(
                model_runner_output=model_runner_output,
                raw_pooler_output=raw_pooler_output,
                finished_mask=finished_mask,
                async_output_copy_stream=self.async_output_copy_stream,
            )

3179
3180
3181
        model_runner_output.pooler_output = _copy_pooler_output_to_cpu(
            raw_pooler_output=raw_pooler_output,
            finished_mask=finished_mask,
3182
        )
3183
3184
        self._sync_device()

3185
        return model_runner_output
3186

3187
    def _pad_for_sequence_parallelism(self, num_scheduled_tokens: int) -> int:
3188
3189
3190
        # Pad tokens to multiple of tensor_parallel_size when
        # enabled collective fusion for SP
        tp_size = self.vllm_config.parallel_config.tensor_parallel_size
3191
        if self.compilation_config.pass_config.enable_sp and tp_size > 1:
3192
3193
3194
            return round_up(num_scheduled_tokens, tp_size)
        return num_scheduled_tokens

Patrick von Platen's avatar
Patrick von Platen committed
3195
3196
3197
3198
3199
3200
3201
3202
3203
3204
3205
    def _prepare_mm_inputs(
        self, num_tokens: int
    ) -> tuple[torch.Tensor | None, torch.Tensor]:
        if self.model.requires_raw_input_tokens:
            input_ids = self.input_ids.gpu[:num_tokens]
        else:
            input_ids = None

        inputs_embeds = self.inputs_embeds.gpu[:num_tokens]
        return input_ids, inputs_embeds

3206
    def _preprocess(
3207
3208
        self,
        scheduler_output: "SchedulerOutput",
3209
        num_input_tokens: int,  # Padded
3210
        intermediate_tensors: IntermediateTensors | None = None,
3211
    ) -> tuple[
3212
3213
        torch.Tensor | None,
        torch.Tensor | None,
3214
        torch.Tensor,
3215
        IntermediateTensors | None,
3216
        dict[str, Any],
3217
        ECConnectorOutput | None,
3218
    ]:
3219
        num_scheduled_tokens = scheduler_output.total_num_scheduled_tokens
3220
        is_first_rank = get_pp_group().is_first_rank
3221
        is_encoder_decoder = self.model_config.is_encoder_decoder
3222

3223
3224
        # _prepare_inputs may reorder the batch, so we must gather multi
        # modal outputs after that to ensure the correct order
3225
3226
        ec_connector_output = None

3227
        if self.supports_mm_inputs and is_first_rank and not is_encoder_decoder:
3228
            # Run the multimodal encoder if any.
3229
3230
3231
3232
3233
3234
            with self.maybe_get_ec_connector_output(
                scheduler_output,
                encoder_cache=self.encoder_cache,
            ) as ec_connector_output:
                self._execute_mm_encoder(scheduler_output)
                mm_embeds, is_mm_embed = self._gather_mm_embeddings(scheduler_output)
3235

3236
3237
3238
            # NOTE(woosuk): To unify token ids and soft tokens (vision
            # embeddings), we always use embeddings (rather than token ids)
            # as input to the multimodal model, even when the input is text.
3239
            inputs_embeds_scheduled = self.model.embed_input_ids(
3240
3241
3242
                self.input_ids.gpu[:num_scheduled_tokens],
                multimodal_embeddings=mm_embeds,
                is_multimodal=is_mm_embed,
3243
            )
3244

3245
            # TODO(woosuk): Avoid the copy. Optimize.
3246
            self.inputs_embeds.gpu[:num_scheduled_tokens].copy_(inputs_embeds_scheduled)
3247

Patrick von Platen's avatar
Patrick von Platen committed
3248
            input_ids, inputs_embeds = self._prepare_mm_inputs(num_input_tokens)
3249
            model_kwargs = {
3250
                **self._init_model_kwargs(),
3251
3252
                **self._extract_mm_kwargs(scheduler_output),
            }
3253
        elif self.enable_prompt_embeds and is_first_rank:
3254
3255
3256
3257
3258
3259
3260
3261
3262
3263
3264
3265
            # Get the input embeddings for the tokens that are not input embeds,
            # then put them into the appropriate positions.
            # TODO(qthequartermasterman): Since even when prompt embeds are
            # enabled, (a) not all requests will use prompt embeds, and (b)
            # after the initial prompt is processed, the rest of the generated
            # tokens will be token ids, it is not desirable to have the
            # embedding layer outside of the CUDA graph all the time. The v0
            # engine avoids this by "double compiling" the CUDA graph, once
            # with input_ids and again with inputs_embeds, for all num_tokens.
            # If a batch only has token ids, then including the embedding layer
            # in the CUDA graph will be more performant (like in the else case
            # below).
3266
3267
3268
            token_ids_idx = (
                self.is_token_ids.gpu[:num_scheduled_tokens]
                .nonzero(as_tuple=False)
3269
                .squeeze(1)
3270
            )
3271
3272
3273
            # Some tokens ids may need to become embeds
            if token_ids_idx.numel() > 0:
                token_ids = self.input_ids.gpu[token_ids_idx]
3274
                tokens_to_embeds = self.model.embed_input_ids(input_ids=token_ids)
3275
3276
3277
                self.inputs_embeds.gpu[token_ids_idx] = tokens_to_embeds

            inputs_embeds = self.inputs_embeds.gpu[:num_input_tokens]
3278
            model_kwargs = self._init_model_kwargs()
3279
            input_ids = None
3280
        else:
3281
3282
3283
3284
            # For text-only models, we use token ids as input.
            # While it is possible to use embeddings as input just like the
            # multimodal models, it is not desirable for performance since
            # then the embedding layer is not included in the CUDA graph.
3285
            input_ids = self.input_ids.gpu[:num_input_tokens]
3286
            inputs_embeds = None
3287
            model_kwargs = self._init_model_kwargs()
3288

3289
        if self.uses_mrope:
3290
            positions = self.mrope_positions.gpu[:, :num_input_tokens]
3291
3292
        elif self.uses_xdrope_dim > 0:
            positions = self.xdrope_positions.gpu[:, :num_input_tokens]
3293
        else:
3294
            positions = self.positions[:num_input_tokens]
3295
            if num_input_tokens > num_scheduled_tokens:
3296
                self.positions[num_scheduled_tokens:num_input_tokens].zero_()
3297

3298
        if is_first_rank:
3299
3300
            intermediate_tensors = None
        else:
3301
            assert intermediate_tensors is not None
3302
            intermediate_tensors = self.sync_and_slice_intermediate_tensors(
3303
3304
                num_input_tokens, intermediate_tensors, True
            )
3305

3306
        if is_encoder_decoder and scheduler_output.scheduled_encoder_inputs:
3307
3308
3309
3310
3311
3312
3313
            # Run the encoder, just like we do with other multimodal inputs.
            # For an encoder-decoder model, our processing here is a bit
            # simpler, because the outputs are just passed to the decoder.
            # We are not doing any prompt replacement. We also will only
            # ever have a single encoder input.
            encoder_outputs = self._execute_mm_encoder(scheduler_output)
            model_kwargs.update({"encoder_outputs": encoder_outputs})
3314

3315
3316
3317
3318
3319
3320
        return (
            input_ids,
            inputs_embeds,
            positions,
            intermediate_tensors,
            model_kwargs,
3321
            ec_connector_output,
3322
        )
3323

3324
    def _sample(
3325
        self,
3326
3327
        logits: torch.Tensor | None,
        spec_decode_metadata: SpecDecodeMetadata | None,
3328
    ) -> SamplerOutput:
3329
        # Sample the next token and get logprobs if needed.
3330
        sampling_metadata = self.input_batch.sampling_metadata
3331
3332
3333
        # Update output token ids with tokens sampled in last step
        # if async scheduling and required by current sampling params.
        self.input_batch.update_async_output_token_ids()
3334
        if spec_decode_metadata is None:
3335
            return self.sampler(
3336
3337
3338
                logits=logits,
                sampling_metadata=sampling_metadata,
            )
3339

3340
3341
3342
3343
3344
3345
        # Update spec_token_ids with real draft tokens from pre step only when
        # output_token_ids is needed (penalties or bad_words are in use).
        if self.use_async_scheduling and self._draft_token_req_ids is not None:
            draft_token_ids_cpu, _ = self._get_draft_token_ids_cpu()
            self.input_batch.update_async_spec_token_ids(draft_token_ids_cpu)

3346
        sampler_output = self.rejection_sampler(
3347
3348
            spec_decode_metadata,
            None,  # draft_probs
3349
            logits,
3350
3351
            sampling_metadata,
        )
3352
3353
3354
        return sampler_output

    def _bookkeeping_sync(
3355
3356
3357
        self,
        scheduler_output: "SchedulerOutput",
        sampler_output: SamplerOutput,
3358
        logits: torch.Tensor | None,
3359
3360
        hidden_states: torch.Tensor,
        num_scheduled_tokens: int,
3361
        spec_decode_metadata: SpecDecodeMetadata | None,
3362
    ) -> tuple[
3363
        dict[str, int],
3364
        LogprobsLists | None,
3365
        list[list[int]],
3366
        dict[str, LogprobsTensors | None],
3367
3368
3369
        list[str],
        dict[str, int],
        list[int],
3370
    ]:
3371
3372
3373
3374
        num_nans_in_logits = {}
        if envs.VLLM_COMPUTE_NANS_IN_LOGITS:
            num_nans_in_logits = self._get_nans_in_logits(logits)

3375
3376
3377
3378
        num_reqs = self.input_batch.num_reqs
        discard_sampled_tokens_req_indices = np.nonzero(
            self.discard_request_mask.np[:num_reqs]
        )[0]
3379
3380
3381
3382
        for i in discard_sampled_tokens_req_indices:
            gen = self.input_batch.generators.get(int(i))
            if gen is not None:
                gen.set_offset(gen.get_offset() - 4)
3383

3384
3385
3386
        # Copy some objects so they don't get modified after returning.
        # This is important when using async scheduling.
        req_ids_output_copy = self.input_batch.req_ids.copy()
3387
        req_id_to_index_output_copy = self.input_batch.req_id_to_index.copy()
3388
3389

        num_sampled_tokens = sampler_output.sampled_token_ids.shape[0]
3390
        sampled_token_ids = sampler_output.sampled_token_ids
3391
        logprobs_tensors = sampler_output.logprobs_tensors
3392
        invalid_req_indices = []
3393
        logprobs_lists = None
3394
3395
3396
3397
3398
3399
        if not self.use_async_scheduling:
            # Get the valid generated tokens.
            max_gen_len = sampled_token_ids.shape[-1]
            if max_gen_len == 1:
                # No spec decode tokens.
                valid_sampled_token_ids = self._to_list(sampled_token_ids)
3400
3401
3402
                # Mask out the sampled tokens that should not be sampled.
                for i in discard_sampled_tokens_req_indices:
                    valid_sampled_token_ids[int(i)].clear()
3403
3404
3405

                if logprobs_tensors is not None:
                    logprobs_lists = logprobs_tensors.tolists()
3406
3407
            else:
                # Includes spec decode tokens.
3408
                valid_sampled_token_ids, logprobs_lists = RejectionSampler.parse_output(
3409
3410
                    sampled_token_ids,
                    self.input_batch.vocab_size,
3411
                    discard_sampled_tokens_req_indices,
3412
                    logprobs_tensors=logprobs_tensors,
3413
                )
3414
        else:
3415
            valid_sampled_token_ids = []
3416
            invalid_req_indices = discard_sampled_tokens_req_indices.tolist()
3417
3418
3419
3420
3421
            invalid_req_indices_set = set(invalid_req_indices)

            # Cache the sampled tokens on the GPU and avoid CPU sync.
            # These will be copied into input_ids in the next step
            # when preparing inputs.
3422
3423
3424
3425
            # With spec decoding, this is done in propose_draft_token_ids().
            if self.input_batch.prev_sampled_token_ids is None:
                assert sampled_token_ids.shape[-1] == 1
                self.input_batch.prev_sampled_token_ids = sampled_token_ids
3426
3427
3428
3429
3430
            self.input_batch.prev_req_id_to_index = {
                req_id: i
                for i, req_id in enumerate(self.input_batch.req_ids)
                if i not in invalid_req_indices_set
            }
3431

3432
3433
3434
3435
3436
        # Cache the sampled tokens in the model runner, so that the scheduler
        # doesn't need to send them back.
        # NOTE(woosuk): As an exception, when using PP, the scheduler sends
        # the sampled tokens back, because there's no direct communication
        # between the first-stage worker and the last-stage worker.
3437
        req_ids = self.input_batch.req_ids
3438
3439
        for req_idx in range(num_sampled_tokens):
            if self.use_async_scheduling:
3440
                sampled_ids = [-1] if req_idx not in invalid_req_indices_set else None
3441
3442
            else:
                sampled_ids = valid_sampled_token_ids[req_idx]
3443

3444
            num_sampled_ids: int = len(sampled_ids) if sampled_ids else 0
3445

3446
            if not sampled_ids:
3447
3448
3449
                continue

            start_idx = self.input_batch.num_tokens_no_spec[req_idx]
3450
            end_idx = start_idx + num_sampled_ids
3451
3452
3453
3454
            assert end_idx <= self.max_model_len, (
                "Sampled token IDs exceed the max model length. "
                f"Total number of tokens: {end_idx} > max_model_len: "
                f"{self.max_model_len}"
3455
            )
3456

3457
3458
            self.input_batch.token_ids_cpu[req_idx, start_idx:end_idx] = sampled_ids
            self.input_batch.is_token_ids[req_idx, start_idx:end_idx] = True
3459
            self.input_batch.num_tokens_no_spec[req_idx] = end_idx
3460

3461
            req_id = req_ids[req_idx]
3462
3463
3464
            req_state = self.requests[req_id]
            req_state.output_token_ids.extend(sampled_ids)

3465
3466
3467
3468
3469
3470
        # Compute prompt logprobs if needed.
        prompt_logprobs_dict = self._get_prompt_logprobs_dict(
            hidden_states[:num_scheduled_tokens],
            scheduler_output.num_scheduled_tokens,
        )

3471
3472
3473
3474
3475
3476
3477
3478
3479
3480
        return (
            num_nans_in_logits,
            logprobs_lists,
            valid_sampled_token_ids,
            prompt_logprobs_dict,
            req_ids_output_copy,
            req_id_to_index_output_copy,
            invalid_req_indices,
        )

3481
3482
3483
3484
3485
3486
3487
3488
3489
3490
3491
3492
3493
3494
3495
    @contextmanager
    def synchronize_input_prep(self):
        if self.prepare_inputs_event is None:
            yield
            return

        # Ensure prior step has finished with reused CPU tensors.
        # This is required in the async scheduling case because
        # the CPU->GPU transfer happens async.
        self.prepare_inputs_event.synchronize()
        try:
            yield
        finally:
            self.prepare_inputs_event.record()

3496
3497
    def _model_forward(
        self,
3498
3499
3500
3501
        input_ids: torch.Tensor | None = None,
        positions: torch.Tensor | None = None,
        intermediate_tensors: IntermediateTensors | None = None,
        inputs_embeds: torch.Tensor | None = None,
3502
3503
3504
3505
3506
        **model_kwargs: dict[str, Any],
    ) -> Any:
        """Helper method to call the model forward pass.

        This method can be overridden by subclasses for model execution.
3507
        Motivation: We can inspect only this method versus
3508
3509
3510
3511
3512
3513
3514
3515
3516
3517
3518
3519
3520
3521
3522
3523
3524
3525
3526
3527
        the whole execute_model, which has additional logic.

        Args:
            input_ids: Input token IDs
            positions: Token positions
            intermediate_tensors: Tensors from previous pipeline stages
            inputs_embeds: Input embeddings (alternative to input_ids)
            **model_kwargs: Additional model arguments

        Returns:
            Model output tensor
        """
        return self.model(
            input_ids=input_ids,
            positions=positions,
            intermediate_tensors=intermediate_tensors,
            inputs_embeds=inputs_embeds,
            **model_kwargs,
        )

3528
3529
3530
3531
3532
3533
3534
3535
3536
3537
3538
3539
3540
3541
3542
3543
3544
3545
3546
3547
3548
    @staticmethod
    def _is_uniform_decode(
        max_num_scheduled_tokens: int,
        uniform_decode_query_len: int,
        num_tokens: int,
        num_reqs: int,
        force_uniform_decode: bool | None = None,
    ) -> bool:
        """
        Checks if it's a decode batch with same amount scheduled tokens
        across all requests.
        """
        return (
            (
                (max_num_scheduled_tokens == uniform_decode_query_len)
                and (num_tokens == max_num_scheduled_tokens * num_reqs)
            )
            if force_uniform_decode is None
            else force_uniform_decode
        )

3549
3550
3551
3552
3553
3554
3555
3556
3557
3558
3559
3560
3561
    def _determine_batch_execution_and_padding(
        self,
        num_tokens: int,
        num_reqs: int,
        num_scheduled_tokens_np: np.ndarray,
        max_num_scheduled_tokens: int,
        use_cascade_attn: bool,
        allow_microbatching: bool = True,
        force_eager: bool = False,
        # For cudagraph capture TODO(lucas): Refactor how we capture cudagraphs (will
        # be improved in model runner v2)
        force_uniform_decode: bool | None = None,
        force_has_lora: bool | None = None,
3562
        force_num_active_loras: int | None = None,
3563
        num_encoder_reqs: int = 0,
3564
    ) -> tuple[
3565
3566
        CUDAGraphMode,
        BatchDescriptor,
3567
        bool,
3568
3569
        torch.Tensor | None,
        CUDAGraphStat | None,
3570
    ]:
3571
3572
3573
3574
3575
3576
        uniform_decode = self._is_uniform_decode(
            max_num_scheduled_tokens=max_num_scheduled_tokens,
            uniform_decode_query_len=self.uniform_decode_query_len,
            num_tokens=num_tokens,
            num_reqs=num_reqs,
            force_uniform_decode=force_uniform_decode,
3577
        )
3578
3579
3580
3581
3582
        # Encoder-decoder models only support CG for decoder_step > 0 (no enc_output
        # is present). Also, chunked-prefill is disabled, so batch are uniform.
        has_encoder_output = (
            self.model_config.is_encoder_decoder and num_encoder_reqs > 0
        )
3583

3584
3585
3586
3587
3588
        # Compute LoRA state for cudagraph dispatch
        num_active_loras = (
            force_num_active_loras
            if force_num_active_loras is not None
            else len(self.input_batch.lora_id_to_lora_request)
3589
        )
3590
        has_lora = num_active_loras > 0 if force_has_lora is None else force_has_lora
3591

3592
        num_tokens_padded = self._pad_for_sequence_parallelism(num_tokens)
3593
3594
3595

        def dispatch_cudagraph(num_tokens, disable_full=False, valid_modes=None):
            return self.cudagraph_dispatcher.dispatch(
3596
3597
3598
                num_tokens=num_tokens,
                has_lora=has_lora,
                uniform_decode=uniform_decode,
3599
                num_active_loras=num_active_loras,
3600
3601
                valid_modes={CUDAGraphMode.NONE} if force_eager else valid_modes,
                invalid_modes={CUDAGraphMode.FULL} if disable_full else None,
3602
3603
            )

3604
        cudagraph_mode, batch_descriptor = dispatch_cudagraph(
3605
            num_tokens_padded, disable_full=use_cascade_attn or has_encoder_output
3606
        )
3607
        num_tokens_padded = batch_descriptor.num_tokens
3608
3609
3610
3611
3612
3613
3614
3615
3616
        if self.compilation_config.pass_config.enable_sp:
            assert (
                batch_descriptor.num_tokens
                % self.vllm_config.parallel_config.tensor_parallel_size
                == 0
            ), (
                "Sequence parallelism requires num_tokens to be "
                "a multiple of tensor parallel size"
            )
3617
3618
3619

        # Extra coordination when running data-parallel since we need to coordinate
        # across ranks
3620
        should_ubatch, num_tokens_across_dp = False, None
3621
        if self.vllm_config.parallel_config.data_parallel_size > 1:
3622
3623
3624
3625
3626
3627
3628
3629
3630
3631
            should_ubatch, num_tokens_across_dp, synced_cudagraph_mode = (
                coordinate_batch_across_dp(
                    num_tokens_unpadded=num_tokens,
                    parallel_config=self.parallel_config,
                    allow_microbatching=allow_microbatching,
                    num_tokens_padded=num_tokens_padded,
                    uniform_decode=uniform_decode,
                    num_scheduled_tokens_per_request=num_scheduled_tokens_np,
                    cudagraph_mode=cudagraph_mode.value,
                )
3632
3633
            )

3634
            # Extract DP-synced values
3635
3636
3637
            if num_tokens_across_dp is not None:
                dp_rank = self.parallel_config.data_parallel_rank
                num_tokens_padded = int(num_tokens_across_dp[dp_rank].item())
3638
3639
3640
                # Re-dispatch with DP padding so we have the correct batch_descriptor
                cudagraph_mode, batch_descriptor = dispatch_cudagraph(
                    num_tokens_padded,
3641
                    valid_modes={CUDAGraphMode(synced_cudagraph_mode)},
3642
                )
3643
3644
3645
3646
                # Assert to make sure the agreed upon token count is correct otherwise
                # num_tokens_across_dp will no-longer be valid
                assert batch_descriptor.num_tokens == num_tokens_padded

3647
3648
3649
3650
3651
3652
3653
3654
3655
3656
3657
3658
        cudagraph_stats = None
        if self.vllm_config.observability_config.cudagraph_metrics:
            cudagraph_stats = CUDAGraphStat(
                num_unpadded_tokens=num_tokens,
                num_padded_tokens=batch_descriptor.num_tokens,
                num_paddings=batch_descriptor.num_tokens - num_tokens,
                runtime_mode=str(cudagraph_mode),
            )

        return (
            cudagraph_mode,
            batch_descriptor,
3659
            should_ubatch,
3660
3661
3662
            num_tokens_across_dp,
            cudagraph_stats,
        )
3663

3664
3665
3666
3667
3668
3669
3670
3671
3672
3673
3674
3675
3676
3677
3678
3679
3680
3681
3682
3683
3684
3685
3686
3687
3688
3689
3690
3691
3692
3693
3694
3695
3696
3697
3698
3699
    def _register_layerwise_nvtx_hooks(self) -> None:
        """
        Register layerwise NVTX hooks if --enable-layerwise-nvtx-tracing is enabled
        to trace detailed information of each layer or module in the model.
        """

        if (
            self.vllm_config.observability_config.enable_layerwise_nvtx_tracing
            and not self.layerwise_nvtx_hooks_registered
        ):
            if self.compilation_config.cudagraph_mode != CUDAGraphMode.NONE:
                logger.debug_once(
                    "layerwise NVTX tracing is not supported when CUDA graph is "
                    "turned off; you may observe part or all of the model "
                    "missing NVTX markers"
                )

            # In STOCK_TORCH_COMPILE mode, after registering hooks here,
            # the __call__ function of nn.module will be recompiled with
            # fullgraph=True. Since nvtx.range_push/pop are not traceable
            # by torch dynamo, we can't register hook functions here
            # because hook functions will also be traced by torch dynamo.
            if (
                self.vllm_config.compilation_config.mode
                == CompilationMode.STOCK_TORCH_COMPILE
            ):
                logger.debug_once(
                    "layerwise NVTX tracing is not supported when "
                    "CompilationMode is STOCK_TORCH_COMPILE, skipping "
                    "function hooks registration"
                )
            else:
                pyt_hooks = PytHooks()
                pyt_hooks.register_hooks(self.model, self.model.__class__.__name__)
                self.layerwise_nvtx_hooks_registered = True

3700
3701
3702
3703
3704
3705
3706
3707
3708
3709
3710
3711
3712
3713
3714
3715
3716
3717
3718
3719
3720
3721
3722
3723
3724
3725
3726
3727
3728
3729
3730
3731
3732
3733
3734
3735
3736
3737
3738
3739
3740
3741
3742
3743
3744
3745
3746
3747
3748
3749
3750
3751
3752
3753
3754
3755
3756
3757
3758
3759
3760
3761
3762
3763
3764
3765
3766
3767
3768
3769
3770
3771
3772
3773
    def _get_slot_mappings(
        self,
        num_tokens_padded: int,
        num_reqs_padded: int,
        num_tokens_unpadded: int,
        ubatch_slices: "UBatchSlices | None" = None,
    ) -> tuple[
        dict[int, torch.Tensor] | None,
        dict[str, torch.Tensor] | list[dict[str, torch.Tensor]] | None,
    ]:
        """
        Build slot mappings in both formats needed by the system.

        Args:
            num_tokens_padded: Total number of tokens (padded)
            num_reqs_padded: Total number of requests (padded)
            num_tokens_unpadded: Actual number of tokens (unpadded)
            ubatch_slices: Optional ubatch slicing info for DBO

        Returns:
            A tuple of:
            - slot_mappings_by_gid: dict[int, torch.Tensor] for attention metadata
            - slot_mappings_by_layer: dict[str, torch.Tensor] or list for ForwardContext
        """
        if not (
            hasattr(self, "kv_cache_config")
            and self.kv_cache_config is not None
            and len(self.kv_cache_config.kv_cache_groups) > 0
        ):
            return None, None

        def _get_slot_mapping(kv_cache_gid: int):
            assert num_reqs_padded is not None and num_tokens_padded is not None
            kv_cache_spec = self.kv_cache_config.kv_cache_groups[
                kv_cache_gid
            ].kv_cache_spec
            if isinstance(kv_cache_spec, EncoderOnlyAttentionSpec):
                slot_mapping = torch.zeros(
                    (num_tokens_padded,),
                    dtype=torch.int64,
                    device=self.device,
                )
            else:
                blk_table = self.input_batch.block_table[kv_cache_gid]
                slot_mapping = blk_table.slot_mapping.gpu[:num_tokens_padded]

            # Fill unused with -1. Needed for reshape_and_cache in full cuda
            # graph mode. `blk_table_tensor` -1 to match mamba PAD_SLOT_ID
            slot_mapping[num_tokens_unpadded:num_tokens_padded].fill_(-1)

            return slot_mapping

        slot_mappings_by_gid = {
            gid: _get_slot_mapping(gid)
            for gid, _ in enumerate(self.kv_cache_config.kv_cache_groups)
        }

        slot_mappings_by_layer: dict[str, torch.Tensor] = {}
        for gid, kv_cache_group in enumerate(self.kv_cache_config.kv_cache_groups):
            slot_mapping = slot_mappings_by_gid[gid]
            for layer_name in kv_cache_group.layer_names:
                slot_mappings_by_layer[layer_name] = slot_mapping

        if ubatch_slices is not None:
            result: list[dict[str, torch.Tensor]] = []
            for ubatch in ubatch_slices:
                sliced_mappings: dict[str, torch.Tensor] = {}
                for layer_name, slot_mapping in slot_mappings_by_layer.items():
                    sliced_mappings[layer_name] = slot_mapping[ubatch.token_slice]
                result.append(sliced_mappings)
            return slot_mappings_by_gid, result

        return slot_mappings_by_gid, slot_mappings_by_layer

3774
3775
3776
3777
    @torch.inference_mode()
    def execute_model(
        self,
        scheduler_output: "SchedulerOutput",
3778
        intermediate_tensors: IntermediateTensors | None = None,
3779
    ) -> ModelRunnerOutput | AsyncModelRunnerOutput | IntermediateTensors | None:
3780
3781
3782
3783
3784
        if self.execute_model_state is not None:
            raise RuntimeError(
                "State error: sample_tokens() must be called "
                "after execute_model() returns None."
            )
3785

3786
        if self.routed_experts_initialized:
3787
3788
3789
3790
3791
3792
            capturer = RoutedExpertsCapturer.get_instance()
            if capturer is not None:
                capturer.clear_buffer()  # noqa
            else:
                logger.error("RoutedExpertsCapturer not initialized.")

3793
3794
3795
3796
3797
3798
3799
3800
3801
3802
3803
3804
3805
3806
3807
3808
3809
        # If ngram_gpu is used, we need to copy the scheduler_output to avoid
        # the modification has influence on the scheduler_output in engine core process.
        # The replace is much faster than deepcopy.
        if (
            self.speculative_config is not None
            and self.speculative_config.use_ngram_gpu()
        ):
            num_scheduled_tokens_copy = scheduler_output.num_scheduled_tokens.copy()
            spec_decode_tokens_copy = (
                scheduler_output.scheduled_spec_decode_tokens.copy()
            )
            scheduler_output = replace(
                scheduler_output,
                num_scheduled_tokens=num_scheduled_tokens_copy,
                scheduled_spec_decode_tokens=spec_decode_tokens_copy,
            )

3810
3811
3812
3813
        if has_kv_transfer_group():
            kv_connector_metadata = scheduler_output.kv_connector_metadata
            assert kv_connector_metadata is not None
            get_kv_transfer_group().handle_preemptions(kv_connector_metadata)
3814

3815
        num_scheduled_tokens = scheduler_output.total_num_scheduled_tokens
3816
3817
3818
3819
3820
        with (
            record_function_or_nullcontext("gpu_model_runner: preprocess"),
            self.synchronize_input_prep(),
        ):
            # Update persistent batch states.
3821
            deferred_state_corrections_fn = self._update_states(scheduler_output)
3822

3823
            if has_ec_transfer() and not get_ec_transfer().is_consumer:
3824
                with self.maybe_get_ec_connector_output(
3825
                    scheduler_output,
3826
3827
3828
3829
3830
3831
3832
3833
3834
3835
3836
3837
3838
3839
3840
3841
3842
3843
3844
3845
3846
3847
3848
3849
3850
3851
3852
3853
                    encoder_cache=self.encoder_cache,
                ) as ec_connector_output:
                    self._execute_mm_encoder(scheduler_output)
                    return make_empty_encoder_model_runner_output(scheduler_output)

            if not num_scheduled_tokens:
                if (
                    self.parallel_config.distributed_executor_backend
                    == "external_launcher"
                    and self.parallel_config.data_parallel_size > 1
                ):
                    # this is a corner case when both external launcher
                    # and DP are enabled, num_scheduled_tokens could be
                    # 0, and has_unfinished_requests in the outer loop
                    # returns True. before returning early here we call
                    # dummy run to ensure coordinate_batch_across_dp
                    # is called into to avoid out of sync issues.
                    self._dummy_run(1)
                if not has_kv_transfer_group():
                    # Return empty ModelRunnerOutput if no work to do.
                    return EMPTY_MODEL_RUNNER_OUTPUT
                return self.kv_connector_no_forward(scheduler_output, self.vllm_config)

            if self.cache_config.kv_sharing_fast_prefill:
                assert not self.num_prompt_logprobs, (
                    "--kv-sharing-fast-prefill produces incorrect "
                    "logprobs for prompt tokens, tokens, please disable "
                    "it when the requests need prompt logprobs"
3854
3855
                )

3856
3857
3858
3859
3860
3861
            num_reqs = self.input_batch.num_reqs
            req_ids = self.input_batch.req_ids
            tokens = [scheduler_output.num_scheduled_tokens[i] for i in req_ids]
            num_scheduled_tokens_np = np.array(tokens, dtype=np.int32)
            max_num_scheduled_tokens = int(num_scheduled_tokens_np.max())
            num_tokens_unpadded = scheduler_output.total_num_scheduled_tokens
3862

3863
3864
3865
3866
            logits_indices, spec_decode_metadata = self._prepare_inputs(
                scheduler_output,
                num_scheduled_tokens_np,
            )
3867

3868
3869
3870
3871
3872
            cascade_attn_prefix_lens = None
            # Disable cascade attention when using microbatching (DBO)
            if self.cascade_attn_enabled and not self.parallel_config.use_ubatching:
                # Pre-compute cascade attention prefix lengths
                cascade_attn_prefix_lens = self._compute_cascade_attn_prefix_lens(
3873
                    num_scheduled_tokens_np,
3874
3875
                    self.input_batch.num_computed_tokens_cpu[:num_reqs],
                    scheduler_output.num_common_prefix_blocks,
3876
3877
                )

3878
3879
3880
3881
3882
3883
3884
3885
3886
3887
3888
3889
3890
3891
            (
                cudagraph_mode,
                batch_desc,
                should_ubatch,
                num_tokens_across_dp,
                cudagraph_stats,
            ) = self._determine_batch_execution_and_padding(
                num_tokens=num_tokens_unpadded,
                num_reqs=num_reqs,
                num_scheduled_tokens_np=num_scheduled_tokens_np,
                max_num_scheduled_tokens=max_num_scheduled_tokens,
                use_cascade_attn=cascade_attn_prefix_lens is not None,
                num_encoder_reqs=len(scheduler_output.scheduled_encoder_inputs),
            )
3892

3893
3894
3895
3896
3897
3898
3899
3900
3901
3902
3903
3904
3905
3906
3907
3908
3909
3910
3911
3912
3913
3914
3915
3916
3917
3918
3919
            logger.debug(
                "Running batch with cudagraph_mode: %s, batch_descriptor: %s, "
                "should_ubatch: %s, num_tokens_across_dp: %s",
                cudagraph_mode,
                batch_desc,
                should_ubatch,
                num_tokens_across_dp,
            )

            num_tokens_padded = batch_desc.num_tokens
            num_reqs_padded = (
                batch_desc.num_reqs if batch_desc.num_reqs is not None else num_reqs
            )
            ubatch_slices, ubatch_slices_padded = maybe_create_ubatch_slices(
                should_ubatch,
                num_scheduled_tokens_np,
                num_tokens_padded,
                num_reqs_padded,
                self.parallel_config.num_ubatches,
            )

            logger.debug(
                "ubatch_slices: %s, ubatch_slices_padded: %s",
                ubatch_slices,
                ubatch_slices_padded,
            )

3920
3921
3922
3923
3924
3925
3926
3927
3928
3929
3930
            # True if any attention backend handles KV cache update separately
            # from forward() (i.e., forward_includes_kv_cache_update=False). When true,
            # slot_mappings must use padded dimensions to match the key/value tensors.
            has_separate_kv_update = not all(
                all(
                    g.backend.forward_includes_kv_cache_update
                    for g in self.attn_groups[id]
                )
                for id, spec in enumerate(self.kv_cache_config.kv_cache_groups)
                if not isinstance(spec.kv_cache_spec, EncoderOnlyAttentionSpec)
            )
3931
3932
            pad_attn = cudagraph_mode == CUDAGraphMode.FULL

3933
            if self.cache_config.mamba_cache_mode == "align":
3934
3935
3936
3937
3938
3939
                # preprocess_mamba reads req_state.num_computed_tokens (CPU)
                # to decide copy operations, so we must apply deferred
                # corrections before it runs.
                if deferred_state_corrections_fn:
                    deferred_state_corrections_fn()
                    deferred_state_corrections_fn = None
3940
3941
3942
3943
3944
3945
3946
3947
3948
                mamba_utils.preprocess_mamba(
                    scheduler_output,
                    self.kv_cache_config,
                    self.cache_config,
                    self.mamba_state_idx,
                    self.input_batch,
                    self.requests,
                    self.compilation_config.static_forward_context,
                    self.model.get_mamba_state_copy_func(),
3949
                    self._get_mamba_copy_bufs(),
3950
                )
3951
3952
3953
3954
3955
3956
3957
3958
                # preprocess_mamba resets num_accepted_tokens_cpu to 1
                # for requests whose state was copied to a new block.
                # Re-sync to GPU so the mamba kernel reads from the
                # correct initial state slot (init_token_idx = 0).
                self.num_accepted_tokens.np[:num_reqs] = (
                    self.input_batch.num_accepted_tokens_cpu[:num_reqs]
                )
                self.num_accepted_tokens.copy_to_gpu(num_reqs)
3959

3960
3961
3962
            use_spec_decode = len(scheduler_output.scheduled_spec_decode_tokens) > 0
            ubatch_slices_attn = ubatch_slices_padded if pad_attn else ubatch_slices

3963
3964
3965
3966
3967
3968
3969
3970
3971
3972
3973
            slot_mappings_by_group, slot_mappings = self._get_slot_mappings(
                num_tokens_padded=num_tokens_padded
                if pad_attn or has_separate_kv_update
                else num_tokens_unpadded,
                num_reqs_padded=(
                    num_reqs_padded if pad_attn or has_separate_kv_update else num_reqs
                ),
                num_tokens_unpadded=num_tokens_unpadded,
                ubatch_slices=ubatch_slices_padded,
            )

3974
3975
3976
3977
3978
3979
3980
3981
3982
3983
3984
3985
            attn_metadata, spec_decode_common_attn_metadata = (
                self._build_attention_metadata(
                    num_tokens=num_tokens_unpadded,
                    num_tokens_padded=num_tokens_padded if pad_attn else None,
                    num_reqs=num_reqs,
                    num_reqs_padded=num_reqs_padded if pad_attn else None,
                    max_query_len=max_num_scheduled_tokens,
                    ubatch_slices=ubatch_slices_attn,
                    logits_indices=logits_indices,
                    use_spec_decode=use_spec_decode,
                    num_scheduled_tokens=scheduler_output.num_scheduled_tokens,
                    cascade_attn_prefix_lens=cascade_attn_prefix_lens,
3986
                    slot_mappings=slot_mappings_by_group,
3987
                )
3988
            )
3989

3990
3991
3992
3993
3994
3995
3996
3997
3998
            (
                input_ids,
                inputs_embeds,
                positions,
                intermediate_tensors,
                model_kwargs,
                ec_connector_output,
            ) = self._preprocess(
                scheduler_output, num_tokens_padded, intermediate_tensors
3999
            )
4000

4001
        # Set cudagraph mode to none if calc_kv_scales is true.
4002
4003
4004
        # KV scales calculation involves dynamic operations that are incompatible
        # with CUDA graph capture.
        if self.calculate_kv_scales:
4005
            cudagraph_mode = CUDAGraphMode.NONE
4006
4007
            # Mark KV scales as calculated after the first forward pass
            self.calculate_kv_scales = False
4008

4009
4010
4011
4012
4013
4014
4015
        # Encoder-decoder models can only compile the pure decode steps where no
        # encoder inputs are present. Use eager for the first pass.
        num_encoder_reqs = len(scheduler_output.scheduled_encoder_inputs)
        has_encoder_input = (
            self.model_config.is_encoder_decoder and num_encoder_reqs > 0
        )

4016
4017
        # Run the model.
        # Use persistent buffers for CUDA graphs.
4018
4019
4020
        # When spec decode is enabled, defer connector finalization
        # (wait_for_save + clear metadata) until after draft model runs.
        defer_kv_connector_finalize = self.speculative_config is not None
4021
4022
        with (
            set_forward_context(
4023
4024
                attn_metadata,
                self.vllm_config,
4025
                num_tokens=num_tokens_padded,
4026
                num_tokens_across_dp=num_tokens_across_dp,
4027
4028
                cudagraph_runtime_mode=cudagraph_mode,
                batch_descriptor=batch_desc,
4029
                ubatch_slices=ubatch_slices_padded,
4030
                slot_mapping=slot_mappings,
4031
                skip_compiled=has_encoder_input,
4032
            ),
4033
            record_function_or_nullcontext("gpu_model_runner: forward"),
4034
            self.maybe_get_kv_connector_output(
4035
4036
                scheduler_output,
                defer_finalize=defer_kv_connector_finalize,
4037
            ) as kv_connector_output,
4038
        ):
4039
            model_output = self._model_forward(
4040
4041
4042
4043
4044
4045
4046
                input_ids=input_ids,
                positions=positions,
                intermediate_tensors=intermediate_tensors,
                inputs_embeds=inputs_embeds,
                **model_kwargs,
            )

4047
        with record_function_or_nullcontext("gpu_model_runner: postprocess"):
4048
            if self.use_aux_hidden_state_outputs:
4049
                # True when EAGLE 3 is used.
4050
4051
                hidden_states, aux_hidden_states = model_output
            else:
4052
                # Common case.
4053
4054
4055
                hidden_states = model_output
                aux_hidden_states = None

4056
4057
4058
4059
4060
            if not self.broadcast_pp_output:
                # Common case.
                if not get_pp_group().is_last_rank:
                    # Return the intermediate tensors.
                    assert isinstance(hidden_states, IntermediateTensors)
4061
                    hidden_states.kv_connector_output = kv_connector_output
4062
                    self.kv_connector_output = kv_connector_output
4063
                    return hidden_states
4064

4065
                if self.is_pooling_model:
4066
                    # Return the pooling output.
4067
4068
4069
4070
4071
                    return self._pool(
                        hidden_states,
                        num_scheduled_tokens,
                        num_scheduled_tokens_np,
                        kv_connector_output,
4072
                    )
4073
4074

                sample_hidden_states = hidden_states[logits_indices]
4075
                logits = self.model.compute_logits(sample_hidden_states)
4076
4077
4078
4079
            else:
                # Rare case.
                assert not self.is_pooling_model

4080
                sample_hidden_states = hidden_states[logits_indices]
4081
                if not get_pp_group().is_last_rank:
4082
                    all_gather_tensors = {
4083
                        "residual": not is_residual_scattered_for_sp(
4084
                            self.vllm_config, num_tokens_padded
4085
                        )
4086
                    }
4087
                    get_pp_group().send_tensor_dict(
4088
4089
                        hidden_states.tensors,
                        all_gather_group=get_tp_group(),
4090
4091
                        all_gather_tensors=all_gather_tensors,
                    )
4092
4093
                    logits = None
                else:
4094
                    logits = self.model.compute_logits(sample_hidden_states)
4095

4096
                model_output_broadcast_data: dict[str, Any] = {}
4097
4098
4099
                if logits is not None:
                    model_output_broadcast_data["logits"] = logits.contiguous()

4100
                broadcasted = get_pp_group().broadcast_tensor_dict(
4101
4102
                    model_output_broadcast_data, src=len(get_pp_group().ranks) - 1
                )
4103
4104
                assert broadcasted is not None
                logits = broadcasted["logits"]
4105

4106
4107
4108
4109
4110
4111
4112
4113
        self.execute_model_state = ExecuteModelState(
            scheduler_output,
            logits,
            spec_decode_metadata,
            spec_decode_common_attn_metadata,
            hidden_states,
            sample_hidden_states,
            aux_hidden_states,
4114
            ec_connector_output,
4115
            cudagraph_stats,
4116
            slot_mappings,
4117
        )
4118
        self.kv_connector_output = kv_connector_output
4119
4120
4121
4122
4123
4124

        # Now the batch has been launched we can wait for corrections from the
        # previous model forward without breaking async scheduling.
        if deferred_state_corrections_fn:
            deferred_state_corrections_fn()

4125
4126
4127
4128
4129
4130
4131
        return None

    @torch.inference_mode
    def sample_tokens(
        self, grammar_output: "GrammarOutput | None"
    ) -> ModelRunnerOutput | AsyncModelRunnerOutput | IntermediateTensors:
        if self.execute_model_state is None:
4132
4133
            kv_connector_output = self.kv_connector_output
            self.kv_connector_output = None
4134
4135
4136
            # receive sampled token ids from the last PP rank.
            if self.use_async_scheduling and get_pp_group().world_size > 1:
                self._pp_receive_prev_sampled_token_ids_to_input_batch()
4137
            if not kv_connector_output:
4138
                return None  # type: ignore[return-value]
4139
4140
4141
4142
4143
4144
4145
4146
4147

            # In case of PP with kv transfer, we need to pass through the
            # kv_connector_output
            if kv_connector_output.is_empty():
                return EMPTY_MODEL_RUNNER_OUTPUT

            output = copy(EMPTY_MODEL_RUNNER_OUTPUT)
            output.kv_connector_output = kv_connector_output
            return output
4148
4149
4150
4151
4152
4153
4154
4155
4156
4157

        # Unpack ephemeral state.
        (
            scheduler_output,
            logits,
            spec_decode_metadata,
            spec_decode_common_attn_metadata,
            hidden_states,
            sample_hidden_states,
            aux_hidden_states,
4158
            ec_connector_output,
4159
            cudagraph_stats,
4160
            slot_mappings,
4161
4162
4163
4164
4165
4166
4167
4168
4169
        ) = self.execute_model_state
        # Clear ephemeral state.
        self.execute_model_state = None

        # Apply structured output bitmasks if present.
        if grammar_output is not None:
            apply_grammar_bitmask(
                scheduler_output, grammar_output, self.input_batch, logits
            )
4170

4171
        with record_function_or_nullcontext("gpu_model_runner: sample"):
4172
4173
            sampler_output = self._sample(logits, spec_decode_metadata)

4174
4175
4176
        self._update_states_after_model_execute(
            sampler_output.sampled_token_ids, scheduler_output
        )
4177
4178
        if self.use_async_scheduling:
            pp = get_pp_group()
4179
4180
4181
4182
            # For torchrun external_launcher PP mode with broadcast_pp_output=True,
            # PP outputs have been broadcasted to all ranks at logits computation.
            # Therefore, here is no need to send sampled token ids again in this case.
            if not self.broadcast_pp_output and pp.world_size > 1 and pp.is_last_rank:
4183
4184
4185
                self._pp_broadcast_prev_sampled_token_ids(
                    sampler_output.sampled_token_ids
                )
4186

4187
4188
        self._draft_token_ids = None
        self._draft_token_req_ids = None
4189
        self.valid_sampled_token_count_gpu = None
4190
4191
        self.input_batch.prev_sampled_token_ids = None

4192
        def propose_draft_token_ids(sampled_token_ids):
4193
            assert spec_decode_common_attn_metadata is not None
4194
            with record_function_or_nullcontext("gpu_model_runner: draft"):
4195
4196
4197
4198
4199
4200
4201
4202
4203
                self._draft_token_ids = self.propose_draft_token_ids(
                    scheduler_output,
                    sampled_token_ids,
                    self.input_batch.sampling_metadata,
                    hidden_states,
                    sample_hidden_states,
                    aux_hidden_states,
                    spec_decode_metadata,
                    spec_decode_common_attn_metadata,
4204
                    slot_mappings,
4205
                )
4206
                self._copy_draft_token_ids_to_cpu(scheduler_output)
4207

4208
        spec_config = self.speculative_config
4209
4210
4211
4212
4213
        propose_drafts_after_bookkeeping = False
        if spec_config is not None:
            input_fits_in_drafter = spec_decode_common_attn_metadata is not None and (
                spec_decode_common_attn_metadata.max_seq_len + self.num_spec_tokens
                <= self.effective_drafter_max_model_len
4214
            )
4215
            use_gpu_toks = (
4216
4217
4218
                spec_config.use_eagle()
                or spec_config.uses_draft_model()
                or spec_config.uses_extract_hidden_states()
4219
4220
4221
            ) and not spec_config.disable_padded_drafter_batch
            if use_gpu_toks:
                # EAGLE/DraftModel speculative decoding can use the GPU sampled tokens
4222
                # as inputs, and does not need to wait for bookkeeping to finish.
4223
4224
                assert isinstance(
                    self.drafter,
4225
4226
4227
4228
                    EagleProposer
                    | DFlashProposer
                    | DraftModelProposer
                    | ExtractHiddenStatesProposer,
4229
                )
4230
4231
4232
4233
4234
4235
4236
4237
4238
4239
4240
4241
                sampled_token_ids = sampler_output.sampled_token_ids
                if input_fits_in_drafter:
                    propose_draft_token_ids(sampled_token_ids)
                elif self.valid_sampled_token_count_event is not None:
                    assert spec_decode_common_attn_metadata is not None
                    next_token_ids, valid_sampled_tokens_count = (
                        self.drafter.prepare_next_token_ids_padded(
                            sampled_token_ids,
                            self.requests,
                            self.input_batch,
                            self.discard_request_mask.gpu,
                        )
4242
                    )
4243
4244
4245
                    self._copy_valid_sampled_token_count(
                        next_token_ids, valid_sampled_tokens_count
                    )
4246
4247
4248
4249
4250
4251
4252
4253
4254
4255
4256
4257
4258
4259
4260
4261
4262
4263
4264
4265
4266
4267
4268
4269
4270
4271
                    self._draft_token_ids = torch.zeros(
                        1, device=self.device, dtype=torch.int32
                    ).expand(len(self.input_batch.req_ids), self.num_spec_tokens)
                    self._copy_draft_token_ids_to_cpu(scheduler_output, zeros_only=True)
            elif (
                spec_config.use_ngram_gpu()
                and not spec_config.disable_padded_drafter_batch
            ):
                assert isinstance(self.drafter, NgramProposerGPU)
                sampled_token_ids = sampler_output.sampled_token_ids
                if input_fits_in_drafter:
                    propose_draft_token_ids(sampled_token_ids)
                elif self.valid_sampled_token_count_event is not None:
                    assert spec_decode_common_attn_metadata is not None
                    next_token_ids, valid_sampled_tokens_count, _ = (
                        self.drafter.update_token_ids_ngram(
                            sampled_token_ids,
                            self.input_batch,
                            self.token_ids_gpu_tensor,
                            self.num_tokens_no_spec_gpu,
                            self.discard_request_mask.gpu,
                        )
                    )
                    self._copy_valid_sampled_token_count(
                        next_token_ids, valid_sampled_tokens_count
                    )
4272
4273
4274
4275
4276
4277
4278
4279
                    # Since we couldn't run the drafter,
                    # just use zeros for the draft tokens.
                    self._draft_token_ids = torch.zeros(
                        1, device=self.device, dtype=torch.int32
                    ).expand(len(self.input_batch.req_ids), self.num_spec_tokens)
                    self._copy_draft_token_ids_to_cpu(scheduler_output, zeros_only=True)
            else:
                propose_drafts_after_bookkeeping = input_fits_in_drafter
4280

4281
        with record_function_or_nullcontext("gpu_model_runner: bookkeep"):
4282
4283
4284
4285
4286
4287
4288
4289
            (
                num_nans_in_logits,
                logprobs_lists,
                valid_sampled_token_ids,
                prompt_logprobs_dict,
                req_ids_output_copy,
                req_id_to_index_output_copy,
                invalid_req_indices,
4290
4291
4292
4293
4294
            ) = self._bookkeeping_sync(
                scheduler_output,
                sampler_output,
                logits,
                hidden_states,
4295
                scheduler_output.total_num_scheduled_tokens,
4296
                spec_decode_metadata,
4297
            )
4298

4299
        if propose_drafts_after_bookkeeping:
4300
4301
4302
            # ngram and other speculative decoding methods use the sampled
            # tokens on the CPU, so they are run after bookkeeping.
            propose_draft_token_ids(valid_sampled_token_ids)
4303

4304
4305
4306
4307
4308
        # Finalize KV connector (wait_for_save + clear metadata) after
        # draft model runs. Deferred from target model forward to allow
        # draft model to also save its KV cache.
        if spec_config is not None:
            self.finalize_kv_connector()
4309

4310
        with record_function_or_nullcontext("gpu_model_runner: eplb"):
4311
            self.eplb_step()
4312

4313
4314
4315
4316
        # self.kv_connector_output may be modified during drafting
        kv_connector_output = self.kv_connector_output
        self.kv_connector_output = None

4317
        with record_function_or_nullcontext("gpu_model_runner: ModelRunnerOutput"):
4318
            if self.routed_experts_initialized:
4319
4320
4321
4322
4323
4324
                capturer = RoutedExpertsCapturer.get_instance()
                if capturer is not None:
                    capturer.save_captured_experts(indices=self.slot_mapping)  # noqa
                else:
                    logger.error("RoutedExpertsCapturer not initialized.")

4325
4326
4327
4328
4329
4330
4331
            output = ModelRunnerOutput(
                req_ids=req_ids_output_copy,
                req_id_to_index=req_id_to_index_output_copy,
                sampled_token_ids=valid_sampled_token_ids,
                logprobs=logprobs_lists,
                prompt_logprobs_dict=prompt_logprobs_dict,
                kv_connector_output=kv_connector_output,
4332
4333
4334
                ec_connector_output=ec_connector_output
                if self.supports_mm_inputs
                else None,
4335
                num_nans_in_logits=num_nans_in_logits,
4336
                cudagraph_stats=cudagraph_stats,
4337
            )
4338

4339
4340
        if not self.use_async_scheduling:
            return output
4341

4342
4343
4344
4345
4346
4347
4348
4349
4350
        with record_function_or_nullcontext(
            "gpu_model_runner: AsyncGPUModelRunnerOutput"
        ):
            async_output = AsyncGPUModelRunnerOutput(
                model_runner_output=output,
                sampled_token_ids=sampler_output.sampled_token_ids,
                logprobs_tensors=sampler_output.logprobs_tensors,
                invalid_req_indices=invalid_req_indices,
                async_output_copy_stream=self.async_output_copy_stream,
4351
                vocab_size=self.input_batch.vocab_size,
4352
4353
4354
4355
4356
            )
        with record_function_or_nullcontext(
            "gpu_model_runner: set_async_sampled_token_ids"
        ):
            # Save ref of sampled_token_ids CPU tensor if the batch contains
4357
            # any requests with sampling params that require output ids.
4358
4359
4360
4361
            self.input_batch.set_async_sampled_token_ids(
                async_output.sampled_token_ids_cpu,
                async_output.async_copy_ready_event,
            )
4362
4363
4364

        return async_output

4365
4366
4367
4368
4369
4370
4371
4372
4373
4374
4375
4376
4377
4378
4379
4380
4381
4382
4383
4384
4385
4386
4387
4388
4389
4390
4391
4392
4393
4394
4395
4396
4397
4398
4399
4400
4401
4402
4403
    def _pp_broadcast_prev_sampled_token_ids(
        self, sampled_token_ids: torch.Tensor
    ) -> None:
        """Broadcast sampled token ids (GPU) from last PP stage"""
        pp = get_pp_group()
        assert pp.is_last_rank
        # `prev_sampled_token_ids` is expected to have shape [num_reqs, 1].
        assert sampled_token_ids.dim() == 2 and sampled_token_ids.shape[-1] == 1, (
            "PP+async expects sampled_token_ids to have shape [num_reqs, 1]"
        )
        torch.distributed.broadcast(
            sampled_token_ids, src=pp.rank, group=pp.device_group
        )

    def _pp_receive_prev_sampled_token_ids_to_input_batch(self) -> None:
        """Receive sampled token ids broadcast from last PP stage"""
        pp = get_pp_group()
        assert not pp.is_last_rank
        num_reqs = self.input_batch.num_reqs
        # `prev_sampled_token_ids` is expected to have shape [num_reqs, 1].
        recv = torch.empty((num_reqs, 1), dtype=torch.int32, device=self.device)
        torch.distributed.broadcast(recv, src=pp.last_rank, group=pp.device_group)
        self.input_batch.prev_sampled_token_ids = recv

        # construct `prev_req_id_to_index` here so `_prepare_input_ids`
        # can map req_id -> previous batch row
        discard_req_indices = np.nonzero(self.discard_request_mask.np[:num_reqs])[0]
        discard_req_indices_set = set(discard_req_indices)
        prev_req_id_to_index: dict[str, int] = {}
        for i, req_id in enumerate(self.input_batch.req_ids):
            if i in discard_req_indices_set:
                continue
            prev_req_id_to_index[req_id] = i
            # PP+async scheduling: advance per-request local cached output length by
            # appending a placeholder (-1) token id.
            if (req_state := self.requests.get(req_id)) is not None:
                req_state.output_token_ids.append(-1)
        self.input_batch.prev_req_id_to_index = prev_req_id_to_index

4404
    def take_draft_token_ids(self) -> DraftTokenIds | None:
4405
        if not self.num_spec_tokens or not self._draft_token_req_ids:
4406
            return None
4407
        draft_token_ids, req_ids = self._get_draft_token_ids_cpu()
4408
        return DraftTokenIds(req_ids, draft_token_ids)
4409

4410
4411
4412
    def _copy_draft_token_ids_to_cpu(
        self, scheduler_output: "SchedulerOutput", zeros_only: bool = False
    ) -> None:
4413
4414
4415
4416
4417
4418
        # Check if we need to copy draft tokens to CPU. In async scheduling,
        # we only copy when needed for structured output, penalties or bad_words.
        if self.use_async_scheduling and not (
            scheduler_output.has_structured_output_requests
            or self.input_batch.sampling_metadata.output_token_ids
        ):
4419
4420
4421
            return
        # We must also set the corresponding request ids.
        self._draft_token_req_ids = self.input_batch.req_ids.copy()
4422

4423
4424
4425
4426
4427
4428
4429
4430
4431
4432
4433
4434
4435
4436
4437
4438
4439
4440
4441
4442
        draft_token_ids: torch.Tensor = self._draft_token_ids
        if not torch.is_tensor(draft_token_ids):
            return
        assert self.draft_token_ids_event is not None
        assert self.draft_token_ids_copy_stream is not None
        assert self.draft_token_ids_cpu is not None
        default_stream = torch.cuda.current_stream()
        num_reqs = draft_token_ids.shape[0]
        with torch.cuda.stream(self.draft_token_ids_copy_stream):
            if not zeros_only:
                # Trigger async copy of draft token ids to cpu.
                self.draft_token_ids_copy_stream.wait_stream(default_stream)
                self.draft_token_ids_cpu[:num_reqs].copy_(
                    draft_token_ids, non_blocking=True
                )
            else:
                # No copy needed, just zero-out cpu tensor.
                self.draft_token_ids_cpu[:num_reqs] = 0
            self.draft_token_ids_event.record()

4443
    def _get_draft_token_ids_cpu(self) -> tuple[list[list[int]], list[str]]:
4444
        if isinstance(self._draft_token_ids, list):
4445
4446
4447
4448
            return self._draft_token_ids, self.input_batch.req_ids
        req_ids = self._draft_token_req_ids
        if req_ids is None:
            return [], []
4449
4450
4451
        assert self.draft_token_ids_event is not None
        assert self.draft_token_ids_cpu is not None
        self.draft_token_ids_event.synchronize()
4452
        return self.draft_token_ids_cpu[: len(req_ids)].tolist(), req_ids
4453

4454
4455
4456
4457
4458
4459
4460
4461
4462
4463
4464
4465
4466
    def _copy_valid_sampled_token_count(
        self, next_token_ids: torch.Tensor, valid_sampled_tokens_count: torch.Tensor
    ) -> None:
        if self.valid_sampled_token_count_event is None:
            return

        default_stream = torch.cuda.current_stream()
        # Initialize a new stream to overlap the copy operation with
        # prepare_input of draft model.
        with torch.cuda.stream(self.valid_sampled_token_count_copy_stream):
            self.valid_sampled_token_count_copy_stream.wait_stream(default_stream)  # type: ignore
            counts = valid_sampled_tokens_count
            counts_cpu = self.valid_sampled_token_count_cpu
4467
            assert counts_cpu is not None
4468
4469
4470
            counts_cpu[: counts.shape[0]].copy_(counts, non_blocking=True)
            self.valid_sampled_token_count_event.record()

4471
4472
4473
        if self.use_async_spec_decode:
            # Stash for GPU-side correction in _prepare_inputs.
            self.valid_sampled_token_count_gpu = valid_sampled_tokens_count
4474
4475
4476
4477
4478
        self.input_batch.prev_sampled_token_ids = next_token_ids.unsqueeze(1)

    def _get_valid_sampled_token_count(self) -> list[int]:
        # Wait until valid_sampled_tokens_count is copied to cpu,
        prev_sampled_token_ids = self.input_batch.prev_sampled_token_ids
4479
4480
        sampled_count_event = self.valid_sampled_token_count_event
        if sampled_count_event is None or prev_sampled_token_ids is None:
4481
4482
4483
            return []

        counts_cpu = self.valid_sampled_token_count_cpu
4484
4485
        assert counts_cpu is not None
        sampled_count_event.synchronize()
4486
4487
        return counts_cpu[: prev_sampled_token_ids.shape[0]].tolist()

4488
4489
4490
    def propose_draft_token_ids(
        self,
        scheduler_output: "SchedulerOutput",
4491
        sampled_token_ids: torch.Tensor | list[list[int]],
4492
4493
4494
        sampling_metadata: SamplingMetadata,
        hidden_states: torch.Tensor,
        sample_hidden_states: torch.Tensor,
4495
4496
        aux_hidden_states: list[torch.Tensor] | None,
        spec_decode_metadata: SpecDecodeMetadata | None,
4497
        common_attn_metadata: CommonAttentionMetadata,
4498
        slot_mappings: dict[str, torch.Tensor] | list[dict[str, torch.Tensor]] | None,
4499
    ) -> list[list[int]] | torch.Tensor:
4500
        num_scheduled_tokens = scheduler_output.total_num_scheduled_tokens
4501
4502
4503
        spec_config = self.speculative_config
        assert spec_config is not None
        if spec_config.method == "ngram":
4504
4505
            from vllm.v1.spec_decode.ngram_proposer import NgramProposer

4506
            assert isinstance(sampled_token_ids, list)
4507
            assert isinstance(self.drafter, NgramProposer)
4508
            draft_token_ids = self.drafter.propose(
4509
                sampled_token_ids,
4510
4511
                self.input_batch.num_tokens_no_spec,
                self.input_batch.token_ids_cpu,
4512
                slot_mappings=slot_mappings,
4513
            )
4514
4515
4516
4517
4518
4519
4520
4521
4522
4523
4524
4525
4526
4527
4528
4529
4530
4531
4532
4533
4534
4535
4536
4537
4538
4539
4540
4541
4542
4543
4544
4545
4546
4547
4548
4549
4550
        elif spec_config.use_ngram_gpu():
            assert isinstance(self.drafter, NgramProposerGPU)
            (
                next_token_ids,
                valid_sampled_tokens_count,
                valid_sampled_token_ids_gpu,
            ) = self.drafter.update_token_ids_ngram(
                sampled_token_ids,
                self.input_batch,
                self.token_ids_gpu_tensor,
                self.num_tokens_no_spec_gpu,
                self.discard_request_mask.gpu,
            )
            self._copy_valid_sampled_token_count(
                next_token_ids, valid_sampled_tokens_count
            )

            batch_size = next_token_ids.shape[0]

            draft_token_ids, num_valid_draft_tokens = self.drafter.propose(
                self.num_tokens_no_spec_gpu[:batch_size],
                self.token_ids_gpu_tensor[:batch_size],
                valid_sampled_token_ids_gpu,
                valid_sampled_tokens_count,
            )

            # Cache valid draft counts for scheduler-side trimming.
            self._num_valid_draft_tokens = num_valid_draft_tokens

            # Async D2H copy on a dedicated stream.
            copy_num_valid_draft_tokens(
                self._num_valid_draft_tokens_cpu,
                self._num_valid_draft_tokens_copy_stream,
                self._num_valid_draft_tokens_event,
                self._num_valid_draft_tokens,
                self.input_batch.num_reqs,
            )
4551
        elif spec_config.method == "suffix":
4552
4553
            assert isinstance(sampled_token_ids, list)
            assert isinstance(self.drafter, SuffixDecodingProposer)
4554
4555
4556
            draft_token_ids = self.drafter.propose(
                self.input_batch, sampled_token_ids, slot_mappings=slot_mappings
            )
4557
        elif spec_config.method == "medusa":
4558
            assert isinstance(sampled_token_ids, list)
4559
            assert isinstance(self.drafter, MedusaProposer)
4560

4561
4562
            if sample_hidden_states.shape[0] == len(sampled_token_ids):
                # The input to the target model does not include draft tokens.
4563
4564
4565
4566
                hidden_states = sample_hidden_states
            else:
                indices = []
                offset = 0
4567
4568
4569
                assert spec_decode_metadata is not None, (
                    "No spec decode metadata for medusa"
                )
4570
                for num_draft, tokens in zip(
4571
4572
                    spec_decode_metadata.num_draft_tokens, sampled_token_ids
                ):
4573
                    indices.append(offset + len(tokens) - 1)
4574
                    offset += num_draft + 1
4575
                indices = torch.tensor(indices, device=self.device)
4576
4577
                hidden_states = sample_hidden_states[indices]

4578
            draft_token_ids = self.drafter.propose(
4579
4580
                target_hidden_states=hidden_states,
                sampling_metadata=sampling_metadata,
4581
                slot_mappings=slot_mappings,
4582
            )
4583
4584
4585
4586
4587
4588
4589
4590
4591
4592
4593
4594
        elif spec_config.uses_extract_hidden_states():
            assert isinstance(self.drafter, ExtractHiddenStatesProposer)
            assert isinstance(sampled_token_ids, torch.Tensor), (
                "sampled_token_ids should be a torch.Tensor for "
                "extract_hidden_states method."
            )
            if not self.use_aux_hidden_state_outputs or aux_hidden_states is None:
                raise ValueError(
                    "aux_hidden_states are required when using `extract_hidden_states`"
                )
            target_hidden_states = [h[:num_scheduled_tokens] for h in aux_hidden_states]

4595
            draft_token_ids = self.drafter.propose(
4596
4597
4598
4599
4600
4601
4602
4603
4604
4605
4606
4607
4608
4609
4610
4611
4612
                sampled_token_ids=sampled_token_ids,
                target_hidden_states=target_hidden_states,
                common_attn_metadata=common_attn_metadata,
                slot_mappings=slot_mappings,
            )
            next_token_ids, valid_sampled_tokens_count = (
                self.drafter.prepare_next_token_ids_padded(
                    sampled_token_ids,
                    self.requests,
                    self.input_batch,
                    self.discard_request_mask.gpu,
                )
            )
            self._copy_valid_sampled_token_count(
                next_token_ids, valid_sampled_tokens_count
            )

4613
4614
4615
4616
4617
4618
4619
4620
        elif (
            spec_config.use_eagle()
            or spec_config.use_dflash()
            or spec_config.uses_draft_model()
        ):
            assert isinstance(
                self.drafter, EagleProposer | DFlashProposer | DraftModelProposer
            )
4621

4622
            if spec_config.disable_padded_drafter_batch:
4623
4624
4625
                # When padded-batch is disabled, the sampled_token_ids should be
                # the cpu-side list[list[int]] of valid sampled tokens for each
                # request, with invalid requests having empty lists.
4626
4627
                assert isinstance(sampled_token_ids, list), (
                    "sampled_token_ids should be a python list when"
4628
                    "padded-batch is disabled."
4629
                )
4630
                next_token_ids = self.drafter.prepare_next_token_ids_cpu(
4631
4632
4633
4634
4635
                    sampled_token_ids,
                    self.requests,
                    self.input_batch,
                    scheduler_output.num_scheduled_tokens,
                )
4636
4637
4638
4639
4640
            else:
                # When using padded-batch, the sampled_token_ids should be
                # the gpu tensor of sampled tokens for each request, of shape
                # (num_reqs, num_spec_tokens + 1) with rejected tokens having
                # value -1.
4641
4642
                assert isinstance(sampled_token_ids, torch.Tensor), (
                    "sampled_token_ids should be a torch.Tensor when"
4643
                    "padded-batch is enabled."
4644
4645
                )
                next_token_ids, valid_sampled_tokens_count = (
4646
4647
4648
4649
                    self.drafter.prepare_next_token_ids_padded(
                        sampled_token_ids,
                        self.requests,
                        self.input_batch,
4650
                        self.discard_request_mask.gpu,
4651
                    )
4652
                )
4653
4654
4655
                self._copy_valid_sampled_token_count(
                    next_token_ids, valid_sampled_tokens_count
                )
Jiayi Yao's avatar
Jiayi Yao committed
4656

4657
            num_rejected_tokens_gpu = None
4658
            if spec_decode_metadata is None:
4659
                token_indices_to_sample = None
4660
                # input_ids can be None for multimodal models.
4661
                target_token_ids = self.input_ids.gpu[:num_scheduled_tokens]
4662
                target_positions = self._get_positions(num_scheduled_tokens)
4663
                if self.use_aux_hidden_state_outputs:
Wentao Ye's avatar
Wentao Ye committed
4664
                    assert aux_hidden_states is not None
4665
                    target_hidden_states = torch.cat(
4666
4667
                        [h[:num_scheduled_tokens] for h in aux_hidden_states], dim=-1
                    )
4668
4669
                else:
                    target_hidden_states = hidden_states[:num_scheduled_tokens]
4670
            else:
4671
                if spec_config.disable_padded_drafter_batch:
4672
                    token_indices_to_sample = None
4673
4674
4675
4676
4677
                    common_attn_metadata, token_indices = self.drafter.prepare_inputs(
                        common_attn_metadata,
                        sampled_token_ids,
                        spec_decode_metadata.num_draft_tokens,
                    )
4678
4679
4680
4681
4682
4683
4684
4685
4686
                    target_token_ids = self.input_ids.gpu[token_indices]
                    target_positions = self._get_positions(token_indices)
                    if self.use_aux_hidden_state_outputs:
                        assert aux_hidden_states is not None
                        target_hidden_states = torch.cat(
                            [h[token_indices] for h in aux_hidden_states], dim=-1
                        )
                    else:
                        target_hidden_states = hidden_states[token_indices]
4687
                else:
4688
4689
4690
4691
4692
4693
4694
4695
                    (
                        common_attn_metadata,
                        token_indices_to_sample,
                        num_rejected_tokens_gpu,
                    ) = self.drafter.prepare_inputs_padded(
                        common_attn_metadata,
                        spec_decode_metadata,
                        valid_sampled_tokens_count,
4696
                    )
4697
4698
4699
4700
4701
4702
4703
4704
4705
4706
4707
                    total_num_tokens = common_attn_metadata.num_actual_tokens
                    # When padding the batch, token_indices is just a range
                    target_token_ids = self.input_ids.gpu[:total_num_tokens]
                    target_positions = self._get_positions(total_num_tokens)
                    if self.use_aux_hidden_state_outputs:
                        assert aux_hidden_states is not None
                        target_hidden_states = torch.cat(
                            [h[:total_num_tokens] for h in aux_hidden_states], dim=-1
                        )
                    else:
                        target_hidden_states = hidden_states[:total_num_tokens]
4708

4709
            if self.supports_mm_inputs and self.drafter.supports_mm_inputs:
4710
4711
4712
4713
4714
4715
                mm_embed_inputs = self._gather_mm_embeddings(
                    scheduler_output,
                    shift_computed_tokens=1,
                )
            else:
                mm_embed_inputs = None
4716

4717
            draft_token_ids = self.drafter.propose(
4718
4719
4720
4721
                target_token_ids=target_token_ids,
                target_positions=target_positions,
                target_hidden_states=target_hidden_states,
                next_token_ids=next_token_ids,
4722
                token_indices_to_sample=token_indices_to_sample,
4723
                sampling_metadata=sampling_metadata,
4724
                common_attn_metadata=common_attn_metadata,
4725
                mm_embed_inputs=mm_embed_inputs,
4726
                num_rejected_tokens_gpu=num_rejected_tokens_gpu,
4727
                slot_mappings=slot_mappings,
4728
            )
4729

4730
        return draft_token_ids
4731

4732
4733
4734
    def update_config(self, overrides: dict[str, Any]) -> None:
        allowed_config_names = {"load_config", "model_config"}
        for config_name, config_overrides in overrides.items():
4735
4736
            assert config_name in allowed_config_names, (
                f"Config `{config_name}` not supported. "
4737
                f"Allowed configs: {allowed_config_names}"
4738
            )
4739
4740
4741
4742
            config = getattr(self, config_name)
            new_config = update_config(config, config_overrides)
            setattr(self, config_name, new_config)

4743
    @instrument(span_name="Loading (GPU)")
4744
    def load_model(self, load_dummy_weights: bool = False) -> None:
4745
4746
        """
        Args:
4747
            load_dummy_weights: load dummy weights instead of real weights.
4748
        """
4749
4750
4751
4752
4753
        logger.info_once(
            "Starting to load model %s...",
            self.model_config.model,
            scope="global",
        )
4754

4755
4756
4757
4758
        if self.parallel_config.enable_eplb:
            self.eplb_state = EplbState(self.parallel_config, self.device)
            eplb_models = 0

4759
4760
4761
        try:
            with DeviceMemoryProfiler() as m:
                time_before_load = time.perf_counter()
4762
4763
                if load_dummy_weights:
                    self.load_config.load_format = "dummy"
4764
4765
4766
4767
4768
4769
4770
                model_loader = get_model_loader(self.load_config)
                self.model = model_loader.load_model(
                    vllm_config=self.vllm_config, model_config=self.model_config
                )
                if self.lora_config:
                    self.model = self.load_lora_model(
                        self.model, self.vllm_config, self.device
4771
                    )
4772
4773
4774
4775
4776
4777
4778
4779
                if hasattr(self, "drafter"):
                    logger.info_once("Loading drafter model...")
                    self.drafter.load_model(self.model)
                    if (
                        hasattr(self.drafter, "model")
                        and is_mixture_of_experts(self.drafter.model)
                        and self.parallel_config.enable_eplb
                    ):
4780
4781
4782
                        assert not self.parallel_config.enable_elastic_ep, (
                            "Elastic EP is not supported with drafter model."
                        )
4783
4784
4785
4786
4787
4788
4789
4790
4791
4792
4793
4794
4795
4796
4797
4798
                        spec_config = self.vllm_config.speculative_config
                        assert spec_config is not None
                        assert spec_config.draft_model_config is not None
                        logger.info_once(
                            "EPLB is enabled for drafter model %s.",
                            spec_config.draft_model_config.model,
                        )
                        if self.eplb_state is None:
                            self.eplb_state = EplbState(
                                self.parallel_config, self.device
                            )
                        self.eplb_state.add_model(
                            self.drafter.model,
                            spec_config.draft_model_config,
                        )
                        eplb_models += 1
4799

4800
4801
4802
4803
4804
4805
                if self.use_aux_hidden_state_outputs:
                    if not supports_eagle3(self.get_model()):
                        raise RuntimeError(
                            "Model does not support EAGLE3 interface but "
                            "aux_hidden_state_outputs was requested"
                        )
4806

4807
4808
4809
4810
4811
4812
4813
4814
4815
                    # Try to get auxiliary layers from speculative config,
                    # otherwise use model's default layers
                    aux_layers = self._get_eagle3_aux_layers_from_config()
                    if aux_layers:
                        logger.info(
                            "Using auxiliary layers from speculative config: %s",
                            aux_layers,
                        )
                    else:
4816
4817
4818
                        aux_layers = (
                            self.model.get_eagle3_default_aux_hidden_state_layers()
                        )
4819
4820
4821
4822
4823
4824
4825
4826
4827
4828
4829
4830
4831
4832
4833

                    self.model.set_aux_hidden_state_layers(aux_layers)
                time_after_load = time.perf_counter()
            self.model_memory_usage = m.consumed_memory
        except torch.cuda.OutOfMemoryError as e:
            msg = (
                "Failed to load model - not enough GPU memory. "
                "Try lowering --gpu-memory-utilization to free memory for weights, "
                "increasing --tensor-parallel-size, or using --quantization. "
                "See https://docs.vllm.ai/en/latest/configuration/conserving_memory/ "
                "for more tips."
            )
            combined_msg = f"{msg} (original error: {e})"
            logger.error(combined_msg)
            raise e
4834
        logger.info_once(
4835
4836
            "Model loading took %s GiB memory and %.6f seconds",
            format_gib(self.model_memory_usage),
4837
            time_after_load - time_before_load,
4838
            scope="local",
4839
        )
4840
4841
4842
4843
4844
4845
        if not load_dummy_weights:
            prepare_communication_buffer_for_model(self.model)
            if (drafter := getattr(self, "drafter", None)) and (
                drafter_model := getattr(drafter, "model", None)
            ):
                prepare_communication_buffer_for_model(drafter_model)
4846
        mm_config = self.model_config.multimodal_config
4847
        self.is_multimodal_pruning_enabled = (
4848
            supports_multimodal_pruning(self.get_model())
4849
4850
            and mm_config is not None
            and mm_config.is_multimodal_pruning_enabled()
4851
        )
4852
4853
4854
        self.requires_sequential_video_encoding = hasattr(
            self.get_model(), "requires_sequential_video_encoding"
        )  # Temporary hack for dynamic res video w/o support for bs>1 yet
4855

4856
4857
4858
4859
4860
        if (
            is_mixture_of_experts(self.model)
            and self.parallel_config.enable_eplb
            and not load_dummy_weights
        ):
4861
4862
4863
            logger.info_once("EPLB is enabled for model %s.", self.model_config.model)
            assert self.eplb_state is not None
            self.eplb_state.add_model(
4864
                self.model,
4865
                self.model_config,
4866
            )
4867
            if self.eplb_state.is_async:
4868
                self.eplb_state.start_async_loop()
4869

4870
        if (
4871
4872
            self.vllm_config.compilation_config.mode
            == CompilationMode.STOCK_TORCH_COMPILE
4873
        ):
4874
            backend = self.vllm_config.compilation_config.init_backend(self.vllm_config)
4875
            compilation_counter.stock_torch_compile_count += 1
4876
            self.model.compile(fullgraph=True, backend=backend)
4877
            return
4878
        # for other compilation modes, cudagraph behavior is controlled by
Jiayi Yan's avatar
Jiayi Yan committed
4879
        # CudagraphWrapper and CudagraphDispatcher of vllm.
4880
4881

        # wrap the model with full cudagraph wrapper if needed.
4882
4883
        cudagraph_mode = self.compilation_config.cudagraph_mode
        assert cudagraph_mode is not None
4884
4885
4886
4887
        if (
            cudagraph_mode.has_full_cudagraphs()
            and not self.parallel_config.use_ubatching
        ):
4888
4889
4890
            self.model = CUDAGraphWrapper(
                self.model, self.vllm_config, runtime_mode=CUDAGraphMode.FULL
            )
4891
        elif self.parallel_config.use_ubatching:
4892
            if cudagraph_mode.has_full_cudagraphs():
4893
4894
4895
                self.model = UBatchWrapper(
                    self.model, self.vllm_config, CUDAGraphMode.FULL, self.device
                )
4896
            else:
4897
4898
4899
                self.model = UBatchWrapper(
                    self.model, self.vllm_config, CUDAGraphMode.NONE, self.device
                )
4900

4901
4902
        get_offloader().post_init()

4903
    def _get_eagle3_aux_layers_from_config(self) -> tuple[int, ...] | None:
4904
4905
4906
4907
4908
4909
4910
4911
4912
4913
4914
4915
4916
4917
4918
        """Extract Eagle3 auxiliary layer indices from speculative config.

        These indices specify which hidden states from the base model should
        be used as auxiliary inputs for the Eagle3 drafter model during
        speculative decoding.

        Returns:
            Tuple of layer indices if found in draft model config,
            None otherwise.
        """
        if not (self.speculative_config and self.speculative_config.draft_model_config):
            return None

        hf_config = self.speculative_config.draft_model_config.hf_config

4919
4920
4921
4922
4923
4924
        layer_ids = getattr(hf_config, "eagle_aux_hidden_state_layer_ids", None)
        if not layer_ids:
            dflash_config = getattr(hf_config, "dflash_config", None)
            if dflash_config and isinstance(dflash_config, dict):
                layer_ids = dflash_config.get("target_layer_ids")

4925
4926
4927
4928
4929
        if layer_ids and isinstance(layer_ids, (list, tuple)):
            return tuple(layer_ids)

        return None

4930
4931
4932
4933
4934
4935
4936
4937
4938
4939
4940
4941
4942
    def reload_weights(
        self,
        weights_iterator: Iterable[tuple[str, torch.Tensor]] | None = None,
        weights_path: str | None = None,
        is_checkpoint_format: bool = True,
    ) -> None:
        """
        Reload weights from a weights iterator or from disk

        :param weights_iterator: weights to load into model
        :param weights_path: path to load weights from if weights_iterator is not
            provided. Use path of original model if neither is provided.
        :param is_checkpoint_format: set to False if weights have already been processed
Jiayi Yan's avatar
Jiayi Yan committed
4943
            into kernel format (repacking, renaming, etc.)
4944
4945
4946
4947
4948
4949
4950
4951
4952
4953
4954
4955
4956
4957
4958
4959
4960
4961
4962
4963
4964
4965
4966
4967
4968
4969
4970
4971
4972
4973
4974
        """
        # TODO(@kylesayrs): generalize to all runners and loaders
        # argument validation
        if weights_iterator is None and not is_checkpoint_format:
            logger.warning(
                "Reloading from disk means that weights will be in checkpoint format. "
                "Please use `is_checkpoint_format=True` "
                "to avoid weight reloading errors"
            )

        model = self.get_model()
        weights_to_load = {name for name, _ in model.named_parameters()}
        counter_before_reloading = time.perf_counter()

        # load weights from disk if none are provided
        if weights_iterator is None:
            model_loader = get_model_loader(self.load_config)
            if not hasattr(model_loader, "get_all_weights"):
                raise NotImplementedError(
                    f"Model reloading with `{self.load_config.load_format}` format"
                )

            if weights_path is not None:
                self.model_config.model = weights_path
            weights_iterator = model_loader.get_all_weights(self.model_config, model)
            weights_iterator = cast(
                Iterable[tuple[str, torch.Tensor]], weights_iterator
            )

        # begin loading weights
        logger.info_once("Reloading weights inplace...", scope="local")
4975
4976
4977
4978
4979
        if is_checkpoint_format:
            # load weights from checkpoint/ original model format
            initialize_layerwise_reload(model)
            loaded_weights = model.load_weights(weights_iterator)
            finalize_layerwise_reload(model, self.model_config)
4980

4981
4982
4983
4984
4985
4986
4987
4988
4989
4990
4991
4992
        else:
            # load weights from kernel format
            logger.warning_once(
                "Reloading with `is_checkpoint_format=True` requires that "
                "weights be in kernel format and already sharded",
                scope="local",
            )
            loaded_weights = set()
            for name, loaded_weight in weights_iterator:
                param = model.get_parameter(name)  # TODO: buffers?
                param.copy_(loaded_weight)
                loaded_weights.add(name)
4993
4994
4995
4996
4997
4998
4999
5000

        # logging and validation
        counter_after_reloading = time.perf_counter()
        diff_seconds = counter_after_reloading - counter_before_reloading
        logger.info_once(
            "Reloading and processing weights took %.2f seconds",
            diff_seconds,
            scope="local",
5001
        )
5002
5003
5004
5005
5006
5007
5008
        if self.model_config.quantization is None and loaded_weights is not None:
            weights_not_loaded = weights_to_load - loaded_weights
            if weights_not_loaded:
                logger.warning(
                    "Following weights were not loaded from checkpoint: %s",
                    weights_not_loaded,
                )
5009

5010
5011
5012
    def _get_prompt_logprobs_dict(
        self,
        hidden_states: torch.Tensor,
5013
        num_scheduled_tokens: dict[str, int],
5014
    ) -> dict[str, LogprobsTensors | None]:
5015
        num_prompt_logprobs_dict = self.num_prompt_logprobs
5016
5017
5018
        if not num_prompt_logprobs_dict:
            return {}

5019
        in_progress_dict = self.input_batch.in_progress_prompt_logprobs_cpu
5020
        prompt_logprobs_dict: dict[str, LogprobsTensors | None] = {}
5021
5022
5023
5024
5025

        # Since prompt logprobs are a rare feature, prioritize simple,
        # maintainable loop over optimal performance.
        completed_prefill_reqs = []
        for req_id, num_prompt_logprobs in num_prompt_logprobs_dict.items():
5026
5027
5028
5029
            num_tokens = num_scheduled_tokens.get(req_id)
            if num_tokens is None:
                # This can happen if the request was preempted in prefill stage.
                continue
5030
5031
5032

            # Get metadata for this request.
            request = self.requests[req_id]
5033
5034
5035
5036
            if request.prompt_token_ids is None:
                # Prompt logprobs is incompatible with prompt embeddings
                continue

5037
5038
            num_prompt_tokens = len(request.prompt_token_ids)
            prompt_token_ids = torch.tensor(request.prompt_token_ids).to(
5039
5040
                self.device, non_blocking=True
            )
5041

5042
5043
5044
5045
5046
5047
            # Set up target LogprobsTensors object.
            logprobs_tensors = in_progress_dict.get(req_id)
            if not logprobs_tensors:
                # Create empty logprobs CPU tensors for the entire prompt.
                # If chunked, we'll copy in slice by slice.
                logprobs_tensors = LogprobsTensors.empty_cpu(
5048
5049
                    num_prompt_tokens - 1, num_prompt_logprobs + 1
                )
5050
5051
                in_progress_dict[req_id] = logprobs_tensors

5052
            # Determine number of logits to retrieve.
5053
5054
            start_idx = request.num_computed_tokens
            start_tok = start_idx + 1
5055
            num_remaining_tokens = num_prompt_tokens - start_tok
5056
            if num_tokens <= num_remaining_tokens:
5057
                # This is a chunk, more tokens remain.
5058
5059
5060
                # In the == case, there are no more prompt logprobs to produce
                # but we want to defer returning them to the next step where we
                # have new generated tokens to return.
5061
5062
5063
5064
5065
                num_logits = num_tokens
            else:
                # This is the last chunk of prompt tokens to return.
                num_logits = num_remaining_tokens
                completed_prefill_reqs.append(req_id)
5066
5067
5068
5069
5070
5071
5072
                prompt_logprobs_dict[req_id] = logprobs_tensors

            if num_logits <= 0:
                # This can happen for the final chunk if we prefilled exactly
                # (num_prompt_tokens - 1) tokens for this request in the prior
                # step. There are no more prompt logprobs to produce.
                continue
5073
5074
5075
5076
5077

            # Get the logits corresponding to this req's prompt tokens.
            # If this is a partial request (i.e. chunked prefill),
            # then there is prompt logprob generated for each index.
            req_idx = self.input_batch.req_id_to_index[req_id]
5078
            offset = self.query_start_loc.np[req_idx].item()
5079
            prompt_hidden_states = hidden_states[offset : offset + num_logits]
5080
            logits = self.model.compute_logits(prompt_hidden_states)
5081
5082
5083
5084

            # Get the "target" tokens for each index. For prompt at index i,
            # the token at prompt index i+1 is the "sampled" token we want
            # to gather the logprob for.
5085
            tgt_token_ids = prompt_token_ids[start_tok : start_tok + num_logits]
5086
5087

            # Compute prompt logprobs.
5088
            logprobs = self.sampler.compute_logprobs(logits)
5089
            token_ids, logprobs, ranks, _ = self.sampler.gather_logprobs(
5090
5091
                logprobs, num_prompt_logprobs, tgt_token_ids
            )
5092
5093

            # Transfer GPU->CPU async.
5094
5095
            chunk_slice = slice(start_idx, start_idx + num_logits)
            logprobs_tensors.logprob_token_ids[chunk_slice].copy_(
5096
5097
5098
                token_ids, non_blocking=True
            )
            logprobs_tensors.logprobs[chunk_slice].copy_(logprobs, non_blocking=True)
5099
            logprobs_tensors.selected_token_ranks[chunk_slice].copy_(
5100
5101
                ranks, non_blocking=True
            )
5102
5103
5104
5105
5106

        # Remove requests that have completed prefill from the batch
        # num_prompt_logprobs_dict.
        for req_id in completed_prefill_reqs:
            del num_prompt_logprobs_dict[req_id]
5107
            del in_progress_dict[req_id]
5108
5109

        # Must synchronize the non-blocking GPU->CPU transfers.
5110
        if prompt_logprobs_dict:
5111
            self._sync_device()
5112
5113
5114

        return prompt_logprobs_dict

5115
5116
    def _get_nans_in_logits(
        self,
5117
        logits: torch.Tensor | None,
5118
5119
5120
5121
5122
5123
5124
5125
5126
5127
5128
    ) -> dict[str, int]:
        try:
            if logits is None:
                return {req_id: 0 for req_id in self.input_batch.req_ids}

            num_nans_in_logits = {}
            num_nans_for_index = logits.isnan().sum(dim=-1).cpu().numpy()
            for req_id in self.input_batch.req_ids:
                req_index = self.input_batch.req_id_to_index[req_id]
                num_nans_in_logits[req_id] = (
                    int(num_nans_for_index[req_index])
5129
5130
5131
                    if num_nans_for_index is not None and req_index < logits.shape[0]
                    else 0
                )
5132
5133
5134
5135
            return num_nans_in_logits
        except IndexError:
            return {}

5136
    @contextmanager
5137
5138
5139
    def maybe_randomize_inputs(
        self, input_ids: torch.Tensor | None, inputs_embeds: torch.Tensor | None
    ):
5140
5141
5142
5143
        """
        Randomize input_ids if VLLM_RANDOMIZE_DP_DUMMY_INPUTS is set.
        This is to help balance expert-selection
         - during profile_run
5144
         - during DP rank dummy run
5145
        """
5146

5147
5148
5149
5150
        dp_size = self.vllm_config.parallel_config.data_parallel_size
        randomize_inputs = envs.VLLM_RANDOMIZE_DP_DUMMY_INPUTS and dp_size > 1
        if not randomize_inputs:
            yield
5151
        elif input_ids is not None:
5152
5153
5154
5155

            @functools.cache
            def rand_input_ids() -> torch.Tensor:
                return torch.randint_like(
5156
                    self.input_ids.gpu,
5157
5158
                    low=0,
                    high=self.model_config.get_vocab_size(),
5159
                )
5160

5161
            logger.debug_once("Randomizing dummy input_ids for DP Rank")
5162
            input_ids.copy_(rand_input_ids()[: input_ids.size(0)], non_blocking=True)
5163
5164
            yield
            input_ids.fill_(0)
5165
5166
5167
5168
5169
5170
5171
5172
5173
5174
5175
5176
5177
5178
5179
        else:

            @functools.cache
            def rand_inputs_embeds() -> torch.Tensor:
                return torch.randn_like(
                    self.inputs_embeds.gpu,
                )

            assert inputs_embeds is not None
            logger.debug_once("Randomizing dummy inputs_embeds for DP Rank")
            inputs_embeds.copy_(
                rand_inputs_embeds()[: inputs_embeds.size(0)], non_blocking=True
            )
            yield
            inputs_embeds.fill_(0)
5180

5181
5182
5183
5184
5185
5186
    def _get_mm_dummy_batch(
        self,
        modality: str,
        max_items_per_batch: int,
    ) -> BatchedTensorInputs:
        """Dummy data for profiling and precompiling multimodal models."""
5187
5188
        assert self.mm_budget is not None

5189
5190
5191
        # Don't use `max_items_per_batch` here to avoid redundant computation
        dummy_mm_inputs = self.mm_registry.get_dummy_mm_inputs(
            self.model_config,
5192
            mm_counts={modality: 1},
5193
            cache=self.mm_budget.cache,
5194
        )
5195
5196
5197
5198
5199
        dummy_mm_item = dummy_mm_inputs["mm_kwargs"][modality][0]

        # We use the cache so that the item is saved to the cache,
        # but not read from the cache
        assert dummy_mm_item is not None, "Item should not already be cached"
5200

5201
        return next(
5202
5203
            mm_kwargs_batch
            for _, _, mm_kwargs_batch in group_and_batch_mm_kwargs(
5204
                [(modality, dummy_mm_item)] * max_items_per_batch,
5205
5206
5207
5208
                device=self.device,
                pin_memory=self.pin_memory,
            )
        )
5209

5210
5211
5212
5213
    @torch.inference_mode()
    def _dummy_run(
        self,
        num_tokens: int,
5214
        cudagraph_runtime_mode: CUDAGraphMode | None = None,
5215
5216
        force_attention: bool = False,
        uniform_decode: bool = False,
5217
        allow_microbatching: bool = True,
5218
5219
        skip_eplb: bool = False,
        is_profile: bool = False,
5220
        create_mixed_batch: bool = False,
5221
        remove_lora: bool = True,
Rémi Delacourt's avatar
Rémi Delacourt committed
5222
        is_graph_capturing: bool = False,
5223
        num_active_loras: int = 0,
5224
        profile_seq_lens: int | None = None,
5225
    ) -> tuple[torch.Tensor, torch.Tensor]:
5226
5227
5228
5229
5230
5231
5232
        """
        Run a dummy forward pass to warm up/profile run or capture the
        CUDA graph for the model.

        Args:
            num_tokens: Number of tokens to run the dummy forward pass.
            cudagraph_runtime_mode: used to control the behavior.
5233
                - if not set will determine the cudagraph mode based on using
5234
                    the self.cudagraph_dispatcher.
5235
5236
5237
5238
                - CUDAGraphMode.NONE: No cudagraph, for warm up and profile run
                - CUDAGraphMode.PIECEWISE: Piecewise cudagraph.
                - CUDAGraphMode.FULL: Full cudagraph, attention metadata is
                    needed.
5239
            force_attention: If True, always create attention metadata. Used to
5240
5241
5242
5243
                warm up attention backend when mode is NONE.
            uniform_decode: If True, the batch is a uniform decode batch.
            skip_eplb: If True, skip EPLB state update.
            is_profile: If True, this is a profile run.
5244
5245
            create_mixed_batch: If True, create a mixed batch with both decode
                (1 token) and prefill (multiple tokens) requests.
5246
            remove_lora: If False, dummy LoRAs are not destroyed after the run
5247
5248
            num_active_loras: Number of distinct active LoRAs to capture for.
                LoRA is activated when num_active_loras > 0.
5249
5250
5251
            profile_seq_lens: If provided, use this value for seq_lens instead
                of max_query_len. Used to profile attention workspace that
                scales with context length.
5252
        """
5253
5254
        mm_config = self.vllm_config.model_config.multimodal_config
        if mm_config and mm_config.mm_encoder_only:
5255
5256
5257
5258
            # The current dummy run only covers LM execution, so we can skip it.
            # mm encoder dummy run may need to add in the future.
            return torch.tensor([]), torch.tensor([])

5259
5260
        assert (
            cudagraph_runtime_mode is None
5261
            or cudagraph_runtime_mode.is_valid_runtime_mode()
5262
        )
5263

5264
        # If cudagraph_mode.decode_mode() == FULL and
5265
        # cudagraph_mode.separate_routine(). This means that we are using
5266
5267
5268
5269
5270
5271
5272
5273
5274
5275
5276
        # different graphs and/or modes for mixed prefill-decode batches vs.
        # uniform decode batches. A uniform decode batch means that all
        # requests have identical query length, except a potential virtual
        # request (shorter) in the batch account for padding.
        # Uniform decode batch could either be common pure decode, where
        # max_query_len == 1, or speculative decode, where
        # max_query_len == 1 + num_spec_decode_tokens.

        # When setting max_query_len = 1, we switch to and capture the optimized
        # routine of FA2 for pure decode, i.e., Flashdecode + an optimization
        # for GQA/MQA.
5277
        max_query_len = self.uniform_decode_query_len if uniform_decode else num_tokens
5278

5279
5280
5281
        # Set num_scheduled_tokens based on num_tokens and max_num_seqs
        # for dummy run with LoRA so that the num_reqs collectively
        # has num_tokens in total.
5282
        assert num_tokens <= self.max_num_tokens
5283
        max_num_reqs = self.scheduler_config.max_num_seqs
5284
5285
5286
5287
        if create_mixed_batch:
            assert not uniform_decode
            # Create mixed batch:
            # first half decode tokens, second half one prefill
5288
            num_decode_tokens = min(max_num_reqs - 1, num_tokens // 2)
5289
5290
5291
5292
            num_prefill_tokens = num_tokens - num_decode_tokens
            num_reqs = num_decode_tokens + 1

            # Create decode requests (1 token each) followed by prefill request
5293
            num_scheduled_tokens_list = [1] * num_decode_tokens + [num_prefill_tokens]
5294
5295
5296
            # Note: Overriding max_query_len to be the prefill tokens
            max_query_len = num_prefill_tokens
        elif uniform_decode:
5297
            assert not create_mixed_batch
5298
            num_reqs = min(max_num_reqs, cdiv(num_tokens, max_query_len))
5299
5300
            num_scheduled_tokens_list = [max_query_len] * num_reqs
            if num_tokens % max_query_len != 0:
5301
                num_scheduled_tokens_list[-1] = num_tokens % max_query_len
5302
5303
5304
5305
5306
5307
        else:
            num_reqs = min(num_tokens, max_num_reqs)
            min_tokens_per_req = num_tokens // num_reqs
            num_scheduled_tokens_list = [min_tokens_per_req] * num_reqs
            num_scheduled_tokens_list[-1] += num_tokens % num_reqs

5308
5309
        assert sum(num_scheduled_tokens_list) == num_tokens
        assert len(num_scheduled_tokens_list) == num_reqs
5310
        num_scheduled_tokens = np.array(num_scheduled_tokens_list, dtype=np.int32)
5311
5312
        num_tokens_unpadded = int(num_scheduled_tokens.sum())

5313
        num_sampled_tokens = np.ones(num_reqs, dtype=np.int32)
5314

5315
        _cudagraph_mode, batch_desc, should_ubatch, num_tokens_across_dp, _ = (
5316
5317
5318
5319
5320
5321
5322
5323
5324
5325
5326
5327
5328
5329
5330
5331
5332
            self._determine_batch_execution_and_padding(
                num_tokens=num_tokens_unpadded,
                num_reqs=num_reqs,
                num_scheduled_tokens_np=num_scheduled_tokens,
                max_num_scheduled_tokens=max_query_len,
                use_cascade_attn=False,
                allow_microbatching=allow_microbatching,
                force_eager=is_profile
                or (cudagraph_runtime_mode == CUDAGraphMode.NONE),
                # `force_uniform_decode` is used for cudagraph capture; because for
                # capturing mixed prefill-decode batches, we sometimes use
                # num_tokens == num_reqs which looks like a uniform decode batch to the
                # dispatcher; but we actually want to capture a piecewise cudagraph
                force_uniform_decode=uniform_decode,
                # `force_has_lora` is used for cudagraph capture; because LoRA is
                # activated later in the context manager, but we need to know the
                # LoRA state when determining the batch descriptor for capture
5333
5334
5335
5336
                force_has_lora=num_active_loras > 0,
                # `force_num_active_loras` is used for cudagraph capture; because we
                # need to capture graphs for specific num_active_loras counts
                force_num_active_loras=num_active_loras,
5337
5338
            )
        )
5339
5340
5341

        if cudagraph_runtime_mode is None:
            cudagraph_runtime_mode = _cudagraph_mode
5342
        else:
5343
5344
5345
5346
5347
5348
5349
5350
5351
            assert cudagraph_runtime_mode == _cudagraph_mode, (
                f"Cudagraph runtime mode mismatch in dummy_run. "
                f"Expected {_cudagraph_mode}, but got {cudagraph_runtime_mode}."
            )

        num_tokens_padded = batch_desc.num_tokens
        num_reqs_padded = (
            batch_desc.num_reqs if batch_desc.num_reqs is not None else num_reqs
        )
5352
        ubatch_slices, ubatch_slices_padded = maybe_create_ubatch_slices(
5353
5354
5355
5356
5357
5358
5359
5360
5361
5362
            should_ubatch,
            num_scheduled_tokens,
            num_tokens_padded,
            num_reqs_padded,
            self.vllm_config.parallel_config.num_ubatches,
        )
        logger.debug(
            "ubatch_slices: %s, ubatch_slices_padded: %s",
            ubatch_slices,
            ubatch_slices_padded,
5363
        )
5364

5365
        attn_metadata: PerLayerAttnMetadata | None = None
5366

5367
5368
5369
5370
5371
5372
5373
        slot_mappings_by_group, slot_mappings = self._get_slot_mappings(
            num_tokens_padded=num_tokens,
            num_reqs_padded=num_reqs_padded,
            num_tokens_unpadded=num_tokens_unpadded,
            ubatch_slices=ubatch_slices_padded,
        )

5374
5375
5376
5377
5378
5379
5380
5381
        # _dummy_run shares pinned CPU buffers (seq_lens, query_start_loc,
        # etc.) with execute_model.  It must participate in the same event
        # protocol so that back-to-back dummy/real steps don't overwrite
        # pinned memory while a prior non_blocking H2D DMA is still reading.
        with self.synchronize_input_prep():
            # If force_attention is True, we always capture attention.
            # Otherwise, it only happens for cudagraph_runtime_mode=FULL.
            if force_attention or cudagraph_runtime_mode == CUDAGraphMode.FULL:
5382
5383
5384
                if profile_seq_lens is not None:
                    seq_lens = profile_seq_lens  # type: ignore[assignment]
                elif create_mixed_batch:
5385
5386
5387
                    # In the mixed batch mode (used for FI warmup), we use
                    # shorter sequence lengths to run faster.
                    # TODO(luka) better system for describing dummy batches
5388
5389
5390
5391
                    seq_lens = torch.tensor(  # type: ignore[assignment]
                        [1] * num_decode_tokens + [num_prefill_tokens + 1],
                        dtype=torch.int,
                    )
5392
5393
                else:
                    seq_lens = max_query_len  # type: ignore[assignment]
5394
5395
5396
                self.optimistic_seq_lens_cpu[:num_reqs] = seq_lens
                self.optimistic_seq_lens_cpu[num_reqs:].fill_(0)
                self.seq_lens.copy_(self.optimistic_seq_lens_cpu, non_blocking=True)
5397

5398
5399
5400
                cum_num_tokens = self._get_cumsum_and_arange(
                    num_scheduled_tokens, self.query_pos.np
                )
5401
5402
                self.query_start_loc.np[1 : num_reqs + 1] = cum_num_tokens
                self.query_start_loc.copy_to_gpu()
5403

5404
5405
5406
5407
5408
5409
                # Sync block table CPU->GPU so cleared rows from
                # remove_request() are visible to the attention metadata
                # builder. Without this, stale block IDs from finished
                # requests can corrupt Mamba state.
                self.input_batch.block_table.commit_block_table(num_reqs_padded)

5410
5411
5412
5413
5414
5415
5416
5417
5418
                pad_attn = cudagraph_runtime_mode == CUDAGraphMode.FULL
                attn_metadata, _ = self._build_attention_metadata(
                    num_tokens=num_tokens_unpadded,
                    num_tokens_padded=num_tokens_padded if pad_attn else None,
                    num_reqs=num_reqs_padded,
                    max_query_len=max_query_len,
                    ubatch_slices=(ubatch_slices_padded if pad_attn else ubatch_slices),
                    for_cudagraph_capture=is_graph_capturing,
                    slot_mappings=slot_mappings_by_group,
5419
                    use_spec_decode=self.speculative_config is not None,
5420
                )
5421

5422
        with self.maybe_dummy_run_with_lora(
5423
5424
5425
5426
            self.lora_config,
            num_scheduled_tokens,
            num_sampled_tokens,
            remove_lora,
5427
            num_active_loras,
5428
        ):
5429
            # Make sure padding doesn't exceed max_num_tokens
5430
            assert num_tokens_padded <= self.max_num_tokens
5431
            model_kwargs = self._init_model_kwargs()
5432
            if self.supports_mm_inputs and not self.model_config.is_encoder_decoder:
Patrick von Platen's avatar
Patrick von Platen committed
5433
5434
                input_ids, inputs_embeds = self._prepare_mm_inputs(num_tokens_padded)

5435
                model_kwargs = {
5436
                    **model_kwargs,
5437
5438
                    **self._dummy_mm_kwargs(num_reqs),
                }
5439
5440
            elif self.enable_prompt_embeds:
                input_ids = None
5441
                inputs_embeds = self.inputs_embeds.gpu[:num_tokens_padded]
5442
                model_kwargs = self._init_model_kwargs()
5443
            else:
5444
                input_ids = self.input_ids.gpu[:num_tokens_padded]
5445
                inputs_embeds = None
5446

5447
            if self.uses_mrope:
5448
                positions = self.mrope_positions.gpu[:, :num_tokens_padded]
5449
            elif self.uses_xdrope_dim > 0:
5450
                positions = self.xdrope_positions.gpu[:, :num_tokens_padded]
5451
            else:
5452
                positions = self.positions[:num_tokens_padded]
5453
5454
5455
5456
5457
5458
5459
5460
5461

            if get_pp_group().is_first_rank:
                intermediate_tensors = None
            else:
                if self.intermediate_tensors is None:
                    self.intermediate_tensors = (
                        self.model.make_empty_intermediate_tensors(
                            batch_size=self.max_num_tokens,
                            dtype=self.model_config.dtype,
5462
5463
5464
                            device=self.device,
                        )
                    )
5465
5466

                intermediate_tensors = self.sync_and_slice_intermediate_tensors(
5467
                    num_tokens_padded, None, False
5468
                )
5469

5470
            if ubatch_slices_padded is not None:
5471
5472
5473
                # Adjust values to reflect a single ubatch.
                # TODO(sage,lucas): this is cruft that should be addressed in
                #  the padding refactor.
5474
                num_tokens_padded = ubatch_slices_padded[0].num_tokens
5475
                if num_tokens_across_dp is not None:
5476
                    num_tokens_across_dp[:] = num_tokens_padded
5477

5478
            with (
5479
                self.maybe_randomize_inputs(input_ids, inputs_embeds),
5480
                set_forward_context(
5481
5482
                    attn_metadata,
                    self.vllm_config,
5483
                    num_tokens=num_tokens_padded,
5484
5485
                    num_tokens_across_dp=num_tokens_across_dp,
                    cudagraph_runtime_mode=cudagraph_runtime_mode,
5486
                    batch_descriptor=batch_desc,
5487
                    ubatch_slices=ubatch_slices_padded,
5488
                    slot_mapping=slot_mappings,
5489
5490
                ),
            ):
5491
                outputs = self.model(
5492
5493
5494
5495
                    input_ids=input_ids,
                    positions=positions,
                    intermediate_tensors=intermediate_tensors,
                    inputs_embeds=inputs_embeds,
5496
                    **model_kwargs,
5497
                )
5498

5499
5500
5501
5502
            if self.use_aux_hidden_state_outputs:
                hidden_states, _ = outputs
            else:
                hidden_states = outputs
5503

5504
5505
5506
            if self.speculative_config and (
                self.speculative_config.use_eagle()
                or self.speculative_config.uses_draft_model()
5507
                or self.speculative_config.uses_extract_hidden_states()
5508
            ):
5509
5510
                assert isinstance(
                    self.drafter,
5511
5512
5513
5514
                    EagleProposer
                    | DFlashProposer
                    | DraftModelProposer
                    | ExtractHiddenStatesProposer,
5515
                )
5516
                assert self.speculative_config is not None
5517
5518
5519
                # Eagle currently only supports PIECEWISE cudagraphs.
                # Therefore only use cudagraphs if the main model uses PIECEWISE
                # NOTE(lucas): this is a hack, need to clean up.
5520
                use_cudagraphs = (
5521
5522
5523
5524
5525
5526
5527
5528
5529
                    (
                        is_graph_capturing
                        and cudagraph_runtime_mode == CUDAGraphMode.PIECEWISE
                    )
                    or (
                        not is_graph_capturing
                        and cudagraph_runtime_mode != CUDAGraphMode.NONE
                    )
                ) and not self.speculative_config.enforce_eager
5530
5531
5532
5533
5534

                # Note(gnovack) - We need to disable cudagraphs for one of the two
                # lora cases when cudagraph_specialize_lora is enabled. This is a
                # short term mitigation for issue mentioned in
                # https://github.com/vllm-project/vllm/issues/28334
5535
5536
5537
5538
                if (
                    self.compilation_config.cudagraph_specialize_lora
                    and num_active_loras > 0
                ):
5539
5540
5541
5542
5543
                    use_cudagraphs = False

                self.drafter.dummy_run(
                    num_tokens,
                    use_cudagraphs=use_cudagraphs,
Rémi Delacourt's avatar
Rémi Delacourt committed
5544
                    is_graph_capturing=is_graph_capturing,
5545
                    slot_mappings=slot_mappings,
5546
                )
5547

5548
5549
5550
5551
5552
5553
5554
5555
5556
5557
5558
        # We register layerwise NVTX hooks here after the first dynamo tracing is
        # done to avoid nvtx operations in hook functions being traced by
        # torch dynamo and causing graph breaks.
        # Note that for DYNAMO_ONCE and VLLM_COMPILE mode,
        # compiled model's dynamo tracing is only done once and the compiled model's
        # __call__ function is replaced by calling the compiled function.
        # So it's safe to register hooks here. Hooks will be registered to
        # both compiled and uncompiled models but they will never
        # be called on the compiled model execution path.
        self._register_layerwise_nvtx_hooks()

5559
5560
5561
5562
5563
5564
5565
5566
5567
5568
        # This is necessary to avoid blocking DP.
        # For dummy runs, we typically skip EPLB since we don't have any real
        # requests to process.
        # However, in DP settings, there may be cases when some DP ranks do
        # not have any requests to process, so they're executing dummy batches.
        # In such cases, we still have to trigger EPLB to make sure
        # ranks execute the rearrangement in synchronization.
        if not skip_eplb:
            self.eplb_step(is_dummy=True, is_profile=is_profile)

5569
        logit_indices = np.cumsum(num_scheduled_tokens) - 1
5570
5571
5572
5573
        logit_indices_device = torch.from_numpy(logit_indices).to(
            self.device, non_blocking=True
        )
        return hidden_states, hidden_states[logit_indices_device]
5574
5575
5576
5577
5578
5579

    @torch.inference_mode()
    def _dummy_sampler_run(
        self,
        hidden_states: torch.Tensor,
    ) -> torch.Tensor:
5580
5581
5582
        # The dummy hidden states may contain special values,
        # like `inf` or `nan`.
        # To avoid breaking the sampler, we use a random tensor here instead.
5583

5584
5585
        mm_config = self.vllm_config.model_config.multimodal_config
        if mm_config and mm_config.mm_encoder_only:
5586
5587
5588
            # MM Encoder only model no need to run sampler.
            return torch.tensor([])

5589
        hidden_states = torch.rand_like(hidden_states)
5590

5591
        logits = self.model.compute_logits(hidden_states)
5592
5593
        num_reqs = logits.size(0)

5594
        dummy_tensors = lambda v: torch.full((num_reqs,), v, device=self.device)
5595
5596
5597
5598
5599
5600
5601
5602
5603

        dummy_metadata = SamplingMetadata(
            temperature=dummy_tensors(0.5),
            all_greedy=False,
            all_random=False,
            top_p=dummy_tensors(0.9),
            top_k=dummy_tensors(logits.size(1) - 1),
            generators={},
            max_num_logprobs=None,
Vedant V Jhaveri's avatar
Vedant V Jhaveri committed
5604
            logprob_token_ids=None,
5605
5606
5607
5608
5609
5610
            no_penalties=True,
            prompt_token_ids=None,
            frequency_penalties=dummy_tensors(0.1),
            presence_penalties=dummy_tensors(0.1),
            repetition_penalties=dummy_tensors(0.1),
            output_token_ids=[[] for _ in range(num_reqs)],
5611
            spec_token_ids=[[] for _ in range(num_reqs)],
5612
5613
            allowed_token_ids_mask=None,
            bad_words_token_ids={},
5614
            logitsprocs=LogitsProcessors(),
5615
        )
5616
        try:
5617
5618
5619
            sampler_output = self.sampler(
                logits=logits, sampling_metadata=dummy_metadata
            )
5620
        except RuntimeError as e:
5621
            if "out of memory" in str(e):
5622
5623
5624
5625
                raise RuntimeError(
                    "CUDA out of memory occurred when warming up sampler with "
                    f"{num_reqs} dummy requests. Please try lowering "
                    "`max_num_seqs` or `gpu_memory_utilization` when "
5626
5627
                    "initializing the engine."
                ) from e
5628
5629
            else:
                raise e
5630
        if self.speculative_config:
5631
5632
            draft_token_ids = [[0] for _ in range(num_reqs)]
            dummy_spec_decode_metadata = SpecDecodeMetadata.make_dummy(
5633
5634
                draft_token_ids, self.device
            )
5635
5636
5637
5638
5639
5640

            num_tokens = sum(len(ids) for ids in draft_token_ids)
            # draft_probs = torch.randn(
            #     num_tokens, logits.shape[-1], device=self.device,
            #     dtype=logits.dtype)
            draft_probs = None
5641
5642
5643
5644
5645
            logits = torch.randn(
                num_tokens + num_reqs,
                logits.shape[-1],
                device=self.device,
                dtype=logits.dtype,
5646
            )
5647
5648
5649
            self.rejection_sampler(
                dummy_spec_decode_metadata,
                draft_probs,
5650
                logits,
5651
5652
                dummy_metadata,
            )
5653
        return sampler_output
5654

5655
    def _dummy_pooler_run_task(
5656
5657
        self,
        hidden_states: torch.Tensor,
5658
5659
        task: PoolingTask,
    ) -> PoolerOutput:
5660
5661
5662
5663
        num_tokens = hidden_states.shape[0]
        max_num_reqs = self.scheduler_config.max_num_seqs
        num_reqs = min(num_tokens, max_num_reqs)
        min_tokens_per_req = num_tokens // num_reqs
5664
5665
5666
5667
        num_scheduled_tokens_np = np.full(num_reqs, min_tokens_per_req)
        num_scheduled_tokens_np[-1] += num_tokens % num_reqs
        assert np.sum(num_scheduled_tokens_np) == num_tokens
        assert len(num_scheduled_tokens_np) == num_reqs
5668
5669
5670

        req_num_tokens = num_tokens // num_reqs

5671
        dummy_prompt_lens = torch.from_numpy(num_scheduled_tokens_np)
5672
5673
5674
        dummy_token_ids = torch.zeros(
            (num_reqs, req_num_tokens), dtype=torch.int32, device=self.device
        )
5675

5676
        model = cast(VllmModelForPooling, self.get_model())
5677
        dummy_pooling_params = PoolingParams(task=task)
5678
        dummy_pooling_params.verify(self.model_config)
5679
        to_update = model.pooler.get_pooling_updates(task)
5680
5681
        to_update.apply(dummy_pooling_params)

5682
        dummy_metadata = PoolingMetadata(
5683
5684
            prompt_lens=dummy_prompt_lens,
            prompt_token_ids=dummy_token_ids,
5685
            prompt_token_ids_cpu=dummy_token_ids.cpu(),
5686
            pooling_params=[dummy_pooling_params] * num_reqs,
5687
            pooling_states=[PoolingStates() for i in range(num_reqs)],
5688
        )
5689

5690
        dummy_metadata.build_pooling_cursor(
5691
            num_scheduled_tokens_np,
5692
5693
            seq_lens_cpu=dummy_prompt_lens,
            device=hidden_states.device,
5694
        )
5695

5696
        try:
5697
5698
5699
            return model.pooler(
                hidden_states=hidden_states, pooling_metadata=dummy_metadata
            )
5700
        except RuntimeError as e:
5701
            if "out of memory" in str(e):
5702
                raise RuntimeError(
5703
5704
5705
                    "CUDA out of memory occurred when warming up pooler "
                    f"({task=}) with {num_reqs} dummy requests. Please try "
                    "lowering `max_num_seqs` or `gpu_memory_utilization` when "
5706
5707
                    "initializing the engine."
                ) from e
5708
5709
            else:
                raise e
5710
5711
5712
5713
5714
5715

    @torch.inference_mode()
    def _dummy_pooler_run(
        self,
        hidden_states: torch.Tensor,
    ) -> PoolerOutput:
5716
5717
        mm_config = self.vllm_config.model_config.multimodal_config
        if mm_config and mm_config.mm_encoder_only:
5718
5719
5720
            # MM Encoder only model not need to run pooler.
            return torch.tensor([])

5721
        # Find the task that has the largest output for subsequent steps
5722
5723
5724
        supported_pooling_tasks = self.get_supported_pooling_tasks()

        if not supported_pooling_tasks:
5725
5726
5727
5728
5729
5730
            raise RuntimeError(
                f"Model {self.model_config.model} does not support "
                "any pooling tasks. See "
                "https://docs.vllm.ai/en/latest/models/pooling_models.html "
                "to learn more."
            )
5731

5732
        output_size = dict[PoolingTask, float]()
5733
        for task in supported_pooling_tasks:
5734
5735
            # Run a full batch with each task to ensure none of them OOMs
            output = self._dummy_pooler_run_task(hidden_states, task)
5736
            output_size[task] = sum(o.nbytes for o in output if o is not None)
5737
5738
5739
5740
            del output  # Allow GC

        max_task = max(output_size.items(), key=lambda x: x[1])[0]
        return self._dummy_pooler_run_task(hidden_states, max_task)
5741

5742
    def profile_run(self) -> None:
5743
        # Profile with multimodal encoder & encoder cache.
5744
        if self.supports_mm_inputs:
5745
5746
            mm_config = self.model_config.multimodal_config
            if mm_config is not None and mm_config.skip_mm_profiling:
5747
                logger.info(
5748
                    "Skipping memory profiling for multimodal encoder and "
5749
5750
                    "encoder cache."
                )
5751
5752
5753
5754
5755
            else:
                mm_budget = self.mm_budget
                assert mm_budget is not None

                if (encoder_budget := mm_budget.get_encoder_budget()) > 0:
5756
5757
5758
5759
5760
5761
5762
5763
5764
5765
5766
5767
5768
5769
5770
5771
5772
                    if not mm_budget.mm_max_toks_per_item:
                        # All modality limits are 0 — embedding-only mode.
                        # Budget is non-zero for embedding storage, but
                        # there's no encoder to profile.
                        logger.info(
                            "Skipping encoder profiling for embedding-only "
                            "mode (all modality limits=0 with "
                            "enable_mm_embeds=True).",
                        )
                    else:
                        # NOTE: Currently model is profiled with a single
                        # non-text modality with the max possible input
                        # tokens even when it supports multiple.
                        dummy_modality = mm_budget.get_modality_with_max_tokens()
                        max_mm_items_per_batch = mm_budget.mm_max_items_per_batch[
                            dummy_modality
                        ]
5773

5774
                        logger.info_once(
5775
5776
5777
5778
5779
5780
                            "Encoder cache will be initialized with a "
                            "budget of %s tokens, and profiled with "
                            "%s %s items of the maximum feature size.",
                            encoder_budget,
                            max_mm_items_per_batch,
                            dummy_modality,
5781
                            scope="local",
5782
                        )
5783

5784
5785
5786
5787
5788
                        # Create dummy batch of multimodal inputs.
                        batched_dummy_mm_inputs = self._get_mm_dummy_batch(
                            dummy_modality,
                            max_mm_items_per_batch,
                        )
5789

5790
5791
5792
5793
5794
5795
5796
5797
5798
5799
5800
                        # Run multimodal encoder.
                        dummy_encoder_outputs = self.model.embed_multimodal(
                            **batched_dummy_mm_inputs
                        )

                        sanity_check_mm_encoder_outputs(
                            dummy_encoder_outputs,
                            expected_num_items=max_mm_items_per_batch,
                        )
                        for i, output in enumerate(dummy_encoder_outputs):
                            self.encoder_cache[f"tmp_{i}"] = output
5801

5802
        # Add `is_profile` here to pre-allocate communication buffers
5803
5804
5805
        hidden_states, last_hidden_states = self._dummy_run(
            self.max_num_tokens, is_profile=True
        )
5806
        if get_pp_group().is_last_rank:
5807
5808
5809
5810
            if self.is_pooling_model:
                output = self._dummy_pooler_run(hidden_states)
            else:
                output = self._dummy_sampler_run(last_hidden_states)
5811
        else:
5812
            output = None
5813
        self._sync_device()
5814
        del hidden_states, output
5815
        self.encoder_cache.clear()
5816
        gc.collect()
5817

5818
5819
5820
5821
5822
5823
5824
5825
5826
5827
    def _init_minimal_kv_cache_for_profiling(self) -> None:
        from vllm.v1.core.kv_cache_utils import (
            get_kv_cache_config_from_groups,
            get_kv_cache_groups,
        )

        kv_cache_spec = self.get_kv_cache_spec()
        kv_cache_groups = get_kv_cache_groups(self.vllm_config, kv_cache_spec)
        min_blocks = self.compilation_config.max_cudagraph_capture_size or 1

5828
5829
5830
        # Temporarily change num_gpu_blocks_override to allocate a minimal KV cache
        saved_override = self.cache_config.num_gpu_blocks_override
        self.cache_config.num_gpu_blocks_override = min_blocks
5831
        minimal_config = get_kv_cache_config_from_groups(
5832
            self.vllm_config, kv_cache_groups, available_memory=0
5833
        )
5834
        self.cache_config.num_gpu_blocks_override = saved_override
5835

5836
        self.initialize_kv_cache(minimal_config, is_profiling=True)
5837
5838
5839
5840
5841
5842
5843
5844
5845
5846
5847
5848
5849
5850
5851
5852
5853
5854
5855
5856
5857
5858
5859
5860
5861
5862
5863
5864
5865
5866
5867
5868
5869
5870
5871
        self.cache_config.num_gpu_blocks = minimal_config.num_blocks

        logger.debug("Initialized minimal KV cache for CUDA graph profiling")

    @staticmethod
    @contextmanager
    def _freeze_gc():
        gc.collect()
        should_freeze = not envs.VLLM_ENABLE_CUDAGRAPH_GC
        if should_freeze:
            gc.freeze()
        try:
            yield
        finally:
            if should_freeze:
                gc.unfreeze()
                gc.collect()

    def _cleanup_profiling_kv_cache(self) -> None:
        torch.accelerator.synchronize()
        if hasattr(self, "kv_caches") and self.kv_caches:
            for i in range(len(self.kv_caches)):
                self.kv_caches[i] = None  # type: ignore
            self.kv_caches.clear()
        if hasattr(self, "cross_layers_kv_cache"):
            self.cross_layers_kv_cache = None
            self.cross_layers_attn_backend = None
        if hasattr(self, "attn_groups"):
            self.attn_groups.clear()
        if hasattr(self, "kv_cache_config"):
            delattr(self, "kv_cache_config")
        self.cache_config.num_gpu_blocks = None

        for layer in self.compilation_config.static_forward_context.values():
            if hasattr(layer, "kv_cache"):
5872
5873
5874
5875
                kv_cache = layer.kv_cache
                layer.kv_cache = (
                    torch.tensor([]) if isinstance(kv_cache, torch.Tensor) else []
                )
5876
5877
5878
5879
5880
5881
5882
5883
5884
5885
5886
5887
5888
5889
5890
5891
5892
5893
5894
5895
5896
5897
5898
5899
5900
5901
5902
5903
5904
5905
5906
5907
5908
5909
5910
5911
5912
5913
5914
5915
5916
5917
5918
5919
5920
5921
5922
5923
5924
5925
5926
5927
5928
5929
5930
5931
5932
5933
5934
5935
5936
5937
5938
5939
5940
5941
5942
5943
5944
5945
5946
5947
5948
5949
5950
5951
5952
5953
5954
5955
5956
5957
5958
5959
5960
5961
5962

        gc.collect()
        torch.accelerator.empty_cache()

        logger.debug("Cleaned up profiling KV cache and CUDA graphs")

    @torch.inference_mode()
    def profile_cudagraph_memory(self) -> int:
        with set_current_vllm_config(self.vllm_config):
            self._init_minimal_kv_cache_for_profiling()

        saved_num_cudagraph_captured = compilation_counter.num_cudagraph_captured

        capture_descs = self.cudagraph_dispatcher.get_capture_descs()

        total_graphs = sum(len(descs) for _, descs in capture_descs)
        if total_graphs == 0:
            logger.debug("No CUDA graphs will be captured, skipping profiling")
            self._cleanup_profiling_kv_cache()
            return 0

        logger.info(
            "Profiling CUDA graph memory: %s",
            ", ".join(
                f"{mode.name}={len(descs)} (largest={descs[0].num_tokens})"
                for mode, descs in capture_descs
                if descs
            ),
        )

        # Use a temporary pool for profiling to avoid fragmentation in the main pool.
        profiling_pool = current_platform.graph_pool_handle()
        original_pools: dict[int, Any] = {}
        for instance in list(CUDAGraphWrapper._all_instances):
            original_pools[id(instance)] = instance.graph_pool
            instance.graph_pool = profiling_pool

        set_cudagraph_capturing_enabled(True)
        with self._freeze_gc(), graph_capture(device=self.device):
            shared_memory_estimate = {}
            per_graph_estimate = {}
            torch.accelerator.synchronize()
            torch.accelerator.empty_cache()

            for mode, descs in capture_descs:
                profile_descs = descs[:2]
                mem_samples: list[int] = []

                for i, desc in enumerate(profile_descs):
                    mem_before = torch.cuda.mem_get_info()[0]
                    self._warmup_and_capture(
                        desc,
                        cudagraph_runtime_mode=mode,
                        profile_seq_lens=(
                            min(
                                self.max_model_len,
                                self.max_num_tokens // desc.num_tokens,
                            )
                            if mode == CUDAGraphMode.FULL and i == 0
                            else None
                        ),
                    )
                    torch.accelerator.synchronize()
                    free_after = torch.cuda.mem_get_info()[0]
                    mem_samples.append(mem_before - free_after)

                first_capture = mem_samples[0]
                # Use at least 1 MiB per graph for driver overhead
                per_graph = max(mem_samples[1] if len(mem_samples) > 1 else 0, 1 << 20)

                shared_memory_estimate[mode] = first_capture
                per_graph_estimate[mode] = per_graph * (len(descs) - 1)

                logger.debug(
                    "Estimated %s CUDA graph memory: "
                    "%.2f MiB first-capture + (%d-1) × %.2f MiB per-graph",
                    mode.name,
                    first_capture / (1 << 20),
                    len(descs),
                    per_graph / (1 << 20),
                )

        set_cudagraph_capturing_enabled(False)
        CUDAGraphWrapper.clear_all_graphs()
        for instance in list(CUDAGraphWrapper._all_instances):
            if id(instance) in original_pools:
                instance.graph_pool = original_pools[id(instance)]
5963
5964
5965
        for key_set in self.cudagraph_dispatcher.cudagraph_keys.values():
            key_set.clear()
        self.cudagraph_dispatcher.keys_initialized = False
5966
5967
5968
5969
5970
5971
5972
5973
5974
5975
5976
5977
5978
5979
5980
5981
5982
        self.maybe_remove_all_loras(self.lora_config)
        self._cleanup_profiling_kv_cache()
        compilation_counter.num_cudagraph_captured = saved_num_cudagraph_captured

        # FULL and PIECEWISE graphs share the global pool at runtime and are
        # never replayed concurrently, so the pool overlays their memory.
        # Take the max to avoid double-counting the overlap.
        total_estimate = max(shared_memory_estimate.values()) + sum(
            per_graph_estimate.values()
        )
        logger.info(
            "Estimated CUDA graph memory: %.2f GiB total",
            total_estimate / (1 << 30),
        )

        return int(total_estimate)

5983
    @instrument(span_name="Capture model")
5984
    def capture_model(self) -> int:
5985
        if self.compilation_config.cudagraph_mode == CUDAGraphMode.NONE:
5986
            logger.warning(
5987
                "Skipping CUDA graph capture. To turn on CUDA graph capture, "
5988
5989
                "ensure `cudagraph_mode` was not manually set to `NONE`"
            )
5990
            return 0
5991

5992
5993
5994
5995
5996
5997
5998
5999
6000
6001
6002
6003
6004
        # Initialize encoder CUDA graph manager if enabled.
        # Use get_model() to unwrap CUDAGraphWrapper/UBatchWrapper,
        # because @runtime_checkable Protocol isinstance() checks do not
        # work through __getattr__ forwarding.
        if (
            self.compilation_config.cudagraph_mm_encoder
            and self.supports_mm_inputs
            and self.encoder_cudagraph_manager is None
        ):
            from vllm.model_executor.models.interfaces import (
                SupportsEncoderCudaGraph,
                supports_encoder_cudagraph,
            )
6005
6006
6007
            from vllm.v1.worker.gpu.mm.encoder_cudagraph import (
                EncoderCudaGraphManager,
            )
6008
6009
6010
6011
6012
6013
6014
6015
6016
6017
6018

            raw_model = self.get_model()
            if supports_encoder_cudagraph(raw_model):
                self.encoder_cudagraph_manager = EncoderCudaGraphManager(
                    vllm_config=self.vllm_config,
                    device=self.device,
                    dtype=self.dtype,
                    model=cast(SupportsEncoderCudaGraph, raw_model),
                )
                logger.info("Initialized EncoderCudaGraphManager for vision encoder")

6019
6020
        compilation_counter.num_gpu_runner_capture_triggers += 1

6021
6022
        start_time = time.perf_counter()

6023
6024
6025
        # Trigger CUDA graph capture for specific shapes.
        # Capture the large shapes first so that the smaller shapes
        # can reuse the memory pool allocated for the large shapes.
6026
        set_cudagraph_capturing_enabled(True)
6027
6028
6029
        with self._freeze_gc(), graph_capture(device=self.device):
            torch.accelerator.synchronize()
            torch.accelerator.empty_cache()
6030
            start_free_gpu_memory = torch.cuda.mem_get_info()[0]
6031

6032
6033
6034
6035
            for (
                runtime_mode,
                batch_descs,
            ) in self.cudagraph_dispatcher.get_capture_descs():
6036
                self._capture_cudagraphs(
6037
6038
                    batch_descriptors=batch_descs,
                    cudagraph_runtime_mode=runtime_mode,
6039
                )
6040
                torch.accelerator.synchronize()
6041

6042
6043
6044
6045
            # Capture encoder CUDA graphs if enabled
            if self.encoder_cudagraph_manager is not None:
                self.encoder_cudagraph_manager.capture()

6046
            torch.accelerator.synchronize()
6047
6048
            end_free_gpu_memory = torch.cuda.mem_get_info()[0]

6049
6050
6051
        # Disable cudagraph capturing globally, so any unexpected cudagraph
        # capturing will be detected and raise an error after here.
        # Note: We don't put it into graph_capture context manager because
6052
        # we may do lazy capturing in future that still allows capturing
6053
6054
        # after here.
        set_cudagraph_capturing_enabled(False)
6055

6056
6057
6058
        torch.accelerator.synchronize()
        torch.accelerator.empty_cache()

6059
6060
6061
6062
        # Lock workspace to prevent resizing during execution.
        # Max workspace sizes should have been captured during warmup/profiling.
        lock_workspace()

6063
6064
6065
6066
        end_time = time.perf_counter()
        elapsed_time = end_time - start_time
        cuda_graph_size = start_free_gpu_memory - end_free_gpu_memory
        # This usually takes 5~20 seconds.
6067
        logger.info_once(
6068
6069
6070
            "Graph capturing finished in %.0f secs, took %.2f GiB",
            elapsed_time,
            cuda_graph_size / (1 << 30),
6071
            scope="local",
6072
        )
6073
        return cuda_graph_size
6074

6075
6076
6077
6078
6079
6080
6081
6082
6083
6084
6085
6086
6087
6088
6089
6090
6091
6092
6093
6094
6095
6096
6097
6098
6099
6100
6101
6102
6103
6104
6105
6106
6107
6108
    def _warmup_and_capture(
        self,
        desc: BatchDescriptor,
        cudagraph_runtime_mode: CUDAGraphMode,
        profile_seq_lens: int | None = None,
        allow_microbatching: bool = False,
        num_warmups: int | None = None,
    ):
        if num_warmups is None:
            num_warmups = self.compilation_config.cudagraph_num_of_warmups
        force_attention = cudagraph_runtime_mode == CUDAGraphMode.FULL
        for _ in range(num_warmups):
            self._dummy_run(
                desc.num_tokens,
                cudagraph_runtime_mode=CUDAGraphMode.NONE,
                force_attention=force_attention,
                uniform_decode=desc.uniform,
                allow_microbatching=allow_microbatching,
                skip_eplb=True,
                remove_lora=False,
                num_active_loras=desc.num_active_loras,
            )
        self._dummy_run(
            desc.num_tokens,
            cudagraph_runtime_mode=cudagraph_runtime_mode,
            uniform_decode=desc.uniform,
            allow_microbatching=allow_microbatching,
            skip_eplb=True,
            remove_lora=False,
            num_active_loras=desc.num_active_loras,
            is_graph_capturing=True,
            profile_seq_lens=profile_seq_lens,
        )

6109
6110
    def _capture_cudagraphs(
        self,
6111
        batch_descriptors: list[BatchDescriptor],
6112
6113
6114
6115
        cudagraph_runtime_mode: CUDAGraphMode,
    ):
        assert (
            cudagraph_runtime_mode != CUDAGraphMode.NONE
6116
            and cudagraph_runtime_mode.is_valid_runtime_mode()
6117
        ), f"Invalid cudagraph runtime mode: {cudagraph_runtime_mode}"
6118

6119
6120
6121
6122
6123
        if not batch_descriptors:
            return

        uniform_decode = batch_descriptors[0].uniform

6124
6125
        # Only rank 0 should print progress bar during capture
        if is_global_first_rank():
6126
6127
            batch_descriptors = tqdm(
                batch_descriptors,
6128
6129
6130
                disable=not self.load_config.use_tqdm_on_load,
                desc="Capturing CUDA graphs ({}, {})".format(
                    "decode" if uniform_decode else "mixed prefill-decode",
6131
6132
6133
                    cudagraph_runtime_mode.name,
                ),
            )
6134

6135
        # We skip EPLB here since we don't want to record dummy metrics
6136
        for batch_desc in batch_descriptors:
6137
            # We currently only capture ubatched graphs when its a FULL
6138
6139
6140
            # cudagraph, a uniform decode batch, and the number of tokens
            # is above the threshold. Otherwise we just capture a non-ubatched
            # version of the graph
6141
            allow_microbatching = (
6142
                self.parallel_config.use_ubatching
6143
6144
                and cudagraph_runtime_mode == CUDAGraphMode.FULL
                and uniform_decode
6145
6146
                and check_ubatch_thresholds(
                    config=self.vllm_config.parallel_config,
6147
                    num_tokens=batch_desc.num_tokens,
6148
6149
                    uniform_decode=uniform_decode,
                )
6150
            )
6151
6152
            self._warmup_and_capture(
                batch_desc,
6153
6154
6155
                cudagraph_runtime_mode=cudagraph_runtime_mode,
                allow_microbatching=allow_microbatching,
            )
6156
            torch.accelerator.synchronize()
6157
        self.maybe_remove_all_loras(self.lora_config)
6158

6159
6160
6161
6162
6163
    def initialize_attn_backend(
        self,
        kv_cache_config: KVCacheConfig,
        is_profiling: bool = False,
    ) -> None:
6164
6165
6166
        """
        Initialize the attention backends and attention metadata builders.
        """
6167
        assert len(self.attn_groups) == 0, "Attention backends are already initialized"
6168

6169
6170
6171
6172
6173
6174
        class AttentionGroupKey(NamedTuple):
            attn_backend: type[AttentionBackend]
            kv_cache_spec: KVCacheSpec

        def get_attn_backends_for_group(
            kv_cache_group_spec: KVCacheGroupSpec,
6175
        ) -> tuple[dict[AttentionGroupKey, list[str]], set[type[AttentionBackend]]]:
6176
            layer_type = cast(type[Any], AttentionLayerBase)
6177
            layers = get_layers_from_vllm_config(
6178
                self.vllm_config, layer_type, kv_cache_group_spec.layer_names
6179
            )
6180
6181
            attn_backends = {}
            attn_backend_layers = defaultdict(list)
6182
            # Dedupe based on full class name; this is a bit safer than
6183
6184
6185
6186
            # using the class itself as the key because when we create dynamic
            # attention backend subclasses (e.g. ChunkedLocalAttention) unless
            # they are cached correctly, there will be different objects per
            # layer.
6187
            for layer_name in kv_cache_group_spec.layer_names:
6188
                attn_backend = layers[layer_name].get_attn_backend()
6189
6190
6191
6192

                if layer_name in self.kv_sharing_fast_prefill_eligible_layers:
                    attn_backend = create_fast_prefill_custom_backend(
                        "FastPrefill",
6193
                        attn_backend,  # type: ignore[arg-type]
6194
6195
                    )

6196
6197
6198
                full_cls_name = attn_backend.full_cls_name()
                layer_kv_cache_spec = kv_cache_group_spec.kv_cache_spec
                if isinstance(layer_kv_cache_spec, UniformTypeKVCacheSpecs):
6199
                    layer_kv_cache_spec = layer_kv_cache_spec.kv_cache_specs[layer_name]
6200
                key = (full_cls_name, layer_kv_cache_spec)
6201
6202
6203
                attn_backends[key] = AttentionGroupKey(
                    attn_backend, layer_kv_cache_spec
                )
6204
                attn_backend_layers[key].append(layer_name)
6205
6206
6207
6208
            return (
                {attn_backends[k]: v for k, v in attn_backend_layers.items()},
                set(group_key.attn_backend for group_key in attn_backends.values()),
            )
6209
6210

        def create_attn_groups(
6211
            attn_backends_map: dict[AttentionGroupKey, list[str]],
6212
            kv_cache_group_id: int,
6213
6214
        ) -> list[AttentionGroup]:
            attn_groups: list[AttentionGroup] = []
6215
            for (attn_backend, kv_cache_spec), layer_names in attn_backends_map.items():
6216
                attn_group = AttentionGroup(
6217
                    attn_backend,
6218
                    layer_names,
6219
                    kv_cache_spec,
6220
                    kv_cache_group_id,
6221
6222
                )

6223
6224
6225
                attn_groups.append(attn_group)
            return attn_groups

6226
        attention_backend_maps = []
6227
        attention_backend_list = []
6228
        for kv_cache_group_spec in kv_cache_config.kv_cache_groups:
6229
            attn_backends = get_attn_backends_for_group(kv_cache_group_spec)
6230
            attention_backend_maps.append(attn_backends[0])
6231
            attention_backend_list.append(attn_backends[1])
6232
6233

        # Resolve cudagraph_mode before actually initialize metadata_builders
6234
        self._check_and_update_cudagraph_mode(
6235
6236
6237
            attention_backend_list,
            kv_cache_config.kv_cache_groups,
            is_profiling=is_profiling,
6238
        )
6239

6240
6241
6242
        # Check if attention backend supports PCP&DCP and related features.
        check_attention_cp_compatibility(self.vllm_config)

6243
6244
        for i, attn_backend_map in enumerate(attention_backend_maps):
            self.attn_groups.append(create_attn_groups(attn_backend_map, i))
6245

6246
6247
6248
6249
6250
6251
6252
6253
6254
6255
6256
6257
6258
6259
6260
    def initialize_metadata_builders(
        self, kv_cache_config: KVCacheConfig, kernel_block_sizes: list[int]
    ) -> None:
        """
        Create the metadata builders for all KV cache groups and attn groups.
        """
        for kv_cache_group_id in range(len(kv_cache_config.kv_cache_groups)):
            for attn_group in self.attn_groups[kv_cache_group_id]:
                attn_group.create_metadata_builders(
                    self.vllm_config,
                    self.device,
                    kernel_block_sizes[kv_cache_group_id]
                    if kv_cache_group_id < len(kernel_block_sizes)
                    else None,
                    num_metadata_builders=1
6261
6262
                    if not self.parallel_config.use_ubatching
                    else self.parallel_config.num_ubatches,
6263
                )
co63oc's avatar
co63oc committed
6264
        # Calculate reorder batch threshold (if needed)
6265
6266
        # Note (tdoublep): do this *after* constructing builders,
        # because some of them change the threshold at init time.
6267
6268
        self.calculate_reorder_batch_threshold()

6269
6270
6271
6272
6273
        # Initialize drafter attention backend
        if self.speculative_config and (
            self.speculative_config.use_eagle()
            or self.speculative_config.uses_draft_model()
        ):
6274
6275
6276
            assert isinstance(
                self.drafter, EagleProposer | DFlashProposer | DraftModelProposer
            )
6277
6278
            self.drafter.initialize_attn_backend(kv_cache_config, kernel_block_sizes)

6279
    def _check_and_update_cudagraph_mode(
6280
6281
6282
        self,
        attention_backends: list[set[type[AttentionBackend]]],
        kv_cache_groups: list[KVCacheGroupSpec],
6283
        is_profiling: bool = False,
6284
    ) -> None:
6285
        """
6286
        Resolve the cudagraph_mode when there are multiple attention
6287
        groups with potential conflicting CUDA graph support.
6288
6289
6290
        Then initialize the cudagraph_dispatcher based on the resolved
        cudagraph_mode.
        """
6291
        min_cg_support = AttentionCGSupport.ALWAYS
6292
        min_cg_backend_name = None
6293

6294
6295
6296
6297
6298
6299
6300
6301
6302
6303
6304
6305
        for attn_backend_set, kv_cache_group in zip(
            attention_backends, kv_cache_groups
        ):
            for attn_backend in attn_backend_set:
                builder_cls = attn_backend.get_builder_cls()

                cg_support = builder_cls.get_cudagraph_support(
                    self.vllm_config, kv_cache_group.kv_cache_spec
                )
                if cg_support.value < min_cg_support.value:
                    min_cg_support = cg_support
                    min_cg_backend_name = attn_backend.__name__
6306
6307
        # Flexible resolve the cudagraph mode
        cudagraph_mode = self.compilation_config.cudagraph_mode
6308
        assert cudagraph_mode is not None
6309
        # check cudagraph for mixed batch is supported
6310
6311
6312
6313
6314
6315
        if (
            cudagraph_mode.mixed_mode() == CUDAGraphMode.FULL
            and min_cg_support != AttentionCGSupport.ALWAYS
        ):
            msg = (
                f"CUDAGraphMode.{cudagraph_mode.name} is not supported "
6316
                f"with {min_cg_backend_name} backend (support: "
6317
6318
                f"{min_cg_support})"
            )
6319
6320
            if min_cg_support == AttentionCGSupport.NEVER:
                # if not supported any full cudagraphs, just raise it.
6321
6322
                msg += (
                    "; please try cudagraph_mode=PIECEWISE, and "
6323
                    "make sure compilation mode is VLLM_COMPILE"
6324
                )
6325
6326
6327
6328
6329
                raise ValueError(msg)

            # attempt to resolve the full cudagraph related mode
            if self.compilation_config.splitting_ops_contain_attention():
                msg += "; setting cudagraph_mode=FULL_AND_PIECEWISE"
6330
                cudagraph_mode = self.compilation_config.cudagraph_mode = (
6331
                    CUDAGraphMode.FULL_AND_PIECEWISE
6332
                )
6333
6334
            else:
                msg += "; setting cudagraph_mode=FULL_DECODE_ONLY"
6335
                cudagraph_mode = self.compilation_config.cudagraph_mode = (
6336
                    CUDAGraphMode.FULL_DECODE_ONLY
6337
                )
6338
6339
            logger.warning(msg)

6340
        # check that if we are doing decode full-cudagraphs it is supported
6341
6342
6343
6344
6345
6346
        if (
            cudagraph_mode.decode_mode() == CUDAGraphMode.FULL
            and min_cg_support == AttentionCGSupport.NEVER
        ):
            msg = (
                f"CUDAGraphMode.{cudagraph_mode.name} is not supported "
6347
                f"with {min_cg_backend_name} backend (support: "
6348
6349
                f"{min_cg_support})"
            )
6350
            if self.compilation_config.mode == CompilationMode.VLLM_COMPILE and (
6351
6352
6353
6354
6355
                self.compilation_config.splitting_ops_contain_attention()
                or self.compilation_config.use_inductor_graph_partition
            ):
                msg += (
                    "; setting cudagraph_mode=PIECEWISE because "
6356
                    "attention is compiled piecewise"
6357
6358
                )
                cudagraph_mode = self.compilation_config.cudagraph_mode = (
6359
                    CUDAGraphMode.PIECEWISE
6360
                )
6361
            else:
6362
6363
                msg += (
                    "; setting cudagraph_mode=NONE because "
6364
                    "attention is not compiled piecewise"
6365
6366
                )
                cudagraph_mode = self.compilation_config.cudagraph_mode = (
6367
                    CUDAGraphMode.NONE
6368
                )
6369
6370
            logger.warning(msg)

6371
6372
        # check that if we are doing spec-decode + decode full-cudagraphs it is
        # supported
6373
6374
6375
6376
6377
6378
6379
6380
        if (
            cudagraph_mode.decode_mode() == CUDAGraphMode.FULL
            and self.uniform_decode_query_len > 1
            and min_cg_support.value < AttentionCGSupport.UNIFORM_BATCH.value
        ):
            msg = (
                f"CUDAGraphMode.{cudagraph_mode.name} is not supported"
                f" with spec-decode for attention backend "
6381
                f"{min_cg_backend_name} (support: {min_cg_support})"
6382
            )
6383
6384
            if self.compilation_config.splitting_ops_contain_attention():
                msg += "; setting cudagraph_mode=PIECEWISE"
6385
                cudagraph_mode = self.compilation_config.cudagraph_mode = (
6386
                    CUDAGraphMode.PIECEWISE
6387
                )
6388
6389
            else:
                msg += "; setting cudagraph_mode=NONE"
6390
                cudagraph_mode = self.compilation_config.cudagraph_mode = (
6391
                    CUDAGraphMode.NONE
6392
                )
6393
6394
6395
6396
            logger.warning(msg)

        # double check that we can support full cudagraph if they are requested
        # even after automatic downgrades
6397
6398
6399
6400
6401
6402
        if (
            cudagraph_mode.has_full_cudagraphs()
            and min_cg_support == AttentionCGSupport.NEVER
        ):
            raise ValueError(
                f"CUDAGraphMode.{cudagraph_mode.name} is not "
6403
                f"supported with {min_cg_backend_name} backend ("
6404
6405
                f"support:{min_cg_support}) "
                "; please try cudagraph_mode=PIECEWISE, "
6406
                "and make sure compilation mode is VLLM_COMPILE"
6407
            )
6408

6409
6410
6411
6412
        # if we have dedicated decode cudagraphs, and spec-decode is enabled,
        # we need to adjust the cudagraph sizes to be a multiple of the uniform
        # decode query length to avoid: https://github.com/vllm-project/vllm/issues/28207
        # temp-fix: https://github.com/vllm-project/vllm/issues/28207#issuecomment-3504004536
6413
        # Will be removed in the near future when we have separate cudagraph capture
6414
6415
6416
6417
6418
6419
6420
6421
6422
6423
        # sizes for decode and mixed prefill-decode.
        if (
            cudagraph_mode.decode_mode() == CUDAGraphMode.FULL
            and cudagraph_mode.separate_routine()
            and self.uniform_decode_query_len > 1
        ):
            self.compilation_config.adjust_cudagraph_sizes_for_spec_decode(
                self.uniform_decode_query_len, self.parallel_config.tensor_parallel_size
            )

6424
6425
6426
6427
6428
6429
        # For Mamba models with FULL decode cudagraphs, each decode
        # sequence needs one Mamba cache block. The decode cudagraph
        # dispatcher already caps batch sizes at max_num_seqs, so we just
        # need to verify that enough blocks exist. Raising here instead
        # of silently capping cudagraph_capture_sizes avoids unintended
        # restrictions on PIECEWISE (prefill) cudagraphs.
6430
        # See: https://github.com/vllm-project/vllm/issues/34094
6431
        if cudagraph_mode.has_full_cudagraphs() and not is_profiling:
6432
6433
6434
6435
            has_mamba = any(
                isinstance(g.kv_cache_spec, MambaSpec) for g in kv_cache_groups
            )
            if has_mamba and self.kv_cache_config is not None:
6436
6437
6438
6439
6440
6441
6442
6443
6444
6445
6446
                num_blocks = self.kv_cache_config.num_blocks
                if self.max_num_reqs > num_blocks:
                    raise ValueError(
                        f"max_num_seqs ({self.max_num_reqs}) exceeds "
                        f"available Mamba cache blocks ({num_blocks}). "
                        f"Each decode sequence requires one Mamba cache "
                        f"block, so CUDA graph capture cannot proceed. "
                        f"Please lower max_num_seqs to at most "
                        f"{num_blocks} or increase "
                        f"gpu_memory_utilization."
                    )
6447

6448
6449
        # Trigger cudagraph dispatching keys initialization after
        # resolved cudagraph mode.
6450
        self.compilation_config.cudagraph_mode = cudagraph_mode
6451
        self.cudagraph_dispatcher.initialize_cudagraph_keys(
6452
            cudagraph_mode, self.uniform_decode_query_len
6453
        )
6454

6455
6456
6457
6458
6459
        # Initialize drafter's cudagraph dispatcher if using spec decode.
        if self.speculative_config and (
            self.speculative_config.use_eagle()
            or self.speculative_config.uses_extract_hidden_states()
        ):
6460
6461
6462
6463
            assert isinstance(
                self.drafter,
                EagleProposer | DFlashProposer | ExtractHiddenStatesProposer,
            )
6464
6465
            self.drafter.initialize_cudagraph_keys(cudagraph_mode)

6466
6467
    def calculate_reorder_batch_threshold(self) -> None:
        """
6468
6469
6470
6471
        Choose the minimum reorder batch threshold from all attention groups.
        Backends should be able to support lower threshold then what they request
        just may have a performance penalty due to that backend treating decodes
        as prefills.
6472
        """
6473
6474
        min_none_high = lambda a, b: a if b is None else b if a is None else min(a, b)

6475
        reorder_batch_thresholds: list[int | None] = [
6476
6477
6478
            group.get_metadata_builder().reorder_batch_threshold
            for group in self._attn_group_iterator()
        ]
6479
6480
6481
6482
6483
        # If there are no attention groups (attention-free model) or no backend
        # reports a threshold, leave reordering disabled.
        if len(reorder_batch_thresholds) == 0:
            self.reorder_batch_threshold = None
            return
6484
        self.reorder_batch_threshold = reduce(min_none_high, reorder_batch_thresholds)  # type: ignore[assignment]
6485

6486
6487
6488
    def may_reinitialize_input_batch(
        self, kv_cache_config: KVCacheConfig, kernel_block_sizes: list[int]
    ) -> None:
6489
6490
        """
        Re-initialize the input batch if the block sizes are different from
6491
6492
6493
6494
        what it was originally created with. This happens when the final
        block size (determined after model loading) differs from the
        placeholder used during __init__, or when there are multiple
        KV cache groups.
6495
6496
6497

        Args:
            kv_cache_config: The KV cache configuration.
6498
            kernel_block_sizes: The kernel block sizes for each KV cache group.
6499
        """
6500
        block_sizes = []
6501
6502
        max_num_blocks = []
        max_model_len = max(self.max_model_len, self.max_encoder_len)
6503
        for kv_cache_group in kv_cache_config.kv_cache_groups:
6504
6505
            if isinstance(kv_cache_group.kv_cache_spec, EncoderOnlyAttentionSpec):
                continue
6506
6507
            block_size = kv_cache_group.kv_cache_spec.block_size
            block_sizes.append(block_size)
6508
            max_num_blocks_per_req = cdiv(
6509
                max_model_len, block_size * get_total_cp_world_size()
6510
6511
            )
            if isinstance(kv_cache_group.kv_cache_spec, MambaSpec):
6512
                max_num_blocks_per_req = (
6513
6514
6515
6516
6517
                    max_num_blocks_per_req
                    if self.cache_config.enable_prefix_caching
                    else 1
                ) + kv_cache_group.kv_cache_spec.num_speculative_blocks
            max_num_blocks.append(max_num_blocks_per_req)
6518

6519
6520
6521
6522
        if (
            block_sizes != self._init_block_sizes
            or kernel_block_sizes != self._init_kernel_block_sizes
        ):
6523
            assert self.offload_config.uva.cpu_offload_gb == 0, (
6524
6525
                "Cannot re-initialize the input batch when CPU weight "
                "offloading is enabled. See https://github.com/vllm-project/vllm/pull/18298 "  # noqa: E501
6526
6527
                "for more details."
            )
6528
6529
            self._init_block_sizes = block_sizes
            self._init_kernel_block_sizes = kernel_block_sizes
6530
6531
            self.input_batch = InputBatch(
                max_num_reqs=self.max_num_reqs,
6532
                max_model_len=max_model_len,
6533
6534
6535
6536
6537
                max_num_batched_tokens=self.max_num_tokens,
                device=self.device,
                pin_memory=self.pin_memory,
                vocab_size=self.model_config.get_vocab_size(),
                block_sizes=block_sizes,
6538
                kernel_block_sizes=kernel_block_sizes,
6539
                max_num_blocks_per_req=max_num_blocks,
6540
                is_spec_decode=bool(self.vllm_config.speculative_config),
6541
                logitsprocs=self.input_batch.logitsprocs,
6542
                logitsprocs_need_output_token_ids=self.input_batch.logitsprocs_need_output_token_ids,
6543
                is_pooling_model=self.is_pooling_model,
6544
6545
            )

6546
6547
6548
6549
6550
6551
6552
6553
6554
        assert self._init_block_sizes == block_sizes, (
            f"InputBatch block_sizes {self._init_block_sizes} != "
            f"kv_cache block_sizes {block_sizes}"
        )
        assert self._init_kernel_block_sizes == kernel_block_sizes, (
            f"InputBatch kernel_block_sizes {self._init_kernel_block_sizes} "
            f"!= kv_cache kernel_block_sizes {kernel_block_sizes}"
        )

6555
    def _allocate_kv_cache_tensors(
6556
6557
        self, kv_cache_config: KVCacheConfig
    ) -> dict[str, torch.Tensor]:
6558
        """
6559
6560
6561
        Initializes the KV cache buffer with the correct size. The buffer needs
        to be reshaped to the desired shape before being used by the models.

6562
        Args:
6563
            kv_cache_config: The KV cache config
6564
        Returns:
6565
            dict[str, torch.Tensor]: A map between layer names to their
6566
            corresponding memory buffer for KV cache.
6567
        """
6568
6569
        kv_cache_raw_tensors: dict[str, torch.Tensor] = {}
        for kv_cache_tensor in kv_cache_config.kv_cache_tensors:
6570
6571
6572
            tensor = torch.zeros(
                kv_cache_tensor.size, dtype=torch.int8, device=self.device
            )
6573
6574
6575
6576
6577
            for layer_name in kv_cache_tensor.shared_by:
                kv_cache_raw_tensors[layer_name] = tensor

        layer_names = set()
        for group in kv_cache_config.kv_cache_groups:
6578
6579
6580
6581
            for layer_name in group.layer_names:
                if layer_name in self.runner_only_attn_layers:
                    continue
                layer_names.add(layer_name)
6582
6583
6584
        assert layer_names == set(kv_cache_raw_tensors.keys()), (
            "Some layers are not correctly initialized"
        )
6585
6586
        return kv_cache_raw_tensors

6587
6588
6589
    def _attn_group_iterator(self) -> Iterator[AttentionGroup]:
        return itertools.chain.from_iterable(self.attn_groups)

6590
    def _kv_cache_spec_attn_group_iterator(self) -> Iterator[AttentionGroup]:
6591
6592
        if not self.kv_cache_config.kv_cache_groups:
            return
6593
6594
        for attn_groups in self.attn_groups:
            yield from attn_groups
6595

6596
6597
6598
6599
    def _reshape_kv_cache_tensors(
        self,
        kv_cache_config: KVCacheConfig,
        kv_cache_raw_tensors: dict[str, torch.Tensor],
6600
        kernel_block_sizes: list[int],
6601
    ) -> dict[str, torch.Tensor]:
6602
        """
6603
        Reshape the KV cache tensors to the desired shape and dtype.
6604

6605
        Args:
6606
6607
            kv_cache_config: The KV cache config
            kv_cache_raw_tensors: The KV cache buffer of each layer, with
6608
                correct size but uninitialized shape.
6609
            kernel_block_sizes: The kernel block sizes for each KV cache group.
6610
        Returns:
6611
            Dict[str, torch.Tensor]: A map between layer names to their
6612
6613
            corresponding memory buffer for KV cache.
        """
6614
        kv_caches: dict[str, torch.Tensor] = {}
6615
        has_attn, has_mamba = False, False
6616
6617
        for group in self._kv_cache_spec_attn_group_iterator():
            kv_cache_spec = group.kv_cache_spec
6618
            attn_backend = group.backend
6619
6620
6621
6622
            if group.kv_cache_group_id == len(kernel_block_sizes):
                # There may be a last group for layers without kv cache.
                continue
            kernel_block_size = kernel_block_sizes[group.kv_cache_group_id]
6623
            for layer_name in group.layer_names:
6624
6625
                if layer_name in self.runner_only_attn_layers:
                    continue
6626
6627
                raw_tensor = kv_cache_raw_tensors[layer_name]
                assert raw_tensor.numel() % kv_cache_spec.page_size_bytes == 0
6628
                num_blocks = raw_tensor.numel() // kv_cache_spec.page_size_bytes
6629
                if isinstance(kv_cache_spec, AttentionSpec):
6630
                    has_attn = True
6631
6632
                    num_blocks_per_kv_block = (
                        kv_cache_spec.block_size // kernel_block_size
6633
6634
6635
                    )
                    kernel_num_blocks = num_blocks * num_blocks_per_kv_block

6636
                    kv_cache_shape = attn_backend.get_kv_cache_shape(
6637
                        kernel_num_blocks,
6638
                        kernel_block_size,
6639
6640
                        kv_cache_spec.num_kv_heads,
                        kv_cache_spec.head_size,
6641
6642
                        cache_dtype_str=self.cache_config.cache_dtype,
                    )
6643
                    dtype = kv_cache_spec.dtype
6644
                    try:
6645
                        kv_cache_stride_order = attn_backend.get_kv_cache_stride_order()
6646
                        assert len(kv_cache_stride_order) == len(kv_cache_shape)
6647
                    except (AttributeError, NotImplementedError):
6648
                        kv_cache_stride_order = tuple(range(len(kv_cache_shape)))
6649
6650
6651
6652
6653
                    # The allocation respects the backend-defined stride order
                    # to ensure the semantic remains consistent for each
                    # backend. We first obtain the generic kv cache shape and
                    # then permute it according to the stride order which could
                    # result in a non-contiguous tensor.
6654
6655
6656
                    kv_cache_shape = tuple(
                        kv_cache_shape[i] for i in kv_cache_stride_order
                    )
6657
6658
6659
6660
6661
                    # Maintain original KV shape view.
                    inv_order = [
                        kv_cache_stride_order.index(i)
                        for i in range(len(kv_cache_stride_order))
                    ]
6662
6663
6664
6665
6666
6667
                    kv_caches[layer_name] = (
                        kv_cache_raw_tensors[layer_name]
                        .view(dtype)
                        .view(kv_cache_shape)
                        .permute(*inv_order)
                    )
Chen Zhang's avatar
Chen Zhang committed
6668
                elif isinstance(kv_cache_spec, MambaSpec):
6669
                    has_mamba = True
Chen Zhang's avatar
Chen Zhang committed
6670
6671
                    raw_tensor = kv_cache_raw_tensors[layer_name]
                    state_tensors = []
6672
                    storage_offset_bytes = 0
6673
                    for shape, dtype in zip(kv_cache_spec.shapes, kv_cache_spec.dtypes):
6674
6675
                        dtype_size = get_dtype_size(dtype)
                        num_element_per_page = (
6676
6677
                            kv_cache_spec.page_size_bytes // dtype_size
                        )
Chen Zhang's avatar
Chen Zhang committed
6678
                        target_shape = (num_blocks, *shape)
6679
6680
                        stride = torch.empty(target_shape).stride()
                        target_stride = (num_element_per_page, *stride[1:])
6681
                        assert storage_offset_bytes % dtype_size == 0
6682
6683
6684
6685
                        tensor = torch.as_strided(
                            raw_tensor.view(dtype),
                            size=target_shape,
                            stride=target_stride,
6686
                            storage_offset=storage_offset_bytes // dtype_size,
6687
                        )
Chen Zhang's avatar
Chen Zhang committed
6688
                        state_tensors.append(tensor)
6689
                        storage_offset_bytes += stride[0] * dtype_size
6690
6691

                    kv_caches[layer_name] = state_tensors
6692
                else:
6693
                    raise NotImplementedError
6694
6695

        if has_attn and has_mamba:
6696
            self._update_hybrid_attention_mamba_layout(kv_caches, kernel_block_sizes)
6697

6698
6699
        return kv_caches

6700
    def _update_hybrid_attention_mamba_layout(
6701
        self, kv_caches: dict[str, torch.Tensor], kernel_block_sizes: list[int]
6702
    ) -> None:
6703
        """
6704
6705
        Update the layout of attention layers from (2, num_blocks, ...) to
        (num_blocks, 2, ...).
6706
6707

        Args:
6708
            kv_caches: The KV cache buffer of each layer.
6709
            kernel_block_sizes: The kernel block sizes for each KV cache group.
6710
6711
        """

6712
6713
        for group in self._kv_cache_spec_attn_group_iterator():
            kv_cache_spec = group.kv_cache_spec
6714
6715
6716
6717
6718
6719
6720
6721
6722
6723
6724
6725
            if not isinstance(kv_cache_spec, AttentionSpec):
                continue
            block_dim = group.backend.get_kv_cache_block_dim(
                kernel_block_sizes[group.kv_cache_group_id],
                kv_cache_spec.num_kv_heads,
                kv_cache_spec.head_size,
                cache_dtype_str=self.cache_config.cache_dtype,
            )
            # block_dim: 0 means (num_blocks, 2, ...); 1 means (2, num_blocks, ...).
            if block_dim == 0:
                continue
            assert block_dim == 1
6726
            for layer_name in group.layer_names:
6727
                kv_cache = kv_caches[layer_name]
6728
6729
6730
6731
6732
                hidden_size = kv_cache.shape[2:].numel()
                kv_cache.as_strided_(
                    size=kv_cache.shape,
                    stride=(hidden_size, 2 * hidden_size, *kv_cache.stride()[2:]),
                )
6733

6734
    def initialize_kv_cache_tensors(
6735
        self, kv_cache_config: KVCacheConfig, kernel_block_sizes: list[int]
6736
    ) -> dict[str, torch.Tensor]:
6737
6738
6739
6740
6741
        """
        Initialize the memory buffer for KV cache.

        Args:
            kv_cache_config: The KV cache config
6742
6743
            kernel_block_sizes: The kernel block sizes for each KV cache group.

6744
        Returns:
6745
            Dict[str, torch.Tensor]: A map between layer names to their
6746
6747
            corresponding memory buffer for KV cache.
        """
6748
6749
6750
6751
6752
6753
6754
6755
6756
6757
6758
6759
6760
6761
6762
6763
6764
6765
6766
6767
6768
6769
6770
6771

        # Try creating KV caches optimized for kv-connector transfers
        cache_dtype = self.cache_config.cache_dtype
        if self.use_uniform_kv_cache(self.attn_groups, cache_dtype):
            kv_caches, cross_layers_kv_cache, attn_backend = (
                self.allocate_uniform_kv_caches(
                    kv_cache_config,
                    self.attn_groups,
                    cache_dtype,
                    self.device,
                    kernel_block_sizes,
                )
            )
            self.cross_layers_kv_cache = cross_layers_kv_cache
            self.cross_layers_attn_backend = attn_backend
        else:
            # Fallback to the general case
            # Initialize the memory buffer for KV cache
            kv_cache_raw_tensors = self._allocate_kv_cache_tensors(kv_cache_config)

            # Change the memory buffer to the desired shape
            kv_caches = self._reshape_kv_cache_tensors(
                kv_cache_config, kv_cache_raw_tensors, kernel_block_sizes
            )
6772

6773
        # Set up cross-layer KV cache sharing
6774
6775
        for layer_name, target_layer_name in self.shared_kv_cache_layers.items():
            logger.debug("%s reuses KV cache of %s", layer_name, target_layer_name)
6776
6777
            kv_caches[layer_name] = kv_caches[target_layer_name]

6778
6779
6780
6781
6782
6783
6784
6785
6786
        num_attn_module = (
            2 if self.model_config.hf_config.model_type == "longcat_flash" else 1
        )
        bind_kv_cache(
            kv_caches,
            self.compilation_config.static_forward_context,
            self.kv_caches,
            num_attn_module,
        )
6787
6788
6789
        return kv_caches

    def maybe_add_kv_sharing_layers_to_kv_cache_groups(
6790
6791
        self, kv_cache_config: KVCacheConfig
    ) -> None:
6792
6793
6794
6795
6796
6797
6798
6799
6800
6801
6802
6803
6804
6805
6806
6807
6808
6809
        """
        Add layers that re-use KV cache to KV cache group of its target layer.
        Mapping of KV cache tensors happens in `initialize_kv_cache_tensors()`
        """
        if not self.shared_kv_cache_layers:
            # No cross-layer KV sharing, return
            return

        add_kv_sharing_layers_to_kv_cache_groups(
            self.shared_kv_cache_layers,
            kv_cache_config.kv_cache_groups,
            self.runner_only_attn_layers,
        )

        if self.cache_config.kv_sharing_fast_prefill:
            # In You Only Cache Once (https://arxiv.org/abs/2405.05254) or other
            # similar KV sharing setups, only the layers that generate KV caches
            # are involved in the prefill phase, enabling prefill to early exit.
6810
            attn_layers = get_layers_from_vllm_config(self.vllm_config, Attention)
6811
6812
            for layer_name in reversed(attn_layers):
                if layer_name in self.shared_kv_cache_layers:
6813
                    self.kv_sharing_fast_prefill_eligible_layers.add(layer_name)
6814
6815
                else:
                    break
6816

6817
6818
6819
6820
6821
    def initialize_kv_cache(
        self,
        kv_cache_config: KVCacheConfig,
        is_profiling: bool = False,
    ) -> None:
6822
6823
6824
6825
6826
6827
        """
        Initialize KV cache based on `kv_cache_config`.
        Args:
            kv_cache_config: Configuration for the KV cache, including the KV
            cache size of each layer
        """
6828
        kv_cache_config = deepcopy(kv_cache_config)
6829
        self.kv_cache_config = kv_cache_config
6830
        self._mamba_copy_bufs = None
6831
        self.may_add_encoder_only_layers_to_kv_cache_config()
6832
        self.maybe_add_kv_sharing_layers_to_kv_cache_groups(kv_cache_config)
6833
        self.initialize_attn_backend(kv_cache_config, is_profiling=is_profiling)
6834
6835
6836
6837
6838
        # The kernel block size for all KV cache groups. For example, if
        # kv_cache_manager uses block_size 256 for a given group, but the attention
        # backends for that group only supports block_size 64, we will return
        # kernel_block_size 64 and split the 256-token-block to 4 blocks with 64
        # tokens each.
6839
6840
6841
        kernel_block_sizes = prepare_kernel_block_sizes(
            kv_cache_config, self.attn_groups
        )
6842
        self._kernel_block_sizes = kernel_block_sizes
6843
6844
6845
6846

        # create metadata builders
        self.initialize_metadata_builders(kv_cache_config, kernel_block_sizes)

6847
        # Reinitialize need to after initialize_attn_backend
6848
6849
6850
6851
        self.may_reinitialize_input_batch(kv_cache_config, kernel_block_sizes)
        kv_caches = self.initialize_kv_cache_tensors(
            kv_cache_config, kernel_block_sizes
        )
6852

6853
6854
6855
        if (
            self.speculative_config
            and self.speculative_config.uses_extract_hidden_states()
6856
        ):
6857
            assert isinstance(self.drafter, ExtractHiddenStatesProposer)
6858
6859
6860
6861
            # validate all draft model layers belong to the same kv cache
            # group
            self.drafter.validate_same_kv_cache_group(kv_cache_config)

Robert Shaw's avatar
Robert Shaw committed
6862
        if has_kv_transfer_group():
6863
            kv_transfer_group = get_kv_transfer_group()
6864
6865
6866
6867
6868
6869
6870
            if self.cross_layers_kv_cache is not None:
                assert self.cross_layers_attn_backend is not None
                kv_transfer_group.register_cross_layers_kv_cache(
                    self.cross_layers_kv_cache, self.cross_layers_attn_backend
                )
            else:
                kv_transfer_group.register_kv_caches(kv_caches)
6871
            kv_transfer_group.set_host_xfer_buffer_ops(copy_kv_blocks)
Robert Shaw's avatar
Robert Shaw committed
6872

6873
6874
6875
6876
6877
6878
    def _get_attention_kv_cache_gid(self) -> int:
        """Find the KV cache group index for attention layers."""
        for gid, group in enumerate(self.kv_cache_config.kv_cache_groups):
            if isinstance(group.kv_cache_spec, AttentionSpec):
                return gid
        return 0
6879
6880
6881
6882
6883
6884
6885

    def init_routed_experts_capturer(self):
        logger.info(
            "Initializing routed experts capturer, enable_return_routed_experts: %s",
            self.model_config.enable_return_routed_experts,
        )
        routed_experts_capturer = RoutedExpertsCapturer.create()
6886
6887
6888
6889
6890
6891
6892
6893
        self.routed_experts_attn_gid = self._get_attention_kv_cache_gid()
        min_block_size = min(
            [
                group.kv_cache_spec.block_size
                for group in self.kv_cache_config.kv_cache_groups
            ]
        )
        num_groups = len(self.kv_cache_config.kv_cache_groups)
6894
        self.max_num_kv_tokens = (
6895
6896
6897
6898
6899
6900
6901
            self.kv_cache_config.num_blocks // num_groups
        ) * min_block_size
        dcp_size = self.vllm_config.parallel_config.decode_context_parallel_size
        pcp_size = self.vllm_config.parallel_config.prefill_context_parallel_size
        if pcp_size * dcp_size > 1:
            self.max_num_kv_tokens *= pcp_size * dcp_size

6902
6903
6904
        routed_experts_capturer.init_buffer(
            max_num_batched_tokens=self.scheduler_config.max_num_batched_tokens,
            max_num_kv_tokens=self.max_num_kv_tokens,
6905
            vllm_config=self.vllm_config,
6906
        )
6907
        self._bind_routed_experts_capturer(routed_experts_capturer)
6908
        self.routed_experts_initialized = True
6909
6910
6911
6912
6913
6914
6915
6916
6917
6918
6919
6920
6921
6922
6923

    def _bind_routed_experts_capturer(self, capturer: RoutedExpertsCapturer) -> None:
        from vllm.model_executor.layers.fused_moe.layer import FusedMoE
        from vllm.model_executor.layers.fused_moe.router.base_router import (
            BaseRouter,
        )

        for module in self.compilation_config.static_forward_context.values():
            if isinstance(module, FusedMoE) and isinstance(module.router, BaseRouter):
                layer_id = module.layer_id

                def _capture_fn(topk_ids, _layer_id=layer_id, _capturer=capturer):
                    _capturer.capture(_layer_id, topk_ids)

                module.router.set_capture_fn(_capture_fn)
6924

6925
6926
6927
6928
6929
    def may_add_encoder_only_layers_to_kv_cache_config(self) -> None:
        """
        Add encoder-only layers to the KV cache config.
        """
        block_size = self.vllm_config.cache_config.block_size
6930
        encoder_only_attn_specs: dict[AttentionSpec, list[str]] = defaultdict(list)
6931
6932
6933
        attn_layers = get_layers_from_vllm_config(self.vllm_config, Attention)
        for layer_name, attn_module in attn_layers.items():
            if attn_module.attn_type == AttentionType.ENCODER_ONLY:
6934
                attn_spec: AttentionSpec = EncoderOnlyAttentionSpec(
6935
6936
6937
                    block_size=block_size,
                    num_kv_heads=attn_module.num_kv_heads,
                    head_size=attn_module.head_size,
6938
6939
                    dtype=self.kv_cache_dtype,
                )
6940
6941
6942
                encoder_only_attn_specs[attn_spec].append(layer_name)
                self.runner_only_attn_layers.add(layer_name)
        if len(encoder_only_attn_specs) > 0:
6943
6944
6945
            assert len(encoder_only_attn_specs) == 1, (
                "Only support one encoder-only attention spec now"
            )
6946
6947
            spec, layer_names = encoder_only_attn_specs.popitem()
            self.kv_cache_config.kv_cache_groups.append(
6948
6949
                KVCacheGroupSpec(layer_names=layer_names, kv_cache_spec=spec)
            )
6950

6951
    def get_kv_cache_spec(self) -> dict[str, KVCacheSpec]:
6952
        """
6953
        Generates the KVCacheSpec by parsing the kv cache format from each
6954
6955
        Attention module in the static forward context.
        Returns:
6956
            KVCacheSpec: A dictionary mapping layer names to their KV cache
6957
6958
            format. Layers that do not need KV cache are not included.
        """
6959
        if has_ec_transfer() and not get_ec_transfer().is_consumer:
6960
            return {}
6961
        kv_cache_spec: dict[str, KVCacheSpec] = {}
6962
6963
        layer_type = cast(type[Any], AttentionLayerBase)
        attn_layers = get_layers_from_vllm_config(self.vllm_config, layer_type)
Chen Zhang's avatar
Chen Zhang committed
6964
        for layer_name, attn_module in attn_layers.items():
6965
6966
6967
6968
6969
6970
6971
6972
6973
6974
6975
6976
6977
6978
6979
            if isinstance(attn_module, Attention) and (
                kv_tgt_layer := attn_module.kv_sharing_target_layer_name
            ):
                # The layer doesn't need its own KV cache and will use that of
                # the target layer. We skip creating a KVCacheSpec for it, so
                # that KV cache management logic will act as this layer does
                # not exist, and doesn't allocate KV cache for the layer. This
                # enables the memory saving of cross-layer kv sharing, allowing
                # a given amount of memory to accommodate longer context lengths
                # or enable more requests to be processed simultaneously.
                self.shared_kv_cache_layers[layer_name] = kv_tgt_layer
                continue
            # Skip modules that don't need KV cache (eg encoder-only attention)
            if spec := attn_module.get_kv_cache_spec(self.vllm_config):
                kv_cache_spec[layer_name] = spec
6980

6981
        return kv_cache_spec
6982

6983
    def _to_list(self, sampled_token_ids: torch.Tensor) -> list[list[int]]:
6984
6985
6986
6987
6988
6989
6990
6991
        # This is a short term mitigation for issue mentioned in
        # https://github.com/vllm-project/vllm/issues/22754.
        # `tolist` would trigger a cuda wise stream sync, which
        # would block other copy ops from other cuda streams.
        # A cuda event sync would avoid such a situation. Since
        # this is in the critical path of every single model
        # forward loop, this has caused perf issue for a disagg
        # setup.
6992
        pinned = self.sampled_token_ids_pinned_cpu[: sampled_token_ids.shape[0]]
6993
6994
6995
        pinned.copy_(sampled_token_ids, non_blocking=True)
        self.transfer_event.record()
        self.transfer_event.synchronize()
6996
        return pinned.tolist()
6997
6998
6999
7000
7001
7002
7003
7004
7005
7006
7007
7008
7009
7010
7011
7012
7013
7014
7015
7016
7017
7018
7019
7020
7021
7022
7023
7024
7025
7026
7027
7028
7029
7030
7031
7032
7033
7034
7035
7036

    def get_encoder_timing_stats(self) -> dict[str, dict[str, float | int]]:
        """
        Get encoder timing stats for all requests and clear the registry.

        Returns:
            Dictionary mapping request_id to stats dict.
        """
        with self._encoder_timing_lock:
            stats = {
                req_id: stats_obj.to_dict()
                for req_id, stats_obj in self.encoder_timing_registry.items()
            }
            self.encoder_timing_registry.clear()
            return stats

    @contextmanager
    def timed_encoder_operation(
        self,
        should_time: bool,
        group_lora_refs: list[tuple[str, Any]],
        current_item_idx: int,
        num_items: int,
    ):
        """
        Context manager to time encoder forward operations.

        Args:
            should_time: Whether timing is enabled
            group_lora_refs: Full list of (request_id, pos_info) tuples
            current_item_idx: Starting index for this group
            num_items: Number of items in this group
        """
        if not should_time:
            yield
            return

        group_refs = group_lora_refs[current_item_idx : current_item_idx + num_items]
        group_request_ids = {req_id for req_id, _ in group_refs}

7037
        torch.accelerator.synchronize()
7038
7039
7040
7041
7042
        start_time = time.perf_counter()

        try:
            yield
        finally:
7043
            torch.accelerator.synchronize()
7044
7045
7046
7047
7048
7049
7050
7051
7052
7053
            elapsed = time.perf_counter() - start_time

            per_request_time = elapsed / max(len(group_request_ids), 1)

            with self._encoder_timing_lock:
                for req_id in group_request_ids:
                    if req_id not in self.encoder_timing_registry:
                        self.encoder_timing_registry[req_id] = EncoderTimingStats()

                    stats = self.encoder_timing_registry[req_id]
7054
                    stats.encoder_forward_secs += per_request_time
7055
7056
7057
7058
7059
7060
7061
                    stats.num_encoder_calls += 1


@dataclass
class EncoderTimingStats:
    """Per-request timing statistics for encoder forward pass."""

7062
    encoder_forward_secs: float = 0.0
7063
7064
7065
7066
7067
7068
7069
    """Time spent in vision encoder forward pass (seconds)."""

    num_encoder_calls: int = 0
    """Number of times encoder was called for this request."""

    def to_dict(self) -> dict[str, float | int]:
        return {
7070
            "encoder_forward_secs": self.encoder_forward_secs,
7071
7072
            "num_encoder_calls": self.num_encoder_calls,
        }