gpu_model_runner.py 291 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 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
112
113
from vllm.utils.torch_utils import (
    get_dtype_size,
    kv_cache_dtype_str_to_dtype,
)
114
115
from vllm.v1.attention.backend import (
    AttentionBackend,
116
    AttentionCGSupport,
117
    AttentionMetadata,
118
    AttentionMetadataBuilder,
119
    AttentionType,
120
    CommonAttentionMetadata,
121
)
122
from vllm.v1.attention.backends.gdn_attn import GDNAttentionMetadataBuilder
123
from vllm.v1.attention.backends.mamba2_attn import Mamba2AttentionMetadataBuilder
124
from vllm.v1.attention.backends.utils import (
125
    create_fast_prefill_custom_backend,
126
    get_dcp_local_seq_lens,
127
128
    reorder_batch_to_split_decodes_and_prefills,
)
129
from vllm.v1.core.sched.output import NewRequestData
130
from vllm.v1.cudagraph_dispatcher import CudagraphDispatcher
131
132
133
134
135
136
137
138
139
140
141
142
143
144
145
146
147
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,
148
    ECConnectorOutput,
149
    KVConnectorOutput,
150
151
152
153
154
    LogprobsLists,
    LogprobsTensors,
    ModelRunnerOutput,
    PoolerOutput,
    SamplerOutput,
155
    make_empty_encoder_model_runner_output,
156
)
157
from vllm.v1.pool.metadata import PoolingMetadata, PoolingStates
158
from vllm.v1.sample.logits_processor import LogitsProcessors, build_logitsprocs
159
from vllm.v1.sample.logits_processor.interface import LogitsProcessor
160
from vllm.v1.sample.metadata import SamplingMetadata
161
from vllm.v1.sample.rejection_sampler import RejectionSampler
162
from vllm.v1.sample.sampler import Sampler
163
from vllm.v1.spec_decode.draft_model import DraftModelProposer
164
from vllm.v1.spec_decode.eagle import EagleProposer
165
from vllm.v1.spec_decode.extract_hidden_states import ExtractHiddenStatesProposer
166
from vllm.v1.spec_decode.medusa import MedusaProposer
167
from vllm.v1.spec_decode.metadata import SpecDecodeMetadata
168
169
170
171
172
173
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,
)
174
from vllm.v1.spec_decode.suffix_decoding import SuffixDecodingProposer
175
from vllm.v1.structured_output.utils import apply_grammar_bitmask
176
from vllm.v1.utils import CpuGpuBuffer, record_function_or_nullcontext
177
178
179
180
181
from vllm.v1.worker import mamba_utils
from vllm.v1.worker.cp_utils import (
    check_attention_cp_compatibility,
    get_total_cp_world_size,
)
182
from vllm.v1.worker.dp_utils import coordinate_batch_across_dp
183
from vllm.v1.worker.ec_connector_model_runner_mixin import ECConnectorModelRunnerMixin
184
from vllm.v1.worker.gpu.pool.late_interaction_runner import LateInteractionRunner
185
from vllm.v1.worker.gpu_input_batch import CachedRequestState, InputBatch
186
from vllm.v1.worker.gpu_ubatch_wrapper import UBatchWrapper
187
from vllm.v1.worker.kv_connector_model_runner_mixin import KVConnectorModelRunnerMixin
188
from vllm.v1.worker.lora_model_runner_mixin import LoRAModelRunnerMixin
189
190
191
from vllm.v1.worker.ubatch_utils import (
    UBatchSlices,
    check_ubatch_thresholds,
192
    maybe_create_ubatch_slices,
193
    split_attn_metadata,
194
)
195
from vllm.v1.worker.utils import is_residual_scattered_for_sp
196
from vllm.v1.worker.workspace import lock_workspace
197

198
199
from .utils import (
    AttentionGroup,
200
    KVBlockZeroer,
201
202
    add_kv_sharing_layers_to_kv_cache_groups,
    bind_kv_cache,
203
    prepare_kernel_block_sizes,
204
205
    sanity_check_mm_encoder_outputs,
)
206

207
if TYPE_CHECKING:
208
    from vllm.v1.core.sched.output import GrammarOutput, SchedulerOutput
209
    from vllm.v1.spec_decode.ngram_proposer import NgramProposer
210
    from vllm.v1.worker.gpu.mm.encoder_cudagraph import EncoderCudaGraphManager
211
212
213

logger = init_logger(__name__)

214
215
AttnMetadataDict: TypeAlias = dict[str, AttentionMetadata]
# list when ubatching is enabled
216
PerLayerAttnMetadata: TypeAlias = list[AttnMetadataDict] | AttnMetadataDict
217

218

219
220
221
222
223
224
# Wrapper for ModelRunnerOutput to support overlapped execution.
class AsyncGPUModelRunnerOutput(AsyncModelRunnerOutput):
    def __init__(
        self,
        model_runner_output: ModelRunnerOutput,
        sampled_token_ids: torch.Tensor,
225
        logprobs_tensors: LogprobsTensors | None,
226
227
        invalid_req_indices: list[int],
        async_output_copy_stream: torch.cuda.Stream,
228
        vocab_size: int,
229
230
231
232
233
    ):
        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.
234
        self.async_copy_ready_event = torch.Event()
235
236
237
238

        # 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
239
        self.vocab_size = vocab_size
240
        self._logprobs_tensors = logprobs_tensors
241
242
243
244
245

        # 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)
246
            self.sampled_token_ids_cpu = self._sampled_token_ids.to(
247
248
                "cpu", non_blocking=True
            )
249
250
251
252
253
            self._logprobs_tensors_cpu = (
                self._logprobs_tensors.to_cpu_nonblocking()
                if self._logprobs_tensors
                else None
            )
254
            self.async_copy_ready_event.record()
255
256
257

    def get_output(self) -> ModelRunnerOutput:
        """Copy the device tensors to the host and return a ModelRunnerOutput.
258

259
260
        This function blocks until the copy is finished.
        """
261
        max_gen_len = self.sampled_token_ids_cpu.shape[-1]
262
        self.async_copy_ready_event.synchronize()
263

264
265
        # Release the device tensors once the copy has completed.
        del self._logprobs_tensors
266
        del self._sampled_token_ids
267
        if max_gen_len == 1:
268
            valid_sampled_token_ids = self.sampled_token_ids_cpu.tolist()
269
270
            for i in self._invalid_req_indices:
                valid_sampled_token_ids[i].clear()
271
272
273
            logprobs_lists = None
            if self._logprobs_tensors_cpu is not None:
                logprobs_lists = self._logprobs_tensors_cpu.tolists()
274
        else:
275
            valid_sampled_token_ids, logprobs_lists = RejectionSampler.parse_output(
276
277
                self.sampled_token_ids_cpu,
                self.vocab_size,
278
                self._invalid_req_indices,
279
                logprobs_tensors=self._logprobs_tensors_cpu,
280
            )
281
282
283

        output = self._model_runner_output
        output.sampled_token_ids = valid_sampled_token_ids
284
        output.logprobs = logprobs_lists
285
286
287
        return output


288
289
290
291
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
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


333
334
335
336
337
338
339
340
341
342
343
344
345
346
347
348
349
350
351
352
353
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)
354
355
356
            self._model_runner_output.pooler_output = _copy_pooler_output_to_cpu(
                raw_pooler_output=self._raw_pooler_output,
                finished_mask=finished_mask,
357
358
359
360
361
362
363
364
365
366
367
368
369
370
            )
            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


371
372
373
374
375
376
377
378
379
380
381
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
382
    ec_connector_output: ECConnectorOutput | None
383
    cudagraph_stats: CUDAGraphStat | None
384
    slot_mappings: dict[str, torch.Tensor] | list[dict[str, torch.Tensor]] | None
385
386


387
388
389
class GPUModelRunner(
    LoRAModelRunnerMixin, KVConnectorModelRunnerMixin, ECConnectorModelRunnerMixin
):
390
391
    def __init__(
        self,
392
        vllm_config: VllmConfig,
393
        device: torch.device,
394
    ):
395
396
397
        self.vllm_config = vllm_config
        self.model_config = vllm_config.model_config
        self.cache_config = vllm_config.cache_config
398
        self.offload_config = vllm_config.offload_config
399
        self.compilation_config = vllm_config.compilation_config
400
401
402
403
404
405
        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
406

407
408
409
410
        model_config = self.model_config
        cache_config = self.cache_config
        scheduler_config = self.scheduler_config
        parallel_config = self.parallel_config
411
        self.device = device
412
413
        self.pin_memory = is_pin_memory_available()
        self.dtype = self.model_config.dtype
414

415
416
417
        self.kv_cache_dtype = kv_cache_dtype_str_to_dtype(
            cache_config.cache_dtype, self.model_config
        )
418

419
        self.is_pooling_model = model_config.runner_type == "pooling"
420
        self.enable_prompt_embeds = model_config.enable_prompt_embeds
421
        self.is_multimodal_raw_input_only_model = (
422
423
            model_config.is_multimodal_raw_input_only_model
        )
424
        # These will be overridden in load_model()
425
        self.is_multimodal_pruning_enabled = False
426
        self.requires_sequential_video_encoding = False
427
428
429
        # Set to True after init_routed_experts_capturer() completes.
        # Prevents routed experts code from running during profiling/dummy run.
        self.routed_experts_initialized = False
430
        self.max_model_len = model_config.max_model_len
431
432
433

        # Always set to false after the first forward pass
        self.calculate_kv_scales = self.cache_config.calculate_kv_scales
434
        self.dcp_world_size = self.parallel_config.decode_context_parallel_size
435
        self.dcp_rank = 0 if self.dcp_world_size <= 1 else get_dcp_group().rank_in_group
436
        self.max_num_tokens = scheduler_config.max_num_batched_tokens
437
        self.max_num_reqs = scheduler_config.max_num_seqs
438

439
440
        # Broadcast PP output for external_launcher (torchrun)
        # to make sure we are synced across pp ranks
441
        # TODO: Support overlapping micro-batches
442
443
        # https://github.com/vllm-project/vllm/issues/18019
        self.broadcast_pp_output = (
444
            self.parallel_config.distributed_executor_backend == "external_launcher"
445
            and len(get_pp_group().ranks) > 1
446
        )
447

448
        # Model-related.
449
        self.num_query_heads = model_config.get_num_attention_heads(parallel_config)
450
        self.inputs_embeds_size = model_config.get_inputs_embeds_size()
451
        self.attention_chunk_size = model_config.attention_chunk_size
452
        # Only relevant for models using ALiBi (e.g, MPT)
453
        self.use_alibi = model_config.uses_alibi
454

455
        self.cascade_attn_enabled = not self.model_config.disable_cascade_attn
456
        self.is_mm_prefix_lm = self.model_config.is_mm_prefix_lm
457

458
        # Multi-modal data support
459
        self.mm_registry = MULTIMODAL_REGISTRY
460
        self.uses_mrope = model_config.uses_mrope
461
        self.uses_xdrope_dim = model_config.uses_xdrope_dim
462
        self.supports_mm_inputs = self.mm_registry.supports_multimodal_inputs(
463
            model_config
464
        )
465

466
467
468
        if self.model_config.is_encoder_decoder:
            # Maximum length of the encoder input, only for encoder-decoder
            # models.
469
            self.max_encoder_len = scheduler_config.max_num_encoder_input_tokens
470
471
472
        else:
            self.max_encoder_len = 0

473
474
475
        # Async scheduling
        self.use_async_scheduling = self.scheduler_config.async_scheduling

476
        # Sampler
477
        self.sampler = Sampler(logprobs_mode=self.model_config.logprobs_mode)
478

479
        self.eplb_state: EplbState | None = None
480
481
        # NOTE(yongji): flag to temporarily disable EPLB during scaling up/down
        self.eep_eplb_suppressed = False
482
483
484
485
486
487
        """
        State of the expert parallelism load balancer.

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

488
        # Lazy initializations
489
        # self.model: nn.Module  # Set after load_model
490
        # Initialize in initialize_kv_cache
491
        self.kv_caches: list[torch.Tensor] = []
492
493
494
        # Initialize in initialize_kv_cache_tensors
        self.cross_layers_kv_cache: torch.Tensor | None = None
        self.cross_layers_attn_backend: type[AttentionBackend] | None = None
495
496
        # indexes: [kv_cache_group_id][attn_group]
        self.attn_groups: list[list[AttentionGroup]] = []
497
498
        # self.kv_cache_config: KVCacheConfig

499
500
        # mm_hash ->  encoder_output
        self.encoder_cache: dict[str, torch.Tensor] = {}
501
        self.late_interaction_runner = LateInteractionRunner()
502

503
504
505
        # Encoder CUDA graph manager (initialized after model load if enabled)
        self.encoder_cudagraph_manager: EncoderCudaGraphManager | None = None

506
        self.use_aux_hidden_state_outputs = False
507
508
509
510
511
        # 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:
512
            self.drafter: (
513
                NgramProposer  # noqa: F823
514
                | NgramProposerGPU
515
516
517
518
                | SuffixDecodingProposer
                | EagleProposer
                | DraftModelProposer
                | MedusaProposer
519
                | ExtractHiddenStatesProposer
520
            )
521
            if self.speculative_config.method == "ngram":
522
523
                from vllm.v1.spec_decode.ngram_proposer import NgramProposer

524
                self.drafter = NgramProposer(self.vllm_config)
525
526
527
528
529
530
            elif self.speculative_config.uses_draft_model():
                self.drafter = DraftModelProposer(
                    vllm_config=self.vllm_config,
                    device=self.device,
                    runner=self,
                )
531
532
533
534
535
536
537
538
539
540
541
542
543
544
545
546
547
            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
                )
548
549
            elif self.speculative_config.method == "suffix":
                self.drafter = SuffixDecodingProposer(self.vllm_config)
550
            elif self.speculative_config.use_eagle():
551
                self.drafter = EagleProposer(self.vllm_config, self.device, self)
552
                if self.speculative_config.method == "eagle3":
553
554
555
                    self.use_aux_hidden_state_outputs = (
                        self.drafter.eagle3_use_aux_hidden_state
                    )
556
557
            elif self.speculative_config.method == "medusa":
                self.drafter = MedusaProposer(
558
                    vllm_config=self.vllm_config, device=self.device
559
                )
560
561
562
563
564
            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
565
            else:
566
567
568
569
                raise ValueError(
                    "Unknown speculative decoding method: "
                    f"{self.speculative_config.method}"
                )
570
            self.rejection_sampler = RejectionSampler(self.sampler)
571

572
573
574
        self.num_spec_tokens = 0
        if self.speculative_config:
            self.num_spec_tokens = self.speculative_config.num_speculative_tokens
575
576
577
578
579
            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
580

581
        # Request states.
582
        self.requests: dict[str, CachedRequestState] = {}
583
584
585
        # NOTE(rob): num_prompt_logprobs only includes reqs
        # that are currently in the prefill phase.
        self.num_prompt_logprobs: dict[str, int] = {}
586
        self.comm_stream = torch.cuda.Stream()
587

588
589
590
591
592
593
594
595
596
        # 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.
597
598
599
600
        logits_processors = model_config.logits_processors
        custom_logitsprocs: Sequence[str | type[LogitsProcessor]] = (
            tuple(logits_processors) if logits_processors is not None else ()
        )
601
602
603
604
605
        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]
606
607
        self.input_batch = InputBatch(
            max_num_reqs=self.max_num_reqs,
608
            # We need to use the encoder length for encoder-decoder
609
610
            # because of KV cache for cross-attention.
            max_model_len=max(self.max_model_len, self.max_encoder_len),
611
612
613
            max_num_batched_tokens=self.max_num_tokens,
            device=self.device,
            pin_memory=self.pin_memory,
614
            vocab_size=self.model_config.get_vocab_size(),
615
616
            block_sizes=[placeholder_block_size],
            kernel_block_sizes=[placeholder_block_size],
617
            is_spec_decode=bool(self.vllm_config.speculative_config),
618
            logitsprocs=build_logitsprocs(
619
620
621
                self.vllm_config,
                self.device,
                self.pin_memory,
622
                self.is_pooling_model,
623
                custom_logitsprocs,
624
            ),
625
626
627
            # We currently don't know whether a particular custom logits processor
            # uses output token ids so we set this conservatively.
            logitsprocs_need_output_token_ids=bool(custom_logitsprocs),
628
            is_pooling_model=self.is_pooling_model,
629
            cp_kv_cache_interleave_size=self.parallel_config.cp_kv_cache_interleave_size,
630
        )
631

632
633
634
635
636
        # 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.
637
        self.prepare_inputs_event: torch.Event | None = None
638
639
        if self.use_async_scheduling:
            self.async_output_copy_stream = torch.cuda.Stream()
640
            self.prepare_inputs_event = torch.Event()
641

642
643
644
645
646
647
648
649
650
651
652
        # 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 = []

653
        # Cache the device properties.
654
        self._init_device_properties()
655

656
657
658
659
        # Encoder timing registry for observability
        self.encoder_timing_registry: dict[str, EncoderTimingStats] = {}
        self._encoder_timing_lock = threading.Lock()

660
        # Persistent buffers for CUDA graphs.
661
662
663
664
665
        self.input_ids = self._make_buffer(self.max_num_tokens, dtype=torch.int32)
        self.positions = self._make_buffer(self.max_num_tokens, dtype=torch.int64)
        self.query_start_loc = self._make_buffer(
            self.max_num_reqs + 1, dtype=torch.int32
        )
666
        self.seq_lens = self._make_buffer(self.max_num_reqs, dtype=torch.int32)
667
        self.encoder_seq_lens = self._make_buffer(self.max_num_reqs, dtype=torch.int32)
668
669
670
671
        if self.dcp_world_size > 1:
            self.dcp_local_seq_lens = self._make_buffer(
                self.max_num_reqs, dtype=torch.int32
            )
672
673
674
        # 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.
675
        self.inputs_embeds = self._make_buffer(
676
            self.max_num_tokens, self.inputs_embeds_size, dtype=self.dtype, numpy=False
677
678
        )
        self.is_token_ids = self._make_buffer(self.max_num_tokens, dtype=torch.bool)
679
680
        self.discard_request_mask = self._make_buffer(
            self.max_num_reqs, dtype=torch.bool
681
682
683
684
685
686
687
        )
        self.num_decode_draft_tokens = self._make_buffer(
            self.max_num_reqs, dtype=torch.int32
        )
        self.num_accepted_tokens = self._make_buffer(
            self.max_num_reqs, dtype=torch.int64
        )
688

689
690
        # Only relevant for multimodal models
        if self.supports_mm_inputs:
691
692
693
694
695
696
697
            # 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
698

699
        # Only relevant for models using M-RoPE (e.g, Qwen2-VL)
700
        if self.uses_mrope:
Roger Wang's avatar
Roger Wang committed
701
702
703
704
            # 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
705
706
707
708
709
710

            # 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
711
            self.mrope_positions = self._make_buffer(
712
713
                (3, self.max_num_tokens + 1), dtype=torch.int64
            )
714

715
716
717
718
719
720
721
        # 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
            )

722
        # None in the first PP rank. The rest are set after load_model.
723
        self.intermediate_tensors: IntermediateTensors | None = None
724

725
        # OPTIMIZATION: Cache the tensors rather than creating them every step.
726
        # Keep in int64 to avoid overflow with long context
727
728
729
730
        self.arange_np = np.arange(
            max(self.max_num_reqs + 1, self.max_model_len, self.max_num_tokens),
            dtype=np.int64,
        )
731

732
733
734
735
736
        # 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] = {}
737
738
739
740
741
        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(
742
743
                self.max_num_tokens, dtype=torch.int32, device=self.device
            )
744

745
        self.uniform_decode_query_len = 1 + self.num_spec_tokens
746
747
748
749

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

750
        self.mm_budget = (
751
            MultiModalBudget(self.vllm_config, self.mm_registry)
752
753
754
            if self.supports_mm_inputs
            else None
        )
755

756
        self.reorder_batch_threshold: int | None = None
757

758
759
760
761
762
        # 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()

763
        # Cached outputs.
764
        self._draft_token_ids: list[list[int]] | torch.Tensor | None = None
765
766
767
768
769
770
771
772
773
774
775
776
777
778
779
        # 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()

780
        self._draft_token_req_ids: list[str] | None = None
781
        self.transfer_event = torch.Event()
782
        self.sampled_token_ids_pinned_cpu = torch.empty(
783
            (self.max_num_reqs, 1),
784
785
            dtype=torch.int64,
            device="cpu",
786
787
            pin_memory=self.pin_memory,
        )
788

789
790
        # Pre-allocated tensor for copying valid sampled token counts to CPU,
        # with dedicated stream for overlapping and event for coordination.
791
        self.valid_sampled_token_count_event: torch.Event | None = None
792
        self.valid_sampled_token_count_copy_stream: torch.cuda.Stream | None = None
793
794
795
796
797
798
        # 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
799
        self.num_accepted_tokens_event: torch.Event | None = None
800
801
        if self.num_spec_tokens:
            self.draft_token_ids_event = torch.Event()
802
            self.num_accepted_tokens_event = torch.Event()
803
804
805
806
807
808
809
810
811
812
813
814
815
816
817
818
            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,
                    dtype=torch.int64,
                    device="cpu",
                    pin_memory=self.pin_memory,
                )
819

820
821
822
823
        # Model weight offloader
        # Make sure this is called before any get_offloader call
        set_offloader(create_offloader(self.offload_config))

824
825
        # Ephemeral state transferred between execute_model() and sample_tokens().
        self.execute_model_state: ExecuteModelState | None = None
826
        self.kv_connector_output: KVConnectorOutput | None = None
827
        self.mamba_state_idx: dict[str, int] = {}
828
        self._mamba_copy_bufs: mamba_utils.MambaCopyBuffers | None = None
829
        self.layerwise_nvtx_hooks_registered = False
830

831
832
833
834
835
836
837
    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

838
    def reset_mm_cache(self) -> None:
839
840
841
842
        """
        Clear the multi-modal cache that was used during profiling,
        but no longer needed during inference.
        """
843
844
        if self.mm_budget:
            self.mm_budget.reset_cache()
845
        self.late_interaction_runner.clear()
846

847
848
849
850
851
852
853
    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()
854
        self.late_interaction_runner.clear()
855

856
857
858
859
860
861
862
863
864
865
866
867
868
869
870
871
872
873
874
875
876
877
878
879
880
881
882
883
884
885
886
887
888
889
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.
        """
        if not self.cache_config.cache_dtype.startswith("fp8"):
            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)

900
901
902
903
    def _get_positions(self, num_tokens: Any):
        if isinstance(num_tokens, int):
            if self.uses_mrope:
                return self.mrope_positions.gpu[:, :num_tokens]
904
905
            if self.uses_xdrope_dim > 0:
                return self.xdrope_positions.gpu[:, :num_tokens]
906
907
908
909
            return self.positions.gpu[:num_tokens]
        else:
            if self.uses_mrope:
                return self.mrope_positions.gpu[:, num_tokens]
910
911
            if self.uses_xdrope_dim > 0:
                return self.xdrope_positions.gpu[:, num_tokens]
912
913
            return self.positions.gpu[num_tokens]

914
    def _make_buffer(
915
        self, *size: int | torch.SymInt, dtype: torch.dtype, numpy: bool = True
916
917
918
919
920
921
922
923
    ) -> CpuGpuBuffer:
        return CpuGpuBuffer(
            *size,
            dtype=dtype,
            device=self.device,
            pin_memory=self.pin_memory,
            with_numpy=numpy,
        )
924

925
926
927
928
929
930
931
932
933
934
    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

935
    def _init_model_kwargs(self):
936
937
        model_kwargs = dict[str, Any]()

938
        if not self.is_pooling_model:
939
940
            return model_kwargs

941
942
        num_reqs = self.input_batch.num_reqs
        pooling_params = self.input_batch.get_pooling_params()
943
944
945

        token_type_id_requests = dict[int, Any]()
        for i, param in enumerate(pooling_params):
946
947
948
949
950
            if (
                param.extra_kwargs is not None
                and (token_types := param.extra_kwargs.get("compressed_token_type_ids"))
                is not None
            ):
951
952
953
954
955
                token_type_id_requests[i] = token_types

        if len(token_type_id_requests) == 0:
            return model_kwargs

956
        seq_lens = self.seq_lens.gpu[:num_reqs]
957
958
959
960
961
962
963
964
        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(
965
966
            device=self.device
        )
967
968
        return model_kwargs

969
    def _may_reorder_batch(self, scheduler_output: "SchedulerOutput") -> None:
970
971
        """
        Update the order of requests in the batch based on the attention
972
        backend's needs. For example, some attention backends (namely MLA) may
973
974
975
976
977
978
        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
979
        # Attention free models have zero kv_cache_groups, however models
980
981
982
983
984
985
986
        # 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

987
988
989
990
        if self.reorder_batch_threshold is not None:
            reorder_batch_to_split_decodes_and_prefills(
                self.input_batch,
                scheduler_output,
991
992
                decode_threshold=self.reorder_batch_threshold,
            )
993

994
995
996
997
998
999
1000
1001
1002
1003
1004
1005
1006
1007
1008
1009
1010
1011
1012
1013
    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)

1014
1015
    # Note: used for model runner override.
    def _init_device_properties(self) -> None:
1016
        """Initialize attributes from torch.cuda.get_device_properties"""
1017
1018

        self.num_sms = num_compute_units(self.device.index)
1019
1020
1021

    # Note: used for model runner override.
    def _sync_device(self) -> None:
1022
        torch.accelerator.synchronize()
1023

1024
    def _update_states(self, scheduler_output: "SchedulerOutput") -> None:
1025
1026
1027
1028
1029
1030
        """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.

1031
1032
        The SamplingMetadata is updated and copied to the GPU if there is a
        new/resumed/paused/finished request in the batch.
1033
1034
        """
        # Remove finished requests from the cached states.
1035
1036
        for req_id in scheduler_output.finished_req_ids:
            self.requests.pop(req_id, None)
1037
            self.num_prompt_logprobs.pop(req_id, None)
1038
1039
1040
        self.late_interaction_runner.on_requests_finished(
            scheduler_output.finished_req_ids
        )
1041
1042
1043
1044
1045
1046
1047
        # 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:
1048
            self.input_batch.remove_request(req_id)
1049

1050
1051
1052
1053
1054
        # 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)

1055
        # Free the cached encoder outputs.
1056
1057
        for mm_hash in scheduler_output.free_encoder_mm_hashes:
            self.encoder_cache.pop(mm_hash, None)
1058

1059
1060
1061
1062
1063
1064
1065
        # 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()
1066
1067
1068
1069
1070
1071
1072
1073
        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)
1074
1075
1076
1077
1078
        # 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:
1079
            self.input_batch.remove_request(req_id)
1080

1081
1082
1083
1084
1085
1086
1087
        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] = []

1088
        reqs_to_add: list[CachedRequestState] = []
1089
        # Add new requests to the cached states.
1090
1091
        for new_req_data in scheduler_output.scheduled_new_reqs:
            req_id = new_req_data.req_id
1092
1093
1094
1095
1096
1097
            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

1098
            sampling_params = new_req_data.sampling_params
1099
            pooling_params = new_req_data.pooling_params
1100

1101
1102
1103
1104
            if (
                sampling_params
                and sampling_params.sampling_type == SamplingType.RANDOM_SEED
            ):
1105
1106
1107
1108
1109
                generator = torch.Generator(device=self.device)
                generator.manual_seed(sampling_params.seed)
            else:
                generator = None

1110
1111
            if self.is_pooling_model:
                assert pooling_params is not None
1112
1113
                task = pooling_params.task
                assert task is not None, "You did not set `task` in the API"
1114

1115
                model = cast(VllmModelForPooling, self.get_model())
1116
                to_update = model.pooler.get_pooling_updates(task)
1117
1118
                to_update.apply(pooling_params)

1119
            req_state = CachedRequestState(
1120
                req_id=req_id,
1121
                prompt_token_ids=new_req_data.prompt_token_ids,
1122
                prompt_embeds=new_req_data.prompt_embeds,
1123
                mm_features=new_req_data.mm_features,
1124
                sampling_params=sampling_params,
1125
                pooling_params=pooling_params,
1126
                generator=generator,
1127
1128
                block_ids=new_req_data.block_ids,
                num_computed_tokens=new_req_data.num_computed_tokens,
1129
                output_token_ids=[],
1130
                lora_request=new_req_data.lora_request,
1131
            )
1132
            self.requests[req_id] = req_state
1133
            self.late_interaction_runner.register_request(req_id, pooling_params)
1134

1135
1136
1137
1138
1139
1140
1141
            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
                )

1142
            # Only relevant for models using M-RoPE (e.g, Qwen2-VL)
1143
            if self.uses_mrope:
1144
                self._init_mrope_positions(req_state)
1145

1146
1147
1148
1149
            # Only relevant for models using XD-RoPE (e.g, HunYuan-VL)
            if self.uses_xdrope_dim > 0:
                self._init_xdrope_positions(req_state)

1150
            reqs_to_add.append(req_state)
1151
1152
1153
            # Track new requests for ngram_gpu full tensor copy
            if is_ngram_gpu:
                ngram_gpu_new_reqs.append(req_state)
1154

1155
        # Update the states of the running/resumed requests.
1156
        is_last_rank = get_pp_group().is_last_rank
1157
        req_data = scheduler_output.scheduled_cached_reqs
1158
        scheduled_spec_tokens = scheduler_output.scheduled_spec_decode_tokens
1159

1160
1161
1162
1163
1164
1165
1166
1167
1168
1169
1170
1171
1172
1173
1174
1175
        # 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,
            )

1176
1177
1178
1179
        # Wait until valid_sampled_tokens_count is copied to cpu,
        # then use it to update actual num_computed_tokens of each request.
        valid_sampled_token_count = self._get_valid_sampled_token_count()

1180
        for i, req_id in enumerate(req_data.req_ids):
1181
            req_state = self.requests[req_id]
1182
1183
            num_computed_tokens = req_data.num_computed_tokens[i]
            new_block_ids = req_data.new_block_ids[i]
1184
            resumed_from_preemption = req_id in req_data.resumed_req_ids
1185
            num_output_tokens = req_data.num_output_tokens[i]
1186
            req_index = self.input_batch.req_id_to_index.get(req_id)
1187

1188
1189
1190
1191
            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:
1192
                # first step: num_computed_tokens = 0, spec_tokens = [],
1193
                # prev_num_draft_len = 0.
Jiayi Yan's avatar
Jiayi Yan committed
1194
                # second step: num_computed_tokens = 100(prompt length),
1195
1196
1197
                # 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.
1198
                # num_computed_tokens in first step and second step doesn't contain
1199
1200
1201
                # 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.
1202
1203
1204
1205
1206
1207
1208
1209
1210
                if req_index is None:
                    req_state.prev_num_draft_len = 0
                else:
                    assert self.input_batch.prev_req_id_to_index is not None
                    prev_req_index = self.input_batch.prev_req_id_to_index[req_id]
                    num_accepted = valid_sampled_token_count[prev_req_index] - 1
                    num_rejected = req_state.prev_num_draft_len - num_accepted
                    num_computed_tokens -= num_rejected
                    req_state.output_token_ids.extend([-1] * num_accepted)
1211

1212
1213
1214
                    if is_ngram_gpu and num_accepted > 0 and req_index is not None:
                        self.input_batch.num_tokens_no_spec[req_index] += num_accepted

1215
            # Update the cached states.
1216
            req_state.num_computed_tokens = num_computed_tokens
1217
1218

            if not is_last_rank:
1219
1220
1221
1222
1223
1224
1225
1226
1227
1228
1229
1230
1231
1232
1233
1234
1235
1236
1237
1238
                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:]
                        )
1239
1240
1241
1242
1243
            elif num_output_tokens < len(req_state.output_token_ids):
                # Some output tokens were discarded due to a sync-KV-load
                # failure. Align the cached state.
                del req_state.output_token_ids[num_output_tokens:]
                if req_index is not None:
1244
1245
1246
1247
                    end_idx = (
                        self.input_batch.num_prompt_tokens[req_index]
                        + num_output_tokens
                    )
1248
                    self.input_batch.num_tokens_no_spec[req_index] = end_idx
1249

1250
            # Update the block IDs.
1251
            if not resumed_from_preemption:
1252
1253
                if new_block_ids is not None:
                    # Append the new blocks to the existing block IDs.
1254
                    for block_ids, new_ids in zip(req_state.block_ids, new_block_ids):
1255
                        block_ids.extend(new_ids)
1256
            else:
1257
                assert req_index is None
1258
                assert new_block_ids is not None
1259
1260
                # The request is resumed from preemption.
                # Replace the existing block IDs with the new ones.
1261
                req_state.block_ids = new_block_ids
1262
1263
1264
1265
1266

            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.
1267
1268
1269
1270
1271
1272
1273

                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:]

1274
                reqs_to_add.append(req_state)
1275
1276
1277
                # Track resumed requests for ngram_gpu full tensor copy
                if is_ngram_gpu:
                    ngram_gpu_new_reqs.append(req_state)
1278
1279
1280
                continue

            # Update the persistent batch.
1281
            self.input_batch.num_computed_tokens_cpu[req_index] = num_computed_tokens
1282
            if new_block_ids is not None:
1283
                self.input_batch.block_table.append_row(new_block_ids, req_index)
1284
1285
1286
1287
1288
1289
1290

            # 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)
1291
                self.input_batch.token_ids_cpu[
1292
1293
1294
                    req_index, start_token_index:end_token_index
                ] = new_token_ids
                self.input_batch.num_tokens_no_spec[req_index] = end_token_index
1295

1296
            # Add spec_token_ids to token_ids_cpu.
1297
            self.input_batch.update_req_spec_token_ids(req_state, scheduled_spec_tokens)
1298
1299
1300
1301
1302
            # 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
1303

1304
1305
        # Add the new or resumed requests to the persistent batch.
        # The smaller empty indices are filled first.
1306
1307
        for request in reqs_to_add:
            self.input_batch.add_request(request)
1308
            self.input_batch.update_req_spec_token_ids(request, scheduled_spec_tokens)
1309

1310
1311
1312
1313
1314
1315
        # 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()
1316

1317
1318
1319
1320
1321
1322
1323
1324
1325
1326
1327
1328
        # 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,
            )

1329
    def _update_states_after_model_execute(
1330
        self, output_token_ids: torch.Tensor, scheduler_output: "SchedulerOutput"
1331
    ) -> None:
1332
1333
1334
1335
1336
1337
1338
1339
        """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.
        """
1340
        if not self.speculative_config or not self.model_config.is_hybrid:
1341
1342
1343
            return

        # Find the number of accepted tokens for each sequence.
1344
1345
        num_reqs = output_token_ids.size(0)
        self.num_accepted_tokens.gpu[:num_reqs] = (
1346
1347
1348
1349
1350
            (
                torch.cat(
                    [
                        output_token_ids,
                        torch.full(
1351
                            (num_reqs, 1),
1352
1353
1354
1355
1356
1357
1358
1359
1360
1361
1362
                            -1,
                            device=output_token_ids.device,
                        ),
                    ],
                    dim=1,
                )
                == -1
            )
            .int()
            .argmax(-1)
        )
1363

1364
        if self.cache_config.mamba_cache_mode == "align":
1365
1366
1367
1368
            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
1369
1370
1371
1372
1373
1374
1375
1376
            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(),
1377
                self._get_mamba_copy_bufs(),
1378
            )
1379
1380
1381
1382
        else:
            self.input_batch.num_accepted_tokens_cpu_tensor[:num_reqs].copy_(
                self.num_accepted_tokens.gpu[:num_reqs], non_blocking=True
            )
1383
1384
            assert self.num_accepted_tokens_event is not None
            self.num_accepted_tokens_event.record()
1385

1386
1387
1388
1389
1390
1391
1392
1393
1394
1395
1396
1397
1398
1399
1400
1401
1402
1403
1404
    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
1405
        self.late_interaction_runner.register_request(req_id, req_state.pooling_params)
1406
1407
1408
1409
1410
1411
1412
1413
1414
1415
1416
1417
1418
1419
1420
        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

1421
    def _init_mrope_positions(self, req_state: CachedRequestState):
1422
1423
        model = self.get_model()
        assert supports_mrope(model), "M-RoPE support is not implemented."
1424
1425
1426
1427
        assert req_state.prompt_token_ids is not None, (
            "M-RoPE requires prompt_token_ids to be available."
        )
        mrope_model = cast(SupportsMRoPE, model)
1428
1429

        req_state.mrope_positions, req_state.mrope_position_delta = (
1430
            mrope_model.get_mrope_input_positions(
1431
                req_state.prompt_token_ids,
1432
                req_state.mm_features,
1433
            )
1434
        )
1435

1436
1437
1438
1439
1440
1441
1442
1443
1444
1445
1446
1447
1448
    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,
        )

1449
    def _extract_mm_kwargs(
1450
        self,
1451
1452
        scheduler_output: "SchedulerOutput",
    ) -> BatchedTensorInputs:
1453
        if not scheduler_output or not self.is_multimodal_raw_input_only_model:
1454
            return {}
1455

1456
        mm_kwargs = list[tuple[str, MultiModalKwargsItem]]()
1457
        for req in scheduler_output.scheduled_new_reqs:
1458
1459
            for feature in req.mm_features:
                if feature.data is not None:
1460
                    mm_kwargs.append((feature.modality, feature.data))
1461

1462
1463
        # Input all modalities at once
        mm_kwargs_combined: BatchedTensorInputs = {}
1464
        for _, _, mm_kwargs_batch in group_and_batch_mm_kwargs(
1465
1466
1467
            mm_kwargs,
            device=self.device,
            pin_memory=self.pin_memory,
1468
        ):
1469
            mm_kwargs_combined.update(mm_kwargs_batch)
1470

1471
        return mm_kwargs_combined
1472

1473
    def _dummy_mm_kwargs(self, num_seqs: int) -> BatchedTensorInputs:
1474
        if not self.is_multimodal_raw_input_only_model:
1475
            return {}
1476

1477
1478
1479
        mm_budget = self.mm_budget
        assert mm_budget is not None

1480
1481
1482
        if not mm_budget.mm_max_toks_per_item:
            return {}  # No tower modalities (embed-only mode)

1483
1484
        dummy_modality = mm_budget.get_modality_with_max_tokens()
        return self._get_mm_dummy_batch(dummy_modality, num_seqs)
1485

1486
1487
1488
    def _get_cumsum_and_arange(
        self,
        num_tokens: np.ndarray,
1489
        cumsum_dtype: np.dtype | None = None,
1490
1491
1492
1493
1494
1495
1496
1497
1498
1499
1500
1501
1502
1503
1504
1505
    ) -> tuple[np.ndarray, np.ndarray]:
        """Get the cumulative sum and batched arange of the given array.
        # E.g., [2, 5, 3] -> ([2, 7, 10], [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])
        """
        # 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]
        arange = self.arange_np[:total_num_tokens] - cumsums_offsets

        return cu_num_tokens, arange

1506
    def _prepare_input_ids(
1507
1508
1509
1510
        self,
        scheduler_output: "SchedulerOutput",
        total_num_scheduled_tokens: int,
        cu_num_tokens: np.ndarray,
1511
    ) -> None:
1512
        """Prepare the input IDs for the current batch.
1513

1514
1515
1516
1517
1518
1519
1520
        Carefully handles the `prev_sampled_token_ids` which can be cached
        from the previous engine iteration, in which case those tokens on the
        GPU need to be copied into the corresponding slots into input_ids."""

        if self.input_batch.prev_sampled_token_ids is None:
            # Normal scheduling case
            self.input_ids.copy_to_gpu(total_num_scheduled_tokens)
1521
1522
1523
            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)
1524
1525
1526
1527
1528
1529
1530
            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.
        prev_req_id_to_index = self.input_batch.prev_req_id_to_index
        assert prev_req_id_to_index is not None
1531
1532
1533
1534
        sample_flattened_indices: list[int] = []
        spec_flattened_indices: list[int] = []
        prev_common_req_indices: list[int] = []
        prev_draft_token_indices: list[int] = []
1535
1536
        indices_match = True
        max_flattened_index = -1
1537
1538
1539
        total_num_spec_tokens = 0
        scheduled_spec_tokens = scheduler_output.scheduled_spec_decode_tokens

1540
1541
1542
1543
1544
        for req_id, cur_index in self.input_batch.req_id_to_index.items():
            if (prev_index := prev_req_id_to_index.get(req_id)) is not None:
                prev_common_req_indices.append(prev_index)
                # We need to compute the flattened input_ids index of the
                # last token in each common request.
1545
1546
                draft_len = len(scheduled_spec_tokens.get(req_id, ()))
                total_num_spec_tokens += draft_len
1547
                flattened_index = cu_num_tokens[cur_index].item() - 1
1548
1549
1550
1551
1552
1553
1554
1555
1556
1557
1558
1559
1560
1561
1562
                # 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))
1563
                indices_match &= prev_index == flattened_index
1564
                max_flattened_index = max(max_flattened_index, flattened_index)
Jiayi Yan's avatar
Jiayi Yan committed
1565
        num_common_tokens = len(sample_flattened_indices)
1566
        total_without_spec = total_num_scheduled_tokens - total_num_spec_tokens
Jiayi Yan's avatar
Jiayi Yan committed
1567
        if num_common_tokens < total_without_spec:
1568
1569
1570
            # If not all requests are decodes from the last iteration,
            # We need to copy the input_ids_cpu to the GPU first.
            self.input_ids.copy_to_gpu(total_num_scheduled_tokens)
1571
1572
1573
            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
1574
        if num_common_tokens == 0:
1575
            # No requests in common with the previous iteration
1576
            # So input_ids.cpu will have all the input ids.
1577
            return
Jiayi Yan's avatar
Jiayi Yan committed
1578
        if indices_match and max_flattened_index == (num_common_tokens - 1):
1579
1580
1581
1582
            # 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
1583
1584
            self.input_ids.gpu[:num_common_tokens].copy_(
                self.input_batch.prev_sampled_token_ids[:num_common_tokens, 0],
1585
1586
                non_blocking=True,
            )
1587
            if self.enable_prompt_embeds:
Jiayi Yan's avatar
Jiayi Yan committed
1588
                self.is_token_ids.gpu[:num_common_tokens] = True
1589
            return
1590
        # Upload the index tensors asynchronously so the scatter can be non-blocking.
1591
1592
        sampled_tokens_index_tensor = torch.tensor(
            sample_flattened_indices, dtype=torch.int64, pin_memory=self.pin_memory
1593
        ).to(self.device, non_blocking=True)
1594
        prev_common_req_indices_tensor = torch.tensor(
1595
1596
            prev_common_req_indices, dtype=torch.int64, pin_memory=self.pin_memory
        ).to(self.device, non_blocking=True)
1597
1598
        self.input_ids.gpu.scatter_(
            dim=0,
1599
            index=sampled_tokens_index_tensor,
1600
            src=self.input_batch.prev_sampled_token_ids[
1601
1602
1603
                prev_common_req_indices_tensor, 0
            ],
        )
1604

1605
1606
1607
1608
1609
1610
1611
1612
1613
1614
1615
1616
1617
1618
1619
1620
1621
1622
1623
1624
1625
1626
        # 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],
        )

1627
1628
    def _get_encoder_seq_lens(
        self,
1629
        num_scheduled_tokens: dict[str, int],
1630
1631
        kv_cache_spec: KVCacheSpec,
        num_reqs: int,
1632
        for_cudagraph_capture: bool = False,
1633
    ) -> tuple[torch.Tensor | None, np.ndarray | None]:
1634
        if not isinstance(kv_cache_spec, CrossAttentionSpec):
1635
            return None, None
1636

1637
1638
        # Zero out buffer for padding requests that are not actually scheduled (CGs)
        self.encoder_seq_lens.np[:num_reqs] = 0
1639

1640
1641
        # Build encoder_seq_lens array mapping request indices to
        # encoder lengths for inputs scheduled in this batch
1642
        for req_id in num_scheduled_tokens:
1643
            req_index = self.input_batch.req_id_to_index[req_id]
1644
1645
1646
1647
1648
1649
1650
1651
1652
1653
1654
1655
            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
1656
1657
1658
1659
1660
1661
1662
1663
1664
        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
1665
1666
1667
1668

        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]
1669

1670
        return encoder_seq_lens, encoder_seq_lens_cpu
1671

1672
    def _prepare_inputs(
1673
1674
1675
        self,
        scheduler_output: "SchedulerOutput",
        num_scheduled_tokens: np.ndarray,
1676
1677
    ) -> tuple[
        torch.Tensor,
1678
        SpecDecodeMetadata | None,
1679
    ]:
1680
1681
        """
        :return: tuple[
1682
            logits_indices, spec_decode_metadata,
1683
1684
        ]
        """
1685
1686
1687
1688
1689
1690
1691
        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.
1692
        self.input_batch.block_table.commit_block_table(num_reqs)
1693
1694
1695

        # Get request indices.
        # E.g., [2, 5, 3] -> [0, 0, 1, 1, 1, 1, 1, 2, 2, 2]
1696
        req_indices = np.repeat(self.arange_np[:num_reqs], num_scheduled_tokens)
1697

1698
1699
        # cu_num_tokens: [2, 5, 3] -> [2, 7, 10]
        # arange: [0, 1, 0, 1, 2, 3, 4, 0, 1, 2]
1700
        cu_num_tokens, arange = self._get_cumsum_and_arange(num_scheduled_tokens)
1701
1702

        # Get positions.
1703
        positions_np = self.positions.np[:total_num_scheduled_tokens]
1704
1705
1706
1707
1708
        np.add(
            self.input_batch.num_computed_tokens_cpu[req_indices],
            arange,
            out=positions_np,
        )
1709

1710
1711
        # Calculate M-RoPE positions.
        # Only relevant for models using M-RoPE (e.g, Qwen2-VL)
1712
        if self.uses_mrope:
1713
1714
            self._calc_mrope_positions(scheduler_output)

1715
1716
1717
1718
1719
        # 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)

1720
1721
1722
1723
        # 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.
1724
1725
1726
        token_indices = (
            positions_np + req_indices * self.input_batch.token_ids_cpu.shape[1]
        )
1727
        token_indices_tensor = torch.from_numpy(token_indices)
1728

1729
1730
1731
        # 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.
1732
1733
1734
1735
1736
1737
        torch.index_select(
            self.input_batch.token_ids_cpu_tensor.flatten(),
            0,
            token_indices_tensor,
            out=self.input_ids.cpu[:total_num_scheduled_tokens],
        )
1738
        if self.enable_prompt_embeds:
1739
            is_token_ids = self.input_batch.is_token_ids_tensor.flatten()
1740
1741
1742
1743
            torch.index_select(
                is_token_ids,
                0,
                token_indices_tensor,
1744
1745
                out=self.is_token_ids.cpu[:total_num_scheduled_tokens],
            )
1746
1747
1748
1749
1750
1751
1752
1753
1754
1755
1756
1757
1758
1759
1760
1761
1762
1763
1764
1765
1766
1767
1768
1769
1770
1771
1772
1773
1774
1775
1776
1777
1778

        # 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:
1779
1780
1781
                    self.inputs_embeds.cpu[
                        output_idx : output_idx + actual_num_sched
                    ].copy_(req_embeds[start_pos:actual_end])
1782
1783

                output_idx += num_sched
1784

1785
1786
        self.input_batch.block_table.compute_slot_mapping(req_indices, positions_np)
        self.input_batch.block_table.commit_slot_mapping(total_num_scheduled_tokens)
1787
1788

        # Prepare the attention metadata.
1789
        self.query_start_loc.np[0] = 0
1790
        self.query_start_loc.np[1 : num_reqs + 1] = cu_num_tokens
1791
1792
        # Note: pad query_start_loc to be non-decreasing, as kernels
        # like FlashAttention requires that
1793
        self.query_start_loc.np[num_reqs + 1 :].fill(cu_num_tokens[-1])
1794
        self.query_start_loc.copy_to_gpu()
1795
        query_start_loc = self.query_start_loc.gpu[: num_reqs + 1]
1796

1797
        self.seq_lens.np[:num_reqs] = (
1798
1799
            self.input_batch.num_computed_tokens_cpu[:num_reqs] + num_scheduled_tokens
        )
1800
        # Fill unused with 0 for full cuda graph mode.
1801
1802
        self.seq_lens.np[num_reqs:].fill(0)
        self.seq_lens.copy_to_gpu()
1803

1804
        num_tokens = [self.requests[r].num_tokens for r in self.input_batch.req_ids]
1805
1806
        num_tokens_np = np.array(num_tokens, dtype=np.int32)

1807
        # Record which requests should not be sampled,
1808
        # so that we could clear the sampled tokens before returning
1809
1810
        self.discard_request_mask.np[:num_reqs] = (
            self.seq_lens.np[:num_reqs] < num_tokens_np
1811
        )
1812
        self.discard_request_mask.copy_to_gpu(num_reqs)
1813

1814
        # Copy the tensors to the GPU.
1815
1816
1817
1818
1819
        self._prepare_input_ids(
            scheduler_output,
            total_num_scheduled_tokens,
            cu_num_tokens,
        )
1820

1821
        if self.uses_mrope:
1822
            # Only relevant for models using M-RoPE (e.g, Qwen2-VL)
1823
1824
            self.mrope_positions.gpu[:, :total_num_scheduled_tokens].copy_(
                self.mrope_positions.cpu[:, :total_num_scheduled_tokens],
1825
1826
                non_blocking=True,
            )
1827
1828
1829
1830
1831
1832
        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,
            )
1833
1834
        else:
            # Common case (1D positions)
1835
            self.positions.copy_to_gpu(total_num_scheduled_tokens)
1836

1837
        use_spec_decode = len(scheduler_output.scheduled_spec_decode_tokens) > 0
1838
1839
1840
1841
1842
1843
1844
1845
        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
1846
            num_sampled_tokens = np.ones(num_reqs, dtype=np.int32)
1847
1848
1849
1850
1851
        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)
1852
1853
1854
            # 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)
1855
1856
1857
1858
            for (
                req_id,
                draft_token_ids,
            ) in scheduler_output.scheduled_spec_decode_tokens.items():
1859
1860
                req_idx = self.input_batch.req_id_to_index[req_id]
                num_draft_tokens[req_idx] = len(draft_token_ids)
1861
1862
1863
1864
1865
                if (
                    self.input_batch.num_computed_tokens_cpu[req_idx]
                    >= self.input_batch.num_prompt_tokens[req_idx]
                ):
                    num_decode_draft_tokens[req_idx] = len(draft_token_ids)
1866
            spec_decode_metadata = self._calc_spec_decode_metadata(
1867
1868
                num_draft_tokens, cu_num_tokens
            )
1869
            logits_indices = spec_decode_metadata.logits_indices
1870
            num_sampled_tokens = num_draft_tokens + 1
1871
            # For DECODE only cuda graph of some attention backends (e.g., GDN).
1872
            self.num_decode_draft_tokens.np[:num_reqs] = num_decode_draft_tokens
1873
1874
            self.num_decode_draft_tokens.np[num_reqs:].fill(-1)
            self.num_decode_draft_tokens.copy_to_gpu()
1875

1876
1877
1878
1879
1880
        # Hot-Swap lora model
        if self.lora_config:
            assert (
                np.sum(num_sampled_tokens)
                <= self.vllm_config.scheduler_config.max_num_batched_tokens
1881
            )
1882
1883
1884
1885
1886
1887
1888
1889
1890
1891
1892
            self.set_active_loras(
                self.input_batch, num_scheduled_tokens, num_sampled_tokens
            )

        return (
            logits_indices,
            spec_decode_metadata,
        )

    def _build_attention_metadata(
        self,
1893
        num_tokens: int,
1894
        num_reqs: int,
1895
1896
1897
        max_query_len: int,
        num_tokens_padded: int | None = None,
        num_reqs_padded: int | None = None,
1898
1899
1900
1901
        ubatch_slices: UBatchSlices | None = None,
        logits_indices: torch.Tensor | None = None,
        use_spec_decode: bool = False,
        for_cudagraph_capture: bool = False,
1902
        num_scheduled_tokens: dict[str, int] | None = None,
1903
        cascade_attn_prefix_lens: list[list[int]] | None = None,
1904
        slot_mappings: dict[int, torch.Tensor] | None = None,
1905
1906
1907
1908
    ) -> tuple[PerLayerAttnMetadata, CommonAttentionMetadata | None]:
        """
        :return: tuple[attn_metadata, spec_decode_common_attn_metadata]
        """
1909
1910
1911
1912
        # Attention metadata is not needed for attention free models
        if len(self.kv_cache_config.kv_cache_groups) == 0:
            return {}, None

1913
1914
        num_tokens_padded = num_tokens_padded or num_tokens
        num_reqs_padded = num_reqs_padded or num_reqs
1915
        assert num_reqs_padded is not None and num_tokens_padded is not None
1916

1917
1918
1919
        attn_metadata: PerLayerAttnMetadata = {}
        if ubatch_slices is not None:
            attn_metadata = [dict() for _ in range(len(ubatch_slices))]
1920

1921
1922
1923
1924
1925
1926
1927
1928
        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:
            max_seq_len = self.seq_lens.np[:num_reqs].max().item()

1929
        if use_spec_decode:
1930
1931
            if self.num_accepted_tokens_event is not None:
                self.num_accepted_tokens_event.synchronize()
1932
            self.num_accepted_tokens.np[:num_reqs] = (
1933
1934
                self.input_batch.num_accepted_tokens_cpu[:num_reqs]
            )
1935
1936
            self.num_accepted_tokens.np[num_reqs:].fill(1)
            self.num_accepted_tokens.copy_to_gpu()
1937

1938
        kv_cache_groups = self.kv_cache_config.kv_cache_groups
1939

1940
        def _get_block_table(kv_cache_gid: int):
1941
1942
1943
            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):
1944
                blk_table_tensor = torch.zeros(
1945
                    (num_reqs_padded, 1),
1946
                    dtype=torch.int32,
1947
1948
                    device=self.device,
                )
1949
            else:
1950
                blk_table = self.input_batch.block_table[kv_cache_gid]
1951
                blk_table_tensor = blk_table.get_device_tensor(num_reqs_padded)
1952

1953
1954
1955
            # Fill unused with -1. Needed for reshape_and_cache in full cuda
            # graph mode. `blk_table_tensor` -1 to match mamba PAD_SLOT_ID
            blk_table_tensor[num_reqs:num_reqs_padded].fill_(-1)
1956
            return blk_table_tensor
1957

1958
1959
1960
        assert slot_mappings is not None
        block_table_gid_0 = _get_block_table(0)
        slot_mapping_gid_0 = slot_mappings[0]
1961

1962
1963
1964
1965
        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()
1966
1967
1968
1969
1970
1971
1972
1973
1974
1975
1976
        # Compute is_prefilling: True if request is still in prefill phase
        # (num_computed_tokens < num_prompt_tokens). Used by mamba backends to
        # distinguish actual decodes from short extends.
        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
        ]
        is_prefilling = num_computed_tokens_cpu < num_prompt_tokens_cpu

1977
1978
1979
1980
1981
        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],
            seq_lens=self.seq_lens.gpu[:num_reqs_padded],
            _seq_lens_cpu=self.seq_lens.cpu[:num_reqs_padded],
1982
            _num_computed_tokens_cpu=num_computed_tokens_cpu,
1983
1984
1985
1986
1987
1988
1989
            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,
1990
            is_prefilling=is_prefilling,
1991
1992
1993
1994
1995
1996
1997
1998
1999
2000
2001
2002
2003
2004
2005
2006
2007
2008
2009
2010
2011
2012
2013
        )

        if self.dcp_world_size > 1:
            self.dcp_local_seq_lens.cpu[:num_reqs] = get_dcp_local_seq_lens(
                self.seq_lens.cpu[:num_reqs],
                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
            )

2014
2015
2016
2017
2018
2019
2020
2021
2022
        # 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
        ] = {}

2023
2024
2025
2026
2027
2028
2029
        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]
2030
            builder = attn_group.get_metadata_builder(ubid or 0)
2031
2032
2033
2034
            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))
2035

2036
2037
2038
2039
2040
2041
2042
            cascade_attn_prefix_len = (
                cascade_attn_prefix_lens[kv_cache_gid][attn_gid]
                if cascade_attn_prefix_lens
                else 0
            )

            extra_attn_metadata_args = {}
2043
2044
2045
            if use_spec_decode and isinstance(
                builder, (Mamba2AttentionMetadataBuilder, GDNAttentionMetadataBuilder)
            ):
2046
2047
2048
2049
2050
2051
2052
2053
2054
2055
2056
2057
                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
                )
2058
2059
2060
2061
2062
2063
2064
2065
2066
            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,
                )
2067
2068
2069
2070
2071
2072
            else:
                attn_metadata_i = builder.build(
                    common_prefix_len=cascade_attn_prefix_len,
                    common_attn_metadata=common_attn_metadata,
                    **extra_attn_metadata_args,
                )
2073
2074
                if builder.supports_update_block_table:
                    cached_attn_metadata[cache_key] = attn_metadata_i
2075
2076
2077
2078
2079
2080
2081
2082
2083
2084
2085
2086
2087
2088
2089
2090
2091
2092
2093
2094
2095
2096
2097

            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,
2098
                for_cudagraph_capture=for_cudagraph_capture,
2099
            )
2100
            if kv_cache_gid > 0:
2101
2102
                cm.block_table_tensor = _get_block_table(kv_cache_gid)
                cm.slot_mapping = slot_mappings[kv_cache_gid]
2103

2104
            if self.speculative_config and spec_decode_common_attn_metadata is None:
2105
                if isinstance(self.drafter, EagleProposer):
2106
                    if self.drafter.kv_cache_gid == kv_cache_gid:
2107
                        spec_decode_common_attn_metadata = cm
2108
                else:
2109
                    spec_decode_common_attn_metadata = cm
2110

2111
            for attn_gid in range(len(self.attn_groups[kv_cache_gid])):
2112
                if ubatch_slices is not None:
2113
2114
2115
                    for ubid, _cm in enumerate(split_attn_metadata(ubatch_slices, cm)):
                        _build_attn_group_metadata(kv_cache_gid, attn_gid, _cm, ubid)

2116
                else:
2117
                    _build_attn_group_metadata(kv_cache_gid, attn_gid, cm)
2118

2119
2120
2121
2122
2123
2124
2125
2126
2127
2128
2129
2130
2131
2132
2133
2134
2135
2136
2137
2138
        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]

2139
2140
2141
2142
2143
2144
2145
2146
2147
2148
        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)
            )

2149
        return attn_metadata, spec_decode_common_attn_metadata
2150

2151
2152
2153
    def _compute_cascade_attn_prefix_lens(
        self,
        num_scheduled_tokens: np.ndarray,
2154
        num_computed_tokens: np.ndarray,
2155
2156
2157
2158
2159
2160
2161
        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
        """
2162

2163
2164
2165
2166
2167
2168
2169
2170
2171
2172
2173
2174
2175
2176
        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,
2177
                        num_computed_tokens,
2178
2179
2180
2181
2182
2183
2184
2185
                        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
2186

2187
2188
2189
    def _compute_cascade_attn_prefix_len(
        self,
        num_scheduled_tokens: np.ndarray,
2190
        num_computed_tokens: np.ndarray,
2191
        num_common_prefix_blocks: int,
2192
2193
        kv_cache_spec: KVCacheSpec,
        attn_metadata_builder: AttentionMetadataBuilder,
2194
2195
2196
2197
2198
2199
2200
2201
2202
2203
2204
2205
2206
2207
2208
2209
2210
2211
    ) -> 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.
        """
2212

2213
        common_prefix_len = num_common_prefix_blocks * kv_cache_spec.block_size
2214
2215
2216
2217
2218
2219
2220
2221
2222
2223
2224
2225
2226
2227
2228
2229
2230
2231
2232
2233
2234
2235
2236
2237
2238
2239
2240
2241
2242
2243
2244
2245
2246
2247
2248
2249
2250
        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]
2251
        # Request 3's num_computed_tokens: 3 (i.e., [A, B, C])
2252
2253
2254
2255
2256
        # 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.
2257
        common_prefix_len = min(common_prefix_len, num_computed_tokens.min())
2258
        # common_prefix_len should be a multiple of the block size.
2259
2260
2261
2262
2263
2264
2265
2266
2267
2268
2269
        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
        )
2270
2271
        assert isinstance(kv_cache_spec, AttentionSpec)
        use_cascade = attn_metadata_builder.use_cascade_attention(
2272
2273
2274
            common_prefix_len=common_prefix_len,
            query_lens=num_scheduled_tokens,
            num_query_heads=self.num_query_heads,
2275
            num_kv_heads=kv_cache_spec.num_kv_heads,
2276
            use_alibi=self.use_alibi,
2277
            use_sliding_window=use_sliding_window,
2278
            use_local_attention=use_local_attention,
2279
            num_sms=self.num_sms,
2280
            dcp_world_size=self.dcp_world_size,
2281
2282
2283
        )
        return common_prefix_len if use_cascade else 0

2284
2285
    def _calc_mrope_positions(self, scheduler_output: "SchedulerOutput"):
        mrope_pos_ptr = 0
2286
        for index, req_id in enumerate(self.input_batch.req_ids):
2287
2288
2289
            req = self.requests[req_id]
            assert req.mrope_positions is not None

2290
2291
            num_computed_tokens = self.input_batch.num_computed_tokens_cpu[index]
            num_scheduled_tokens = scheduler_output.num_scheduled_tokens[req_id]
2292
            num_prompt_tokens = length_from_prompt_token_ids_or_embeds(
2293
2294
                req.prompt_token_ids, req.prompt_embeds
            )
2295
2296

            if num_computed_tokens + num_scheduled_tokens > num_prompt_tokens:
2297
2298
                prompt_part_len = max(0, num_prompt_tokens - num_computed_tokens)
                completion_part_len = max(0, num_scheduled_tokens - prompt_part_len)
2299
2300
2301
2302
2303
2304
2305
2306
2307
2308
2309
2310
2311
            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

2312
2313
2314
                self.mrope_positions.cpu[:, dst_start:dst_end] = req.mrope_positions[
                    :, src_start:src_end
                ]
2315
2316
2317
2318
2319
2320
2321
                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

2322
                assert req.mrope_position_delta is not None
2323
                MRotaryEmbedding.get_next_input_positions_tensor(
2324
                    out=self.mrope_positions.np,
2325
2326
2327
2328
2329
                    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,
                )
2330
2331
2332

                mrope_pos_ptr += completion_part_len

2333
2334
2335
2336
2337
2338
2339
2340
2341
2342
2343
2344
2345
2346
2347
2348
2349
2350
2351
2352
2353
2354
2355
2356
2357
2358
2359
2360
2361
2362
2363
2364
2365
2366
2367
2368
2369
2370
2371
2372
2373
2374
2375
2376
2377
2378
2379
    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

2380
2381
    def _calc_spec_decode_metadata(
        self,
2382
2383
2384
2385
2386
2387
2388
2389
2390
2391
2392
2393
2394
2395
2396
2397
        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
2398
2399
2400
2401

        # Step 1. cu_num_sampled_tokens: [4, 5, 8, 9, 11]
        # arange: [0, 1, 2, 3, 0, 0, 1, 2, 0, 0, 1]
        cu_num_sampled_tokens, arange = self._get_cumsum_and_arange(
2402
2403
            num_sampled_tokens, cumsum_dtype=np.int32
        )
2404
        # Step 2. [0, 0, 0, 0, 103, 104, 104, 104, 206, 207, 207]
2405
        logits_indices = np.repeat(
2406
2407
            cu_num_scheduled_tokens - num_sampled_tokens, num_sampled_tokens
        )
2408
        # Step 3. [0, 1, 2, 3, 103, 104, 105, 106, 206, 207, 208]
2409
2410
2411
2412
2413
2414
        logits_indices += arange

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

        # Compute the draft logits indices.
2415
2416
2417
        # cu_num_draft_tokens: [3, 3, 5, 5, 6]
        # arange: [0, 1, 2, 0, 1, 0]
        cu_num_draft_tokens, arange = self._get_cumsum_and_arange(
2418
2419
            num_draft_tokens, cumsum_dtype=np.int32
        )
2420
2421
        # [0, 0, 0, 5, 5, 9]
        target_logits_indices = np.repeat(
2422
2423
            cu_num_sampled_tokens - num_sampled_tokens, num_draft_tokens
        )
2424
2425
2426
2427
2428
        # [0, 1, 2, 5, 6, 9]
        target_logits_indices += arange

        # TODO: Optimize the CPU -> GPU copy.
        cu_num_draft_tokens = torch.from_numpy(cu_num_draft_tokens).to(
2429
2430
            self.device, non_blocking=True
        )
2431
2432
2433
        cu_num_sampled_tokens = torch.from_numpy(cu_num_sampled_tokens).to(
            self.device, non_blocking=True
        )
2434
2435
2436
        logits_indices = torch.from_numpy(logits_indices).to(
            self.device, non_blocking=True
        )
2437
        target_logits_indices = torch.from_numpy(target_logits_indices).to(
2438
2439
            self.device, non_blocking=True
        )
2440
        bonus_logits_indices = torch.from_numpy(bonus_logits_indices).to(
2441
2442
            self.device, non_blocking=True
        )
2443

2444
2445
        # Compute the draft token ids.
        # draft_token_indices:      [  1,   2,   3, 105, 106, 208]
2446
        draft_token_ids = self.input_ids.gpu[logits_indices]
2447
2448
        draft_token_ids = draft_token_ids[target_logits_indices + 1]

2449
        return SpecDecodeMetadata(
2450
2451
2452
            draft_token_ids=draft_token_ids,
            num_draft_tokens=num_draft_tokens.tolist(),
            cu_num_draft_tokens=cu_num_draft_tokens,
2453
            cu_num_sampled_tokens=cu_num_sampled_tokens,
2454
2455
2456
2457
2458
            target_logits_indices=target_logits_indices,
            bonus_logits_indices=bonus_logits_indices,
            logits_indices=logits_indices,
        )

2459
2460
2461
2462
2463
2464
2465
    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
2466
        self.kv_sharing_fast_prefill_logits_indices[:num_logits].copy_(logits_indices)
2467
2468
2469
2470
2471
        # 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_(
2472
2473
            logits_indices[-1].item()
        )
2474
2475
        # Dispatch for the decoder portion of the model.
        _, batch_desc = self.cudagraph_dispatcher.dispatch(
2476
            num_logits, invalid_modes={CUDAGraphMode.FULL}
2477
2478
        )
        num_logits_padded = batch_desc.num_tokens
2479
2480
2481
        logits_indices_padded = self.kv_sharing_fast_prefill_logits_indices[
            :num_logits_padded
        ]
2482
2483
        return logits_indices_padded

2484
    def _batch_mm_inputs_from_scheduler(
2485
2486
        self,
        scheduler_output: "SchedulerOutput",
2487
2488
    ) -> tuple[
        list[str],
2489
        list[tuple[str, MultiModalKwargsItem]],
2490
2491
        list[tuple[str, PlaceholderRange]],
    ]:
2492
        """Batch multimodal inputs from scheduled encoder inputs.
2493
2494
2495

        Args:
            scheduler_output: The scheduler output containing scheduled encoder
2496
                inputs.
2497
2498

        Returns:
2499
            A tuple of (mm_hashes, mm_kwargs, mm_lora_refs) where:
2500
2501
            - mm_hashes: List of multimodal hashes for each item
            - mm_kwargs: List of multimodal kwargs for each item
2502
            - mm_lora_refs: List of (req_id, placeholder_range) for each item
2503
        """
2504
2505
        scheduled_encoder_inputs = scheduler_output.scheduled_encoder_inputs
        if not scheduled_encoder_inputs:
2506
            return [], [], []
2507
2508

        mm_hashes = list[str]()
2509
        mm_kwargs = list[tuple[str, MultiModalKwargsItem]]()
2510
2511
2512
        # Multimodal LoRA reference info to map each multimodal item
        # back to its request & position
        mm_lora_refs = list[tuple[str, PlaceholderRange]]()
2513
2514
        for req_id, encoder_input_ids in scheduled_encoder_inputs.items():
            req_state = self.requests[req_id]
2515
2516

            for mm_input_id in encoder_input_ids:
2517
                mm_feature = req_state.mm_features[mm_input_id]
2518
2519
                if mm_feature.data is None:
                    continue
2520
2521

                mm_hashes.append(mm_feature.identifier)
2522
                mm_kwargs.append((mm_feature.modality, mm_feature.data))
2523
                mm_lora_refs.append((req_id, mm_feature.mm_position))
2524

2525
        return mm_hashes, mm_kwargs, mm_lora_refs
2526

2527
2528
2529
    def _execute_mm_encoder(
        self, scheduler_output: "SchedulerOutput"
    ) -> list[torch.Tensor]:
2530
2531
2532
        mm_hashes, mm_kwargs, mm_lora_refs = self._batch_mm_inputs_from_scheduler(
            scheduler_output
        )
2533
2534

        if not mm_kwargs:
2535
            return []
2536

2537
2538
2539
2540
2541
2542
        should_time = bool(
            self.observability_config
            and self.observability_config.enable_mm_processor_stats
            and scheduler_output.scheduled_encoder_inputs
        )

2543
2544
2545
2546
2547
2548
2549
        # 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.
2550
        model = cast(SupportsMultiModal, self.model)
2551
2552
2553
2554
2555
2556
2557
2558
2559
2560
2561
2562
2563
2564
2565

        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]
2566
                    pos_info.get_num_embeds()
2567
2568
2569
2570
2571
2572
2573
2574
2575
2576
2577
2578
2579
2580
2581
2582
2583
2584
2585
2586
2587
2588
2589
2590
2591
2592
2593
2594
2595
2596
2597
2598
2599
2600
2601
2602
2603
2604
2605
2606
2607
                )
                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,
                )

2608
        encoder_outputs: list[torch.Tensor] = []
2609
2610
        # Track the current index in mm_kwargs/mm_lora_refs to map groups to request IDs
        current_item_idx = 0
2611
        for modality, num_items, mm_kwargs_batch in group_and_batch_mm_kwargs(
2612
2613
2614
            mm_kwargs,
            device=self.device,
            pin_memory=self.pin_memory,
2615
        ):
2616
            batch_outputs: MultiModalEmbeddings
2617

2618
            # EVS and dynamic res video related change.
2619
            # (ekhvedchenia): Temporary hack to limit peak memory usage when
2620
            # processing multimodal data. This solves the issue with scheduler
2621
2622
2623
2624
            # 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)
2625
2626
2627
            # dynamic res video for nemotron temporarily uses this hack via
            # requires_sequential_video_encoding
            # because it doesn't yet support video batching.
2628
2629
2630
            # TODO(ywang96): Fix memory profiling to take EVS into account and
            # remove this hack.
            if (
2631
2632
2633
2634
                (
                    self.is_multimodal_pruning_enabled
                    or self.requires_sequential_video_encoding
                )
2635
2636
2637
                and modality == "video"
                and num_items > 1
            ):
2638
                batch_outputs_lst = list[torch.Tensor]()
2639
2640
2641
2642
2643
2644
                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(
2645
                            group_and_batch_mm_kwargs(
2646
2647
2648
2649
                                [video_mm_kwargs_item],
                                device=self.device,
                                pin_memory=self.pin_memory,
                            )
2650
                        )
2651

2652
2653
2654
                        micro_batch_outputs = model.embed_multimodal(
                            **micro_batch_mm_inputs
                        )
2655

2656
                        batch_outputs_lst.extend(micro_batch_outputs)
2657

2658
                batch_outputs = batch_outputs_lst
2659
2660
            else:
                # Run the encoder.
2661
                # `batch_outputs` is either of the following:
2662
2663
2664
2665
2666
                # 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.
2667
2668
2669
2670

                with self.timed_encoder_operation(
                    should_time, mm_lora_refs, current_item_idx, num_items
                ):
2671
2672
2673
2674
2675
2676
2677
2678
2679
2680
2681
2682
2683
                    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)
2684

2685
2686
            sanity_check_mm_encoder_outputs(batch_outputs, expected_num_items=num_items)
            encoder_outputs.extend(batch_outputs)
2687

2688
2689
            current_item_idx += num_items

2690
        # Cache the encoder outputs by mm_hash
2691
        for mm_hash, output in zip(mm_hashes, encoder_outputs):
2692
            self.encoder_cache[mm_hash] = output
2693
2694
            logger.debug("Finish execute for mm hash %s", mm_hash)
            self.maybe_save_ec_to_connector(self.encoder_cache, mm_hash)
2695

2696
2697
        return encoder_outputs

2698
    def _gather_mm_embeddings(
2699
2700
        self,
        scheduler_output: "SchedulerOutput",
2701
        shift_computed_tokens: int = 0,
2702
2703
2704
    ) -> tuple[list[torch.Tensor], torch.Tensor]:
        total_num_scheduled_tokens = scheduler_output.total_num_scheduled_tokens

2705
2706
2707
2708
2709
        # 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]

2710
        mm_embeds = list[torch.Tensor]()
2711
        is_mm_embed = is_mm_embed_buf.cpu
2712
2713
2714
        is_mm_embed[:total_num_scheduled_tokens] = False

        req_start_idx = 0
2715
        should_sync_mrope_positions = False
2716
        should_sync_xdrope_positions = False
2717

2718
        for req_id in self.input_batch.req_ids:
2719
2720
            mm_embeds_req: list[torch.Tensor] = []

2721
            num_scheduled_tokens = scheduler_output.num_scheduled_tokens[req_id]
2722
            req_state = self.requests[req_id]
2723
            num_computed_tokens = req_state.num_computed_tokens + shift_computed_tokens
2724

2725
2726
            for mm_feature in req_state.mm_features:
                pos_info = mm_feature.mm_position
2727
2728
                start_pos = pos_info.offset
                num_encoder_tokens = pos_info.length
2729
2730
2731
2732
2733
2734
2735
2736
2737
2738
2739
2740
2741
2742
2743
2744

                # 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,
2745
2746
                    num_encoder_tokens,
                )
2747
                assert start_idx < end_idx
2748
2749
2750
2751
2752
2753
2754
                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
2755

2756
                mm_hash = mm_feature.identifier
2757
                encoder_output = self.encoder_cache.get(mm_hash, None)
2758
                assert encoder_output is not None, f"Encoder cache miss for {mm_hash}."
2759
2760
2761

                if (is_embed := pos_info.is_embed) is not None:
                    is_embed = is_embed[start_idx:end_idx]
2762
2763
2764
                    mm_embeds_item = encoder_output[curr_embeds_start:curr_embeds_end]
                else:
                    mm_embeds_item = encoder_output[start_idx:end_idx]
2765

2766
                req_start_pos = req_start_idx + start_pos - num_computed_tokens
2767
2768
2769
2770
2771
2772
2773
2774
2775
                # 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
2776
2777
2778
                mm_embeds_req.append(mm_embeds_item)

            if self.is_multimodal_pruning_enabled and self.uses_mrope:
2779
                assert req_state.mrope_positions is not None
2780
2781
2782
2783
2784
2785
2786
                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,
2787
2788
                    )
                )
2789
2790
2791
2792
                req_state.mrope_positions.copy_(new_mrope_positions)
                req_state.mrope_position_delta = new_delta

            mm_embeds.extend(mm_embeds_req)
2793
2794
            req_start_idx += num_scheduled_tokens

2795
        is_mm_embed = is_mm_embed_buf.copy_to_gpu(total_num_scheduled_tokens)
2796
2797
2798

        if should_sync_mrope_positions:
            self._calc_mrope_positions(scheduler_output)
2799
            self.mrope_positions.copy_to_gpu(total_num_scheduled_tokens)
2800

2801
2802
2803
2804
        if should_sync_xdrope_positions:
            self._calc_xdrope_positions(scheduler_output)
            self.xdrope_positions.copy_to_gpu(total_num_scheduled_tokens)

2805
        return mm_embeds, is_mm_embed
2806

2807
    def get_model(self) -> nn.Module:
2808
2809
        if not hasattr(self, "model"):
            raise ValueError("Cannot get model before model has been initialized")
2810
        if isinstance(self.model, (CUDAGraphWrapper, UBatchWrapper)):
2811
            # get raw model out of the cudagraph wrapper.
2812
            return self.model.unwrap()
2813
2814
        return self.model

2815
2816
2817
2818
2819
2820
2821
2822
2823
2824
2825
2826
2827
    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")

2828
2829
2830
        if supports_realtime(model):
            supported_tasks.append("realtime")

2831
2832
        return supported_tasks

2833
2834
2835
2836
2837
    def get_supported_pooling_tasks(self) -> list[PoolingTask]:
        model = self.get_model()
        if not is_pooling_model(model):
            return []

2838
        return list(model.pooler.get_supported_tasks())
2839

2840
2841
2842
2843
2844
2845
2846
2847
2848
2849
    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)

2850
    def sync_and_slice_intermediate_tensors(
2851
2852
        self,
        num_tokens: int,
2853
        intermediate_tensors: IntermediateTensors | None,
2854
2855
        sync_self: bool,
    ) -> IntermediateTensors:
2856
2857
2858
        assert self.intermediate_tensors is not None

        tp = self.vllm_config.parallel_config.tensor_parallel_size
2859
        is_rs = is_residual_scattered_for_sp(self.vllm_config, num_tokens)
2860
2861
2862
2863
2864
2865

        # 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():
2866
                is_scattered = k == "residual" and is_rs
2867
                copy_len = num_tokens // tp if is_scattered else num_tokens
2868
                self.intermediate_tensors[k][:copy_len].copy_(
2869
2870
2871
2872
2873
2874
2875
2876
2877
2878
2879
2880
2881
                    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:
2882
2883
2884
        """
        Step for the EPLB (Expert Parallelism Load Balancing) state.
        """
2885
        if not self.parallel_config.enable_eplb or self.eep_eplb_suppressed:
2886
2887
2888
            return

        assert self.eplb_state is not None
2889
2890
        model = self.get_model()
        assert is_mixture_of_experts(model)
2891
2892
2893
        self.eplb_state.step(
            is_dummy,
            is_profile,
2894
            log_stats=self.parallel_config.eplb_config.log_balancedness,
2895
2896
        )

2897
2898
2899
2900
2901
2902
2903
2904
2905
2906
2907
2908
2909
2910
2911
2912
2913
    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,
        )

2914
2915
2916
2917
2918
    def _pool(
        self,
        hidden_states: torch.Tensor,
        num_scheduled_tokens: int,
        num_scheduled_tokens_np: np.ndarray,
2919
2920
2921
2922
        kv_connector_output: KVConnectorOutput | None,
    ) -> ModelRunnerOutput | AsyncModelRunnerOutput:
        num_reqs = self.input_batch.num_reqs
        assert num_reqs == len(self.input_batch.pooling_params), (
2923
2924
            "Either all or none of the requests in a batch must be pooling request"
        )
2925

2926
        hidden_states = hidden_states[:num_scheduled_tokens]
2927
        seq_lens_cpu = self.seq_lens.cpu[:num_reqs]
2928

2929
        pooling_metadata = self.input_batch.get_pooling_metadata()
2930
        pooling_metadata.build_pooling_cursor(
2931
2932
2933
2934
            num_scheduled_tokens_np,
            seq_lens_cpu,
            device=hidden_states.device,
            query_start_loc_gpu=self.query_start_loc.gpu[: num_reqs + 1],
2935
        )
2936

2937
2938
        model = cast(VllmModelForPooling, self.model)
        raw_pooler_output: PoolerOutput = model.pooler(
2939
            hidden_states=hidden_states, pooling_metadata=pooling_metadata
2940
        )
2941
2942
2943
2944
2945

        finished_mask = [
            seq_len == prompt_len
            for seq_len, prompt_len in zip(seq_lens_cpu, pooling_metadata.prompt_lens)
        ]
2946
2947
2948
2949
2950
2951
        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,
        )
2952
2953
2954
2955
2956
2957
2958
2959
2960
2961
2962
2963
2964
2965
2966
2967
2968
2969
2970

        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,
            )

2971
2972
2973
        model_runner_output.pooler_output = _copy_pooler_output_to_cpu(
            raw_pooler_output=raw_pooler_output,
            finished_mask=finished_mask,
2974
        )
2975
2976
        self._sync_device()

2977
        return model_runner_output
2978

2979
    def _pad_for_sequence_parallelism(self, num_scheduled_tokens: int) -> int:
2980
2981
2982
        # Pad tokens to multiple of tensor_parallel_size when
        # enabled collective fusion for SP
        tp_size = self.vllm_config.parallel_config.tensor_parallel_size
2983
        if self.compilation_config.pass_config.enable_sp and tp_size > 1:
2984
2985
2986
            return round_up(num_scheduled_tokens, tp_size)
        return num_scheduled_tokens

Patrick von Platen's avatar
Patrick von Platen committed
2987
2988
2989
2990
2991
2992
2993
2994
2995
2996
2997
    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

2998
    def _preprocess(
2999
3000
        self,
        scheduler_output: "SchedulerOutput",
3001
        num_input_tokens: int,  # Padded
3002
        intermediate_tensors: IntermediateTensors | None = None,
3003
    ) -> tuple[
3004
3005
        torch.Tensor | None,
        torch.Tensor | None,
3006
        torch.Tensor,
3007
        IntermediateTensors | None,
3008
        dict[str, Any],
3009
        ECConnectorOutput | None,
3010
    ]:
3011
        num_scheduled_tokens = scheduler_output.total_num_scheduled_tokens
3012
        is_first_rank = get_pp_group().is_first_rank
3013
        is_encoder_decoder = self.model_config.is_encoder_decoder
3014

3015
3016
        # _prepare_inputs may reorder the batch, so we must gather multi
        # modal outputs after that to ensure the correct order
3017
3018
        ec_connector_output = None

3019
        if self.supports_mm_inputs and is_first_rank and not is_encoder_decoder:
3020
            # Run the multimodal encoder if any.
3021
3022
3023
3024
3025
3026
            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)
3027

3028
3029
3030
            # 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.
3031
            inputs_embeds_scheduled = self.model.embed_input_ids(
3032
3033
3034
                self.input_ids.gpu[:num_scheduled_tokens],
                multimodal_embeddings=mm_embeds,
                is_multimodal=is_mm_embed,
3035
            )
3036

3037
            # TODO(woosuk): Avoid the copy. Optimize.
3038
            self.inputs_embeds.gpu[:num_scheduled_tokens].copy_(inputs_embeds_scheduled)
3039

Patrick von Platen's avatar
Patrick von Platen committed
3040
            input_ids, inputs_embeds = self._prepare_mm_inputs(num_input_tokens)
3041
            model_kwargs = {
3042
                **self._init_model_kwargs(),
3043
3044
                **self._extract_mm_kwargs(scheduler_output),
            }
3045
        elif self.enable_prompt_embeds and is_first_rank:
3046
3047
3048
3049
3050
3051
3052
3053
3054
3055
3056
3057
            # 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).
3058
3059
3060
            token_ids_idx = (
                self.is_token_ids.gpu[:num_scheduled_tokens]
                .nonzero(as_tuple=False)
3061
                .squeeze(1)
3062
            )
3063
3064
3065
            # Some tokens ids may need to become embeds
            if token_ids_idx.numel() > 0:
                token_ids = self.input_ids.gpu[token_ids_idx]
3066
                tokens_to_embeds = self.model.embed_input_ids(input_ids=token_ids)
3067
3068
3069
                self.inputs_embeds.gpu[token_ids_idx] = tokens_to_embeds

            inputs_embeds = self.inputs_embeds.gpu[:num_input_tokens]
3070
            model_kwargs = self._init_model_kwargs()
3071
            input_ids = None
3072
        else:
3073
3074
3075
3076
            # 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.
3077
            input_ids = self.input_ids.gpu[:num_input_tokens]
3078
            inputs_embeds = None
3079
            model_kwargs = self._init_model_kwargs()
3080

3081
        if self.uses_mrope:
3082
            positions = self.mrope_positions.gpu[:, :num_input_tokens]
3083
3084
        elif self.uses_xdrope_dim > 0:
            positions = self.xdrope_positions.gpu[:, :num_input_tokens]
3085
        else:
3086
            positions = self.positions.gpu[:num_input_tokens]
3087

3088
        if is_first_rank:
3089
3090
            intermediate_tensors = None
        else:
3091
            assert intermediate_tensors is not None
3092
            intermediate_tensors = self.sync_and_slice_intermediate_tensors(
3093
3094
                num_input_tokens, intermediate_tensors, True
            )
3095

3096
        if is_encoder_decoder and scheduler_output.scheduled_encoder_inputs:
3097
3098
3099
3100
3101
3102
3103
            # 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})
3104

3105
3106
3107
3108
3109
3110
        return (
            input_ids,
            inputs_embeds,
            positions,
            intermediate_tensors,
            model_kwargs,
3111
            ec_connector_output,
3112
        )
3113

3114
    def _sample(
3115
        self,
3116
3117
        logits: torch.Tensor | None,
        spec_decode_metadata: SpecDecodeMetadata | None,
3118
    ) -> SamplerOutput:
3119
        # Sample the next token and get logprobs if needed.
3120
        sampling_metadata = self.input_batch.sampling_metadata
3121
3122
3123
        # 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()
3124
        if spec_decode_metadata is None:
3125
            return self.sampler(
3126
3127
3128
                logits=logits,
                sampling_metadata=sampling_metadata,
            )
3129

3130
3131
3132
3133
3134
3135
        # 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)

3136
        sampler_output = self.rejection_sampler(
3137
3138
            spec_decode_metadata,
            None,  # draft_probs
3139
            logits,
3140
3141
            sampling_metadata,
        )
3142
3143
3144
        return sampler_output

    def _bookkeeping_sync(
3145
3146
3147
        self,
        scheduler_output: "SchedulerOutput",
        sampler_output: SamplerOutput,
3148
        logits: torch.Tensor | None,
3149
3150
        hidden_states: torch.Tensor,
        num_scheduled_tokens: int,
3151
        spec_decode_metadata: SpecDecodeMetadata | None,
3152
    ) -> tuple[
3153
        dict[str, int],
3154
        LogprobsLists | None,
3155
        list[list[int]],
3156
        dict[str, LogprobsTensors | None],
3157
3158
3159
        list[str],
        dict[str, int],
        list[int],
3160
    ]:
3161
3162
3163
3164
        num_nans_in_logits = {}
        if envs.VLLM_COMPUTE_NANS_IN_LOGITS:
            num_nans_in_logits = self._get_nans_in_logits(logits)

3165
3166
3167
3168
        num_reqs = self.input_batch.num_reqs
        discard_sampled_tokens_req_indices = np.nonzero(
            self.discard_request_mask.np[:num_reqs]
        )[0]
3169
3170
3171
3172
        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)
3173

3174
3175
3176
        # 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()
3177
        req_id_to_index_output_copy = self.input_batch.req_id_to_index.copy()
3178
3179

        num_sampled_tokens = sampler_output.sampled_token_ids.shape[0]
3180
        sampled_token_ids = sampler_output.sampled_token_ids
3181
        logprobs_tensors = sampler_output.logprobs_tensors
3182
        invalid_req_indices = []
3183
        logprobs_lists = None
3184
3185
3186
3187
3188
3189
        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)
3190
3191
3192
                # 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()
3193
3194
3195

                if logprobs_tensors is not None:
                    logprobs_lists = logprobs_tensors.tolists()
3196
3197
            else:
                # Includes spec decode tokens.
3198
                valid_sampled_token_ids, logprobs_lists = RejectionSampler.parse_output(
3199
3200
                    sampled_token_ids,
                    self.input_batch.vocab_size,
3201
                    discard_sampled_tokens_req_indices,
3202
                    logprobs_tensors=logprobs_tensors,
3203
                )
3204
        else:
3205
            valid_sampled_token_ids = []
3206
            invalid_req_indices = discard_sampled_tokens_req_indices.tolist()
3207
3208
3209
3210
3211
            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.
3212
3213
3214
3215
            # 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
3216
3217
3218
3219
3220
            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
            }
3221

3222
3223
3224
3225
3226
        # 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.
3227
        req_ids = self.input_batch.req_ids
3228
3229
        for req_idx in range(num_sampled_tokens):
            if self.use_async_scheduling:
3230
                sampled_ids = [-1] if req_idx not in invalid_req_indices_set else None
3231
3232
            else:
                sampled_ids = valid_sampled_token_ids[req_idx]
3233

3234
            num_sampled_ids: int = len(sampled_ids) if sampled_ids else 0
3235

3236
            if not sampled_ids:
3237
3238
3239
                continue

            start_idx = self.input_batch.num_tokens_no_spec[req_idx]
3240
            end_idx = start_idx + num_sampled_ids
3241
3242
3243
3244
            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}"
3245
            )
3246

3247
3248
            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
3249
            self.input_batch.num_tokens_no_spec[req_idx] = end_idx
3250

3251
            req_id = req_ids[req_idx]
3252
3253
3254
            req_state = self.requests[req_id]
            req_state.output_token_ids.extend(sampled_ids)

3255
3256
3257
3258
3259
3260
        # Compute prompt logprobs if needed.
        prompt_logprobs_dict = self._get_prompt_logprobs_dict(
            hidden_states[:num_scheduled_tokens],
            scheduler_output.num_scheduled_tokens,
        )

3261
3262
3263
3264
3265
3266
3267
3268
3269
3270
        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,
        )

3271
3272
3273
3274
3275
3276
3277
3278
3279
3280
3281
3282
3283
3284
3285
    @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()

3286
3287
    def _model_forward(
        self,
3288
3289
3290
3291
        input_ids: torch.Tensor | None = None,
        positions: torch.Tensor | None = None,
        intermediate_tensors: IntermediateTensors | None = None,
        inputs_embeds: torch.Tensor | None = None,
3292
3293
3294
3295
3296
        **model_kwargs: dict[str, Any],
    ) -> Any:
        """Helper method to call the model forward pass.

        This method can be overridden by subclasses for model execution.
3297
        Motivation: We can inspect only this method versus
3298
3299
3300
3301
3302
3303
3304
3305
3306
3307
3308
3309
3310
3311
3312
3313
3314
3315
3316
3317
        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,
        )

3318
3319
3320
3321
3322
3323
3324
3325
3326
3327
3328
3329
3330
3331
3332
3333
3334
3335
3336
3337
3338
    @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
        )

3339
3340
3341
3342
3343
3344
3345
3346
3347
3348
3349
3350
3351
    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,
3352
        force_num_active_loras: int | None = None,
3353
        num_encoder_reqs: int = 0,
3354
    ) -> tuple[
3355
3356
        CUDAGraphMode,
        BatchDescriptor,
3357
        bool,
3358
3359
        torch.Tensor | None,
        CUDAGraphStat | None,
3360
    ]:
3361
3362
3363
3364
3365
3366
        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,
3367
        )
3368
3369
3370
3371
3372
        # 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
        )
3373

3374
3375
3376
3377
3378
        # 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)
3379
        )
3380
        has_lora = num_active_loras > 0 if force_has_lora is None else force_has_lora
3381

3382
        num_tokens_padded = self._pad_for_sequence_parallelism(num_tokens)
3383
3384
3385

        def dispatch_cudagraph(num_tokens, disable_full=False, valid_modes=None):
            return self.cudagraph_dispatcher.dispatch(
3386
3387
3388
                num_tokens=num_tokens,
                has_lora=has_lora,
                uniform_decode=uniform_decode,
3389
                num_active_loras=num_active_loras,
3390
3391
                valid_modes={CUDAGraphMode.NONE} if force_eager else valid_modes,
                invalid_modes={CUDAGraphMode.FULL} if disable_full else None,
3392
3393
            )

3394
        cudagraph_mode, batch_descriptor = dispatch_cudagraph(
3395
            num_tokens_padded, disable_full=use_cascade_attn or has_encoder_output
3396
        )
3397
        num_tokens_padded = batch_descriptor.num_tokens
3398
3399
3400
3401
3402
3403
3404
3405
3406
        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"
            )
3407
3408
3409

        # Extra coordination when running data-parallel since we need to coordinate
        # across ranks
3410
        should_ubatch, num_tokens_across_dp = False, None
3411
        if self.vllm_config.parallel_config.data_parallel_size > 1:
3412
3413
3414
3415
3416
3417
3418
3419
3420
3421
            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,
                )
3422
3423
            )

3424
            # Extract DP-synced values
3425
3426
3427
            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())
3428
3429
3430
                # Re-dispatch with DP padding so we have the correct batch_descriptor
                cudagraph_mode, batch_descriptor = dispatch_cudagraph(
                    num_tokens_padded,
3431
                    valid_modes={CUDAGraphMode(synced_cudagraph_mode)},
3432
                )
3433
3434
3435
3436
                # 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

3437
3438
3439
3440
3441
3442
3443
3444
3445
3446
3447
3448
        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,
3449
            should_ubatch,
3450
3451
3452
            num_tokens_across_dp,
            cudagraph_stats,
        )
3453

3454
3455
3456
3457
3458
3459
3460
3461
3462
3463
3464
3465
3466
3467
3468
3469
3470
3471
3472
3473
3474
3475
3476
3477
3478
3479
3480
3481
3482
3483
3484
3485
3486
3487
3488
3489
    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

3490
3491
3492
3493
3494
3495
3496
3497
3498
3499
3500
3501
3502
3503
3504
3505
3506
3507
3508
3509
3510
3511
3512
3513
3514
3515
3516
3517
3518
3519
3520
3521
3522
3523
3524
3525
3526
3527
3528
3529
3530
3531
3532
3533
3534
3535
3536
3537
3538
3539
3540
3541
3542
3543
3544
3545
3546
3547
3548
3549
3550
3551
3552
3553
3554
3555
3556
3557
3558
3559
3560
3561
3562
3563
    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

3564
3565
3566
3567
    @torch.inference_mode()
    def execute_model(
        self,
        scheduler_output: "SchedulerOutput",
3568
        intermediate_tensors: IntermediateTensors | None = None,
3569
    ) -> ModelRunnerOutput | AsyncModelRunnerOutput | IntermediateTensors | None:
3570
3571
3572
3573
3574
        if self.execute_model_state is not None:
            raise RuntimeError(
                "State error: sample_tokens() must be called "
                "after execute_model() returns None."
            )
3575

3576
        if self.routed_experts_initialized:
3577
3578
3579
3580
3581
3582
            capturer = RoutedExpertsCapturer.get_instance()
            if capturer is not None:
                capturer.clear_buffer()  # noqa
            else:
                logger.error("RoutedExpertsCapturer not initialized.")

3583
3584
3585
3586
3587
3588
3589
3590
3591
3592
3593
3594
3595
3596
3597
3598
3599
        # 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,
            )

3600
3601
3602
3603
        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)
3604

3605
        num_scheduled_tokens = scheduler_output.total_num_scheduled_tokens
3606
3607
3608
3609
3610
3611
        with (
            record_function_or_nullcontext("gpu_model_runner: preprocess"),
            self.synchronize_input_prep(),
        ):
            # Update persistent batch states.
            self._update_states(scheduler_output)
3612

3613
            if has_ec_transfer() and not get_ec_transfer().is_consumer:
3614
                with self.maybe_get_ec_connector_output(
3615
                    scheduler_output,
3616
3617
3618
3619
3620
3621
3622
3623
3624
3625
3626
3627
3628
3629
3630
3631
3632
3633
3634
3635
3636
3637
3638
3639
3640
3641
3642
3643
                    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"
3644
3645
                )

3646
3647
3648
3649
3650
3651
            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
3652

3653
3654
3655
3656
            logits_indices, spec_decode_metadata = self._prepare_inputs(
                scheduler_output,
                num_scheduled_tokens_np,
            )
3657

3658
3659
3660
3661
3662
            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(
3663
                    num_scheduled_tokens_np,
3664
3665
                    self.input_batch.num_computed_tokens_cpu[:num_reqs],
                    scheduler_output.num_common_prefix_blocks,
3666
3667
                )

3668
3669
3670
3671
3672
3673
3674
3675
3676
3677
3678
3679
3680
3681
            (
                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),
            )
3682

3683
3684
3685
3686
3687
3688
3689
3690
3691
3692
3693
3694
3695
3696
3697
3698
3699
3700
3701
3702
3703
3704
3705
3706
3707
3708
3709
            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,
            )

3710
3711
3712
3713
3714
3715
3716
3717
3718
3719
3720
            # 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)
            )
3721
3722
            pad_attn = cudagraph_mode == CUDAGraphMode.FULL

3723
3724
3725
3726
3727
3728
3729
3730
3731
3732
            if self.cache_config.mamba_cache_mode == "align":
                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(),
3733
                    self._get_mamba_copy_bufs(),
3734
3735
                )

3736
3737
3738
            use_spec_decode = len(scheduler_output.scheduled_spec_decode_tokens) > 0
            ubatch_slices_attn = ubatch_slices_padded if pad_attn else ubatch_slices

3739
3740
3741
3742
3743
3744
3745
3746
3747
3748
3749
            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,
            )

3750
3751
3752
3753
3754
3755
3756
3757
3758
3759
3760
3761
            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,
3762
                    slot_mappings=slot_mappings_by_group,
3763
                )
3764
            )
3765

3766
3767
3768
3769
3770
3771
3772
3773
3774
            (
                input_ids,
                inputs_embeds,
                positions,
                intermediate_tensors,
                model_kwargs,
                ec_connector_output,
            ) = self._preprocess(
                scheduler_output, num_tokens_padded, intermediate_tensors
3775
            )
3776

3777
        # Set cudagraph mode to none if calc_kv_scales is true.
3778
3779
3780
        # KV scales calculation involves dynamic operations that are incompatible
        # with CUDA graph capture.
        if self.calculate_kv_scales:
3781
            cudagraph_mode = CUDAGraphMode.NONE
3782
3783
            # Mark KV scales as calculated after the first forward pass
            self.calculate_kv_scales = False
3784

3785
3786
3787
3788
3789
3790
3791
        # 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
        )

3792
3793
        # Run the model.
        # Use persistent buffers for CUDA graphs.
3794
3795
3796
        # 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
3797
3798
        with (
            set_forward_context(
3799
3800
                attn_metadata,
                self.vllm_config,
3801
                num_tokens=num_tokens_padded,
3802
                num_tokens_across_dp=num_tokens_across_dp,
3803
3804
                cudagraph_runtime_mode=cudagraph_mode,
                batch_descriptor=batch_desc,
3805
                ubatch_slices=ubatch_slices_padded,
3806
                slot_mapping=slot_mappings,
3807
                skip_compiled=has_encoder_input,
3808
            ),
3809
            record_function_or_nullcontext("gpu_model_runner: forward"),
3810
            self.maybe_get_kv_connector_output(
3811
3812
                scheduler_output,
                defer_finalize=defer_kv_connector_finalize,
3813
            ) as kv_connector_output,
3814
        ):
3815
            model_output = self._model_forward(
3816
3817
3818
3819
3820
3821
3822
                input_ids=input_ids,
                positions=positions,
                intermediate_tensors=intermediate_tensors,
                inputs_embeds=inputs_embeds,
                **model_kwargs,
            )

3823
        with record_function_or_nullcontext("gpu_model_runner: postprocess"):
3824
            if self.use_aux_hidden_state_outputs:
3825
                # True when EAGLE 3 is used.
3826
3827
                hidden_states, aux_hidden_states = model_output
            else:
3828
                # Common case.
3829
3830
3831
                hidden_states = model_output
                aux_hidden_states = None

3832
3833
3834
3835
3836
            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)
3837
                    hidden_states.kv_connector_output = kv_connector_output
3838
                    self.kv_connector_output = kv_connector_output
3839
                    return hidden_states
3840

3841
                if self.is_pooling_model:
3842
                    # Return the pooling output.
3843
3844
3845
3846
3847
                    return self._pool(
                        hidden_states,
                        num_scheduled_tokens,
                        num_scheduled_tokens_np,
                        kv_connector_output,
3848
                    )
3849
3850

                sample_hidden_states = hidden_states[logits_indices]
3851
                logits = self.model.compute_logits(sample_hidden_states)
3852
3853
3854
3855
            else:
                # Rare case.
                assert not self.is_pooling_model

3856
                sample_hidden_states = hidden_states[logits_indices]
3857
                if not get_pp_group().is_last_rank:
3858
                    all_gather_tensors = {
3859
                        "residual": not is_residual_scattered_for_sp(
3860
                            self.vllm_config, num_tokens_padded
3861
                        )
3862
                    }
3863
                    get_pp_group().send_tensor_dict(
3864
3865
                        hidden_states.tensors,
                        all_gather_group=get_tp_group(),
3866
3867
                        all_gather_tensors=all_gather_tensors,
                    )
3868
3869
                    logits = None
                else:
3870
                    logits = self.model.compute_logits(sample_hidden_states)
3871

3872
                model_output_broadcast_data: dict[str, Any] = {}
3873
3874
3875
                if logits is not None:
                    model_output_broadcast_data["logits"] = logits.contiguous()

3876
                broadcasted = get_pp_group().broadcast_tensor_dict(
3877
3878
                    model_output_broadcast_data, src=len(get_pp_group().ranks) - 1
                )
3879
3880
                assert broadcasted is not None
                logits = broadcasted["logits"]
3881

3882
3883
3884
3885
3886
3887
3888
3889
        self.execute_model_state = ExecuteModelState(
            scheduler_output,
            logits,
            spec_decode_metadata,
            spec_decode_common_attn_metadata,
            hidden_states,
            sample_hidden_states,
            aux_hidden_states,
3890
            ec_connector_output,
3891
            cudagraph_stats,
3892
            slot_mappings,
3893
        )
3894
        self.kv_connector_output = kv_connector_output
3895
3896
3897
3898
3899
3900
3901
        return None

    @torch.inference_mode
    def sample_tokens(
        self, grammar_output: "GrammarOutput | None"
    ) -> ModelRunnerOutput | AsyncModelRunnerOutput | IntermediateTensors:
        if self.execute_model_state is None:
3902
3903
            kv_connector_output = self.kv_connector_output
            self.kv_connector_output = None
3904
3905
3906
            # 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()
3907
            if not kv_connector_output:
3908
                return None  # type: ignore[return-value]
3909
3910
3911
3912
3913
3914
3915
3916
3917

            # 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
3918
3919
3920
3921
3922
3923
3924
3925
3926
3927

        # Unpack ephemeral state.
        (
            scheduler_output,
            logits,
            spec_decode_metadata,
            spec_decode_common_attn_metadata,
            hidden_states,
            sample_hidden_states,
            aux_hidden_states,
3928
            ec_connector_output,
3929
            cudagraph_stats,
3930
            slot_mappings,
3931
3932
3933
3934
3935
3936
3937
3938
3939
        ) = 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
            )
3940

3941
        with record_function_or_nullcontext("gpu_model_runner: sample"):
3942
3943
            sampler_output = self._sample(logits, spec_decode_metadata)

3944
3945
3946
        self._update_states_after_model_execute(
            sampler_output.sampled_token_ids, scheduler_output
        )
3947
3948
        if self.use_async_scheduling:
            pp = get_pp_group()
3949
3950
3951
3952
            # 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:
3953
3954
3955
                self._pp_broadcast_prev_sampled_token_ids(
                    sampler_output.sampled_token_ids
                )
3956

3957
3958
        self._draft_token_ids = None
        self._draft_token_req_ids = None
3959
3960
        self.input_batch.prev_sampled_token_ids = None

3961
        def propose_draft_token_ids(sampled_token_ids):
3962
            assert spec_decode_common_attn_metadata is not None
3963
            with record_function_or_nullcontext("gpu_model_runner: draft"):
3964
3965
3966
3967
3968
3969
3970
3971
3972
                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,
3973
                    slot_mappings,
3974
                )
3975
                self._copy_draft_token_ids_to_cpu(scheduler_output)
3976

3977
        spec_config = self.speculative_config
3978
3979
3980
3981
3982
        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
3983
            )
3984
            use_gpu_toks = (
3985
3986
3987
                spec_config.use_eagle()
                or spec_config.uses_draft_model()
                or spec_config.uses_extract_hidden_states()
3988
3989
3990
            ) and not spec_config.disable_padded_drafter_batch
            if use_gpu_toks:
                # EAGLE/DraftModel speculative decoding can use the GPU sampled tokens
3991
                # as inputs, and does not need to wait for bookkeeping to finish.
3992
3993
3994
3995
                assert isinstance(
                    self.drafter,
                    EagleProposer | DraftModelProposer | ExtractHiddenStatesProposer,
                )
3996
3997
3998
3999
4000
4001
4002
4003
4004
4005
4006
4007
4008
                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(
                            spec_decode_common_attn_metadata,
                            sampled_token_ids,
                            self.requests,
                            self.input_batch,
                            self.discard_request_mask.gpu,
                        )
4009
                    )
4010
4011
4012
                    self._copy_valid_sampled_token_count(
                        next_token_ids, valid_sampled_tokens_count
                    )
4013
4014
4015
4016
4017
4018
4019
4020
4021
4022
4023
4024
4025
4026
4027
4028
4029
4030
4031
4032
4033
4034
4035
4036
4037
4038
                    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
                    )
4039
4040
4041
4042
4043
4044
4045
4046
                    # 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
4047

4048
        with record_function_or_nullcontext("gpu_model_runner: bookkeep"):
4049
4050
4051
4052
4053
4054
4055
4056
            (
                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,
4057
4058
4059
4060
4061
            ) = self._bookkeeping_sync(
                scheduler_output,
                sampler_output,
                logits,
                hidden_states,
4062
                scheduler_output.total_num_scheduled_tokens,
4063
                spec_decode_metadata,
4064
            )
4065

4066
        if propose_drafts_after_bookkeeping:
4067
4068
4069
            # 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)
4070

4071
4072
4073
4074
4075
        # 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()
4076

4077
        with record_function_or_nullcontext("gpu_model_runner: eplb"):
4078
            self.eplb_step()
4079

4080
4081
4082
4083
        # self.kv_connector_output may be modified during drafting
        kv_connector_output = self.kv_connector_output
        self.kv_connector_output = None

4084
        with record_function_or_nullcontext("gpu_model_runner: ModelRunnerOutput"):
4085
            if self.routed_experts_initialized:
4086
4087
4088
4089
4090
4091
                capturer = RoutedExpertsCapturer.get_instance()
                if capturer is not None:
                    capturer.save_captured_experts(indices=self.slot_mapping)  # noqa
                else:
                    logger.error("RoutedExpertsCapturer not initialized.")

4092
4093
4094
4095
4096
4097
4098
            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,
4099
4100
4101
                ec_connector_output=ec_connector_output
                if self.supports_mm_inputs
                else None,
4102
                num_nans_in_logits=num_nans_in_logits,
4103
                cudagraph_stats=cudagraph_stats,
4104
            )
4105

4106
4107
        if not self.use_async_scheduling:
            return output
4108

4109
4110
4111
4112
4113
4114
4115
4116
4117
        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,
4118
                vocab_size=self.input_batch.vocab_size,
4119
4120
4121
4122
4123
            )
        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
4124
            # any requests with sampling params that require output ids.
4125
4126
4127
4128
            self.input_batch.set_async_sampled_token_ids(
                async_output.sampled_token_ids_cpu,
                async_output.async_copy_ready_event,
            )
4129
4130
4131

        return async_output

4132
4133
4134
4135
4136
4137
4138
4139
4140
4141
4142
4143
4144
4145
4146
4147
4148
4149
4150
4151
4152
4153
4154
4155
4156
4157
4158
4159
4160
4161
4162
4163
4164
4165
4166
4167
4168
4169
4170
    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

4171
    def take_draft_token_ids(self) -> DraftTokenIds | None:
4172
        if not self.num_spec_tokens or not self._draft_token_req_ids:
4173
            return None
4174
        draft_token_ids, req_ids = self._get_draft_token_ids_cpu()
4175
        return DraftTokenIds(req_ids, draft_token_ids)
4176

4177
4178
4179
    def _copy_draft_token_ids_to_cpu(
        self, scheduler_output: "SchedulerOutput", zeros_only: bool = False
    ) -> None:
4180
4181
4182
4183
4184
4185
        # 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
        ):
4186
4187
4188
            return
        # We must also set the corresponding request ids.
        self._draft_token_req_ids = self.input_batch.req_ids.copy()
4189

4190
4191
4192
4193
4194
4195
4196
4197
4198
4199
4200
4201
4202
4203
4204
4205
4206
4207
4208
4209
        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()

4210
    def _get_draft_token_ids_cpu(self) -> tuple[list[list[int]], list[str]]:
4211
        if isinstance(self._draft_token_ids, list):
4212
4213
4214
4215
            return self._draft_token_ids, self.input_batch.req_ids
        req_ids = self._draft_token_req_ids
        if req_ids is None:
            return [], []
4216
4217
4218
        assert self.draft_token_ids_event is not None
        assert self.draft_token_ids_cpu is not None
        self.draft_token_ids_event.synchronize()
4219
        return self.draft_token_ids_cpu[: len(req_ids)].tolist(), req_ids
4220

4221
4222
4223
4224
4225
4226
4227
4228
4229
4230
4231
4232
4233
    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
4234
            assert counts_cpu is not None
4235
4236
4237
4238
4239
4240
4241
4242
            counts_cpu[: counts.shape[0]].copy_(counts, non_blocking=True)
            self.valid_sampled_token_count_event.record()

        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
4243
4244
        sampled_count_event = self.valid_sampled_token_count_event
        if sampled_count_event is None or prev_sampled_token_ids is None:
4245
4246
4247
            return []

        counts_cpu = self.valid_sampled_token_count_cpu
4248
4249
        assert counts_cpu is not None
        sampled_count_event.synchronize()
4250
4251
        return counts_cpu[: prev_sampled_token_ids.shape[0]].tolist()

4252
4253
4254
    def propose_draft_token_ids(
        self,
        scheduler_output: "SchedulerOutput",
4255
        sampled_token_ids: torch.Tensor | list[list[int]],
4256
4257
4258
        sampling_metadata: SamplingMetadata,
        hidden_states: torch.Tensor,
        sample_hidden_states: torch.Tensor,
4259
4260
        aux_hidden_states: list[torch.Tensor] | None,
        spec_decode_metadata: SpecDecodeMetadata | None,
4261
        common_attn_metadata: CommonAttentionMetadata,
4262
        slot_mappings: dict[str, torch.Tensor] | list[dict[str, torch.Tensor]] | None,
4263
    ) -> list[list[int]] | torch.Tensor:
4264
        num_scheduled_tokens = scheduler_output.total_num_scheduled_tokens
4265
4266
4267
        spec_config = self.speculative_config
        assert spec_config is not None
        if spec_config.method == "ngram":
4268
4269
            from vllm.v1.spec_decode.ngram_proposer import NgramProposer

4270
            assert isinstance(sampled_token_ids, list)
4271
            assert isinstance(self.drafter, NgramProposer)
4272
            draft_token_ids = self.drafter.propose(
4273
                sampled_token_ids,
4274
4275
                self.input_batch.num_tokens_no_spec,
                self.input_batch.token_ids_cpu,
4276
                slot_mappings=slot_mappings,
4277
            )
4278
4279
4280
4281
4282
4283
4284
4285
4286
4287
4288
4289
4290
4291
4292
4293
4294
4295
4296
4297
4298
4299
4300
4301
4302
4303
4304
4305
4306
4307
4308
4309
4310
4311
4312
4313
4314
        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,
            )
4315
        elif spec_config.method == "suffix":
4316
4317
            assert isinstance(sampled_token_ids, list)
            assert isinstance(self.drafter, SuffixDecodingProposer)
4318
4319
4320
            draft_token_ids = self.drafter.propose(
                self.input_batch, sampled_token_ids, slot_mappings=slot_mappings
            )
4321
        elif spec_config.method == "medusa":
4322
            assert isinstance(sampled_token_ids, list)
4323
            assert isinstance(self.drafter, MedusaProposer)
4324

4325
4326
            if sample_hidden_states.shape[0] == len(sampled_token_ids):
                # The input to the target model does not include draft tokens.
4327
4328
4329
4330
                hidden_states = sample_hidden_states
            else:
                indices = []
                offset = 0
4331
4332
4333
                assert spec_decode_metadata is not None, (
                    "No spec decode metadata for medusa"
                )
4334
                for num_draft, tokens in zip(
4335
4336
                    spec_decode_metadata.num_draft_tokens, sampled_token_ids
                ):
4337
                    indices.append(offset + len(tokens) - 1)
4338
                    offset += num_draft + 1
4339
                indices = torch.tensor(indices, device=self.device)
4340
4341
                hidden_states = sample_hidden_states[indices]

4342
            draft_token_ids = self.drafter.propose(
4343
4344
                target_hidden_states=hidden_states,
                sampling_metadata=sampling_metadata,
4345
                slot_mappings=slot_mappings,
4346
            )
4347
4348
4349
4350
4351
4352
4353
4354
4355
4356
4357
4358
        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]

4359
            draft_token_ids = self.drafter.propose(
4360
4361
4362
4363
4364
4365
4366
4367
4368
4369
4370
4371
4372
4373
4374
4375
4376
4377
                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(
                    common_attn_metadata,
                    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
            )

4378
4379
        elif spec_config.use_eagle() or spec_config.uses_draft_model():
            assert isinstance(self.drafter, EagleProposer | DraftModelProposer)
4380

4381
            if spec_config.disable_padded_drafter_batch:
4382
4383
4384
                # 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.
4385
4386
                assert isinstance(sampled_token_ids, list), (
                    "sampled_token_ids should be a python list when"
4387
                    "padded-batch is disabled."
4388
                )
4389
                next_token_ids = self.drafter.prepare_next_token_ids_cpu(
4390
4391
4392
4393
4394
                    sampled_token_ids,
                    self.requests,
                    self.input_batch,
                    scheduler_output.num_scheduled_tokens,
                )
4395
4396
4397
4398
4399
            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.
4400
4401
                assert isinstance(sampled_token_ids, torch.Tensor), (
                    "sampled_token_ids should be a torch.Tensor when"
4402
                    "padded-batch is enabled."
4403
4404
                )
                next_token_ids, valid_sampled_tokens_count = (
4405
4406
4407
4408
4409
                    self.drafter.prepare_next_token_ids_padded(
                        common_attn_metadata,
                        sampled_token_ids,
                        self.requests,
                        self.input_batch,
4410
                        self.discard_request_mask.gpu,
4411
                    )
4412
                )
4413
4414
4415
                self._copy_valid_sampled_token_count(
                    next_token_ids, valid_sampled_tokens_count
                )
Jiayi Yao's avatar
Jiayi Yao committed
4416

4417
            num_rejected_tokens_gpu = None
4418
            if spec_decode_metadata is None:
4419
                token_indices_to_sample = None
4420
                # input_ids can be None for multimodal models.
4421
                target_token_ids = self.input_ids.gpu[:num_scheduled_tokens]
4422
                target_positions = self._get_positions(num_scheduled_tokens)
4423
                if self.use_aux_hidden_state_outputs:
Wentao Ye's avatar
Wentao Ye committed
4424
                    assert aux_hidden_states is not None
4425
                    target_hidden_states = torch.cat(
4426
4427
                        [h[:num_scheduled_tokens] for h in aux_hidden_states], dim=-1
                    )
4428
4429
                else:
                    target_hidden_states = hidden_states[:num_scheduled_tokens]
4430
            else:
4431
                if spec_config.disable_padded_drafter_batch:
4432
                    token_indices_to_sample = None
4433
4434
4435
4436
4437
                    common_attn_metadata, token_indices = self.drafter.prepare_inputs(
                        common_attn_metadata,
                        sampled_token_ids,
                        spec_decode_metadata.num_draft_tokens,
                    )
4438
4439
4440
4441
4442
4443
4444
4445
4446
                    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]
4447
                else:
4448
4449
4450
4451
4452
4453
4454
4455
                    (
                        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,
4456
                    )
4457
4458
4459
4460
4461
4462
4463
4464
4465
4466
4467
                    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]
4468

4469
            if self.supports_mm_inputs and self.drafter.supports_mm_inputs:
4470
4471
4472
4473
4474
4475
                mm_embed_inputs = self._gather_mm_embeddings(
                    scheduler_output,
                    shift_computed_tokens=1,
                )
            else:
                mm_embed_inputs = None
4476

4477
            draft_token_ids = self.drafter.propose(
4478
4479
4480
4481
                target_token_ids=target_token_ids,
                target_positions=target_positions,
                target_hidden_states=target_hidden_states,
                next_token_ids=next_token_ids,
4482
                token_indices_to_sample=token_indices_to_sample,
4483
                sampling_metadata=sampling_metadata,
4484
                common_attn_metadata=common_attn_metadata,
4485
                mm_embed_inputs=mm_embed_inputs,
4486
                num_rejected_tokens_gpu=num_rejected_tokens_gpu,
4487
                slot_mappings=slot_mappings,
4488
            )
4489

4490
        return draft_token_ids
4491

4492
4493
4494
    def update_config(self, overrides: dict[str, Any]) -> None:
        allowed_config_names = {"load_config", "model_config"}
        for config_name, config_overrides in overrides.items():
4495
4496
            assert config_name in allowed_config_names, (
                f"Config `{config_name}` not supported. "
4497
                f"Allowed configs: {allowed_config_names}"
4498
            )
4499
4500
4501
4502
            config = getattr(self, config_name)
            new_config = update_config(config, config_overrides)
            setattr(self, config_name, new_config)

4503
    @instrument(span_name="Loading (GPU)")
4504
    def load_model(self, load_dummy_weights: bool = False) -> None:
4505
4506
        """
        Args:
4507
            load_dummy_weights: load dummy weights instead of real weights.
4508
        """
4509
4510
4511
4512
4513
        logger.info_once(
            "Starting to load model %s...",
            self.model_config.model,
            scope="global",
        )
4514

4515
4516
4517
4518
        if self.parallel_config.enable_eplb:
            self.eplb_state = EplbState(self.parallel_config, self.device)
            eplb_models = 0

4519
4520
4521
        try:
            with DeviceMemoryProfiler() as m:
                time_before_load = time.perf_counter()
4522
4523
                if load_dummy_weights:
                    self.load_config.load_format = "dummy"
4524
4525
4526
4527
4528
4529
4530
                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
4531
                    )
4532
4533
4534
4535
4536
4537
4538
4539
                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
                    ):
4540
4541
4542
                        assert not self.parallel_config.enable_elastic_ep, (
                            "Elastic EP is not supported with drafter model."
                        )
4543
4544
4545
4546
4547
4548
4549
4550
4551
4552
4553
4554
4555
4556
4557
4558
                        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
4559

4560
4561
4562
4563
4564
4565
                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"
                        )
4566

4567
4568
4569
4570
4571
4572
4573
4574
4575
                    # 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:
4576
4577
4578
                        aux_layers = (
                            self.model.get_eagle3_default_aux_hidden_state_layers()
                        )
4579
4580
4581
4582
4583
4584
4585
4586
4587
4588
4589
4590
4591
4592
4593

                    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
4594
        logger.info_once(
4595
4596
            "Model loading took %s GiB memory and %.6f seconds",
            format_gib(self.model_memory_usage),
4597
            time_after_load - time_before_load,
4598
            scope="local",
4599
        )
4600
4601
4602
4603
4604
4605
        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)
4606
        mm_config = self.model_config.multimodal_config
4607
        self.is_multimodal_pruning_enabled = (
4608
            supports_multimodal_pruning(self.get_model())
4609
4610
            and mm_config is not None
            and mm_config.is_multimodal_pruning_enabled()
4611
        )
4612
4613
4614
        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
4615

4616
4617
4618
4619
4620
        if (
            is_mixture_of_experts(self.model)
            and self.parallel_config.enable_eplb
            and not load_dummy_weights
        ):
4621
4622
4623
            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(
4624
                self.model,
4625
                self.model_config,
4626
            )
4627
            if self.eplb_state.is_async:
4628
                self.eplb_state.start_async_loop()
4629

4630
        if (
4631
4632
            self.vllm_config.compilation_config.mode
            == CompilationMode.STOCK_TORCH_COMPILE
4633
        ):
4634
            backend = self.vllm_config.compilation_config.init_backend(self.vllm_config)
4635
            compilation_counter.stock_torch_compile_count += 1
4636
            self.model.compile(fullgraph=True, backend=backend)
4637
            return
4638
        # for other compilation modes, cudagraph behavior is controlled by
Jiayi Yan's avatar
Jiayi Yan committed
4639
        # CudagraphWrapper and CudagraphDispatcher of vllm.
4640
4641

        # wrap the model with full cudagraph wrapper if needed.
4642
4643
        cudagraph_mode = self.compilation_config.cudagraph_mode
        assert cudagraph_mode is not None
4644
4645
4646
4647
        if (
            cudagraph_mode.has_full_cudagraphs()
            and not self.parallel_config.use_ubatching
        ):
4648
4649
4650
            self.model = CUDAGraphWrapper(
                self.model, self.vllm_config, runtime_mode=CUDAGraphMode.FULL
            )
4651
        elif self.parallel_config.use_ubatching:
4652
            if cudagraph_mode.has_full_cudagraphs():
4653
4654
4655
                self.model = UBatchWrapper(
                    self.model, self.vllm_config, CUDAGraphMode.FULL, self.device
                )
4656
            else:
4657
4658
4659
                self.model = UBatchWrapper(
                    self.model, self.vllm_config, CUDAGraphMode.NONE, self.device
                )
4660

4661
4662
        get_offloader().post_init()

4663
    def _get_eagle3_aux_layers_from_config(self) -> tuple[int, ...] | None:
4664
4665
4666
4667
4668
4669
4670
4671
4672
4673
4674
4675
4676
4677
4678
4679
4680
4681
4682
4683
4684
4685
4686
        """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
        if not hasattr(hf_config, "eagle_aux_hidden_state_layer_ids"):
            return None

        layer_ids = hf_config.eagle_aux_hidden_state_layer_ids
        if layer_ids and isinstance(layer_ids, (list, tuple)):
            return tuple(layer_ids)

        return None

4687
4688
4689
4690
4691
4692
4693
4694
4695
4696
4697
4698
4699
    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
4700
            into kernel format (repacking, renaming, etc.)
4701
4702
4703
4704
4705
4706
4707
4708
4709
4710
4711
4712
4713
4714
4715
4716
4717
4718
4719
4720
4721
4722
4723
4724
4725
4726
4727
4728
4729
4730
4731
4732
4733
4734
4735
4736
4737
4738
4739
4740
4741
4742
4743
4744
4745
4746
4747
4748
4749
4750
4751
4752
4753
4754
4755
4756
4757
4758
4759
4760
4761
        """
        # 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")
        load_device = (
            self.vllm_config.load_config.device or self.vllm_config.device_config.device
        )
        with torch.device(load_device):
            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)

            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)

        # 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",
4762
        )
4763
4764
4765
4766
4767
4768
4769
        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,
                )
4770

4771
4772
4773
    def _get_prompt_logprobs_dict(
        self,
        hidden_states: torch.Tensor,
4774
        num_scheduled_tokens: dict[str, int],
4775
    ) -> dict[str, LogprobsTensors | None]:
4776
        num_prompt_logprobs_dict = self.num_prompt_logprobs
4777
4778
4779
        if not num_prompt_logprobs_dict:
            return {}

4780
        in_progress_dict = self.input_batch.in_progress_prompt_logprobs_cpu
4781
        prompt_logprobs_dict: dict[str, LogprobsTensors | None] = {}
4782
4783
4784
4785
4786

        # 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():
4787
4788
4789
4790
            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
4791
4792
4793

            # Get metadata for this request.
            request = self.requests[req_id]
4794
4795
4796
4797
            if request.prompt_token_ids is None:
                # Prompt logprobs is incompatible with prompt embeddings
                continue

4798
4799
            num_prompt_tokens = len(request.prompt_token_ids)
            prompt_token_ids = torch.tensor(request.prompt_token_ids).to(
4800
4801
                self.device, non_blocking=True
            )
4802

4803
4804
4805
4806
4807
4808
            # 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(
4809
4810
                    num_prompt_tokens - 1, num_prompt_logprobs + 1
                )
4811
4812
                in_progress_dict[req_id] = logprobs_tensors

4813
            # Determine number of logits to retrieve.
4814
4815
            start_idx = request.num_computed_tokens
            start_tok = start_idx + 1
4816
            num_remaining_tokens = num_prompt_tokens - start_tok
4817
            if num_tokens <= num_remaining_tokens:
4818
                # This is a chunk, more tokens remain.
4819
4820
4821
                # 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.
4822
4823
4824
4825
4826
                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)
4827
4828
4829
4830
4831
4832
4833
                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
4834
4835
4836
4837
4838

            # 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]
4839
            offset = self.query_start_loc.np[req_idx].item()
4840
            prompt_hidden_states = hidden_states[offset : offset + num_logits]
4841
            logits = self.model.compute_logits(prompt_hidden_states)
4842
4843
4844
4845

            # 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.
4846
            tgt_token_ids = prompt_token_ids[start_tok : start_tok + num_logits]
4847
4848

            # Compute prompt logprobs.
4849
            logprobs = self.sampler.compute_logprobs(logits)
4850
            token_ids, logprobs, ranks, _ = self.sampler.gather_logprobs(
4851
4852
                logprobs, num_prompt_logprobs, tgt_token_ids
            )
4853
4854

            # Transfer GPU->CPU async.
4855
4856
            chunk_slice = slice(start_idx, start_idx + num_logits)
            logprobs_tensors.logprob_token_ids[chunk_slice].copy_(
4857
4858
4859
                token_ids, non_blocking=True
            )
            logprobs_tensors.logprobs[chunk_slice].copy_(logprobs, non_blocking=True)
4860
            logprobs_tensors.selected_token_ranks[chunk_slice].copy_(
4861
4862
                ranks, non_blocking=True
            )
4863
4864
4865
4866
4867

        # 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]
4868
            del in_progress_dict[req_id]
4869
4870

        # Must synchronize the non-blocking GPU->CPU transfers.
4871
        if prompt_logprobs_dict:
4872
            self._sync_device()
4873
4874
4875

        return prompt_logprobs_dict

4876
4877
    def _get_nans_in_logits(
        self,
4878
        logits: torch.Tensor | None,
4879
4880
4881
4882
4883
4884
4885
4886
4887
4888
4889
    ) -> 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])
4890
4891
4892
                    if num_nans_for_index is not None and req_index < logits.shape[0]
                    else 0
                )
4893
4894
4895
4896
            return num_nans_in_logits
        except IndexError:
            return {}

4897
    @contextmanager
4898
4899
4900
    def maybe_randomize_inputs(
        self, input_ids: torch.Tensor | None, inputs_embeds: torch.Tensor | None
    ):
4901
4902
4903
4904
        """
        Randomize input_ids if VLLM_RANDOMIZE_DP_DUMMY_INPUTS is set.
        This is to help balance expert-selection
         - during profile_run
4905
         - during DP rank dummy run
4906
        """
4907

4908
4909
4910
4911
        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
4912
        elif input_ids is not None:
4913
4914
4915
4916

            @functools.cache
            def rand_input_ids() -> torch.Tensor:
                return torch.randint_like(
4917
                    self.input_ids.gpu,
4918
4919
                    low=0,
                    high=self.model_config.get_vocab_size(),
4920
                )
4921

4922
            logger.debug_once("Randomizing dummy input_ids for DP Rank")
4923
            input_ids.copy_(rand_input_ids()[: input_ids.size(0)], non_blocking=True)
4924
4925
            yield
            input_ids.fill_(0)
4926
4927
4928
4929
4930
4931
4932
4933
4934
4935
4936
4937
4938
4939
4940
        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)
4941

4942
4943
4944
4945
4946
4947
    def _get_mm_dummy_batch(
        self,
        modality: str,
        max_items_per_batch: int,
    ) -> BatchedTensorInputs:
        """Dummy data for profiling and precompiling multimodal models."""
4948
4949
        assert self.mm_budget is not None

4950
4951
4952
        # 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,
4953
            mm_counts={modality: 1},
4954
            cache=self.mm_budget.cache,
4955
        )
4956
4957
4958
4959
4960
        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"
4961

4962
        return next(
4963
4964
            mm_kwargs_batch
            for _, _, mm_kwargs_batch in group_and_batch_mm_kwargs(
4965
                [(modality, dummy_mm_item)] * max_items_per_batch,
4966
4967
4968
4969
                device=self.device,
                pin_memory=self.pin_memory,
            )
        )
4970

4971
4972
4973
4974
    @torch.inference_mode()
    def _dummy_run(
        self,
        num_tokens: int,
4975
        cudagraph_runtime_mode: CUDAGraphMode | None = None,
4976
4977
        force_attention: bool = False,
        uniform_decode: bool = False,
4978
        allow_microbatching: bool = True,
4979
4980
        skip_eplb: bool = False,
        is_profile: bool = False,
4981
        create_mixed_batch: bool = False,
4982
        remove_lora: bool = True,
Rémi Delacourt's avatar
Rémi Delacourt committed
4983
        is_graph_capturing: bool = False,
4984
        num_active_loras: int = 0,
4985
        profile_seq_lens: int | None = None,
4986
    ) -> tuple[torch.Tensor, torch.Tensor]:
4987
4988
4989
4990
4991
4992
4993
        """
        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.
4994
                - if not set will determine the cudagraph mode based on using
4995
                    the self.cudagraph_dispatcher.
4996
4997
4998
4999
                - CUDAGraphMode.NONE: No cudagraph, for warm up and profile run
                - CUDAGraphMode.PIECEWISE: Piecewise cudagraph.
                - CUDAGraphMode.FULL: Full cudagraph, attention metadata is
                    needed.
5000
            force_attention: If True, always create attention metadata. Used to
5001
5002
5003
5004
                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.
5005
5006
            create_mixed_batch: If True, create a mixed batch with both decode
                (1 token) and prefill (multiple tokens) requests.
5007
            remove_lora: If False, dummy LoRAs are not destroyed after the run
5008
5009
            num_active_loras: Number of distinct active LoRAs to capture for.
                LoRA is activated when num_active_loras > 0.
5010
5011
5012
            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.
5013
        """
5014
5015
        mm_config = self.vllm_config.model_config.multimodal_config
        if mm_config and mm_config.mm_encoder_only:
5016
5017
5018
5019
            # 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([])

5020
5021
        assert (
            cudagraph_runtime_mode is None
5022
            or cudagraph_runtime_mode.is_valid_runtime_mode()
5023
        )
5024

5025
        # If cudagraph_mode.decode_mode() == FULL and
5026
        # cudagraph_mode.separate_routine(). This means that we are using
5027
5028
5029
5030
5031
5032
5033
5034
5035
5036
5037
        # 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.
5038
        max_query_len = self.uniform_decode_query_len if uniform_decode else num_tokens
5039

5040
5041
5042
        # 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.
5043
        assert num_tokens <= self.max_num_tokens
5044
        max_num_reqs = self.scheduler_config.max_num_seqs
5045
5046
5047
5048
        if create_mixed_batch:
            assert not uniform_decode
            # Create mixed batch:
            # first half decode tokens, second half one prefill
5049
            num_decode_tokens = min(max_num_reqs - 1, num_tokens // 2)
5050
5051
5052
5053
            num_prefill_tokens = num_tokens - num_decode_tokens
            num_reqs = num_decode_tokens + 1

            # Create decode requests (1 token each) followed by prefill request
5054
            num_scheduled_tokens_list = [1] * num_decode_tokens + [num_prefill_tokens]
5055
5056
5057
            # Note: Overriding max_query_len to be the prefill tokens
            max_query_len = num_prefill_tokens
        elif uniform_decode:
5058
            assert not create_mixed_batch
5059
            num_reqs = min(max_num_reqs, cdiv(num_tokens, max_query_len))
5060
5061
            num_scheduled_tokens_list = [max_query_len] * num_reqs
            if num_tokens % max_query_len != 0:
5062
                num_scheduled_tokens_list[-1] = num_tokens % max_query_len
5063
5064
5065
5066
5067
5068
        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

5069
5070
        assert sum(num_scheduled_tokens_list) == num_tokens
        assert len(num_scheduled_tokens_list) == num_reqs
5071
        num_scheduled_tokens = np.array(num_scheduled_tokens_list, dtype=np.int32)
5072
5073
        num_tokens_unpadded = int(num_scheduled_tokens.sum())

5074
        num_sampled_tokens = np.ones(num_reqs, dtype=np.int32)
5075

5076
        _cudagraph_mode, batch_desc, should_ubatch, num_tokens_across_dp, _ = (
5077
5078
5079
5080
5081
5082
5083
5084
5085
5086
5087
5088
5089
5090
5091
5092
5093
            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
5094
5095
5096
5097
                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,
5098
5099
            )
        )
5100
5101
5102

        if cudagraph_runtime_mode is None:
            cudagraph_runtime_mode = _cudagraph_mode
5103
        else:
5104
5105
5106
5107
5108
5109
5110
5111
5112
            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
        )
5113
        ubatch_slices, ubatch_slices_padded = maybe_create_ubatch_slices(
5114
5115
5116
5117
5118
5119
5120
5121
5122
5123
            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,
5124
        )
5125

5126
        attn_metadata: PerLayerAttnMetadata | None = None
5127

5128
5129
5130
5131
5132
5133
5134
        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,
        )

5135
5136
5137
5138
5139
5140
5141
5142
        # _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:
5143
5144
5145
                if profile_seq_lens is not None:
                    seq_lens = profile_seq_lens  # type: ignore[assignment]
                elif create_mixed_batch:
5146
5147
5148
                    # 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
5149
                    seq_lens = [1] * num_decode_tokens + [num_prefill_tokens + 1]  # type: ignore[assignment]
5150
5151
5152
5153
5154
                else:
                    seq_lens = max_query_len  # type: ignore[assignment]
                self.seq_lens.np[:num_reqs] = seq_lens
                self.seq_lens.np[num_reqs:] = 0
                self.seq_lens.copy_to_gpu()
5155

5156
5157
5158
                cum_num_tokens, _ = self._get_cumsum_and_arange(num_scheduled_tokens)
                self.query_start_loc.np[1 : num_reqs + 1] = cum_num_tokens
                self.query_start_loc.copy_to_gpu()
5159

5160
5161
5162
5163
5164
5165
5166
5167
5168
                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,
5169
                    use_spec_decode=self.speculative_config is not None,
5170
                )
5171

5172
        with self.maybe_dummy_run_with_lora(
5173
5174
5175
5176
            self.lora_config,
            num_scheduled_tokens,
            num_sampled_tokens,
            remove_lora,
5177
            num_active_loras,
5178
        ):
5179
            # Make sure padding doesn't exceed max_num_tokens
5180
            assert num_tokens_padded <= self.max_num_tokens
5181
            model_kwargs = self._init_model_kwargs()
5182
            if self.supports_mm_inputs and not self.model_config.is_encoder_decoder:
Patrick von Platen's avatar
Patrick von Platen committed
5183
5184
                input_ids, inputs_embeds = self._prepare_mm_inputs(num_tokens_padded)

5185
                model_kwargs = {
5186
                    **model_kwargs,
5187
5188
                    **self._dummy_mm_kwargs(num_reqs),
                }
5189
5190
            elif self.enable_prompt_embeds:
                input_ids = None
5191
                inputs_embeds = self.inputs_embeds.gpu[:num_tokens_padded]
5192
                model_kwargs = self._init_model_kwargs()
5193
            else:
5194
                input_ids = self.input_ids.gpu[:num_tokens_padded]
5195
                inputs_embeds = None
5196

5197
            if self.uses_mrope:
5198
                positions = self.mrope_positions.gpu[:, :num_tokens_padded]
5199
            elif self.uses_xdrope_dim > 0:
5200
                positions = self.xdrope_positions.gpu[:, :num_tokens_padded]
5201
            else:
5202
                positions = self.positions.gpu[:num_tokens_padded]
5203
5204
5205
5206
5207
5208
5209
5210
5211

            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,
5212
5213
5214
                            device=self.device,
                        )
                    )
5215
5216

                intermediate_tensors = self.sync_and_slice_intermediate_tensors(
5217
                    num_tokens_padded, None, False
5218
                )
5219

5220
            if ubatch_slices_padded is not None:
5221
5222
5223
                # Adjust values to reflect a single ubatch.
                # TODO(sage,lucas): this is cruft that should be addressed in
                #  the padding refactor.
5224
                num_tokens_padded = ubatch_slices_padded[0].num_tokens
5225
                if num_tokens_across_dp is not None:
5226
                    num_tokens_across_dp[:] = num_tokens_padded
5227

5228
            with (
5229
                self.maybe_randomize_inputs(input_ids, inputs_embeds),
5230
                set_forward_context(
5231
5232
                    attn_metadata,
                    self.vllm_config,
5233
                    num_tokens=num_tokens_padded,
5234
5235
                    num_tokens_across_dp=num_tokens_across_dp,
                    cudagraph_runtime_mode=cudagraph_runtime_mode,
5236
                    batch_descriptor=batch_desc,
5237
                    ubatch_slices=ubatch_slices_padded,
5238
                    slot_mapping=slot_mappings,
5239
5240
                ),
            ):
5241
                outputs = self.model(
5242
5243
5244
5245
                    input_ids=input_ids,
                    positions=positions,
                    intermediate_tensors=intermediate_tensors,
                    inputs_embeds=inputs_embeds,
5246
                    **model_kwargs,
5247
                )
5248

5249
5250
5251
5252
            if self.use_aux_hidden_state_outputs:
                hidden_states, _ = outputs
            else:
                hidden_states = outputs
5253

5254
5255
5256
            if self.speculative_config and (
                self.speculative_config.use_eagle()
                or self.speculative_config.uses_draft_model()
5257
                or self.speculative_config.uses_extract_hidden_states()
5258
            ):
5259
5260
5261
5262
                assert isinstance(
                    self.drafter,
                    EagleProposer | DraftModelProposer | ExtractHiddenStatesProposer,
                )
5263
                assert self.speculative_config is not None
5264
5265
5266
                # 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.
5267
                use_cudagraphs = (
5268
5269
5270
5271
5272
5273
5274
5275
5276
                    (
                        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
5277
5278
5279
5280
5281

                # 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
5282
5283
5284
5285
                if (
                    self.compilation_config.cudagraph_specialize_lora
                    and num_active_loras > 0
                ):
5286
5287
5288
5289
5290
                    use_cudagraphs = False

                self.drafter.dummy_run(
                    num_tokens,
                    use_cudagraphs=use_cudagraphs,
Rémi Delacourt's avatar
Rémi Delacourt committed
5291
                    is_graph_capturing=is_graph_capturing,
5292
                    slot_mappings=slot_mappings,
5293
                )
5294

5295
5296
5297
5298
5299
5300
5301
5302
5303
5304
5305
        # 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()

5306
5307
5308
5309
5310
5311
5312
5313
5314
5315
        # 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)

5316
        logit_indices = np.cumsum(num_scheduled_tokens) - 1
5317
5318
5319
5320
        logit_indices_device = torch.from_numpy(logit_indices).to(
            self.device, non_blocking=True
        )
        return hidden_states, hidden_states[logit_indices_device]
5321
5322
5323
5324
5325
5326

    @torch.inference_mode()
    def _dummy_sampler_run(
        self,
        hidden_states: torch.Tensor,
    ) -> torch.Tensor:
5327
5328
5329
        # The dummy hidden states may contain special values,
        # like `inf` or `nan`.
        # To avoid breaking the sampler, we use a random tensor here instead.
5330

5331
5332
        mm_config = self.vllm_config.model_config.multimodal_config
        if mm_config and mm_config.mm_encoder_only:
5333
5334
5335
            # MM Encoder only model no need to run sampler.
            return torch.tensor([])

5336
        hidden_states = torch.rand_like(hidden_states)
5337

5338
        logits = self.model.compute_logits(hidden_states)
5339
5340
        num_reqs = logits.size(0)

5341
        dummy_tensors = lambda v: torch.full((num_reqs,), v, device=self.device)
5342
5343
5344
5345
5346
5347
5348
5349
5350
5351
5352
5353
5354
5355
5356

        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,
            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)],
5357
            spec_token_ids=[[] for _ in range(num_reqs)],
5358
5359
            allowed_token_ids_mask=None,
            bad_words_token_ids={},
5360
            logitsprocs=LogitsProcessors(),
5361
        )
5362
        try:
5363
5364
5365
            sampler_output = self.sampler(
                logits=logits, sampling_metadata=dummy_metadata
            )
5366
        except RuntimeError as e:
5367
            if "out of memory" in str(e):
5368
5369
5370
5371
                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 "
5372
5373
                    "initializing the engine."
                ) from e
5374
5375
            else:
                raise e
5376
        if self.speculative_config:
5377
5378
            draft_token_ids = [[0] for _ in range(num_reqs)]
            dummy_spec_decode_metadata = SpecDecodeMetadata.make_dummy(
5379
5380
                draft_token_ids, self.device
            )
5381
5382
5383
5384
5385
5386

            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
5387
5388
5389
5390
5391
            logits = torch.randn(
                num_tokens + num_reqs,
                logits.shape[-1],
                device=self.device,
                dtype=logits.dtype,
5392
            )
5393
5394
5395
            self.rejection_sampler(
                dummy_spec_decode_metadata,
                draft_probs,
5396
                logits,
5397
5398
                dummy_metadata,
            )
5399
        return sampler_output
5400

5401
    def _dummy_pooler_run_task(
5402
5403
        self,
        hidden_states: torch.Tensor,
5404
5405
        task: PoolingTask,
    ) -> PoolerOutput:
5406
5407
5408
5409
        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
5410
5411
5412
5413
        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
5414
5415
5416

        req_num_tokens = num_tokens // num_reqs

5417
        dummy_prompt_lens = torch.from_numpy(num_scheduled_tokens_np)
5418
5419
5420
        dummy_token_ids = torch.zeros(
            (num_reqs, req_num_tokens), dtype=torch.int32, device=self.device
        )
5421

5422
        model = cast(VllmModelForPooling, self.get_model())
5423
        dummy_pooling_params = PoolingParams(task=task)
5424
        dummy_pooling_params.verify(self.model_config)
5425
        to_update = model.pooler.get_pooling_updates(task)
5426
5427
        to_update.apply(dummy_pooling_params)

5428
        dummy_metadata = PoolingMetadata(
5429
5430
5431
            prompt_lens=dummy_prompt_lens,
            prompt_token_ids=dummy_token_ids,
            pooling_params=[dummy_pooling_params] * num_reqs,
5432
            pooling_states=[PoolingStates() for i in range(num_reqs)],
5433
        )
5434

5435
        dummy_metadata.build_pooling_cursor(
5436
            num_scheduled_tokens_np,
5437
5438
            seq_lens_cpu=dummy_prompt_lens,
            device=hidden_states.device,
5439
        )
5440

5441
        try:
5442
5443
5444
            return model.pooler(
                hidden_states=hidden_states, pooling_metadata=dummy_metadata
            )
5445
        except RuntimeError as e:
5446
            if "out of memory" in str(e):
5447
                raise RuntimeError(
5448
5449
5450
                    "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 "
5451
5452
                    "initializing the engine."
                ) from e
5453
5454
            else:
                raise e
5455
5456
5457
5458
5459
5460

    @torch.inference_mode()
    def _dummy_pooler_run(
        self,
        hidden_states: torch.Tensor,
    ) -> PoolerOutput:
5461
5462
        mm_config = self.vllm_config.model_config.multimodal_config
        if mm_config and mm_config.mm_encoder_only:
5463
5464
5465
            # MM Encoder only model not need to run pooler.
            return torch.tensor([])

5466
        # Find the task that has the largest output for subsequent steps
5467
5468
5469
        supported_pooling_tasks = self.get_supported_pooling_tasks()

        if not supported_pooling_tasks:
5470
5471
5472
5473
5474
5475
            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."
            )
5476

5477
        output_size = dict[PoolingTask, float]()
5478
        for task in supported_pooling_tasks:
5479
5480
            # Run a full batch with each task to ensure none of them OOMs
            output = self._dummy_pooler_run_task(hidden_states, task)
5481
            output_size[task] = sum(o.nbytes for o in output if o is not None)
5482
5483
5484
5485
            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)
5486

5487
    def profile_run(self) -> None:
5488
        # Profile with multimodal encoder & encoder cache.
5489
        if self.supports_mm_inputs:
5490
5491
            mm_config = self.model_config.multimodal_config
            if mm_config is not None and mm_config.skip_mm_profiling:
5492
                logger.info(
5493
                    "Skipping memory profiling for multimodal encoder and "
5494
5495
                    "encoder cache."
                )
5496
5497
5498
5499
5500
            else:
                mm_budget = self.mm_budget
                assert mm_budget is not None

                if (encoder_budget := mm_budget.get_encoder_budget()) > 0:
5501
5502
5503
5504
5505
5506
5507
5508
5509
5510
5511
5512
5513
5514
5515
5516
5517
                    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
                        ]
5518

5519
                        logger.info_once(
5520
5521
5522
5523
5524
5525
                            "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,
5526
                            scope="local",
5527
                        )
5528

5529
5530
5531
5532
5533
                        # Create dummy batch of multimodal inputs.
                        batched_dummy_mm_inputs = self._get_mm_dummy_batch(
                            dummy_modality,
                            max_mm_items_per_batch,
                        )
5534

5535
5536
5537
5538
5539
5540
5541
5542
5543
5544
5545
                        # 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
5546

5547
        # Add `is_profile` here to pre-allocate communication buffers
5548
5549
5550
        hidden_states, last_hidden_states = self._dummy_run(
            self.max_num_tokens, is_profile=True
        )
5551
        if get_pp_group().is_last_rank:
5552
5553
5554
5555
            if self.is_pooling_model:
                output = self._dummy_pooler_run(hidden_states)
            else:
                output = self._dummy_sampler_run(last_hidden_states)
5556
        else:
5557
            output = None
5558
        self._sync_device()
5559
        del hidden_states, output
5560
        self.encoder_cache.clear()
5561
        gc.collect()
5562

5563
5564
5565
5566
5567
5568
5569
5570
5571
5572
    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

5573
5574
5575
        # 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
5576
        minimal_config = get_kv_cache_config_from_groups(
5577
            self.vllm_config, kv_cache_groups, available_memory=0
5578
        )
5579
        self.cache_config.num_gpu_blocks_override = saved_override
5580
5581
5582
5583
5584
5585
5586
5587
5588
5589
5590
5591
5592
5593
5594
5595
5596
5597
5598
5599
5600
5601
5602
5603
5604
5605
5606
5607
5608
5609
5610
5611
5612
5613
5614
5615
5616
5617
5618
5619
5620
5621
5622
5623
5624
5625
5626
5627
5628
5629
5630
5631
5632
5633
5634
5635
5636
5637
5638
5639
5640
5641
5642
5643
5644
5645
5646
5647
5648
5649
5650
5651
5652
5653
5654
5655
5656
5657
5658
5659
5660
5661
5662
5663
5664
5665
5666
5667
5668
5669
5670
5671
5672
5673
5674
5675
5676
5677
5678
5679
5680
5681
5682
5683
5684
5685
5686
5687
5688
5689
5690
5691
5692
5693
5694
5695
5696
5697
5698
5699
5700
5701
5702
5703
5704

        self.initialize_kv_cache(minimal_config)
        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"):
                layer.kv_cache = []

        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)]
5705
5706
5707
        for key_set in self.cudagraph_dispatcher.cudagraph_keys.values():
            key_set.clear()
        self.cudagraph_dispatcher.keys_initialized = False
5708
5709
5710
5711
5712
5713
5714
5715
5716
5717
5718
5719
5720
5721
5722
5723
5724
        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)

5725
    @instrument(span_name="Capture model")
5726
    def capture_model(self) -> int:
5727
        if self.compilation_config.cudagraph_mode == CUDAGraphMode.NONE:
5728
            logger.warning(
5729
                "Skipping CUDA graph capture. To turn on CUDA graph capture, "
5730
5731
                "ensure `cudagraph_mode` was not manually set to `NONE`"
            )
5732
            return 0
5733

5734
5735
5736
5737
5738
5739
5740
5741
5742
5743
5744
5745
5746
5747
5748
5749
5750
5751
5752
5753
5754
5755
5756
5757
5758
5759
5760
        # 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,
            )
            from vllm.v1.worker.gpu.mm.encoder_cudagraph import (
                EncoderCudaGraphManager,
            )

            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")

5761
5762
        compilation_counter.num_gpu_runner_capture_triggers += 1

5763
5764
        start_time = time.perf_counter()

5765
5766
5767
        # 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.
5768
        set_cudagraph_capturing_enabled(True)
5769
5770
5771
        with self._freeze_gc(), graph_capture(device=self.device):
            torch.accelerator.synchronize()
            torch.accelerator.empty_cache()
5772
            start_free_gpu_memory = torch.cuda.mem_get_info()[0]
5773

5774
5775
5776
5777
            for (
                runtime_mode,
                batch_descs,
            ) in self.cudagraph_dispatcher.get_capture_descs():
5778
                self._capture_cudagraphs(
5779
5780
                    batch_descriptors=batch_descs,
                    cudagraph_runtime_mode=runtime_mode,
5781
                )
5782
                torch.accelerator.synchronize()
5783

5784
5785
5786
5787
            # Capture encoder CUDA graphs if enabled
            if self.encoder_cudagraph_manager is not None:
                self.encoder_cudagraph_manager.capture()

5788
            torch.accelerator.synchronize()
5789
5790
            end_free_gpu_memory = torch.cuda.mem_get_info()[0]

5791
5792
5793
        # 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
5794
        # we may do lazy capturing in future that still allows capturing
5795
5796
        # after here.
        set_cudagraph_capturing_enabled(False)
5797

5798
5799
5800
        torch.accelerator.synchronize()
        torch.accelerator.empty_cache()

5801
5802
5803
5804
        # Lock workspace to prevent resizing during execution.
        # Max workspace sizes should have been captured during warmup/profiling.
        lock_workspace()

5805
5806
5807
5808
        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.
5809
        logger.info_once(
5810
5811
5812
            "Graph capturing finished in %.0f secs, took %.2f GiB",
            elapsed_time,
            cuda_graph_size / (1 << 30),
5813
            scope="local",
5814
        )
5815
        return cuda_graph_size
5816

5817
5818
5819
5820
5821
5822
5823
5824
5825
5826
5827
5828
5829
5830
5831
5832
5833
5834
5835
5836
5837
5838
5839
5840
5841
5842
5843
5844
5845
5846
5847
5848
5849
5850
    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,
        )

5851
5852
    def _capture_cudagraphs(
        self,
5853
        batch_descriptors: list[BatchDescriptor],
5854
5855
5856
5857
        cudagraph_runtime_mode: CUDAGraphMode,
    ):
        assert (
            cudagraph_runtime_mode != CUDAGraphMode.NONE
5858
            and cudagraph_runtime_mode.is_valid_runtime_mode()
5859
        ), f"Invalid cudagraph runtime mode: {cudagraph_runtime_mode}"
5860

5861
5862
5863
5864
5865
        if not batch_descriptors:
            return

        uniform_decode = batch_descriptors[0].uniform

5866
5867
        # Only rank 0 should print progress bar during capture
        if is_global_first_rank():
5868
5869
            batch_descriptors = tqdm(
                batch_descriptors,
5870
5871
5872
                disable=not self.load_config.use_tqdm_on_load,
                desc="Capturing CUDA graphs ({}, {})".format(
                    "decode" if uniform_decode else "mixed prefill-decode",
5873
5874
5875
                    cudagraph_runtime_mode.name,
                ),
            )
5876

5877
        # We skip EPLB here since we don't want to record dummy metrics
5878
        for batch_desc in batch_descriptors:
5879
            # We currently only capture ubatched graphs when its a FULL
5880
5881
5882
            # 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
5883
            allow_microbatching = (
5884
                self.parallel_config.use_ubatching
5885
5886
                and cudagraph_runtime_mode == CUDAGraphMode.FULL
                and uniform_decode
5887
5888
                and check_ubatch_thresholds(
                    config=self.vllm_config.parallel_config,
5889
                    num_tokens=batch_desc.num_tokens,
5890
5891
                    uniform_decode=uniform_decode,
                )
5892
            )
5893
5894
            self._warmup_and_capture(
                batch_desc,
5895
5896
5897
                cudagraph_runtime_mode=cudagraph_runtime_mode,
                allow_microbatching=allow_microbatching,
            )
5898
            torch.accelerator.synchronize()
5899
        self.maybe_remove_all_loras(self.lora_config)
5900

5901
5902
5903
5904
    def initialize_attn_backend(self, kv_cache_config: KVCacheConfig) -> None:
        """
        Initialize the attention backends and attention metadata builders.
        """
5905
        assert len(self.attn_groups) == 0, "Attention backends are already initialized"
5906

5907
5908
5909
5910
5911
5912
        class AttentionGroupKey(NamedTuple):
            attn_backend: type[AttentionBackend]
            kv_cache_spec: KVCacheSpec

        def get_attn_backends_for_group(
            kv_cache_group_spec: KVCacheGroupSpec,
5913
        ) -> tuple[dict[AttentionGroupKey, list[str]], set[type[AttentionBackend]]]:
5914
            layer_type = cast(type[Any], AttentionLayerBase)
5915
            layers = get_layers_from_vllm_config(
5916
                self.vllm_config, layer_type, kv_cache_group_spec.layer_names
5917
            )
5918
5919
            attn_backends = {}
            attn_backend_layers = defaultdict(list)
5920
            # Dedupe based on full class name; this is a bit safer than
5921
5922
5923
5924
            # 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.
5925
            for layer_name in kv_cache_group_spec.layer_names:
5926
                attn_backend = layers[layer_name].get_attn_backend()
5927
5928
5929
5930

                if layer_name in self.kv_sharing_fast_prefill_eligible_layers:
                    attn_backend = create_fast_prefill_custom_backend(
                        "FastPrefill",
5931
                        attn_backend,  # type: ignore[arg-type]
5932
5933
                    )

5934
5935
5936
                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):
5937
                    layer_kv_cache_spec = layer_kv_cache_spec.kv_cache_specs[layer_name]
5938
                key = (full_cls_name, layer_kv_cache_spec)
5939
5940
5941
                attn_backends[key] = AttentionGroupKey(
                    attn_backend, layer_kv_cache_spec
                )
5942
                attn_backend_layers[key].append(layer_name)
5943
5944
5945
5946
            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()),
            )
5947
5948

        def create_attn_groups(
5949
            attn_backends_map: dict[AttentionGroupKey, list[str]],
5950
            kv_cache_group_id: int,
5951
5952
        ) -> list[AttentionGroup]:
            attn_groups: list[AttentionGroup] = []
5953
            for (attn_backend, kv_cache_spec), layer_names in attn_backends_map.items():
5954
                attn_group = AttentionGroup(
5955
                    attn_backend,
5956
                    layer_names,
5957
                    kv_cache_spec,
5958
                    kv_cache_group_id,
5959
5960
                )

5961
5962
5963
                attn_groups.append(attn_group)
            return attn_groups

5964
        attention_backend_maps = []
5965
        attention_backend_list = []
5966
        for kv_cache_group_spec in kv_cache_config.kv_cache_groups:
5967
            attn_backends = get_attn_backends_for_group(kv_cache_group_spec)
5968
            attention_backend_maps.append(attn_backends[0])
5969
            attention_backend_list.append(attn_backends[1])
5970
5971

        # Resolve cudagraph_mode before actually initialize metadata_builders
5972
5973
5974
        self._check_and_update_cudagraph_mode(
            attention_backend_list, kv_cache_config.kv_cache_groups
        )
5975

5976
5977
5978
        # Check if attention backend supports PCP&DCP and related features.
        check_attention_cp_compatibility(self.vllm_config)

5979
5980
        for i, attn_backend_map in enumerate(attention_backend_maps):
            self.attn_groups.append(create_attn_groups(attn_backend_map, i))
5981

5982
5983
5984
5985
5986
5987
5988
5989
5990
5991
5992
5993
5994
5995
5996
    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
5997
5998
                    if not self.parallel_config.use_ubatching
                    else self.parallel_config.num_ubatches,
5999
                )
co63oc's avatar
co63oc committed
6000
        # Calculate reorder batch threshold (if needed)
6001
6002
        # Note (tdoublep): do this *after* constructing builders,
        # because some of them change the threshold at init time.
6003
6004
        self.calculate_reorder_batch_threshold()

6005
6006
6007
6008
6009
6010
6011
6012
        # Initialize drafter attention backend
        if self.speculative_config and (
            self.speculative_config.use_eagle()
            or self.speculative_config.uses_draft_model()
        ):
            assert isinstance(self.drafter, EagleProposer | DraftModelProposer)
            self.drafter.initialize_attn_backend(kv_cache_config, kernel_block_sizes)

6013
    def _check_and_update_cudagraph_mode(
6014
6015
6016
        self,
        attention_backends: list[set[type[AttentionBackend]]],
        kv_cache_groups: list[KVCacheGroupSpec],
6017
    ) -> None:
6018
        """
6019
        Resolve the cudagraph_mode when there are multiple attention
6020
        groups with potential conflicting CUDA graph support.
6021
6022
6023
        Then initialize the cudagraph_dispatcher based on the resolved
        cudagraph_mode.
        """
6024
        min_cg_support = AttentionCGSupport.ALWAYS
6025
        min_cg_backend_name = None
6026

6027
6028
6029
6030
6031
6032
6033
6034
6035
6036
6037
6038
        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__
6039
6040
        # Flexible resolve the cudagraph mode
        cudagraph_mode = self.compilation_config.cudagraph_mode
6041
        assert cudagraph_mode is not None
6042
        # check cudagraph for mixed batch is supported
6043
6044
6045
6046
6047
6048
        if (
            cudagraph_mode.mixed_mode() == CUDAGraphMode.FULL
            and min_cg_support != AttentionCGSupport.ALWAYS
        ):
            msg = (
                f"CUDAGraphMode.{cudagraph_mode.name} is not supported "
6049
                f"with {min_cg_backend_name} backend (support: "
6050
6051
                f"{min_cg_support})"
            )
6052
6053
            if min_cg_support == AttentionCGSupport.NEVER:
                # if not supported any full cudagraphs, just raise it.
6054
6055
                msg += (
                    "; please try cudagraph_mode=PIECEWISE, and "
6056
                    "make sure compilation mode is VLLM_COMPILE"
6057
                )
6058
6059
6060
6061
6062
                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"
6063
                cudagraph_mode = self.compilation_config.cudagraph_mode = (
6064
                    CUDAGraphMode.FULL_AND_PIECEWISE
6065
                )
6066
6067
            else:
                msg += "; setting cudagraph_mode=FULL_DECODE_ONLY"
6068
                cudagraph_mode = self.compilation_config.cudagraph_mode = (
6069
                    CUDAGraphMode.FULL_DECODE_ONLY
6070
                )
6071
6072
            logger.warning(msg)

6073
        # check that if we are doing decode full-cudagraphs it is supported
6074
6075
6076
6077
6078
6079
        if (
            cudagraph_mode.decode_mode() == CUDAGraphMode.FULL
            and min_cg_support == AttentionCGSupport.NEVER
        ):
            msg = (
                f"CUDAGraphMode.{cudagraph_mode.name} is not supported "
6080
                f"with {min_cg_backend_name} backend (support: "
6081
6082
                f"{min_cg_support})"
            )
6083
            if self.compilation_config.mode == CompilationMode.VLLM_COMPILE and (
6084
6085
6086
6087
6088
                self.compilation_config.splitting_ops_contain_attention()
                or self.compilation_config.use_inductor_graph_partition
            ):
                msg += (
                    "; setting cudagraph_mode=PIECEWISE because "
6089
                    "attention is compiled piecewise"
6090
6091
                )
                cudagraph_mode = self.compilation_config.cudagraph_mode = (
6092
                    CUDAGraphMode.PIECEWISE
6093
                )
6094
            else:
6095
6096
                msg += (
                    "; setting cudagraph_mode=NONE because "
6097
                    "attention is not compiled piecewise"
6098
6099
                )
                cudagraph_mode = self.compilation_config.cudagraph_mode = (
6100
                    CUDAGraphMode.NONE
6101
                )
6102
6103
            logger.warning(msg)

6104
6105
        # check that if we are doing spec-decode + decode full-cudagraphs it is
        # supported
6106
6107
6108
6109
6110
6111
6112
6113
        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 "
6114
                f"{min_cg_backend_name} (support: {min_cg_support})"
6115
            )
6116
6117
            if self.compilation_config.splitting_ops_contain_attention():
                msg += "; setting cudagraph_mode=PIECEWISE"
6118
                cudagraph_mode = self.compilation_config.cudagraph_mode = (
6119
                    CUDAGraphMode.PIECEWISE
6120
                )
6121
6122
            else:
                msg += "; setting cudagraph_mode=NONE"
6123
                cudagraph_mode = self.compilation_config.cudagraph_mode = (
6124
                    CUDAGraphMode.NONE
6125
                )
6126
6127
6128
6129
            logger.warning(msg)

        # double check that we can support full cudagraph if they are requested
        # even after automatic downgrades
6130
6131
6132
6133
6134
6135
        if (
            cudagraph_mode.has_full_cudagraphs()
            and min_cg_support == AttentionCGSupport.NEVER
        ):
            raise ValueError(
                f"CUDAGraphMode.{cudagraph_mode.name} is not "
6136
                f"supported with {min_cg_backend_name} backend ("
6137
6138
                f"support:{min_cg_support}) "
                "; please try cudagraph_mode=PIECEWISE, "
6139
                "and make sure compilation mode is VLLM_COMPILE"
6140
            )
6141

6142
6143
6144
6145
        # 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
6146
        # Will be removed in the near future when we have separate cudagraph capture
6147
6148
6149
6150
6151
6152
6153
6154
6155
6156
        # 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
            )

6157
6158
6159
6160
6161
6162
6163
6164
6165
6166
6167
6168
6169
6170
6171
6172
        # If the model has Mamba layers and cudagraph mode includes FULL
        # decode, cap cudagraph capture sizes to the number of available
        # Mamba cache blocks. Each decode request needs one conv_state
        # cache line, so capture batch sizes cannot exceed num_blocks.
        # Only FULL decode graphs are affected because PIECEWISE captures
        # run GDN/Mamba ops eagerly (prefill path, no causal_conv1d_update).
        # See: https://github.com/vllm-project/vllm/issues/34094
        if cudagraph_mode.has_full_cudagraphs():
            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:
                self.compilation_config.adjust_cudagraph_sizes_for_mamba_cache(
                    self.kv_cache_config.num_blocks
                )

6173
6174
        # Trigger cudagraph dispatching keys initialization after
        # resolved cudagraph mode.
6175
        self.compilation_config.cudagraph_mode = cudagraph_mode
6176
        self.cudagraph_dispatcher.initialize_cudagraph_keys(
6177
            cudagraph_mode, self.uniform_decode_query_len
6178
        )
6179

6180
6181
6182
6183
6184
6185
        # 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()
        ):
            assert isinstance(self.drafter, EagleProposer | ExtractHiddenStatesProposer)
6186
6187
            self.drafter.initialize_cudagraph_keys(cudagraph_mode)

6188
6189
    def calculate_reorder_batch_threshold(self) -> None:
        """
6190
6191
6192
6193
        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.
6194
        """
6195
6196
        min_none_high = lambda a, b: a if b is None else b if a is None else min(a, b)

6197
        reorder_batch_thresholds: list[int | None] = [
6198
6199
6200
            group.get_metadata_builder().reorder_batch_threshold
            for group in self._attn_group_iterator()
        ]
6201
6202
6203
6204
6205
        # 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
6206
        self.reorder_batch_threshold = reduce(min_none_high, reorder_batch_thresholds)  # type: ignore[assignment]
6207

6208
6209
6210
    def may_reinitialize_input_batch(
        self, kv_cache_config: KVCacheConfig, kernel_block_sizes: list[int]
    ) -> None:
6211
6212
        """
        Re-initialize the input batch if the block sizes are different from
6213
6214
6215
6216
        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.
6217
6218
6219

        Args:
            kv_cache_config: The KV cache configuration.
6220
            kernel_block_sizes: The kernel block sizes for each KV cache group.
6221
        """
6222
        block_sizes = []
6223
6224
        max_num_blocks = []
        max_model_len = max(self.max_model_len, self.max_encoder_len)
6225
        for kv_cache_group in kv_cache_config.kv_cache_groups:
6226
6227
            if isinstance(kv_cache_group.kv_cache_spec, EncoderOnlyAttentionSpec):
                continue
6228
6229
            block_size = kv_cache_group.kv_cache_spec.block_size
            block_sizes.append(block_size)
6230
            max_num_blocks_per_req = cdiv(
6231
                max_model_len, block_size * get_total_cp_world_size()
6232
6233
            )
            if isinstance(kv_cache_group.kv_cache_spec, MambaSpec):
6234
                max_num_blocks_per_req = (
6235
6236
6237
6238
6239
                    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)
6240

6241
6242
6243
6244
        if (
            block_sizes != self._init_block_sizes
            or kernel_block_sizes != self._init_kernel_block_sizes
        ):
6245
            assert self.offload_config.uva.cpu_offload_gb == 0, (
6246
6247
                "Cannot re-initialize the input batch when CPU weight "
                "offloading is enabled. See https://github.com/vllm-project/vllm/pull/18298 "  # noqa: E501
6248
6249
                "for more details."
            )
6250
6251
            self._init_block_sizes = block_sizes
            self._init_kernel_block_sizes = kernel_block_sizes
6252
6253
            self.input_batch = InputBatch(
                max_num_reqs=self.max_num_reqs,
6254
                max_model_len=max_model_len,
6255
6256
6257
6258
6259
                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,
6260
                kernel_block_sizes=kernel_block_sizes,
6261
                max_num_blocks_per_req=max_num_blocks,
6262
                is_spec_decode=bool(self.vllm_config.speculative_config),
6263
                logitsprocs=self.input_batch.logitsprocs,
6264
                logitsprocs_need_output_token_ids=self.input_batch.logitsprocs_need_output_token_ids,
6265
                is_pooling_model=self.is_pooling_model,
6266
6267
            )

6268
6269
6270
6271
6272
6273
6274
6275
6276
        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}"
        )

6277
    def _allocate_kv_cache_tensors(
6278
6279
        self, kv_cache_config: KVCacheConfig
    ) -> dict[str, torch.Tensor]:
6280
        """
6281
6282
6283
        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.

6284
        Args:
6285
            kv_cache_config: The KV cache config
6286
        Returns:
6287
            dict[str, torch.Tensor]: A map between layer names to their
6288
            corresponding memory buffer for KV cache.
6289
        """
6290
6291
        kv_cache_raw_tensors: dict[str, torch.Tensor] = {}
        for kv_cache_tensor in kv_cache_config.kv_cache_tensors:
6292
6293
6294
            tensor = torch.zeros(
                kv_cache_tensor.size, dtype=torch.int8, device=self.device
            )
6295
6296
6297
6298
6299
            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:
6300
6301
6302
6303
            for layer_name in group.layer_names:
                if layer_name in self.runner_only_attn_layers:
                    continue
                layer_names.add(layer_name)
6304
6305
6306
        assert layer_names == set(kv_cache_raw_tensors.keys()), (
            "Some layers are not correctly initialized"
        )
6307
6308
        return kv_cache_raw_tensors

6309
6310
6311
    def _attn_group_iterator(self) -> Iterator[AttentionGroup]:
        return itertools.chain.from_iterable(self.attn_groups)

6312
    def _kv_cache_spec_attn_group_iterator(self) -> Iterator[AttentionGroup]:
6313
6314
        if not self.kv_cache_config.kv_cache_groups:
            return
6315
6316
        for attn_groups in self.attn_groups:
            yield from attn_groups
6317

6318
6319
6320
6321
    def _reshape_kv_cache_tensors(
        self,
        kv_cache_config: KVCacheConfig,
        kv_cache_raw_tensors: dict[str, torch.Tensor],
6322
        kernel_block_sizes: list[int],
6323
    ) -> dict[str, torch.Tensor]:
6324
        """
6325
        Reshape the KV cache tensors to the desired shape and dtype.
6326

6327
        Args:
6328
6329
            kv_cache_config: The KV cache config
            kv_cache_raw_tensors: The KV cache buffer of each layer, with
6330
                correct size but uninitialized shape.
6331
            kernel_block_sizes: The kernel block sizes for each KV cache group.
6332
        Returns:
6333
            Dict[str, torch.Tensor]: A map between layer names to their
6334
6335
            corresponding memory buffer for KV cache.
        """
6336
        kv_caches: dict[str, torch.Tensor] = {}
6337
        has_attn, has_mamba = False, False
6338
6339
        for group in self._kv_cache_spec_attn_group_iterator():
            kv_cache_spec = group.kv_cache_spec
6340
            attn_backend = group.backend
6341
6342
6343
6344
            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]
6345
            for layer_name in group.layer_names:
6346
6347
                if layer_name in self.runner_only_attn_layers:
                    continue
6348
6349
                raw_tensor = kv_cache_raw_tensors[layer_name]
                assert raw_tensor.numel() % kv_cache_spec.page_size_bytes == 0
6350
                num_blocks = raw_tensor.numel() // kv_cache_spec.page_size_bytes
6351
                if isinstance(kv_cache_spec, AttentionSpec):
6352
                    has_attn = True
6353
6354
                    num_blocks_per_kv_block = (
                        kv_cache_spec.block_size // kernel_block_size
6355
6356
6357
                    )
                    kernel_num_blocks = num_blocks * num_blocks_per_kv_block

6358
                    kv_cache_shape = attn_backend.get_kv_cache_shape(
6359
                        kernel_num_blocks,
6360
                        kernel_block_size,
6361
6362
                        kv_cache_spec.num_kv_heads,
                        kv_cache_spec.head_size,
6363
6364
                        cache_dtype_str=self.cache_config.cache_dtype,
                    )
6365
                    dtype = kv_cache_spec.dtype
6366
                    try:
6367
                        kv_cache_stride_order = attn_backend.get_kv_cache_stride_order()
6368
                        assert len(kv_cache_stride_order) == len(kv_cache_shape)
6369
                    except (AttributeError, NotImplementedError):
6370
                        kv_cache_stride_order = tuple(range(len(kv_cache_shape)))
6371
6372
6373
6374
6375
                    # 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.
6376
6377
6378
                    kv_cache_shape = tuple(
                        kv_cache_shape[i] for i in kv_cache_stride_order
                    )
6379
6380
6381
6382
6383
                    # Maintain original KV shape view.
                    inv_order = [
                        kv_cache_stride_order.index(i)
                        for i in range(len(kv_cache_stride_order))
                    ]
6384
6385
6386
6387
6388
6389
                    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
6390
                elif isinstance(kv_cache_spec, MambaSpec):
6391
                    has_mamba = True
Chen Zhang's avatar
Chen Zhang committed
6392
6393
                    raw_tensor = kv_cache_raw_tensors[layer_name]
                    state_tensors = []
6394
                    storage_offset_bytes = 0
6395
                    for shape, dtype in zip(kv_cache_spec.shapes, kv_cache_spec.dtypes):
6396
6397
                        dtype_size = get_dtype_size(dtype)
                        num_element_per_page = (
6398
6399
                            kv_cache_spec.page_size_bytes // dtype_size
                        )
Chen Zhang's avatar
Chen Zhang committed
6400
                        target_shape = (num_blocks, *shape)
6401
6402
                        stride = torch.empty(target_shape).stride()
                        target_stride = (num_element_per_page, *stride[1:])
6403
                        assert storage_offset_bytes % dtype_size == 0
6404
6405
6406
6407
                        tensor = torch.as_strided(
                            raw_tensor.view(dtype),
                            size=target_shape,
                            stride=target_stride,
6408
                            storage_offset=storage_offset_bytes // dtype_size,
6409
                        )
Chen Zhang's avatar
Chen Zhang committed
6410
                        state_tensors.append(tensor)
6411
                        storage_offset_bytes += stride[0] * dtype_size
6412
6413

                    kv_caches[layer_name] = state_tensors
6414
                else:
6415
                    raise NotImplementedError
6416
6417

        if has_attn and has_mamba:
6418
            self._update_hybrid_attention_mamba_layout(kv_caches)
6419

6420
6421
        return kv_caches

6422
    def _update_hybrid_attention_mamba_layout(
6423
6424
        self, kv_caches: dict[str, torch.Tensor]
    ) -> None:
6425
        """
6426
6427
        Update the layout of attention layers from (2, num_blocks, ...) to
        (num_blocks, 2, ...).
6428
6429

        Args:
6430
            kv_caches: The KV cache buffer of each layer.
6431
6432
        """

6433
6434
        for group in self._kv_cache_spec_attn_group_iterator():
            kv_cache_spec = group.kv_cache_spec
6435
            for layer_name in group.layer_names:
6436
                kv_cache = kv_caches[layer_name]
6437
6438
6439
6440
                if isinstance(kv_cache_spec, AttentionSpec) and kv_cache.shape[0] == 2:
                    assert kv_cache.shape[1] != 2, (
                        "Fail to determine whether the layout is "
                        "(2, num_blocks, ...) or (num_blocks, 2, ...) for "
6441
                        f"a tensor of shape {kv_cache.shape}"
6442
                    )
6443
                    hidden_size = kv_cache.shape[2:].numel()
6444
6445
6446
6447
                    kv_cache.as_strided_(
                        size=kv_cache.shape,
                        stride=(hidden_size, 2 * hidden_size, *kv_cache.stride()[2:]),
                    )
6448

6449
    def initialize_kv_cache_tensors(
6450
        self, kv_cache_config: KVCacheConfig, kernel_block_sizes: list[int]
6451
    ) -> dict[str, torch.Tensor]:
6452
6453
6454
6455
6456
        """
        Initialize the memory buffer for KV cache.

        Args:
            kv_cache_config: The KV cache config
6457
6458
            kernel_block_sizes: The kernel block sizes for each KV cache group.

6459
        Returns:
6460
            Dict[str, torch.Tensor]: A map between layer names to their
6461
6462
            corresponding memory buffer for KV cache.
        """
6463
6464
6465
6466
6467
6468
6469
6470
6471
6472
6473
6474
6475
6476
6477
6478
6479
6480
6481
6482
6483
6484
6485
6486

        # 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
            )
6487

6488
        # Set up cross-layer KV cache sharing
6489
6490
        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)
6491
6492
            kv_caches[layer_name] = kv_caches[target_layer_name]

6493
6494
6495
6496
6497
6498
6499
6500
6501
        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,
        )
6502
6503
6504
        return kv_caches

    def maybe_add_kv_sharing_layers_to_kv_cache_groups(
6505
6506
        self, kv_cache_config: KVCacheConfig
    ) -> None:
6507
6508
6509
6510
6511
6512
6513
6514
6515
6516
6517
6518
6519
6520
6521
6522
6523
6524
        """
        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.
6525
            attn_layers = get_layers_from_vllm_config(self.vllm_config, Attention)
6526
6527
            for layer_name in reversed(attn_layers):
                if layer_name in self.shared_kv_cache_layers:
6528
                    self.kv_sharing_fast_prefill_eligible_layers.add(layer_name)
6529
6530
                else:
                    break
6531

6532
6533
6534
6535
6536
6537
6538
    def initialize_kv_cache(self, kv_cache_config: KVCacheConfig) -> None:
        """
        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
        """
6539
        kv_cache_config = deepcopy(kv_cache_config)
6540
        self.kv_cache_config = kv_cache_config
6541
        self._mamba_copy_bufs = None
6542
        self.may_add_encoder_only_layers_to_kv_cache_config()
6543
        self.maybe_add_kv_sharing_layers_to_kv_cache_groups(kv_cache_config)
6544
        self.initialize_attn_backend(kv_cache_config)
6545
6546
6547
6548
6549
        # 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.
6550
6551
6552
        kernel_block_sizes = prepare_kernel_block_sizes(
            kv_cache_config, self.attn_groups
        )
6553
        self._kernel_block_sizes = kernel_block_sizes
6554
6555
6556
6557

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

6558
        # Reinitialize need to after initialize_attn_backend
6559
6560
6561
6562
        self.may_reinitialize_input_batch(kv_cache_config, kernel_block_sizes)
        kv_caches = self.initialize_kv_cache_tensors(
            kv_cache_config, kernel_block_sizes
        )
6563

6564
6565
6566
        if (
            self.speculative_config
            and self.speculative_config.uses_extract_hidden_states()
6567
        ):
6568
            assert isinstance(self.drafter, ExtractHiddenStatesProposer)
6569
6570
6571
6572
            # 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
6573
        if has_kv_transfer_group():
6574
            kv_transfer_group = get_kv_transfer_group()
6575
6576
6577
6578
6579
6580
6581
            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)
6582
            kv_transfer_group.set_host_xfer_buffer_ops(copy_kv_blocks)
Robert Shaw's avatar
Robert Shaw committed
6583

6584
6585
6586
6587
6588
6589
    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
6590
6591
6592
6593
6594
6595
6596

    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()
6597
6598
6599
6600
6601
6602
6603
6604
        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)
6605
        self.max_num_kv_tokens = (
6606
6607
6608
6609
6610
6611
6612
            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

6613
6614
6615
        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,
6616
            vllm_config=self.vllm_config,
6617
        )
6618
        self._bind_routed_experts_capturer(routed_experts_capturer)
6619
        self.routed_experts_initialized = True
6620
6621
6622
6623
6624
6625
6626
6627
6628
6629
6630
6631
6632
6633
6634

    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)
6635

6636
6637
6638
6639
6640
    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
6641
        encoder_only_attn_specs: dict[AttentionSpec, list[str]] = defaultdict(list)
6642
6643
6644
        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:
6645
                attn_spec: AttentionSpec = EncoderOnlyAttentionSpec(
6646
6647
6648
                    block_size=block_size,
                    num_kv_heads=attn_module.num_kv_heads,
                    head_size=attn_module.head_size,
6649
6650
                    dtype=self.kv_cache_dtype,
                )
6651
6652
6653
                encoder_only_attn_specs[attn_spec].append(layer_name)
                self.runner_only_attn_layers.add(layer_name)
        if len(encoder_only_attn_specs) > 0:
6654
6655
6656
            assert len(encoder_only_attn_specs) == 1, (
                "Only support one encoder-only attention spec now"
            )
6657
6658
            spec, layer_names = encoder_only_attn_specs.popitem()
            self.kv_cache_config.kv_cache_groups.append(
6659
6660
                KVCacheGroupSpec(layer_names=layer_names, kv_cache_spec=spec)
            )
6661

6662
    def get_kv_cache_spec(self) -> dict[str, KVCacheSpec]:
6663
        """
6664
        Generates the KVCacheSpec by parsing the kv cache format from each
6665
6666
        Attention module in the static forward context.
        Returns:
6667
            KVCacheSpec: A dictionary mapping layer names to their KV cache
6668
6669
            format. Layers that do not need KV cache are not included.
        """
6670
        if has_ec_transfer() and not get_ec_transfer().is_consumer:
6671
            return {}
6672
        kv_cache_spec: dict[str, KVCacheSpec] = {}
6673
6674
        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
6675
        for layer_name, attn_module in attn_layers.items():
6676
6677
6678
6679
6680
6681
6682
6683
6684
6685
6686
6687
6688
6689
6690
            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
6691

6692
        return kv_cache_spec
6693

6694
    def _to_list(self, sampled_token_ids: torch.Tensor) -> list[list[int]]:
6695
6696
6697
6698
6699
6700
6701
6702
        # 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.
6703
        pinned = self.sampled_token_ids_pinned_cpu[: sampled_token_ids.shape[0]]
6704
6705
6706
        pinned.copy_(sampled_token_ids, non_blocking=True)
        self.transfer_event.record()
        self.transfer_event.synchronize()
6707
        return pinned.tolist()
6708
6709
6710
6711
6712
6713
6714
6715
6716
6717
6718
6719
6720
6721
6722
6723
6724
6725
6726
6727
6728
6729
6730
6731
6732
6733
6734
6735
6736
6737
6738
6739
6740
6741
6742
6743
6744
6745
6746
6747

    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}

6748
        torch.accelerator.synchronize()
6749
6750
6751
6752
6753
        start_time = time.perf_counter()

        try:
            yield
        finally:
6754
            torch.accelerator.synchronize()
6755
6756
6757
6758
6759
6760
6761
6762
6763
6764
            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]
6765
                    stats.encoder_forward_secs += per_request_time
6766
6767
6768
6769
6770
6771
6772
                    stats.num_encoder_calls += 1


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

6773
    encoder_forward_secs: float = 0.0
6774
6775
6776
6777
6778
6779
6780
    """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 {
6781
            "encoder_forward_secs": self.encoder_forward_secs,
6782
6783
            "num_encoder_calls": self.num_encoder_calls,
        }