gpu_model_runner.py 301 KB
Newer Older
1
# SPDX-License-Identifier: Apache-2.0
2
# SPDX-FileCopyrightText: Copyright contributors to the vLLM project
3

4
import functools
5
import gc
6
import itertools
7
import threading
8
import time
9
from collections import defaultdict
10
from collections.abc import Callable, Iterable, Iterator, Sequence
11
from contextlib import contextmanager
12
from copy import copy, deepcopy
13
from dataclasses import dataclass, replace
14
from functools import reduce
15
from typing import TYPE_CHECKING, Any, NamedTuple, TypeAlias, cast
16
17
18
19
20

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

23
import vllm.envs as envs
24
from vllm.compilation.counter import compilation_counter
25
from vllm.compilation.cuda_graph import CUDAGraphStat, CUDAGraphWrapper
26
from vllm.compilation.monitor import set_cudagraph_capturing_enabled
27
from vllm.config import (
28
    CompilationMode,
29
30
31
    CUDAGraphMode,
    VllmConfig,
    get_layers_from_vllm_config,
32
    set_current_vllm_config,
33
34
    update_config,
)
35
from vllm.config.cache import CacheConfig
36
from vllm.distributed.ec_transfer import get_ec_transfer, has_ec_transfer
37
from vllm.distributed.eplb.eplb_state import EplbState
38
from vllm.distributed.kv_transfer import get_kv_transfer_group, has_kv_transfer_group
39
from vllm.distributed.kv_transfer.kv_connector.utils import copy_kv_blocks
40
from vllm.distributed.parallel_state import (
41
    get_dcp_group,
42
43
44
45
46
47
    get_pp_group,
    get_tp_group,
    graph_capture,
    is_global_first_rank,
    prepare_communication_buffer_for_model,
)
48
49
50
51
from vllm.forward_context import (
    BatchDescriptor,
    set_forward_context,
)
52
from vllm.logger import init_logger
53
from vllm.lora.layers import LoRAMapping, LoRAMappingType
54
from vllm.model_executor.layers.attention import Attention, MLAAttention
55
from vllm.model_executor.layers.attention_layer_base import AttentionLayerBase
56
57
58
from vllm.model_executor.layers.fused_moe.routed_experts_capturer import (
    RoutedExpertsCapturer,
)
59
60
61
from vllm.model_executor.layers.mamba.ops.ssu_dispatch import (
    initialize_mamba_ssu_backend,
)
62
63
64
65
from vllm.model_executor.layers.rotary_embedding import (
    MRotaryEmbedding,
    XDRotaryEmbedding,
)
66
from vllm.model_executor.model_loader import get_model_loader
67
68
69
70
from vllm.model_executor.model_loader.reload import (
    finalize_layerwise_reload,
    initialize_layerwise_reload,
)
71
from vllm.model_executor.models.interfaces import (
72
    MultiModalEmbeddings,
73
    SupportsMRoPE,
74
    SupportsMultiModal,
75
    SupportsXDRoPE,
76
77
78
79
    is_mixture_of_experts,
    supports_eagle3,
    supports_mrope,
    supports_multimodal_pruning,
80
    supports_realtime,
81
    supports_transcription,
82
    supports_xdrope,
83
)
84
from vllm.model_executor.models.interfaces_base import (
85
86
87
88
    VllmModelForPooling,
    is_pooling_model,
    is_text_generation_model,
)
89
90
91
92
93
from vllm.model_executor.offloader import (
    create_offloader,
    get_offloader,
    set_offloader,
)
94
from vllm.multimodal import MULTIMODAL_REGISTRY
95
from vllm.multimodal.encoder_budget import MultiModalBudget
96
97
98
99
100
from vllm.multimodal.inputs import (
    BatchedTensorInputs,
    MultiModalKwargsItem,
    PlaceholderRange,
)
101
from vllm.multimodal.utils import group_and_batch_mm_kwargs
102
from vllm.platforms import current_platform
103
from vllm.pooling_params import PoolingParams
104
from vllm.sampling_params import SamplingType
105
from vllm.sequence import IntermediateTensors
106
from vllm.tasks import GenerationTask, PoolingTask, SupportedTask
107
from vllm.tracing import instrument
108
from vllm.utils import length_from_prompt_token_ids_or_embeds
109
from vllm.utils.math_utils import cdiv, round_up
110
from vllm.utils.mem_utils import DeviceMemoryProfiler, format_gib
111
from vllm.utils.nvtx_pytorch_hooks import PytHooks
112
from vllm.utils.platform_utils import is_pin_memory_available, num_compute_units
113
114
from vllm.utils.torch_utils import (
    get_dtype_size,
115
    is_quantized_kv_cache,
116
117
    kv_cache_dtype_str_to_dtype,
)
118
119
from vllm.v1.attention.backend import (
    AttentionBackend,
120
    AttentionCGSupport,
121
    AttentionMetadata,
122
    AttentionMetadataBuilder,
123
    AttentionType,
124
    CommonAttentionMetadata,
125
)
126
from vllm.v1.attention.backends.gdn_attn import GDNAttentionMetadataBuilder
127
from vllm.v1.attention.backends.mamba2_attn import Mamba2AttentionMetadataBuilder
128
from vllm.v1.attention.backends.utils import (
129
    NULL_BLOCK_ID,
130
    create_fast_prefill_custom_backend,
131
    get_dcp_local_seq_lens,
132
133
    reorder_batch_to_split_decodes_and_prefills,
)
134
from vllm.v1.core.sched.output import NewRequestData
135
from vllm.v1.cudagraph_dispatcher import CudagraphDispatcher
136
137
138
139
140
141
142
143
144
145
146
147
148
149
150
151
152
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,
153
    ECConnectorOutput,
154
    KVConnectorOutput,
155
156
157
158
159
    LogprobsLists,
    LogprobsTensors,
    ModelRunnerOutput,
    PoolerOutput,
    SamplerOutput,
160
    make_empty_encoder_model_runner_output,
161
)
162
from vllm.v1.pool.metadata import PoolingMetadata, PoolingStates
163
from vllm.v1.sample.logits_processor import LogitsProcessors, build_logitsprocs
164
from vllm.v1.sample.logits_processor.interface import LogitsProcessor
165
from vllm.v1.sample.metadata import SamplingMetadata
166
from vllm.v1.sample.rejection_sampler import RejectionSampler
167
from vllm.v1.sample.sampler import Sampler
168
from vllm.v1.spec_decode.dflash import DFlashProposer
169
from vllm.v1.spec_decode.draft_model import DraftModelProposer
170
from vllm.v1.spec_decode.eagle import EagleProposer
171
from vllm.v1.spec_decode.extract_hidden_states import ExtractHiddenStatesProposer
172
from vllm.v1.spec_decode.medusa import MedusaProposer
173
from vllm.v1.spec_decode.metadata import SpecDecodeMetadata
174
175
176
177
178
179
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,
)
180
from vllm.v1.spec_decode.suffix_decoding import SuffixDecodingProposer
181
from vllm.v1.spec_decode.utils import update_num_computed_tokens_for_batch_change
182
from vllm.v1.structured_output.utils import apply_grammar_bitmask
183
from vllm.v1.utils import CpuGpuBuffer, record_function_or_nullcontext
184
185
186
187
188
from vllm.v1.worker import mamba_utils
from vllm.v1.worker.cp_utils import (
    check_attention_cp_compatibility,
    get_total_cp_world_size,
)
189
from vllm.v1.worker.dp_utils import coordinate_batch_across_dp
190
from vllm.v1.worker.ec_connector_model_runner_mixin import ECConnectorModelRunnerMixin
191
from vllm.v1.worker.gpu.pool.late_interaction_runner import LateInteractionRunner
192
from vllm.v1.worker.gpu_input_batch import CachedRequestState, InputBatch
193
from vllm.v1.worker.gpu_ubatch_wrapper import UBatchWrapper
194
from vllm.v1.worker.kv_connector_model_runner_mixin import KVConnectorModelRunnerMixin
195
from vllm.v1.worker.lora_model_runner_mixin import LoRAModelRunnerMixin
196
197
198
from vllm.v1.worker.ubatch_utils import (
    UBatchSlices,
    check_ubatch_thresholds,
199
    maybe_create_ubatch_slices,
200
    split_attn_metadata,
201
)
202
from vllm.v1.worker.utils import is_residual_scattered_for_sp
203
from vllm.v1.worker.workspace import lock_workspace
204

205
206
from .utils import (
    AttentionGroup,
207
    KVBlockZeroer,
208
209
    add_kv_sharing_layers_to_kv_cache_groups,
    bind_kv_cache,
210
    prepare_kernel_block_sizes,
211
212
    sanity_check_mm_encoder_outputs,
)
213

214
if TYPE_CHECKING:
215
    from vllm.v1.core.sched.output import GrammarOutput, SchedulerOutput
216
    from vllm.v1.spec_decode.ngram_proposer import NgramProposer
217
    from vllm.v1.worker.encoder_cudagraph import EncoderCudaGraphManager
218
219
220

logger = init_logger(__name__)

221
222
AttnMetadataDict: TypeAlias = dict[str, AttentionMetadata]
# list when ubatching is enabled
223
PerLayerAttnMetadata: TypeAlias = list[AttnMetadataDict] | AttnMetadataDict
224

225

226
227
228
229
230
231
# Wrapper for ModelRunnerOutput to support overlapped execution.
class AsyncGPUModelRunnerOutput(AsyncModelRunnerOutput):
    def __init__(
        self,
        model_runner_output: ModelRunnerOutput,
        sampled_token_ids: torch.Tensor,
232
        logprobs_tensors: LogprobsTensors | None,
233
234
        invalid_req_indices: list[int],
        async_output_copy_stream: torch.cuda.Stream,
235
        vocab_size: int,
236
237
238
239
240
    ):
        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.
241
        self.async_copy_ready_event = torch.Event()
242
243
244
245

        # 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
246
        self.vocab_size = vocab_size
247
        self._logprobs_tensors = logprobs_tensors
248
249
250
251
252

        # 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)
253
            self.sampled_token_ids_cpu = self._sampled_token_ids.to(
254
255
                "cpu", non_blocking=True
            )
256
257
258
259
260
            self._logprobs_tensors_cpu = (
                self._logprobs_tensors.to_cpu_nonblocking()
                if self._logprobs_tensors
                else None
            )
261
            self.async_copy_ready_event.record()
262
263
264

    def get_output(self) -> ModelRunnerOutput:
        """Copy the device tensors to the host and return a ModelRunnerOutput.
265

266
267
        This function blocks until the copy is finished.
        """
268
        max_gen_len = self.sampled_token_ids_cpu.shape[-1]
269
        self.async_copy_ready_event.synchronize()
270

271
272
        # Release the device tensors once the copy has completed.
        del self._logprobs_tensors
273
        del self._sampled_token_ids
274
        if max_gen_len == 1:
275
            valid_sampled_token_ids = self.sampled_token_ids_cpu.tolist()
276
277
            for i in self._invalid_req_indices:
                valid_sampled_token_ids[i].clear()
278
279
280
            logprobs_lists = None
            if self._logprobs_tensors_cpu is not None:
                logprobs_lists = self._logprobs_tensors_cpu.tolists()
281
        else:
282
            valid_sampled_token_ids, logprobs_lists = RejectionSampler.parse_output(
283
284
                self.sampled_token_ids_cpu,
                self.vocab_size,
285
                self._invalid_req_indices,
286
                logprobs_tensors=self._logprobs_tensors_cpu,
287
            )
288
289
290

        output = self._model_runner_output
        output.sampled_token_ids = valid_sampled_token_ids
291
        output.logprobs = logprobs_lists
292
293
294
        return output


295
296
297
298
299
300
301
302
303
304
305
306
307
308
309
310
311
312
313
314
315
316
317
318
319
320
321
322
323
324
325
326
327
328
329
330
331
332
333
334
335
336
337
338
339
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


340
341
342
343
344
345
346
347
348
349
350
351
352
353
354
355
356
357
358
359
360
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)
361
362
363
            self._model_runner_output.pooler_output = _copy_pooler_output_to_cpu(
                raw_pooler_output=self._raw_pooler_output,
                finished_mask=finished_mask,
364
365
366
367
368
369
370
371
372
373
374
375
376
377
            )
            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


378
379
380
381
382
383
384
385
386
387
388
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
389
    ec_connector_output: ECConnectorOutput | None
390
    cudagraph_stats: CUDAGraphStat | None
391
    slot_mappings: dict[str, torch.Tensor] | list[dict[str, torch.Tensor]] | None
392
393


394
395
396
class GPUModelRunner(
    LoRAModelRunnerMixin, KVConnectorModelRunnerMixin, ECConnectorModelRunnerMixin
):
397
398
    def __init__(
        self,
399
        vllm_config: VllmConfig,
400
        device: torch.device,
401
    ):
402
403
404
        self.vllm_config = vllm_config
        self.model_config = vllm_config.model_config
        self.cache_config = vllm_config.cache_config
405
        self.offload_config = vllm_config.offload_config
406
        self.compilation_config = vllm_config.compilation_config
407
408
409
410
411
412
        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
413

414
415
416
417
        model_config = self.model_config
        cache_config = self.cache_config
        scheduler_config = self.scheduler_config
        parallel_config = self.parallel_config
418
        self.device = device
419
420
        self.pin_memory = is_pin_memory_available()
        self.dtype = self.model_config.dtype
421

422
423
424
        self.kv_cache_dtype = kv_cache_dtype_str_to_dtype(
            cache_config.cache_dtype, self.model_config
        )
425

426
        self.is_pooling_model = model_config.runner_type == "pooling"
427
        self.enable_prompt_embeds = model_config.enable_prompt_embeds
428
        self.is_multimodal_raw_input_only_model = (
429
430
            model_config.is_multimodal_raw_input_only_model
        )
431
        # These will be overridden in load_model()
432
        self.is_multimodal_pruning_enabled = False
433
        self.requires_sequential_video_encoding = False
434
435
436
        # Set to True after init_routed_experts_capturer() completes.
        # Prevents routed experts code from running during profiling/dummy run.
        self.routed_experts_initialized = False
437
        self.max_model_len = model_config.max_model_len
438
439
440

        # Always set to false after the first forward pass
        self.calculate_kv_scales = self.cache_config.calculate_kv_scales
441
        self.dcp_world_size = self.parallel_config.decode_context_parallel_size
442
        self.dcp_rank = 0 if self.dcp_world_size <= 1 else get_dcp_group().rank_in_group
443
        self.max_num_tokens = scheduler_config.max_num_batched_tokens
444
        self.max_num_reqs = scheduler_config.max_num_seqs
445

446
447
        # Broadcast PP output for external_launcher (torchrun)
        # to make sure we are synced across pp ranks
448
        # TODO: Support overlapping micro-batches
449
450
        # https://github.com/vllm-project/vllm/issues/18019
        self.broadcast_pp_output = (
451
            self.parallel_config.distributed_executor_backend == "external_launcher"
452
            and len(get_pp_group().ranks) > 1
453
        )
454

455
        # Model-related.
456
        self.num_query_heads = model_config.get_num_attention_heads(parallel_config)
457
        self.inputs_embeds_size = model_config.get_inputs_embeds_size()
458
        # Only relevant for models using ALiBi (e.g, MPT)
459
        self.use_alibi = model_config.uses_alibi
460

461
        self.cascade_attn_enabled = not self.model_config.disable_cascade_attn
462
        self.is_mm_prefix_lm = self.model_config.is_mm_prefix_lm
463

464
        # Multi-modal data support
465
        self.mm_registry = MULTIMODAL_REGISTRY
466
        self.uses_mrope = model_config.uses_mrope
467
        self.uses_xdrope_dim = model_config.uses_xdrope_dim
468
        self.supports_mm_inputs = self.mm_registry.supports_multimodal_inputs(
469
            model_config
470
        )
471

472
473
474
        if self.model_config.is_encoder_decoder:
            # Maximum length of the encoder input, only for encoder-decoder
            # models.
475
            self.max_encoder_len = scheduler_config.max_num_encoder_input_tokens
476
477
478
        else:
            self.max_encoder_len = 0

479
480
481
        # Async scheduling
        self.use_async_scheduling = self.scheduler_config.async_scheduling

482
        # Sampler
483
        self.sampler = Sampler(logprobs_mode=self.model_config.logprobs_mode)
484

485
        self.eplb_state: EplbState | None = None
486
487
        # NOTE(yongji): flag to temporarily disable EPLB during scaling up/down
        self.eep_eplb_suppressed = False
488
489
490
491
492
493
        """
        State of the expert parallelism load balancer.

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

494
        # Lazy initializations
495
        # self.model: nn.Module  # Set after load_model
496
        # Initialize in initialize_kv_cache
497
        self.kv_caches: list[torch.Tensor] = []
498
499
500
        # Initialize in initialize_kv_cache_tensors
        self.cross_layers_kv_cache: torch.Tensor | None = None
        self.cross_layers_attn_backend: type[AttentionBackend] | None = None
501
502
        # indexes: [kv_cache_group_id][attn_group]
        self.attn_groups: list[list[AttentionGroup]] = []
503
504
        # self.kv_cache_config: KVCacheConfig

505
506
        # mm_hash ->  encoder_output
        self.encoder_cache: dict[str, torch.Tensor] = {}
507
        self.late_interaction_runner = LateInteractionRunner()
508

509
510
511
        # Encoder CUDA graph manager (initialized after model load if enabled)
        self.encoder_cudagraph_manager: EncoderCudaGraphManager | None = None

512
        self.use_aux_hidden_state_outputs = False
513
514
515
516
517
        # 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:
518
            self.drafter: (
519
                NgramProposer  # noqa: F823
520
                | NgramProposerGPU
521
522
                | SuffixDecodingProposer
                | EagleProposer
523
                | DFlashProposer
524
525
                | DraftModelProposer
                | MedusaProposer
526
                | ExtractHiddenStatesProposer
527
            )
528
            if self.speculative_config.method == "ngram":
529
530
                from vllm.v1.spec_decode.ngram_proposer import NgramProposer

531
                self.drafter = NgramProposer(self.vllm_config)
532
533
534
535
536
537
            elif self.speculative_config.uses_draft_model():
                self.drafter = DraftModelProposer(
                    vllm_config=self.vllm_config,
                    device=self.device,
                    runner=self,
                )
538
539
540
541
542
543
544
545
546
547
548
549
550
551
552
553
554
            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
                )
555
556
557
            elif self.speculative_config.use_dflash():
                self.drafter = DFlashProposer(self.vllm_config, self.device, self)
                self.use_aux_hidden_state_outputs = True
558
559
            elif self.speculative_config.method == "suffix":
                self.drafter = SuffixDecodingProposer(self.vllm_config)
560
            elif self.speculative_config.use_eagle():
561
                self.drafter = EagleProposer(self.vllm_config, self.device, self)
562
                if self.speculative_config.method == "eagle3":
563
564
565
                    self.use_aux_hidden_state_outputs = (
                        self.drafter.eagle3_use_aux_hidden_state
                    )
566
567
            elif self.speculative_config.method == "medusa":
                self.drafter = MedusaProposer(
568
                    vllm_config=self.vllm_config, device=self.device
569
                )
570
571
572
573
574
            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
575
            else:
576
577
578
579
                raise ValueError(
                    "Unknown speculative decoding method: "
                    f"{self.speculative_config.method}"
                )
580
581
582
            self.rejection_sampler = RejectionSampler(
                self.sampler, self.speculative_config, self.device
            )
583

584
        self.num_spec_tokens = 0
585
        self.valid_sampled_token_count_gpu: torch.Tensor | None = None
586
587
        if self.speculative_config:
            self.num_spec_tokens = self.speculative_config.num_speculative_tokens
588
589
590
591
592
            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
593
594
595
        self.use_async_spec_decode = (
            self.use_async_scheduling and self.num_spec_tokens > 0
        )
596

597
        # Request states.
598
        self.requests: dict[str, CachedRequestState] = {}
599
600
601
        # NOTE(rob): num_prompt_logprobs only includes reqs
        # that are currently in the prefill phase.
        self.num_prompt_logprobs: dict[str, int] = {}
602

603
604
605
606
607
608
609
610
611
        # 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.
612
613
614
615
        logits_processors = model_config.logits_processors
        custom_logitsprocs: Sequence[str | type[LogitsProcessor]] = (
            tuple(logits_processors) if logits_processors is not None else ()
        )
616
617
618
619
620
        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]
621
622
        self.input_batch = InputBatch(
            max_num_reqs=self.max_num_reqs,
623
            # We need to use the encoder length for encoder-decoder
624
625
            # because of KV cache for cross-attention.
            max_model_len=max(self.max_model_len, self.max_encoder_len),
626
627
628
            max_num_batched_tokens=self.max_num_tokens,
            device=self.device,
            pin_memory=self.pin_memory,
629
            vocab_size=self.model_config.get_vocab_size(),
630
631
            block_sizes=[placeholder_block_size],
            kernel_block_sizes=[placeholder_block_size],
632
            is_spec_decode=bool(self.vllm_config.speculative_config),
633
            logitsprocs=build_logitsprocs(
634
635
636
                self.vllm_config,
                self.device,
                self.pin_memory,
637
                self.is_pooling_model,
638
                custom_logitsprocs,
639
            ),
640
641
            # We currently don't know whether a particular custom logits processor
            # uses output token ids so we set this conservatively.
642
643
644
645
            # ThinkingTokenBudgetLogitsProcessor also needs output token ids to
            # correctly track think start/end token sequences in async scheduling.
            logitsprocs_need_output_token_ids=bool(custom_logitsprocs)
            or self.vllm_config.reasoning_config is not None,
646
            is_pooling_model=self.is_pooling_model,
647
            cp_kv_cache_interleave_size=self.parallel_config.cp_kv_cache_interleave_size,
648
        )
649

650
651
652
653
654
        # 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.
655
        self.prepare_inputs_event: torch.Event | None = None
656
657
        if self.use_async_scheduling:
            self.async_output_copy_stream = torch.cuda.Stream()
658
            self.prepare_inputs_event = torch.Event()
659

660
661
662
663
664
665
666
667
668
669
670
        # 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 = []

671
        # Cache the device properties.
672
        self._init_device_properties()
673

674
675
676
677
        # Encoder timing registry for observability
        self.encoder_timing_registry: dict[str, EncoderTimingStats] = {}
        self._encoder_timing_lock = threading.Lock()

678
        # Persistent buffers for CUDA graphs.
679
        self.input_ids = self._make_buffer(self.max_num_tokens, dtype=torch.int32)
680
681
682
        self.positions = torch.zeros(
            self.max_num_tokens, dtype=torch.int64, device=self.device
        )
683
684
685
        self.query_start_loc = self._make_buffer(
            self.max_num_reqs + 1, dtype=torch.int32
        )
686
687
688
689
690
691
692
693
694
695
696
697
698
699
700
701
702
703
704
        self.seq_lens = torch.zeros(
            self.max_num_reqs, dtype=torch.int32, device=self.device
        )
        self.optimistic_seq_lens_cpu = torch.zeros(
            self.max_num_reqs, dtype=torch.int32, pin_memory=self.pin_memory
        )
        self.num_computed_tokens = torch.zeros(
            self.max_num_reqs, dtype=torch.int32, device=self.device
        )
        self.prev_num_draft_tokens = self._make_buffer(
            self.max_num_reqs, dtype=torch.int32
        )
        self.req_indices = self._make_buffer(self.max_num_tokens, dtype=torch.int64)
        # Maps current batch position -> previous batch position (-1 for new reqs)
        self.prev_positions = self._make_buffer(self.max_num_reqs, dtype=torch.int64)
        self.num_scheduled_tokens = self._make_buffer(
            self.max_num_reqs, dtype=torch.int32
        )

705
        self.encoder_seq_lens = self._make_buffer(self.max_num_reqs, dtype=torch.int32)
706
707
708
709
        if self.dcp_world_size > 1:
            self.dcp_local_seq_lens = self._make_buffer(
                self.max_num_reqs, dtype=torch.int32
            )
710
711
712
        # 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.
713
        self.inputs_embeds = self._make_buffer(
714
            self.max_num_tokens, self.inputs_embeds_size, dtype=self.dtype, numpy=False
715
716
        )
        self.is_token_ids = self._make_buffer(self.max_num_tokens, dtype=torch.bool)
717
718
        self.discard_request_mask = self._make_buffer(
            self.max_num_reqs, dtype=torch.bool
719
720
721
722
723
        )
        self.num_decode_draft_tokens = self._make_buffer(
            self.max_num_reqs, dtype=torch.int32
        )
        self.num_accepted_tokens = self._make_buffer(
724
            self.max_num_reqs, dtype=torch.int32
725
        )
726
727

        # Only relevant for models using M-RoPE (e.g, Qwen2-VL)
728
        if self.uses_mrope:
Roger Wang's avatar
Roger Wang committed
729
730
731
732
            # 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
733
734
735
736
737
738

            # 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
739
            self.mrope_positions = self._make_buffer(
740
741
                (3, self.max_num_tokens + 1), dtype=torch.int64
            )
742

743
744
745
746
747
748
749
        # 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
            )

750
        # None in the first PP rank. The rest are set after load_model.
751
        self.intermediate_tensors: IntermediateTensors | None = None
752

753
754
755
756
757
758
759
760
        # OPTIMIZATION: Cache the arange tensors rather than creating them
        # every step. Keep in int64 to avoid overflow with long context.
        # - arange_np: immutable [0, 1, 2, ...] used as source for batched computation
        # - query_pos: CpuGpuBuffer for the computed batched arange result
        arange_size = max(self.max_num_reqs + 1, self.max_num_tokens)
        self.arange_np = np.arange(arange_size, dtype=np.int64)
        self.query_pos = self._make_buffer(arange_size, dtype=torch.int64)
        self._arange_scratch = np.empty(arange_size, dtype=np.int64)
761

762
763
764
765
766
        # 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] = {}
767
768
769
770
771
        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(
772
773
                self.max_num_tokens, dtype=torch.int32, device=self.device
            )
774

775
        self.uniform_decode_query_len = 1 + self.num_spec_tokens
776
777
778
779

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

780
        self.mm_budget = (
781
            MultiModalBudget(self.vllm_config, self.mm_registry)
782
783
784
            if self.supports_mm_inputs
            else None
        )
785

786
        self.reorder_batch_threshold: int | None = None
787

788
789
790
791
792
        # 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()

793
        # Cached outputs.
794
        self._draft_token_ids: list[list[int]] | torch.Tensor | None = None
795
796
797
798
799
800
801
802
803
804
805
806
807
808
809
        # 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()

810
        self._draft_token_req_ids: list[str] | None = None
811
        self.transfer_event = torch.Event()
812
        self.sampled_token_ids_pinned_cpu = torch.empty(
813
            (self.max_num_reqs, 1),
814
815
            dtype=torch.int64,
            device="cpu",
816
817
            pin_memory=self.pin_memory,
        )
818

819
820
        # Pre-allocated tensor for copying valid sampled token counts to CPU,
        # with dedicated stream for overlapping and event for coordination.
821
        self.valid_sampled_token_count_event: torch.Event | None = None
822
        self.valid_sampled_token_count_copy_stream: torch.cuda.Stream | None = None
823
824
825
826
827
828
        # 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
829
        self.num_accepted_tokens_event: torch.Event | None = None
830
831
        if self.num_spec_tokens:
            self.draft_token_ids_event = torch.Event()
832
            self.num_accepted_tokens_event = torch.Event()
833
834
835
836
837
838
839
840
841
842
843
844
            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,
845
                    dtype=torch.int32,
846
847
848
                    device="cpu",
                    pin_memory=self.pin_memory,
                )
849

850
851
852
853
        # Model weight offloader
        # Make sure this is called before any get_offloader call
        set_offloader(create_offloader(self.offload_config))

854
855
        # Ephemeral state transferred between execute_model() and sample_tokens().
        self.execute_model_state: ExecuteModelState | None = None
856
        self.kv_connector_output: KVConnectorOutput | None = None
857
        self.mamba_state_idx: dict[str, int] = {}
858
        self._mamba_copy_bufs: mamba_utils.MambaCopyBuffers | None = None
859
        self.layerwise_nvtx_hooks_registered = False
860

861
862
863
864
865
866
867
    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

868
    def reset_mm_cache(self) -> None:
869
870
871
872
        """
        Clear the multi-modal cache that was used during profiling,
        but no longer needed during inference.
        """
873
874
        if self.mm_budget:
            self.mm_budget.reset_cache()
875
        self.late_interaction_runner.clear()
876

877
878
879
880
881
882
883
    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()
884
        self.late_interaction_runner.clear()
885

886
887
888
889
890
891
892
893
894
    @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.
        """
895
        if not is_quantized_kv_cache(self.cache_config.cache_dtype):
896
897
898
899
900
901
902
903
904
905
906
907
908
909
910
911
912
913
914
915
916
917
918
919
920
921
922
923
924
925
926
927
928
929
            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)

930
931
932
933
    def _get_positions(self, num_tokens: Any):
        if isinstance(num_tokens, int):
            if self.uses_mrope:
                return self.mrope_positions.gpu[:, :num_tokens]
934
935
            if self.uses_xdrope_dim > 0:
                return self.xdrope_positions.gpu[:, :num_tokens]
936
            return self.positions[:num_tokens]
937
938
939
        else:
            if self.uses_mrope:
                return self.mrope_positions.gpu[:, num_tokens]
940
941
            if self.uses_xdrope_dim > 0:
                return self.xdrope_positions.gpu[:, num_tokens]
942
            return self.positions[num_tokens]
943

944
    def _make_buffer(
945
        self, *size: int | torch.SymInt, dtype: torch.dtype, numpy: bool = True
946
947
948
949
950
951
952
953
    ) -> CpuGpuBuffer:
        return CpuGpuBuffer(
            *size,
            dtype=dtype,
            device=self.device,
            pin_memory=self.pin_memory,
            with_numpy=numpy,
        )
954

955
956
957
958
959
960
961
962
963
964
    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

965
    def _init_model_kwargs(self):
966
967
        model_kwargs = dict[str, Any]()

968
        if not self.is_pooling_model:
969
970
            return model_kwargs

971
972
        num_reqs = self.input_batch.num_reqs
        pooling_params = self.input_batch.get_pooling_params()
973
974
975

        token_type_id_requests = dict[int, Any]()
        for i, param in enumerate(pooling_params):
976
977
978
979
980
            if (
                param.extra_kwargs is not None
                and (token_types := param.extra_kwargs.get("compressed_token_type_ids"))
                is not None
            ):
981
982
983
984
985
                token_type_id_requests[i] = token_types

        if len(token_type_id_requests) == 0:
            return model_kwargs

986
        seq_lens = self.seq_lens[:num_reqs]
987
988
989
990
991
992
993
994
        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(
995
996
            device=self.device
        )
997
998
        return model_kwargs

999
    def _may_reorder_batch(self, scheduler_output: "SchedulerOutput") -> None:
1000
1001
        """
        Update the order of requests in the batch based on the attention
1002
        backend's needs. For example, some attention backends (namely MLA) may
1003
1004
1005
1006
1007
1008
        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
1009
        # Attention free models have zero kv_cache_groups, however models
1010
1011
1012
1013
1014
1015
1016
        # 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

1017
1018
1019
1020
        if self.reorder_batch_threshold is not None:
            reorder_batch_to_split_decodes_and_prefills(
                self.input_batch,
                scheduler_output,
1021
1022
                decode_threshold=self.reorder_batch_threshold,
            )
1023

1024
1025
1026
1027
1028
1029
1030
1031
1032
1033
1034
1035
1036
1037
1038
1039
1040
1041
1042
1043
    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)

1044
1045
    # Note: used for model runner override.
    def _init_device_properties(self) -> None:
1046
        """Initialize attributes from torch.cuda.get_device_properties"""
1047
1048

        self.num_sms = num_compute_units(self.device.index)
1049
1050
1051

    # Note: used for model runner override.
    def _sync_device(self) -> None:
1052
        torch.accelerator.synchronize()
1053

1054
1055
1056
1057
1058
1059
1060
    def _get_or_create_async_output_copy_stream(self) -> torch.cuda.Stream:
        stream = self.async_output_copy_stream
        if stream is None:
            stream = torch.cuda.Stream()
            self.async_output_copy_stream = stream
        return stream

1061
    def _update_states(self, scheduler_output: "SchedulerOutput") -> Callable | None:
1062
1063
1064
1065
1066
1067
        """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.

1068
1069
        The SamplingMetadata is updated and copied to the GPU if there is a
        new/resumed/paused/finished request in the batch.
1070
1071
        """
        # Remove finished requests from the cached states.
1072
1073
        for req_id in scheduler_output.finished_req_ids:
            self.requests.pop(req_id, None)
1074
            self.num_prompt_logprobs.pop(req_id, None)
1075
1076
1077
        self.late_interaction_runner.on_requests_finished(
            scheduler_output.finished_req_ids
        )
1078
1079
1080
1081
1082
1083
1084
        # 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:
1085
            self.input_batch.remove_request(req_id)
1086

1087
1088
1089
1090
1091
        # 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)

1092
        # Free the cached encoder outputs.
1093
1094
        for mm_hash in scheduler_output.free_encoder_mm_hashes:
            self.encoder_cache.pop(mm_hash, None)
1095

1096
1097
1098
1099
1100
1101
1102
        # 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()
1103
1104
1105
1106
1107
1108
1109
1110
        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)
1111
1112
1113
1114
1115
        # 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:
1116
            self.input_batch.remove_request(req_id)
1117

1118
1119
1120
1121
1122
1123
1124
        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] = []

1125
        reqs_to_add: list[CachedRequestState] = []
1126
1127
        deferred_spec_decode_corrections = []

1128
        # Add new requests to the cached states.
1129
1130
        for new_req_data in scheduler_output.scheduled_new_reqs:
            req_id = new_req_data.req_id
1131
1132
1133
1134
1135
1136
            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

1137
            sampling_params = new_req_data.sampling_params
1138
            pooling_params = new_req_data.pooling_params
1139

1140
1141
1142
1143
            if (
                sampling_params
                and sampling_params.sampling_type == SamplingType.RANDOM_SEED
            ):
1144
1145
1146
1147
1148
                generator = torch.Generator(device=self.device)
                generator.manual_seed(sampling_params.seed)
            else:
                generator = None

1149
1150
            if self.is_pooling_model:
                assert pooling_params is not None
1151
1152
                task = pooling_params.task
                assert task is not None, "You did not set `task` in the API"
1153

1154
                model = cast(VllmModelForPooling, self.get_model())
1155
                to_update = model.pooler.get_pooling_updates(task)
1156
1157
                to_update.apply(pooling_params)

1158
            req_state = CachedRequestState(
1159
                req_id=req_id,
1160
                prompt_token_ids=new_req_data.prompt_token_ids,
1161
                prompt_embeds=new_req_data.prompt_embeds,
1162
                mm_features=new_req_data.mm_features,
1163
                sampling_params=sampling_params,
1164
                pooling_params=pooling_params,
1165
                generator=generator,
1166
1167
                block_ids=new_req_data.block_ids,
                num_computed_tokens=new_req_data.num_computed_tokens,
1168
                output_token_ids=[],
1169
                lora_request=new_req_data.lora_request,
1170
            )
1171
            self.requests[req_id] = req_state
1172
            self.late_interaction_runner.register_request(req_id, pooling_params)
1173

1174
1175
1176
1177
1178
1179
1180
            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
                )

1181
            # Only relevant for models using M-RoPE (e.g, Qwen2-VL)
1182
            if self.uses_mrope:
1183
                self._init_mrope_positions(req_state)
1184

1185
1186
1187
1188
            # Only relevant for models using XD-RoPE (e.g, HunYuan-VL)
            if self.uses_xdrope_dim > 0:
                self._init_xdrope_positions(req_state)

1189
            reqs_to_add.append(req_state)
1190
1191
1192
            # Track new requests for ngram_gpu full tensor copy
            if is_ngram_gpu:
                ngram_gpu_new_reqs.append(req_state)
1193

1194
        # Update the states of the running/resumed requests.
1195
        is_last_rank = get_pp_group().is_last_rank
1196
        req_data = scheduler_output.scheduled_cached_reqs
1197
        scheduled_spec_tokens = scheduler_output.scheduled_spec_decode_tokens
1198

1199
1200
1201
1202
1203
1204
1205
1206
1207
1208
1209
1210
1211
1212
1213
        # 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,
            )
1214
1215
        if self.use_async_spec_decode:
            self.prev_num_draft_tokens.np.fill(0)
1216

1217
        for i, req_id in enumerate(req_data.req_ids):
1218
            req_state = self.requests[req_id]
1219
1220
            num_computed_tokens = req_data.num_computed_tokens[i]
            new_block_ids = req_data.new_block_ids[i]
1221
            resumed_from_preemption = req_id in req_data.resumed_req_ids
1222
            num_output_tokens = req_data.num_output_tokens[i]
1223
            req_index = self.input_batch.req_id_to_index.get(req_id)
1224

1225
1226
1227
1228
            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:
1229
                # first step: num_computed_tokens = 0, spec_tokens = [],
1230
                # prev_num_draft_len = 0.
Jiayi Yan's avatar
Jiayi Yan committed
1231
                # second step: num_computed_tokens = 100(prompt length),
1232
1233
1234
                # 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.
1235
                # num_computed_tokens in first step and second step doesn't contain
1236
1237
1238
                # 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.
1239
1240
1241
                if req_index is None:
                    req_state.prev_num_draft_len = 0
                else:
1242
1243
1244
1245
1246
1247
1248
1249
1250
1251
1252
1253
1254
1255
1256
1257
1258
1259
1260
                    # Optimistically assume all accepted; queue up a correction
                    # to be called after the model forward to preserve async
                    # scheduling. Corrected on GPU in _prepare_inputs.
                    optimistic_num_accepted = req_state.prev_num_draft_len
                    req_state.output_token_ids.extend([-1] * optimistic_num_accepted)

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

                    prev_req_index = (
                        self.input_batch.prev_req_id_to_index.get(req_id)
                        if self.input_batch.prev_req_id_to_index
                        else None
                    )
                    if prev_req_index is not None:
                        self.prev_num_draft_tokens.np[prev_req_index] = (
                            optimistic_num_accepted
                        )
1261

1262
1263
1264
1265
                    if is_ngram_gpu and optimistic_num_accepted > 0:
                        self.input_batch.num_tokens_no_spec[req_index] += (
                            optimistic_num_accepted
                        )
1266

1267
            # Update the cached states.
1268
            req_state.num_computed_tokens = num_computed_tokens
1269
1270

            if not is_last_rank:
1271
1272
1273
1274
1275
1276
1277
1278
1279
1280
1281
1282
1283
1284
1285
1286
1287
1288
1289
1290
                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:]
                        )
1291
1292
            elif num_output_tokens < len(req_state.output_token_ids):
                # Some output tokens were discarded due to a sync-KV-load
1293
1294
                # failure, or output_token_ids was inflated by the optimistic
                # extend above (async spec decode). Align the cached state.
1295
1296
                del req_state.output_token_ids[num_output_tokens:]
                if req_index is not None:
1297
1298
1299
1300
                    end_idx = (
                        self.input_batch.num_prompt_tokens[req_index]
                        + num_output_tokens
                    )
1301
                    self.input_batch.num_tokens_no_spec[req_index] = end_idx
1302

1303
            # Update the block IDs.
1304
            if not resumed_from_preemption:
1305
1306
                if new_block_ids is not None:
                    # Append the new blocks to the existing block IDs.
1307
                    for block_ids, new_ids in zip(req_state.block_ids, new_block_ids):
1308
                        block_ids.extend(new_ids)
1309
            else:
1310
                assert req_index is None
1311
                assert new_block_ids is not None
1312
1313
                # The request is resumed from preemption.
                # Replace the existing block IDs with the new ones.
1314
                req_state.block_ids = new_block_ids
1315
1316
1317
1318
1319

            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.
1320
1321
1322
1323
1324
1325
1326

                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:]

1327
                reqs_to_add.append(req_state)
1328
1329
1330
                # Track resumed requests for ngram_gpu full tensor copy
                if is_ngram_gpu:
                    ngram_gpu_new_reqs.append(req_state)
1331
1332
1333
                continue

            # Update the persistent batch.
1334
            self.input_batch.num_computed_tokens_cpu[req_index] = num_computed_tokens
1335
            if new_block_ids is not None:
1336
                self.input_batch.block_table.append_row(new_block_ids, req_index)
1337
1338
1339
1340
1341
1342
1343

            # 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)
1344
                self.input_batch.token_ids_cpu[
1345
1346
1347
                    req_index, start_token_index:end_token_index
                ] = new_token_ids
                self.input_batch.num_tokens_no_spec[req_index] = end_token_index
1348

1349
            # Add spec_token_ids to token_ids_cpu.
1350
            self.input_batch.update_req_spec_token_ids(req_state, scheduled_spec_tokens)
1351
1352
1353
1354
1355
            # 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
1356

1357
1358
        # Add the new or resumed requests to the persistent batch.
        # The smaller empty indices are filled first.
1359
1360
        for request in reqs_to_add:
            self.input_batch.add_request(request)
1361
            self.input_batch.update_req_spec_token_ids(request, scheduled_spec_tokens)
1362

1363
1364
1365
1366
1367
1368
        # 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()
1369

1370
1371
1372
1373
1374
1375
1376
1377
1378
1379
1380
1381
        # 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,
            )

1382
1383
1384
1385
1386
1387
1388
1389
1390
1391
1392
1393
1394
1395
1396
1397
1398
1399
1400
1401
1402
1403
1404
1405
1406
1407
1408
1409
1410
1411
1412
1413
1414
1415
        if deferred_spec_decode_corrections:

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

            return correct_spec_decode_token_counts
        else:
            return None

1416
    def _update_states_after_model_execute(
1417
        self, output_token_ids: torch.Tensor, scheduler_output: "SchedulerOutput"
1418
    ) -> None:
1419
1420
1421
1422
1423
1424
1425
1426
        """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.
        """
1427
        if not self.speculative_config or not self.model_config.is_hybrid:
1428
1429
            return

1430
1431
1432
        # TODO: Remove .cpu() sync to enable fully async for hybrid model;
        # Use num_computed_tokens.gpu instead of req.num_computed_tokens to
        # support aligned mamba cache mode.
1433
1434
1435
        # Count the number of accepted tokens for each sequence.
        # Valid tokens are contiguous from position 0, so counting non-(-1)
        # tokens gives us the first -1 position (i.e., number of accepted).
1436
        num_reqs = output_token_ids.size(0)
1437
        self.num_accepted_tokens.gpu[:num_reqs] = (output_token_ids != -1).sum(dim=1)
1438

1439
        if self.cache_config.mamba_cache_mode == "align":
1440
1441
1442
1443
            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
1444
1445
1446
1447
1448
1449
1450
1451
            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(),
1452
                self._get_mamba_copy_bufs(),
1453
            )
1454
1455
1456
1457
        else:
            self.input_batch.num_accepted_tokens_cpu_tensor[:num_reqs].copy_(
                self.num_accepted_tokens.gpu[:num_reqs], non_blocking=True
            )
1458
1459
            assert self.num_accepted_tokens_event is not None
            self.num_accepted_tokens_event.record()
1460

1461
1462
1463
1464
1465
1466
1467
1468
1469
1470
1471
1472
1473
1474
1475
1476
1477
1478
1479
    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
1480
        self.late_interaction_runner.register_request(req_id, req_state.pooling_params)
1481
1482
1483
1484
1485
1486
1487
1488
1489
1490
1491
1492
1493
1494
1495
        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

1496
    def _init_mrope_positions(self, req_state: CachedRequestState):
1497
1498
        model = self.get_model()
        assert supports_mrope(model), "M-RoPE support is not implemented."
1499
1500
1501
1502
        assert req_state.prompt_token_ids is not None, (
            "M-RoPE requires prompt_token_ids to be available."
        )
        mrope_model = cast(SupportsMRoPE, model)
1503
1504

        req_state.mrope_positions, req_state.mrope_position_delta = (
1505
            mrope_model.get_mrope_input_positions(
1506
                req_state.prompt_token_ids,
1507
                req_state.mm_features,
1508
            )
1509
        )
1510

1511
1512
1513
1514
1515
1516
1517
1518
1519
1520
1521
1522
1523
    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,
        )

1524
    def _extract_mm_kwargs(
1525
        self,
1526
1527
        scheduler_output: "SchedulerOutput",
    ) -> BatchedTensorInputs:
1528
        if not scheduler_output or not self.is_multimodal_raw_input_only_model:
1529
            return {}
1530

1531
        mm_kwargs = list[tuple[str, MultiModalKwargsItem]]()
1532
        for req in scheduler_output.scheduled_new_reqs:
1533
1534
            for feature in req.mm_features:
                if feature.data is not None:
1535
                    mm_kwargs.append((feature.modality, feature.data))
1536

1537
1538
        # Input all modalities at once
        mm_kwargs_combined: BatchedTensorInputs = {}
1539
        for _, _, mm_kwargs_batch in group_and_batch_mm_kwargs(
1540
1541
1542
            mm_kwargs,
            device=self.device,
            pin_memory=self.pin_memory,
1543
        ):
1544
            mm_kwargs_combined.update(mm_kwargs_batch)
1545

1546
        return mm_kwargs_combined
1547

1548
    def _dummy_mm_kwargs(self, num_seqs: int) -> BatchedTensorInputs:
1549
        if not self.is_multimodal_raw_input_only_model:
1550
            return {}
1551

1552
1553
1554
        mm_budget = self.mm_budget
        assert mm_budget is not None

1555
1556
1557
        if not mm_budget.mm_max_toks_per_item:
            return {}  # No tower modalities (embed-only mode)

1558
1559
        dummy_modality = mm_budget.get_modality_with_max_tokens()
        return self._get_mm_dummy_batch(dummy_modality, num_seqs)
1560

1561
1562
1563
    def _get_cumsum_and_arange(
        self,
        num_tokens: np.ndarray,
1564
        arange_out: np.ndarray,
1565
        cumsum_dtype: np.dtype | None = None,
1566
    ) -> np.ndarray:
1567
        """Get the cumulative sum and batched arange of the given array.
1568
1569
1570
1571
        E.g., [2, 5, 3] -> [2, 7, 10], arange written to
        arange_out[:10] as [0, 1, 0, 1, 2, 3, 4, 0, 1, 2].
        Equivalent to but faster than:
        np.concatenate([np.arange(n) for n in num_tokens])
1572
1573
1574
1575
1576
1577
1578
        """
        # 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]
1579
1580
1581
1582
1583
1584
1585
1586
1587
1588
1589
1590
1591
1592
1593
1594
1595
1596
1597
        np.subtract(
            self.arange_np[:total_num_tokens],
            cumsums_offsets,
            out=arange_out[:total_num_tokens],
        )

        return cu_num_tokens

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

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

        if not prev_req_id_to_index:
            prev_positions.fill(-1)
            return
1598

1599
1600
        for i, req_id in enumerate(self.input_batch.req_ids[:num_reqs]):
            prev_positions[i] = prev_req_id_to_index.get(req_id, -1)
1601

1602
    def _prepare_input_ids(
1603
1604
        self,
        scheduler_output: "SchedulerOutput",
1605
        num_reqs: int,
1606
1607
        total_num_scheduled_tokens: int,
        cu_num_tokens: np.ndarray,
1608
    ) -> None:
1609
        """Prepare the input IDs for the current batch.
1610

1611
1612
        Carefully handles the `prev_sampled_token_ids` which can be cached
        from the previous engine iteration, in which case those tokens on the
1613
1614
1615
1616
1617
        GPU need to be copied into the corresponding slots into input_ids.

        Uses self.prev_positions[:num_reqs] which maps current pos -> prev pos
        (-1 for new requests).
        """
1618
1619
1620
1621

        if self.input_batch.prev_sampled_token_ids is None:
            # Normal scheduling case
            self.input_ids.copy_to_gpu(total_num_scheduled_tokens)
1622
1623
1624
            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)
1625
1626
1627
1628
1629
            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.
1630
1631
        prev_positions = self.prev_positions.np[:num_reqs]
        scheduled_spec_tokens = scheduler_output.scheduled_spec_decode_tokens
1632
1633
1634
        sample_flattened_indices: list[int] = []
        spec_flattened_indices: list[int] = []
        prev_draft_token_indices: list[int] = []
1635
1636
        prev_indices: list[int] = []
        common_indices_match = True
1637
        max_flattened_index = -1
1638
1639
        total_num_spec_tokens = 0

1640
1641
1642
1643
1644
1645
1646
1647
1648
1649
1650
1651
1652
1653
1654
1655
1656
1657
1658
1659
1660
1661
1662
1663
1664
1665
1666
1667
1668
        for cur_index in range(num_reqs):
            prev_index = prev_positions[cur_index]
            if prev_index < 0:
                continue
            prev_indices.append(prev_index)
            req_id = self.input_batch.req_ids[cur_index]
            # We need to compute the flattened input_ids index of the
            # last token in each common request.
            draft_len = len(scheduled_spec_tokens.get(req_id, ()))
            total_num_spec_tokens += draft_len
            flattened_index = cu_num_tokens[cur_index].item() - 1
            # example: cu_num_tokens = [2, 5, 8], draft_tokens = [1, 2, 2]
            # sample_flattened_indices = [0, 2, 5]
            # spec_flattened_indices = [1,   3, 4,    6, 7]
            sample_flattened_indices.append(flattened_index - draft_len)
            spec_flattened_indices.extend(
                range(flattened_index - draft_len + 1, flattened_index + 1)
            )
            start = prev_index * self.num_spec_tokens
            # prev_draft_token_indices is used to find which draft_tokens_id
            # should be copied to input_ids
            # example: prev draft_tokens_id [[1,2], [3,4], [5, 6]]
            # flatten draft_tokens_id [1,2,3,4,5,6]
            # draft_len of each request [1, 2, 1]
            # then prev_draft_token_indices is [0,   2, 3,   4]
            prev_draft_token_indices.extend(range(start, start + draft_len))
            common_indices_match &= prev_index == flattened_index
            max_flattened_index = max(max_flattened_index, flattened_index)

Jiayi Yan's avatar
Jiayi Yan committed
1669
        num_common_tokens = len(sample_flattened_indices)
1670
        total_without_spec = total_num_scheduled_tokens - total_num_spec_tokens
Jiayi Yan's avatar
Jiayi Yan committed
1671
        if num_common_tokens < total_without_spec:
1672
            # If not all requests are decodes from the last iteration,
1673
            # we need to copy the input_ids_cpu to the GPU first.
1674
            self.input_ids.copy_to_gpu(total_num_scheduled_tokens)
1675
1676
1677
            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
1678
        if num_common_tokens == 0:
1679
            # No requests in common with the previous iteration
1680
            # So input_ids.cpu will have all the input ids.
1681
            return
1682
        if common_indices_match and max_flattened_index == (num_common_tokens - 1):
1683
1684
1685
1686
            # 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
1687
1688
            self.input_ids.gpu[:num_common_tokens].copy_(
                self.input_batch.prev_sampled_token_ids[:num_common_tokens, 0],
1689
1690
                non_blocking=True,
            )
1691
            if self.enable_prompt_embeds:
Jiayi Yan's avatar
Jiayi Yan committed
1692
                self.is_token_ids.gpu[:num_common_tokens] = True
1693
            return
1694
        # Upload the index tensors asynchronously so the scatter can be non-blocking.
1695
1696
        sampled_tokens_index_tensor = torch.tensor(
            sample_flattened_indices, dtype=torch.int64, pin_memory=self.pin_memory
1697
        ).to(self.device, non_blocking=True)
1698
        prev_common_req_indices_tensor = torch.tensor(
1699
            prev_indices, dtype=torch.int64, pin_memory=self.pin_memory
1700
        ).to(self.device, non_blocking=True)
1701
1702
        self.input_ids.gpu.scatter_(
            dim=0,
1703
            index=sampled_tokens_index_tensor,
1704
            src=self.input_batch.prev_sampled_token_ids[
1705
1706
1707
                prev_common_req_indices_tensor, 0
            ],
        )
1708

1709
1710
1711
1712
1713
1714
1715
1716
1717
1718
1719
1720
1721
1722
1723
1724
1725
1726
1727
1728
1729
1730
        # 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],
        )

1731
1732
    def _get_encoder_seq_lens(
        self,
1733
        num_scheduled_tokens: dict[str, int],
1734
1735
        kv_cache_spec: KVCacheSpec,
        num_reqs: int,
1736
        for_cudagraph_capture: bool = False,
1737
    ) -> tuple[torch.Tensor | None, np.ndarray | None]:
1738
        if not isinstance(kv_cache_spec, CrossAttentionSpec):
1739
            return None, None
1740

1741
1742
        # Zero out buffer for padding requests that are not actually scheduled (CGs)
        self.encoder_seq_lens.np[:num_reqs] = 0
1743

1744
1745
        # Build encoder_seq_lens array mapping request indices to
        # encoder lengths for inputs scheduled in this batch
1746
        for req_id in num_scheduled_tokens:
1747
            req_index = self.input_batch.req_id_to_index[req_id]
1748
1749
1750
1751
1752
1753
1754
1755
1756
1757
1758
1759
            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
1760
1761
1762
1763
1764
1765
1766
1767
1768
        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
1769
1770
1771
1772

        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]
1773

1774
        return encoder_seq_lens, encoder_seq_lens_cpu
1775

1776
    def _prepare_inputs(
1777
1778
1779
        self,
        scheduler_output: "SchedulerOutput",
        num_scheduled_tokens: np.ndarray,
1780
1781
    ) -> tuple[
        torch.Tensor,
1782
        SpecDecodeMetadata | None,
1783
    ]:
1784
1785
        """
        :return: tuple[
1786
            logits_indices, spec_decode_metadata,
1787
1788
        ]
        """
1789
1790
1791
1792
1793
1794
1795
        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.
1796
        self.input_batch.block_table.commit_block_table(num_reqs)
1797
1798
1799

        # Get request indices.
        # E.g., [2, 5, 3] -> [0, 0, 1, 1, 1, 1, 1, 2, 2, 2]
1800
        req_indices = np.repeat(self.arange_np[:num_reqs], num_scheduled_tokens)
1801

1802
        # cu_num_tokens: [2, 5, 3] -> [2, 7, 10]
1803
1804
1805
1806
        # self.query_pos.np[:10]: [0, 1, 0, 1, 2, 3, 4, 0, 1, 2]
        cu_num_tokens = self._get_cumsum_and_arange(
            num_scheduled_tokens, self.query_pos.np
        )
1807
1808

        # Get positions.
1809
1810
1811
        positions_np = (
            self.input_batch.num_computed_tokens_cpu[req_indices]
            + self.query_pos.np[: cu_num_tokens[-1]]
1812
        )
1813

1814
1815
        # Calculate M-RoPE positions.
        # Only relevant for models using M-RoPE (e.g, Qwen2-VL)
1816
        if self.uses_mrope:
1817
1818
            self._calc_mrope_positions(scheduler_output)

1819
1820
1821
1822
1823
        # 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)

1824
1825
1826
1827
        # 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.
1828
1829
1830
        token_indices = (
            positions_np + req_indices * self.input_batch.token_ids_cpu.shape[1]
        )
1831
        token_indices_tensor = torch.from_numpy(token_indices)
1832

1833
1834
1835
        # 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.
1836
1837
1838
1839
1840
1841
        torch.index_select(
            self.input_batch.token_ids_cpu_tensor.flatten(),
            0,
            token_indices_tensor,
            out=self.input_ids.cpu[:total_num_scheduled_tokens],
        )
1842
        if self.enable_prompt_embeds:
1843
            is_token_ids = self.input_batch.is_token_ids_tensor.flatten()
1844
1845
1846
1847
            torch.index_select(
                is_token_ids,
                0,
                token_indices_tensor,
1848
1849
                out=self.is_token_ids.cpu[:total_num_scheduled_tokens],
            )
1850
1851
1852
1853
1854
1855
1856
1857
1858
1859
1860
1861
1862
1863
1864
1865
1866
1867
1868
1869
1870
1871
1872
1873
1874
1875
1876
1877
1878
1879
1880
1881
1882

        # 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:
1883
1884
1885
                    self.inputs_embeds.cpu[
                        output_idx : output_idx + actual_num_sched
                    ].copy_(req_embeds[start_pos:actual_end])
1886
1887

                output_idx += num_sched
1888
1889

        # Prepare the attention metadata.
1890
        self.query_start_loc.np[0] = 0
1891
        self.query_start_loc.np[1 : num_reqs + 1] = cu_num_tokens
1892
1893
        # Note: pad query_start_loc to be non-decreasing, as kernels
        # like FlashAttention requires that
1894
        self.query_start_loc.np[num_reqs + 1 :].fill(cu_num_tokens[-1])
1895
        self.query_start_loc.copy_to_gpu()
1896
        query_start_loc = self.query_start_loc.gpu[: num_reqs + 1]
1897

1898
1899
1900
1901
1902
1903
1904
1905
        # Compute optimistic seq_lens (assumes all draft tokens from previous
        # iteration accepted). Store in optimistic_seq_lens_cpu for use by
        # _build_attention_metadata (max_seq_len) and discard_request_mask.
        # seq_lens (GPU) will be computed later using the same optimistic values.
        torch.add(
            self.input_batch.num_computed_tokens_cpu_tensor[:num_reqs],
            torch.from_numpy(num_scheduled_tokens),
            out=self.optimistic_seq_lens_cpu[:num_reqs],
1906
        )
1907
1908
1909
1910
1911
1912
        self.optimistic_seq_lens_cpu[num_reqs:].fill_(0)

        # Build prev_positions mapping: current pos -> prev pos (-1 if new).
        # Used for gathering from previous iteration's GPU tensors.
        prev_req_id_to_index = self.input_batch.prev_req_id_to_index
        self._compute_prev_positions(num_reqs)
1913

1914
        num_tokens = [self.requests[r].num_tokens for r in self.input_batch.req_ids]
1915
1916
        num_tokens_np = np.array(num_tokens, dtype=np.int32)

1917
        # Record which requests should not be sampled,
1918
        # so that we could clear the sampled tokens before returning
1919
        self.discard_request_mask.np[:num_reqs] = (
1920
            self.optimistic_seq_lens_cpu[:num_reqs].numpy() < num_tokens_np
1921
        )
1922
        self.discard_request_mask.copy_to_gpu(num_reqs)
1923

1924
1925
1926
1927
        # Sync num_accepted_tokens from CPU (set by
        # _update_states_after_model_execute for hybrid models).
        if self.num_accepted_tokens_event is not None:
            self.num_accepted_tokens_event.synchronize()
1928
1929
1930
1931
1932
1933
1934
1935
1936
1937
1938
1939
1940
1941
1942
1943
1944
1945
            # Async mode: condense() reordered indices, use prev_positions mapping
            if self.use_async_scheduling and prev_req_id_to_index:
                prev_idx = self.prev_positions.np[:num_reqs]
                new_mask = prev_idx < 0
                self.num_accepted_tokens.np[:num_reqs] = (
                    self.input_batch.num_accepted_tokens_cpu[
                        np.where(new_mask, 0, prev_idx)
                    ]
                )
                self.num_accepted_tokens.np[:num_reqs][new_mask] = 1
                self.input_batch.num_accepted_tokens_cpu[:num_reqs] = (
                    self.num_accepted_tokens.np[:num_reqs]
                )
            else:
                # Non-async mode: use values directly
                self.num_accepted_tokens.np[:num_reqs] = (
                    self.input_batch.num_accepted_tokens_cpu[:num_reqs]
                )
1946
1947
1948
1949
1950
1951
1952
1953
1954
1955
1956
1957
1958
1959
1960
1961
1962
1963
1964
1965
1966
1967
1968
1969
1970
1971
1972
1973
1974
1975
1976
1977
1978
1979
1980
1981
1982
1983
1984
1985
1986
1987
1988
1989
1990
1991
1992
1993
1994
1995
1996
1997
1998
1999
2000
2001
2002
            self.num_accepted_tokens.np[num_reqs:].fill(1)
            self.num_accepted_tokens.copy_to_gpu()
        else:
            self.num_accepted_tokens.np.fill(1)
            self.num_accepted_tokens.gpu.fill_(1)

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

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

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

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

2003
        # Copy the tensors to the GPU.
2004
2005
        self._prepare_input_ids(
            scheduler_output,
2006
            num_reqs,
2007
2008
2009
            total_num_scheduled_tokens,
            cu_num_tokens,
        )
2010

2011
        if self.uses_mrope:
2012
            # Only relevant for models using M-RoPE (e.g, Qwen2-VL)
2013
2014
            self.mrope_positions.gpu[:, :total_num_scheduled_tokens].copy_(
                self.mrope_positions.cpu[:, :total_num_scheduled_tokens],
2015
2016
                non_blocking=True,
            )
2017
2018
2019
2020
2021
2022
        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,
            )
2023
2024
2025
2026
2027
2028
2029
2030
        if self.use_async_spec_decode and (self.uses_mrope or self.uses_xdrope_dim > 0):
            drift = self.num_computed_tokens[req_indices_gpu].to(
                torch.int64
            ) - self.input_batch.num_computed_tokens_cpu_tensor[req_indices].to(
                device=self.device, dtype=torch.int64, non_blocking=True
            )
            target = self.mrope_positions if self.uses_mrope else self.xdrope_positions
            target.gpu[:, :total_num_scheduled_tokens] += drift
2031

2032
        use_spec_decode = len(scheduler_output.scheduled_spec_decode_tokens) > 0
2033
2034
2035
2036
2037
2038
2039
2040
        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
2041
            num_sampled_tokens = np.ones(num_reqs, dtype=np.int32)
2042
2043
2044
2045
2046
        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)
2047
2048
2049
            # 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)
2050
2051
2052
2053
            for (
                req_id,
                draft_token_ids,
            ) in scheduler_output.scheduled_spec_decode_tokens.items():
2054
                req_idx = self.input_batch.req_id_to_index[req_id]
2055
2056
                draft_len = len(draft_token_ids)
                num_draft_tokens[req_idx] = draft_len
2057
2058
2059
2060
                if (
                    self.input_batch.num_computed_tokens_cpu[req_idx]
                    >= self.input_batch.num_prompt_tokens[req_idx]
                ):
2061
                    num_decode_draft_tokens[req_idx] = draft_len
2062
            spec_decode_metadata = self._calc_spec_decode_metadata(
2063
2064
                num_draft_tokens, cu_num_tokens
            )
2065
            logits_indices = spec_decode_metadata.logits_indices
2066
            num_sampled_tokens = num_draft_tokens + 1
2067
            # For DECODE only cuda graph of some attention backends (e.g., GDN).
2068
            self.num_decode_draft_tokens.np[:num_reqs] = num_decode_draft_tokens
2069
2070
            self.num_decode_draft_tokens.np[num_reqs:].fill(-1)
            self.num_decode_draft_tokens.copy_to_gpu()
2071

2072
2073
2074
2075
2076
        # Hot-Swap lora model
        if self.lora_config:
            assert (
                np.sum(num_sampled_tokens)
                <= self.vllm_config.scheduler_config.max_num_batched_tokens
2077
            )
2078
2079
2080
2081
2082
2083
2084
2085
2086
2087
2088
            self.set_active_loras(
                self.input_batch, num_scheduled_tokens, num_sampled_tokens
            )

        return (
            logits_indices,
            spec_decode_metadata,
        )

    def _build_attention_metadata(
        self,
2089
        num_tokens: int,
2090
        num_reqs: int,
2091
2092
2093
        max_query_len: int,
        num_tokens_padded: int | None = None,
        num_reqs_padded: int | None = None,
2094
2095
2096
2097
        ubatch_slices: UBatchSlices | None = None,
        logits_indices: torch.Tensor | None = None,
        use_spec_decode: bool = False,
        for_cudagraph_capture: bool = False,
2098
        num_scheduled_tokens: dict[str, int] | None = None,
2099
        cascade_attn_prefix_lens: list[list[int]] | None = None,
2100
        slot_mappings: dict[int, torch.Tensor] | None = None,
2101
2102
2103
2104
    ) -> tuple[PerLayerAttnMetadata, CommonAttentionMetadata | None]:
        """
        :return: tuple[attn_metadata, spec_decode_common_attn_metadata]
        """
2105
2106
2107
2108
        # Attention metadata is not needed for attention free models
        if len(self.kv_cache_config.kv_cache_groups) == 0:
            return {}, None

2109
2110
        num_tokens_padded = num_tokens_padded or num_tokens
        num_reqs_padded = num_reqs_padded or num_reqs
2111
        assert num_reqs_padded is not None and num_tokens_padded is not None
2112

2113
2114
2115
        attn_metadata: PerLayerAttnMetadata = {}
        if ubatch_slices is not None:
            attn_metadata = [dict() for _ in range(len(ubatch_slices))]
2116

2117
2118
2119
2120
2121
2122
        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:
2123
            max_seq_len = self.optimistic_seq_lens_cpu.numpy()[:num_reqs].max().item()
2124

2125
        kv_cache_groups = self.kv_cache_config.kv_cache_groups
2126

2127
        def _get_block_table(kv_cache_gid: int):
2128
2129
2130
            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):
2131
                blk_table_tensor = torch.zeros(
2132
                    (num_reqs_padded, 1),
2133
                    dtype=torch.int32,
2134
2135
                    device=self.device,
                )
2136
            else:
2137
                blk_table = self.input_batch.block_table[kv_cache_gid]
2138
                blk_table_tensor = blk_table.get_device_tensor(num_reqs_padded)
2139

2140
2141
2142
            # Fill unused block table entries with NULL_BLOCK_ID (null block)
            # for CUDAGraph padding. Block 0 is reserved for padding.
            blk_table_tensor[num_reqs:num_reqs_padded].fill_(NULL_BLOCK_ID)
2143
            return blk_table_tensor
2144

2145
2146
2147
        assert slot_mappings is not None
        block_table_gid_0 = _get_block_table(0)
        slot_mapping_gid_0 = slot_mappings[0]
2148

2149
2150
2151
2152
        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()
2153
2154
2155
2156
2157
2158
        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
        ]
2159
        seq_lens_cpu = self.optimistic_seq_lens_cpu[:num_reqs_padded]
2160
        seq_lens_cpu_upper_bound = seq_lens_cpu
2161
2162
2163
2164

        # is_prefilling: True if request is still in prefill phase.
        # Used by mamba backends to distinguish actual decodes from
        # short extends.
2165
2166
        is_prefilling = num_computed_tokens_cpu < num_prompt_tokens_cpu

2167
2168
2169
2170
2171
        if self.use_async_spec_decode:
            # GPU tensors are authoritative in async mode.
            seq_lens_cpu = None
            num_computed_tokens_cpu = None

2172
2173
2174
        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],
2175
2176
            seq_lens=self.seq_lens[:num_reqs_padded],
            _seq_lens_cpu=seq_lens_cpu,
2177
            _num_computed_tokens_cpu=num_computed_tokens_cpu,
2178
            seq_lens_cpu_upper_bound=seq_lens_cpu_upper_bound,
2179
2180
2181
2182
2183
2184
2185
            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,
2186
            is_prefilling=is_prefilling,
2187
2188
2189
2190
        )

        if self.dcp_world_size > 1:
            self.dcp_local_seq_lens.cpu[:num_reqs] = get_dcp_local_seq_lens(
2191
                self.optimistic_seq_lens_cpu[:num_reqs],
2192
2193
2194
2195
2196
2197
2198
2199
2200
2201
2202
2203
2204
2205
2206
2207
2208
2209
                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
            )

2210
2211
2212
2213
2214
2215
2216
2217
2218
        # 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
        ] = {}

2219
2220
2221
2222
2223
2224
2225
        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]
2226
            builder = attn_group.get_metadata_builder(ubid or 0)
2227
2228
2229
2230
            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))
2231

2232
2233
2234
2235
2236
2237
2238
            cascade_attn_prefix_len = (
                cascade_attn_prefix_lens[kv_cache_gid][attn_gid]
                if cascade_attn_prefix_lens
                else 0
            )

            extra_attn_metadata_args = {}
2239
2240
2241
            if use_spec_decode and isinstance(
                builder, (Mamba2AttentionMetadataBuilder, GDNAttentionMetadataBuilder)
            ):
2242
2243
2244
2245
2246
2247
2248
2249
2250
2251
2252
2253
                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
                )
2254
2255
2256
2257
2258
2259
2260
2261
2262
            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,
                )
2263
2264
2265
2266
2267
2268
            else:
                attn_metadata_i = builder.build(
                    common_prefix_len=cascade_attn_prefix_len,
                    common_attn_metadata=common_attn_metadata,
                    **extra_attn_metadata_args,
                )
2269
2270
                if builder.supports_update_block_table:
                    cached_attn_metadata[cache_key] = attn_metadata_i
2271
2272
2273
2274
2275
2276
2277
2278
2279
2280
2281
2282
2283
2284
2285
2286
2287
2288
2289
2290
2291
2292
2293

            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,
2294
                for_cudagraph_capture=for_cudagraph_capture,
2295
            )
2296
            if kv_cache_gid > 0:
2297
2298
                cm.block_table_tensor = _get_block_table(kv_cache_gid)
                cm.slot_mapping = slot_mappings[kv_cache_gid]
2299

2300
            if self.speculative_config and spec_decode_common_attn_metadata is None:
2301
                if isinstance(self.drafter, (EagleProposer, DFlashProposer)):
2302
                    if self.drafter.kv_cache_gid == kv_cache_gid:
2303
                        spec_decode_common_attn_metadata = cm
2304
                else:
2305
                    spec_decode_common_attn_metadata = cm
2306

2307
            for attn_gid in range(len(self.attn_groups[kv_cache_gid])):
2308
                if ubatch_slices is not None:
2309
2310
2311
                    for ubid, _cm in enumerate(split_attn_metadata(ubatch_slices, cm)):
                        _build_attn_group_metadata(kv_cache_gid, attn_gid, _cm, ubid)

2312
                else:
2313
                    _build_attn_group_metadata(kv_cache_gid, attn_gid, cm)
2314

2315
2316
2317
2318
2319
2320
2321
2322
2323
2324
2325
2326
        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

2327
2328
            # Set mm_prefix_range for all attention metadata
            self._set_mm_prefix_range_for_metadata(attn_metadata, req_doc_ranges)
2329

2330
2331
2332
2333
2334
2335
2336
2337
2338
2339
        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)
            )

2340
        return attn_metadata, spec_decode_common_attn_metadata
2341

2342
2343
2344
    def _compute_cascade_attn_prefix_lens(
        self,
        num_scheduled_tokens: np.ndarray,
2345
        num_computed_tokens: np.ndarray,
2346
2347
2348
2349
2350
2351
2352
        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
        """
2353

2354
2355
2356
2357
2358
2359
2360
2361
2362
2363
2364
2365
2366
2367
        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,
2368
                        num_computed_tokens,
2369
2370
2371
2372
2373
2374
2375
2376
                        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
2377

2378
2379
2380
    def _compute_cascade_attn_prefix_len(
        self,
        num_scheduled_tokens: np.ndarray,
2381
        num_computed_tokens: np.ndarray,
2382
        num_common_prefix_blocks: int,
2383
2384
        kv_cache_spec: KVCacheSpec,
        attn_metadata_builder: AttentionMetadataBuilder,
2385
2386
2387
2388
2389
2390
2391
2392
2393
2394
2395
2396
2397
2398
2399
2400
2401
2402
    ) -> 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.
        """
2403

2404
        common_prefix_len = num_common_prefix_blocks * kv_cache_spec.block_size
2405
2406
2407
2408
2409
2410
2411
2412
2413
2414
2415
2416
2417
2418
2419
2420
2421
2422
2423
2424
2425
2426
2427
2428
2429
2430
2431
2432
2433
2434
2435
2436
2437
2438
2439
2440
2441
        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]
2442
        # Request 3's num_computed_tokens: 3 (i.e., [A, B, C])
2443
2444
2445
2446
2447
        # 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.
2448
        common_prefix_len = min(common_prefix_len, num_computed_tokens.min())
2449
        # common_prefix_len should be a multiple of the block size.
2450
2451
2452
2453
2454
2455
2456
2457
2458
2459
2460
        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
        )
2461
2462
        assert isinstance(kv_cache_spec, AttentionSpec)
        use_cascade = attn_metadata_builder.use_cascade_attention(
2463
2464
2465
            common_prefix_len=common_prefix_len,
            query_lens=num_scheduled_tokens,
            num_query_heads=self.num_query_heads,
2466
            num_kv_heads=kv_cache_spec.num_kv_heads,
2467
            use_alibi=self.use_alibi,
2468
            use_sliding_window=use_sliding_window,
2469
            use_local_attention=use_local_attention,
2470
            num_sms=self.num_sms,
2471
            dcp_world_size=self.dcp_world_size,
2472
2473
2474
        )
        return common_prefix_len if use_cascade else 0

2475
2476
    def _calc_mrope_positions(self, scheduler_output: "SchedulerOutput"):
        mrope_pos_ptr = 0
2477
        for index, req_id in enumerate(self.input_batch.req_ids):
2478
2479
2480
            req = self.requests[req_id]
            assert req.mrope_positions is not None

2481
2482
            num_computed_tokens = self.input_batch.num_computed_tokens_cpu[index]
            num_scheduled_tokens = scheduler_output.num_scheduled_tokens[req_id]
2483
            num_prompt_tokens = length_from_prompt_token_ids_or_embeds(
2484
2485
                req.prompt_token_ids, req.prompt_embeds
            )
2486
2487

            if num_computed_tokens + num_scheduled_tokens > num_prompt_tokens:
2488
2489
                prompt_part_len = max(0, num_prompt_tokens - num_computed_tokens)
                completion_part_len = max(0, num_scheduled_tokens - prompt_part_len)
2490
2491
2492
2493
2494
2495
2496
2497
2498
2499
2500
2501
2502
            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

2503
2504
2505
                self.mrope_positions.cpu[:, dst_start:dst_end] = req.mrope_positions[
                    :, src_start:src_end
                ]
2506
2507
2508
2509
2510
2511
2512
                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

2513
                assert req.mrope_position_delta is not None
2514
                MRotaryEmbedding.get_next_input_positions_tensor(
2515
                    out=self.mrope_positions.np,
2516
2517
2518
2519
2520
                    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,
                )
2521
2522
2523

                mrope_pos_ptr += completion_part_len

2524
2525
2526
2527
2528
2529
2530
2531
2532
2533
2534
2535
2536
2537
2538
2539
2540
2541
2542
2543
2544
2545
2546
2547
2548
2549
2550
2551
2552
2553
2554
2555
2556
2557
2558
2559
2560
2561
2562
2563
2564
2565
2566
2567
2568
2569
2570
    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

2571
2572
    def _calc_spec_decode_metadata(
        self,
2573
2574
2575
2576
2577
2578
2579
2580
2581
2582
2583
2584
2585
2586
2587
2588
        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
2589

2590
2591
2592
2593
2594
        # Step 1.
        # cu_num_sampled_tokens: [4, 5, 8, 9, 11]
        # _arange_scratch[:11]: [0, 1, 2, 3, 0, 0, 1, 2, 0, 0, 1]
        cu_num_sampled_tokens = self._get_cumsum_and_arange(
            num_sampled_tokens, self._arange_scratch, cumsum_dtype=np.int32
2595
        )
2596
        # Step 2. [0, 0, 0, 0, 103, 104, 104, 104, 206, 207, 207]
2597
        logits_indices = np.repeat(
2598
2599
            cu_num_scheduled_tokens - num_sampled_tokens, num_sampled_tokens
        )
2600
        # Step 3. [0, 1, 2, 3, 103, 104, 105, 106, 206, 207, 208]
2601
        logits_indices += self._arange_scratch[: cu_num_sampled_tokens[-1]]
2602
2603
2604
2605
2606

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

        # Compute the draft logits indices.
2607
        # cu_num_draft_tokens: [3, 3, 5, 5, 6]
2608
2609
2610
        # _arange_scratch[:6]: [0, 1, 2, 0, 1, 0]
        cu_num_draft_tokens = self._get_cumsum_and_arange(
            num_draft_tokens, self._arange_scratch, cumsum_dtype=np.int32
2611
        )
2612
2613
        # [0, 0, 0, 5, 5, 9]
        target_logits_indices = np.repeat(
2614
2615
            cu_num_sampled_tokens - num_sampled_tokens, num_draft_tokens
        )
2616
        # [0, 1, 2, 5, 6, 9]
2617
        target_logits_indices += self._arange_scratch[: cu_num_draft_tokens[-1]]
2618
2619
2620

        # TODO: Optimize the CPU -> GPU copy.
        cu_num_draft_tokens = torch.from_numpy(cu_num_draft_tokens).to(
2621
2622
            self.device, non_blocking=True
        )
2623
2624
2625
        cu_num_sampled_tokens = torch.from_numpy(cu_num_sampled_tokens).to(
            self.device, non_blocking=True
        )
2626
2627
2628
        logits_indices = torch.from_numpy(logits_indices).to(
            self.device, non_blocking=True
        )
2629
        target_logits_indices = torch.from_numpy(target_logits_indices).to(
2630
2631
            self.device, non_blocking=True
        )
2632
        bonus_logits_indices = torch.from_numpy(bonus_logits_indices).to(
2633
2634
            self.device, non_blocking=True
        )
2635

2636
2637
        # Compute the draft token ids.
        # draft_token_indices:      [  1,   2,   3, 105, 106, 208]
2638
        draft_token_ids = self.input_ids.gpu[logits_indices]
2639
2640
        draft_token_ids = draft_token_ids[target_logits_indices + 1]

2641
        return SpecDecodeMetadata(
2642
2643
2644
            draft_token_ids=draft_token_ids,
            num_draft_tokens=num_draft_tokens.tolist(),
            cu_num_draft_tokens=cu_num_draft_tokens,
2645
            cu_num_sampled_tokens=cu_num_sampled_tokens,
2646
2647
2648
2649
2650
            target_logits_indices=target_logits_indices,
            bonus_logits_indices=bonus_logits_indices,
            logits_indices=logits_indices,
        )

2651
2652
2653
2654
2655
2656
2657
    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
2658
        self.kv_sharing_fast_prefill_logits_indices[:num_logits].copy_(logits_indices)
2659
2660
2661
2662
2663
        # 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_(
2664
2665
            logits_indices[-1].item()
        )
2666
2667
        # Dispatch for the decoder portion of the model.
        _, batch_desc = self.cudagraph_dispatcher.dispatch(
2668
            num_logits, invalid_modes={CUDAGraphMode.FULL}
2669
2670
        )
        num_logits_padded = batch_desc.num_tokens
2671
2672
2673
        logits_indices_padded = self.kv_sharing_fast_prefill_logits_indices[
            :num_logits_padded
        ]
2674
2675
        return logits_indices_padded

2676
    def _batch_mm_inputs_from_scheduler(
2677
2678
        self,
        scheduler_output: "SchedulerOutput",
2679
2680
    ) -> tuple[
        list[str],
2681
        list[tuple[str, MultiModalKwargsItem]],
2682
2683
        list[tuple[str, PlaceholderRange]],
    ]:
2684
        """Batch multimodal inputs from scheduled encoder inputs.
2685
2686
2687

        Args:
            scheduler_output: The scheduler output containing scheduled encoder
2688
                inputs.
2689
2690

        Returns:
2691
            A tuple of (mm_hashes, mm_kwargs, mm_lora_refs) where:
2692
2693
            - mm_hashes: List of multimodal hashes for each item
            - mm_kwargs: List of multimodal kwargs for each item
2694
            - mm_lora_refs: List of (req_id, placeholder_range) for each item
2695
        """
2696
2697
        scheduled_encoder_inputs = scheduler_output.scheduled_encoder_inputs
        if not scheduled_encoder_inputs:
2698
            return [], [], []
2699
2700

        mm_hashes = list[str]()
2701
        mm_kwargs = list[tuple[str, MultiModalKwargsItem]]()
2702
2703
2704
        # Multimodal LoRA reference info to map each multimodal item
        # back to its request & position
        mm_lora_refs = list[tuple[str, PlaceholderRange]]()
2705
2706
        for req_id, encoder_input_ids in scheduled_encoder_inputs.items():
            req_state = self.requests[req_id]
2707
2708

            for mm_input_id in encoder_input_ids:
2709
                mm_feature = req_state.mm_features[mm_input_id]
2710
2711
                if mm_feature.data is None:
                    continue
2712
2713

                mm_hashes.append(mm_feature.identifier)
2714
                mm_kwargs.append((mm_feature.modality, mm_feature.data))
2715
                mm_lora_refs.append((req_id, mm_feature.mm_position))
2716

2717
        return mm_hashes, mm_kwargs, mm_lora_refs
2718

2719
2720
2721
    def _execute_mm_encoder(
        self, scheduler_output: "SchedulerOutput"
    ) -> list[torch.Tensor]:
2722
2723
2724
        mm_hashes, mm_kwargs, mm_lora_refs = self._batch_mm_inputs_from_scheduler(
            scheduler_output
        )
2725
2726

        if not mm_kwargs:
2727
            return []
2728

2729
2730
2731
2732
2733
2734
        should_time = bool(
            self.observability_config
            and self.observability_config.enable_mm_processor_stats
            and scheduler_output.scheduled_encoder_inputs
        )

2735
2736
2737
2738
2739
2740
2741
        # 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.
2742
        model = cast(SupportsMultiModal, self.model)
2743
2744
2745
2746
2747
2748
2749
2750
2751
2752
2753
2754
2755
2756
2757

        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]
2758
                    pos_info.get_num_embeds()
2759
2760
2761
2762
2763
2764
2765
2766
2767
2768
2769
2770
2771
2772
2773
2774
2775
2776
2777
                )
                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)

2778
2779
2780
2781
2782
2783
2784
2785
2786
2787
2788
2789
2790
2791
            # Only set connector mapping if the model actually has a connector.
            # Some multimodal models inherit a stub `get_num_mm_connector_tokens`
            # from `SupportsMultiModal`, which returns None and should not be
            # treated as a signal that connector LoRA is supported.
            mm_mapping = (
                self.model.get_mm_mapping()  # type: ignore[attr-defined]
                if hasattr(self.model, "get_mm_mapping")
                else None
            )
            if (
                mm_mapping is not None
                and mm_mapping.connector
                and hasattr(self.model, "get_num_mm_connector_tokens")
            ):
2792
2793
2794
2795
2796
2797
2798
2799
2800
2801
2802
2803
2804
2805
2806
2807
2808
2809
2810
2811
2812
                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,
                )

2813
        encoder_outputs: list[torch.Tensor] = []
2814
2815
        # Track the current index in mm_kwargs/mm_lora_refs to map groups to request IDs
        current_item_idx = 0
2816
        for modality, num_items, mm_kwargs_batch in group_and_batch_mm_kwargs(
2817
2818
2819
            mm_kwargs,
            device=self.device,
            pin_memory=self.pin_memory,
2820
        ):
2821
            batch_outputs: MultiModalEmbeddings
2822

2823
            # EVS and dynamic res video related change.
2824
            # (ekhvedchenia): Temporary hack to limit peak memory usage when
2825
            # processing multimodal data. This solves the issue with scheduler
2826
2827
2828
2829
            # 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)
2830
2831
2832
            # dynamic res video for nemotron temporarily uses this hack via
            # requires_sequential_video_encoding
            # because it doesn't yet support video batching.
2833
2834
2835
            # TODO(ywang96): Fix memory profiling to take EVS into account and
            # remove this hack.
            if (
2836
2837
2838
2839
                (
                    self.is_multimodal_pruning_enabled
                    or self.requires_sequential_video_encoding
                )
2840
2841
2842
                and modality == "video"
                and num_items > 1
            ):
2843
                batch_outputs_lst = list[torch.Tensor]()
2844
2845
2846
2847
2848
2849
                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(
2850
                            group_and_batch_mm_kwargs(
2851
2852
2853
2854
                                [video_mm_kwargs_item],
                                device=self.device,
                                pin_memory=self.pin_memory,
                            )
2855
                        )
2856

2857
2858
2859
                        micro_batch_outputs = model.embed_multimodal(
                            **micro_batch_mm_inputs
                        )
2860

2861
                        batch_outputs_lst.extend(micro_batch_outputs)
2862

2863
                batch_outputs = batch_outputs_lst
2864
2865
            else:
                # Run the encoder.
2866
                # `batch_outputs` is either of the following:
2867
2868
2869
2870
2871
                # 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.
2872
2873
2874
2875

                with self.timed_encoder_operation(
                    should_time, mm_lora_refs, current_item_idx, num_items
                ):
2876
2877
2878
2879
2880
2881
2882
2883
2884
2885
2886
2887
2888
                    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)
2889

2890
2891
            sanity_check_mm_encoder_outputs(batch_outputs, expected_num_items=num_items)
            encoder_outputs.extend(batch_outputs)
2892

2893
2894
            current_item_idx += num_items

2895
        # Cache the encoder outputs by mm_hash
2896
        for mm_hash, output in zip(mm_hashes, encoder_outputs):
2897
            self.encoder_cache[mm_hash] = output
2898
2899
            logger.debug("Finish execute for mm hash %s", mm_hash)
            self.maybe_save_ec_to_connector(self.encoder_cache, mm_hash)
2900

2901
2902
        return encoder_outputs

2903
    def _gather_mm_embeddings(
2904
2905
        self,
        scheduler_output: "SchedulerOutput",
2906
        shift_computed_tokens: int = 0,
2907
2908
2909
2910
    ) -> tuple[list[torch.Tensor], torch.Tensor]:
        total_num_scheduled_tokens = scheduler_output.total_num_scheduled_tokens

        mm_embeds = list[torch.Tensor]()
2911
2912
2913
        is_mm_embed = torch.zeros(
            total_num_scheduled_tokens, dtype=torch.bool, device="cpu"
        )
2914
2915

        req_start_idx = 0
2916
        should_sync_mrope_positions = False
2917
        should_sync_xdrope_positions = False
2918

2919
        for req_id in self.input_batch.req_ids:
2920
2921
            mm_embeds_req: list[torch.Tensor] = []

2922
            num_scheduled_tokens = scheduler_output.num_scheduled_tokens[req_id]
2923
            req_state = self.requests[req_id]
2924
            num_computed_tokens = req_state.num_computed_tokens + shift_computed_tokens
2925

2926
2927
            for mm_feature in req_state.mm_features:
                pos_info = mm_feature.mm_position
2928
2929
                start_pos = pos_info.offset
                num_encoder_tokens = pos_info.length
2930
2931
2932
2933
2934
2935
2936
2937
2938
2939
2940
2941
2942
2943
2944
2945

                # 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,
2946
2947
                    num_encoder_tokens,
                )
2948
                assert start_idx < end_idx
2949
2950
2951
2952
2953
2954
2955
                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
2956

2957
                mm_hash = mm_feature.identifier
2958
                encoder_output = self.encoder_cache.get(mm_hash, None)
2959
                assert encoder_output is not None, f"Encoder cache miss for {mm_hash}."
2960
2961
2962

                if (is_embed := pos_info.is_embed) is not None:
                    is_embed = is_embed[start_idx:end_idx]
2963
2964
2965
                    mm_embeds_item = encoder_output[curr_embeds_start:curr_embeds_end]
                else:
                    mm_embeds_item = encoder_output[start_idx:end_idx]
2966

2967
                req_start_pos = req_start_idx + start_pos - num_computed_tokens
2968
2969
2970
2971
2972
2973
2974
2975
2976
                # 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
2977
2978
2979
                mm_embeds_req.append(mm_embeds_item)

            if self.is_multimodal_pruning_enabled and self.uses_mrope:
2980
                assert req_state.mrope_positions is not None
2981
2982
2983
2984
2985
2986
2987
                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,
2988
2989
                    )
                )
2990
2991
2992
2993
                req_state.mrope_positions.copy_(new_mrope_positions)
                req_state.mrope_position_delta = new_delta

            mm_embeds.extend(mm_embeds_req)
2994
2995
            req_start_idx += num_scheduled_tokens

2996
2997
        if should_sync_mrope_positions:
            self._calc_mrope_positions(scheduler_output)
2998
            self.mrope_positions.copy_to_gpu(total_num_scheduled_tokens)
2999

3000
3001
3002
3003
        if should_sync_xdrope_positions:
            self._calc_xdrope_positions(scheduler_output)
            self.xdrope_positions.copy_to_gpu(total_num_scheduled_tokens)

3004
        return mm_embeds, is_mm_embed
3005

3006
    def get_model(self) -> nn.Module:
3007
3008
        if not hasattr(self, "model"):
            raise ValueError("Cannot get model before model has been initialized")
3009
        if isinstance(self.model, (CUDAGraphWrapper, UBatchWrapper)):
3010
            # get raw model out of the cudagraph wrapper.
3011
            return self.model.unwrap()
3012
3013
        return self.model

3014
3015
3016
3017
3018
3019
3020
3021
3022
3023
3024
3025
3026
    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")

3027
3028
3029
        if supports_realtime(model):
            supported_tasks.append("realtime")

3030
3031
        return supported_tasks

3032
3033
3034
3035
3036
    def get_supported_pooling_tasks(self) -> list[PoolingTask]:
        model = self.get_model()
        if not is_pooling_model(model):
            return []

3037
        return list(model.pooler.get_supported_tasks())
3038

3039
3040
3041
3042
3043
3044
3045
3046
3047
3048
    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)

3049
    def sync_and_slice_intermediate_tensors(
3050
3051
        self,
        num_tokens: int,
3052
        intermediate_tensors: IntermediateTensors | None,
3053
3054
        sync_self: bool,
    ) -> IntermediateTensors:
3055
3056
3057
        assert self.intermediate_tensors is not None

        tp = self.vllm_config.parallel_config.tensor_parallel_size
3058
        is_rs = is_residual_scattered_for_sp(self.vllm_config, num_tokens)
3059
3060
3061
3062
3063
3064

        # 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():
3065
                is_scattered = k == "residual" and is_rs
3066
                copy_len = num_tokens // tp if is_scattered else num_tokens
3067
                self.intermediate_tensors[k][:copy_len].copy_(
3068
3069
3070
3071
3072
3073
3074
3075
3076
3077
3078
3079
3080
                    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:
3081
3082
3083
        """
        Step for the EPLB (Expert Parallelism Load Balancing) state.
        """
3084
        if not self.parallel_config.enable_eplb or self.eep_eplb_suppressed:
3085
3086
3087
            return

        assert self.eplb_state is not None
3088
3089
        model = self.get_model()
        assert is_mixture_of_experts(model)
3090
3091
3092
        self.eplb_state.step(
            is_dummy,
            is_profile,
3093
            log_stats=self.parallel_config.eplb_config.log_balancedness,
3094
3095
        )

3096
3097
3098
3099
3100
3101
3102
3103
3104
3105
3106
3107
3108
3109
3110
3111
3112
    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,
        )

3113
3114
3115
3116
3117
    def _pool(
        self,
        hidden_states: torch.Tensor,
        num_scheduled_tokens: int,
        num_scheduled_tokens_np: np.ndarray,
3118
3119
3120
3121
        kv_connector_output: KVConnectorOutput | None,
    ) -> ModelRunnerOutput | AsyncModelRunnerOutput:
        num_reqs = self.input_batch.num_reqs
        assert num_reqs == len(self.input_batch.pooling_params), (
3122
3123
            "Either all or none of the requests in a batch must be pooling request"
        )
3124

3125
        hidden_states = hidden_states[:num_scheduled_tokens]
3126
        seq_lens_cpu = self.optimistic_seq_lens_cpu[:num_reqs]
3127

3128
        pooling_metadata = self.input_batch.get_pooling_metadata()
3129
        pooling_metadata.build_pooling_cursor(
3130
3131
3132
3133
            num_scheduled_tokens_np,
            seq_lens_cpu,
            device=hidden_states.device,
            query_start_loc_gpu=self.query_start_loc.gpu[: num_reqs + 1],
3134
        )
3135

3136
3137
        model = cast(VllmModelForPooling, self.model)
        raw_pooler_output: PoolerOutput = model.pooler(
3138
            hidden_states=hidden_states, pooling_metadata=pooling_metadata
3139
        )
3140
3141
3142
3143
3144

        finished_mask = [
            seq_len == prompt_len
            for seq_len, prompt_len in zip(seq_lens_cpu, pooling_metadata.prompt_lens)
        ]
3145
3146
3147
3148
3149
3150
        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,
        )
3151
3152
3153
3154
3155
3156
3157
3158
3159
3160
3161

        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

3162
3163
3164
        if not current_platform.is_cuda_alike():
            # cpu/xpu runners cannot use the CUDA stream/event-based wrapper.
            model_runner_output.pooler_output = _copy_pooler_output_to_cpu(
3165
3166
3167
                raw_pooler_output=raw_pooler_output,
                finished_mask=finished_mask,
            )
3168
3169
            self._sync_device()
            return model_runner_output
3170

3171
3172
        return AsyncGPUPoolingModelRunnerOutput(
            model_runner_output=model_runner_output,
3173
3174
            raw_pooler_output=raw_pooler_output,
            finished_mask=finished_mask,
3175
            async_output_copy_stream=self._get_or_create_async_output_copy_stream(),
3176
        )
3177

3178
    def _pad_for_sequence_parallelism(self, num_scheduled_tokens: int) -> int:
3179
3180
3181
        # Pad tokens to multiple of tensor_parallel_size when
        # enabled collective fusion for SP
        tp_size = self.vllm_config.parallel_config.tensor_parallel_size
3182
        if self.compilation_config.pass_config.enable_sp and tp_size > 1:
3183
3184
3185
            return round_up(num_scheduled_tokens, tp_size)
        return num_scheduled_tokens

Patrick von Platen's avatar
Patrick von Platen committed
3186
3187
3188
3189
3190
3191
3192
3193
3194
3195
3196
    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

3197
    def _preprocess(
3198
3199
        self,
        scheduler_output: "SchedulerOutput",
3200
        num_input_tokens: int,  # Padded
3201
        intermediate_tensors: IntermediateTensors | None = None,
3202
    ) -> tuple[
3203
3204
        torch.Tensor | None,
        torch.Tensor | None,
3205
        torch.Tensor,
3206
        IntermediateTensors | None,
3207
        dict[str, Any],
3208
        ECConnectorOutput | None,
3209
    ]:
3210
        num_scheduled_tokens = scheduler_output.total_num_scheduled_tokens
3211
        is_first_rank = get_pp_group().is_first_rank
3212
        is_encoder_decoder = self.model_config.is_encoder_decoder
3213

3214
3215
        # _prepare_inputs may reorder the batch, so we must gather multi
        # modal outputs after that to ensure the correct order
3216
3217
        ec_connector_output = None

3218
        if self.supports_mm_inputs and is_first_rank and not is_encoder_decoder:
3219
            # Run the multimodal encoder if any.
3220
3221
3222
3223
3224
3225
            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)
3226

3227
3228
3229
            # 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.
3230
            inputs_embeds_scheduled = self.model.embed_input_ids(
3231
3232
3233
                self.input_ids.gpu[:num_scheduled_tokens],
                multimodal_embeddings=mm_embeds,
                is_multimodal=is_mm_embed,
3234
            )
3235

3236
            # TODO(woosuk): Avoid the copy. Optimize.
3237
            self.inputs_embeds.gpu[:num_scheduled_tokens].copy_(inputs_embeds_scheduled)
3238

Patrick von Platen's avatar
Patrick von Platen committed
3239
            input_ids, inputs_embeds = self._prepare_mm_inputs(num_input_tokens)
3240
            model_kwargs = {
3241
                **self._init_model_kwargs(),
3242
3243
                **self._extract_mm_kwargs(scheduler_output),
            }
3244
        elif self.enable_prompt_embeds and is_first_rank:
3245
3246
3247
3248
3249
3250
3251
3252
3253
3254
3255
3256
            # 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).
3257
3258
3259
            token_ids_idx = (
                self.is_token_ids.gpu[:num_scheduled_tokens]
                .nonzero(as_tuple=False)
3260
                .squeeze(1)
3261
            )
3262
3263
3264
            # Some tokens ids may need to become embeds
            if token_ids_idx.numel() > 0:
                token_ids = self.input_ids.gpu[token_ids_idx]
3265
                tokens_to_embeds = self.model.embed_input_ids(input_ids=token_ids)
3266
3267
3268
                self.inputs_embeds.gpu[token_ids_idx] = tokens_to_embeds

            inputs_embeds = self.inputs_embeds.gpu[:num_input_tokens]
3269
            model_kwargs = self._init_model_kwargs()
3270
            input_ids = None
3271
        else:
3272
3273
3274
3275
            # 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.
3276
            input_ids = self.input_ids.gpu[:num_input_tokens]
3277
            inputs_embeds = None
3278
            model_kwargs = self._init_model_kwargs()
3279

3280
        if self.uses_mrope:
3281
            positions = self.mrope_positions.gpu[:, :num_input_tokens]
3282
3283
        elif self.uses_xdrope_dim > 0:
            positions = self.xdrope_positions.gpu[:, :num_input_tokens]
3284
        else:
3285
            positions = self.positions[:num_input_tokens]
3286
            if num_input_tokens > num_scheduled_tokens:
3287
                self.positions[num_scheduled_tokens:num_input_tokens].zero_()
3288

3289
        if is_first_rank:
3290
3291
            intermediate_tensors = None
        else:
3292
            assert intermediate_tensors is not None
3293
            intermediate_tensors = self.sync_and_slice_intermediate_tensors(
3294
3295
                num_input_tokens, intermediate_tensors, True
            )
3296

3297
        if is_encoder_decoder and scheduler_output.scheduled_encoder_inputs:
3298
3299
3300
3301
3302
3303
3304
            # 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})
3305

3306
3307
3308
3309
3310
3311
        return (
            input_ids,
            inputs_embeds,
            positions,
            intermediate_tensors,
            model_kwargs,
3312
            ec_connector_output,
3313
        )
3314

3315
    def _sample(
3316
        self,
3317
3318
        logits: torch.Tensor | None,
        spec_decode_metadata: SpecDecodeMetadata | None,
3319
    ) -> SamplerOutput:
3320
        # Sample the next token and get logprobs if needed.
3321
        sampling_metadata = self.input_batch.sampling_metadata
3322
3323
3324
        # 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()
3325
        if spec_decode_metadata is None:
3326
            return self.sampler(
3327
3328
3329
                logits=logits,
                sampling_metadata=sampling_metadata,
            )
3330

3331
3332
3333
3334
3335
3336
        # 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)

3337
        sampler_output = self.rejection_sampler(
3338
3339
            spec_decode_metadata,
            None,  # draft_probs
3340
            logits,
3341
3342
            sampling_metadata,
        )
3343
3344
3345
        return sampler_output

    def _bookkeeping_sync(
3346
3347
3348
        self,
        scheduler_output: "SchedulerOutput",
        sampler_output: SamplerOutput,
3349
        logits: torch.Tensor | None,
3350
3351
        hidden_states: torch.Tensor,
        num_scheduled_tokens: int,
3352
        spec_decode_metadata: SpecDecodeMetadata | None,
3353
    ) -> tuple[
3354
        dict[str, int],
3355
        LogprobsLists | None,
3356
        list[list[int]],
3357
        dict[str, LogprobsTensors | None],
3358
3359
3360
        list[str],
        dict[str, int],
        list[int],
3361
    ]:
3362
3363
3364
3365
        num_nans_in_logits = {}
        if envs.VLLM_COMPUTE_NANS_IN_LOGITS:
            num_nans_in_logits = self._get_nans_in_logits(logits)

3366
3367
3368
3369
        num_reqs = self.input_batch.num_reqs
        discard_sampled_tokens_req_indices = np.nonzero(
            self.discard_request_mask.np[:num_reqs]
        )[0]
3370
3371
3372
3373
        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)
3374

3375
3376
3377
        # 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()
3378
        req_id_to_index_output_copy = self.input_batch.req_id_to_index.copy()
3379
3380

        num_sampled_tokens = sampler_output.sampled_token_ids.shape[0]
3381
        sampled_token_ids = sampler_output.sampled_token_ids
3382
        logprobs_tensors = sampler_output.logprobs_tensors
3383
        invalid_req_indices = []
3384
        logprobs_lists = None
3385
3386
3387
3388
3389
3390
        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)
3391
3392
3393
                # 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()
3394
3395
3396

                if logprobs_tensors is not None:
                    logprobs_lists = logprobs_tensors.tolists()
3397
3398
            else:
                # Includes spec decode tokens.
3399
                valid_sampled_token_ids, logprobs_lists = RejectionSampler.parse_output(
3400
3401
                    sampled_token_ids,
                    self.input_batch.vocab_size,
3402
                    discard_sampled_tokens_req_indices,
3403
                    logprobs_tensors=logprobs_tensors,
3404
                )
3405
        else:
3406
            valid_sampled_token_ids = []
3407
            invalid_req_indices = discard_sampled_tokens_req_indices.tolist()
3408
3409
3410
3411
3412
            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.
3413
3414
3415
3416
            # 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
3417
3418
3419
3420
3421
            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
            }
3422

3423
3424
3425
3426
3427
        # 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.
3428
        req_ids = self.input_batch.req_ids
3429
3430
        for req_idx in range(num_sampled_tokens):
            if self.use_async_scheduling:
3431
                sampled_ids = [-1] if req_idx not in invalid_req_indices_set else None
3432
3433
            else:
                sampled_ids = valid_sampled_token_ids[req_idx]
3434

3435
            num_sampled_ids: int = len(sampled_ids) if sampled_ids else 0
3436

3437
            if not sampled_ids:
3438
3439
3440
                continue

            start_idx = self.input_batch.num_tokens_no_spec[req_idx]
3441
            end_idx = start_idx + num_sampled_ids
3442
3443
3444
3445
            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}"
3446
            )
3447

3448
3449
            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
3450
            self.input_batch.num_tokens_no_spec[req_idx] = end_idx
3451

3452
            req_id = req_ids[req_idx]
3453
3454
3455
            req_state = self.requests[req_id]
            req_state.output_token_ids.extend(sampled_ids)

3456
3457
3458
3459
3460
3461
        # Compute prompt logprobs if needed.
        prompt_logprobs_dict = self._get_prompt_logprobs_dict(
            hidden_states[:num_scheduled_tokens],
            scheduler_output.num_scheduled_tokens,
        )

3462
3463
3464
3465
3466
3467
3468
3469
3470
3471
        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,
        )

3472
3473
3474
3475
3476
3477
3478
3479
3480
3481
3482
3483
3484
3485
3486
    @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()

3487
3488
    def _model_forward(
        self,
3489
3490
3491
3492
        input_ids: torch.Tensor | None = None,
        positions: torch.Tensor | None = None,
        intermediate_tensors: IntermediateTensors | None = None,
        inputs_embeds: torch.Tensor | None = None,
3493
3494
3495
3496
3497
        **model_kwargs: dict[str, Any],
    ) -> Any:
        """Helper method to call the model forward pass.

        This method can be overridden by subclasses for model execution.
3498
        Motivation: We can inspect only this method versus
3499
3500
3501
3502
3503
3504
3505
3506
3507
3508
3509
3510
3511
3512
3513
3514
3515
3516
3517
3518
        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,
        )

3519
3520
3521
3522
3523
3524
3525
3526
3527
3528
3529
3530
3531
3532
3533
3534
3535
3536
3537
3538
3539
    @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
        )

3540
3541
3542
3543
3544
3545
3546
3547
3548
3549
3550
3551
3552
    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,
3553
        force_num_active_loras: int | None = None,
3554
        num_encoder_reqs: int = 0,
3555
    ) -> tuple[
3556
3557
        CUDAGraphMode,
        BatchDescriptor,
3558
        bool,
3559
3560
        torch.Tensor | None,
        CUDAGraphStat | None,
3561
    ]:
3562
3563
3564
3565
3566
3567
        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,
3568
        )
3569
3570
3571
3572
3573
        # 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
        )
3574

3575
3576
3577
3578
3579
        # 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)
3580
        )
3581
        has_lora = num_active_loras > 0 if force_has_lora is None else force_has_lora
3582

3583
        num_tokens_padded = self._pad_for_sequence_parallelism(num_tokens)
3584
3585
3586

        def dispatch_cudagraph(num_tokens, disable_full=False, valid_modes=None):
            return self.cudagraph_dispatcher.dispatch(
3587
3588
3589
                num_tokens=num_tokens,
                has_lora=has_lora,
                uniform_decode=uniform_decode,
3590
                num_active_loras=num_active_loras,
3591
3592
                valid_modes={CUDAGraphMode.NONE} if force_eager else valid_modes,
                invalid_modes={CUDAGraphMode.FULL} if disable_full else None,
3593
3594
            )

3595
        cudagraph_mode, batch_descriptor = dispatch_cudagraph(
3596
            num_tokens_padded, disable_full=use_cascade_attn or has_encoder_output
3597
        )
3598
        num_tokens_padded = batch_descriptor.num_tokens
3599
3600
3601
3602
3603
3604
3605
3606
3607
        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"
            )
3608
3609
3610

        # Extra coordination when running data-parallel since we need to coordinate
        # across ranks
3611
        should_ubatch, num_tokens_across_dp = False, None
3612
        if self.vllm_config.parallel_config.data_parallel_size > 1:
3613
3614
3615
3616
3617
3618
3619
3620
3621
3622
            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,
                )
3623
3624
            )

3625
            # Extract DP-synced values
3626
3627
3628
            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())
3629
3630
3631
                # Re-dispatch with DP padding so we have the correct batch_descriptor
                cudagraph_mode, batch_descriptor = dispatch_cudagraph(
                    num_tokens_padded,
3632
                    valid_modes={CUDAGraphMode(synced_cudagraph_mode)},
3633
                )
3634
3635
3636
3637
                # 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

3638
3639
3640
3641
3642
3643
3644
3645
3646
3647
3648
3649
        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,
3650
            should_ubatch,
3651
3652
3653
            num_tokens_across_dp,
            cudagraph_stats,
        )
3654

3655
3656
3657
3658
3659
3660
3661
3662
3663
3664
3665
3666
3667
3668
3669
3670
3671
3672
3673
3674
3675
3676
3677
3678
3679
3680
3681
3682
3683
3684
3685
3686
3687
3688
3689
3690
    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

3691
3692
3693
3694
3695
3696
3697
3698
3699
3700
3701
3702
3703
3704
3705
3706
3707
3708
3709
3710
3711
3712
3713
3714
3715
3716
3717
3718
3719
3720
3721
3722
3723
3724
3725
3726
3727
3728
3729
3730
3731
3732
3733
3734
3735
3736
3737
3738
3739
3740
3741
3742
3743
3744
3745
3746
3747
3748
3749
3750
3751
3752
3753
3754
3755
3756
3757
3758
3759
3760
3761
3762
3763
3764
    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

3765
3766
3767
3768
3769
3770
3771
3772
3773
    def _is_all_reqs_chunked_prefill(self) -> bool:
        """Check if all scheduled requests are marked to discard sampled tokens.

        This is true when `discard_request_mask` is set for every scheduled
        request (e.g., for chunked prefill requests that are not the last
        prefill chunk)."""
        num_reqs = self.input_batch.num_reqs
        return bool(self.discard_request_mask.np[:num_reqs].all())

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

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

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

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

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

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

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

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

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

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

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

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

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

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

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

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

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

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

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

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

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

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

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

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

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

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

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

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

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

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

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

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

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

4125
4126
4127
4128
4129
4130
4131
        return None

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

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

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

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

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

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

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

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

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

4208
        spec_config = self.speculative_config
4209
4210
        propose_drafts_after_bookkeeping = False
        if spec_config is not None:
4211
            # Decide whether to run the drafter or zero out draft tokens.
4212
4213
4214
            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
4215
            )
4216
            use_gpu_toks = (
4217
4218
4219
                spec_config.use_eagle()
                or spec_config.uses_draft_model()
                or spec_config.uses_extract_hidden_states()
4220
4221
4222
            ) and not spec_config.disable_padded_drafter_batch
            if use_gpu_toks:
                # EAGLE/DraftModel speculative decoding can use the GPU sampled tokens
4223
                # as inputs, and does not need to wait for bookkeeping to finish.
4224
4225
                assert isinstance(
                    self.drafter,
4226
4227
4228
4229
                    EagleProposer
                    | DFlashProposer
                    | DraftModelProposer
                    | ExtractHiddenStatesProposer,
4230
                )
4231
4232
4233
4234
4235
4236
4237
4238
4239
4240
4241
4242
                sampled_token_ids = sampler_output.sampled_token_ids
                if input_fits_in_drafter:
                    propose_draft_token_ids(sampled_token_ids)
                elif self.valid_sampled_token_count_event is not None:
                    assert spec_decode_common_attn_metadata is not None
                    next_token_ids, valid_sampled_tokens_count = (
                        self.drafter.prepare_next_token_ids_padded(
                            sampled_token_ids,
                            self.requests,
                            self.input_batch,
                            self.discard_request_mask.gpu,
                        )
4243
                    )
4244
4245
4246
                    self._copy_valid_sampled_token_count(
                        next_token_ids, valid_sampled_tokens_count
                    )
4247
4248
4249
4250
4251
4252
4253
4254
4255
4256
4257
4258
4259
4260
4261
4262
4263
4264
4265
4266
4267
4268
            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
                    )
4269
4270
            else:
                propose_drafts_after_bookkeeping = input_fits_in_drafter
4271

4272
4273
4274
4275
4276
4277
4278
4279
4280
4281
4282
            if not input_fits_in_drafter:
                # Zero out draft tokens so the scheduler doesn't schedule
                # stale drafts from the previous step.
                # For Nemotron-H: it is necessary to zero out the draft tokens,
                # otherwise the stale tokens will corrupt Mamba recurrent
                # state and logprobs for sequences near max_model_len.
                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)

4283
        with record_function_or_nullcontext("gpu_model_runner: bookkeep"):
4284
4285
4286
4287
4288
4289
4290
4291
            (
                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,
4292
4293
4294
4295
4296
            ) = self._bookkeeping_sync(
                scheduler_output,
                sampler_output,
                logits,
                hidden_states,
4297
                scheduler_output.total_num_scheduled_tokens,
4298
                spec_decode_metadata,
4299
            )
4300

4301
        if propose_drafts_after_bookkeeping:
4302
4303
4304
            # 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)
4305

4306
4307
4308
4309
4310
        # 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()
4311

4312
        with record_function_or_nullcontext("gpu_model_runner: eplb"):
4313
            self.eplb_step()
4314

4315
4316
4317
4318
        # self.kv_connector_output may be modified during drafting
        kv_connector_output = self.kv_connector_output
        self.kv_connector_output = None

4319
        with record_function_or_nullcontext("gpu_model_runner: ModelRunnerOutput"):
4320
            if self.routed_experts_initialized:
4321
4322
4323
4324
4325
4326
                capturer = RoutedExpertsCapturer.get_instance()
                if capturer is not None:
                    capturer.save_captured_experts(indices=self.slot_mapping)  # noqa
                else:
                    logger.error("RoutedExpertsCapturer not initialized.")

4327
4328
4329
4330
4331
4332
4333
            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,
4334
4335
4336
                ec_connector_output=ec_connector_output
                if self.supports_mm_inputs
                else None,
4337
                num_nans_in_logits=num_nans_in_logits,
4338
                cudagraph_stats=cudagraph_stats,
4339
            )
4340

4341
4342
        if not self.use_async_scheduling:
            return output
4343

4344
4345
4346
4347
4348
4349
4350
4351
        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,
4352
                async_output_copy_stream=self._get_or_create_async_output_copy_stream(),
4353
                vocab_size=self.input_batch.vocab_size,
4354
4355
4356
4357
4358
            )
        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
4359
            # any requests with sampling params that require output ids.
4360
4361
4362
4363
            self.input_batch.set_async_sampled_token_ids(
                async_output.sampled_token_ids_cpu,
                async_output.async_copy_ready_event,
            )
4364
4365
4366

        return async_output

4367
4368
4369
4370
4371
4372
4373
4374
4375
4376
    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]"
        )
4377
4378
4379
4380
4381
4382
        # Skip for chunked prefill: sampled tokens are dummy
        # and will be discarded, no need to broadcast.
        if not self._is_all_reqs_chunked_prefill():
            torch.distributed.broadcast(
                sampled_token_ids, src=pp.rank, group=pp.device_group
            )
4383
4384
4385
4386
4387
4388
4389
4390

    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)
4391
4392
4393
        # skip for chunked prefill.
        if not self._is_all_reqs_chunked_prefill():
            torch.distributed.broadcast(recv, src=pp.last_rank, group=pp.device_group)
4394
4395
4396
4397
4398
4399
4400
4401
4402
4403
4404
4405
4406
4407
4408
4409
4410
        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

4411
    def take_draft_token_ids(self) -> DraftTokenIds | None:
4412
        if not self.num_spec_tokens or not self._draft_token_req_ids:
4413
            return None
4414
        draft_token_ids, req_ids = self._get_draft_token_ids_cpu()
4415
        return DraftTokenIds(req_ids, draft_token_ids)
4416

4417
4418
4419
    def _copy_draft_token_ids_to_cpu(
        self, scheduler_output: "SchedulerOutput", zeros_only: bool = False
    ) -> None:
4420
4421
4422
4423
4424
4425
        # 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
        ):
4426
4427
4428
            return
        # We must also set the corresponding request ids.
        self._draft_token_req_ids = self.input_batch.req_ids.copy()
4429

4430
4431
4432
4433
4434
4435
4436
4437
4438
4439
4440
4441
4442
4443
4444
4445
4446
4447
4448
4449
        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()

4450
    def _get_draft_token_ids_cpu(self) -> tuple[list[list[int]], list[str]]:
4451
        if isinstance(self._draft_token_ids, list):
4452
4453
4454
4455
            return self._draft_token_ids, self.input_batch.req_ids
        req_ids = self._draft_token_req_ids
        if req_ids is None:
            return [], []
4456
4457
4458
        assert self.draft_token_ids_event is not None
        assert self.draft_token_ids_cpu is not None
        self.draft_token_ids_event.synchronize()
4459
        return self.draft_token_ids_cpu[: len(req_ids)].tolist(), req_ids
4460

4461
4462
4463
4464
4465
4466
4467
4468
4469
4470
4471
4472
4473
    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
4474
            assert counts_cpu is not None
4475
4476
4477
            counts_cpu[: counts.shape[0]].copy_(counts, non_blocking=True)
            self.valid_sampled_token_count_event.record()

4478
4479
4480
        if self.use_async_spec_decode:
            # Stash for GPU-side correction in _prepare_inputs.
            self.valid_sampled_token_count_gpu = valid_sampled_tokens_count
4481
4482
4483
4484
4485
        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
4486
4487
        sampled_count_event = self.valid_sampled_token_count_event
        if sampled_count_event is None or prev_sampled_token_ids is None:
4488
4489
4490
            return []

        counts_cpu = self.valid_sampled_token_count_cpu
4491
4492
        assert counts_cpu is not None
        sampled_count_event.synchronize()
4493
4494
        return counts_cpu[: prev_sampled_token_ids.shape[0]].tolist()

4495
4496
4497
    def propose_draft_token_ids(
        self,
        scheduler_output: "SchedulerOutput",
4498
        sampled_token_ids: torch.Tensor | list[list[int]],
4499
4500
4501
        sampling_metadata: SamplingMetadata,
        hidden_states: torch.Tensor,
        sample_hidden_states: torch.Tensor,
4502
4503
        aux_hidden_states: list[torch.Tensor] | None,
        spec_decode_metadata: SpecDecodeMetadata | None,
4504
        common_attn_metadata: CommonAttentionMetadata,
4505
        slot_mappings: dict[str, torch.Tensor] | list[dict[str, torch.Tensor]] | None,
4506
    ) -> list[list[int]] | torch.Tensor:
4507
        num_scheduled_tokens = scheduler_output.total_num_scheduled_tokens
4508
4509
4510
        spec_config = self.speculative_config
        assert spec_config is not None
        if spec_config.method == "ngram":
4511
4512
            from vllm.v1.spec_decode.ngram_proposer import NgramProposer

4513
            assert isinstance(sampled_token_ids, list)
4514
            assert isinstance(self.drafter, NgramProposer)
4515
            draft_token_ids = self.drafter.propose(
4516
                sampled_token_ids,
4517
4518
                self.input_batch.num_tokens_no_spec,
                self.input_batch.token_ids_cpu,
4519
                slot_mappings=slot_mappings,
4520
            )
4521
4522
4523
4524
4525
4526
4527
4528
4529
4530
4531
4532
4533
4534
4535
4536
4537
4538
4539
4540
4541
4542
4543
4544
4545
4546
4547
4548
4549
4550
4551
4552
4553
4554
4555
4556
4557
        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,
            )
4558
        elif spec_config.method == "suffix":
4559
4560
            assert isinstance(sampled_token_ids, list)
            assert isinstance(self.drafter, SuffixDecodingProposer)
4561
4562
4563
            draft_token_ids = self.drafter.propose(
                self.input_batch, sampled_token_ids, slot_mappings=slot_mappings
            )
4564
        elif spec_config.method == "medusa":
4565
            assert isinstance(sampled_token_ids, list)
4566
            assert isinstance(self.drafter, MedusaProposer)
4567

4568
4569
            if sample_hidden_states.shape[0] == len(sampled_token_ids):
                # The input to the target model does not include draft tokens.
4570
4571
4572
4573
                hidden_states = sample_hidden_states
            else:
                indices = []
                offset = 0
4574
4575
4576
                assert spec_decode_metadata is not None, (
                    "No spec decode metadata for medusa"
                )
4577
                for num_draft, tokens in zip(
4578
4579
                    spec_decode_metadata.num_draft_tokens, sampled_token_ids
                ):
4580
                    indices.append(offset + len(tokens) - 1)
4581
                    offset += num_draft + 1
4582
                indices = torch.tensor(indices, device=self.device)
4583
4584
                hidden_states = sample_hidden_states[indices]

4585
            draft_token_ids = self.drafter.propose(
4586
4587
                target_hidden_states=hidden_states,
                sampling_metadata=sampling_metadata,
4588
                slot_mappings=slot_mappings,
4589
            )
4590
4591
4592
4593
4594
4595
4596
4597
4598
4599
4600
4601
        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]

4602
            draft_token_ids = self.drafter.propose(
4603
4604
4605
4606
4607
4608
4609
4610
4611
4612
4613
4614
4615
4616
4617
4618
4619
                sampled_token_ids=sampled_token_ids,
                target_hidden_states=target_hidden_states,
                common_attn_metadata=common_attn_metadata,
                slot_mappings=slot_mappings,
            )
            next_token_ids, valid_sampled_tokens_count = (
                self.drafter.prepare_next_token_ids_padded(
                    sampled_token_ids,
                    self.requests,
                    self.input_batch,
                    self.discard_request_mask.gpu,
                )
            )
            self._copy_valid_sampled_token_count(
                next_token_ids, valid_sampled_tokens_count
            )

4620
4621
4622
4623
4624
4625
4626
4627
        elif (
            spec_config.use_eagle()
            or spec_config.use_dflash()
            or spec_config.uses_draft_model()
        ):
            assert isinstance(
                self.drafter, EagleProposer | DFlashProposer | DraftModelProposer
            )
4628

4629
            if spec_config.disable_padded_drafter_batch:
4630
4631
4632
                # 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.
4633
4634
                assert isinstance(sampled_token_ids, list), (
                    "sampled_token_ids should be a python list when"
4635
                    "padded-batch is disabled."
4636
                )
4637
                next_token_ids = self.drafter.prepare_next_token_ids_cpu(
4638
4639
4640
4641
4642
                    sampled_token_ids,
                    self.requests,
                    self.input_batch,
                    scheduler_output.num_scheduled_tokens,
                )
4643
4644
4645
4646
4647
            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.
4648
4649
                assert isinstance(sampled_token_ids, torch.Tensor), (
                    "sampled_token_ids should be a torch.Tensor when"
4650
                    "padded-batch is enabled."
4651
4652
                )
                next_token_ids, valid_sampled_tokens_count = (
4653
4654
4655
4656
                    self.drafter.prepare_next_token_ids_padded(
                        sampled_token_ids,
                        self.requests,
                        self.input_batch,
4657
                        self.discard_request_mask.gpu,
4658
                    )
4659
                )
4660
4661
4662
                self._copy_valid_sampled_token_count(
                    next_token_ids, valid_sampled_tokens_count
                )
Jiayi Yao's avatar
Jiayi Yao committed
4663

4664
            num_rejected_tokens_gpu = None
4665
            if spec_decode_metadata is None:
4666
                token_indices_to_sample = None
4667
                # input_ids can be None for multimodal models.
4668
                target_token_ids = self.input_ids.gpu[:num_scheduled_tokens]
4669
                target_positions = self._get_positions(num_scheduled_tokens)
4670
                if self.use_aux_hidden_state_outputs:
Wentao Ye's avatar
Wentao Ye committed
4671
                    assert aux_hidden_states is not None
4672
                    target_hidden_states = torch.cat(
4673
4674
                        [h[:num_scheduled_tokens] for h in aux_hidden_states], dim=-1
                    )
4675
4676
                else:
                    target_hidden_states = hidden_states[:num_scheduled_tokens]
4677
            else:
4678
                if spec_config.disable_padded_drafter_batch:
4679
                    token_indices_to_sample = None
4680
4681
4682
4683
4684
                    common_attn_metadata, token_indices = self.drafter.prepare_inputs(
                        common_attn_metadata,
                        sampled_token_ids,
                        spec_decode_metadata.num_draft_tokens,
                    )
4685
4686
4687
4688
4689
4690
4691
4692
4693
                    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]
4694
                else:
4695
4696
4697
4698
4699
4700
4701
4702
                    (
                        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,
4703
                    )
4704
4705
4706
4707
4708
4709
4710
4711
4712
4713
4714
                    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]
4715

4716
            if self.supports_mm_inputs and self.drafter.supports_mm_inputs:
4717
4718
4719
4720
4721
4722
                mm_embed_inputs = self._gather_mm_embeddings(
                    scheduler_output,
                    shift_computed_tokens=1,
                )
            else:
                mm_embed_inputs = None
4723

4724
            draft_token_ids = self.drafter.propose(
4725
4726
4727
4728
                target_token_ids=target_token_ids,
                target_positions=target_positions,
                target_hidden_states=target_hidden_states,
                next_token_ids=next_token_ids,
4729
                token_indices_to_sample=token_indices_to_sample,
4730
                sampling_metadata=sampling_metadata,
4731
                common_attn_metadata=common_attn_metadata,
4732
                mm_embed_inputs=mm_embed_inputs,
4733
                num_rejected_tokens_gpu=num_rejected_tokens_gpu,
4734
                slot_mappings=slot_mappings,
4735
            )
4736

4737
        return draft_token_ids
4738

4739
4740
4741
    def update_config(self, overrides: dict[str, Any]) -> None:
        allowed_config_names = {"load_config", "model_config"}
        for config_name, config_overrides in overrides.items():
4742
4743
            assert config_name in allowed_config_names, (
                f"Config `{config_name}` not supported. "
4744
                f"Allowed configs: {allowed_config_names}"
4745
            )
4746
4747
4748
4749
            config = getattr(self, config_name)
            new_config = update_config(config, config_overrides)
            setattr(self, config_name, new_config)

4750
    @instrument(span_name="Loading (GPU)")
4751
    def load_model(self, load_dummy_weights: bool = False) -> None:
4752
4753
        """
        Args:
4754
            load_dummy_weights: load dummy weights instead of real weights.
4755
        """
4756
4757
4758
4759
4760
        logger.info_once(
            "Starting to load model %s...",
            self.model_config.model,
            scope="global",
        )
4761

4762
4763
4764
4765
        if self.parallel_config.enable_eplb:
            self.eplb_state = EplbState(self.parallel_config, self.device)
            eplb_models = 0

4766
4767
4768
        try:
            with DeviceMemoryProfiler() as m:
                time_before_load = time.perf_counter()
4769
4770
                if load_dummy_weights:
                    self.load_config.load_format = "dummy"
4771
4772
4773
4774
4775
4776
4777
                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
4778
                    )
4779
4780
4781
4782
4783
4784
4785
4786
                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
                    ):
4787
4788
4789
                        assert not self.parallel_config.enable_elastic_ep, (
                            "Elastic EP is not supported with drafter model."
                        )
4790
4791
4792
4793
4794
4795
4796
4797
4798
4799
4800
4801
4802
4803
4804
4805
                        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
4806

4807
4808
4809
4810
4811
4812
                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"
                        )
4813

4814
4815
4816
4817
4818
4819
4820
4821
4822
                    # 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:
4823
4824
4825
                        aux_layers = (
                            self.model.get_eagle3_default_aux_hidden_state_layers()
                        )
4826
4827

                    self.model.set_aux_hidden_state_layers(aux_layers)
4828
4829
4830
4831
4832
4833
4834
4835
4836
4837
4838
4839
4840
4841
4842
4843
4844

                if (
                    is_mixture_of_experts(self.model)
                    and self.parallel_config.enable_eplb
                    and not load_dummy_weights
                ):
                    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(
                        self.model,
                        self.model_config,
                    )
                    eplb_models += 1

4845
4846
4847
4848
4849
4850
4851
4852
4853
4854
4855
4856
4857
                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
4858
        logger.info_once(
4859
4860
            "Model loading took %s GiB memory and %.6f seconds",
            format_gib(self.model_memory_usage),
4861
4862
            time_after_load - time_before_load,
        )
4863
4864
4865
4866
4867
4868
        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)
4869
        mm_config = self.model_config.multimodal_config
4870
        self.is_multimodal_pruning_enabled = (
4871
            supports_multimodal_pruning(self.get_model())
4872
4873
            and mm_config is not None
            and mm_config.is_multimodal_pruning_enabled()
4874
        )
4875
4876
4877
        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
4878

4879
4880
4881
4882
        if (
            is_mixture_of_experts(self.model)
            and self.parallel_config.enable_eplb
            and not load_dummy_weights
4883
4884
            and self.eplb_state is not None
            and self.eplb_state.is_async
4885
        ):
4886
            self.eplb_state.start_async_loop()
4887

4888
        if (
4889
4890
            self.vllm_config.compilation_config.mode
            == CompilationMode.STOCK_TORCH_COMPILE
4891
        ):
4892
4893
4894
            from vllm.env_override import _apply_constrain_to_fx_strides_patch

            _apply_constrain_to_fx_strides_patch()
4895
            backend = self.vllm_config.compilation_config.init_backend(self.vllm_config)
4896
            compilation_counter.stock_torch_compile_count += 1
4897
            self.model.compile(fullgraph=True, backend=backend)
4898
            return
4899
        # for other compilation modes, cudagraph behavior is controlled by
Jiayi Yan's avatar
Jiayi Yan committed
4900
        # CudagraphWrapper and CudagraphDispatcher of vllm.
4901
4902

        # wrap the model with full cudagraph wrapper if needed.
4903
4904
        cudagraph_mode = self.compilation_config.cudagraph_mode
        assert cudagraph_mode is not None
4905
4906
4907
4908
        if (
            cudagraph_mode.has_full_cudagraphs()
            and not self.parallel_config.use_ubatching
        ):
4909
4910
4911
            self.model = CUDAGraphWrapper(
                self.model, self.vllm_config, runtime_mode=CUDAGraphMode.FULL
            )
4912
        elif self.parallel_config.use_ubatching:
4913
            if cudagraph_mode.has_full_cudagraphs():
4914
4915
4916
                self.model = UBatchWrapper(
                    self.model, self.vllm_config, CUDAGraphMode.FULL, self.device
                )
4917
            else:
4918
4919
4920
                self.model = UBatchWrapper(
                    self.model, self.vllm_config, CUDAGraphMode.NONE, self.device
                )
4921

4922
4923
        get_offloader().post_init()

4924
    def _get_eagle3_aux_layers_from_config(self) -> tuple[int, ...] | None:
4925
4926
4927
4928
4929
4930
4931
4932
4933
4934
4935
4936
4937
4938
4939
        """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

4940
4941
4942
4943
4944
4945
        layer_ids = getattr(hf_config, "eagle_aux_hidden_state_layer_ids", None)
        if not layer_ids:
            dflash_config = getattr(hf_config, "dflash_config", None)
            if dflash_config and isinstance(dflash_config, dict):
                layer_ids = dflash_config.get("target_layer_ids")

4946
4947
4948
4949
4950
        if layer_ids and isinstance(layer_ids, (list, tuple)):
            return tuple(layer_ids)

        return None

4951
4952
4953
4954
4955
4956
4957
4958
4959
4960
4961
4962
4963
    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
4964
            into kernel format (repacking, renaming, etc.)
4965
4966
4967
4968
4969
4970
4971
4972
4973
4974
4975
4976
4977
4978
4979
4980
4981
4982
4983
4984
4985
4986
4987
4988
4989
4990
4991
4992
4993
4994
        """
        # 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
4995
        logger.info_once("Reloading weights inplace...")
4996
4997
4998
4999
5000
        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)
5001

5002
5003
5004
5005
5006
5007
5008
5009
5010
5011
5012
        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",
            )
            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)
5013
5014
5015
5016
5017
5018
5019

        # 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,
5020
        )
5021
5022
5023
5024
5025
5026
5027
        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,
                )
5028

5029
5030
5031
    def _get_prompt_logprobs_dict(
        self,
        hidden_states: torch.Tensor,
5032
        num_scheduled_tokens: dict[str, int],
5033
    ) -> dict[str, LogprobsTensors | None]:
5034
        num_prompt_logprobs_dict = self.num_prompt_logprobs
5035
5036
5037
        if not num_prompt_logprobs_dict:
            return {}

5038
        in_progress_dict = self.input_batch.in_progress_prompt_logprobs_cpu
5039
        prompt_logprobs_dict: dict[str, LogprobsTensors | None] = {}
5040
5041
5042
5043
5044

        # 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():
5045
5046
5047
5048
            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
5049
5050
5051

            # Get metadata for this request.
            request = self.requests[req_id]
5052
5053
5054
5055
            if request.prompt_token_ids is None:
                # Prompt logprobs is incompatible with prompt embeddings
                continue

5056
5057
            num_prompt_tokens = len(request.prompt_token_ids)
            prompt_token_ids = torch.tensor(request.prompt_token_ids).to(
5058
5059
                self.device, non_blocking=True
            )
5060

5061
5062
5063
5064
5065
5066
            # 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(
5067
5068
                    num_prompt_tokens - 1, num_prompt_logprobs + 1
                )
5069
5070
                in_progress_dict[req_id] = logprobs_tensors

5071
            # Determine number of logits to retrieve.
5072
5073
            start_idx = request.num_computed_tokens
            start_tok = start_idx + 1
5074
            num_remaining_tokens = num_prompt_tokens - start_tok
5075
            if num_tokens <= num_remaining_tokens:
5076
                # This is a chunk, more tokens remain.
5077
5078
5079
                # 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.
5080
5081
5082
5083
5084
                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)
5085
5086
5087
5088
5089
5090
5091
                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
5092
5093
5094
5095
5096

            # 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]
5097
            offset = self.query_start_loc.np[req_idx].item()
5098
            prompt_hidden_states = hidden_states[offset : offset + num_logits]
5099
            logits = self.model.compute_logits(prompt_hidden_states)
5100
5101
5102
5103

            # 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.
5104
            tgt_token_ids = prompt_token_ids[start_tok : start_tok + num_logits]
5105
5106

            # Compute prompt logprobs.
5107
            logprobs = self.sampler.compute_logprobs(logits)
5108
            token_ids, logprobs, ranks, _ = self.sampler.gather_logprobs(
5109
5110
                logprobs, num_prompt_logprobs, tgt_token_ids
            )
5111
5112

            # Transfer GPU->CPU async.
5113
5114
            chunk_slice = slice(start_idx, start_idx + num_logits)
            logprobs_tensors.logprob_token_ids[chunk_slice].copy_(
5115
5116
5117
                token_ids, non_blocking=True
            )
            logprobs_tensors.logprobs[chunk_slice].copy_(logprobs, non_blocking=True)
5118
            logprobs_tensors.selected_token_ranks[chunk_slice].copy_(
5119
5120
                ranks, non_blocking=True
            )
5121
5122
5123
5124
5125

        # 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]
5126
            del in_progress_dict[req_id]
5127
5128

        # Must synchronize the non-blocking GPU->CPU transfers.
5129
        if prompt_logprobs_dict:
5130
            self._sync_device()
5131
5132
5133

        return prompt_logprobs_dict

5134
5135
    def _get_nans_in_logits(
        self,
5136
        logits: torch.Tensor | None,
5137
5138
5139
5140
5141
5142
5143
5144
5145
5146
5147
    ) -> 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])
5148
5149
5150
                    if num_nans_for_index is not None and req_index < logits.shape[0]
                    else 0
                )
5151
5152
5153
5154
            return num_nans_in_logits
        except IndexError:
            return {}

5155
    @contextmanager
5156
5157
5158
    def maybe_randomize_inputs(
        self, input_ids: torch.Tensor | None, inputs_embeds: torch.Tensor | None
    ):
5159
5160
5161
5162
        """
        Randomize input_ids if VLLM_RANDOMIZE_DP_DUMMY_INPUTS is set.
        This is to help balance expert-selection
         - during profile_run
5163
         - during DP rank dummy run
5164
        """
5165

5166
5167
5168
5169
        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
5170
        elif input_ids is not None:
5171
5172
5173
5174

            @functools.cache
            def rand_input_ids() -> torch.Tensor:
                return torch.randint_like(
5175
                    self.input_ids.gpu,
5176
5177
                    low=0,
                    high=self.model_config.get_vocab_size(),
5178
                )
5179

5180
            logger.debug_once("Randomizing dummy input_ids for DP Rank")
5181
            input_ids.copy_(rand_input_ids()[: input_ids.size(0)], non_blocking=True)
5182
5183
            yield
            input_ids.fill_(0)
5184
5185
5186
5187
5188
5189
5190
5191
5192
5193
5194
5195
5196
5197
5198
        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)
5199

5200
5201
5202
5203
5204
5205
    def _get_mm_dummy_batch(
        self,
        modality: str,
        max_items_per_batch: int,
    ) -> BatchedTensorInputs:
        """Dummy data for profiling and precompiling multimodal models."""
5206
5207
        assert self.mm_budget is not None

5208
5209
5210
        # 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,
5211
            mm_counts={modality: 1},
5212
            cache=self.mm_budget.cache,
5213
        )
5214
5215
5216
5217
5218
        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"
5219

5220
        return next(
5221
5222
            mm_kwargs_batch
            for _, _, mm_kwargs_batch in group_and_batch_mm_kwargs(
5223
                [(modality, dummy_mm_item)] * max_items_per_batch,
5224
5225
5226
5227
                device=self.device,
                pin_memory=self.pin_memory,
            )
        )
5228

5229
5230
5231
5232
    @torch.inference_mode()
    def _dummy_run(
        self,
        num_tokens: int,
5233
        cudagraph_runtime_mode: CUDAGraphMode | None = None,
5234
5235
        force_attention: bool = False,
        uniform_decode: bool = False,
5236
        allow_microbatching: bool = True,
5237
5238
        skip_eplb: bool = False,
        is_profile: bool = False,
5239
        create_mixed_batch: bool = False,
5240
        remove_lora: bool = True,
Rémi Delacourt's avatar
Rémi Delacourt committed
5241
        is_graph_capturing: bool = False,
5242
        num_active_loras: int = 0,
5243
        profile_seq_lens: int | None = None,
5244
    ) -> tuple[torch.Tensor, torch.Tensor]:
5245
5246
5247
5248
5249
5250
5251
        """
        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.
5252
                - if not set will determine the cudagraph mode based on using
5253
                    the self.cudagraph_dispatcher.
5254
5255
5256
5257
                - CUDAGraphMode.NONE: No cudagraph, for warm up and profile run
                - CUDAGraphMode.PIECEWISE: Piecewise cudagraph.
                - CUDAGraphMode.FULL: Full cudagraph, attention metadata is
                    needed.
5258
            force_attention: If True, always create attention metadata. Used to
5259
5260
5261
5262
                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.
5263
5264
            create_mixed_batch: If True, create a mixed batch with both decode
                (1 token) and prefill (multiple tokens) requests.
5265
            remove_lora: If False, dummy LoRAs are not destroyed after the run
5266
5267
            num_active_loras: Number of distinct active LoRAs to capture for.
                LoRA is activated when num_active_loras > 0.
5268
5269
5270
            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.
5271
        """
5272
5273
        mm_config = self.vllm_config.model_config.multimodal_config
        if mm_config and mm_config.mm_encoder_only:
5274
5275
5276
5277
            # 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([])

5278
5279
        assert (
            cudagraph_runtime_mode is None
5280
            or cudagraph_runtime_mode.is_valid_runtime_mode()
5281
        )
5282

5283
        # If cudagraph_mode.decode_mode() == FULL and
5284
        # cudagraph_mode.separate_routine(). This means that we are using
5285
5286
5287
5288
5289
5290
5291
5292
5293
5294
5295
        # 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.
5296
        max_query_len = self.uniform_decode_query_len if uniform_decode else num_tokens
5297

5298
5299
5300
        # 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.
5301
        assert num_tokens <= self.max_num_tokens
5302
        max_num_reqs = self.scheduler_config.max_num_seqs
5303
5304
5305
5306
        if create_mixed_batch:
            assert not uniform_decode
            # Create mixed batch:
            # first half decode tokens, second half one prefill
5307
            num_decode_tokens = min(max_num_reqs - 1, num_tokens // 2)
5308
5309
5310
5311
            num_prefill_tokens = num_tokens - num_decode_tokens
            num_reqs = num_decode_tokens + 1

            # Create decode requests (1 token each) followed by prefill request
5312
            num_scheduled_tokens_list = [1] * num_decode_tokens + [num_prefill_tokens]
5313
5314
5315
            # Note: Overriding max_query_len to be the prefill tokens
            max_query_len = num_prefill_tokens
        elif uniform_decode:
5316
            assert not create_mixed_batch
5317
            num_reqs = min(max_num_reqs, cdiv(num_tokens, max_query_len))
5318
5319
            num_scheduled_tokens_list = [max_query_len] * num_reqs
            if num_tokens % max_query_len != 0:
5320
                num_scheduled_tokens_list[-1] = num_tokens % max_query_len
5321
5322
5323
5324
5325
5326
        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

5327
5328
        assert sum(num_scheduled_tokens_list) == num_tokens
        assert len(num_scheduled_tokens_list) == num_reqs
5329
        num_scheduled_tokens = np.array(num_scheduled_tokens_list, dtype=np.int32)
5330
5331
        num_tokens_unpadded = int(num_scheduled_tokens.sum())

5332
        num_sampled_tokens = np.ones(num_reqs, dtype=np.int32)
5333

5334
        _cudagraph_mode, batch_desc, should_ubatch, num_tokens_across_dp, _ = (
5335
5336
5337
5338
5339
5340
5341
5342
5343
5344
5345
5346
5347
5348
5349
5350
5351
            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
5352
5353
5354
5355
                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,
5356
5357
            )
        )
5358
5359
5360

        if cudagraph_runtime_mode is None:
            cudagraph_runtime_mode = _cudagraph_mode
5361
        else:
5362
5363
5364
5365
5366
5367
5368
5369
5370
            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
        )
5371
        ubatch_slices, ubatch_slices_padded = maybe_create_ubatch_slices(
5372
5373
5374
5375
5376
5377
5378
5379
5380
5381
            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,
5382
        )
5383

5384
        attn_metadata: PerLayerAttnMetadata | None = None
5385

5386
        slot_mappings_by_group, slot_mappings = self._get_slot_mappings(
5387
            num_tokens_padded=num_tokens_padded,
5388
5389
5390
5391
5392
            num_reqs_padded=num_reqs_padded,
            num_tokens_unpadded=num_tokens_unpadded,
            ubatch_slices=ubatch_slices_padded,
        )

5393
5394
5395
5396
5397
5398
        # Dummy runs have no real slot assignments — fill with -1 so
        # concat_and_cache kernels skip the KV write.
        if slot_mappings_by_group is not None:
            for sm in slot_mappings_by_group.values():
                sm.fill_(-1)

5399
5400
5401
5402
5403
5404
5405
5406
        # _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:
5407
5408
5409
                if profile_seq_lens is not None:
                    seq_lens = profile_seq_lens  # type: ignore[assignment]
                elif create_mixed_batch:
5410
5411
5412
                    # 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
5413
5414
5415
5416
                    seq_lens = torch.tensor(  # type: ignore[assignment]
                        [1] * num_decode_tokens + [num_prefill_tokens + 1],
                        dtype=torch.int,
                    )
5417
5418
                else:
                    seq_lens = max_query_len  # type: ignore[assignment]
5419
5420
5421
                self.optimistic_seq_lens_cpu[:num_reqs] = seq_lens
                self.optimistic_seq_lens_cpu[num_reqs:].fill_(0)
                self.seq_lens.copy_(self.optimistic_seq_lens_cpu, non_blocking=True)
5422

5423
5424
5425
                cum_num_tokens = self._get_cumsum_and_arange(
                    num_scheduled_tokens, self.query_pos.np
                )
5426
5427
                self.query_start_loc.np[1 : num_reqs + 1] = cum_num_tokens
                self.query_start_loc.copy_to_gpu()
5428

5429
5430
5431
5432
5433
5434
                # Sync block table CPU->GPU so cleared rows from
                # remove_request() are visible to the attention metadata
                # builder. Without this, stale block IDs from finished
                # requests can corrupt Mamba state.
                self.input_batch.block_table.commit_block_table(num_reqs_padded)

5435
5436
5437
5438
5439
5440
5441
5442
5443
                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,
5444
                    use_spec_decode=self.speculative_config is not None,
5445
                )
5446

5447
        with self.maybe_dummy_run_with_lora(
5448
5449
5450
5451
            self.lora_config,
            num_scheduled_tokens,
            num_sampled_tokens,
            remove_lora,
5452
            num_active_loras,
5453
        ):
5454
            # Make sure padding doesn't exceed max_num_tokens
5455
            assert num_tokens_padded <= self.max_num_tokens
5456
            model_kwargs = self._init_model_kwargs()
5457
            if self.supports_mm_inputs and not self.model_config.is_encoder_decoder:
Patrick von Platen's avatar
Patrick von Platen committed
5458
5459
                input_ids, inputs_embeds = self._prepare_mm_inputs(num_tokens_padded)

5460
                model_kwargs = {
5461
                    **model_kwargs,
5462
5463
                    **self._dummy_mm_kwargs(num_reqs),
                }
5464
5465
            elif self.enable_prompt_embeds:
                input_ids = None
5466
                inputs_embeds = self.inputs_embeds.gpu[:num_tokens_padded]
5467
                model_kwargs = self._init_model_kwargs()
5468
            else:
5469
                input_ids = self.input_ids.gpu[:num_tokens_padded]
5470
                inputs_embeds = None
5471

5472
            if self.uses_mrope:
5473
                positions = self.mrope_positions.gpu[:, :num_tokens_padded]
5474
            elif self.uses_xdrope_dim > 0:
5475
                positions = self.xdrope_positions.gpu[:, :num_tokens_padded]
5476
            else:
5477
                positions = self.positions[:num_tokens_padded]
5478
5479
5480
5481
5482
5483
5484
5485
5486

            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,
5487
5488
5489
                            device=self.device,
                        )
                    )
5490
5491

                intermediate_tensors = self.sync_and_slice_intermediate_tensors(
5492
                    num_tokens_padded, None, False
5493
                )
5494

5495
            if ubatch_slices_padded is not None:
5496
5497
5498
                # Adjust values to reflect a single ubatch.
                # TODO(sage,lucas): this is cruft that should be addressed in
                #  the padding refactor.
5499
                num_tokens_padded = ubatch_slices_padded[0].num_tokens
5500
                if num_tokens_across_dp is not None:
5501
                    num_tokens_across_dp[:] = num_tokens_padded
5502

5503
            with (
5504
                self.maybe_randomize_inputs(input_ids, inputs_embeds),
5505
                set_forward_context(
5506
5507
                    attn_metadata,
                    self.vllm_config,
5508
                    num_tokens=num_tokens_padded,
5509
5510
                    num_tokens_across_dp=num_tokens_across_dp,
                    cudagraph_runtime_mode=cudagraph_runtime_mode,
5511
                    batch_descriptor=batch_desc,
5512
                    ubatch_slices=ubatch_slices_padded,
5513
                    slot_mapping=slot_mappings,
5514
5515
                ),
            ):
5516
                outputs = self.model(
5517
5518
5519
5520
                    input_ids=input_ids,
                    positions=positions,
                    intermediate_tensors=intermediate_tensors,
                    inputs_embeds=inputs_embeds,
5521
                    **model_kwargs,
5522
                )
5523

5524
5525
5526
5527
            if self.use_aux_hidden_state_outputs:
                hidden_states, _ = outputs
            else:
                hidden_states = outputs
5528

5529
5530
5531
            if self.speculative_config and (
                self.speculative_config.use_eagle()
                or self.speculative_config.uses_draft_model()
5532
                or self.speculative_config.uses_extract_hidden_states()
5533
            ):
5534
5535
                assert isinstance(
                    self.drafter,
5536
5537
5538
5539
                    EagleProposer
                    | DFlashProposer
                    | DraftModelProposer
                    | ExtractHiddenStatesProposer,
5540
                )
5541
                assert self.speculative_config is not None
5542
5543
5544
                # 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.
5545
                use_cudagraphs = (
5546
5547
5548
5549
5550
5551
5552
5553
5554
                    (
                        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
5555
5556
5557
5558
5559

                # 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
5560
5561
5562
5563
                if (
                    self.compilation_config.cudagraph_specialize_lora
                    and num_active_loras > 0
                ):
5564
5565
5566
5567
5568
                    use_cudagraphs = False

                self.drafter.dummy_run(
                    num_tokens,
                    use_cudagraphs=use_cudagraphs,
Rémi Delacourt's avatar
Rémi Delacourt committed
5569
                    is_graph_capturing=is_graph_capturing,
5570
                    slot_mappings=slot_mappings,
5571
                )
5572

5573
5574
5575
5576
5577
5578
5579
5580
5581
5582
5583
        # 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()

5584
5585
5586
5587
5588
5589
5590
5591
5592
5593
        # 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)

5594
        logit_indices = np.cumsum(num_scheduled_tokens) - 1
5595
5596
5597
5598
        logit_indices_device = torch.from_numpy(logit_indices).to(
            self.device, non_blocking=True
        )
        return hidden_states, hidden_states[logit_indices_device]
5599
5600
5601
5602
5603
5604

    @torch.inference_mode()
    def _dummy_sampler_run(
        self,
        hidden_states: torch.Tensor,
    ) -> torch.Tensor:
5605
5606
5607
        # The dummy hidden states may contain special values,
        # like `inf` or `nan`.
        # To avoid breaking the sampler, we use a random tensor here instead.
5608

5609
5610
        mm_config = self.vllm_config.model_config.multimodal_config
        if mm_config and mm_config.mm_encoder_only:
5611
5612
5613
            # MM Encoder only model no need to run sampler.
            return torch.tensor([])

5614
        hidden_states = torch.rand_like(hidden_states)
5615

5616
        logits = self.model.compute_logits(hidden_states)
5617
5618
        num_reqs = logits.size(0)

5619
        dummy_tensors = lambda v: torch.full((num_reqs,), v, device=self.device)
5620
5621
5622
5623
5624
5625
5626
5627
5628

        dummy_metadata = SamplingMetadata(
            temperature=dummy_tensors(0.5),
            all_greedy=False,
            all_random=False,
            top_p=dummy_tensors(0.9),
            top_k=dummy_tensors(logits.size(1) - 1),
            generators={},
            max_num_logprobs=None,
Vedant V Jhaveri's avatar
Vedant V Jhaveri committed
5629
            logprob_token_ids=None,
5630
5631
5632
5633
5634
5635
            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)],
5636
            spec_token_ids=[[] for _ in range(num_reqs)],
5637
5638
            allowed_token_ids_mask=None,
            bad_words_token_ids={},
5639
            logitsprocs=LogitsProcessors(),
5640
        )
5641
        try:
5642
5643
5644
            sampler_output = self.sampler(
                logits=logits, sampling_metadata=dummy_metadata
            )
5645
        except RuntimeError as e:
5646
            if "out of memory" in str(e):
5647
5648
5649
5650
                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 "
5651
5652
                    "initializing the engine."
                ) from e
5653
5654
            else:
                raise e
5655
        if self.speculative_config:
5656
5657
            draft_token_ids = [[0] for _ in range(num_reqs)]
            dummy_spec_decode_metadata = SpecDecodeMetadata.make_dummy(
5658
5659
                draft_token_ids, self.device
            )
5660
5661
5662
5663
5664
5665

            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
5666
5667
5668
5669
5670
            logits = torch.randn(
                num_tokens + num_reqs,
                logits.shape[-1],
                device=self.device,
                dtype=logits.dtype,
5671
            )
5672
5673
5674
            self.rejection_sampler(
                dummy_spec_decode_metadata,
                draft_probs,
5675
                logits,
5676
5677
                dummy_metadata,
            )
5678
        return sampler_output
5679

5680
    def _dummy_pooler_run_task(
5681
5682
        self,
        hidden_states: torch.Tensor,
5683
5684
        task: PoolingTask,
    ) -> PoolerOutput:
5685
5686
5687
5688
        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
5689
5690
5691
5692
        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
5693
5694
5695

        req_num_tokens = num_tokens // num_reqs

5696
        dummy_prompt_lens = torch.from_numpy(num_scheduled_tokens_np)
5697
5698
5699
        dummy_token_ids = torch.zeros(
            (num_reqs, req_num_tokens), dtype=torch.int32, device=self.device
        )
5700

5701
        model = cast(VllmModelForPooling, self.get_model())
5702
        dummy_pooling_params = PoolingParams(task=task)
5703
        dummy_pooling_params.verify(self.model_config)
5704
        to_update = model.pooler.get_pooling_updates(task)
5705
5706
        to_update.apply(dummy_pooling_params)

5707
        dummy_metadata = PoolingMetadata(
5708
5709
            prompt_lens=dummy_prompt_lens,
            prompt_token_ids=dummy_token_ids,
5710
            prompt_token_ids_cpu=dummy_token_ids.cpu(),
5711
            pooling_params=[dummy_pooling_params] * num_reqs,
5712
            pooling_states=[PoolingStates() for i in range(num_reqs)],
5713
        )
5714

5715
        dummy_metadata.build_pooling_cursor(
5716
            num_scheduled_tokens_np,
5717
5718
            seq_lens_cpu=dummy_prompt_lens,
            device=hidden_states.device,
5719
        )
5720

5721
        try:
5722
5723
5724
            return model.pooler(
                hidden_states=hidden_states, pooling_metadata=dummy_metadata
            )
5725
        except RuntimeError as e:
5726
            if "out of memory" in str(e):
5727
                raise RuntimeError(
5728
5729
5730
                    "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 "
5731
5732
                    "initializing the engine."
                ) from e
5733
5734
            else:
                raise e
5735
5736
5737
5738
5739
5740

    @torch.inference_mode()
    def _dummy_pooler_run(
        self,
        hidden_states: torch.Tensor,
    ) -> PoolerOutput:
5741
5742
        mm_config = self.vllm_config.model_config.multimodal_config
        if mm_config and mm_config.mm_encoder_only:
5743
5744
5745
            # MM Encoder only model not need to run pooler.
            return torch.tensor([])

5746
        # Find the task that has the largest output for subsequent steps
5747
5748
5749
        supported_pooling_tasks = self.get_supported_pooling_tasks()

        if not supported_pooling_tasks:
5750
5751
5752
5753
5754
5755
            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."
            )
5756

5757
        output_size = dict[PoolingTask, float]()
5758
        for task in supported_pooling_tasks:
5759
5760
            # Run a full batch with each task to ensure none of them OOMs
            output = self._dummy_pooler_run_task(hidden_states, task)
5761
            output_size[task] = sum(o.nbytes for o in output if o is not None)
5762
5763
5764
5765
            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)
5766

5767
    def profile_run(self) -> None:
5768
        # Profile with multimodal encoder & encoder cache.
5769
        if self.supports_mm_inputs:
5770
5771
            mm_config = self.model_config.multimodal_config
            if mm_config is not None and mm_config.skip_mm_profiling:
5772
                logger.info(
5773
                    "Skipping memory profiling for multimodal encoder and "
5774
5775
                    "encoder cache."
                )
5776
5777
5778
5779
5780
            else:
                mm_budget = self.mm_budget
                assert mm_budget is not None

                if (encoder_budget := mm_budget.get_encoder_budget()) > 0:
5781
5782
5783
5784
5785
5786
5787
5788
5789
5790
5791
5792
5793
5794
5795
5796
5797
                    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
                        ]
5798

5799
                        logger.info_once(
5800
5801
5802
5803
5804
5805
5806
                            "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,
                        )
5807

5808
5809
5810
5811
5812
                        # Create dummy batch of multimodal inputs.
                        batched_dummy_mm_inputs = self._get_mm_dummy_batch(
                            dummy_modality,
                            max_mm_items_per_batch,
                        )
5813

5814
5815
5816
5817
5818
5819
5820
5821
5822
5823
5824
                        # 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
5825

5826
        # Add `is_profile` here to pre-allocate communication buffers
5827
5828
5829
        hidden_states, last_hidden_states = self._dummy_run(
            self.max_num_tokens, is_profile=True
        )
5830
        if get_pp_group().is_last_rank:
5831
5832
5833
5834
            if self.is_pooling_model:
                output = self._dummy_pooler_run(hidden_states)
            else:
                output = self._dummy_sampler_run(last_hidden_states)
5835
        else:
5836
            output = None
5837
        self._sync_device()
5838
        del hidden_states, output
5839
        self.encoder_cache.clear()
5840
        gc.collect()
5841

5842
5843
5844
5845
5846
5847
5848
5849
5850
5851
    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

5852
5853
5854
        # 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
5855
        minimal_config = get_kv_cache_config_from_groups(
5856
            self.vllm_config, kv_cache_groups, available_memory=0, suppress_log=True
5857
        )
5858
        self.cache_config.num_gpu_blocks_override = saved_override
5859

5860
        self.initialize_kv_cache(minimal_config, is_profiling=True)
5861
5862
5863
5864
5865
5866
5867
5868
5869
5870
5871
5872
5873
5874
5875
5876
5877
5878
5879
5880
5881
5882
5883
5884
5885
5886
5887
5888
5889
5890
5891
5892
5893
5894
5895
        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"):
5896
5897
5898
5899
                kv_cache = layer.kv_cache
                layer.kv_cache = (
                    torch.tensor([]) if isinstance(kv_cache, torch.Tensor) else []
                )
5900
5901
5902
5903
5904
5905
5906
            # Clean up quantized KV cache scale views
            # (int8_per_token_head, fp8_per_token_head)
            if hasattr(layer, "impl"):
                if hasattr(layer.impl, "_k_scale_cache"):
                    layer.impl._k_scale_cache = None
                if hasattr(layer.impl, "_v_scale_cache"):
                    layer.impl._v_scale_cache = None
5907
5908
5909
5910
5911
5912
5913
5914
5915
5916
5917
5918
5919
5920
5921
5922
5923
5924
5925
5926
5927
5928
5929
5930
5931
5932
5933
5934
5935
5936
5937
5938
5939
5940
5941
5942
5943
5944
5945
5946
5947
5948
5949
5950
5951
5952
5953
5954
5955
5956
5957
5958
5959
5960
5961
5962
5963
5964
5965
5966
5967
5968
5969
5970
5971
5972
5973
5974
5975
5976
5977
5978
5979
5980
5981
5982
5983
5984
5985
5986
5987
5988
5989
5990
5991
5992
5993

        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)]
5994
5995
5996
        for key_set in self.cudagraph_dispatcher.cudagraph_keys.values():
            key_set.clear()
        self.cudagraph_dispatcher.keys_initialized = False
5997
5998
5999
6000
6001
6002
6003
6004
6005
6006
6007
6008
6009
6010
6011
6012
6013
        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)

6014
    @instrument(span_name="Capture model")
6015
    def capture_model(self) -> int:
6016
        if self.compilation_config.cudagraph_mode == CUDAGraphMode.NONE:
6017
            logger.warning(
6018
                "Skipping CUDA graph capture. To turn on CUDA graph capture, "
6019
6020
                "ensure `cudagraph_mode` was not manually set to `NONE`"
            )
6021
            return 0
6022

6023
6024
6025
6026
6027
6028
6029
6030
6031
6032
6033
6034
6035
        # 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,
            )
6036
            from vllm.v1.worker.encoder_cudagraph import (
6037
6038
                EncoderCudaGraphManager,
            )
6039
6040
6041
6042
6043
6044
6045
6046
6047
6048
6049

            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")

6050
6051
        compilation_counter.num_gpu_runner_capture_triggers += 1

6052
6053
        start_time = time.perf_counter()

6054
6055
6056
        # 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.
6057
        set_cudagraph_capturing_enabled(True)
6058
6059
6060
        with self._freeze_gc(), graph_capture(device=self.device):
            torch.accelerator.synchronize()
            torch.accelerator.empty_cache()
6061
            start_free_gpu_memory = torch.cuda.mem_get_info()[0]
6062

6063
6064
6065
6066
            for (
                runtime_mode,
                batch_descs,
            ) in self.cudagraph_dispatcher.get_capture_descs():
6067
                self._capture_cudagraphs(
6068
6069
                    batch_descriptors=batch_descs,
                    cudagraph_runtime_mode=runtime_mode,
6070
                )
6071
                torch.accelerator.synchronize()
6072

6073
6074
6075
6076
            # Capture encoder CUDA graphs if enabled
            if self.encoder_cudagraph_manager is not None:
                self.encoder_cudagraph_manager.capture()

6077
            torch.accelerator.synchronize()
6078
6079
            end_free_gpu_memory = torch.cuda.mem_get_info()[0]

6080
6081
6082
        # 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
6083
        # we may do lazy capturing in future that still allows capturing
6084
6085
        # after here.
        set_cudagraph_capturing_enabled(False)
6086

6087
6088
6089
        torch.accelerator.synchronize()
        torch.accelerator.empty_cache()

6090
6091
6092
6093
        # Lock workspace to prevent resizing during execution.
        # Max workspace sizes should have been captured during warmup/profiling.
        lock_workspace()

6094
6095
6096
6097
        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.
6098
        logger.info_once(
6099
6100
6101
6102
            "Graph capturing finished in %.0f secs, took %.2f GiB",
            elapsed_time,
            cuda_graph_size / (1 << 30),
        )
6103
        return cuda_graph_size
6104

6105
6106
6107
6108
6109
6110
6111
6112
6113
6114
6115
6116
6117
6118
6119
6120
6121
6122
6123
6124
6125
    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,
6126
                profile_seq_lens=profile_seq_lens,
6127
6128
6129
6130
6131
6132
6133
6134
6135
6136
6137
6138
6139
            )
        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,
        )

6140
6141
    def _capture_cudagraphs(
        self,
6142
        batch_descriptors: list[BatchDescriptor],
6143
6144
6145
6146
        cudagraph_runtime_mode: CUDAGraphMode,
    ):
        assert (
            cudagraph_runtime_mode != CUDAGraphMode.NONE
6147
            and cudagraph_runtime_mode.is_valid_runtime_mode()
6148
        ), f"Invalid cudagraph runtime mode: {cudagraph_runtime_mode}"
6149

6150
6151
6152
6153
6154
        if not batch_descriptors:
            return

        uniform_decode = batch_descriptors[0].uniform

6155
6156
        # Only rank 0 should print progress bar during capture
        if is_global_first_rank():
6157
6158
            batch_descriptors = tqdm(
                batch_descriptors,
6159
6160
6161
                disable=not self.load_config.use_tqdm_on_load,
                desc="Capturing CUDA graphs ({}, {})".format(
                    "decode" if uniform_decode else "mixed prefill-decode",
6162
6163
6164
                    cudagraph_runtime_mode.name,
                ),
            )
6165

6166
        # We skip EPLB here since we don't want to record dummy metrics
6167
        for batch_desc in batch_descriptors:
6168
            # We currently only capture ubatched graphs when its a FULL
6169
6170
6171
            # 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
6172
            allow_microbatching = (
6173
                self.parallel_config.use_ubatching
6174
6175
                and cudagraph_runtime_mode == CUDAGraphMode.FULL
                and uniform_decode
6176
6177
                and check_ubatch_thresholds(
                    config=self.vllm_config.parallel_config,
6178
                    num_tokens=batch_desc.num_tokens,
6179
6180
                    uniform_decode=uniform_decode,
                )
6181
            )
6182
6183
            self._warmup_and_capture(
                batch_desc,
6184
6185
6186
                cudagraph_runtime_mode=cudagraph_runtime_mode,
                allow_microbatching=allow_microbatching,
            )
6187
            torch.accelerator.synchronize()
6188
        self.maybe_remove_all_loras(self.lora_config)
6189

6190
6191
6192
6193
6194
    def initialize_attn_backend(
        self,
        kv_cache_config: KVCacheConfig,
        is_profiling: bool = False,
    ) -> None:
6195
6196
6197
        """
        Initialize the attention backends and attention metadata builders.
        """
6198
        assert len(self.attn_groups) == 0, "Attention backends are already initialized"
6199

6200
6201
6202
6203
6204
6205
        class AttentionGroupKey(NamedTuple):
            attn_backend: type[AttentionBackend]
            kv_cache_spec: KVCacheSpec

        def get_attn_backends_for_group(
            kv_cache_group_spec: KVCacheGroupSpec,
6206
        ) -> tuple[dict[AttentionGroupKey, list[str]], set[type[AttentionBackend]]]:
6207
            layer_type = cast(type[Any], AttentionLayerBase)
6208
            layers = get_layers_from_vllm_config(
6209
                self.vllm_config, layer_type, kv_cache_group_spec.layer_names
6210
            )
6211
6212
            attn_backends = {}
            attn_backend_layers = defaultdict(list)
6213
            # Dedupe based on full class name; this is a bit safer than
6214
6215
6216
6217
            # 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.
6218
            for layer_name in kv_cache_group_spec.layer_names:
6219
                attn_backend = layers[layer_name].get_attn_backend()
6220
6221
6222
6223

                if layer_name in self.kv_sharing_fast_prefill_eligible_layers:
                    attn_backend = create_fast_prefill_custom_backend(
                        "FastPrefill",
6224
                        attn_backend,  # type: ignore[arg-type]
6225
6226
                    )

6227
6228
6229
                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):
6230
                    layer_kv_cache_spec = layer_kv_cache_spec.kv_cache_specs[layer_name]
6231
                key = (full_cls_name, layer_kv_cache_spec)
6232
6233
6234
                attn_backends[key] = AttentionGroupKey(
                    attn_backend, layer_kv_cache_spec
                )
6235
                attn_backend_layers[key].append(layer_name)
6236
6237
6238
6239
            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()),
            )
6240
6241

        def create_attn_groups(
6242
            attn_backends_map: dict[AttentionGroupKey, list[str]],
6243
            kv_cache_group_id: int,
6244
6245
        ) -> list[AttentionGroup]:
            attn_groups: list[AttentionGroup] = []
6246
            for (attn_backend, kv_cache_spec), layer_names in attn_backends_map.items():
6247
                attn_group = AttentionGroup(
6248
                    attn_backend,
6249
                    layer_names,
6250
                    kv_cache_spec,
6251
                    kv_cache_group_id,
6252
6253
                )

6254
6255
6256
                attn_groups.append(attn_group)
            return attn_groups

6257
        attention_backend_maps = []
6258
        attention_backend_list = []
6259
        for kv_cache_group_spec in kv_cache_config.kv_cache_groups:
6260
            attn_backends = get_attn_backends_for_group(kv_cache_group_spec)
6261
            attention_backend_maps.append(attn_backends[0])
6262
            attention_backend_list.append(attn_backends[1])
6263
6264

        # Resolve cudagraph_mode before actually initialize metadata_builders
6265
        self._check_and_update_cudagraph_mode(
6266
6267
6268
            attention_backend_list,
            kv_cache_config.kv_cache_groups,
            is_profiling=is_profiling,
6269
        )
6270

6271
6272
6273
        # Check if attention backend supports PCP&DCP and related features.
        check_attention_cp_compatibility(self.vllm_config)

6274
6275
        for i, attn_backend_map in enumerate(attention_backend_maps):
            self.attn_groups.append(create_attn_groups(attn_backend_map, i))
6276

6277
6278
6279
6280
6281
6282
6283
6284
6285
6286
6287
6288
6289
6290
6291
    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
6292
6293
                    if not self.parallel_config.use_ubatching
                    else self.parallel_config.num_ubatches,
6294
                )
co63oc's avatar
co63oc committed
6295
        # Calculate reorder batch threshold (if needed)
6296
6297
        # Note (tdoublep): do this *after* constructing builders,
        # because some of them change the threshold at init time.
6298
6299
        self.calculate_reorder_batch_threshold()

6300
6301
6302
6303
6304
        # Initialize drafter attention backend
        if self.speculative_config and (
            self.speculative_config.use_eagle()
            or self.speculative_config.uses_draft_model()
        ):
6305
6306
6307
            assert isinstance(
                self.drafter, EagleProposer | DFlashProposer | DraftModelProposer
            )
6308
6309
            self.drafter.initialize_attn_backend(kv_cache_config, kernel_block_sizes)

6310
    def _check_and_update_cudagraph_mode(
6311
6312
6313
        self,
        attention_backends: list[set[type[AttentionBackend]]],
        kv_cache_groups: list[KVCacheGroupSpec],
6314
        is_profiling: bool = False,
6315
    ) -> None:
6316
        """
6317
        Resolve the cudagraph_mode when there are multiple attention
6318
        groups with potential conflicting CUDA graph support.
6319
6320
6321
        Then initialize the cudagraph_dispatcher based on the resolved
        cudagraph_mode.
        """
6322
        min_cg_support = AttentionCGSupport.ALWAYS
6323
        min_cg_attn_backend = None
6324

6325
6326
6327
6328
6329
6330
6331
6332
6333
6334
6335
        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
6336
6337
6338
6339
6340
6341
6342
6343
6344
6345
                    min_cg_attn_backend = attn_backend.__name__
        cudagraph_mode = self.compilation_config.resolve_cudagraph_mode_and_sizes(
            min_cg_support,
            min_cg_attn_backend,
            self.uniform_decode_query_len,
            self.parallel_config.tensor_parallel_size,
            self.kv_cache_config,
            self.max_num_reqs,
            is_profiling=is_profiling,
        )
6346
6347
        # Trigger cudagraph dispatching keys initialization after
        # resolved cudagraph mode.
6348
        self.cudagraph_dispatcher.initialize_cudagraph_keys(
6349
            cudagraph_mode, self.uniform_decode_query_len
6350
        )
6351

6352
6353
6354
6355
6356
        # 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()
        ):
6357
6358
6359
6360
            assert isinstance(
                self.drafter,
                EagleProposer | DFlashProposer | ExtractHiddenStatesProposer,
            )
6361
6362
            self.drafter.initialize_cudagraph_keys(cudagraph_mode)

6363
6364
    def calculate_reorder_batch_threshold(self) -> None:
        """
6365
6366
6367
6368
        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.
6369
        """
6370
6371
        min_none_high = lambda a, b: a if b is None else b if a is None else min(a, b)

6372
        reorder_batch_thresholds: list[int | None] = [
6373
6374
6375
            group.get_metadata_builder().reorder_batch_threshold
            for group in self._attn_group_iterator()
        ]
6376
6377
6378
6379
6380
        # 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
6381
        self.reorder_batch_threshold = reduce(min_none_high, reorder_batch_thresholds)  # type: ignore[assignment]
6382

6383
6384
6385
6386
6387
6388
6389
6390
6391
6392
6393
6394
6395
6396
6397
6398
6399
6400
6401
6402
6403
6404
6405
6406
6407
6408
6409
6410
6411
6412
6413
6414
6415
6416
6417
6418
6419
6420
6421
6422
    def _set_mm_prefix_range_for_metadata(
        self,
        attn_metadata: Any,
        req_doc_ranges: dict[int, list[tuple[int, int]]],
    ) -> None:
        """Set mm_prefix_range for all attention metadata objects.

        This method handles both list and non-list attention metadata,
        computing mm_prefix_range_tensor once and sharing it across all
        metadata objects to avoid redundant host-to-device transfers.
        """
        from vllm.v1.attention.backends.triton_attn import (
            TritonAttentionMetadata,
        )

        # Get all metadata objects from either list or dict structure
        metadata_list = []
        if isinstance(attn_metadata, list):
            for ub_metadata in attn_metadata:
                metadata_list.extend(ub_metadata.values())
        else:
            metadata_list.extend(attn_metadata.values())

        # Set mm_prefix_range for all metadata and compute tensor once
        shared_tensor = None
        for metadata in metadata_list:
            metadata.mm_prefix_range = req_doc_ranges  # type: ignore[attr-defined]

            # Only compute tensor for TritonAttentionMetadata
            if isinstance(metadata, TritonAttentionMetadata):
                if shared_tensor is None:
                    shared_tensor = (
                        TritonAttentionMetadata.compute_mm_prefix_range_tensor(
                            req_doc_ranges,
                            metadata.seq_lens.shape[0],  # type: ignore[attr-defined]
                            metadata.seq_lens.device,  # type: ignore[attr-defined]
                        )
                    )
                metadata.mm_prefix_range_tensor = shared_tensor

6423
6424
6425
    def may_reinitialize_input_batch(
        self, kv_cache_config: KVCacheConfig, kernel_block_sizes: list[int]
    ) -> None:
6426
6427
        """
        Re-initialize the input batch if the block sizes are different from
6428
6429
6430
6431
        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.
6432
6433
6434

        Args:
            kv_cache_config: The KV cache configuration.
6435
            kernel_block_sizes: The kernel block sizes for each KV cache group.
6436
        """
6437
        block_sizes = []
6438
6439
        max_num_blocks = []
        max_model_len = max(self.max_model_len, self.max_encoder_len)
6440
        for kv_cache_group in kv_cache_config.kv_cache_groups:
6441
6442
            if isinstance(kv_cache_group.kv_cache_spec, EncoderOnlyAttentionSpec):
                continue
6443
6444
            block_size = kv_cache_group.kv_cache_spec.block_size
            block_sizes.append(block_size)
6445
            max_num_blocks_per_req = cdiv(
6446
                max_model_len, block_size * get_total_cp_world_size()
6447
6448
            )
            if isinstance(kv_cache_group.kv_cache_spec, MambaSpec):
6449
                max_num_blocks_per_req = (
6450
6451
6452
6453
6454
                    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)
6455

6456
6457
6458
6459
6460
6461
        if (
            block_sizes != self._init_block_sizes
            or kernel_block_sizes != self._init_kernel_block_sizes
        ):
            self._init_block_sizes = block_sizes
            self._init_kernel_block_sizes = kernel_block_sizes
6462
6463
            self.input_batch = InputBatch(
                max_num_reqs=self.max_num_reqs,
6464
                max_model_len=max_model_len,
6465
6466
6467
6468
6469
                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,
6470
                kernel_block_sizes=kernel_block_sizes,
6471
                max_num_blocks_per_req=max_num_blocks,
6472
                is_spec_decode=bool(self.vllm_config.speculative_config),
6473
                logitsprocs=self.input_batch.logitsprocs,
6474
                logitsprocs_need_output_token_ids=self.input_batch.logitsprocs_need_output_token_ids,
6475
                is_pooling_model=self.is_pooling_model,
6476
6477
            )

6478
6479
6480
6481
6482
6483
6484
6485
6486
        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}"
        )

6487
    def _allocate_kv_cache_tensors(
6488
6489
        self, kv_cache_config: KVCacheConfig
    ) -> dict[str, torch.Tensor]:
6490
        """
6491
6492
6493
        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.

6494
        Args:
6495
            kv_cache_config: The KV cache config
6496
        Returns:
6497
            dict[str, torch.Tensor]: A map between layer names to their
6498
            corresponding memory buffer for KV cache.
6499
        """
6500
6501
        kv_cache_raw_tensors: dict[str, torch.Tensor] = {}
        for kv_cache_tensor in kv_cache_config.kv_cache_tensors:
6502
6503
6504
            tensor = torch.zeros(
                kv_cache_tensor.size, dtype=torch.int8, device=self.device
            )
6505
6506
6507
6508
6509
            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:
6510
6511
6512
6513
            for layer_name in group.layer_names:
                if layer_name in self.runner_only_attn_layers:
                    continue
                layer_names.add(layer_name)
6514
6515
6516
        assert layer_names == set(kv_cache_raw_tensors.keys()), (
            "Some layers are not correctly initialized"
        )
6517
6518
        return kv_cache_raw_tensors

6519
6520
6521
    def _attn_group_iterator(self) -> Iterator[AttentionGroup]:
        return itertools.chain.from_iterable(self.attn_groups)

6522
    def _kv_cache_spec_attn_group_iterator(self) -> Iterator[AttentionGroup]:
6523
6524
        if not self.kv_cache_config.kv_cache_groups:
            return
6525
6526
        for attn_groups in self.attn_groups:
            yield from attn_groups
6527

6528
6529
6530
6531
    def _reshape_kv_cache_tensors(
        self,
        kv_cache_config: KVCacheConfig,
        kv_cache_raw_tensors: dict[str, torch.Tensor],
6532
        kernel_block_sizes: list[int],
6533
    ) -> dict[str, torch.Tensor]:
6534
        """
6535
        Reshape the KV cache tensors to the desired shape and dtype.
6536

6537
        Args:
6538
6539
            kv_cache_config: The KV cache config
            kv_cache_raw_tensors: The KV cache buffer of each layer, with
6540
                correct size but uninitialized shape.
6541
            kernel_block_sizes: The kernel block sizes for each KV cache group.
6542
        Returns:
6543
            Dict[str, torch.Tensor]: A map between layer names to their
6544
6545
            corresponding memory buffer for KV cache.
        """
6546
        kv_caches: dict[str, torch.Tensor] = {}
6547
        has_attn, has_mamba = False, False
6548
6549
        for group in self._kv_cache_spec_attn_group_iterator():
            kv_cache_spec = group.kv_cache_spec
6550
            attn_backend = group.backend
6551
6552
6553
6554
            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]
6555
            for layer_name in group.layer_names:
6556
6557
                if layer_name in self.runner_only_attn_layers:
                    continue
6558
6559
                raw_tensor = kv_cache_raw_tensors[layer_name]
                assert raw_tensor.numel() % kv_cache_spec.page_size_bytes == 0
6560
                num_blocks = raw_tensor.numel() // kv_cache_spec.page_size_bytes
6561
                if isinstance(kv_cache_spec, AttentionSpec):
6562
                    has_attn = True
6563
6564
                    num_blocks_per_kv_block = (
                        kv_cache_spec.block_size // kernel_block_size
6565
6566
6567
                    )
                    kernel_num_blocks = num_blocks * num_blocks_per_kv_block

6568
                    kv_cache_shape = attn_backend.get_kv_cache_shape(
6569
                        kernel_num_blocks,
6570
                        kernel_block_size,
6571
6572
                        kv_cache_spec.num_kv_heads,
                        kv_cache_spec.head_size,
6573
6574
                        cache_dtype_str=self.cache_config.cache_dtype,
                    )
6575
                    dtype = kv_cache_spec.dtype
6576
                    try:
6577
                        kv_cache_stride_order = attn_backend.get_kv_cache_stride_order()
6578
                        assert len(kv_cache_stride_order) == len(kv_cache_shape)
6579
                    except (AttributeError, NotImplementedError):
6580
                        kv_cache_stride_order = tuple(range(len(kv_cache_shape)))
6581
6582
6583
6584
6585
                    # 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.
6586
6587
6588
                    kv_cache_shape = tuple(
                        kv_cache_shape[i] for i in kv_cache_stride_order
                    )
6589
6590
6591
6592
6593
                    # Maintain original KV shape view.
                    inv_order = [
                        kv_cache_stride_order.index(i)
                        for i in range(len(kv_cache_stride_order))
                    ]
6594
6595
6596
6597
6598
6599
                    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
6600
                elif isinstance(kv_cache_spec, MambaSpec):
6601
                    has_mamba = True
Chen Zhang's avatar
Chen Zhang committed
6602
6603
                    raw_tensor = kv_cache_raw_tensors[layer_name]
                    state_tensors = []
6604
                    storage_offset_bytes = 0
6605
                    for shape, dtype in zip(kv_cache_spec.shapes, kv_cache_spec.dtypes):
6606
6607
                        dtype_size = get_dtype_size(dtype)
                        num_element_per_page = (
6608
6609
                            kv_cache_spec.page_size_bytes // dtype_size
                        )
Chen Zhang's avatar
Chen Zhang committed
6610
                        target_shape = (num_blocks, *shape)
6611
6612
                        stride = torch.empty(target_shape).stride()
                        target_stride = (num_element_per_page, *stride[1:])
6613
                        assert storage_offset_bytes % dtype_size == 0
6614
6615
6616
6617
                        tensor = torch.as_strided(
                            raw_tensor.view(dtype),
                            size=target_shape,
                            stride=target_stride,
6618
                            storage_offset=storage_offset_bytes // dtype_size,
6619
                        )
Chen Zhang's avatar
Chen Zhang committed
6620
                        state_tensors.append(tensor)
6621
                        storage_offset_bytes += stride[0] * dtype_size
6622
6623

                    kv_caches[layer_name] = state_tensors
6624
                else:
6625
                    raise NotImplementedError
6626
6627

        if has_attn and has_mamba:
6628
            self._update_hybrid_attention_mamba_layout(kv_caches, kernel_block_sizes)
6629

6630
6631
        return kv_caches

6632
    def _update_hybrid_attention_mamba_layout(
6633
        self, kv_caches: dict[str, torch.Tensor], kernel_block_sizes: list[int]
6634
    ) -> None:
6635
        """
6636
6637
        Update the layout of attention layers from (2, num_blocks, ...) to
        (num_blocks, 2, ...).
6638
6639

        Args:
6640
            kv_caches: The KV cache buffer of each layer.
6641
            kernel_block_sizes: The kernel block sizes for each KV cache group.
6642
6643
        """

6644
6645
        for group in self._kv_cache_spec_attn_group_iterator():
            kv_cache_spec = group.kv_cache_spec
6646
6647
6648
6649
6650
6651
6652
6653
6654
6655
6656
6657
            if not isinstance(kv_cache_spec, AttentionSpec):
                continue
            block_dim = group.backend.get_kv_cache_block_dim(
                kernel_block_sizes[group.kv_cache_group_id],
                kv_cache_spec.num_kv_heads,
                kv_cache_spec.head_size,
                cache_dtype_str=self.cache_config.cache_dtype,
            )
            # block_dim: 0 means (num_blocks, 2, ...); 1 means (2, num_blocks, ...).
            if block_dim == 0:
                continue
            assert block_dim == 1
6658
            for layer_name in group.layer_names:
6659
                kv_cache = kv_caches[layer_name]
6660
6661
6662
6663
6664
                hidden_size = kv_cache.shape[2:].numel()
                kv_cache.as_strided_(
                    size=kv_cache.shape,
                    stride=(hidden_size, 2 * hidden_size, *kv_cache.stride()[2:]),
                )
6665

6666
    def initialize_kv_cache_tensors(
6667
        self, kv_cache_config: KVCacheConfig, kernel_block_sizes: list[int]
6668
    ) -> dict[str, torch.Tensor]:
6669
6670
6671
6672
6673
        """
        Initialize the memory buffer for KV cache.

        Args:
            kv_cache_config: The KV cache config
6674
6675
            kernel_block_sizes: The kernel block sizes for each KV cache group.

6676
        Returns:
6677
            Dict[str, torch.Tensor]: A map between layer names to their
6678
6679
            corresponding memory buffer for KV cache.
        """
6680
6681
6682
6683
6684
6685
6686
6687
6688
6689
6690
6691
6692
6693
6694
6695
6696
6697
6698
6699
6700
6701
6702
6703

        # 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
            )
6704

6705
        # Set up cross-layer KV cache sharing
6706
6707
        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)
6708
6709
            kv_caches[layer_name] = kv_caches[target_layer_name]

6710
6711
6712
6713
6714
6715
6716
6717
6718
        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,
        )
6719
6720
6721
        return kv_caches

    def maybe_add_kv_sharing_layers_to_kv_cache_groups(
6722
6723
        self, kv_cache_config: KVCacheConfig
    ) -> None:
6724
6725
6726
6727
6728
6729
6730
6731
6732
6733
6734
6735
6736
6737
6738
6739
6740
6741
        """
        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.
6742
            attn_layers = get_layers_from_vllm_config(self.vllm_config, Attention)
6743
6744
            for layer_name in reversed(attn_layers):
                if layer_name in self.shared_kv_cache_layers:
6745
                    self.kv_sharing_fast_prefill_eligible_layers.add(layer_name)
6746
6747
                else:
                    break
6748

6749
6750
6751
6752
6753
    def initialize_kv_cache(
        self,
        kv_cache_config: KVCacheConfig,
        is_profiling: bool = False,
    ) -> None:
6754
6755
6756
6757
6758
6759
        """
        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
        """
6760
        kv_cache_config = deepcopy(kv_cache_config)
6761
        self.kv_cache_config = kv_cache_config
6762
        self._mamba_copy_bufs = None
6763
        self.may_add_encoder_only_layers_to_kv_cache_config()
6764
        self.maybe_add_kv_sharing_layers_to_kv_cache_groups(kv_cache_config)
6765
        self.initialize_attn_backend(kv_cache_config, is_profiling=is_profiling)
roikoren755's avatar
roikoren755 committed
6766
6767
6768
        initialize_mamba_ssu_backend(
            self.vllm_config.mamba_config, self.kv_cache_config
        )
6769
6770
6771
6772
6773
        # 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.
6774
6775
6776
        kernel_block_sizes = prepare_kernel_block_sizes(
            kv_cache_config, self.attn_groups
        )
6777
        self._kernel_block_sizes = kernel_block_sizes
6778
6779
6780
6781

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

6782
        # Reinitialize need to after initialize_attn_backend
6783
6784
6785
6786
        self.may_reinitialize_input_batch(kv_cache_config, kernel_block_sizes)
        kv_caches = self.initialize_kv_cache_tensors(
            kv_cache_config, kernel_block_sizes
        )
6787

6788
6789
6790
        if (
            self.speculative_config
            and self.speculative_config.uses_extract_hidden_states()
6791
        ):
6792
            assert isinstance(self.drafter, ExtractHiddenStatesProposer)
6793
6794
6795
6796
            # validate all draft model layers belong to the same kv cache
            # group
            self.drafter.validate_same_kv_cache_group(kv_cache_config)

6797
        if has_kv_transfer_group() and not is_profiling:
6798
            kv_transfer_group = get_kv_transfer_group()
6799
6800
6801
6802
6803
6804
6805
            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)
6806
            kv_transfer_group.set_host_xfer_buffer_ops(copy_kv_blocks)
Robert Shaw's avatar
Robert Shaw committed
6807

6808
6809
6810
6811
6812
6813
    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
6814
6815
6816
6817
6818
6819
6820

    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()
6821
6822
6823
6824
6825
6826
6827
6828
        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)
6829
        self.max_num_kv_tokens = (
6830
6831
6832
6833
6834
6835
6836
            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

6837
6838
6839
        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,
6840
            vllm_config=self.vllm_config,
6841
        )
6842
        self._bind_routed_experts_capturer(routed_experts_capturer)
6843
        self.routed_experts_initialized = True
6844
6845
6846
6847
6848
6849
6850
6851
6852
6853
6854
6855
6856
6857
6858

    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)
6859

6860
6861
6862
6863
6864
    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
6865
        encoder_only_attn_specs: dict[AttentionSpec, list[str]] = defaultdict(list)
6866
6867
6868
        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:
6869
                attn_spec: AttentionSpec = EncoderOnlyAttentionSpec(
6870
6871
6872
                    block_size=block_size,
                    num_kv_heads=attn_module.num_kv_heads,
                    head_size=attn_module.head_size,
6873
6874
                    dtype=self.kv_cache_dtype,
                )
6875
6876
6877
                encoder_only_attn_specs[attn_spec].append(layer_name)
                self.runner_only_attn_layers.add(layer_name)
        if len(encoder_only_attn_specs) > 0:
6878
6879
6880
            assert len(encoder_only_attn_specs) == 1, (
                "Only support one encoder-only attention spec now"
            )
6881
6882
            spec, layer_names = encoder_only_attn_specs.popitem()
            self.kv_cache_config.kv_cache_groups.append(
6883
6884
                KVCacheGroupSpec(layer_names=layer_names, kv_cache_spec=spec)
            )
6885

6886
    def get_kv_cache_spec(self) -> dict[str, KVCacheSpec]:
6887
        """
6888
        Generates the KVCacheSpec by parsing the kv cache format from each
6889
6890
        Attention module in the static forward context.
        Returns:
6891
            KVCacheSpec: A dictionary mapping layer names to their KV cache
6892
6893
            format. Layers that do not need KV cache are not included.
        """
6894
        if has_ec_transfer() and not get_ec_transfer().is_consumer:
6895
            return {}
6896
        kv_cache_spec: dict[str, KVCacheSpec] = {}
6897
6898
        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
6899
        for layer_name, attn_module in attn_layers.items():
6900
6901
6902
6903
6904
6905
6906
6907
6908
6909
6910
6911
6912
6913
6914
            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
6915

6916
        return kv_cache_spec
6917

6918
    def _to_list(self, sampled_token_ids: torch.Tensor) -> list[list[int]]:
6919
6920
6921
6922
6923
6924
6925
6926
        # 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.
6927
        pinned = self.sampled_token_ids_pinned_cpu[: sampled_token_ids.shape[0]]
6928
6929
6930
        pinned.copy_(sampled_token_ids, non_blocking=True)
        self.transfer_event.record()
        self.transfer_event.synchronize()
6931
        return pinned.tolist()
6932
6933
6934
6935
6936
6937
6938
6939
6940
6941
6942
6943
6944
6945
6946
6947
6948
6949
6950
6951
6952
6953
6954
6955
6956
6957
6958
6959
6960
6961
6962
6963
6964
6965
6966
6967
6968
6969
6970
6971

    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}

6972
        torch.accelerator.synchronize()
6973
6974
6975
6976
6977
        start_time = time.perf_counter()

        try:
            yield
        finally:
6978
            torch.accelerator.synchronize()
6979
6980
6981
6982
6983
6984
6985
6986
6987
6988
            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]
6989
                    stats.encoder_forward_secs += per_request_time
6990
6991
6992
6993
6994
6995
6996
                    stats.num_encoder_calls += 1


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

6997
    encoder_forward_secs: float = 0.0
6998
6999
7000
7001
7002
7003
7004
    """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 {
7005
            "encoder_forward_secs": self.encoder_forward_secs,
7006
7007
            "num_encoder_calls": self.num_encoder_calls,
        }