gpu_model_runner.py 270 KB
Newer Older
1
# SPDX-License-Identifier: Apache-2.0
2
# SPDX-FileCopyrightText: Copyright contributors to the vLLM project
3

4
import functools
5
import gc
6
import itertools
7
import threading
8
import time
9
from collections import defaultdict
10
from collections.abc import Iterable, Iterator, Sequence
11
from contextlib import contextmanager
12
from copy import copy, deepcopy
13
from dataclasses import dataclass
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
32
33
    CUDAGraphMode,
    VllmConfig,
    get_layers_from_vllm_config,
    update_config,
)
34
from vllm.distributed.ec_transfer import get_ec_transfer, has_ec_transfer
35
from vllm.distributed.eplb.eplb_state import EplbState
36
from vllm.distributed.kv_transfer import get_kv_transfer_group, has_kv_transfer_group
37
from vllm.distributed.kv_transfer.kv_connector.utils import copy_kv_blocks
38
from vllm.distributed.parallel_state import (
39
    get_dcp_group,
40
41
42
43
44
45
    get_pp_group,
    get_tp_group,
    graph_capture,
    is_global_first_rank,
    prepare_communication_buffer_for_model,
)
46
47
48
49
from vllm.forward_context import (
    BatchDescriptor,
    set_forward_context,
)
50
from vllm.logger import init_logger
51
from vllm.lora.layers import LoRAMapping, LoRAMappingType
52
from vllm.model_executor.layers.attention import Attention, MLAAttention
53
from vllm.model_executor.layers.attention_layer_base import AttentionLayerBase
54
55
56
from vllm.model_executor.layers.fused_moe.routed_experts_capturer import (
    RoutedExpertsCapturer,
)
57
58
59
60
from vllm.model_executor.layers.rotary_embedding import (
    MRotaryEmbedding,
    XDRotaryEmbedding,
)
61
from vllm.model_executor.model_loader import get_model_loader
62
63
64
65
from vllm.model_executor.model_loader.reload import (
    finalize_layerwise_reload,
    initialize_layerwise_reload,
)
66
from vllm.model_executor.models.interfaces import (
67
    MultiModalEmbeddings,
68
    SupportsMRoPE,
69
    SupportsMultiModal,
70
    SupportsXDRoPE,
71
72
73
74
    is_mixture_of_experts,
    supports_eagle3,
    supports_mrope,
    supports_multimodal_pruning,
75
    supports_realtime,
76
    supports_transcription,
77
    supports_xdrope,
78
)
79
from vllm.model_executor.models.interfaces_base import (
80
81
82
83
    VllmModelForPooling,
    is_pooling_model,
    is_text_generation_model,
)
84
85
86
87
88
from vllm.model_executor.offloader import (
    create_offloader,
    get_offloader,
    set_offloader,
)
89
from vllm.multimodal import MULTIMODAL_REGISTRY
90
from vllm.multimodal.encoder_budget import MultiModalBudget
91
92
93
94
95
from vllm.multimodal.inputs import (
    BatchedTensorInputs,
    MultiModalKwargsItem,
    PlaceholderRange,
)
96
from vllm.multimodal.utils import group_mm_kwargs_by_modality
97
from vllm.pooling_params import PoolingParams
98
from vllm.sampling_params import SamplingType
99
from vllm.sequence import IntermediateTensors
100
from vllm.tasks import GenerationTask, PoolingTask, SupportedTask
101
from vllm.tracing import instrument
102
from vllm.utils import length_from_prompt_token_ids_or_embeds
103
from vllm.utils.math_utils import cdiv, round_up
104
from vllm.utils.mem_utils import DeviceMemoryProfiler, format_gib
105
from vllm.utils.nvtx_pytorch_hooks import PytHooks
106
from vllm.utils.platform_utils import is_pin_memory_available, num_compute_units
107
108
109
110
from vllm.utils.torch_utils import (
    get_dtype_size,
    kv_cache_dtype_str_to_dtype,
)
111
112
from vllm.v1.attention.backend import (
    AttentionBackend,
113
    AttentionCGSupport,
114
    AttentionMetadata,
115
    AttentionMetadataBuilder,
116
    AttentionType,
117
    CommonAttentionMetadata,
118
)
119
from vllm.v1.attention.backends.gdn_attn import GDNAttentionMetadataBuilder
120
from vllm.v1.attention.backends.mamba2_attn import Mamba2AttentionMetadataBuilder
121
from vllm.v1.attention.backends.utils import (
122
    create_fast_prefill_custom_backend,
123
    get_dcp_local_seq_lens,
124
125
    reorder_batch_to_split_decodes_and_prefills,
)
126
from vllm.v1.core.sched.output import NewRequestData
127
from vllm.v1.cudagraph_dispatcher import CudagraphDispatcher
128
129
130
131
132
133
134
135
136
137
138
139
140
141
142
143
144
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,
145
    ECConnectorOutput,
146
    KVConnectorOutput,
147
148
149
150
151
    LogprobsLists,
    LogprobsTensors,
    ModelRunnerOutput,
    PoolerOutput,
    SamplerOutput,
152
    make_empty_encoder_model_runner_output,
153
)
154
from vllm.v1.pool.metadata import PoolingMetadata, PoolingStates
155
from vllm.v1.sample.logits_processor import LogitsProcessors, build_logitsprocs
156
from vllm.v1.sample.logits_processor.interface import LogitsProcessor
157
from vllm.v1.sample.metadata import SamplingMetadata
158
from vllm.v1.sample.rejection_sampler import RejectionSampler
159
from vllm.v1.sample.sampler import Sampler
160
from vllm.v1.spec_decode.draft_model import DraftModelProposer
161
from vllm.v1.spec_decode.eagle import EagleProposer
162
from vllm.v1.spec_decode.extract_hidden_states import ExtractHiddenStatesProposer
163
from vllm.v1.spec_decode.medusa import MedusaProposer
164
from vllm.v1.spec_decode.metadata import SpecDecodeMetadata
165
from vllm.v1.spec_decode.suffix_decoding import SuffixDecodingProposer
166
from vllm.v1.structured_output.utils import apply_grammar_bitmask
167
from vllm.v1.utils import CpuGpuBuffer, record_function_or_nullcontext
168
169
170
171
172
from vllm.v1.worker import mamba_utils
from vllm.v1.worker.cp_utils import (
    check_attention_cp_compatibility,
    get_total_cp_world_size,
)
173
from vllm.v1.worker.dp_utils import coordinate_batch_across_dp
174
from vllm.v1.worker.ec_connector_model_runner_mixin import ECConnectorModelRunnerMixin
175
from vllm.v1.worker.gpu_input_batch import CachedRequestState, InputBatch
176
from vllm.v1.worker.gpu_ubatch_wrapper import UBatchWrapper
177
from vllm.v1.worker.kv_connector_model_runner_mixin import KVConnectorModelRunnerMixin
178
from vllm.v1.worker.lora_model_runner_mixin import LoRAModelRunnerMixin
179
180
181
from vllm.v1.worker.ubatch_utils import (
    UBatchSlices,
    check_ubatch_thresholds,
182
    maybe_create_ubatch_slices,
183
    split_attn_metadata,
184
)
185
from vllm.v1.worker.utils import is_residual_scattered_for_sp
186
from vllm.v1.worker.workspace import lock_workspace
187

188
189
190
191
from .utils import (
    AttentionGroup,
    add_kv_sharing_layers_to_kv_cache_groups,
    bind_kv_cache,
192
    prepare_kernel_block_sizes,
193
194
    sanity_check_mm_encoder_outputs,
)
195

196
if TYPE_CHECKING:
197
    from vllm.v1.core.sched.output import GrammarOutput, SchedulerOutput
198
    from vllm.v1.spec_decode.ngram_proposer import NgramProposer
199
200
201

logger = init_logger(__name__)

202
203
AttnMetadataDict: TypeAlias = dict[str, AttentionMetadata]
# list when ubatching is enabled
204
PerLayerAttnMetadata: TypeAlias = list[AttnMetadataDict] | AttnMetadataDict
205

206

207
208
209
210
211
212
# Wrapper for ModelRunnerOutput to support overlapped execution.
class AsyncGPUModelRunnerOutput(AsyncModelRunnerOutput):
    def __init__(
        self,
        model_runner_output: ModelRunnerOutput,
        sampled_token_ids: torch.Tensor,
213
        logprobs_tensors: LogprobsTensors | None,
214
215
        invalid_req_indices: list[int],
        async_output_copy_stream: torch.cuda.Stream,
216
        vocab_size: int,
217
218
219
220
221
    ):
        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.
222
        self.async_copy_ready_event = torch.Event()
223
224
225
226

        # 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
227
        self.vocab_size = vocab_size
228
        self._logprobs_tensors = logprobs_tensors
229
230
231
232
233

        # 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)
234
            self.sampled_token_ids_cpu = self._sampled_token_ids.to(
235
236
                "cpu", non_blocking=True
            )
237
238
239
240
241
            self._logprobs_tensors_cpu = (
                self._logprobs_tensors.to_cpu_nonblocking()
                if self._logprobs_tensors
                else None
            )
242
            self.async_copy_ready_event.record()
243
244
245

    def get_output(self) -> ModelRunnerOutput:
        """Copy the device tensors to the host and return a ModelRunnerOutput.
246

247
248
        This function blocks until the copy is finished.
        """
249
        max_gen_len = self.sampled_token_ids_cpu.shape[-1]
250
        self.async_copy_ready_event.synchronize()
251

252
253
        # Release the device tensors once the copy has completed.
        del self._logprobs_tensors
254
        del self._sampled_token_ids
255
        if max_gen_len == 1:
256
            valid_sampled_token_ids = self.sampled_token_ids_cpu.tolist()
257
258
            for i in self._invalid_req_indices:
                valid_sampled_token_ids[i].clear()
259
260
261
            logprobs_lists = None
            if self._logprobs_tensors_cpu is not None:
                logprobs_lists = self._logprobs_tensors_cpu.tolists()
262
        else:
263
            valid_sampled_token_ids, logprobs_lists = RejectionSampler.parse_output(
264
265
                self.sampled_token_ids_cpu,
                self.vocab_size,
266
                self._invalid_req_indices,
267
                logprobs_tensors=self._logprobs_tensors_cpu,
268
            )
269
270
271

        output = self._model_runner_output
        output.sampled_token_ids = valid_sampled_token_ids
272
        output.logprobs = logprobs_lists
273
274
275
        return output


276
277
278
279
280
281
282
283
284
285
286
287
288
289
290
291
292
293
294
295
296
297
298
299
300
301
302
303
304
305
306
307
308
309
310
311
312
313
314
315
316
317
318
319
320
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


321
322
323
324
325
326
327
328
329
330
331
332
333
334
335
336
337
338
339
340
341
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)
342
343
344
            self._model_runner_output.pooler_output = _copy_pooler_output_to_cpu(
                raw_pooler_output=self._raw_pooler_output,
                finished_mask=finished_mask,
345
346
347
348
349
350
351
352
353
354
355
356
357
358
            )
            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


359
360
361
362
363
364
365
366
367
368
369
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
370
    ec_connector_output: ECConnectorOutput | None
371
    cudagraph_stats: CUDAGraphStat | None
372
    slot_mappings: dict[str, torch.Tensor] | list[dict[str, torch.Tensor]] | None
373
374


375
376
377
class GPUModelRunner(
    LoRAModelRunnerMixin, KVConnectorModelRunnerMixin, ECConnectorModelRunnerMixin
):
378
379
    def __init__(
        self,
380
        vllm_config: VllmConfig,
381
        device: torch.device,
382
    ):
383
384
385
        self.vllm_config = vllm_config
        self.model_config = vllm_config.model_config
        self.cache_config = vllm_config.cache_config
386
        self.offload_config = vllm_config.offload_config
387
        self.compilation_config = vllm_config.compilation_config
388
389
390
391
392
393
        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
394

395
396
397
398
        model_config = self.model_config
        cache_config = self.cache_config
        scheduler_config = self.scheduler_config
        parallel_config = self.parallel_config
399
        self.device = device
400
401
        self.pin_memory = is_pin_memory_available()
        self.dtype = self.model_config.dtype
402

403
404
405
        self.kv_cache_dtype = kv_cache_dtype_str_to_dtype(
            cache_config.cache_dtype, self.model_config
        )
406

407
        self.is_pooling_model = model_config.runner_type == "pooling"
408
        self.enable_prompt_embeds = model_config.enable_prompt_embeds
409
        self.is_multimodal_raw_input_only_model = (
410
411
            model_config.is_multimodal_raw_input_only_model
        )
412
413
        # This will be overridden in load_model()
        self.is_multimodal_pruning_enabled = False
414
        self.max_model_len = model_config.max_model_len
415
416
417

        # Always set to false after the first forward pass
        self.calculate_kv_scales = self.cache_config.calculate_kv_scales
418
        self.dcp_world_size = self.parallel_config.decode_context_parallel_size
419
        self.dcp_rank = 0 if self.dcp_world_size <= 1 else get_dcp_group().rank_in_group
420
        self.max_num_tokens = scheduler_config.max_num_batched_tokens
421
        self.max_num_reqs = scheduler_config.max_num_seqs
422

423
424
425
426
427
        # Broadcast PP output for external_launcher (torchrun)
        # to make sure we are synced across pp ranks
        # TODO: Support overlapping mirco-batches
        # https://github.com/vllm-project/vllm/issues/18019
        self.broadcast_pp_output = (
428
            self.parallel_config.distributed_executor_backend == "external_launcher"
429
            and len(get_pp_group().ranks) > 1
430
        )
431

432
        # Model-related.
433
        self.num_query_heads = model_config.get_num_attention_heads(parallel_config)
434
        self.inputs_embeds_size = model_config.get_inputs_embeds_size()
435
        self.attention_chunk_size = model_config.attention_chunk_size
436
        # Only relevant for models using ALiBi (e.g, MPT)
437
        self.use_alibi = model_config.uses_alibi
438

439
        self.cascade_attn_enabled = not self.model_config.disable_cascade_attn
440
        self.is_mm_prefix_lm = self.model_config.is_mm_prefix_lm
441

442
        # Multi-modal data support
443
        self.mm_registry = MULTIMODAL_REGISTRY
444
        self.uses_mrope = model_config.uses_mrope
445
        self.uses_xdrope_dim = model_config.uses_xdrope_dim
446
        self.supports_mm_inputs = self.mm_registry.supports_multimodal_inputs(
447
            model_config
448
        )
449

450
451
452
        if self.model_config.is_encoder_decoder:
            # Maximum length of the encoder input, only for encoder-decoder
            # models.
453
            self.max_encoder_len = scheduler_config.max_num_encoder_input_tokens
454
455
456
        else:
            self.max_encoder_len = 0

457
458
459
        # Async scheduling
        self.use_async_scheduling = self.scheduler_config.async_scheduling

460
        # Sampler
461
        self.sampler = Sampler(logprobs_mode=self.model_config.logprobs_mode)
462

463
        self.eplb_state: EplbState | None = None
464
465
        # NOTE(yongji): flag to temporarily disable EPLB during scaling up/down
        self.eep_eplb_suppressed = False
466
467
468
469
470
471
        """
        State of the expert parallelism load balancer.

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

472
        # Lazy initializations
473
        # self.model: nn.Module  # Set after load_model
474
        # Initialize in initialize_kv_cache
475
        self.kv_caches: list[torch.Tensor] = []
476
477
478
        # Initialize in initialize_kv_cache_tensors
        self.cross_layers_kv_cache: torch.Tensor | None = None
        self.cross_layers_attn_backend: type[AttentionBackend] | None = None
479
480
        # indexes: [kv_cache_group_id][attn_group]
        self.attn_groups: list[list[AttentionGroup]] = []
481
482
        # self.kv_cache_config: KVCacheConfig

483
484
        # mm_hash ->  encoder_output
        self.encoder_cache: dict[str, torch.Tensor] = {}
485

486
        self.use_aux_hidden_state_outputs = False
487
488
489
490
491
        # 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:
492
            self.drafter: (
493
                NgramProposer  # noqa: F823
494
495
496
497
                | SuffixDecodingProposer
                | EagleProposer
                | DraftModelProposer
                | MedusaProposer
498
                | ExtractHiddenStatesProposer
499
            )
500
            if self.speculative_config.method == "ngram":
501
502
                from vllm.v1.spec_decode.ngram_proposer import NgramProposer

503
                self.drafter = NgramProposer(self.vllm_config)
504
505
506
507
508
509
            elif self.speculative_config.uses_draft_model():
                self.drafter = DraftModelProposer(
                    vllm_config=self.vllm_config,
                    device=self.device,
                    runner=self,
                )
510
511
            elif self.speculative_config.method == "suffix":
                self.drafter = SuffixDecodingProposer(self.vllm_config)
512
            elif self.speculative_config.use_eagle():
513
                self.drafter = EagleProposer(self.vllm_config, self.device, self)
514
                if self.speculative_config.method == "eagle3":
515
516
517
                    self.use_aux_hidden_state_outputs = (
                        self.drafter.eagle3_use_aux_hidden_state
                    )
518
519
            elif self.speculative_config.method == "medusa":
                self.drafter = MedusaProposer(
520
                    vllm_config=self.vllm_config, device=self.device
521
                )
522
523
524
525
526
            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
527
            else:
528
529
530
531
                raise ValueError(
                    "Unknown speculative decoding method: "
                    f"{self.speculative_config.method}"
                )
532
            self.rejection_sampler = RejectionSampler(self.sampler)
533

534
535
536
        self.num_spec_tokens = 0
        if self.speculative_config:
            self.num_spec_tokens = self.speculative_config.num_speculative_tokens
537
538
539
540
541
            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
542

543
        # Request states.
544
        self.requests: dict[str, CachedRequestState] = {}
545
546
547
        # NOTE(rob): num_prompt_logprobs only includes reqs
        # that are currently in the prefill phase.
        self.num_prompt_logprobs: dict[str, int] = {}
548
        self.comm_stream = torch.cuda.Stream()
549

550
551
552
553
554
555
556
557
558
        # 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.
559
560
561
562
        logits_processors = model_config.logits_processors
        custom_logitsprocs: Sequence[str | type[LogitsProcessor]] = (
            tuple(logits_processors) if logits_processors is not None else ()
        )
563
564
        self.input_batch = InputBatch(
            max_num_reqs=self.max_num_reqs,
565
566
567
            # We need to use the encoder length for encoder-decoer
            # because of KV cache for cross-attention.
            max_model_len=max(self.max_model_len, self.max_encoder_len),
568
569
570
            max_num_batched_tokens=self.max_num_tokens,
            device=self.device,
            pin_memory=self.pin_memory,
571
            vocab_size=self.model_config.get_vocab_size(),
572
573
            block_sizes=[self.cache_config.block_size],
            kernel_block_sizes=[self.cache_config.block_size],
574
            is_spec_decode=bool(self.vllm_config.speculative_config),
575
            logitsprocs=build_logitsprocs(
576
577
578
                self.vllm_config,
                self.device,
                self.pin_memory,
579
                self.is_pooling_model,
580
                custom_logitsprocs,
581
            ),
582
583
584
            # We currently don't know whether a particular custom logits processor
            # uses output token ids so we set this conservatively.
            logitsprocs_need_output_token_ids=bool(custom_logitsprocs),
585
            is_pooling_model=self.is_pooling_model,
586
            cp_kv_cache_interleave_size=self.parallel_config.cp_kv_cache_interleave_size,
587
        )
588

589
590
591
592
593
        # 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.
594
        self.prepare_inputs_event: torch.Event | None = None
595
596
        if self.use_async_scheduling:
            self.async_output_copy_stream = torch.cuda.Stream()
597
            self.prepare_inputs_event = torch.Event()
598

599
        # Cache the device properties.
600
        self._init_device_properties()
601

602
603
604
605
        # Encoder timing registry for observability
        self.encoder_timing_registry: dict[str, EncoderTimingStats] = {}
        self._encoder_timing_lock = threading.Lock()

606
        # Persistent buffers for CUDA graphs.
607
608
609
610
611
        self.input_ids = self._make_buffer(self.max_num_tokens, dtype=torch.int32)
        self.positions = self._make_buffer(self.max_num_tokens, dtype=torch.int64)
        self.query_start_loc = self._make_buffer(
            self.max_num_reqs + 1, dtype=torch.int32
        )
612
        self.seq_lens = self._make_buffer(self.max_num_reqs, dtype=torch.int32)
613
        self.encoder_seq_lens = self._make_buffer(self.max_num_reqs, dtype=torch.int32)
614
615
616
617
        if self.dcp_world_size > 1:
            self.dcp_local_seq_lens = self._make_buffer(
                self.max_num_reqs, dtype=torch.int32
            )
618
619
620
        # 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.
621
        self.inputs_embeds = self._make_buffer(
622
            self.max_num_tokens, self.inputs_embeds_size, dtype=self.dtype, numpy=False
623
624
        )
        self.is_token_ids = self._make_buffer(self.max_num_tokens, dtype=torch.bool)
625
626
        self.discard_request_mask = self._make_buffer(
            self.max_num_reqs, dtype=torch.bool
627
628
629
630
631
632
633
        )
        self.num_decode_draft_tokens = self._make_buffer(
            self.max_num_reqs, dtype=torch.int32
        )
        self.num_accepted_tokens = self._make_buffer(
            self.max_num_reqs, dtype=torch.int64
        )
634

635
636
        # Only relevant for multimodal models
        if self.supports_mm_inputs:
637
638
639
640
641
642
643
            # Double buffer to avoid race condition: previous iteration's async
            # copy may still be reading from CPU while current iteration writes.
            self.is_mm_embed_buffers = [
                self._make_buffer(self.max_num_tokens, dtype=torch.bool),
                self._make_buffer(self.max_num_tokens, dtype=torch.bool),
            ]
            self.is_mm_embed_idx = 0
644

645
        # Only relevant for models using M-RoPE (e.g, Qwen2-VL)
646
        if self.uses_mrope:
Roger Wang's avatar
Roger Wang committed
647
648
649
650
            # 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
651
652
653
654
655
656

            # 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
657
            self.mrope_positions = self._make_buffer(
658
659
                (3, self.max_num_tokens + 1), dtype=torch.int64
            )
660

661
662
663
664
665
666
667
        # 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
            )

668
        # None in the first PP rank. The rest are set after load_model.
669
        self.intermediate_tensors: IntermediateTensors | None = None
670

671
        # OPTIMIZATION: Cache the tensors rather than creating them every step.
672
        # Keep in int64 to avoid overflow with long context
673
674
675
676
        self.arange_np = np.arange(
            max(self.max_num_reqs + 1, self.max_model_len, self.max_num_tokens),
            dtype=np.int64,
        )
677

678
679
680
681
682
        # 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] = {}
683
684
685
686
687
        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(
688
689
                self.max_num_tokens, dtype=torch.int32, device=self.device
            )
690

691
        self.uniform_decode_query_len = 1 + self.num_spec_tokens
692
693
694
695

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

696
        self.mm_budget = (
697
            MultiModalBudget(self.vllm_config, self.mm_registry)
698
699
700
            if self.supports_mm_inputs
            else None
        )
701

702
        self.reorder_batch_threshold: int | None = None
703

704
705
706
707
708
        # 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()

709
        # Cached outputs.
710
        self._draft_token_ids: list[list[int]] | torch.Tensor | None = None
711
        self._draft_token_req_ids: list[str] | None = None
712
        self.transfer_event = torch.Event()
713
        self.sampled_token_ids_pinned_cpu = torch.empty(
714
            (self.max_num_reqs, 1),
715
716
            dtype=torch.int64,
            device="cpu",
717
718
            pin_memory=self.pin_memory,
        )
719

720
721
        # Pre-allocated tensor for copying valid sampled token counts to CPU,
        # with dedicated stream for overlapping and event for coordination.
722
        self.valid_sampled_token_count_event: torch.Event | None = None
723
        self.valid_sampled_token_count_copy_stream: torch.cuda.Stream | None = None
724
725
726
727
728
729
        # 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
730
        self.num_accepted_tokens_event: torch.Event | None = None
731
732
        if self.num_spec_tokens:
            self.draft_token_ids_event = torch.Event()
733
            self.num_accepted_tokens_event = torch.Event()
734
735
736
737
738
739
740
741
742
743
744
745
746
747
748
749
            self.draft_token_ids_copy_stream = torch.cuda.Stream()
            self.draft_token_ids_cpu = torch.empty(
                (self.max_num_reqs, self.num_spec_tokens),
                dtype=torch.int64,
                device="cpu",
                pin_memory=self.pin_memory,
            )
            if self.use_async_scheduling:
                self.valid_sampled_token_count_event = torch.Event()
                self.valid_sampled_token_count_copy_stream = torch.cuda.Stream()
                self.valid_sampled_token_count_cpu = torch.empty(
                    self.max_num_reqs,
                    dtype=torch.int64,
                    device="cpu",
                    pin_memory=self.pin_memory,
                )
750

751
752
753
754
        # Model weight offloader
        # Make sure this is called before any get_offloader call
        set_offloader(create_offloader(self.offload_config))

755
756
        # Ephemeral state transferred between execute_model() and sample_tokens().
        self.execute_model_state: ExecuteModelState | None = None
757
        self.kv_connector_output: KVConnectorOutput | None = None
758
        self.mamba_state_idx: dict[str, int] = {}
759
        self._mamba_copy_bufs: mamba_utils.MambaCopyBuffers | None = None
760
        self.layerwise_nvtx_hooks_registered = False
761

762
763
764
765
766
767
768
    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

769
    def reset_mm_cache(self) -> None:
770
771
772
773
        """
        Clear the multi-modal cache that was used during profiling,
        but no longer needed during inference.
        """
774
775
776
        if self.mm_budget:
            self.mm_budget.reset_cache()

777
778
779
780
781
782
783
784
    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()

785
786
787
788
789
790
791
792
793
794
795
796
797
798
799
800
801
802
803
804
805
806
807
808
809
810
811
812
813
814
815
816
817
818
819
820
821
822
823
824
825
826
827
828
    @torch.inference_mode()
    def init_fp8_kv_scales(self) -> None:
        """
        Re-initialize the KV cache and FP8 scales after waking from sleep.
        1. Zero out the KV cache tensors to remove garbage data from re-allocation.
        2. Reset Attention layer scaling factors (_k_scale, _v_scale) to 1.0.
          If these are left at 0.0 (default after wake_up), all KV cache values
          become effectively zero, causing gibberish output.
        """
        if not self.cache_config.cache_dtype.startswith("fp8"):
            return

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

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

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

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

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

829
830
831
832
    def _get_positions(self, num_tokens: Any):
        if isinstance(num_tokens, int):
            if self.uses_mrope:
                return self.mrope_positions.gpu[:, :num_tokens]
833
834
            if self.uses_xdrope_dim > 0:
                return self.xdrope_positions.gpu[:, :num_tokens]
835
836
837
838
            return self.positions.gpu[:num_tokens]
        else:
            if self.uses_mrope:
                return self.mrope_positions.gpu[:, num_tokens]
839
840
            if self.uses_xdrope_dim > 0:
                return self.xdrope_positions.gpu[:, num_tokens]
841
842
            return self.positions.gpu[num_tokens]

843
    def _make_buffer(
844
        self, *size: int | torch.SymInt, dtype: torch.dtype, numpy: bool = True
845
846
847
848
849
850
851
852
    ) -> CpuGpuBuffer:
        return CpuGpuBuffer(
            *size,
            dtype=dtype,
            device=self.device,
            pin_memory=self.pin_memory,
            with_numpy=numpy,
        )
853

854
855
856
857
858
859
860
861
862
863
    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

864
    def _init_model_kwargs(self):
865
866
        model_kwargs = dict[str, Any]()

867
        if not self.is_pooling_model:
868
869
            return model_kwargs

870
871
        num_reqs = self.input_batch.num_reqs
        pooling_params = self.input_batch.get_pooling_params()
872
873
874

        token_type_id_requests = dict[int, Any]()
        for i, param in enumerate(pooling_params):
875
876
877
878
879
            if (
                param.extra_kwargs is not None
                and (token_types := param.extra_kwargs.get("compressed_token_type_ids"))
                is not None
            ):
880
881
882
883
884
                token_type_id_requests[i] = token_types

        if len(token_type_id_requests) == 0:
            return model_kwargs

885
        seq_lens = self.seq_lens.gpu[:num_reqs]
886
887
888
889
890
891
892
893
        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(
894
895
            device=self.device
        )
896
897
        return model_kwargs

898
    def _may_reorder_batch(self, scheduler_output: "SchedulerOutput") -> None:
899
900
        """
        Update the order of requests in the batch based on the attention
901
        backend's needs. For example, some attention backends (namely MLA) may
902
903
904
905
906
907
        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
908
        # Attention free models have zero kv_cache_groups, however models
909
910
911
912
913
914
915
        # 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

916
917
918
919
        if self.reorder_batch_threshold is not None:
            reorder_batch_to_split_decodes_and_prefills(
                self.input_batch,
                scheduler_output,
920
921
                decode_threshold=self.reorder_batch_threshold,
            )
922

923
924
    # Note: used for model runner override.
    def _init_device_properties(self) -> None:
925
        """Initialize attributes from torch.cuda.get_device_properties"""
926
927

        self.num_sms = num_compute_units(self.device.index)
928
929
930

    # Note: used for model runner override.
    def _sync_device(self) -> None:
931
        torch.accelerator.synchronize()
932

933
    def _update_states(self, scheduler_output: "SchedulerOutput") -> None:
934
935
936
937
938
939
        """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.

940
941
        The SamplingMetadata is updated and copied to the GPU if there is a
        new/resumed/paused/finished request in the batch.
942
943
        """
        # Remove finished requests from the cached states.
944
945
        for req_id in scheduler_output.finished_req_ids:
            self.requests.pop(req_id, None)
946
            self.num_prompt_logprobs.pop(req_id, None)
947
948
949
950
951
952
953
        # 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:
954
            self.input_batch.remove_request(req_id)
955
956

        # Free the cached encoder outputs.
957
958
        for mm_hash in scheduler_output.free_encoder_mm_hashes:
            self.encoder_cache.pop(mm_hash, None)
959

960
961
962
963
964
965
966
        # 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()
967
968
969
970
971
972
973
974
        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)
975
976
977
978
979
        # 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:
980
            self.input_batch.remove_request(req_id)
981

982
        reqs_to_add: list[CachedRequestState] = []
983
        # Add new requests to the cached states.
984
985
        for new_req_data in scheduler_output.scheduled_new_reqs:
            req_id = new_req_data.req_id
986
987
988
989
990
991
            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

992
            sampling_params = new_req_data.sampling_params
993
            pooling_params = new_req_data.pooling_params
994

995
996
997
998
            if (
                sampling_params
                and sampling_params.sampling_type == SamplingType.RANDOM_SEED
            ):
999
1000
1001
1002
1003
                generator = torch.Generator(device=self.device)
                generator.manual_seed(sampling_params.seed)
            else:
                generator = None

1004
1005
            if self.is_pooling_model:
                assert pooling_params is not None
1006
1007
                task = pooling_params.task
                assert task is not None, "You did not set `task` in the API"
1008

1009
                model = cast(VllmModelForPooling, self.get_model())
1010
                to_update = model.pooler.get_pooling_updates(task)
1011
1012
                to_update.apply(pooling_params)

1013
            req_state = CachedRequestState(
1014
                req_id=req_id,
1015
                prompt_token_ids=new_req_data.prompt_token_ids,
1016
                prompt_embeds=new_req_data.prompt_embeds,
1017
                mm_features=new_req_data.mm_features,
1018
                sampling_params=sampling_params,
1019
                pooling_params=pooling_params,
1020
                generator=generator,
1021
1022
                block_ids=new_req_data.block_ids,
                num_computed_tokens=new_req_data.num_computed_tokens,
1023
                output_token_ids=[],
1024
                lora_request=new_req_data.lora_request,
1025
            )
1026
1027
            self.requests[req_id] = req_state

1028
1029
1030
1031
1032
1033
1034
            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
                )

1035
            # Only relevant for models using M-RoPE (e.g, Qwen2-VL)
1036
            if self.uses_mrope:
1037
                self._init_mrope_positions(req_state)
1038

1039
1040
1041
1042
            # Only relevant for models using XD-RoPE (e.g, HunYuan-VL)
            if self.uses_xdrope_dim > 0:
                self._init_xdrope_positions(req_state)

1043
            reqs_to_add.append(req_state)
1044

1045
        # Update the states of the running/resumed requests.
1046
        is_last_rank = get_pp_group().is_last_rank
1047
        req_data = scheduler_output.scheduled_cached_reqs
1048
        scheduled_spec_tokens = scheduler_output.scheduled_spec_decode_tokens
1049
1050
1051
1052
1053

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

1054
        for i, req_id in enumerate(req_data.req_ids):
1055
            req_state = self.requests[req_id]
1056
1057
            num_computed_tokens = req_data.num_computed_tokens[i]
            new_block_ids = req_data.new_block_ids[i]
1058
            resumed_from_preemption = req_id in req_data.resumed_req_ids
1059
            num_output_tokens = req_data.num_output_tokens[i]
1060
            req_index = self.input_batch.req_id_to_index.get(req_id)
1061

1062
1063
1064
1065
1066
1067
            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:
                # fist step: num_computed_tokens = 0, spec_tokens = [],
                # prev_num_draft_len = 0.
Jiayi Yan's avatar
Jiayi Yan committed
1068
                # second step: num_computed_tokens = 100(prompt length),
1069
1070
1071
1072
1073
1074
1075
                # 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.
                # num_computed_tokens in first step and second step does't contain
                # 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.
1076
1077
1078
1079
1080
1081
1082
1083
1084
                if req_index is None:
                    req_state.prev_num_draft_len = 0
                else:
                    assert self.input_batch.prev_req_id_to_index is not None
                    prev_req_index = self.input_batch.prev_req_id_to_index[req_id]
                    num_accepted = valid_sampled_token_count[prev_req_index] - 1
                    num_rejected = req_state.prev_num_draft_len - num_accepted
                    num_computed_tokens -= num_rejected
                    req_state.output_token_ids.extend([-1] * num_accepted)
1085

1086
            # Update the cached states.
1087
            req_state.num_computed_tokens = num_computed_tokens
1088
1089

            if not is_last_rank:
1090
1091
1092
1093
1094
1095
1096
1097
1098
1099
1100
1101
1102
1103
1104
1105
1106
1107
1108
1109
                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:]
                        )
1110
1111
1112
1113
1114
            elif num_output_tokens < len(req_state.output_token_ids):
                # Some output tokens were discarded due to a sync-KV-load
                # failure. Align the cached state.
                del req_state.output_token_ids[num_output_tokens:]
                if req_index is not None:
1115
1116
1117
1118
                    end_idx = (
                        self.input_batch.num_prompt_tokens[req_index]
                        + num_output_tokens
                    )
1119
                    self.input_batch.num_tokens_no_spec[req_index] = end_idx
1120

1121
            # Update the block IDs.
1122
            if not resumed_from_preemption:
1123
1124
                if new_block_ids is not None:
                    # Append the new blocks to the existing block IDs.
1125
                    for block_ids, new_ids in zip(req_state.block_ids, new_block_ids):
1126
                        block_ids.extend(new_ids)
1127
            else:
1128
                assert req_index is None
1129
                assert new_block_ids is not None
1130
1131
                # The request is resumed from preemption.
                # Replace the existing block IDs with the new ones.
1132
                req_state.block_ids = new_block_ids
1133
1134
1135
1136
1137

            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.
1138
1139
1140
1141
1142
1143
1144

                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:]

1145
                reqs_to_add.append(req_state)
1146
1147
1148
                continue

            # Update the persistent batch.
1149
            self.input_batch.num_computed_tokens_cpu[req_index] = num_computed_tokens
1150
            if new_block_ids is not None:
1151
                self.input_batch.block_table.append_row(new_block_ids, req_index)
1152
1153
1154
1155
1156
1157
1158

            # 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)
1159
                self.input_batch.token_ids_cpu[
1160
1161
1162
                    req_index, start_token_index:end_token_index
                ] = new_token_ids
                self.input_batch.num_tokens_no_spec[req_index] = end_token_index
1163

1164
            # Add spec_token_ids to token_ids_cpu.
1165
            self.input_batch.update_req_spec_token_ids(req_state, scheduled_spec_tokens)
1166

1167
1168
        # Add the new or resumed requests to the persistent batch.
        # The smaller empty indices are filled first.
1169
1170
        for request in reqs_to_add:
            self.input_batch.add_request(request)
1171
            self.input_batch.update_req_spec_token_ids(request, scheduled_spec_tokens)
1172

1173
1174
1175
1176
1177
1178
        # 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()
1179

1180
    def _update_states_after_model_execute(
1181
        self, output_token_ids: torch.Tensor, scheduler_output: "SchedulerOutput"
1182
    ) -> None:
1183
1184
1185
1186
1187
1188
1189
1190
        """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.
        """
1191
        if not self.speculative_config or not self.model_config.is_hybrid:
1192
1193
1194
            return

        # Find the number of accepted tokens for each sequence.
1195
1196
        num_reqs = output_token_ids.size(0)
        self.num_accepted_tokens.gpu[:num_reqs] = (
1197
1198
1199
1200
1201
            (
                torch.cat(
                    [
                        output_token_ids,
                        torch.full(
1202
                            (num_reqs, 1),
1203
1204
1205
1206
1207
1208
1209
1210
1211
1212
1213
                            -1,
                            device=output_token_ids.device,
                        ),
                    ],
                    dim=1,
                )
                == -1
            )
            .int()
            .argmax(-1)
        )
1214
        if self.cache_config.mamba_cache_mode == "align":
1215
1216
1217
1218
1219
            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

1220
1221
1222
1223
1224
1225
1226
1227
            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(),
1228
                self._get_mamba_copy_bufs(),
1229
            )
1230
1231
1232
1233
        else:
            self.input_batch.num_accepted_tokens_cpu_tensor[:num_reqs].copy_(
                self.num_accepted_tokens.gpu[:num_reqs], non_blocking=True
            )
1234
1235
            assert self.num_accepted_tokens_event is not None
            self.num_accepted_tokens_event.record()
1236

1237
1238
1239
1240
1241
1242
1243
1244
1245
1246
1247
1248
1249
1250
1251
1252
1253
1254
1255
1256
1257
1258
1259
1260
1261
1262
1263
1264
1265
1266
1267
1268
1269
1270
    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
        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

1271
    def _init_mrope_positions(self, req_state: CachedRequestState):
1272
1273
        model = self.get_model()
        assert supports_mrope(model), "M-RoPE support is not implemented."
1274
1275
1276
1277
        assert req_state.prompt_token_ids is not None, (
            "M-RoPE requires prompt_token_ids to be available."
        )
        mrope_model = cast(SupportsMRoPE, model)
1278
1279

        req_state.mrope_positions, req_state.mrope_position_delta = (
1280
            mrope_model.get_mrope_input_positions(
1281
                req_state.prompt_token_ids,
1282
                req_state.mm_features,
1283
            )
1284
        )
1285

1286
1287
1288
1289
1290
1291
1292
1293
1294
1295
1296
1297
1298
    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,
        )

1299
    def _extract_mm_kwargs(
1300
        self,
1301
1302
        scheduler_output: "SchedulerOutput",
    ) -> BatchedTensorInputs:
1303
        if not scheduler_output or not self.is_multimodal_raw_input_only_model:
1304
            return {}
1305

1306
        mm_kwargs = list[tuple[str, MultiModalKwargsItem]]()
1307
        for req in scheduler_output.scheduled_new_reqs:
1308
1309
            for feature in req.mm_features:
                if feature.data is not None:
1310
                    mm_kwargs.append((feature.modality, feature.data))
1311

1312
1313
1314
        # Input all modalities at once
        mm_kwargs_combined: BatchedTensorInputs = {}
        for _, _, mm_kwargs_group in group_mm_kwargs_by_modality(
1315
1316
1317
            mm_kwargs,
            device=self.device,
            pin_memory=self.pin_memory,
1318
1319
        ):
            mm_kwargs_combined.update(mm_kwargs_group)
1320

1321
        return mm_kwargs_combined
1322

1323
    def _dummy_mm_kwargs(self, num_seqs: int) -> BatchedTensorInputs:
1324
        if not self.is_multimodal_raw_input_only_model:
1325
            return {}
1326

1327
1328
1329
        mm_budget = self.mm_budget
        assert mm_budget is not None

1330
1331
1332
        if not mm_budget.mm_max_toks_per_item:
            return {}  # No tower modalities (embed-only mode)

1333
1334
        dummy_modality = mm_budget.get_modality_with_max_tokens()
        return self._get_mm_dummy_batch(dummy_modality, num_seqs)
1335

1336
1337
1338
    def _get_cumsum_and_arange(
        self,
        num_tokens: np.ndarray,
1339
        cumsum_dtype: np.dtype | None = None,
1340
1341
1342
1343
1344
1345
1346
1347
1348
1349
1350
1351
1352
1353
1354
1355
    ) -> tuple[np.ndarray, np.ndarray]:
        """Get the cumulative sum and batched arange of the given array.
        # E.g., [2, 5, 3] -> ([2, 7, 10], [0, 1, 0, 1, 2, 3, 4, 0, 1, 2])
        # Equivalent to but faster than:
        # np.concatenate([np.arange(n) for n in num_tokens])
        """
        # Step 1. [2, 5, 3] -> [2, 7, 10]
        cu_num_tokens = np.cumsum(num_tokens, dtype=cumsum_dtype)
        total_num_tokens = cu_num_tokens[-1]
        # Step 2. [2, 7, 10] -> [0, 0, 2, 2, 2, 2, 2, 7, 7, 7]
        cumsums_offsets = np.repeat(cu_num_tokens - num_tokens, num_tokens)
        # Step 3. [0, 1, 0, 1, 2, 3, 4, 0, 1, 2]
        arange = self.arange_np[:total_num_tokens] - cumsums_offsets

        return cu_num_tokens, arange

1356
    def _prepare_input_ids(
1357
1358
1359
1360
        self,
        scheduler_output: "SchedulerOutput",
        total_num_scheduled_tokens: int,
        cu_num_tokens: np.ndarray,
1361
    ) -> None:
1362
        """Prepare the input IDs for the current batch.
1363

1364
1365
1366
1367
1368
1369
1370
        Carefully handles the `prev_sampled_token_ids` which can be cached
        from the previous engine iteration, in which case those tokens on the
        GPU need to be copied into the corresponding slots into input_ids."""

        if self.input_batch.prev_sampled_token_ids is None:
            # Normal scheduling case
            self.input_ids.copy_to_gpu(total_num_scheduled_tokens)
1371
1372
1373
            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)
1374
1375
1376
1377
1378
1379
1380
            return

        # Async scheduling case, where some decode requests from the previous
        # iteration won't have entries in input_ids_cpu and need to be copied
        # on the GPU from prev_sampled_token_ids.
        prev_req_id_to_index = self.input_batch.prev_req_id_to_index
        assert prev_req_id_to_index is not None
1381
1382
1383
1384
        sample_flattened_indices: list[int] = []
        spec_flattened_indices: list[int] = []
        prev_common_req_indices: list[int] = []
        prev_draft_token_indices: list[int] = []
1385
1386
        indices_match = True
        max_flattened_index = -1
1387
1388
1389
        total_num_spec_tokens = 0
        scheduled_spec_tokens = scheduler_output.scheduled_spec_decode_tokens

1390
1391
1392
1393
1394
        for req_id, cur_index in self.input_batch.req_id_to_index.items():
            if (prev_index := prev_req_id_to_index.get(req_id)) is not None:
                prev_common_req_indices.append(prev_index)
                # We need to compute the flattened input_ids index of the
                # last token in each common request.
1395
1396
                draft_len = len(scheduled_spec_tokens.get(req_id, ()))
                total_num_spec_tokens += draft_len
1397
                flattened_index = cu_num_tokens[cur_index].item() - 1
1398
1399
1400
1401
1402
1403
1404
1405
1406
1407
1408
1409
1410
1411
1412
                # 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))
1413
                indices_match &= prev_index == flattened_index
1414
                max_flattened_index = max(max_flattened_index, flattened_index)
Jiayi Yan's avatar
Jiayi Yan committed
1415
        num_common_tokens = len(sample_flattened_indices)
1416
        total_without_spec = total_num_scheduled_tokens - total_num_spec_tokens
Jiayi Yan's avatar
Jiayi Yan committed
1417
        if num_common_tokens < total_without_spec:
1418
1419
1420
            # If not all requests are decodes from the last iteration,
            # We need to copy the input_ids_cpu to the GPU first.
            self.input_ids.copy_to_gpu(total_num_scheduled_tokens)
1421
1422
1423
            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
1424
        if num_common_tokens == 0:
1425
            # No requests in common with the previous iteration
1426
            # So input_ids.cpu will have all the input ids.
1427
            return
Jiayi Yan's avatar
Jiayi Yan committed
1428
        if indices_match and max_flattened_index == (num_common_tokens - 1):
1429
1430
1431
1432
            # 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
1433
1434
            self.input_ids.gpu[:num_common_tokens].copy_(
                self.input_batch.prev_sampled_token_ids[:num_common_tokens, 0],
1435
1436
                non_blocking=True,
            )
1437
            if self.enable_prompt_embeds:
Jiayi Yan's avatar
Jiayi Yan committed
1438
                self.is_token_ids.gpu[:num_common_tokens] = True
1439
            return
1440
        # Upload the index tensors asynchronously so the scatter can be non-blocking.
1441
1442
        sampled_tokens_index_tensor = torch.tensor(
            sample_flattened_indices, dtype=torch.int64, pin_memory=self.pin_memory
1443
        ).to(self.device, non_blocking=True)
1444
        prev_common_req_indices_tensor = torch.tensor(
1445
1446
            prev_common_req_indices, dtype=torch.int64, pin_memory=self.pin_memory
        ).to(self.device, non_blocking=True)
1447
1448
        self.input_ids.gpu.scatter_(
            dim=0,
1449
            index=sampled_tokens_index_tensor,
1450
            src=self.input_batch.prev_sampled_token_ids[
1451
1452
1453
                prev_common_req_indices_tensor, 0
            ],
        )
1454

1455
1456
1457
1458
1459
1460
1461
1462
1463
1464
1465
1466
1467
1468
1469
1470
1471
1472
1473
1474
1475
1476
        # 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],
        )

1477
1478
    def _get_encoder_seq_lens(
        self,
1479
        num_scheduled_tokens: dict[str, int],
1480
1481
        kv_cache_spec: KVCacheSpec,
        num_reqs: int,
1482
        for_cudagraph_capture: bool = False,
1483
    ) -> tuple[torch.Tensor | None, np.ndarray | None]:
1484
        if not isinstance(kv_cache_spec, CrossAttentionSpec):
1485
            return None, None
1486

1487
1488
        # Zero out buffer for padding requests that are not actually scheduled (CGs)
        self.encoder_seq_lens.np[:num_reqs] = 0
1489

1490
1491
        # Build encoder_seq_lens array mapping request indices to
        # encoder lengths for inputs scheduled in this batch
1492
        for req_id in num_scheduled_tokens:
1493
            req_index = self.input_batch.req_id_to_index[req_id]
1494
1495
1496
1497
1498
1499
1500
1501
1502
1503
1504
1505
            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
1506
1507
1508
1509
1510
1511
1512
1513
1514
        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
1515
1516
1517
1518

        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]
1519

1520
        return encoder_seq_lens, encoder_seq_lens_cpu
1521

1522
    def _prepare_inputs(
1523
1524
1525
        self,
        scheduler_output: "SchedulerOutput",
        num_scheduled_tokens: np.ndarray,
1526
1527
    ) -> tuple[
        torch.Tensor,
1528
        SpecDecodeMetadata | None,
1529
    ]:
1530
1531
        """
        :return: tuple[
1532
            logits_indices, spec_decode_metadata,
1533
1534
        ]
        """
1535
1536
1537
1538
1539
1540
1541
        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.
1542
        self.input_batch.block_table.commit_block_table(num_reqs)
1543
1544
1545

        # Get request indices.
        # E.g., [2, 5, 3] -> [0, 0, 1, 1, 1, 1, 1, 2, 2, 2]
1546
        req_indices = np.repeat(self.arange_np[:num_reqs], num_scheduled_tokens)
1547

1548
1549
        # cu_num_tokens: [2, 5, 3] -> [2, 7, 10]
        # arange: [0, 1, 0, 1, 2, 3, 4, 0, 1, 2]
1550
        cu_num_tokens, arange = self._get_cumsum_and_arange(num_scheduled_tokens)
1551
1552

        # Get positions.
1553
        positions_np = self.positions.np[:total_num_scheduled_tokens]
1554
1555
1556
1557
1558
        np.add(
            self.input_batch.num_computed_tokens_cpu[req_indices],
            arange,
            out=positions_np,
        )
1559

1560
1561
        # Calculate M-RoPE positions.
        # Only relevant for models using M-RoPE (e.g, Qwen2-VL)
1562
        if self.uses_mrope:
1563
1564
            self._calc_mrope_positions(scheduler_output)

1565
1566
1567
1568
1569
        # 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)

1570
1571
1572
1573
        # 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.
1574
1575
1576
        token_indices = (
            positions_np + req_indices * self.input_batch.token_ids_cpu.shape[1]
        )
1577
        token_indices_tensor = torch.from_numpy(token_indices)
1578

1579
1580
1581
        # 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.
1582
1583
1584
1585
1586
1587
        torch.index_select(
            self.input_batch.token_ids_cpu_tensor.flatten(),
            0,
            token_indices_tensor,
            out=self.input_ids.cpu[:total_num_scheduled_tokens],
        )
1588
        if self.enable_prompt_embeds:
1589
            is_token_ids = self.input_batch.is_token_ids_tensor.flatten()
1590
1591
1592
1593
            torch.index_select(
                is_token_ids,
                0,
                token_indices_tensor,
1594
1595
                out=self.is_token_ids.cpu[:total_num_scheduled_tokens],
            )
1596
1597
1598
1599
1600
1601
1602
1603
1604
1605
1606
1607
1608
1609
1610
1611
1612
1613
1614
1615
1616
1617
1618
1619
1620
1621
1622
1623
1624
1625
1626
1627
1628

        # 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:
1629
1630
1631
                    self.inputs_embeds.cpu[
                        output_idx : output_idx + actual_num_sched
                    ].copy_(req_embeds[start_pos:actual_end])
1632
1633

                output_idx += num_sched
1634

1635
1636
        self.input_batch.block_table.compute_slot_mapping(req_indices, positions_np)
        self.input_batch.block_table.commit_slot_mapping(total_num_scheduled_tokens)
1637
1638

        # Prepare the attention metadata.
1639
        self.query_start_loc.np[0] = 0
1640
        self.query_start_loc.np[1 : num_reqs + 1] = cu_num_tokens
1641
1642
        # Note: pad query_start_loc to be non-decreasing, as kernels
        # like FlashAttention requires that
1643
        self.query_start_loc.np[num_reqs + 1 :].fill(cu_num_tokens[-1])
1644
        self.query_start_loc.copy_to_gpu()
1645
        query_start_loc = self.query_start_loc.gpu[: num_reqs + 1]
1646

1647
        self.seq_lens.np[:num_reqs] = (
1648
1649
            self.input_batch.num_computed_tokens_cpu[:num_reqs] + num_scheduled_tokens
        )
1650
        # Fill unused with 0 for full cuda graph mode.
1651
1652
        self.seq_lens.np[num_reqs:].fill(0)
        self.seq_lens.copy_to_gpu()
1653

1654
        num_tokens = [self.requests[r].num_tokens for r in self.input_batch.req_ids]
1655
1656
        num_tokens_np = np.array(num_tokens, dtype=np.int32)

1657
        # Record which requests should not be sampled,
1658
        # so that we could clear the sampled tokens before returning
1659
1660
        self.discard_request_mask.np[:num_reqs] = (
            self.seq_lens.np[:num_reqs] < num_tokens_np
1661
        )
1662
        self.discard_request_mask.copy_to_gpu(num_reqs)
1663

1664
        # Copy the tensors to the GPU.
1665
1666
1667
1668
1669
        self._prepare_input_ids(
            scheduler_output,
            total_num_scheduled_tokens,
            cu_num_tokens,
        )
1670

1671
        if self.uses_mrope:
1672
            # Only relevant for models using M-RoPE (e.g, Qwen2-VL)
1673
1674
            self.mrope_positions.gpu[:, :total_num_scheduled_tokens].copy_(
                self.mrope_positions.cpu[:, :total_num_scheduled_tokens],
1675
1676
                non_blocking=True,
            )
1677
1678
1679
1680
1681
1682
        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,
            )
1683
1684
        else:
            # Common case (1D positions)
1685
            self.positions.copy_to_gpu(total_num_scheduled_tokens)
1686

1687
        use_spec_decode = len(scheduler_output.scheduled_spec_decode_tokens) > 0
1688
1689
1690
1691
1692
1693
1694
1695
        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
1696
            num_sampled_tokens = np.ones(num_reqs, dtype=np.int32)
1697
1698
1699
1700
1701
        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)
1702
1703
1704
            # 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)
1705
1706
1707
1708
            for (
                req_id,
                draft_token_ids,
            ) in scheduler_output.scheduled_spec_decode_tokens.items():
1709
1710
                req_idx = self.input_batch.req_id_to_index[req_id]
                num_draft_tokens[req_idx] = len(draft_token_ids)
1711
1712
1713
1714
1715
                if (
                    self.input_batch.num_computed_tokens_cpu[req_idx]
                    >= self.input_batch.num_prompt_tokens[req_idx]
                ):
                    num_decode_draft_tokens[req_idx] = len(draft_token_ids)
1716
            spec_decode_metadata = self._calc_spec_decode_metadata(
1717
1718
                num_draft_tokens, cu_num_tokens
            )
1719
            logits_indices = spec_decode_metadata.logits_indices
1720
            num_sampled_tokens = num_draft_tokens + 1
1721
            # For DECODE only cuda graph of some attention backends (e.g., GDN).
1722
            self.num_decode_draft_tokens.np[:num_reqs] = num_decode_draft_tokens
1723
1724
            self.num_decode_draft_tokens.np[num_reqs:].fill(-1)
            self.num_decode_draft_tokens.copy_to_gpu()
1725

1726
1727
1728
1729
1730
        # Hot-Swap lora model
        if self.lora_config:
            assert (
                np.sum(num_sampled_tokens)
                <= self.vllm_config.scheduler_config.max_num_batched_tokens
1731
            )
1732
1733
1734
1735
1736
1737
1738
1739
1740
1741
1742
            self.set_active_loras(
                self.input_batch, num_scheduled_tokens, num_sampled_tokens
            )

        return (
            logits_indices,
            spec_decode_metadata,
        )

    def _build_attention_metadata(
        self,
1743
        num_tokens: int,
1744
        num_reqs: int,
1745
1746
1747
        max_query_len: int,
        num_tokens_padded: int | None = None,
        num_reqs_padded: int | None = None,
1748
1749
1750
1751
        ubatch_slices: UBatchSlices | None = None,
        logits_indices: torch.Tensor | None = None,
        use_spec_decode: bool = False,
        for_cudagraph_capture: bool = False,
1752
        num_scheduled_tokens: dict[str, int] | None = None,
1753
        cascade_attn_prefix_lens: list[list[int]] | None = None,
1754
        slot_mappings: dict[int, torch.Tensor] | None = None,
1755
1756
1757
1758
    ) -> tuple[PerLayerAttnMetadata, CommonAttentionMetadata | None]:
        """
        :return: tuple[attn_metadata, spec_decode_common_attn_metadata]
        """
1759
1760
1761
1762
        # Attention metadata is not needed for attention free models
        if len(self.kv_cache_config.kv_cache_groups) == 0:
            return {}, None

1763
1764
        num_tokens_padded = num_tokens_padded or num_tokens
        num_reqs_padded = num_reqs_padded or num_reqs
1765
        assert num_reqs_padded is not None and num_tokens_padded is not None
1766

1767
1768
1769
        attn_metadata: PerLayerAttnMetadata = {}
        if ubatch_slices is not None:
            attn_metadata = [dict() for _ in range(len(ubatch_slices))]
1770

1771
1772
1773
1774
1775
1776
1777
1778
        if for_cudagraph_capture:
            # For some attention backends (e.g. FA) with sliding window models we need
            # to make sure the backend see a max_seq_len that is larger to the sliding
            # window size when capturing to make sure the correct kernel is selected.
            max_seq_len = self.max_model_len
        else:
            max_seq_len = self.seq_lens.np[:num_reqs].max().item()

1779
        if use_spec_decode:
1780
1781
            if self.num_accepted_tokens_event is not None:
                self.num_accepted_tokens_event.synchronize()
1782
            self.num_accepted_tokens.np[:num_reqs] = (
1783
1784
                self.input_batch.num_accepted_tokens_cpu[:num_reqs]
            )
1785
1786
            self.num_accepted_tokens.np[num_reqs:].fill(1)
            self.num_accepted_tokens.copy_to_gpu()
1787

1788
        kv_cache_groups = self.kv_cache_config.kv_cache_groups
1789

1790
        def _get_block_table(kv_cache_gid: int):
1791
1792
1793
            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):
1794
                blk_table_tensor = torch.zeros(
1795
                    (num_reqs_padded, 1),
1796
                    dtype=torch.int32,
1797
1798
                    device=self.device,
                )
1799
            else:
1800
                blk_table = self.input_batch.block_table[kv_cache_gid]
1801
                blk_table_tensor = blk_table.get_device_tensor(num_reqs_padded)
1802

1803
1804
1805
            # Fill unused with -1. Needed for reshape_and_cache in full cuda
            # graph mode. `blk_table_tensor` -1 to match mamba PAD_SLOT_ID
            blk_table_tensor[num_reqs:num_reqs_padded].fill_(-1)
1806
            return blk_table_tensor
1807

1808
1809
1810
        assert slot_mappings is not None
        block_table_gid_0 = _get_block_table(0)
        slot_mapping_gid_0 = slot_mappings[0]
1811

1812
1813
        if self.model_config.enable_return_routed_experts:
            self.slot_mapping = slot_mapping_gid_0[:num_tokens].cpu().numpy()
1814
1815
1816
1817
1818
1819
1820
1821
1822
1823
1824
1825
1826
1827
1828
1829
1830
1831
1832
1833
1834
1835
1836
1837
1838
1839
1840
1841
1842
1843
1844
1845
1846
1847
1848
1849
1850
1851
        cm_base = CommonAttentionMetadata(
            query_start_loc=self.query_start_loc.gpu[: num_reqs_padded + 1],
            query_start_loc_cpu=self.query_start_loc.cpu[: num_reqs_padded + 1],
            seq_lens=self.seq_lens.gpu[:num_reqs_padded],
            _seq_lens_cpu=self.seq_lens.cpu[:num_reqs_padded],
            _num_computed_tokens_cpu=self.input_batch.num_computed_tokens_cpu_tensor[
                :num_reqs_padded
            ],
            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,
        )

        if self.dcp_world_size > 1:
            self.dcp_local_seq_lens.cpu[:num_reqs] = get_dcp_local_seq_lens(
                self.seq_lens.cpu[:num_reqs],
                self.dcp_world_size,
                self.dcp_rank,
                self.parallel_config.cp_kv_cache_interleave_size,
            )
            self.dcp_local_seq_lens.cpu[num_reqs:].fill_(0)
            self.dcp_local_seq_lens.copy_to_gpu(num_reqs_padded)

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

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

1852
1853
1854
1855
1856
1857
1858
1859
1860
        # 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
        ] = {}

1861
1862
1863
1864
1865
1866
1867
        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]
1868
            builder = attn_group.get_metadata_builder(ubid or 0)
1869
1870
1871
1872
            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))
1873

1874
1875
1876
1877
1878
1879
1880
            cascade_attn_prefix_len = (
                cascade_attn_prefix_lens[kv_cache_gid][attn_gid]
                if cascade_attn_prefix_lens
                else 0
            )

            extra_attn_metadata_args = {}
1881
1882
1883
            if use_spec_decode and isinstance(
                builder, (Mamba2AttentionMetadataBuilder, GDNAttentionMetadataBuilder)
            ):
1884
1885
1886
1887
1888
1889
1890
1891
1892
1893
1894
1895
                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
                )
1896
1897
1898
1899
1900
1901
1902
1903
1904
            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,
                )
1905
1906
1907
1908
1909
1910
            else:
                attn_metadata_i = builder.build(
                    common_prefix_len=cascade_attn_prefix_len,
                    common_attn_metadata=common_attn_metadata,
                    **extra_attn_metadata_args,
                )
1911
1912
                if builder.supports_update_block_table:
                    cached_attn_metadata[cache_key] = attn_metadata_i
1913
1914
1915
1916
1917
1918
1919
1920
1921
1922
1923
1924
1925
1926
1927
1928
1929
1930
1931
1932
1933
1934
1935

            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,
1936
                for_cudagraph_capture=for_cudagraph_capture,
1937
            )
1938
            if kv_cache_gid > 0:
1939
1940
                cm.block_table_tensor = _get_block_table(kv_cache_gid)
                cm.slot_mapping = slot_mappings[kv_cache_gid]
1941

1942
            if self.speculative_config and spec_decode_common_attn_metadata is None:
1943
                if isinstance(self.drafter, EagleProposer):
1944
                    if self.drafter.kv_cache_gid == kv_cache_gid:
1945
                        spec_decode_common_attn_metadata = cm
1946
                else:
1947
                    spec_decode_common_attn_metadata = cm
1948

1949
            for attn_gid in range(len(self.attn_groups[kv_cache_gid])):
1950
                if ubatch_slices is not None:
1951
1952
1953
                    for ubid, _cm in enumerate(split_attn_metadata(ubatch_slices, cm)):
                        _build_attn_group_metadata(kv_cache_gid, attn_gid, _cm, ubid)

1954
                else:
1955
                    _build_attn_group_metadata(kv_cache_gid, attn_gid, cm)
1956

1957
1958
1959
1960
1961
1962
1963
1964
1965
1966
1967
1968
1969
1970
1971
1972
1973
1974
1975
1976
        if self.is_mm_prefix_lm:
            req_doc_ranges = {}
            for req_id in self.input_batch.req_ids:
                image_doc_ranges = []
                req_state = self.requests[req_id]
                for mm_feature in req_state.mm_features:
                    pos_info = mm_feature.mm_position
                    img_doc_range = pos_info.extract_embeds_range()
                    image_doc_ranges.extend(img_doc_range)
                req_idx = self.input_batch.req_id_to_index[req_id]
                req_doc_ranges[req_idx] = image_doc_ranges

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

1977
1978
1979
1980
1981
1982
1983
1984
1985
1986
        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)
            )

1987
        return attn_metadata, spec_decode_common_attn_metadata
1988

1989
1990
1991
    def _compute_cascade_attn_prefix_lens(
        self,
        num_scheduled_tokens: np.ndarray,
1992
        num_computed_tokens: np.ndarray,
1993
1994
1995
1996
1997
1998
1999
        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
        """
2000

2001
2002
2003
2004
2005
2006
2007
2008
2009
2010
2011
2012
2013
2014
        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,
2015
                        num_computed_tokens,
2016
2017
2018
2019
2020
2021
2022
2023
                        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
2024

2025
2026
2027
    def _compute_cascade_attn_prefix_len(
        self,
        num_scheduled_tokens: np.ndarray,
2028
        num_computed_tokens: np.ndarray,
2029
        num_common_prefix_blocks: int,
2030
2031
        kv_cache_spec: KVCacheSpec,
        attn_metadata_builder: AttentionMetadataBuilder,
2032
2033
2034
2035
2036
2037
2038
2039
2040
2041
2042
2043
2044
2045
2046
2047
2048
2049
    ) -> 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.
        """
2050

2051
        common_prefix_len = num_common_prefix_blocks * kv_cache_spec.block_size
2052
2053
2054
2055
2056
2057
2058
2059
2060
2061
2062
2063
2064
2065
2066
2067
2068
2069
2070
2071
2072
2073
2074
2075
2076
2077
2078
2079
2080
2081
2082
2083
2084
2085
2086
2087
2088
        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]
2089
        # Request 3's num_computed_tokens: 3 (i.e., [A, B, C])
2090
2091
2092
2093
2094
        # 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.
2095
        common_prefix_len = min(common_prefix_len, num_computed_tokens.min())
2096
        # common_prefix_len should be a multiple of the block size.
2097
2098
2099
2100
2101
2102
2103
2104
2105
2106
2107
        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
        )
2108
2109
        assert isinstance(kv_cache_spec, AttentionSpec)
        use_cascade = attn_metadata_builder.use_cascade_attention(
2110
2111
2112
            common_prefix_len=common_prefix_len,
            query_lens=num_scheduled_tokens,
            num_query_heads=self.num_query_heads,
2113
            num_kv_heads=kv_cache_spec.num_kv_heads,
2114
            use_alibi=self.use_alibi,
2115
            use_sliding_window=use_sliding_window,
2116
            use_local_attention=use_local_attention,
2117
            num_sms=self.num_sms,
2118
            dcp_world_size=self.dcp_world_size,
2119
2120
2121
        )
        return common_prefix_len if use_cascade else 0

2122
2123
    def _calc_mrope_positions(self, scheduler_output: "SchedulerOutput"):
        mrope_pos_ptr = 0
2124
        for index, req_id in enumerate(self.input_batch.req_ids):
2125
2126
2127
            req = self.requests[req_id]
            assert req.mrope_positions is not None

2128
2129
            num_computed_tokens = self.input_batch.num_computed_tokens_cpu[index]
            num_scheduled_tokens = scheduler_output.num_scheduled_tokens[req_id]
2130
            num_prompt_tokens = length_from_prompt_token_ids_or_embeds(
2131
2132
                req.prompt_token_ids, req.prompt_embeds
            )
2133
2134

            if num_computed_tokens + num_scheduled_tokens > num_prompt_tokens:
2135
2136
                prompt_part_len = max(0, num_prompt_tokens - num_computed_tokens)
                completion_part_len = max(0, num_scheduled_tokens - prompt_part_len)
2137
2138
2139
2140
2141
2142
2143
2144
2145
2146
2147
2148
2149
            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

2150
2151
2152
                self.mrope_positions.cpu[:, dst_start:dst_end] = req.mrope_positions[
                    :, src_start:src_end
                ]
2153
2154
2155
2156
2157
2158
2159
                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

2160
                assert req.mrope_position_delta is not None
2161
                MRotaryEmbedding.get_next_input_positions_tensor(
2162
                    out=self.mrope_positions.np,
2163
2164
2165
2166
2167
                    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,
                )
2168
2169
2170

                mrope_pos_ptr += completion_part_len

2171
2172
2173
2174
2175
2176
2177
2178
2179
2180
2181
2182
2183
2184
2185
2186
2187
2188
2189
2190
2191
2192
2193
2194
2195
2196
2197
2198
2199
2200
2201
2202
2203
2204
2205
2206
2207
2208
2209
2210
2211
2212
2213
2214
2215
2216
2217
    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

2218
2219
    def _calc_spec_decode_metadata(
        self,
2220
2221
2222
2223
2224
2225
2226
2227
2228
2229
2230
2231
2232
2233
2234
2235
        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
2236
2237
2238
2239

        # Step 1. cu_num_sampled_tokens: [4, 5, 8, 9, 11]
        # arange: [0, 1, 2, 3, 0, 0, 1, 2, 0, 0, 1]
        cu_num_sampled_tokens, arange = self._get_cumsum_and_arange(
2240
2241
            num_sampled_tokens, cumsum_dtype=np.int32
        )
2242
        # Step 2. [0, 0, 0, 0, 103, 104, 104, 104, 206, 207, 207]
2243
        logits_indices = np.repeat(
2244
2245
            cu_num_scheduled_tokens - num_sampled_tokens, num_sampled_tokens
        )
2246
        # Step 3. [0, 1, 2, 3, 103, 104, 105, 106, 206, 207, 208]
2247
2248
2249
2250
2251
2252
        logits_indices += arange

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

        # Compute the draft logits indices.
2253
2254
2255
        # cu_num_draft_tokens: [3, 3, 5, 5, 6]
        # arange: [0, 1, 2, 0, 1, 0]
        cu_num_draft_tokens, arange = self._get_cumsum_and_arange(
2256
2257
            num_draft_tokens, cumsum_dtype=np.int32
        )
2258
2259
        # [0, 0, 0, 5, 5, 9]
        target_logits_indices = np.repeat(
2260
2261
            cu_num_sampled_tokens - num_sampled_tokens, num_draft_tokens
        )
2262
2263
2264
2265
2266
        # [0, 1, 2, 5, 6, 9]
        target_logits_indices += arange

        # TODO: Optimize the CPU -> GPU copy.
        cu_num_draft_tokens = torch.from_numpy(cu_num_draft_tokens).to(
2267
2268
            self.device, non_blocking=True
        )
2269
2270
2271
        cu_num_sampled_tokens = torch.from_numpy(cu_num_sampled_tokens).to(
            self.device, non_blocking=True
        )
2272
2273
2274
        logits_indices = torch.from_numpy(logits_indices).to(
            self.device, non_blocking=True
        )
2275
        target_logits_indices = torch.from_numpy(target_logits_indices).to(
2276
2277
            self.device, non_blocking=True
        )
2278
        bonus_logits_indices = torch.from_numpy(bonus_logits_indices).to(
2279
2280
            self.device, non_blocking=True
        )
2281

2282
2283
        # Compute the draft token ids.
        # draft_token_indices:      [  1,   2,   3, 105, 106, 208]
2284
        draft_token_ids = self.input_ids.gpu[logits_indices]
2285
2286
        draft_token_ids = draft_token_ids[target_logits_indices + 1]

2287
        return SpecDecodeMetadata(
2288
2289
2290
            draft_token_ids=draft_token_ids,
            num_draft_tokens=num_draft_tokens.tolist(),
            cu_num_draft_tokens=cu_num_draft_tokens,
2291
            cu_num_sampled_tokens=cu_num_sampled_tokens,
2292
2293
2294
2295
2296
            target_logits_indices=target_logits_indices,
            bonus_logits_indices=bonus_logits_indices,
            logits_indices=logits_indices,
        )

2297
2298
2299
2300
2301
2302
2303
    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
2304
        self.kv_sharing_fast_prefill_logits_indices[:num_logits].copy_(logits_indices)
2305
2306
2307
2308
2309
        # 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_(
2310
2311
            logits_indices[-1].item()
        )
2312
2313
        # Dispatch for the decoder portion of the model.
        _, batch_desc = self.cudagraph_dispatcher.dispatch(
2314
            num_logits, invalid_modes={CUDAGraphMode.FULL}
2315
2316
        )
        num_logits_padded = batch_desc.num_tokens
2317
2318
2319
        logits_indices_padded = self.kv_sharing_fast_prefill_logits_indices[
            :num_logits_padded
        ]
2320
2321
        return logits_indices_padded

2322
    def _batch_mm_inputs_from_scheduler(
2323
2324
        self,
        scheduler_output: "SchedulerOutput",
2325
2326
    ) -> tuple[
        list[str],
2327
        list[tuple[str, MultiModalKwargsItem]],
2328
2329
        list[tuple[str, PlaceholderRange]],
    ]:
2330
        """Batch multimodal inputs from scheduled encoder inputs.
2331
2332
2333

        Args:
            scheduler_output: The scheduler output containing scheduled encoder
2334
                inputs.
2335
2336

        Returns:
2337
            A tuple of (mm_hashes, mm_kwargs, mm_lora_refs) where:
2338
2339
            - mm_hashes: List of multimodal hashes for each item
            - mm_kwargs: List of multimodal kwargs for each item
2340
            - mm_lora_refs: List of (req_id, placeholder_range) for each item
2341
        """
2342
2343
        scheduled_encoder_inputs = scheduler_output.scheduled_encoder_inputs
        if not scheduled_encoder_inputs:
2344
            return [], [], []
2345
2346

        mm_hashes = list[str]()
2347
        mm_kwargs = list[tuple[str, MultiModalKwargsItem]]()
2348
2349
2350
        # Multimodal LoRA reference info to map each multimodal item
        # back to its request & position
        mm_lora_refs = list[tuple[str, PlaceholderRange]]()
2351
2352
        for req_id, encoder_input_ids in scheduled_encoder_inputs.items():
            req_state = self.requests[req_id]
2353
2354

            for mm_input_id in encoder_input_ids:
2355
                mm_feature = req_state.mm_features[mm_input_id]
2356
2357
                if mm_feature.data is None:
                    continue
2358
2359

                mm_hashes.append(mm_feature.identifier)
2360
                mm_kwargs.append((mm_feature.modality, mm_feature.data))
2361
                mm_lora_refs.append((req_id, mm_feature.mm_position))
2362

2363
        return mm_hashes, mm_kwargs, mm_lora_refs
2364

2365
2366
2367
    def _execute_mm_encoder(
        self, scheduler_output: "SchedulerOutput"
    ) -> list[torch.Tensor]:
2368
2369
2370
        mm_hashes, mm_kwargs, mm_lora_refs = self._batch_mm_inputs_from_scheduler(
            scheduler_output
        )
2371
2372

        if not mm_kwargs:
2373
            return []
2374

2375
2376
2377
2378
2379
2380
        should_time = bool(
            self.observability_config
            and self.observability_config.enable_mm_processor_stats
            and scheduler_output.scheduled_encoder_inputs
        )

2381
2382
2383
2384
2385
2386
2387
        # 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.
2388
        model = cast(SupportsMultiModal, self.model)
2389
2390
2391
2392
2393
2394
2395
2396
2397
2398
2399
2400
2401
2402
2403

        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]
2404
                    pos_info.get_num_embeds()
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
2442
2443
2444
2445
                )
                prompt_lora_mapping.append(lora_id)
                token_lora_mapping.extend([lora_id] * num_tokens)
                encoder_token_counts.append(num_tokens)

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

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

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

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

                self.lora_manager.set_active_adapters(
                    lora_requests,
                    connector_mapping,
                )

2446
        encoder_outputs: list[torch.Tensor] = []
2447
2448
        # Track the current index in mm_kwargs/mm_lora_refs to map groups to request IDs
        current_item_idx = 0
2449
        for modality, num_items, mm_kwargs_group in group_mm_kwargs_by_modality(
2450
2451
2452
            mm_kwargs,
            device=self.device,
            pin_memory=self.pin_memory,
2453
        ):
2454
            curr_group_outputs: MultiModalEmbeddings
2455
2456

            # EVS-related change.
2457
            # (ekhvedchenia): Temporary hack to limit peak memory usage when
2458
            # processing multimodal data. This solves the issue with scheduler
2459
2460
2461
2462
            # 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)
2463
2464
2465
2466
2467
2468
2469
            # TODO(ywang96): Fix memory profiling to take EVS into account and
            # remove this hack.
            if (
                self.is_multimodal_pruning_enabled
                and modality == "video"
                and num_items > 1
            ):
2470
                curr_group_outputs_lst = list[torch.Tensor]()
2471
2472
2473
2474
2475
2476
2477
2478
2479
2480
2481
                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(
                            group_mm_kwargs_by_modality(
                                [video_mm_kwargs_item],
                                device=self.device,
                                pin_memory=self.pin_memory,
                            )
2482
                        )
2483

2484
2485
2486
                        micro_batch_outputs = model.embed_multimodal(
                            **micro_batch_mm_inputs
                        )
2487

2488
                        curr_group_outputs_lst.extend(micro_batch_outputs)
2489
2490

                curr_group_outputs = curr_group_outputs_lst
2491
2492
2493
2494
2495
2496
2497
2498
            else:
                # Run the encoder.
                # `curr_group_outputs` is either of the following:
                # 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.
2499
2500
2501
2502
2503

                with self.timed_encoder_operation(
                    should_time, mm_lora_refs, current_item_idx, num_items
                ):
                    curr_group_outputs = model.embed_multimodal(**mm_kwargs_group)
2504

2505
2506
            sanity_check_mm_encoder_outputs(
                curr_group_outputs,
2507
                expected_num_items=num_items,
2508
            )
2509
            encoder_outputs.extend(curr_group_outputs)
2510

2511
2512
            current_item_idx += num_items

2513
        # Cache the encoder outputs by mm_hash
2514
        for mm_hash, output in zip(mm_hashes, encoder_outputs):
2515
            self.encoder_cache[mm_hash] = output
2516
2517
            logger.debug("Finish execute for mm hash %s", mm_hash)
            self.maybe_save_ec_to_connector(self.encoder_cache, mm_hash)
2518

2519
2520
        return encoder_outputs

2521
    def _gather_mm_embeddings(
2522
2523
        self,
        scheduler_output: "SchedulerOutput",
2524
        shift_computed_tokens: int = 0,
2525
2526
2527
    ) -> tuple[list[torch.Tensor], torch.Tensor]:
        total_num_scheduled_tokens = scheduler_output.total_num_scheduled_tokens

2528
2529
2530
2531
2532
        # Swap to the other buffer to avoid race condition with previous
        # iteration's async copy that may still be reading from CPU.
        self.is_mm_embed_idx = 1 - self.is_mm_embed_idx
        is_mm_embed_buf = self.is_mm_embed_buffers[self.is_mm_embed_idx]

2533
        mm_embeds = list[torch.Tensor]()
2534
        is_mm_embed = is_mm_embed_buf.cpu
2535
2536
2537
        is_mm_embed[:total_num_scheduled_tokens] = False

        req_start_idx = 0
2538
        should_sync_mrope_positions = False
2539
        should_sync_xdrope_positions = False
2540

2541
        for req_id in self.input_batch.req_ids:
2542
2543
            mm_embeds_req: list[torch.Tensor] = []

2544
            num_scheduled_tokens = scheduler_output.num_scheduled_tokens[req_id]
2545
            req_state = self.requests[req_id]
2546
            num_computed_tokens = req_state.num_computed_tokens + shift_computed_tokens
2547

2548
2549
            for mm_feature in req_state.mm_features:
                pos_info = mm_feature.mm_position
2550
2551
                start_pos = pos_info.offset
                num_encoder_tokens = pos_info.length
2552
2553
2554
2555
2556
2557
2558
2559
2560
2561
2562
2563
2564
2565
2566
2567

                # 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,
2568
2569
                    num_encoder_tokens,
                )
2570
                assert start_idx < end_idx
2571
2572
2573
2574
2575
2576
2577
                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
2578

2579
                mm_hash = mm_feature.identifier
2580
                encoder_output = self.encoder_cache.get(mm_hash, None)
2581
                assert encoder_output is not None, f"Encoder cache miss for {mm_hash}."
2582
2583
2584

                if (is_embed := pos_info.is_embed) is not None:
                    is_embed = is_embed[start_idx:end_idx]
2585
2586
2587
                    mm_embeds_item = encoder_output[curr_embeds_start:curr_embeds_end]
                else:
                    mm_embeds_item = encoder_output[start_idx:end_idx]
2588

2589
                req_start_pos = req_start_idx + start_pos - num_computed_tokens
2590
2591
2592
2593
2594
2595
2596
2597
2598
                # 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
2599
2600
2601
                mm_embeds_req.append(mm_embeds_item)

            if self.is_multimodal_pruning_enabled and self.uses_mrope:
2602
                assert req_state.mrope_positions is not None
2603
2604
2605
2606
2607
2608
2609
                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,
2610
2611
                    )
                )
2612
2613
2614
2615
                req_state.mrope_positions.copy_(new_mrope_positions)
                req_state.mrope_position_delta = new_delta

            mm_embeds.extend(mm_embeds_req)
2616
2617
            req_start_idx += num_scheduled_tokens

2618
        is_mm_embed = is_mm_embed_buf.copy_to_gpu(total_num_scheduled_tokens)
2619
2620
2621

        if should_sync_mrope_positions:
            self._calc_mrope_positions(scheduler_output)
2622
            self.mrope_positions.copy_to_gpu(total_num_scheduled_tokens)
2623

2624
2625
2626
2627
        if should_sync_xdrope_positions:
            self._calc_xdrope_positions(scheduler_output)
            self.xdrope_positions.copy_to_gpu(total_num_scheduled_tokens)

2628
        return mm_embeds, is_mm_embed
2629

2630
    def get_model(self) -> nn.Module:
2631
2632
        if not hasattr(self, "model"):
            raise ValueError("Cannot get model before model has been initialized")
2633
        if isinstance(self.model, (CUDAGraphWrapper, UBatchWrapper)):
2634
            # get raw model out of the cudagraph wrapper.
2635
            return self.model.unwrap()
2636
2637
        return self.model

2638
2639
2640
2641
2642
2643
2644
2645
2646
2647
2648
2649
2650
    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")

2651
2652
2653
        if supports_realtime(model):
            supported_tasks.append("realtime")

2654
2655
        return supported_tasks

2656
2657
2658
2659
2660
    def get_supported_pooling_tasks(self) -> list[PoolingTask]:
        model = self.get_model()
        if not is_pooling_model(model):
            return []

2661
2662
        supported_tasks = list(model.pooler.get_supported_tasks())

2663
2664
2665
2666
        if "score" in supported_tasks:
            num_labels = getattr(self.model_config.hf_config, "num_labels", 0)
            if num_labels != 1:
                supported_tasks.remove("score")
2667
                logger.debug_once("Score API is only enabled for num_labels == 1.")
2668
2669

        return supported_tasks
2670

2671
2672
2673
2674
2675
2676
2677
2678
2679
2680
    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)

2681
    def sync_and_slice_intermediate_tensors(
2682
2683
        self,
        num_tokens: int,
2684
        intermediate_tensors: IntermediateTensors | None,
2685
2686
        sync_self: bool,
    ) -> IntermediateTensors:
2687
2688
2689
        assert self.intermediate_tensors is not None

        tp = self.vllm_config.parallel_config.tensor_parallel_size
2690
        is_rs = is_residual_scattered_for_sp(self.vllm_config, num_tokens)
2691
2692
2693
2694
2695
2696

        # 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():
2697
                is_scattered = k == "residual" and is_rs
2698
                copy_len = num_tokens // tp if is_scattered else num_tokens
2699
                self.intermediate_tensors[k][:copy_len].copy_(
2700
2701
2702
2703
2704
2705
2706
2707
2708
2709
2710
2711
2712
                    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:
2713
2714
2715
        """
        Step for the EPLB (Expert Parallelism Load Balancing) state.
        """
2716
        if not self.parallel_config.enable_eplb or self.eep_eplb_suppressed:
2717
2718
2719
            return

        assert self.eplb_state is not None
2720
2721
        model = self.get_model()
        assert is_mixture_of_experts(model)
2722
2723
2724
        self.eplb_state.step(
            is_dummy,
            is_profile,
2725
            log_stats=self.parallel_config.eplb_config.log_balancedness,
2726
2727
        )

2728
2729
2730
2731
2732
2733
2734
2735
2736
2737
2738
2739
2740
2741
2742
2743
2744
    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,
        )

2745
2746
2747
2748
2749
    def _pool(
        self,
        hidden_states: torch.Tensor,
        num_scheduled_tokens: int,
        num_scheduled_tokens_np: np.ndarray,
2750
2751
2752
2753
        kv_connector_output: KVConnectorOutput | None,
    ) -> ModelRunnerOutput | AsyncModelRunnerOutput:
        num_reqs = self.input_batch.num_reqs
        assert num_reqs == len(self.input_batch.pooling_params), (
2754
2755
            "Either all or none of the requests in a batch must be pooling request"
        )
2756

2757
        hidden_states = hidden_states[:num_scheduled_tokens]
2758
        seq_lens_cpu = self.seq_lens.cpu[:num_reqs]
2759

2760
        pooling_metadata = self.input_batch.get_pooling_metadata()
2761
        pooling_metadata.build_pooling_cursor(
2762
            num_scheduled_tokens_np, seq_lens_cpu, device=hidden_states.device
2763
        )
2764

2765
2766
        model = cast(VllmModelForPooling, self.model)
        raw_pooler_output: PoolerOutput = model.pooler(
2767
            hidden_states=hidden_states, pooling_metadata=pooling_metadata
2768
        )
2769
2770
2771
2772
2773
2774
2775
2776
2777
2778
2779
2780
2781
2782
2783
2784
2785
2786
2787
2788
2789
2790
2791
2792

        finished_mask = [
            seq_len == prompt_len
            for seq_len, prompt_len in zip(seq_lens_cpu, pooling_metadata.prompt_lens)
        ]

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

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

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

2793
2794
2795
        model_runner_output.pooler_output = _copy_pooler_output_to_cpu(
            raw_pooler_output=raw_pooler_output,
            finished_mask=finished_mask,
2796
        )
2797
2798
        self._sync_device()

2799
        return model_runner_output
2800

2801
    def _pad_for_sequence_parallelism(self, num_scheduled_tokens: int) -> int:
2802
2803
2804
        # Pad tokens to multiple of tensor_parallel_size when
        # enabled collective fusion for SP
        tp_size = self.vllm_config.parallel_config.tensor_parallel_size
2805
        if self.compilation_config.pass_config.enable_sp and tp_size > 1:
2806
2807
2808
            return round_up(num_scheduled_tokens, tp_size)
        return num_scheduled_tokens

Patrick von Platen's avatar
Patrick von Platen committed
2809
2810
2811
2812
2813
2814
2815
2816
2817
2818
2819
    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

2820
    def _preprocess(
2821
2822
        self,
        scheduler_output: "SchedulerOutput",
2823
        num_input_tokens: int,  # Padded
2824
        intermediate_tensors: IntermediateTensors | None = None,
2825
    ) -> tuple[
2826
2827
        torch.Tensor | None,
        torch.Tensor | None,
2828
        torch.Tensor,
2829
        IntermediateTensors | None,
2830
        dict[str, Any],
2831
        ECConnectorOutput | None,
2832
    ]:
2833
        num_scheduled_tokens = scheduler_output.total_num_scheduled_tokens
2834
        is_first_rank = get_pp_group().is_first_rank
2835
        is_encoder_decoder = self.model_config.is_encoder_decoder
2836

2837
2838
        # _prepare_inputs may reorder the batch, so we must gather multi
        # modal outputs after that to ensure the correct order
2839
2840
        ec_connector_output = None

2841
        if self.supports_mm_inputs and is_first_rank and not is_encoder_decoder:
2842
            # Run the multimodal encoder if any.
2843
2844
2845
2846
2847
2848
            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)
2849

2850
2851
2852
            # 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.
2853
            inputs_embeds_scheduled = self.model.embed_input_ids(
2854
2855
2856
                self.input_ids.gpu[:num_scheduled_tokens],
                multimodal_embeddings=mm_embeds,
                is_multimodal=is_mm_embed,
2857
            )
2858

2859
            # TODO(woosuk): Avoid the copy. Optimize.
2860
            self.inputs_embeds.gpu[:num_scheduled_tokens].copy_(inputs_embeds_scheduled)
2861

Patrick von Platen's avatar
Patrick von Platen committed
2862
            input_ids, inputs_embeds = self._prepare_mm_inputs(num_input_tokens)
2863
            model_kwargs = {
2864
                **self._init_model_kwargs(),
2865
2866
                **self._extract_mm_kwargs(scheduler_output),
            }
2867
        elif self.enable_prompt_embeds and is_first_rank:
2868
2869
2870
2871
2872
2873
2874
2875
2876
2877
2878
2879
            # 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).
2880
2881
2882
            token_ids_idx = (
                self.is_token_ids.gpu[:num_scheduled_tokens]
                .nonzero(as_tuple=False)
2883
                .squeeze(1)
2884
            )
2885
2886
2887
            # Some tokens ids may need to become embeds
            if token_ids_idx.numel() > 0:
                token_ids = self.input_ids.gpu[token_ids_idx]
2888
                tokens_to_embeds = self.model.embed_input_ids(input_ids=token_ids)
2889
2890
2891
                self.inputs_embeds.gpu[token_ids_idx] = tokens_to_embeds

            inputs_embeds = self.inputs_embeds.gpu[:num_input_tokens]
2892
            model_kwargs = self._init_model_kwargs()
2893
            input_ids = None
2894
        else:
2895
2896
2897
2898
            # 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.
2899
            input_ids = self.input_ids.gpu[:num_input_tokens]
2900
            inputs_embeds = None
2901
            model_kwargs = self._init_model_kwargs()
2902

2903
        if self.uses_mrope:
2904
            positions = self.mrope_positions.gpu[:, :num_input_tokens]
2905
2906
        elif self.uses_xdrope_dim > 0:
            positions = self.xdrope_positions.gpu[:, :num_input_tokens]
2907
        else:
2908
            positions = self.positions.gpu[:num_input_tokens]
2909

2910
        if is_first_rank:
2911
2912
            intermediate_tensors = None
        else:
2913
            assert intermediate_tensors is not None
2914
            intermediate_tensors = self.sync_and_slice_intermediate_tensors(
2915
2916
                num_input_tokens, intermediate_tensors, True
            )
2917

2918
        if is_encoder_decoder and scheduler_output.scheduled_encoder_inputs:
2919
2920
2921
2922
2923
2924
2925
            # 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})
2926

2927
2928
2929
2930
2931
2932
        return (
            input_ids,
            inputs_embeds,
            positions,
            intermediate_tensors,
            model_kwargs,
2933
            ec_connector_output,
2934
        )
2935

2936
    def _sample(
2937
        self,
2938
2939
        logits: torch.Tensor | None,
        spec_decode_metadata: SpecDecodeMetadata | None,
2940
    ) -> SamplerOutput:
2941
        # Sample the next token and get logprobs if needed.
2942
        sampling_metadata = self.input_batch.sampling_metadata
2943
2944
2945
        # 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()
2946
        if spec_decode_metadata is None:
2947
            return self.sampler(
2948
2949
2950
                logits=logits,
                sampling_metadata=sampling_metadata,
            )
2951

2952
2953
2954
2955
2956
2957
        # 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)

2958
        sampler_output = self.rejection_sampler(
2959
2960
            spec_decode_metadata,
            None,  # draft_probs
2961
            logits,
2962
2963
            sampling_metadata,
        )
2964
2965
2966
        return sampler_output

    def _bookkeeping_sync(
2967
2968
2969
        self,
        scheduler_output: "SchedulerOutput",
        sampler_output: SamplerOutput,
2970
        logits: torch.Tensor | None,
2971
2972
        hidden_states: torch.Tensor,
        num_scheduled_tokens: int,
2973
        spec_decode_metadata: SpecDecodeMetadata | None,
2974
    ) -> tuple[
2975
        dict[str, int],
2976
        LogprobsLists | None,
2977
        list[list[int]],
2978
        dict[str, LogprobsTensors | None],
2979
2980
2981
        list[str],
        dict[str, int],
        list[int],
2982
    ]:
2983
2984
2985
2986
        num_nans_in_logits = {}
        if envs.VLLM_COMPUTE_NANS_IN_LOGITS:
            num_nans_in_logits = self._get_nans_in_logits(logits)

2987
2988
2989
2990
        num_reqs = self.input_batch.num_reqs
        discard_sampled_tokens_req_indices = np.nonzero(
            self.discard_request_mask.np[:num_reqs]
        )[0]
2991
2992
2993
2994
        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)
2995

2996
2997
2998
        # 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()
2999
        req_id_to_index_output_copy = self.input_batch.req_id_to_index.copy()
3000
3001

        num_sampled_tokens = sampler_output.sampled_token_ids.shape[0]
3002
        sampled_token_ids = sampler_output.sampled_token_ids
3003
        logprobs_tensors = sampler_output.logprobs_tensors
3004
        invalid_req_indices = []
3005
        logprobs_lists = None
3006
3007
3008
3009
3010
3011
        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)
3012
3013
3014
                # 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()
3015
3016
3017

                if logprobs_tensors is not None:
                    logprobs_lists = logprobs_tensors.tolists()
3018
3019
            else:
                # Includes spec decode tokens.
3020
                valid_sampled_token_ids, logprobs_lists = RejectionSampler.parse_output(
3021
3022
                    sampled_token_ids,
                    self.input_batch.vocab_size,
3023
                    discard_sampled_tokens_req_indices,
3024
                    logprobs_tensors=logprobs_tensors,
3025
                )
3026
        else:
3027
            valid_sampled_token_ids = []
3028
            invalid_req_indices = discard_sampled_tokens_req_indices.tolist()
3029
3030
3031
3032
3033
            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.
3034
3035
3036
3037
            # 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
3038
3039
3040
3041
3042
            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
            }
3043

3044
3045
3046
3047
3048
        # 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.
3049
        req_ids = self.input_batch.req_ids
3050
3051
        for req_idx in range(num_sampled_tokens):
            if self.use_async_scheduling:
3052
                sampled_ids = [-1] if req_idx not in invalid_req_indices_set else None
3053
3054
            else:
                sampled_ids = valid_sampled_token_ids[req_idx]
3055

3056
            num_sampled_ids: int = len(sampled_ids) if sampled_ids else 0
3057

3058
            if not sampled_ids:
3059
3060
3061
                continue

            start_idx = self.input_batch.num_tokens_no_spec[req_idx]
3062
            end_idx = start_idx + num_sampled_ids
3063
3064
3065
3066
            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}"
3067
            )
3068

3069
3070
            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
3071
            self.input_batch.num_tokens_no_spec[req_idx] = end_idx
3072

3073
            req_id = req_ids[req_idx]
3074
3075
3076
            req_state = self.requests[req_id]
            req_state.output_token_ids.extend(sampled_ids)

3077
3078
3079
3080
3081
3082
        # Compute prompt logprobs if needed.
        prompt_logprobs_dict = self._get_prompt_logprobs_dict(
            hidden_states[:num_scheduled_tokens],
            scheduler_output.num_scheduled_tokens,
        )

3083
3084
3085
3086
3087
3088
3089
3090
3091
3092
        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,
        )

3093
3094
3095
3096
3097
3098
3099
3100
3101
3102
3103
3104
3105
3106
3107
    @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()

3108
3109
    def _model_forward(
        self,
3110
3111
3112
3113
        input_ids: torch.Tensor | None = None,
        positions: torch.Tensor | None = None,
        intermediate_tensors: IntermediateTensors | None = None,
        inputs_embeds: torch.Tensor | None = None,
3114
3115
3116
3117
3118
        **model_kwargs: dict[str, Any],
    ) -> Any:
        """Helper method to call the model forward pass.

        This method can be overridden by subclasses for model execution.
3119
        Motivation: We can inspect only this method versus
3120
3121
3122
3123
3124
3125
3126
3127
3128
3129
3130
3131
3132
3133
3134
3135
3136
3137
3138
3139
        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,
        )

3140
3141
3142
3143
3144
3145
3146
3147
3148
3149
3150
3151
3152
3153
3154
3155
3156
3157
3158
3159
3160
    @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
        )

3161
3162
3163
3164
3165
3166
3167
3168
3169
3170
3171
3172
3173
    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,
3174
        force_num_active_loras: int | None = None,
3175
        num_encoder_reqs: int = 0,
3176
    ) -> tuple[
3177
3178
        CUDAGraphMode,
        BatchDescriptor,
3179
        bool,
3180
3181
        torch.Tensor | None,
        CUDAGraphStat | None,
3182
    ]:
3183
3184
3185
3186
3187
3188
        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,
3189
        )
3190
3191
3192
3193
3194
        # 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
        )
3195

3196
3197
3198
3199
3200
        # 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)
3201
        )
3202
        has_lora = num_active_loras > 0 if force_has_lora is None else force_has_lora
3203

3204
        num_tokens_padded = self._pad_for_sequence_parallelism(num_tokens)
3205
3206
3207

        def dispatch_cudagraph(num_tokens, disable_full=False, valid_modes=None):
            return self.cudagraph_dispatcher.dispatch(
3208
3209
3210
                num_tokens=num_tokens,
                has_lora=has_lora,
                uniform_decode=uniform_decode,
3211
                num_active_loras=num_active_loras,
3212
3213
                valid_modes={CUDAGraphMode.NONE} if force_eager else valid_modes,
                invalid_modes={CUDAGraphMode.FULL} if disable_full else None,
3214
3215
            )

3216
        cudagraph_mode, batch_descriptor = dispatch_cudagraph(
3217
            num_tokens_padded, disable_full=use_cascade_attn or has_encoder_output
3218
        )
3219
        num_tokens_padded = batch_descriptor.num_tokens
3220
3221
3222
3223
3224
3225
3226
3227
3228
        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"
            )
3229
3230
3231

        # Extra coordination when running data-parallel since we need to coordinate
        # across ranks
3232
        should_ubatch, num_tokens_across_dp = False, None
3233
        if self.vllm_config.parallel_config.data_parallel_size > 1:
3234
3235
3236
3237
3238
3239
3240
3241
3242
3243
            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,
                )
3244
3245
            )

3246
            # Extract DP-synced values
3247
3248
3249
            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())
3250
3251
3252
                # Re-dispatch with DP padding so we have the correct batch_descriptor
                cudagraph_mode, batch_descriptor = dispatch_cudagraph(
                    num_tokens_padded,
3253
                    valid_modes={CUDAGraphMode(synced_cudagraph_mode)},
3254
                )
3255
3256
3257
3258
                # 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

3259
3260
3261
3262
3263
3264
3265
3266
3267
3268
3269
3270
        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,
3271
            should_ubatch,
3272
3273
3274
            num_tokens_across_dp,
            cudagraph_stats,
        )
3275

3276
3277
3278
3279
3280
3281
3282
3283
3284
3285
3286
3287
3288
3289
3290
3291
3292
3293
3294
3295
3296
3297
3298
3299
3300
3301
3302
3303
3304
3305
3306
3307
3308
3309
3310
3311
    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

3312
3313
3314
3315
3316
3317
3318
3319
3320
3321
3322
3323
3324
3325
3326
3327
3328
3329
3330
3331
3332
3333
3334
3335
3336
3337
3338
3339
3340
3341
3342
3343
3344
3345
3346
3347
3348
3349
3350
3351
3352
3353
3354
3355
3356
3357
3358
3359
3360
3361
3362
3363
3364
3365
3366
3367
3368
3369
3370
3371
3372
3373
3374
3375
3376
3377
3378
3379
3380
3381
3382
3383
3384
3385
    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

3386
3387
3388
3389
    @torch.inference_mode()
    def execute_model(
        self,
        scheduler_output: "SchedulerOutput",
3390
        intermediate_tensors: IntermediateTensors | None = None,
3391
    ) -> ModelRunnerOutput | AsyncModelRunnerOutput | IntermediateTensors | None:
3392
3393
3394
3395
3396
        if self.execute_model_state is not None:
            raise RuntimeError(
                "State error: sample_tokens() must be called "
                "after execute_model() returns None."
            )
3397

3398
3399
3400
3401
3402
3403
3404
        if self.vllm_config.model_config.enable_return_routed_experts:
            capturer = RoutedExpertsCapturer.get_instance()
            if capturer is not None:
                capturer.clear_buffer()  # noqa
            else:
                logger.error("RoutedExpertsCapturer not initialized.")

3405
3406
3407
3408
3409
        if scheduler_output.preempted_req_ids and has_kv_transfer_group():
            get_kv_transfer_group().handle_preemptions(
                scheduler_output.preempted_req_ids
            )

3410
        num_scheduled_tokens = scheduler_output.total_num_scheduled_tokens
3411
3412
3413
3414
3415
3416
        with (
            record_function_or_nullcontext("gpu_model_runner: preprocess"),
            self.synchronize_input_prep(),
        ):
            # Update persistent batch states.
            self._update_states(scheduler_output)
3417

3418
            if has_ec_transfer() and not get_ec_transfer().is_consumer:
3419
                with self.maybe_get_ec_connector_output(
3420
                    scheduler_output,
3421
3422
3423
3424
3425
3426
3427
3428
3429
3430
3431
3432
3433
3434
3435
3436
3437
3438
3439
3440
3441
3442
3443
3444
3445
3446
3447
3448
                    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"
3449
3450
                )

3451
3452
3453
3454
3455
3456
            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
3457

3458
3459
3460
3461
            logits_indices, spec_decode_metadata = self._prepare_inputs(
                scheduler_output,
                num_scheduled_tokens_np,
            )
3462

3463
3464
3465
3466
3467
            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(
3468
                    num_scheduled_tokens_np,
3469
3470
                    self.input_batch.num_computed_tokens_cpu[:num_reqs],
                    scheduler_output.num_common_prefix_blocks,
3471
3472
                )

3473
3474
3475
3476
3477
3478
3479
3480
3481
3482
3483
3484
3485
3486
            (
                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),
            )
3487

3488
3489
3490
3491
3492
3493
3494
3495
3496
3497
3498
3499
3500
3501
3502
3503
3504
3505
3506
3507
3508
3509
3510
3511
3512
3513
3514
            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,
            )

3515
3516
3517
3518
3519
3520
3521
3522
3523
3524
3525
            # 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)
            )
3526
3527
            pad_attn = cudagraph_mode == CUDAGraphMode.FULL

3528
3529
3530
3531
3532
3533
3534
3535
3536
3537
            if self.cache_config.mamba_cache_mode == "align":
                mamba_utils.preprocess_mamba(
                    scheduler_output,
                    self.kv_cache_config,
                    self.cache_config,
                    self.mamba_state_idx,
                    self.input_batch,
                    self.requests,
                    self.compilation_config.static_forward_context,
                    self.model.get_mamba_state_copy_func(),
3538
                    self._get_mamba_copy_bufs(),
3539
3540
                )

3541
3542
3543
            use_spec_decode = len(scheduler_output.scheduled_spec_decode_tokens) > 0
            ubatch_slices_attn = ubatch_slices_padded if pad_attn else ubatch_slices

3544
3545
3546
3547
3548
3549
3550
3551
3552
3553
3554
            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,
            )

3555
3556
3557
3558
3559
3560
3561
3562
3563
3564
3565
3566
            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,
3567
                    slot_mappings=slot_mappings_by_group,
3568
                )
3569
            )
3570

3571
3572
3573
3574
3575
3576
3577
3578
3579
            (
                input_ids,
                inputs_embeds,
                positions,
                intermediate_tensors,
                model_kwargs,
                ec_connector_output,
            ) = self._preprocess(
                scheduler_output, num_tokens_padded, intermediate_tensors
3580
            )
3581

3582
        # Set cudagraph mode to none if calc_kv_scales is true.
3583
3584
3585
        # KV scales calculation involves dynamic operations that are incompatible
        # with CUDA graph capture.
        if self.calculate_kv_scales:
3586
            cudagraph_mode = CUDAGraphMode.NONE
3587
3588
            # Mark KV scales as calculated after the first forward pass
            self.calculate_kv_scales = False
3589

3590
3591
3592
3593
3594
3595
3596
        # 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
        )

3597
3598
        # Run the model.
        # Use persistent buffers for CUDA graphs.
3599
3600
3601
        # When spec decode is enabled, delay clearing connector metadata
        # until after draft model runs in sample_tokens.
        clear_kv_metadata = self.speculative_config is None
3602
3603
        with (
            set_forward_context(
3604
3605
                attn_metadata,
                self.vllm_config,
3606
                num_tokens=num_tokens_padded,
3607
                num_tokens_across_dp=num_tokens_across_dp,
3608
3609
                cudagraph_runtime_mode=cudagraph_mode,
                batch_descriptor=batch_desc,
3610
                ubatch_slices=ubatch_slices_padded,
3611
                slot_mapping=slot_mappings,
3612
                skip_compiled=has_encoder_input,
3613
            ),
3614
            record_function_or_nullcontext("gpu_model_runner: forward"),
3615
3616
3617
            self.maybe_get_kv_connector_output(
                scheduler_output, clear_metadata=clear_kv_metadata
            ) as kv_connector_output,
3618
        ):
3619
            model_output = self._model_forward(
3620
3621
3622
3623
3624
3625
3626
                input_ids=input_ids,
                positions=positions,
                intermediate_tensors=intermediate_tensors,
                inputs_embeds=inputs_embeds,
                **model_kwargs,
            )

3627
        with record_function_or_nullcontext("gpu_model_runner: postprocess"):
3628
            if self.use_aux_hidden_state_outputs:
3629
                # True when EAGLE 3 is used.
3630
3631
                hidden_states, aux_hidden_states = model_output
            else:
3632
                # Common case.
3633
3634
3635
                hidden_states = model_output
                aux_hidden_states = None

3636
3637
3638
3639
3640
            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)
3641
                    hidden_states.kv_connector_output = kv_connector_output
3642
                    self.kv_connector_output = kv_connector_output
3643
                    return hidden_states
3644

3645
                if self.is_pooling_model:
3646
                    # Return the pooling output.
3647
3648
3649
3650
3651
                    return self._pool(
                        hidden_states,
                        num_scheduled_tokens,
                        num_scheduled_tokens_np,
                        kv_connector_output,
3652
                    )
3653
3654

                sample_hidden_states = hidden_states[logits_indices]
3655
                logits = self.model.compute_logits(sample_hidden_states)
3656
3657
3658
3659
            else:
                # Rare case.
                assert not self.is_pooling_model

3660
                sample_hidden_states = hidden_states[logits_indices]
3661
                if not get_pp_group().is_last_rank:
3662
                    all_gather_tensors = {
3663
                        "residual": not is_residual_scattered_for_sp(
3664
                            self.vllm_config, num_tokens_padded
3665
                        )
3666
                    }
3667
                    get_pp_group().send_tensor_dict(
3668
3669
                        hidden_states.tensors,
                        all_gather_group=get_tp_group(),
3670
3671
                        all_gather_tensors=all_gather_tensors,
                    )
3672
3673
                    logits = None
                else:
3674
                    logits = self.model.compute_logits(sample_hidden_states)
3675

3676
                model_output_broadcast_data: dict[str, Any] = {}
3677
3678
3679
                if logits is not None:
                    model_output_broadcast_data["logits"] = logits.contiguous()

3680
                broadcasted = get_pp_group().broadcast_tensor_dict(
3681
3682
                    model_output_broadcast_data, src=len(get_pp_group().ranks) - 1
                )
3683
3684
                assert broadcasted is not None
                logits = broadcasted["logits"]
3685

3686
3687
3688
3689
3690
3691
3692
3693
        self.execute_model_state = ExecuteModelState(
            scheduler_output,
            logits,
            spec_decode_metadata,
            spec_decode_common_attn_metadata,
            hidden_states,
            sample_hidden_states,
            aux_hidden_states,
3694
            ec_connector_output,
3695
            cudagraph_stats,
3696
            slot_mappings,
3697
        )
3698
        self.kv_connector_output = kv_connector_output
3699
3700
3701
3702
3703
3704
3705
        return None

    @torch.inference_mode
    def sample_tokens(
        self, grammar_output: "GrammarOutput | None"
    ) -> ModelRunnerOutput | AsyncModelRunnerOutput | IntermediateTensors:
        if self.execute_model_state is None:
3706
3707
            kv_connector_output = self.kv_connector_output
            self.kv_connector_output = None
3708
3709
3710
            # 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()
3711
            if not kv_connector_output:
3712
                return None  # type: ignore[return-value]
3713
3714
3715
3716
3717
3718
3719
3720
3721

            # 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
3722
3723
3724
3725
3726
3727
3728
3729
3730
3731

        # Unpack ephemeral state.
        (
            scheduler_output,
            logits,
            spec_decode_metadata,
            spec_decode_common_attn_metadata,
            hidden_states,
            sample_hidden_states,
            aux_hidden_states,
3732
            ec_connector_output,
3733
            cudagraph_stats,
3734
            slot_mappings,
3735
3736
3737
3738
3739
3740
3741
3742
3743
        ) = 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
            )
3744

3745
        with record_function_or_nullcontext("gpu_model_runner: sample"):
3746
3747
            sampler_output = self._sample(logits, spec_decode_metadata)

3748
3749
3750
        self._update_states_after_model_execute(
            sampler_output.sampled_token_ids, scheduler_output
        )
3751
3752
        if self.use_async_scheduling:
            pp = get_pp_group()
3753
3754
3755
3756
            # 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:
3757
3758
3759
                self._pp_broadcast_prev_sampled_token_ids(
                    sampler_output.sampled_token_ids
                )
3760

3761
3762
        self._draft_token_ids = None
        self._draft_token_req_ids = None
3763
3764
        self.input_batch.prev_sampled_token_ids = None

3765
        def propose_draft_token_ids(sampled_token_ids):
3766
            assert spec_decode_common_attn_metadata is not None
3767
            with record_function_or_nullcontext("gpu_model_runner: draft"):
3768
3769
3770
3771
3772
3773
3774
3775
3776
                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,
3777
                    slot_mappings,
3778
                )
3779
                self._copy_draft_token_ids_to_cpu(scheduler_output)
3780

3781
        spec_config = self.speculative_config
3782
3783
3784
3785
3786
        propose_drafts_after_bookkeeping = False
        if spec_config is not None:
            input_fits_in_drafter = spec_decode_common_attn_metadata is not None and (
                spec_decode_common_attn_metadata.max_seq_len + self.num_spec_tokens
                <= self.effective_drafter_max_model_len
3787
            )
3788
            use_gpu_toks = (
3789
3790
3791
                spec_config.use_eagle()
                or spec_config.uses_draft_model()
                or spec_config.uses_extract_hidden_states()
3792
3793
3794
            ) and not spec_config.disable_padded_drafter_batch
            if use_gpu_toks:
                # EAGLE/DraftModel speculative decoding can use the GPU sampled tokens
3795
                # as inputs, and does not need to wait for bookkeeping to finish.
3796
3797
3798
3799
                assert isinstance(
                    self.drafter,
                    EagleProposer | DraftModelProposer | ExtractHiddenStatesProposer,
                )
3800
3801
3802
3803
3804
3805
3806
3807
3808
3809
3810
3811
3812
                sampled_token_ids = sampler_output.sampled_token_ids
                if input_fits_in_drafter:
                    propose_draft_token_ids(sampled_token_ids)
                elif self.valid_sampled_token_count_event is not None:
                    assert spec_decode_common_attn_metadata is not None
                    next_token_ids, valid_sampled_tokens_count = (
                        self.drafter.prepare_next_token_ids_padded(
                            spec_decode_common_attn_metadata,
                            sampled_token_ids,
                            self.requests,
                            self.input_batch,
                            self.discard_request_mask.gpu,
                        )
3813
                    )
3814
3815
3816
3817
3818
3819
3820
3821
3822
3823
3824
                    self._copy_valid_sampled_token_count(
                        next_token_ids, valid_sampled_tokens_count
                    )
                    # Since we couldn't run the drafter,
                    # just use zeros for the draft tokens.
                    self._draft_token_ids = torch.zeros(
                        1, device=self.device, dtype=torch.int32
                    ).expand(len(self.input_batch.req_ids), self.num_spec_tokens)
                    self._copy_draft_token_ids_to_cpu(scheduler_output, zeros_only=True)
            else:
                propose_drafts_after_bookkeeping = input_fits_in_drafter
3825

3826
        with record_function_or_nullcontext("gpu_model_runner: bookkeep"):
3827
3828
3829
3830
3831
3832
3833
3834
            (
                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,
3835
3836
3837
3838
3839
            ) = self._bookkeeping_sync(
                scheduler_output,
                sampler_output,
                logits,
                hidden_states,
3840
                scheduler_output.total_num_scheduled_tokens,
3841
                spec_decode_metadata,
3842
            )
3843

3844
        if propose_drafts_after_bookkeeping:
3845
3846
3847
            # 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)
3848

3849
3850
3851
3852
3853
3854
        # Clear KV connector metadata after draft model runs (if spec decode).
        # This was deferred from target model forward to allow draft model
        # to also save its KV cache.
        if self.speculative_config is not None:
            self.clear_kv_connector_metadata()

3855
        with record_function_or_nullcontext("gpu_model_runner: eplb"):
3856
            self.eplb_step()
3857

3858
3859
3860
3861
        # self.kv_connector_output may be modified during drafting
        kv_connector_output = self.kv_connector_output
        self.kv_connector_output = None

3862
        with record_function_or_nullcontext("gpu_model_runner: ModelRunnerOutput"):
3863
3864
3865
3866
3867
3868
3869
            if self.model_config.enable_return_routed_experts:
                capturer = RoutedExpertsCapturer.get_instance()
                if capturer is not None:
                    capturer.save_captured_experts(indices=self.slot_mapping)  # noqa
                else:
                    logger.error("RoutedExpertsCapturer not initialized.")

3870
3871
3872
3873
3874
3875
3876
            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,
3877
3878
3879
                ec_connector_output=ec_connector_output
                if self.supports_mm_inputs
                else None,
3880
                num_nans_in_logits=num_nans_in_logits,
3881
                cudagraph_stats=cudagraph_stats,
3882
            )
3883

3884
3885
        if not self.use_async_scheduling:
            return output
3886

3887
3888
3889
3890
3891
3892
3893
3894
3895
        with record_function_or_nullcontext(
            "gpu_model_runner: AsyncGPUModelRunnerOutput"
        ):
            async_output = AsyncGPUModelRunnerOutput(
                model_runner_output=output,
                sampled_token_ids=sampler_output.sampled_token_ids,
                logprobs_tensors=sampler_output.logprobs_tensors,
                invalid_req_indices=invalid_req_indices,
                async_output_copy_stream=self.async_output_copy_stream,
3896
                vocab_size=self.input_batch.vocab_size,
3897
3898
3899
3900
3901
            )
        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
3902
            # any requests with sampling params that require output ids.
3903
3904
3905
3906
            self.input_batch.set_async_sampled_token_ids(
                async_output.sampled_token_ids_cpu,
                async_output.async_copy_ready_event,
            )
3907
3908
3909

        return async_output

3910
3911
3912
3913
3914
3915
3916
3917
3918
3919
3920
3921
3922
3923
3924
3925
3926
3927
3928
3929
3930
3931
3932
3933
3934
3935
3936
3937
3938
3939
3940
3941
3942
3943
3944
3945
3946
3947
3948
    def _pp_broadcast_prev_sampled_token_ids(
        self, sampled_token_ids: torch.Tensor
    ) -> None:
        """Broadcast sampled token ids (GPU) from last PP stage"""
        pp = get_pp_group()
        assert pp.is_last_rank
        # `prev_sampled_token_ids` is expected to have shape [num_reqs, 1].
        assert sampled_token_ids.dim() == 2 and sampled_token_ids.shape[-1] == 1, (
            "PP+async expects sampled_token_ids to have shape [num_reqs, 1]"
        )
        torch.distributed.broadcast(
            sampled_token_ids, src=pp.rank, group=pp.device_group
        )

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

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

3949
    def take_draft_token_ids(self) -> DraftTokenIds | None:
3950
        if not self.num_spec_tokens or not self._draft_token_req_ids:
3951
            return None
3952
        draft_token_ids, req_ids = self._get_draft_token_ids_cpu()
3953
        return DraftTokenIds(req_ids, draft_token_ids)
3954

3955
3956
3957
    def _copy_draft_token_ids_to_cpu(
        self, scheduler_output: "SchedulerOutput", zeros_only: bool = False
    ) -> None:
3958
3959
3960
3961
3962
3963
        # 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
        ):
3964
3965
3966
            return
        # We must also set the corresponding request ids.
        self._draft_token_req_ids = self.input_batch.req_ids.copy()
3967

3968
3969
3970
3971
3972
3973
3974
3975
3976
3977
3978
3979
3980
3981
3982
3983
3984
3985
3986
3987
        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()

3988
    def _get_draft_token_ids_cpu(self) -> tuple[list[list[int]], list[str]]:
3989
        if isinstance(self._draft_token_ids, list):
3990
3991
3992
3993
            return self._draft_token_ids, self.input_batch.req_ids
        req_ids = self._draft_token_req_ids
        if req_ids is None:
            return [], []
3994
3995
3996
        assert self.draft_token_ids_event is not None
        assert self.draft_token_ids_cpu is not None
        self.draft_token_ids_event.synchronize()
3997
        return self.draft_token_ids_cpu[: len(req_ids)].tolist(), req_ids
3998

3999
4000
4001
4002
4003
4004
4005
4006
4007
4008
4009
4010
4011
    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
4012
            assert counts_cpu is not None
4013
4014
4015
4016
4017
4018
4019
4020
            counts_cpu[: counts.shape[0]].copy_(counts, non_blocking=True)
            self.valid_sampled_token_count_event.record()

        self.input_batch.prev_sampled_token_ids = next_token_ids.unsqueeze(1)

    def _get_valid_sampled_token_count(self) -> list[int]:
        # Wait until valid_sampled_tokens_count is copied to cpu,
        prev_sampled_token_ids = self.input_batch.prev_sampled_token_ids
4021
4022
        sampled_count_event = self.valid_sampled_token_count_event
        if sampled_count_event is None or prev_sampled_token_ids is None:
4023
4024
4025
            return []

        counts_cpu = self.valid_sampled_token_count_cpu
4026
4027
        assert counts_cpu is not None
        sampled_count_event.synchronize()
4028
4029
        return counts_cpu[: prev_sampled_token_ids.shape[0]].tolist()

4030
4031
4032
    def propose_draft_token_ids(
        self,
        scheduler_output: "SchedulerOutput",
4033
        sampled_token_ids: torch.Tensor | list[list[int]],
4034
4035
4036
        sampling_metadata: SamplingMetadata,
        hidden_states: torch.Tensor,
        sample_hidden_states: torch.Tensor,
4037
4038
        aux_hidden_states: list[torch.Tensor] | None,
        spec_decode_metadata: SpecDecodeMetadata | None,
4039
        common_attn_metadata: CommonAttentionMetadata,
4040
        slot_mappings: dict[str, torch.Tensor] | list[dict[str, torch.Tensor]] | None,
4041
    ) -> list[list[int]] | torch.Tensor:
4042
        num_scheduled_tokens = scheduler_output.total_num_scheduled_tokens
4043
4044
4045
        spec_config = self.speculative_config
        assert spec_config is not None
        if spec_config.method == "ngram":
4046
4047
            from vllm.v1.spec_decode.ngram_proposer import NgramProposer

4048
            assert isinstance(sampled_token_ids, list)
4049
            assert isinstance(self.drafter, NgramProposer)
4050
            draft_token_ids = self.drafter.propose(
4051
                sampled_token_ids,
4052
4053
                self.input_batch.num_tokens_no_spec,
                self.input_batch.token_ids_cpu,
4054
                slot_mappings=slot_mappings,
4055
            )
4056
        elif spec_config.method == "suffix":
4057
4058
            assert isinstance(sampled_token_ids, list)
            assert isinstance(self.drafter, SuffixDecodingProposer)
4059
4060
4061
            draft_token_ids = self.drafter.propose(
                self.input_batch, sampled_token_ids, slot_mappings=slot_mappings
            )
4062
        elif spec_config.method == "medusa":
4063
            assert isinstance(sampled_token_ids, list)
4064
            assert isinstance(self.drafter, MedusaProposer)
4065

4066
4067
            if sample_hidden_states.shape[0] == len(sampled_token_ids):
                # The input to the target model does not include draft tokens.
4068
4069
4070
4071
                hidden_states = sample_hidden_states
            else:
                indices = []
                offset = 0
4072
4073
4074
                assert spec_decode_metadata is not None, (
                    "No spec decode metadata for medusa"
                )
4075
                for num_draft, tokens in zip(
4076
4077
                    spec_decode_metadata.num_draft_tokens, sampled_token_ids
                ):
4078
                    indices.append(offset + len(tokens) - 1)
4079
                    offset += num_draft + 1
4080
                indices = torch.tensor(indices, device=self.device)
4081
4082
                hidden_states = sample_hidden_states[indices]

4083
            draft_token_ids = self.drafter.propose(
4084
4085
                target_hidden_states=hidden_states,
                sampling_metadata=sampling_metadata,
4086
                slot_mappings=slot_mappings,
4087
            )
4088
4089
4090
4091
4092
4093
4094
4095
4096
4097
4098
4099
4100
4101
4102
4103
4104
4105
4106
4107
4108
4109
4110
4111
4112
4113
4114
4115
4116
4117
4118
4119
4120
4121
4122
4123
4124
4125
4126
4127
4128
4129
        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]

            draft_token_ids, drafter_kv_connector_output = self.drafter.propose(
                sampled_token_ids=sampled_token_ids,
                target_hidden_states=target_hidden_states,
                common_attn_metadata=common_attn_metadata,
                scheduler_output=scheduler_output,
                slot_mappings=slot_mappings,
            )
            # Combine KVConnectorOutputs or select the non-empty one
            if self.kv_connector_output and drafter_kv_connector_output:
                self.kv_connector_output = KVConnectorOutput.merge(
                    self.kv_connector_output, drafter_kv_connector_output
                )
            else:
                self.kv_connector_output = (
                    self.kv_connector_output or drafter_kv_connector_output
                )

            next_token_ids, valid_sampled_tokens_count = (
                self.drafter.prepare_next_token_ids_padded(
                    common_attn_metadata,
                    sampled_token_ids,
                    self.requests,
                    self.input_batch,
                    self.discard_request_mask.gpu,
                )
            )
            self._copy_valid_sampled_token_count(
                next_token_ids, valid_sampled_tokens_count
            )

4130
4131
        elif spec_config.use_eagle() or spec_config.uses_draft_model():
            assert isinstance(self.drafter, EagleProposer | DraftModelProposer)
4132

4133
            if spec_config.disable_padded_drafter_batch:
4134
4135
4136
                # 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.
4137
4138
                assert isinstance(sampled_token_ids, list), (
                    "sampled_token_ids should be a python list when"
4139
                    "padded-batch is disabled."
4140
                )
4141
                next_token_ids = self.drafter.prepare_next_token_ids_cpu(
4142
4143
4144
4145
4146
                    sampled_token_ids,
                    self.requests,
                    self.input_batch,
                    scheduler_output.num_scheduled_tokens,
                )
4147
4148
4149
4150
4151
            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.
4152
4153
                assert isinstance(sampled_token_ids, torch.Tensor), (
                    "sampled_token_ids should be a torch.Tensor when"
4154
                    "padded-batch is enabled."
4155
4156
                )
                next_token_ids, valid_sampled_tokens_count = (
4157
4158
4159
4160
4161
                    self.drafter.prepare_next_token_ids_padded(
                        common_attn_metadata,
                        sampled_token_ids,
                        self.requests,
                        self.input_batch,
4162
                        self.discard_request_mask.gpu,
4163
                    )
4164
                )
4165
4166
4167
                self._copy_valid_sampled_token_count(
                    next_token_ids, valid_sampled_tokens_count
                )
Jiayi Yao's avatar
Jiayi Yao committed
4168

4169
            num_rejected_tokens_gpu = None
4170
            if spec_decode_metadata is None:
4171
                token_indices_to_sample = None
4172
                # input_ids can be None for multimodal models.
4173
                target_token_ids = self.input_ids.gpu[:num_scheduled_tokens]
4174
                target_positions = self._get_positions(num_scheduled_tokens)
4175
                if self.use_aux_hidden_state_outputs:
Wentao Ye's avatar
Wentao Ye committed
4176
                    assert aux_hidden_states is not None
4177
                    target_hidden_states = torch.cat(
4178
4179
                        [h[:num_scheduled_tokens] for h in aux_hidden_states], dim=-1
                    )
4180
4181
                else:
                    target_hidden_states = hidden_states[:num_scheduled_tokens]
4182
            else:
4183
                if spec_config.disable_padded_drafter_batch:
4184
                    token_indices_to_sample = None
4185
4186
4187
4188
4189
                    common_attn_metadata, token_indices = self.drafter.prepare_inputs(
                        common_attn_metadata,
                        sampled_token_ids,
                        spec_decode_metadata.num_draft_tokens,
                    )
4190
4191
4192
4193
4194
4195
4196
4197
4198
                    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]
4199
                else:
4200
4201
4202
4203
4204
4205
4206
4207
                    (
                        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,
4208
                    )
4209
4210
4211
4212
4213
4214
4215
4216
4217
4218
4219
                    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]
4220

4221
            if self.supports_mm_inputs and self.drafter.supports_mm_inputs:
4222
4223
4224
4225
4226
4227
                mm_embed_inputs = self._gather_mm_embeddings(
                    scheduler_output,
                    shift_computed_tokens=1,
                )
            else:
                mm_embed_inputs = None
4228

4229
            draft_token_ids = self.drafter.propose(
4230
4231
4232
4233
                target_token_ids=target_token_ids,
                target_positions=target_positions,
                target_hidden_states=target_hidden_states,
                next_token_ids=next_token_ids,
4234
                token_indices_to_sample=token_indices_to_sample,
4235
                sampling_metadata=sampling_metadata,
4236
                common_attn_metadata=common_attn_metadata,
4237
                mm_embed_inputs=mm_embed_inputs,
4238
                num_rejected_tokens_gpu=num_rejected_tokens_gpu,
4239
                slot_mappings=slot_mappings,
4240
            )
4241

4242
        return draft_token_ids
4243

4244
4245
4246
    def update_config(self, overrides: dict[str, Any]) -> None:
        allowed_config_names = {"load_config", "model_config"}
        for config_name, config_overrides in overrides.items():
4247
4248
            assert config_name in allowed_config_names, (
                f"Config `{config_name}` not supported. "
4249
                f"Allowed configs: {allowed_config_names}"
4250
            )
4251
4252
4253
4254
            config = getattr(self, config_name)
            new_config = update_config(config, config_overrides)
            setattr(self, config_name, new_config)

4255
    @instrument(span_name="Loading (GPU)")
4256
    def load_model(self, load_dummy_weights: bool = False) -> None:
4257
4258
        """
        Args:
4259
            load_dummy_weights: load dummy weights instead of real weights.
4260
        """
4261
4262
4263
4264
4265
        logger.info_once(
            "Starting to load model %s...",
            self.model_config.model,
            scope="global",
        )
4266

4267
4268
4269
4270
        if self.parallel_config.enable_eplb:
            self.eplb_state = EplbState(self.parallel_config, self.device)
            eplb_models = 0

4271
4272
4273
        try:
            with DeviceMemoryProfiler() as m:
                time_before_load = time.perf_counter()
4274
4275
                if load_dummy_weights:
                    self.load_config.load_format = "dummy"
4276
4277
4278
4279
4280
4281
4282
                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
4283
                    )
4284
4285
4286
4287
4288
4289
4290
4291
                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
                    ):
4292
4293
4294
                        assert not self.parallel_config.enable_elastic_ep, (
                            "Elastic EP is not supported with drafter model."
                        )
4295
4296
4297
4298
4299
4300
4301
4302
4303
4304
4305
4306
4307
4308
4309
4310
                        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
4311

4312
4313
4314
4315
4316
4317
                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"
                        )
4318

4319
4320
4321
4322
4323
4324
4325
4326
4327
4328
4329
4330
4331
4332
4333
4334
4335
4336
4337
4338
4339
4340
4341
4342
4343
                    # 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:
                        aux_layers = self.model.get_eagle3_aux_hidden_state_layers()

                    self.model.set_aux_hidden_state_layers(aux_layers)
                time_after_load = time.perf_counter()
            self.model_memory_usage = m.consumed_memory
        except torch.cuda.OutOfMemoryError as e:
            msg = (
                "Failed to load model - not enough GPU memory. "
                "Try lowering --gpu-memory-utilization to free memory for weights, "
                "increasing --tensor-parallel-size, or using --quantization. "
                "See https://docs.vllm.ai/en/latest/configuration/conserving_memory/ "
                "for more tips."
            )
            combined_msg = f"{msg} (original error: {e})"
            logger.error(combined_msg)
            raise e
4344
        logger.info_once(
4345
4346
            "Model loading took %s GiB memory and %.6f seconds",
            format_gib(self.model_memory_usage),
4347
            time_after_load - time_before_load,
4348
            scope="local",
4349
        )
4350
4351
4352
4353
4354
4355
        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)
4356
        mm_config = self.model_config.multimodal_config
4357
        self.is_multimodal_pruning_enabled = (
4358
            supports_multimodal_pruning(self.get_model())
4359
4360
            and mm_config is not None
            and mm_config.is_multimodal_pruning_enabled()
4361
        )
4362

4363
4364
4365
4366
4367
        if (
            is_mixture_of_experts(self.model)
            and self.parallel_config.enable_eplb
            and not load_dummy_weights
        ):
4368
4369
4370
            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(
4371
                self.model,
4372
                self.model_config,
4373
            )
4374
            if self.eplb_state.is_async:
4375
                self.eplb_state.start_async_loop()
4376

4377
        if (
4378
4379
            self.vllm_config.compilation_config.mode
            == CompilationMode.STOCK_TORCH_COMPILE
4380
        ):
4381
            backend = self.vllm_config.compilation_config.init_backend(self.vllm_config)
4382
            compilation_counter.stock_torch_compile_count += 1
4383
            self.model.compile(fullgraph=True, backend=backend)
4384
            return
4385
        # for other compilation modes, cudagraph behavior is controlled by
Jiayi Yan's avatar
Jiayi Yan committed
4386
        # CudagraphWrapper and CudagraphDispatcher of vllm.
4387
4388

        # wrap the model with full cudagraph wrapper if needed.
4389
4390
        cudagraph_mode = self.compilation_config.cudagraph_mode
        assert cudagraph_mode is not None
4391
4392
4393
4394
        if (
            cudagraph_mode.has_full_cudagraphs()
            and not self.parallel_config.use_ubatching
        ):
4395
4396
4397
            self.model = CUDAGraphWrapper(
                self.model, self.vllm_config, runtime_mode=CUDAGraphMode.FULL
            )
4398
        elif self.parallel_config.use_ubatching:
4399
            if cudagraph_mode.has_full_cudagraphs():
4400
4401
4402
                self.model = UBatchWrapper(
                    self.model, self.vllm_config, CUDAGraphMode.FULL, self.device
                )
4403
            else:
4404
4405
4406
                self.model = UBatchWrapper(
                    self.model, self.vllm_config, CUDAGraphMode.NONE, self.device
                )
4407

4408
4409
        get_offloader().post_init()

4410
    def _get_eagle3_aux_layers_from_config(self) -> tuple[int, ...] | None:
4411
4412
4413
4414
4415
4416
4417
4418
4419
4420
4421
4422
4423
4424
4425
4426
4427
4428
4429
4430
4431
4432
4433
        """Extract Eagle3 auxiliary layer indices from speculative config.

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

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

        hf_config = self.speculative_config.draft_model_config.hf_config
        if not hasattr(hf_config, "eagle_aux_hidden_state_layer_ids"):
            return None

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

        return None

4434
4435
4436
4437
4438
4439
4440
4441
4442
4443
4444
4445
4446
    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
4447
            into kernel format (repacking, renaming, etc.)
4448
4449
4450
4451
4452
4453
4454
4455
4456
4457
4458
4459
4460
4461
4462
4463
4464
4465
4466
4467
4468
4469
4470
4471
4472
4473
4474
4475
4476
4477
4478
4479
4480
4481
4482
4483
4484
4485
4486
4487
4488
4489
4490
4491
4492
4493
4494
4495
4496
4497
4498
4499
4500
4501
4502
4503
4504
4505
4506
4507
4508
        """
        # TODO(@kylesayrs): generalize to all runners and loaders
        # argument validation
        if weights_iterator is None and not is_checkpoint_format:
            logger.warning(
                "Reloading from disk means that weights will be in checkpoint format. "
                "Please use `is_checkpoint_format=True` "
                "to avoid weight reloading errors"
            )

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

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

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

        # begin loading weights
        logger.info_once("Reloading weights inplace...", scope="local")
        load_device = (
            self.vllm_config.load_config.device or self.vllm_config.device_config.device
        )
        with torch.device(load_device):
            if is_checkpoint_format:
                # load weights from checkpoint/ original model format
                initialize_layerwise_reload(model)
                loaded_weights = model.load_weights(weights_iterator)
                finalize_layerwise_reload(model, self.model_config)

            else:
                # load weights from kernel format
                logger.warning_once(
                    "Reloading with `is_checkpoint_format=True` requires that "
                    "weights be in kernel format and already sharded",
                    scope="local",
                )
                loaded_weights = set()
                for name, loaded_weight in weights_iterator:
                    param = model.get_parameter(name)  # TODO: buffers?
                    param.copy_(loaded_weight)
                    loaded_weights.add(name)

        # logging and validation
        counter_after_reloading = time.perf_counter()
        diff_seconds = counter_after_reloading - counter_before_reloading
        logger.info_once(
            "Reloading and processing weights took %.2f seconds",
            diff_seconds,
            scope="local",
4509
        )
4510
4511
4512
4513
4514
4515
4516
        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,
                )
4517

4518
4519
4520
    def _get_prompt_logprobs_dict(
        self,
        hidden_states: torch.Tensor,
4521
        num_scheduled_tokens: dict[str, int],
4522
    ) -> dict[str, LogprobsTensors | None]:
4523
        num_prompt_logprobs_dict = self.num_prompt_logprobs
4524
4525
4526
        if not num_prompt_logprobs_dict:
            return {}

4527
        in_progress_dict = self.input_batch.in_progress_prompt_logprobs_cpu
4528
        prompt_logprobs_dict: dict[str, LogprobsTensors | None] = {}
4529
4530
4531
4532
4533

        # 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():
4534
4535
4536
4537
            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
4538
4539
4540

            # Get metadata for this request.
            request = self.requests[req_id]
4541
4542
4543
4544
            if request.prompt_token_ids is None:
                # Prompt logprobs is incompatible with prompt embeddings
                continue

4545
4546
            num_prompt_tokens = len(request.prompt_token_ids)
            prompt_token_ids = torch.tensor(request.prompt_token_ids).to(
4547
4548
                self.device, non_blocking=True
            )
4549

4550
4551
4552
4553
4554
4555
            # 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(
4556
4557
                    num_prompt_tokens - 1, num_prompt_logprobs + 1
                )
4558
4559
                in_progress_dict[req_id] = logprobs_tensors

4560
            # Determine number of logits to retrieve.
4561
4562
            start_idx = request.num_computed_tokens
            start_tok = start_idx + 1
4563
            num_remaining_tokens = num_prompt_tokens - start_tok
4564
            if num_tokens <= num_remaining_tokens:
4565
                # This is a chunk, more tokens remain.
4566
4567
4568
                # 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.
4569
4570
4571
4572
4573
                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)
4574
4575
4576
4577
4578
4579
4580
                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
4581
4582
4583
4584
4585

            # 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]
4586
            offset = self.query_start_loc.np[req_idx].item()
4587
            prompt_hidden_states = hidden_states[offset : offset + num_logits]
4588
            logits = self.model.compute_logits(prompt_hidden_states)
4589
4590
4591
4592

            # 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.
4593
            tgt_token_ids = prompt_token_ids[start_tok : start_tok + num_logits]
4594
4595

            # Compute prompt logprobs.
4596
            logprobs = self.sampler.compute_logprobs(logits)
4597
            token_ids, logprobs, ranks, _ = self.sampler.gather_logprobs(
4598
4599
                logprobs, num_prompt_logprobs, tgt_token_ids
            )
4600
4601

            # Transfer GPU->CPU async.
4602
4603
            chunk_slice = slice(start_idx, start_idx + num_logits)
            logprobs_tensors.logprob_token_ids[chunk_slice].copy_(
4604
4605
4606
                token_ids, non_blocking=True
            )
            logprobs_tensors.logprobs[chunk_slice].copy_(logprobs, non_blocking=True)
4607
            logprobs_tensors.selected_token_ranks[chunk_slice].copy_(
4608
4609
                ranks, non_blocking=True
            )
4610
4611
4612
4613
4614

        # 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]
4615
            del in_progress_dict[req_id]
4616
4617

        # Must synchronize the non-blocking GPU->CPU transfers.
4618
        if prompt_logprobs_dict:
4619
            self._sync_device()
4620
4621
4622

        return prompt_logprobs_dict

4623
4624
    def _get_nans_in_logits(
        self,
4625
        logits: torch.Tensor | None,
4626
4627
4628
4629
4630
4631
4632
4633
4634
4635
4636
    ) -> 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])
4637
4638
4639
                    if num_nans_for_index is not None and req_index < logits.shape[0]
                    else 0
                )
4640
4641
4642
4643
            return num_nans_in_logits
        except IndexError:
            return {}

4644
    @contextmanager
4645
4646
4647
    def maybe_randomize_inputs(
        self, input_ids: torch.Tensor | None, inputs_embeds: torch.Tensor | None
    ):
4648
4649
4650
4651
        """
        Randomize input_ids if VLLM_RANDOMIZE_DP_DUMMY_INPUTS is set.
        This is to help balance expert-selection
         - during profile_run
4652
         - during DP rank dummy run
4653
        """
4654

4655
4656
4657
4658
        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
4659
        elif input_ids is not None:
4660
4661
4662
4663

            @functools.cache
            def rand_input_ids() -> torch.Tensor:
                return torch.randint_like(
4664
                    self.input_ids.gpu,
4665
4666
                    low=0,
                    high=self.model_config.get_vocab_size(),
4667
                )
4668

4669
            logger.debug_once("Randomizing dummy input_ids for DP Rank")
4670
            input_ids.copy_(rand_input_ids()[: input_ids.size(0)], non_blocking=True)
4671
4672
            yield
            input_ids.fill_(0)
4673
4674
4675
4676
4677
4678
4679
4680
4681
4682
4683
4684
4685
4686
4687
        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)
4688

4689
4690
4691
4692
4693
4694
    def _get_mm_dummy_batch(
        self,
        modality: str,
        max_items_per_batch: int,
    ) -> BatchedTensorInputs:
        """Dummy data for profiling and precompiling multimodal models."""
4695
4696
        assert self.mm_budget is not None

4697
4698
4699
        # 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,
4700
            mm_counts={modality: 1},
4701
            cache=self.mm_budget.cache,
4702
        )
4703
4704
4705
4706
4707
        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"
4708

4709
4710
4711
        return next(
            mm_kwargs_group
            for _, _, mm_kwargs_group in group_mm_kwargs_by_modality(
4712
                [(modality, dummy_mm_item)] * max_items_per_batch,
4713
4714
4715
4716
                device=self.device,
                pin_memory=self.pin_memory,
            )
        )
4717

4718
4719
4720
4721
    @torch.inference_mode()
    def _dummy_run(
        self,
        num_tokens: int,
4722
        cudagraph_runtime_mode: CUDAGraphMode | None = None,
4723
4724
        force_attention: bool = False,
        uniform_decode: bool = False,
4725
        allow_microbatching: bool = True,
4726
4727
        skip_eplb: bool = False,
        is_profile: bool = False,
4728
        create_mixed_batch: bool = False,
4729
        remove_lora: bool = True,
Rémi Delacourt's avatar
Rémi Delacourt committed
4730
        is_graph_capturing: bool = False,
4731
        num_active_loras: int = 0,
4732
    ) -> tuple[torch.Tensor, torch.Tensor]:
4733
4734
4735
4736
4737
4738
4739
        """
        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.
4740
                - if not set will determine the cudagraph mode based on using
4741
                    the self.cudagraph_dispatcher.
4742
4743
4744
4745
                - CUDAGraphMode.NONE: No cudagraph, for warm up and profile run
                - CUDAGraphMode.PIECEWISE: Piecewise cudagraph.
                - CUDAGraphMode.FULL: Full cudagraph, attention metadata is
                    needed.
4746
            force_attention: If True, always create attention metadata. Used to
4747
4748
4749
4750
                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.
4751
4752
            create_mixed_batch: If True, create a mixed batch with both decode
                (1 token) and prefill (multiple tokens) requests.
4753
            remove_lora: If False, dummy LoRAs are not destroyed after the run
4754
4755
            num_active_loras: Number of distinct active LoRAs to capture for.
                LoRA is activated when num_active_loras > 0.
4756
        """
4757
4758
        mm_config = self.vllm_config.model_config.multimodal_config
        if mm_config and mm_config.mm_encoder_only:
4759
4760
4761
4762
            # 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([])

4763
4764
        assert (
            cudagraph_runtime_mode is None
4765
            or cudagraph_runtime_mode.is_valid_runtime_mode()
4766
        )
4767

4768
        # If cudagraph_mode.decode_mode() == FULL and
4769
        # cudagraph_mode.separate_routine(). This means that we are using
4770
4771
4772
4773
4774
4775
4776
4777
4778
4779
4780
        # 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.
4781
        max_query_len = self.uniform_decode_query_len if uniform_decode else num_tokens
4782

4783
4784
4785
        # 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.
4786
        assert num_tokens <= self.max_num_tokens
4787
        max_num_reqs = self.scheduler_config.max_num_seqs
4788
4789
4790
4791
        if create_mixed_batch:
            assert not uniform_decode
            # Create mixed batch:
            # first half decode tokens, second half one prefill
4792
            num_decode_tokens = min(max_num_reqs - 1, num_tokens // 2)
4793
4794
4795
4796
            num_prefill_tokens = num_tokens - num_decode_tokens
            num_reqs = num_decode_tokens + 1

            # Create decode requests (1 token each) followed by prefill request
4797
            num_scheduled_tokens_list = [1] * num_decode_tokens + [num_prefill_tokens]
4798
4799
4800
            # Note: Overriding max_query_len to be the prefill tokens
            max_query_len = num_prefill_tokens
        elif uniform_decode:
4801
            assert not create_mixed_batch
4802
            num_reqs = min(max_num_reqs, cdiv(num_tokens, max_query_len))
4803
4804
            num_scheduled_tokens_list = [max_query_len] * num_reqs
            if num_tokens % max_query_len != 0:
4805
                num_scheduled_tokens_list[-1] = num_tokens % max_query_len
4806
4807
4808
4809
4810
4811
        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

4812
4813
        assert sum(num_scheduled_tokens_list) == num_tokens
        assert len(num_scheduled_tokens_list) == num_reqs
4814
        num_scheduled_tokens = np.array(num_scheduled_tokens_list, dtype=np.int32)
4815
4816
        num_tokens_unpadded = int(num_scheduled_tokens.sum())

4817
        num_sampled_tokens = np.ones(num_reqs, dtype=np.int32)
4818

4819
        _cudagraph_mode, batch_desc, should_ubatch, num_tokens_across_dp, _ = (
4820
4821
4822
4823
4824
4825
4826
4827
4828
4829
4830
4831
4832
4833
4834
4835
4836
            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
4837
4838
4839
4840
                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,
4841
4842
            )
        )
4843
4844
4845

        if cudagraph_runtime_mode is None:
            cudagraph_runtime_mode = _cudagraph_mode
4846
        else:
4847
4848
4849
4850
4851
4852
4853
4854
4855
            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
        )
4856
        ubatch_slices, ubatch_slices_padded = maybe_create_ubatch_slices(
4857
4858
4859
4860
4861
4862
4863
4864
4865
4866
            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,
4867
        )
4868

4869
        attn_metadata: PerLayerAttnMetadata | None = None
4870

4871
4872
4873
4874
4875
4876
4877
        slot_mappings_by_group, slot_mappings = self._get_slot_mappings(
            num_tokens_padded=num_tokens,
            num_reqs_padded=num_reqs_padded,
            num_tokens_unpadded=num_tokens_unpadded,
            ubatch_slices=ubatch_slices_padded,
        )

4878
4879
4880
4881
4882
4883
4884
4885
4886
4887
4888
4889
4890
4891
4892
4893
4894
4895
        # _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:
                if create_mixed_batch:
                    # 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
                    seq_lens = [1] * num_decode_tokens + [num_prefill_tokens + 1]
                else:
                    seq_lens = max_query_len  # type: ignore[assignment]
                self.seq_lens.np[:num_reqs] = seq_lens
                self.seq_lens.np[num_reqs:] = 0
                self.seq_lens.copy_to_gpu()
4896

4897
4898
4899
                cum_num_tokens, _ = self._get_cumsum_and_arange(num_scheduled_tokens)
                self.query_start_loc.np[1 : num_reqs + 1] = cum_num_tokens
                self.query_start_loc.copy_to_gpu()
4900

4901
4902
4903
4904
4905
4906
4907
4908
4909
                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,
4910
                    use_spec_decode=self.speculative_config is not None,
4911
                )
4912

4913
        with self.maybe_dummy_run_with_lora(
4914
4915
4916
4917
            self.lora_config,
            num_scheduled_tokens,
            num_sampled_tokens,
            remove_lora,
4918
            num_active_loras,
4919
        ):
4920
            # Make sure padding doesn't exceed max_num_tokens
4921
            assert num_tokens_padded <= self.max_num_tokens
4922
            model_kwargs = self._init_model_kwargs()
4923
            if self.supports_mm_inputs and not self.model_config.is_encoder_decoder:
Patrick von Platen's avatar
Patrick von Platen committed
4924
4925
                input_ids, inputs_embeds = self._prepare_mm_inputs(num_tokens_padded)

4926
                model_kwargs = {
4927
                    **model_kwargs,
4928
4929
                    **self._dummy_mm_kwargs(num_reqs),
                }
4930
4931
            elif self.enable_prompt_embeds:
                input_ids = None
4932
                inputs_embeds = self.inputs_embeds.gpu[:num_tokens_padded]
4933
                model_kwargs = self._init_model_kwargs()
4934
            else:
4935
                input_ids = self.input_ids.gpu[:num_tokens_padded]
4936
                inputs_embeds = None
4937

4938
            if self.uses_mrope:
4939
                positions = self.mrope_positions.gpu[:, :num_tokens_padded]
4940
            elif self.uses_xdrope_dim > 0:
4941
                positions = self.xdrope_positions.gpu[:, :num_tokens_padded]
4942
            else:
4943
                positions = self.positions.gpu[:num_tokens_padded]
4944
4945
4946
4947
4948
4949
4950
4951
4952

            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,
4953
4954
4955
                            device=self.device,
                        )
                    )
4956
4957

                intermediate_tensors = self.sync_and_slice_intermediate_tensors(
4958
                    num_tokens_padded, None, False
4959
                )
4960

4961
            if ubatch_slices_padded is not None:
4962
4963
4964
                # Adjust values to reflect a single ubatch.
                # TODO(sage,lucas): this is cruft that should be addressed in
                #  the padding refactor.
4965
                num_tokens_padded = ubatch_slices_padded[0].num_tokens
4966
                if num_tokens_across_dp is not None:
4967
                    num_tokens_across_dp[:] = num_tokens_padded
4968

4969
            with (
4970
                self.maybe_randomize_inputs(input_ids, inputs_embeds),
4971
                set_forward_context(
4972
4973
                    attn_metadata,
                    self.vllm_config,
4974
                    num_tokens=num_tokens_padded,
4975
4976
                    num_tokens_across_dp=num_tokens_across_dp,
                    cudagraph_runtime_mode=cudagraph_runtime_mode,
4977
                    batch_descriptor=batch_desc,
4978
                    ubatch_slices=ubatch_slices_padded,
4979
                    slot_mapping=slot_mappings,
4980
4981
                ),
            ):
4982
                outputs = self.model(
4983
4984
4985
4986
                    input_ids=input_ids,
                    positions=positions,
                    intermediate_tensors=intermediate_tensors,
                    inputs_embeds=inputs_embeds,
4987
                    **model_kwargs,
4988
                )
4989

4990
4991
4992
4993
            if self.use_aux_hidden_state_outputs:
                hidden_states, _ = outputs
            else:
                hidden_states = outputs
4994

4995
4996
4997
            if self.speculative_config and (
                self.speculative_config.use_eagle()
                or self.speculative_config.uses_draft_model()
4998
                or self.speculative_config.uses_extract_hidden_states()
4999
            ):
5000
5001
5002
5003
                assert isinstance(
                    self.drafter,
                    EagleProposer | DraftModelProposer | ExtractHiddenStatesProposer,
                )
5004
                assert self.speculative_config is not None
5005
5006
5007
                # 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.
5008
                use_cudagraphs = (
5009
5010
5011
5012
5013
5014
5015
5016
5017
                    (
                        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
5018
5019
5020
5021
5022

                # 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
5023
5024
5025
5026
                if (
                    self.compilation_config.cudagraph_specialize_lora
                    and num_active_loras > 0
                ):
5027
5028
5029
5030
5031
                    use_cudagraphs = False

                self.drafter.dummy_run(
                    num_tokens,
                    use_cudagraphs=use_cudagraphs,
Rémi Delacourt's avatar
Rémi Delacourt committed
5032
                    is_graph_capturing=is_graph_capturing,
5033
                    slot_mappings=slot_mappings,
5034
                )
5035

5036
5037
5038
5039
5040
5041
5042
5043
5044
5045
5046
        # 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()

5047
5048
5049
5050
5051
5052
5053
5054
5055
5056
        # 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)

5057
        logit_indices = np.cumsum(num_scheduled_tokens) - 1
5058
5059
5060
5061
        logit_indices_device = torch.from_numpy(logit_indices).to(
            self.device, non_blocking=True
        )
        return hidden_states, hidden_states[logit_indices_device]
5062
5063
5064
5065
5066
5067

    @torch.inference_mode()
    def _dummy_sampler_run(
        self,
        hidden_states: torch.Tensor,
    ) -> torch.Tensor:
5068
5069
5070
        # The dummy hidden states may contain special values,
        # like `inf` or `nan`.
        # To avoid breaking the sampler, we use a random tensor here instead.
5071

5072
5073
        mm_config = self.vllm_config.model_config.multimodal_config
        if mm_config and mm_config.mm_encoder_only:
5074
5075
5076
            # MM Encoder only model no need to run sampler.
            return torch.tensor([])

5077
        hidden_states = torch.rand_like(hidden_states)
5078

5079
        logits = self.model.compute_logits(hidden_states)
5080
5081
        num_reqs = logits.size(0)

5082
        dummy_tensors = lambda v: torch.full((num_reqs,), v, device=self.device)
5083
5084
5085
5086
5087
5088
5089
5090
5091
5092
5093
5094
5095
5096
5097

        dummy_metadata = SamplingMetadata(
            temperature=dummy_tensors(0.5),
            all_greedy=False,
            all_random=False,
            top_p=dummy_tensors(0.9),
            top_k=dummy_tensors(logits.size(1) - 1),
            generators={},
            max_num_logprobs=None,
            no_penalties=True,
            prompt_token_ids=None,
            frequency_penalties=dummy_tensors(0.1),
            presence_penalties=dummy_tensors(0.1),
            repetition_penalties=dummy_tensors(0.1),
            output_token_ids=[[] for _ in range(num_reqs)],
5098
            spec_token_ids=[[] for _ in range(num_reqs)],
5099
5100
            allowed_token_ids_mask=None,
            bad_words_token_ids={},
5101
            logitsprocs=LogitsProcessors(),
5102
        )
5103
        try:
5104
5105
5106
            sampler_output = self.sampler(
                logits=logits, sampling_metadata=dummy_metadata
            )
5107
        except RuntimeError as e:
5108
            if "out of memory" in str(e):
5109
5110
5111
5112
                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 "
5113
5114
                    "initializing the engine."
                ) from e
5115
5116
            else:
                raise e
5117
        if self.speculative_config:
5118
5119
            draft_token_ids = [[0] for _ in range(num_reqs)]
            dummy_spec_decode_metadata = SpecDecodeMetadata.make_dummy(
5120
5121
                draft_token_ids, self.device
            )
5122
5123
5124
5125
5126
5127

            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
5128
5129
5130
5131
5132
            logits = torch.randn(
                num_tokens + num_reqs,
                logits.shape[-1],
                device=self.device,
                dtype=logits.dtype,
5133
            )
5134
5135
5136
            self.rejection_sampler(
                dummy_spec_decode_metadata,
                draft_probs,
5137
                logits,
5138
5139
                dummy_metadata,
            )
5140
        return sampler_output
5141

5142
    def _dummy_pooler_run_task(
5143
5144
        self,
        hidden_states: torch.Tensor,
5145
5146
        task: PoolingTask,
    ) -> PoolerOutput:
5147
5148
5149
5150
        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
5151
5152
5153
5154
        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
5155
5156
5157

        req_num_tokens = num_tokens // num_reqs

5158
        dummy_prompt_lens = torch.from_numpy(num_scheduled_tokens_np)
5159
5160
5161
        dummy_token_ids = torch.zeros(
            (num_reqs, req_num_tokens), dtype=torch.int32, device=self.device
        )
5162

5163
        model = cast(VllmModelForPooling, self.get_model())
5164
        dummy_pooling_params = PoolingParams(task=task)
5165
        dummy_pooling_params.verify(self.model_config)
5166
        to_update = model.pooler.get_pooling_updates(task)
5167
5168
        to_update.apply(dummy_pooling_params)

5169
        dummy_metadata = PoolingMetadata(
5170
5171
5172
            prompt_lens=dummy_prompt_lens,
            prompt_token_ids=dummy_token_ids,
            pooling_params=[dummy_pooling_params] * num_reqs,
5173
            pooling_states=[PoolingStates() for i in range(num_reqs)],
5174
        )
5175

5176
        dummy_metadata.build_pooling_cursor(
5177
            num_scheduled_tokens_np,
5178
5179
            seq_lens_cpu=dummy_prompt_lens,
            device=hidden_states.device,
5180
        )
5181

5182
        try:
5183
5184
5185
            return model.pooler(
                hidden_states=hidden_states, pooling_metadata=dummy_metadata
            )
5186
        except RuntimeError as e:
5187
            if "out of memory" in str(e):
5188
                raise RuntimeError(
5189
5190
5191
                    "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 "
5192
5193
                    "initializing the engine."
                ) from e
5194
5195
            else:
                raise e
5196
5197
5198
5199
5200
5201

    @torch.inference_mode()
    def _dummy_pooler_run(
        self,
        hidden_states: torch.Tensor,
    ) -> PoolerOutput:
5202
5203
        mm_config = self.vllm_config.model_config.multimodal_config
        if mm_config and mm_config.mm_encoder_only:
5204
5205
5206
            # MM Encoder only model not need to run pooler.
            return torch.tensor([])

5207
        # Find the task that has the largest output for subsequent steps
5208
5209
5210
        supported_pooling_tasks = self.get_supported_pooling_tasks()

        if not supported_pooling_tasks:
5211
5212
5213
5214
5215
5216
            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."
            )
5217

5218
        output_size = dict[PoolingTask, float]()
5219
        for task in supported_pooling_tasks:
5220
5221
            # Run a full batch with each task to ensure none of them OOMs
            output = self._dummy_pooler_run_task(hidden_states, task)
5222
            output_size[task] = sum(o.nbytes for o in output if o is not None)
5223
5224
5225
5226
            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)
5227

5228
    def profile_run(self) -> None:
5229
        # Profile with multimodal encoder & encoder cache.
5230
        if self.supports_mm_inputs:
5231
5232
            mm_config = self.model_config.multimodal_config
            if mm_config is not None and mm_config.skip_mm_profiling:
5233
                logger.info(
5234
                    "Skipping memory profiling for multimodal encoder and "
5235
5236
                    "encoder cache."
                )
5237
5238
5239
5240
5241
            else:
                mm_budget = self.mm_budget
                assert mm_budget is not None

                if (encoder_budget := mm_budget.get_encoder_budget()) > 0:
5242
5243
5244
5245
5246
5247
5248
5249
5250
5251
5252
5253
5254
5255
5256
5257
5258
                    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
                        ]
5259

5260
5261
5262
5263
5264
5265
5266
5267
                        logger.info(
                            "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,
                        )
5268

5269
5270
5271
5272
5273
                        # Create dummy batch of multimodal inputs.
                        batched_dummy_mm_inputs = self._get_mm_dummy_batch(
                            dummy_modality,
                            max_mm_items_per_batch,
                        )
5274

5275
5276
5277
5278
5279
5280
5281
5282
5283
5284
5285
                        # 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
5286

5287
        # Add `is_profile` here to pre-allocate communication buffers
5288
5289
5290
        hidden_states, last_hidden_states = self._dummy_run(
            self.max_num_tokens, is_profile=True
        )
5291
        if get_pp_group().is_last_rank:
5292
5293
5294
5295
            if self.is_pooling_model:
                output = self._dummy_pooler_run(hidden_states)
            else:
                output = self._dummy_sampler_run(last_hidden_states)
5296
        else:
5297
            output = None
5298
        self._sync_device()
5299
        del hidden_states, output
5300
        self.encoder_cache.clear()
5301
        gc.collect()
5302

5303
    @instrument(span_name="Capture model")
5304
    def capture_model(self) -> int:
5305
        if self.compilation_config.cudagraph_mode == CUDAGraphMode.NONE:
5306
            logger.warning(
5307
                "Skipping CUDA graph capture. To turn on CUDA graph capture, "
5308
5309
                "ensure `cudagraph_mode` was not manually set to `NONE`"
            )
5310
            return 0
5311

5312
5313
        compilation_counter.num_gpu_runner_capture_triggers += 1

5314
5315
        start_time = time.perf_counter()

5316
5317
5318
5319
5320
5321
5322
5323
5324
5325
5326
5327
5328
5329
        @contextmanager
        def freeze_gc():
            # Optimize garbage collection during CUDA graph capture.
            # Clean up, then freeze all remaining objects from being included
            # in future collections.
            gc.collect()
            should_freeze = not envs.VLLM_ENABLE_CUDAGRAPH_GC
            if should_freeze:
                gc.freeze()
            try:
                yield
            finally:
                if should_freeze:
                    gc.unfreeze()
5330
                    gc.collect()
5331

5332
5333
5334
        # 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.
5335
        set_cudagraph_capturing_enabled(True)
5336
        with freeze_gc(), graph_capture(device=self.device):
5337
            start_free_gpu_memory = torch.cuda.mem_get_info()[0]
5338

5339
5340
5341
5342
            for (
                runtime_mode,
                batch_descs,
            ) in self.cudagraph_dispatcher.get_capture_descs():
5343
                self._capture_cudagraphs(
5344
5345
                    batch_descriptors=batch_descs,
                    cudagraph_runtime_mode=runtime_mode,
5346
                )
5347

5348
            torch.accelerator.synchronize()
5349
5350
            end_free_gpu_memory = torch.cuda.mem_get_info()[0]

5351
5352
5353
        # 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
5354
        # we may do lazy capturing in future that still allows capturing
5355
5356
        # after here.
        set_cudagraph_capturing_enabled(False)
5357

5358
5359
5360
5361
        # Lock workspace to prevent resizing during execution.
        # Max workspace sizes should have been captured during warmup/profiling.
        lock_workspace()

5362
5363
5364
5365
        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.
5366
        logger.info_once(
5367
5368
5369
            "Graph capturing finished in %.0f secs, took %.2f GiB",
            elapsed_time,
            cuda_graph_size / (1 << 30),
5370
            scope="local",
5371
        )
5372
        return cuda_graph_size
5373

5374
5375
    def _capture_cudagraphs(
        self,
5376
        batch_descriptors: list[BatchDescriptor],
5377
5378
5379
5380
        cudagraph_runtime_mode: CUDAGraphMode,
    ):
        assert (
            cudagraph_runtime_mode != CUDAGraphMode.NONE
5381
            and cudagraph_runtime_mode.is_valid_runtime_mode()
5382
        ), f"Invalid cudagraph runtime mode: {cudagraph_runtime_mode}"
5383

5384
5385
5386
5387
5388
5389
5390
5391
5392
5393
5394
5395
5396
5397
        if not batch_descriptors:
            return

        uniform_decode = batch_descriptors[0].uniform
        force_attention = cudagraph_runtime_mode == CUDAGraphMode.FULL

        dummy_run = functools.partial(
            self._dummy_run,
            uniform_decode=uniform_decode,
            skip_eplb=True,
            remove_lora=False,
            force_attention=force_attention,
        )

5398
5399
        # Only rank 0 should print progress bar during capture
        if is_global_first_rank():
5400
5401
            batch_descriptors = tqdm(
                batch_descriptors,
5402
5403
5404
                disable=not self.load_config.use_tqdm_on_load,
                desc="Capturing CUDA graphs ({}, {})".format(
                    "decode" if uniform_decode else "mixed prefill-decode",
5405
5406
5407
                    cudagraph_runtime_mode.name,
                ),
            )
5408

5409
        # We skip EPLB here since we don't want to record dummy metrics
5410
5411
        for batch_desc in batch_descriptors:
            num_tokens = batch_desc.num_tokens
5412
            num_active_loras = batch_desc.num_active_loras
5413

5414
            # We currently only capture ubatched graphs when its a FULL
5415
5416
5417
            # 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
5418
            allow_microbatching = (
5419
                self.parallel_config.use_ubatching
5420
5421
                and cudagraph_runtime_mode == CUDAGraphMode.FULL
                and uniform_decode
5422
5423
5424
5425
5426
                and check_ubatch_thresholds(
                    config=self.vllm_config.parallel_config,
                    num_tokens=num_tokens,
                    uniform_decode=uniform_decode,
                )
5427
            )
5428

5429
5430
            for _ in range(self.compilation_config.cudagraph_num_of_warmups):
                # Use CUDAGraphRuntimeStyle.NONE (default) for warmup.
5431
                # But be careful, warm up with `NONE` is orthogonal to
5432
5433
5434
                # if we want to warm up attention or not. This is
                # different from the case where `FULL` implies capture
                # attention while `PIECEWISE` implies no attention.
5435

5436
                dummy_run(
5437
5438
5439
                    num_tokens,
                    cudagraph_runtime_mode=CUDAGraphMode.NONE,
                    allow_microbatching=allow_microbatching,
5440
                    num_active_loras=num_active_loras,
5441
                )
5442
5443
5444

            # Capture run
            dummy_run(
5445
5446
5447
                num_tokens,
                cudagraph_runtime_mode=cudagraph_runtime_mode,
                allow_microbatching=allow_microbatching,
5448
                num_active_loras=num_active_loras,
Rémi Delacourt's avatar
Rémi Delacourt committed
5449
                is_graph_capturing=True,
5450
            )
5451
        self.maybe_remove_all_loras(self.lora_config)
5452

5453
5454
5455
5456
    def initialize_attn_backend(self, kv_cache_config: KVCacheConfig) -> None:
        """
        Initialize the attention backends and attention metadata builders.
        """
5457
        assert len(self.attn_groups) == 0, "Attention backends are already initialized"
5458

5459
5460
5461
5462
5463
5464
        class AttentionGroupKey(NamedTuple):
            attn_backend: type[AttentionBackend]
            kv_cache_spec: KVCacheSpec

        def get_attn_backends_for_group(
            kv_cache_group_spec: KVCacheGroupSpec,
5465
        ) -> tuple[dict[AttentionGroupKey, list[str]], set[type[AttentionBackend]]]:
5466
            layer_type = cast(type[Any], AttentionLayerBase)
5467
            layers = get_layers_from_vllm_config(
5468
                self.vllm_config, layer_type, kv_cache_group_spec.layer_names
5469
            )
5470
5471
            attn_backends = {}
            attn_backend_layers = defaultdict(list)
5472
            # Dedupe based on full class name; this is a bit safer than
5473
5474
5475
5476
            # 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.
5477
            for layer_name in kv_cache_group_spec.layer_names:
5478
                attn_backend = layers[layer_name].get_attn_backend()
5479
5480
5481
5482

                if layer_name in self.kv_sharing_fast_prefill_eligible_layers:
                    attn_backend = create_fast_prefill_custom_backend(
                        "FastPrefill",
5483
                        attn_backend,  # type: ignore[arg-type]
5484
5485
                    )

5486
5487
5488
                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):
5489
                    layer_kv_cache_spec = layer_kv_cache_spec.kv_cache_specs[layer_name]
5490
                key = (full_cls_name, layer_kv_cache_spec)
5491
5492
5493
                attn_backends[key] = AttentionGroupKey(
                    attn_backend, layer_kv_cache_spec
                )
5494
                attn_backend_layers[key].append(layer_name)
5495
5496
5497
5498
            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()),
            )
5499
5500

        def create_attn_groups(
5501
            attn_backends_map: dict[AttentionGroupKey, list[str]],
5502
            kv_cache_group_id: int,
5503
5504
        ) -> list[AttentionGroup]:
            attn_groups: list[AttentionGroup] = []
5505
            for (attn_backend, kv_cache_spec), layer_names in attn_backends_map.items():
5506
                attn_group = AttentionGroup(
5507
                    attn_backend,
5508
                    layer_names,
5509
                    kv_cache_spec,
5510
                    kv_cache_group_id,
5511
5512
                )

5513
5514
5515
                attn_groups.append(attn_group)
            return attn_groups

5516
        attention_backend_maps = []
5517
        attention_backend_list = []
5518
        for kv_cache_group_spec in kv_cache_config.kv_cache_groups:
5519
            attn_backends = get_attn_backends_for_group(kv_cache_group_spec)
5520
            attention_backend_maps.append(attn_backends[0])
5521
            attention_backend_list.append(attn_backends[1])
5522
5523

        # Resolve cudagraph_mode before actually initialize metadata_builders
5524
5525
5526
        self._check_and_update_cudagraph_mode(
            attention_backend_list, kv_cache_config.kv_cache_groups
        )
5527

5528
5529
5530
        # Check if attention backend supports PCP&DCP and related features.
        check_attention_cp_compatibility(self.vllm_config)

5531
5532
        for i, attn_backend_map in enumerate(attention_backend_maps):
            self.attn_groups.append(create_attn_groups(attn_backend_map, i))
5533

5534
5535
5536
5537
5538
5539
5540
5541
5542
5543
5544
5545
5546
5547
5548
    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
5549
5550
                    if not self.parallel_config.use_ubatching
                    else self.parallel_config.num_ubatches,
5551
                )
co63oc's avatar
co63oc committed
5552
        # Calculate reorder batch threshold (if needed)
5553
5554
        # Note (tdoublep): do this *after* constructing builders,
        # because some of them change the threshold at init time.
5555
5556
        self.calculate_reorder_batch_threshold()

5557
5558
5559
5560
5561
5562
5563
5564
        # Initialize drafter attention backend
        if self.speculative_config and (
            self.speculative_config.use_eagle()
            or self.speculative_config.uses_draft_model()
        ):
            assert isinstance(self.drafter, EagleProposer | DraftModelProposer)
            self.drafter.initialize_attn_backend(kv_cache_config, kernel_block_sizes)

5565
    def _check_and_update_cudagraph_mode(
5566
5567
5568
        self,
        attention_backends: list[set[type[AttentionBackend]]],
        kv_cache_groups: list[KVCacheGroupSpec],
5569
    ) -> None:
5570
        """
5571
        Resolve the cudagraph_mode when there are multiple attention
5572
        groups with potential conflicting CUDA graph support.
5573
5574
5575
        Then initialize the cudagraph_dispatcher based on the resolved
        cudagraph_mode.
        """
5576
        min_cg_support = AttentionCGSupport.ALWAYS
5577
        min_cg_backend_name = None
5578

5579
5580
5581
5582
5583
5584
5585
5586
5587
5588
5589
5590
        for attn_backend_set, kv_cache_group in zip(
            attention_backends, kv_cache_groups
        ):
            for attn_backend in attn_backend_set:
                builder_cls = attn_backend.get_builder_cls()

                cg_support = builder_cls.get_cudagraph_support(
                    self.vllm_config, kv_cache_group.kv_cache_spec
                )
                if cg_support.value < min_cg_support.value:
                    min_cg_support = cg_support
                    min_cg_backend_name = attn_backend.__name__
5591
5592
        # Flexible resolve the cudagraph mode
        cudagraph_mode = self.compilation_config.cudagraph_mode
5593
        assert cudagraph_mode is not None
5594
        # check cudagraph for mixed batch is supported
5595
5596
5597
5598
5599
5600
        if (
            cudagraph_mode.mixed_mode() == CUDAGraphMode.FULL
            and min_cg_support != AttentionCGSupport.ALWAYS
        ):
            msg = (
                f"CUDAGraphMode.{cudagraph_mode.name} is not supported "
5601
                f"with {min_cg_backend_name} backend (support: "
5602
5603
                f"{min_cg_support})"
            )
5604
5605
            if min_cg_support == AttentionCGSupport.NEVER:
                # if not supported any full cudagraphs, just raise it.
5606
5607
                msg += (
                    "; please try cudagraph_mode=PIECEWISE, and "
5608
                    "make sure compilation mode is VLLM_COMPILE"
5609
                )
5610
5611
5612
5613
5614
                raise ValueError(msg)

            # attempt to resolve the full cudagraph related mode
            if self.compilation_config.splitting_ops_contain_attention():
                msg += "; setting cudagraph_mode=FULL_AND_PIECEWISE"
5615
                cudagraph_mode = self.compilation_config.cudagraph_mode = (
5616
                    CUDAGraphMode.FULL_AND_PIECEWISE
5617
                )
5618
5619
            else:
                msg += "; setting cudagraph_mode=FULL_DECODE_ONLY"
5620
                cudagraph_mode = self.compilation_config.cudagraph_mode = (
5621
                    CUDAGraphMode.FULL_DECODE_ONLY
5622
                )
5623
5624
            logger.warning(msg)

5625
        # check that if we are doing decode full-cudagraphs it is supported
5626
5627
5628
5629
5630
5631
        if (
            cudagraph_mode.decode_mode() == CUDAGraphMode.FULL
            and min_cg_support == AttentionCGSupport.NEVER
        ):
            msg = (
                f"CUDAGraphMode.{cudagraph_mode.name} is not supported "
5632
                f"with {min_cg_backend_name} backend (support: "
5633
5634
                f"{min_cg_support})"
            )
5635
            if self.compilation_config.mode == CompilationMode.VLLM_COMPILE and (
5636
5637
5638
5639
5640
                self.compilation_config.splitting_ops_contain_attention()
                or self.compilation_config.use_inductor_graph_partition
            ):
                msg += (
                    "; setting cudagraph_mode=PIECEWISE because "
5641
                    "attention is compiled piecewise"
5642
5643
                )
                cudagraph_mode = self.compilation_config.cudagraph_mode = (
5644
                    CUDAGraphMode.PIECEWISE
5645
                )
5646
            else:
5647
5648
                msg += (
                    "; setting cudagraph_mode=NONE because "
5649
                    "attention is not compiled piecewise"
5650
5651
                )
                cudagraph_mode = self.compilation_config.cudagraph_mode = (
5652
                    CUDAGraphMode.NONE
5653
                )
5654
5655
            logger.warning(msg)

5656
5657
        # check that if we are doing spec-decode + decode full-cudagraphs it is
        # supported
5658
5659
5660
5661
5662
5663
5664
5665
        if (
            cudagraph_mode.decode_mode() == CUDAGraphMode.FULL
            and self.uniform_decode_query_len > 1
            and min_cg_support.value < AttentionCGSupport.UNIFORM_BATCH.value
        ):
            msg = (
                f"CUDAGraphMode.{cudagraph_mode.name} is not supported"
                f" with spec-decode for attention backend "
5666
                f"{min_cg_backend_name} (support: {min_cg_support})"
5667
            )
5668
5669
            if self.compilation_config.splitting_ops_contain_attention():
                msg += "; setting cudagraph_mode=PIECEWISE"
5670
                cudagraph_mode = self.compilation_config.cudagraph_mode = (
5671
                    CUDAGraphMode.PIECEWISE
5672
                )
5673
5674
            else:
                msg += "; setting cudagraph_mode=NONE"
5675
                cudagraph_mode = self.compilation_config.cudagraph_mode = (
5676
                    CUDAGraphMode.NONE
5677
                )
5678
5679
5680
5681
            logger.warning(msg)

        # double check that we can support full cudagraph if they are requested
        # even after automatic downgrades
5682
5683
5684
5685
5686
5687
        if (
            cudagraph_mode.has_full_cudagraphs()
            and min_cg_support == AttentionCGSupport.NEVER
        ):
            raise ValueError(
                f"CUDAGraphMode.{cudagraph_mode.name} is not "
5688
                f"supported with {min_cg_backend_name} backend ("
5689
5690
                f"support:{min_cg_support}) "
                "; please try cudagraph_mode=PIECEWISE, "
5691
                "and make sure compilation mode is VLLM_COMPILE"
5692
            )
5693

5694
5695
5696
5697
        # if we have dedicated decode cudagraphs, and spec-decode is enabled,
        # we need to adjust the cudagraph sizes to be a multiple of the uniform
        # decode query length to avoid: https://github.com/vllm-project/vllm/issues/28207
        # temp-fix: https://github.com/vllm-project/vllm/issues/28207#issuecomment-3504004536
5698
        # Will be removed in the near future when we have separate cudagraph capture
5699
5700
5701
5702
5703
5704
5705
5706
5707
5708
        # sizes for decode and mixed prefill-decode.
        if (
            cudagraph_mode.decode_mode() == CUDAGraphMode.FULL
            and cudagraph_mode.separate_routine()
            and self.uniform_decode_query_len > 1
        ):
            self.compilation_config.adjust_cudagraph_sizes_for_spec_decode(
                self.uniform_decode_query_len, self.parallel_config.tensor_parallel_size
            )

5709
5710
5711
5712
5713
5714
5715
5716
5717
5718
5719
5720
5721
5722
5723
5724
        # If the model has Mamba layers and cudagraph mode includes FULL
        # decode, cap cudagraph capture sizes to the number of available
        # Mamba cache blocks. Each decode request needs one conv_state
        # cache line, so capture batch sizes cannot exceed num_blocks.
        # Only FULL decode graphs are affected because PIECEWISE captures
        # run GDN/Mamba ops eagerly (prefill path, no causal_conv1d_update).
        # See: https://github.com/vllm-project/vllm/issues/34094
        if cudagraph_mode.has_full_cudagraphs():
            has_mamba = any(
                isinstance(g.kv_cache_spec, MambaSpec) for g in kv_cache_groups
            )
            if has_mamba and self.kv_cache_config is not None:
                self.compilation_config.adjust_cudagraph_sizes_for_mamba_cache(
                    self.kv_cache_config.num_blocks
                )

5725
5726
        # Trigger cudagraph dispatching keys initialization after
        # resolved cudagraph mode.
5727
        self.compilation_config.cudagraph_mode = cudagraph_mode
5728
        self.cudagraph_dispatcher.initialize_cudagraph_keys(
5729
            cudagraph_mode, self.uniform_decode_query_len
5730
        )
5731

5732
5733
5734
5735
5736
5737
        # Initialize drafter's cudagraph dispatcher if using spec decode.
        if self.speculative_config and (
            self.speculative_config.use_eagle()
            or self.speculative_config.uses_extract_hidden_states()
        ):
            assert isinstance(self.drafter, EagleProposer | ExtractHiddenStatesProposer)
5738
5739
            self.drafter.initialize_cudagraph_keys(cudagraph_mode)

5740
5741
    def calculate_reorder_batch_threshold(self) -> None:
        """
5742
5743
5744
5745
        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.
5746
        """
5747
5748
        min_none_high = lambda a, b: a if b is None else b if a is None else min(a, b)

5749
        reorder_batch_thresholds: list[int | None] = [
5750
5751
5752
            group.get_metadata_builder().reorder_batch_threshold
            for group in self._attn_group_iterator()
        ]
5753
5754
5755
5756
5757
        # 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
5758
        self.reorder_batch_threshold = reduce(min_none_high, reorder_batch_thresholds)  # type: ignore[assignment]
5759

5760
5761
5762
    def may_reinitialize_input_batch(
        self, kv_cache_config: KVCacheConfig, kernel_block_sizes: list[int]
    ) -> None:
5763
5764
5765
5766
5767
5768
5769
        """
        Re-initialize the input batch if the block sizes are different from
        `[self.cache_config.block_size]`. This usually happens when there
        are multiple KV cache groups.

        Args:
            kv_cache_config: The KV cache configuration.
5770
            kernel_block_sizes: The kernel block sizes for each KV cache group.
5771
        """
5772
        block_sizes = []
5773
5774
        max_num_blocks = []
        max_model_len = max(self.max_model_len, self.max_encoder_len)
5775
        for kv_cache_group in kv_cache_config.kv_cache_groups:
5776
5777
            if isinstance(kv_cache_group.kv_cache_spec, EncoderOnlyAttentionSpec):
                continue
5778
5779
            block_size = kv_cache_group.kv_cache_spec.block_size
            block_sizes.append(block_size)
5780
            max_num_blocks_per_req = cdiv(
5781
                max_model_len, block_size * get_total_cp_world_size()
5782
5783
            )
            if isinstance(kv_cache_group.kv_cache_spec, MambaSpec):
5784
                max_num_blocks_per_req = (
5785
5786
5787
5788
5789
                    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)
5790
5791
5792
5793

        if block_sizes != [self.cache_config.block_size] or kernel_block_sizes != [
            self.cache_config.block_size
        ]:
5794
            assert self.offload_config.uva.cpu_offload_gb == 0, (
5795
5796
                "Cannot re-initialize the input batch when CPU weight "
                "offloading is enabled. See https://github.com/vllm-project/vllm/pull/18298 "  # noqa: E501
5797
5798
                "for more details."
            )
5799
5800
            self.input_batch = InputBatch(
                max_num_reqs=self.max_num_reqs,
5801
                max_model_len=max_model_len,
5802
5803
5804
5805
5806
                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,
5807
                kernel_block_sizes=kernel_block_sizes,
5808
                max_num_blocks_per_req=max_num_blocks,
5809
                is_spec_decode=bool(self.vllm_config.speculative_config),
5810
                logitsprocs=self.input_batch.logitsprocs,
5811
                logitsprocs_need_output_token_ids=self.input_batch.logitsprocs_need_output_token_ids,
5812
                is_pooling_model=self.is_pooling_model,
5813
5814
            )

5815
    def _allocate_kv_cache_tensors(
5816
5817
        self, kv_cache_config: KVCacheConfig
    ) -> dict[str, torch.Tensor]:
5818
        """
5819
5820
5821
        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.

5822
        Args:
5823
            kv_cache_config: The KV cache config
5824
        Returns:
5825
            dict[str, torch.Tensor]: A map between layer names to their
5826
            corresponding memory buffer for KV cache.
5827
        """
5828
5829
        kv_cache_raw_tensors: dict[str, torch.Tensor] = {}
        for kv_cache_tensor in kv_cache_config.kv_cache_tensors:
5830
5831
5832
            tensor = torch.zeros(
                kv_cache_tensor.size, dtype=torch.int8, device=self.device
            )
5833
5834
5835
5836
5837
            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:
5838
5839
5840
5841
            for layer_name in group.layer_names:
                if layer_name in self.runner_only_attn_layers:
                    continue
                layer_names.add(layer_name)
5842
5843
5844
        assert layer_names == set(kv_cache_raw_tensors.keys()), (
            "Some layers are not correctly initialized"
        )
5845
5846
        return kv_cache_raw_tensors

5847
5848
5849
    def _attn_group_iterator(self) -> Iterator[AttentionGroup]:
        return itertools.chain.from_iterable(self.attn_groups)

5850
    def _kv_cache_spec_attn_group_iterator(self) -> Iterator[AttentionGroup]:
5851
5852
        if not self.kv_cache_config.kv_cache_groups:
            return
5853
5854
        for attn_groups in self.attn_groups:
            yield from attn_groups
5855

5856
5857
5858
5859
    def _reshape_kv_cache_tensors(
        self,
        kv_cache_config: KVCacheConfig,
        kv_cache_raw_tensors: dict[str, torch.Tensor],
5860
        kernel_block_sizes: list[int],
5861
    ) -> dict[str, torch.Tensor]:
5862
        """
5863
        Reshape the KV cache tensors to the desired shape and dtype.
5864

5865
        Args:
5866
5867
            kv_cache_config: The KV cache config
            kv_cache_raw_tensors: The KV cache buffer of each layer, with
5868
                correct size but uninitialized shape.
5869
            kernel_block_sizes: The kernel block sizes for each KV cache group.
5870
        Returns:
5871
            Dict[str, torch.Tensor]: A map between layer names to their
5872
5873
            corresponding memory buffer for KV cache.
        """
5874
        kv_caches: dict[str, torch.Tensor] = {}
5875
        has_attn, has_mamba = False, False
5876
5877
        for group in self._kv_cache_spec_attn_group_iterator():
            kv_cache_spec = group.kv_cache_spec
5878
            attn_backend = group.backend
5879
5880
5881
5882
            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]
5883
            for layer_name in group.layer_names:
5884
5885
                if layer_name in self.runner_only_attn_layers:
                    continue
5886
5887
                raw_tensor = kv_cache_raw_tensors[layer_name]
                assert raw_tensor.numel() % kv_cache_spec.page_size_bytes == 0
5888
                num_blocks = raw_tensor.numel() // kv_cache_spec.page_size_bytes
5889
                if isinstance(kv_cache_spec, AttentionSpec):
5890
                    has_attn = True
5891
5892
                    num_blocks_per_kv_block = (
                        kv_cache_spec.block_size // kernel_block_size
5893
5894
5895
                    )
                    kernel_num_blocks = num_blocks * num_blocks_per_kv_block

5896
                    kv_cache_shape = attn_backend.get_kv_cache_shape(
5897
                        kernel_num_blocks,
5898
                        kernel_block_size,
5899
5900
                        kv_cache_spec.num_kv_heads,
                        kv_cache_spec.head_size,
5901
5902
                        cache_dtype_str=self.cache_config.cache_dtype,
                    )
5903
                    dtype = kv_cache_spec.dtype
5904
                    try:
5905
                        kv_cache_stride_order = attn_backend.get_kv_cache_stride_order()
5906
                        assert len(kv_cache_stride_order) == len(kv_cache_shape)
5907
                    except (AttributeError, NotImplementedError):
5908
                        kv_cache_stride_order = tuple(range(len(kv_cache_shape)))
5909
5910
5911
5912
5913
                    # 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.
5914
5915
5916
                    kv_cache_shape = tuple(
                        kv_cache_shape[i] for i in kv_cache_stride_order
                    )
5917
5918
5919
5920
5921
                    # Maintain original KV shape view.
                    inv_order = [
                        kv_cache_stride_order.index(i)
                        for i in range(len(kv_cache_stride_order))
                    ]
5922
5923
5924
5925
5926
5927
                    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
5928
                elif isinstance(kv_cache_spec, MambaSpec):
5929
                    has_mamba = True
Chen Zhang's avatar
Chen Zhang committed
5930
5931
                    raw_tensor = kv_cache_raw_tensors[layer_name]
                    state_tensors = []
5932
                    storage_offset_bytes = 0
5933
                    for shape, dtype in zip(kv_cache_spec.shapes, kv_cache_spec.dtypes):
5934
5935
                        dtype_size = get_dtype_size(dtype)
                        num_element_per_page = (
5936
5937
                            kv_cache_spec.page_size_bytes // dtype_size
                        )
Chen Zhang's avatar
Chen Zhang committed
5938
                        target_shape = (num_blocks, *shape)
5939
5940
                        stride = torch.empty(target_shape).stride()
                        target_stride = (num_element_per_page, *stride[1:])
5941
                        assert storage_offset_bytes % dtype_size == 0
5942
5943
5944
5945
                        tensor = torch.as_strided(
                            raw_tensor.view(dtype),
                            size=target_shape,
                            stride=target_stride,
5946
                            storage_offset=storage_offset_bytes // dtype_size,
5947
                        )
Chen Zhang's avatar
Chen Zhang committed
5948
                        state_tensors.append(tensor)
5949
                        storage_offset_bytes += stride[0] * dtype_size
5950
5951

                    kv_caches[layer_name] = state_tensors
5952
                else:
5953
                    raise NotImplementedError
5954
5955

        if has_attn and has_mamba:
5956
            self._update_hybrid_attention_mamba_layout(kv_caches)
5957

5958
5959
        return kv_caches

5960
    def _update_hybrid_attention_mamba_layout(
5961
5962
        self, kv_caches: dict[str, torch.Tensor]
    ) -> None:
5963
        """
5964
5965
        Update the layout of attention layers from (2, num_blocks, ...) to
        (num_blocks, 2, ...).
5966
5967

        Args:
5968
            kv_caches: The KV cache buffer of each layer.
5969
5970
        """

5971
5972
        for group in self._kv_cache_spec_attn_group_iterator():
            kv_cache_spec = group.kv_cache_spec
5973
            for layer_name in group.layer_names:
5974
                kv_cache = kv_caches[layer_name]
5975
5976
5977
5978
                if isinstance(kv_cache_spec, AttentionSpec) and kv_cache.shape[0] == 2:
                    assert kv_cache.shape[1] != 2, (
                        "Fail to determine whether the layout is "
                        "(2, num_blocks, ...) or (num_blocks, 2, ...) for "
5979
                        f"a tensor of shape {kv_cache.shape}"
5980
                    )
5981
                    hidden_size = kv_cache.shape[2:].numel()
5982
5983
5984
5985
                    kv_cache.as_strided_(
                        size=kv_cache.shape,
                        stride=(hidden_size, 2 * hidden_size, *kv_cache.stride()[2:]),
                    )
5986

5987
    def initialize_kv_cache_tensors(
5988
        self, kv_cache_config: KVCacheConfig, kernel_block_sizes: list[int]
5989
    ) -> dict[str, torch.Tensor]:
5990
5991
5992
5993
5994
        """
        Initialize the memory buffer for KV cache.

        Args:
            kv_cache_config: The KV cache config
5995
5996
            kernel_block_sizes: The kernel block sizes for each KV cache group.

5997
        Returns:
5998
            Dict[str, torch.Tensor]: A map between layer names to their
5999
6000
            corresponding memory buffer for KV cache.
        """
6001
6002
6003
6004
6005
6006
6007
6008
6009
6010
6011
6012
6013
6014
6015
6016
6017
6018
6019
6020
6021
6022
6023
6024

        # 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
            )
6025

6026
        # Set up cross-layer KV cache sharing
6027
6028
        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)
6029
6030
            kv_caches[layer_name] = kv_caches[target_layer_name]

6031
6032
6033
6034
6035
6036
6037
6038
6039
        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,
        )
6040
6041
6042
        return kv_caches

    def maybe_add_kv_sharing_layers_to_kv_cache_groups(
6043
6044
        self, kv_cache_config: KVCacheConfig
    ) -> None:
6045
6046
6047
6048
6049
6050
6051
6052
6053
6054
6055
6056
6057
6058
6059
6060
6061
6062
        """
        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.
6063
            attn_layers = get_layers_from_vllm_config(self.vllm_config, Attention)
6064
6065
            for layer_name in reversed(attn_layers):
                if layer_name in self.shared_kv_cache_layers:
6066
                    self.kv_sharing_fast_prefill_eligible_layers.add(layer_name)
6067
6068
                else:
                    break
6069

6070
6071
6072
6073
6074
6075
6076
    def initialize_kv_cache(self, kv_cache_config: KVCacheConfig) -> None:
        """
        Initialize KV cache based on `kv_cache_config`.
        Args:
            kv_cache_config: Configuration for the KV cache, including the KV
            cache size of each layer
        """
6077
        kv_cache_config = deepcopy(kv_cache_config)
6078
        self.kv_cache_config = kv_cache_config
6079
        self._mamba_copy_bufs = None
6080
        self.may_add_encoder_only_layers_to_kv_cache_config()
6081
        self.maybe_add_kv_sharing_layers_to_kv_cache_groups(kv_cache_config)
6082
        self.initialize_attn_backend(kv_cache_config)
6083
6084
6085
6086
6087
        # 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.
6088
6089
6090
        kernel_block_sizes = prepare_kernel_block_sizes(
            kv_cache_config, self.attn_groups
        )
6091
6092
6093
6094

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

6095
        # Reinitialize need to after initialize_attn_backend
6096
6097
6098
6099
        self.may_reinitialize_input_batch(kv_cache_config, kernel_block_sizes)
        kv_caches = self.initialize_kv_cache_tensors(
            kv_cache_config, kernel_block_sizes
        )
6100

6101
6102
6103
        if (
            self.speculative_config
            and self.speculative_config.uses_extract_hidden_states()
6104
        ):
6105
            assert isinstance(self.drafter, ExtractHiddenStatesProposer)
6106
6107
6108
6109
            # validate all draft model layers belong to the same kv cache
            # group
            self.drafter.validate_same_kv_cache_group(kv_cache_config)

Robert Shaw's avatar
Robert Shaw committed
6110
        if has_kv_transfer_group():
6111
            kv_transfer_group = get_kv_transfer_group()
6112
6113
6114
6115
6116
6117
6118
            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)
6119
            kv_transfer_group.set_host_xfer_buffer_ops(copy_kv_blocks)
Robert Shaw's avatar
Robert Shaw committed
6120

6121
6122
6123
6124
6125
6126
6127
6128
6129
6130
6131
6132
6133
6134
6135
6136
6137
        if self.model_config.enable_return_routed_experts:
            self.init_routed_experts_capturer()

    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()
        block_size = self.cache_config.block_size
        self.max_num_kv_tokens = (
            self.kv_cache_config.num_blocks // len(self.kv_cache_config.kv_cache_groups)
            + 1
        ) * block_size
        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,
6138
            vllm_config=self.vllm_config,
6139
        )
6140
6141
6142
6143
6144
6145
6146
6147
6148
6149
6150
6151
6152
6153
6154
6155
        self._bind_routed_experts_capturer(routed_experts_capturer)

    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)
6156

6157
6158
6159
6160
6161
    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
6162
        encoder_only_attn_specs: dict[AttentionSpec, list[str]] = defaultdict(list)
6163
6164
6165
        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:
6166
                attn_spec: AttentionSpec = EncoderOnlyAttentionSpec(
6167
6168
6169
                    block_size=block_size,
                    num_kv_heads=attn_module.num_kv_heads,
                    head_size=attn_module.head_size,
6170
6171
                    dtype=self.kv_cache_dtype,
                )
6172
6173
6174
                encoder_only_attn_specs[attn_spec].append(layer_name)
                self.runner_only_attn_layers.add(layer_name)
        if len(encoder_only_attn_specs) > 0:
6175
6176
6177
            assert len(encoder_only_attn_specs) == 1, (
                "Only support one encoder-only attention spec now"
            )
6178
6179
            spec, layer_names = encoder_only_attn_specs.popitem()
            self.kv_cache_config.kv_cache_groups.append(
6180
6181
                KVCacheGroupSpec(layer_names=layer_names, kv_cache_spec=spec)
            )
6182

6183
    def get_kv_cache_spec(self) -> dict[str, KVCacheSpec]:
6184
        """
6185
        Generates the KVCacheSpec by parsing the kv cache format from each
6186
6187
        Attention module in the static forward context.
        Returns:
6188
            KVCacheSpec: A dictionary mapping layer names to their KV cache
6189
6190
            format. Layers that do not need KV cache are not included.
        """
6191
        if has_ec_transfer() and not get_ec_transfer().is_consumer:
6192
            return {}
6193
        kv_cache_spec: dict[str, KVCacheSpec] = {}
6194
6195
        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
6196
        for layer_name, attn_module in attn_layers.items():
6197
6198
6199
6200
6201
6202
6203
6204
6205
6206
6207
6208
6209
6210
6211
            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
6212

6213
        return kv_cache_spec
6214

6215
    def _to_list(self, sampled_token_ids: torch.Tensor) -> list[list[int]]:
6216
6217
6218
6219
6220
6221
6222
6223
        # 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.
6224
        pinned = self.sampled_token_ids_pinned_cpu[: sampled_token_ids.shape[0]]
6225
6226
6227
        pinned.copy_(sampled_token_ids, non_blocking=True)
        self.transfer_event.record()
        self.transfer_event.synchronize()
6228
        return pinned.tolist()
6229
6230
6231
6232
6233
6234
6235
6236
6237
6238
6239
6240
6241
6242
6243
6244
6245
6246
6247
6248
6249
6250
6251
6252
6253
6254
6255
6256
6257
6258
6259
6260
6261
6262
6263
6264
6265
6266
6267
6268

    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}

6269
        torch.accelerator.synchronize()
6270
6271
6272
6273
6274
        start_time = time.perf_counter()

        try:
            yield
        finally:
6275
            torch.accelerator.synchronize()
6276
6277
6278
6279
6280
6281
6282
6283
6284
6285
            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]
6286
                    stats.encoder_forward_secs += per_request_time
6287
6288
6289
6290
6291
6292
6293
                    stats.num_encoder_calls += 1


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

6294
    encoder_forward_secs: float = 0.0
6295
6296
6297
6298
6299
6300
6301
    """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 {
6302
            "encoder_forward_secs": self.encoder_forward_secs,
6303
6304
            "num_encoder_calls": self.num_encoder_calls,
        }