gpu_model_runner.py 259 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 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.attention.layer import Attention, MLAAttention
25
from vllm.compilation.counter import compilation_counter
26
from vllm.compilation.cuda_graph import CUDAGraphStat, CUDAGraphWrapper
27
from vllm.compilation.monitor import set_cudagraph_capturing_enabled
28
from vllm.config import (
29
    CompilationMode,
30
31
32
33
34
    CUDAGraphMode,
    VllmConfig,
    get_layers_from_vllm_config,
    update_config,
)
35
from vllm.distributed.ec_transfer import get_ec_transfer, has_ec_transfer
36
from vllm.distributed.eplb.eplb_state import EplbState
37
from vllm.distributed.kv_transfer import get_kv_transfer_group, has_kv_transfer_group
38
from vllm.distributed.kv_transfer.kv_connector.utils import copy_kv_blocks
39
from vllm.distributed.parallel_state import (
40
    get_dcp_group,
41
42
43
44
45
46
    get_pp_group,
    get_tp_group,
    graph_capture,
    is_global_first_rank,
    prepare_communication_buffer_for_model,
)
47
48
49
50
from vllm.forward_context import (
    BatchDescriptor,
    set_forward_context,
)
51
from vllm.logger import init_logger
52
from vllm.lora.layers import LoRAMapping, LoRAMappingType
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 TensorizerLoader, get_model_loader
62
from vllm.model_executor.models.interfaces import (
63
    MultiModalEmbeddings,
64
    SupportsMRoPE,
65
    SupportsMultiModal,
66
    SupportsXDRoPE,
67
68
69
70
71
    is_mixture_of_experts,
    supports_eagle3,
    supports_mrope,
    supports_multimodal_pruning,
    supports_transcription,
72
    supports_xdrope,
73
)
74
from vllm.model_executor.models.interfaces_base import (
75
76
77
78
    VllmModelForPooling,
    is_pooling_model,
    is_text_generation_model,
)
79
from vllm.multimodal import MULTIMODAL_REGISTRY
80
81
82
83
84
from vllm.multimodal.inputs import (
    BatchedTensorInputs,
    MultiModalKwargsItem,
    PlaceholderRange,
)
85
from vllm.multimodal.utils import group_mm_kwargs_by_modality
86
from vllm.pooling_params import PoolingParams
87
from vllm.sampling_params import SamplingType
88
from vllm.sequence import IntermediateTensors
89
from vllm.tasks import GenerationTask, PoolingTask, SupportedTask
90
from vllm.utils import length_from_prompt_token_ids_or_embeds
91
from vllm.utils.jsontree import json_map_leaves
92
from vllm.utils.math_utils import cdiv, round_up
93
from vllm.utils.mem_utils import DeviceMemoryProfiler, format_gib
94
from vllm.utils.nvtx_pytorch_hooks import PytHooks
95
from vllm.utils.platform_utils import is_pin_memory_available
96
97
98
99
from vllm.utils.torch_utils import (
    get_dtype_size,
    kv_cache_dtype_str_to_dtype,
)
100
101
from vllm.v1.attention.backend import (
    AttentionBackend,
102
    AttentionCGSupport,
103
    AttentionMetadata,
104
    AttentionMetadataBuilder,
105
    AttentionType,
106
    CommonAttentionMetadata,
107
108
    MultipleOf,
)
109
from vllm.v1.attention.backends.gdn_attn import GDNAttentionMetadataBuilder
110
from vllm.v1.attention.backends.utils import (
111
    create_fast_prefill_custom_backend,
112
    get_dcp_local_seq_lens,
113
114
    reorder_batch_to_split_decodes_and_prefills,
)
115
from vllm.v1.core.sched.output import NewRequestData
116
from vllm.v1.cudagraph_dispatcher import CudagraphDispatcher
117
118
119
120
121
122
123
124
125
126
127
128
129
130
131
132
133
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,
134
    ECConnectorOutput,
135
    KVConnectorOutput,
136
137
138
139
140
    LogprobsLists,
    LogprobsTensors,
    ModelRunnerOutput,
    PoolerOutput,
    SamplerOutput,
141
    make_empty_encoder_model_runner_output,
142
)
143
from vllm.v1.pool.metadata import PoolingMetadata, PoolingStates
144
from vllm.v1.sample.logits_processor import LogitsProcessors, build_logitsprocs
145
from vllm.v1.sample.logits_processor.interface import LogitsProcessor
146
from vllm.v1.sample.metadata import SamplingMetadata
147
from vllm.v1.sample.rejection_sampler import RejectionSampler
148
from vllm.v1.sample.sampler import Sampler
149
from vllm.v1.spec_decode.draft_model import DraftModelProposer
150
from vllm.v1.spec_decode.eagle import EagleProposer
151
from vllm.v1.spec_decode.medusa import MedusaProposer
152
from vllm.v1.spec_decode.metadata import SpecDecodeMetadata
153
from vllm.v1.spec_decode.ngram_proposer import NgramProposer
154
from vllm.v1.spec_decode.suffix_decoding import SuffixDecodingProposer
155
from vllm.v1.structured_output.utils import apply_grammar_bitmask
156
from vllm.v1.utils import CpuGpuBuffer, record_function_or_nullcontext
157
158
159
160
161
from vllm.v1.worker import mamba_utils
from vllm.v1.worker.cp_utils import (
    check_attention_cp_compatibility,
    get_total_cp_world_size,
)
162
from vllm.v1.worker.dp_utils import coordinate_batch_across_dp
163
from vllm.v1.worker.ec_connector_model_runner_mixin import ECConnectorModelRunnerMixin
164
from vllm.v1.worker.gpu_input_batch import CachedRequestState, InputBatch
165
from vllm.v1.worker.gpu_ubatch_wrapper import UBatchWrapper
166
from vllm.v1.worker.kv_connector_model_runner_mixin import KVConnectorModelRunnerMixin
167
from vllm.v1.worker.lora_model_runner_mixin import LoRAModelRunnerMixin
168
169
170
from vllm.v1.worker.ubatch_utils import (
    UBatchSlices,
    check_ubatch_thresholds,
171
    maybe_create_ubatch_slices,
172
    split_attn_metadata,
173
)
174
from vllm.v1.worker.utils import is_residual_scattered_for_sp
175
from vllm.v1.worker.workspace import lock_workspace
176

177
178
179
180
181
182
183
from .utils import (
    AttentionGroup,
    MultiModalBudget,
    add_kv_sharing_layers_to_kv_cache_groups,
    bind_kv_cache,
    sanity_check_mm_encoder_outputs,
)
184

185
if TYPE_CHECKING:
186
    from vllm.model_executor.model_loader.tensorizer import TensorizerConfig
187
    from vllm.v1.core.sched.output import GrammarOutput, SchedulerOutput
188
189
190

logger = init_logger(__name__)

191
192
AttnMetadataDict: TypeAlias = dict[str, AttentionMetadata]
# list when ubatching is enabled
193
PerLayerAttnMetadata: TypeAlias = list[AttnMetadataDict] | AttnMetadataDict
194

195

196
197
198
199
200
201
# Wrapper for ModelRunnerOutput to support overlapped execution.
class AsyncGPUModelRunnerOutput(AsyncModelRunnerOutput):
    def __init__(
        self,
        model_runner_output: ModelRunnerOutput,
        sampled_token_ids: torch.Tensor,
202
        logprobs_tensors: LogprobsTensors | None,
203
204
        invalid_req_indices: list[int],
        async_output_copy_stream: torch.cuda.Stream,
205
        vocab_size: int,
206
207
208
209
210
    ):
        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.
211
        self.async_copy_ready_event = torch.Event()
212
213
214
215

        # 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
216
        self.vocab_size = vocab_size
217
        self._logprobs_tensors = logprobs_tensors
218
219
220
221
222

        # 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)
223
            self.sampled_token_ids_cpu = self._sampled_token_ids.to(
224
225
                "cpu", non_blocking=True
            )
226
227
228
229
230
            self._logprobs_tensors_cpu = (
                self._logprobs_tensors.to_cpu_nonblocking()
                if self._logprobs_tensors
                else None
            )
231
            self.async_copy_ready_event.record()
232
233
234

    def get_output(self) -> ModelRunnerOutput:
        """Copy the device tensors to the host and return a ModelRunnerOutput.
235

236
237
        This function blocks until the copy is finished.
        """
238
        max_gen_len = self.sampled_token_ids_cpu.shape[-1]
239
        self.async_copy_ready_event.synchronize()
240

241
242
        # Release the device tensors once the copy has completed.
        del self._logprobs_tensors
243
        del self._sampled_token_ids
244
        if max_gen_len == 1:
245
            valid_sampled_token_ids = self.sampled_token_ids_cpu.tolist()
246
247
            for i in self._invalid_req_indices:
                valid_sampled_token_ids[i].clear()
248
249
250
            logprobs_lists = None
            if self._logprobs_tensors_cpu is not None:
                logprobs_lists = self._logprobs_tensors_cpu.tolists()
251
        else:
252
            valid_sampled_token_ids, logprobs_lists = RejectionSampler.parse_output(
253
254
                self.sampled_token_ids_cpu,
                self.vocab_size,
255
                self._invalid_req_indices,
256
                logprobs_tensors=self._logprobs_tensors_cpu,
257
            )
258
259
260

        output = self._model_runner_output
        output.sampled_token_ids = valid_sampled_token_ids
261
        output.logprobs = logprobs_lists
262
263
264
        return output


265
266
267
268
269
270
271
272
273
274
275
276
277
278
279
280
281
282
283
284
285
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)
286
            raw_pooler_output_cpu = json_map_leaves(
287
288
289
290
                lambda x: None if x is None else x.to("cpu", non_blocking=True),
                self._raw_pooler_output,
            )
            self.async_copy_ready_event.record()
291
292
293
294
            self._model_runner_output.pooler_output = [
                out if include else None
                for out, include in zip(raw_pooler_output_cpu, finished_mask)
            ]
295
296
297
298
299
300
301
302
303
304
305
306

    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


307
308
309
310
311
312
313
314
315
316
317
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
318
    ec_connector_output: ECConnectorOutput | None
319
    cudagraph_stats: CUDAGraphStat | None
320
    slot_mappings: dict[str, torch.Tensor] | list[dict[str, torch.Tensor]] | None
321
322


323
324
325
class GPUModelRunner(
    LoRAModelRunnerMixin, KVConnectorModelRunnerMixin, ECConnectorModelRunnerMixin
):
326
327
    def __init__(
        self,
328
        vllm_config: VllmConfig,
329
        device: torch.device,
330
    ):
331
332
333
        self.vllm_config = vllm_config
        self.model_config = vllm_config.model_config
        self.cache_config = vllm_config.cache_config
334
        self.compilation_config = vllm_config.compilation_config
335
336
337
338
339
340
        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
341

342
        from vllm.model_executor.models.utils import set_cpu_offload_max_bytes
343
344

        set_cpu_offload_max_bytes(int(self.cache_config.cpu_offload_gb * 1024**3))
345

346
347
348
349
        model_config = self.model_config
        cache_config = self.cache_config
        scheduler_config = self.scheduler_config
        parallel_config = self.parallel_config
350
        self.device = device
351
352
        self.pin_memory = is_pin_memory_available()
        self.dtype = self.model_config.dtype
353

354
355
356
        self.kv_cache_dtype = kv_cache_dtype_str_to_dtype(
            cache_config.cache_dtype, self.model_config
        )
357

358
        self.is_pooling_model = model_config.runner_type == "pooling"
359
        self.enable_prompt_embeds = model_config.enable_prompt_embeds
360
        self.is_multimodal_raw_input_only_model = (
361
362
            model_config.is_multimodal_raw_input_only_model
        )
363
364
        # This will be overridden in load_model()
        self.is_multimodal_pruning_enabled = False
365
        self.max_model_len = model_config.max_model_len
366
367
368

        # Always set to false after the first forward pass
        self.calculate_kv_scales = self.cache_config.calculate_kv_scales
369
        self.dcp_world_size = self.parallel_config.decode_context_parallel_size
370
        self.dcp_rank = 0 if self.dcp_world_size <= 1 else get_dcp_group().rank_in_group
371
        self.max_num_tokens = scheduler_config.max_num_batched_tokens
372
        self.max_num_reqs = scheduler_config.max_num_seqs
373

374
375
376
377
378
        # 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 = (
379
            self.parallel_config.distributed_executor_backend == "external_launcher"
380
            and len(get_pp_group().ranks) > 1
381
        )
382

383
        # Model-related.
384
        self.num_query_heads = model_config.get_num_attention_heads(parallel_config)
385
        self.inputs_embeds_size = model_config.get_inputs_embeds_size()
386
        self.attention_chunk_size = model_config.attention_chunk_size
387
        # Only relevant for models using ALiBi (e.g, MPT)
388
        self.use_alibi = model_config.uses_alibi
389

390
        self.cascade_attn_enabled = not self.model_config.disable_cascade_attn
391
        self.is_mm_prefix_lm = self.model_config.is_mm_prefix_lm
392

393
        # Multi-modal data support
394
        self.mm_registry = MULTIMODAL_REGISTRY
395
        self.uses_mrope = model_config.uses_mrope
396
        self.uses_xdrope_dim = model_config.uses_xdrope_dim
397
        self.supports_mm_inputs = self.mm_registry.supports_multimodal_inputs(
398
            model_config
399
        )
400

401
402
403
        if self.model_config.is_encoder_decoder:
            # Maximum length of the encoder input, only for encoder-decoder
            # models.
404
            self.max_encoder_len = scheduler_config.max_num_encoder_input_tokens
405
406
407
        else:
            self.max_encoder_len = 0

408
409
410
        # Async scheduling
        self.use_async_scheduling = self.scheduler_config.async_scheduling

411
        # Sampler
412
        self.sampler = Sampler(logprobs_mode=self.model_config.logprobs_mode)
413

414
        self.eplb_state: EplbState | None = None
415
416
417
418
419
420
        """
        State of the expert parallelism load balancer.

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

421
        # Lazy initializations
422
        # self.model: nn.Module  # Set after load_model
423
        # Initialize in initialize_kv_cache
424
        self.kv_caches: list[torch.Tensor] = []
425
426
427
        # Initialize in initialize_kv_cache_tensors
        self.cross_layers_kv_cache: torch.Tensor | None = None
        self.cross_layers_attn_backend: type[AttentionBackend] | None = None
428
429
        # indexes: [kv_cache_group_id][attn_group]
        self.attn_groups: list[list[AttentionGroup]] = []
430
431
        # self.kv_cache_config: KVCacheConfig

432
433
        # mm_hash ->  encoder_output
        self.encoder_cache: dict[str, torch.Tensor] = {}
434

435
        self.use_aux_hidden_state_outputs = False
436
437
438
439
440
        # 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:
441
            self.drafter: (
442
443
444
445
446
                NgramProposer
                | SuffixDecodingProposer
                | EagleProposer
                | DraftModelProposer
                | MedusaProposer
447
            )
448
449
            if self.speculative_config.method == "ngram":
                self.drafter = NgramProposer(self.vllm_config)
450
451
452
453
454
455
            elif self.speculative_config.uses_draft_model():
                self.drafter = DraftModelProposer(
                    vllm_config=self.vllm_config,
                    device=self.device,
                    runner=self,
                )
456
457
            elif self.speculative_config.method == "suffix":
                self.drafter = SuffixDecodingProposer(self.vllm_config)
458
            elif self.speculative_config.use_eagle():
459
                self.drafter = EagleProposer(self.vllm_config, self.device, self)
460
                if self.speculative_config.method == "eagle3":
461
462
463
                    self.use_aux_hidden_state_outputs = (
                        self.drafter.eagle3_use_aux_hidden_state
                    )
464
465
            elif self.speculative_config.method == "medusa":
                self.drafter = MedusaProposer(
466
                    vllm_config=self.vllm_config, device=self.device
467
                )
468
            else:
469
470
471
472
                raise ValueError(
                    "Unknown speculative decoding method: "
                    f"{self.speculative_config.method}"
                )
473
            self.rejection_sampler = RejectionSampler(self.sampler)
474

475
476
477
        self.num_spec_tokens = 0
        if self.speculative_config:
            self.num_spec_tokens = self.speculative_config.num_speculative_tokens
478
479
480
481
482
            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
483

484
        # Request states.
485
        self.requests: dict[str, CachedRequestState] = {}
486
487
488
        # NOTE(rob): num_prompt_logprobs only includes reqs
        # that are currently in the prefill phase.
        self.num_prompt_logprobs: dict[str, int] = {}
489
        self.comm_stream = torch.cuda.Stream()
490

491
492
493
494
495
496
497
498
499
        # 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.
500
501
502
503
        logits_processors = model_config.logits_processors
        custom_logitsprocs: Sequence[str | type[LogitsProcessor]] = (
            tuple(logits_processors) if logits_processors is not None else ()
        )
504
505
        self.input_batch = InputBatch(
            max_num_reqs=self.max_num_reqs,
506
507
508
            # 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),
509
510
511
            max_num_batched_tokens=self.max_num_tokens,
            device=self.device,
            pin_memory=self.pin_memory,
512
            vocab_size=self.model_config.get_vocab_size(),
513
            block_sizes=[self.cache_config.block_size],
514
            kernel_block_sizes=[self.cache_config.block_size],
515
            is_spec_decode=bool(self.vllm_config.speculative_config),
516
            logitsprocs=build_logitsprocs(
517
518
519
                self.vllm_config,
                self.device,
                self.pin_memory,
520
                self.is_pooling_model,
521
                custom_logitsprocs,
522
            ),
523
524
525
            # 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),
526
            is_pooling_model=self.is_pooling_model,
527
            cp_kv_cache_interleave_size=self.parallel_config.cp_kv_cache_interleave_size,
528
        )
529

530
531
532
533
534
        # 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.
535
        self.prepare_inputs_event: torch.Event | None = None
536
537
        if self.use_async_scheduling:
            self.async_output_copy_stream = torch.cuda.Stream()
538
            self.prepare_inputs_event = torch.Event()
539

540
        # self.cudagraph_batch_sizes sorts in ascending order.
541
542
543
544
        if (
            self.compilation_config.cudagraph_capture_sizes
            and self.compilation_config.cudagraph_mode != CUDAGraphMode.NONE
        ):
545
546
            self.cudagraph_batch_sizes = sorted(
                self.compilation_config.cudagraph_capture_sizes
547
            )
548

549
        # Cache the device properties.
550
        self._init_device_properties()
551

552
553
554
555
        # Encoder timing registry for observability
        self.encoder_timing_registry: dict[str, EncoderTimingStats] = {}
        self._encoder_timing_lock = threading.Lock()

556
        # Persistent buffers for CUDA graphs.
557
558
559
560
561
        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
        )
562
        self.seq_lens = self._make_buffer(self.max_num_reqs, dtype=torch.int32)
563
        self.encoder_seq_lens = self._make_buffer(self.max_num_reqs, dtype=torch.int32)
564
565
566
567
        if self.dcp_world_size > 1:
            self.dcp_local_seq_lens = self._make_buffer(
                self.max_num_reqs, dtype=torch.int32
            )
568
569
570
        # 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.
571
        self.inputs_embeds = self._make_buffer(
572
            self.max_num_tokens, self.inputs_embeds_size, dtype=self.dtype, numpy=False
573
574
        )
        self.is_token_ids = self._make_buffer(self.max_num_tokens, dtype=torch.bool)
575
576
        self.discard_request_mask = self._make_buffer(
            self.max_num_reqs, dtype=torch.bool
577
578
579
580
581
582
583
        )
        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
        )
584

585
586
        # Only relevant for multimodal models
        if self.supports_mm_inputs:
587
588
589
590
591
592
593
            # 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
594

595
        # Only relevant for models using M-RoPE (e.g, Qwen2-VL)
596
        if self.uses_mrope:
Roger Wang's avatar
Roger Wang committed
597
598
599
600
            # 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
601
602
603
604
605
606

            # 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
607
            self.mrope_positions = self._make_buffer(
608
609
                (3, self.max_num_tokens + 1), dtype=torch.int64
            )
610

611
612
613
614
615
616
617
        # 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
            )

618
        # None in the first PP rank. The rest are set after load_model.
619
        self.intermediate_tensors: IntermediateTensors | None = None
620

621
        # OPTIMIZATION: Cache the tensors rather than creating them every step.
622
        # Keep in int64 to avoid overflow with long context
623
624
625
626
        self.arange_np = np.arange(
            max(self.max_num_reqs + 1, self.max_model_len, self.max_num_tokens),
            dtype=np.int64,
        )
627

628
629
630
631
632
        # 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] = {}
633
634
635
636
637
        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(
638
639
                self.max_num_tokens, dtype=torch.int32, device=self.device
            )
640

641
        self.uniform_decode_query_len = 1 + self.num_spec_tokens
642
643
644
645

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

646
        self.mm_budget = (
647
            MultiModalBudget(self.vllm_config, self.mm_registry)
648
649
650
            if self.supports_mm_inputs
            else None
        )
651

652
        self.reorder_batch_threshold: int | None = None
653

654
655
656
657
658
        # 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()

659
        # Cached outputs.
660
        self._draft_token_ids: list[list[int]] | torch.Tensor | None = None
661
        self._draft_token_req_ids: list[str] | None = None
662
        self.transfer_event = torch.Event()
663
        self.sampled_token_ids_pinned_cpu = torch.empty(
664
            (self.max_num_reqs, 1),
665
666
            dtype=torch.int64,
            device="cpu",
667
668
            pin_memory=self.pin_memory,
        )
669

670
671
        # Pre-allocated tensor for copying valid sampled token counts to CPU,
        # with dedicated stream for overlapping and event for coordination.
672
        self.valid_sampled_token_count_event: torch.Event | None = None
673
        self.valid_sampled_token_count_copy_stream: torch.cuda.Stream | None = None
674
675
676
677
678
679
680
681
682
683
684
685
686
687
688
689
690
691
692
693
694
695
696
697
        # 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
        if self.num_spec_tokens:
            self.draft_token_ids_event = torch.Event()
            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,
                )
698

699
700
        # Ephemeral state transferred between execute_model() and sample_tokens().
        self.execute_model_state: ExecuteModelState | None = None
701
        self.kv_connector_output: KVConnectorOutput | None = None
702
        self.mamba_state_idx: dict[str, int] = {}
703
        self.layerwise_nvtx_hooks_registered = False
704

705
706
707
708
709
710
711
    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

712
713
714
715
    def reset_mm_cache(self) -> None:
        if self.mm_budget:
            self.mm_budget.reset_cache()

716
717
718
719
720
721
722
723
724
725
726
727
728
729
730
731
732
733
734
735
736
737
738
739
740
741
742
743
744
745
746
747
748
749
750
751
752
753
754
755
756
757
758
759
    @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)

760
761
762
763
    def _get_positions(self, num_tokens: Any):
        if isinstance(num_tokens, int):
            if self.uses_mrope:
                return self.mrope_positions.gpu[:, :num_tokens]
764
765
            if self.uses_xdrope_dim > 0:
                return self.xdrope_positions.gpu[:, :num_tokens]
766
767
768
769
            return self.positions.gpu[:num_tokens]
        else:
            if self.uses_mrope:
                return self.mrope_positions.gpu[:, num_tokens]
770
771
            if self.uses_xdrope_dim > 0:
                return self.xdrope_positions.gpu[:, num_tokens]
772
773
            return self.positions.gpu[num_tokens]

774
    def _make_buffer(
775
        self, *size: int | torch.SymInt, dtype: torch.dtype, numpy: bool = True
776
777
778
779
780
781
782
783
    ) -> CpuGpuBuffer:
        return CpuGpuBuffer(
            *size,
            dtype=dtype,
            device=self.device,
            pin_memory=self.pin_memory,
            with_numpy=numpy,
        )
784

785
    def _init_model_kwargs(self):
786
787
        model_kwargs = dict[str, Any]()

788
        if not self.is_pooling_model:
789
790
            return model_kwargs

791
792
        num_reqs = self.input_batch.num_reqs
        pooling_params = self.input_batch.get_pooling_params()
793
794
795

        token_type_id_requests = dict[int, Any]()
        for i, param in enumerate(pooling_params):
796
797
798
799
800
            if (
                param.extra_kwargs is not None
                and (token_types := param.extra_kwargs.get("compressed_token_type_ids"))
                is not None
            ):
801
802
803
804
805
                token_type_id_requests[i] = token_types

        if len(token_type_id_requests) == 0:
            return model_kwargs

806
        seq_lens = self.seq_lens.gpu[:num_reqs]
807
808
809
810
811
812
813
814
        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(
815
816
            device=self.device
        )
817
818
        return model_kwargs

819
    def _may_reorder_batch(self, scheduler_output: "SchedulerOutput") -> None:
820
821
        """
        Update the order of requests in the batch based on the attention
822
        backend's needs. For example, some attention backends (namely MLA) may
823
824
825
826
827
828
        want to separate requests based on if the attention computation will be
        compute-bound or memory-bound.

        Args:
            scheduler_output: The scheduler output.
        """
829
830
831
832
833
834
835
836
        # Attention free models have zero kv_cache_goups, however models
        # 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

837
838
839
840
        if self.reorder_batch_threshold is not None:
            reorder_batch_to_split_decodes_and_prefills(
                self.input_batch,
                scheduler_output,
841
842
                decode_threshold=self.reorder_batch_threshold,
            )
843

844
845
    # Note: used for model runner override.
    def _init_device_properties(self) -> None:
846
        """Initialize attributes from torch.cuda.get_device_properties"""
847
848
849
850
851
852
853
        self.device_properties = torch.cuda.get_device_properties(self.device)
        self.num_sms = self.device_properties.multi_processor_count

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

854
    def _update_states(self, scheduler_output: "SchedulerOutput") -> None:
855
856
857
858
859
860
        """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.

861
862
        The SamplingMetadata is updated and copied to the GPU if there is a
        new/resumed/paused/finished request in the batch.
863
864
        """
        # Remove finished requests from the cached states.
865
866
        for req_id in scheduler_output.finished_req_ids:
            self.requests.pop(req_id, None)
867
            self.num_prompt_logprobs.pop(req_id, None)
868
869
870
871
872
873
874
        # 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:
875
            self.input_batch.remove_request(req_id)
876
877

        # Free the cached encoder outputs.
878
879
        for mm_hash in scheduler_output.free_encoder_mm_hashes:
            self.encoder_cache.pop(mm_hash, None)
880

881
882
883
884
885
886
887
        # 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()
888
889
890
891
892
893
894
895
        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)
896
897
898
899
900
        # 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:
901
            self.input_batch.remove_request(req_id)
902

903
        reqs_to_add: list[CachedRequestState] = []
904
        # Add new requests to the cached states.
905
906
        for new_req_data in scheduler_output.scheduled_new_reqs:
            req_id = new_req_data.req_id
907
908
909
910
911
912
            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

913
            sampling_params = new_req_data.sampling_params
914
            pooling_params = new_req_data.pooling_params
915

916
917
918
919
            if (
                sampling_params
                and sampling_params.sampling_type == SamplingType.RANDOM_SEED
            ):
920
921
922
923
924
                generator = torch.Generator(device=self.device)
                generator.manual_seed(sampling_params.seed)
            else:
                generator = None

925
926
            if self.is_pooling_model:
                assert pooling_params is not None
927
928
                task = pooling_params.task
                assert task is not None, "You did not set `task` in the API"
929

930
                model = cast(VllmModelForPooling, self.get_model())
931
                to_update = model.pooler.get_pooling_updates(task)
932
933
                to_update.apply(pooling_params)

934
            req_state = CachedRequestState(
935
                req_id=req_id,
936
                prompt_token_ids=new_req_data.prompt_token_ids,
937
                prompt_embeds=new_req_data.prompt_embeds,
938
                mm_features=new_req_data.mm_features,
939
                sampling_params=sampling_params,
940
                pooling_params=pooling_params,
941
                generator=generator,
942
943
                block_ids=new_req_data.block_ids,
                num_computed_tokens=new_req_data.num_computed_tokens,
944
                output_token_ids=[],
945
                lora_request=new_req_data.lora_request,
946
            )
947
948
            self.requests[req_id] = req_state

949
950
951
952
953
954
955
            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
                )

956
            # Only relevant for models using M-RoPE (e.g, Qwen2-VL)
957
            if self.uses_mrope:
958
                self._init_mrope_positions(req_state)
959

960
961
962
963
            # Only relevant for models using XD-RoPE (e.g, HunYuan-VL)
            if self.uses_xdrope_dim > 0:
                self._init_xdrope_positions(req_state)

964
            reqs_to_add.append(req_state)
965

966
        # Update the states of the running/resumed requests.
967
        is_last_rank = get_pp_group().is_last_rank
968
        req_data = scheduler_output.scheduled_cached_reqs
969
        scheduled_spec_tokens = scheduler_output.scheduled_spec_decode_tokens
970
971
972
973
974

        # 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()

975
        for i, req_id in enumerate(req_data.req_ids):
976
            req_state = self.requests[req_id]
977
978
            num_computed_tokens = req_data.num_computed_tokens[i]
            new_block_ids = req_data.new_block_ids[i]
979
            resumed_from_preemption = req_id in req_data.resumed_req_ids
980
            num_output_tokens = req_data.num_output_tokens[i]
981
            req_index = self.input_batch.req_id_to_index.get(req_id)
982

983
984
985
986
987
988
989
990
991
992
993
994
995
996
            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.
                # second step: num_computed_tokens = 100(prompt lenth),
                # 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.
997
998
999
1000
1001
1002
1003
1004
1005
                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)
1006

1007
            # Update the cached states.
1008
            req_state.num_computed_tokens = num_computed_tokens
1009
1010
1011
1012
1013
1014
1015
1016

            if not is_last_rank:
                # 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.
                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.
1017
1018
1019
                num_new_tokens = (
                    num_computed_tokens + len(new_token_ids) - req_state.num_tokens
                )
1020
1021
1022
1023
                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:
1024
                    req_state.output_token_ids.extend(new_token_ids[-num_new_tokens:])
1025
1026
1027
1028
1029
            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:
1030
1031
1032
1033
                    end_idx = (
                        self.input_batch.num_prompt_tokens[req_index]
                        + num_output_tokens
                    )
1034
                    self.input_batch.num_tokens_no_spec[req_index] = end_idx
1035

1036
            # Update the block IDs.
1037
            if not resumed_from_preemption:
1038
1039
                if new_block_ids is not None:
                    # Append the new blocks to the existing block IDs.
1040
                    for block_ids, new_ids in zip(req_state.block_ids, new_block_ids):
1041
                        block_ids.extend(new_ids)
1042
            else:
1043
                assert req_index is None
1044
                assert new_block_ids is not None
1045
1046
                # The request is resumed from preemption.
                # Replace the existing block IDs with the new ones.
1047
                req_state.block_ids = new_block_ids
1048
1049
1050
1051
1052

            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.
1053
1054
1055
1056
1057
1058
1059

                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:]

1060
                reqs_to_add.append(req_state)
1061
1062
1063
                continue

            # Update the persistent batch.
1064
            self.input_batch.num_computed_tokens_cpu[req_index] = num_computed_tokens
1065
            if new_block_ids is not None:
1066
                self.input_batch.block_table.append_row(new_block_ids, req_index)
1067
1068
1069
1070
1071
1072
1073

            # 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)
1074
                self.input_batch.token_ids_cpu[
1075
1076
1077
                    req_index, start_token_index:end_token_index
                ] = new_token_ids
                self.input_batch.num_tokens_no_spec[req_index] = end_token_index
1078

1079
            # Add spec_token_ids to token_ids_cpu.
1080
            self.input_batch.update_req_spec_token_ids(req_state, scheduled_spec_tokens)
1081

1082
1083
        # Add the new or resumed requests to the persistent batch.
        # The smaller empty indices are filled first.
1084
1085
        for request in reqs_to_add:
            self.input_batch.add_request(request)
1086
            self.input_batch.update_req_spec_token_ids(request, scheduled_spec_tokens)
1087

1088
1089
1090
1091
1092
1093
        # 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()
1094

1095
    def _update_states_after_model_execute(
1096
        self, output_token_ids: torch.Tensor, scheduler_output: "SchedulerOutput"
1097
    ) -> None:
1098
1099
1100
1101
1102
1103
1104
1105
        """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.
        """
1106
        if not self.speculative_config or not self.model_config.is_hybrid:
1107
1108
1109
            return

        # Find the number of accepted tokens for each sequence.
1110
1111
1112
1113
1114
1115
1116
1117
1118
1119
1120
1121
1122
1123
1124
1125
1126
1127
1128
1129
        num_accepted_tokens = (
            (
                torch.cat(
                    [
                        output_token_ids,
                        torch.full(
                            (output_token_ids.size(0), 1),
                            -1,
                            device=output_token_ids.device,
                        ),
                    ],
                    dim=1,
                )
                == -1
            )
            .int()
            .argmax(-1)
            .cpu()
            .numpy()
        )
1130
1131
        for i, num_tokens in enumerate(num_accepted_tokens):
            self.input_batch.num_accepted_tokens_cpu[i] = num_tokens
1132
1133
1134
1135
1136
1137
1138
1139
1140
1141
        if self.cache_config.mamba_cache_mode == "align":
            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(),
            )
1142

1143
1144
1145
1146
1147
1148
1149
1150
1151
1152
1153
1154
1155
1156
1157
1158
1159
1160
1161
1162
1163
1164
1165
1166
1167
1168
1169
1170
1171
1172
1173
1174
1175
1176
    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

1177
    def _init_mrope_positions(self, req_state: CachedRequestState):
1178
1179
        model = self.get_model()
        assert supports_mrope(model), "M-RoPE support is not implemented."
1180
1181
1182
1183
        assert req_state.prompt_token_ids is not None, (
            "M-RoPE requires prompt_token_ids to be available."
        )
        mrope_model = cast(SupportsMRoPE, model)
1184
1185

        req_state.mrope_positions, req_state.mrope_position_delta = (
1186
            mrope_model.get_mrope_input_positions(
1187
                req_state.prompt_token_ids,
1188
                req_state.mm_features,
1189
            )
1190
        )
1191

1192
1193
1194
1195
1196
1197
1198
1199
1200
1201
1202
1203
1204
    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,
        )

1205
    def _extract_mm_kwargs(
1206
        self,
1207
1208
        scheduler_output: "SchedulerOutput",
    ) -> BatchedTensorInputs:
1209
        if not scheduler_output or not self.is_multimodal_raw_input_only_model:
1210
            return {}
1211

1212
1213
        mm_kwargs = list[MultiModalKwargsItem]()
        for req in scheduler_output.scheduled_new_reqs:
1214
1215
1216
            for feature in req.mm_features:
                if feature.data is not None:
                    mm_kwargs.append(feature.data)
1217

1218
1219
1220
        # Input all modalities at once
        mm_kwargs_combined: BatchedTensorInputs = {}
        for _, _, mm_kwargs_group in group_mm_kwargs_by_modality(
1221
1222
1223
            mm_kwargs,
            device=self.device,
            pin_memory=self.pin_memory,
1224
1225
        ):
            mm_kwargs_combined.update(mm_kwargs_group)
1226

1227
        return mm_kwargs_combined
1228

1229
    def _dummy_mm_kwargs(self, num_seqs: int) -> BatchedTensorInputs:
1230
        if not self.is_multimodal_raw_input_only_model:
1231
            return {}
1232

1233
1234
1235
1236
1237
        mm_budget = self.mm_budget
        assert mm_budget is not None

        dummy_modality = mm_budget.get_modality_with_max_tokens()
        return self._get_mm_dummy_batch(dummy_modality, num_seqs)
1238

1239
1240
1241
    def _get_cumsum_and_arange(
        self,
        num_tokens: np.ndarray,
1242
        cumsum_dtype: np.dtype | None = None,
1243
1244
1245
1246
1247
1248
1249
1250
1251
1252
1253
1254
1255
1256
1257
1258
    ) -> 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

1259
    def _prepare_input_ids(
1260
1261
1262
1263
        self,
        scheduler_output: "SchedulerOutput",
        total_num_scheduled_tokens: int,
        cu_num_tokens: np.ndarray,
1264
    ) -> None:
1265
        """Prepare the input IDs for the current batch.
1266

1267
1268
1269
1270
1271
1272
1273
        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)
1274
1275
1276
            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)
1277
1278
1279
1280
1281
1282
1283
            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
1284
1285
1286
1287
        sample_flattened_indices: list[int] = []
        spec_flattened_indices: list[int] = []
        prev_common_req_indices: list[int] = []
        prev_draft_token_indices: list[int] = []
1288
1289
        indices_match = True
        max_flattened_index = -1
1290
1291
1292
        total_num_spec_tokens = 0
        scheduled_spec_tokens = scheduler_output.scheduled_spec_decode_tokens

1293
1294
1295
1296
1297
        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.
1298
1299
                draft_len = len(scheduled_spec_tokens.get(req_id, ()))
                total_num_spec_tokens += draft_len
1300
                flattened_index = cu_num_tokens[cur_index].item() - 1
1301
1302
1303
1304
1305
1306
1307
1308
1309
1310
1311
1312
1313
1314
1315
                # 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))
1316
                indices_match &= prev_index == flattened_index
1317
                max_flattened_index = max(max_flattened_index, flattened_index)
1318
1319
1320
        num_commmon_tokens = len(sample_flattened_indices)
        total_without_spec = total_num_scheduled_tokens - total_num_spec_tokens
        if num_commmon_tokens < total_without_spec:
1321
1322
1323
            # 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)
1324
1325
1326
            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)
1327
1328
        if num_commmon_tokens == 0:
            # No requests in common with the previous iteration
1329
            # So input_ids.cpu will have all the input ids.
1330
1331
1332
1333
1334
1335
1336
            return
        if indices_match and max_flattened_index == (num_commmon_tokens - 1):
            # 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.
            self.input_ids.gpu[:num_commmon_tokens].copy_(
1337
1338
1339
                self.input_batch.prev_sampled_token_ids[:num_commmon_tokens, 0],
                non_blocking=True,
            )
1340
1341
            if self.enable_prompt_embeds:
                self.is_token_ids.gpu[:num_commmon_tokens] = True
1342
            return
1343
        # Upload the index tensors asynchronously so the scatter can be non-blocking.
1344
1345
        sampled_tokens_index_tensor = torch.tensor(
            sample_flattened_indices, dtype=torch.int64, pin_memory=self.pin_memory
1346
        ).to(self.device, non_blocking=True)
1347
        prev_common_req_indices_tensor = torch.tensor(
1348
1349
            prev_common_req_indices, dtype=torch.int64, pin_memory=self.pin_memory
        ).to(self.device, non_blocking=True)
1350
1351
        self.input_ids.gpu.scatter_(
            dim=0,
1352
            index=sampled_tokens_index_tensor,
1353
            src=self.input_batch.prev_sampled_token_ids[
1354
1355
1356
                prev_common_req_indices_tensor, 0
            ],
        )
1357

1358
1359
1360
1361
1362
1363
1364
1365
1366
1367
1368
1369
1370
1371
1372
1373
1374
1375
1376
1377
1378
1379
        # 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],
        )

1380
1381
    def _get_encoder_seq_lens(
        self,
1382
        num_scheduled_tokens: dict[str, int],
1383
1384
        kv_cache_spec: KVCacheSpec,
        num_reqs: int,
1385
    ) -> tuple[torch.Tensor | None, np.ndarray | None]:
1386
        if not isinstance(kv_cache_spec, CrossAttentionSpec):
1387
            return None, None
1388

1389
1390
        # Zero out buffer for padding requests that are not actually scheduled (CGs)
        self.encoder_seq_lens.np[:num_reqs] = 0
1391
1392
        # Build encoder_seq_lens array mapping request indices to
        # encoder lengths for inputs scheduled in this batch
1393
        for req_id in num_scheduled_tokens:
1394
            req_index = self.input_batch.req_id_to_index[req_id]
1395
1396
1397
1398
1399
1400
1401
1402
1403
1404
1405
1406
1407
1408
1409
1410
            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

        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]
1411

1412
        return encoder_seq_lens, encoder_seq_lens_cpu
1413

1414
    def _prepare_inputs(
1415
1416
1417
        self,
        scheduler_output: "SchedulerOutput",
        num_scheduled_tokens: np.ndarray,
1418
1419
    ) -> tuple[
        torch.Tensor,
1420
        SpecDecodeMetadata | None,
1421
    ]:
1422
1423
        """
        :return: tuple[
1424
            logits_indices, spec_decode_metadata,
1425
1426
        ]
        """
1427
1428
1429
1430
1431
1432
1433
        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.
1434
        self.input_batch.block_table.commit_block_table(num_reqs)
1435
1436
1437

        # Get request indices.
        # E.g., [2, 5, 3] -> [0, 0, 1, 1, 1, 1, 1, 2, 2, 2]
1438
        req_indices = np.repeat(self.arange_np[:num_reqs], num_scheduled_tokens)
1439

1440
1441
        # cu_num_tokens: [2, 5, 3] -> [2, 7, 10]
        # arange: [0, 1, 0, 1, 2, 3, 4, 0, 1, 2]
1442
        cu_num_tokens, arange = self._get_cumsum_and_arange(num_scheduled_tokens)
1443
1444

        # Get positions.
1445
        positions_np = self.positions.np[:total_num_scheduled_tokens]
1446
1447
1448
1449
1450
        np.add(
            self.input_batch.num_computed_tokens_cpu[req_indices],
            arange,
            out=positions_np,
        )
1451

1452
1453
        # Calculate M-RoPE positions.
        # Only relevant for models using M-RoPE (e.g, Qwen2-VL)
1454
        if self.uses_mrope:
1455
1456
            self._calc_mrope_positions(scheduler_output)

1457
1458
1459
1460
1461
        # 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)

1462
1463
1464
1465
        # 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.
1466
1467
1468
        token_indices = (
            positions_np + req_indices * self.input_batch.token_ids_cpu.shape[1]
        )
1469
        token_indices_tensor = torch.from_numpy(token_indices)
1470

1471
1472
1473
        # 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.
1474
1475
1476
1477
1478
1479
        torch.index_select(
            self.input_batch.token_ids_cpu_tensor.flatten(),
            0,
            token_indices_tensor,
            out=self.input_ids.cpu[:total_num_scheduled_tokens],
        )
1480
        if self.enable_prompt_embeds:
1481
            is_token_ids = self.input_batch.is_token_ids_tensor.flatten()
1482
1483
1484
1485
            torch.index_select(
                is_token_ids,
                0,
                token_indices_tensor,
1486
1487
                out=self.is_token_ids.cpu[:total_num_scheduled_tokens],
            )
1488
1489
1490
1491
1492
1493
1494
1495
1496
1497
1498
1499
1500
1501
1502
1503
1504
1505
1506
1507
1508
1509
1510
1511
1512
1513
1514
1515
1516
1517
1518
1519
1520

        # 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:
1521
1522
1523
                    self.inputs_embeds.cpu[
                        output_idx : output_idx + actual_num_sched
                    ].copy_(req_embeds[start_pos:actual_end])
1524
1525

                output_idx += num_sched
1526

1527
1528
        self.input_batch.block_table.compute_slot_mapping(req_indices, positions_np)
        self.input_batch.block_table.commit_slot_mapping(total_num_scheduled_tokens)
1529
1530

        # Prepare the attention metadata.
1531
        self.query_start_loc.np[0] = 0
1532
        self.query_start_loc.np[1 : num_reqs + 1] = cu_num_tokens
1533
1534
        # Note: pad query_start_loc to be non-decreasing, as kernels
        # like FlashAttention requires that
1535
        self.query_start_loc.np[num_reqs + 1 :].fill(cu_num_tokens[-1])
1536
        self.query_start_loc.copy_to_gpu()
1537
        query_start_loc = self.query_start_loc.gpu[: num_reqs + 1]
1538

1539
        self.seq_lens.np[:num_reqs] = (
1540
1541
            self.input_batch.num_computed_tokens_cpu[:num_reqs] + num_scheduled_tokens
        )
1542
        # Fill unused with 0 for full cuda graph mode.
1543
1544
        self.seq_lens.np[num_reqs:].fill(0)
        self.seq_lens.copy_to_gpu()
1545

1546
        num_tokens = [self.requests[r].num_tokens for r in self.input_batch.req_ids]
1547
1548
        num_tokens_np = np.array(num_tokens, dtype=np.int32)

1549
        # Record which requests should not be sampled,
1550
        # so that we could clear the sampled tokens before returning
1551
1552
        self.discard_request_mask.np[:num_reqs] = (
            self.seq_lens.np[:num_reqs] < num_tokens_np
1553
        )
1554
        self.discard_request_mask.copy_to_gpu(num_reqs)
1555

1556
        # Copy the tensors to the GPU.
1557
1558
1559
1560
1561
        self._prepare_input_ids(
            scheduler_output,
            total_num_scheduled_tokens,
            cu_num_tokens,
        )
1562

1563
        if self.uses_mrope:
1564
            # Only relevant for models using M-RoPE (e.g, Qwen2-VL)
1565
1566
            self.mrope_positions.gpu[:, :total_num_scheduled_tokens].copy_(
                self.mrope_positions.cpu[:, :total_num_scheduled_tokens],
1567
1568
                non_blocking=True,
            )
1569
1570
1571
1572
1573
1574
        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,
            )
1575
1576
        else:
            # Common case (1D positions)
1577
            self.positions.copy_to_gpu(total_num_scheduled_tokens)
1578

1579
        use_spec_decode = len(scheduler_output.scheduled_spec_decode_tokens) > 0
1580
1581
1582
1583
1584
1585
1586
1587
        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
1588
            num_sampled_tokens = np.ones(num_reqs, dtype=np.int32)
1589
1590
1591
1592
1593
        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)
1594
1595
1596
            # 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)
1597
1598
1599
1600
            for (
                req_id,
                draft_token_ids,
            ) in scheduler_output.scheduled_spec_decode_tokens.items():
1601
1602
                req_idx = self.input_batch.req_id_to_index[req_id]
                num_draft_tokens[req_idx] = len(draft_token_ids)
1603
1604
1605
1606
1607
                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)
1608
            spec_decode_metadata = self._calc_spec_decode_metadata(
1609
1610
                num_draft_tokens, cu_num_tokens
            )
1611
            logits_indices = spec_decode_metadata.logits_indices
1612
            num_sampled_tokens = num_draft_tokens + 1
1613
            # For DECODE only cuda graph of some attention backends (e.g., GDN).
1614
            self.num_decode_draft_tokens.np[:num_reqs] = num_decode_draft_tokens
1615
1616
            self.num_decode_draft_tokens.np[num_reqs:].fill(-1)
            self.num_decode_draft_tokens.copy_to_gpu()
1617

1618
1619
1620
1621
1622
        # Hot-Swap lora model
        if self.lora_config:
            assert (
                np.sum(num_sampled_tokens)
                <= self.vllm_config.scheduler_config.max_num_batched_tokens
1623
            )
1624
1625
1626
1627
1628
1629
1630
1631
1632
1633
1634
            self.set_active_loras(
                self.input_batch, num_scheduled_tokens, num_sampled_tokens
            )

        return (
            logits_indices,
            spec_decode_metadata,
        )

    def _build_attention_metadata(
        self,
1635
        num_tokens: int,
1636
        num_reqs: int,
1637
1638
1639
        max_query_len: int,
        num_tokens_padded: int | None = None,
        num_reqs_padded: int | None = None,
1640
1641
1642
1643
        ubatch_slices: UBatchSlices | None = None,
        logits_indices: torch.Tensor | None = None,
        use_spec_decode: bool = False,
        for_cudagraph_capture: bool = False,
1644
        num_scheduled_tokens: dict[str, int] | None = None,
1645
        cascade_attn_prefix_lens: list[list[int]] | None = None,
1646
        slot_mappings: dict[int, torch.Tensor] | None = None,
1647
1648
1649
1650
    ) -> tuple[PerLayerAttnMetadata, CommonAttentionMetadata | None]:
        """
        :return: tuple[attn_metadata, spec_decode_common_attn_metadata]
        """
1651
1652
1653
1654
        # Attention metadata is not needed for attention free models
        if len(self.kv_cache_config.kv_cache_groups) == 0:
            return {}, None

1655
1656
        num_tokens_padded = num_tokens_padded or num_tokens
        num_reqs_padded = num_reqs_padded or num_reqs
1657
        assert num_reqs_padded is not None and num_tokens_padded is not None
1658

1659
1660
1661
        attn_metadata: PerLayerAttnMetadata = {}
        if ubatch_slices is not None:
            attn_metadata = [dict() for _ in range(len(ubatch_slices))]
1662

1663
1664
1665
1666
1667
1668
1669
1670
        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()

1671
1672
        if use_spec_decode:
            self.num_accepted_tokens.np[:num_reqs] = (
1673
1674
                self.input_batch.num_accepted_tokens_cpu[:num_reqs]
            )
1675
1676
            self.num_accepted_tokens.np[num_reqs:].fill(1)
            self.num_accepted_tokens.copy_to_gpu()
1677

1678
        kv_cache_groups = self.kv_cache_config.kv_cache_groups
1679

1680
        def _get_block_table(kv_cache_gid: int):
1681
1682
1683
            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):
1684
                blk_table_tensor = torch.zeros(
1685
                    (num_reqs_padded, 1),
1686
                    dtype=torch.int32,
1687
1688
                    device=self.device,
                )
1689
            else:
1690
                blk_table = self.input_batch.block_table[kv_cache_gid]
1691
                blk_table_tensor = blk_table.get_device_tensor(num_reqs_padded)
1692

1693
1694
1695
            # 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)
1696
            return blk_table_tensor
1697

1698
1699
1700
        assert slot_mappings is not None
        block_table_gid_0 = _get_block_table(0)
        slot_mapping_gid_0 = slot_mappings[0]
1701

1702
1703
        if self.model_config.enable_return_routed_experts:
            self.slot_mapping = slot_mapping_gid_0[:num_tokens].cpu().numpy()
1704
1705
1706
1707
1708
1709
1710
1711
1712
1713
1714
1715
1716
1717
1718
1719
1720
1721
1722
1723
1724
1725
1726
1727
1728
1729
1730
1731
1732
1733
1734
1735
1736
1737
1738
1739
1740
1741
        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
            )

1742
1743
1744
1745
1746
1747
1748
1749
1750
        # 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
        ] = {}

1751
1752
1753
1754
1755
1756
1757
        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]
1758
            builder = attn_group.get_metadata_builder(ubid or 0)
1759
1760
1761
1762
            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))
1763

1764
1765
1766
1767
1768
1769
1770
1771
1772
1773
1774
1775
1776
1777
1778
1779
1780
1781
1782
1783
            cascade_attn_prefix_len = (
                cascade_attn_prefix_lens[kv_cache_gid][attn_gid]
                if cascade_attn_prefix_lens
                else 0
            )

            extra_attn_metadata_args = {}
            if use_spec_decode and isinstance(builder, GDNAttentionMetadataBuilder):
                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
                )
1784
1785
1786
1787
1788
1789
1790
1791
1792
            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,
                )
1793
1794
1795
1796
1797
1798
            else:
                attn_metadata_i = builder.build(
                    common_prefix_len=cascade_attn_prefix_len,
                    common_attn_metadata=common_attn_metadata,
                    **extra_attn_metadata_args,
                )
1799
1800
                if builder.supports_update_block_table:
                    cached_attn_metadata[cache_key] = attn_metadata_i
1801
1802
1803
1804
1805
1806
1807
1808
1809
1810
1811
1812
1813
1814
1815
1816
1817
1818
1819
1820
1821
1822
1823

            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,
1824
            )
1825
            if kv_cache_gid > 0:
1826
1827
                cm.block_table_tensor = _get_block_table(kv_cache_gid)
                cm.slot_mapping = slot_mappings[kv_cache_gid]
1828

1829
            if self.speculative_config and spec_decode_common_attn_metadata is None:
1830
                if isinstance(self.drafter, EagleProposer):
1831
                    if self.drafter.attn_layer_names[0] in kv_cache_group.layer_names:
1832
                        spec_decode_common_attn_metadata = cm
1833
                else:
1834
                    spec_decode_common_attn_metadata = cm
1835

1836
            for attn_gid in range(len(self.attn_groups[kv_cache_gid])):
1837
                if ubatch_slices is not None:
1838
1839
1840
                    for ubid, _cm in enumerate(split_attn_metadata(ubatch_slices, cm)):
                        _build_attn_group_metadata(kv_cache_gid, attn_gid, _cm, ubid)

1841
                else:
1842
                    _build_attn_group_metadata(kv_cache_gid, attn_gid, cm)
1843

1844
1845
1846
1847
1848
1849
1850
1851
1852
1853
1854
1855
1856
1857
1858
1859
1860
1861
1862
1863
        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]

1864
1865
1866
1867
1868
1869
1870
1871
1872
1873
        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)
            )

1874
        return attn_metadata, spec_decode_common_attn_metadata
1875

1876
1877
1878
    def _compute_cascade_attn_prefix_lens(
        self,
        num_scheduled_tokens: np.ndarray,
1879
        num_computed_tokens: np.ndarray,
1880
1881
1882
1883
1884
1885
1886
        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
        """
1887

1888
1889
1890
1891
1892
1893
1894
1895
1896
1897
1898
1899
1900
1901
        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,
1902
                        num_computed_tokens,
1903
1904
1905
1906
1907
1908
1909
1910
                        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
1911

1912
1913
1914
    def _compute_cascade_attn_prefix_len(
        self,
        num_scheduled_tokens: np.ndarray,
1915
        num_computed_tokens: np.ndarray,
1916
        num_common_prefix_blocks: int,
1917
1918
        kv_cache_spec: KVCacheSpec,
        attn_metadata_builder: AttentionMetadataBuilder,
1919
1920
1921
1922
1923
1924
1925
1926
1927
1928
1929
1930
1931
1932
1933
1934
1935
1936
    ) -> 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.
        """
1937

1938
        common_prefix_len = num_common_prefix_blocks * kv_cache_spec.block_size
1939
1940
1941
1942
1943
1944
1945
1946
1947
1948
1949
1950
1951
1952
1953
1954
1955
1956
1957
1958
1959
1960
1961
1962
1963
1964
1965
1966
1967
1968
1969
1970
1971
1972
1973
1974
1975
        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]
1976
        # Request 3's num_computed_tokens: 3 (i.e., [A, B, C])
1977
1978
1979
1980
1981
        # 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.
1982
        common_prefix_len = min(common_prefix_len, num_computed_tokens.min())
1983
        # common_prefix_len should be a multiple of the block size.
1984
1985
1986
1987
1988
1989
1990
1991
1992
1993
1994
        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
        )
1995
1996
        assert isinstance(kv_cache_spec, AttentionSpec)
        use_cascade = attn_metadata_builder.use_cascade_attention(
1997
1998
1999
            common_prefix_len=common_prefix_len,
            query_lens=num_scheduled_tokens,
            num_query_heads=self.num_query_heads,
2000
            num_kv_heads=kv_cache_spec.num_kv_heads,
2001
            use_alibi=self.use_alibi,
2002
            use_sliding_window=use_sliding_window,
2003
            use_local_attention=use_local_attention,
2004
            num_sms=self.num_sms,
2005
            dcp_world_size=self.dcp_world_size,
2006
2007
2008
        )
        return common_prefix_len if use_cascade else 0

2009
2010
    def _calc_mrope_positions(self, scheduler_output: "SchedulerOutput"):
        mrope_pos_ptr = 0
2011
        for index, req_id in enumerate(self.input_batch.req_ids):
2012
2013
2014
            req = self.requests[req_id]
            assert req.mrope_positions is not None

2015
2016
            num_computed_tokens = self.input_batch.num_computed_tokens_cpu[index]
            num_scheduled_tokens = scheduler_output.num_scheduled_tokens[req_id]
2017
            num_prompt_tokens = length_from_prompt_token_ids_or_embeds(
2018
2019
                req.prompt_token_ids, req.prompt_embeds
            )
2020
2021

            if num_computed_tokens + num_scheduled_tokens > num_prompt_tokens:
2022
2023
                prompt_part_len = max(0, num_prompt_tokens - num_computed_tokens)
                completion_part_len = max(0, num_scheduled_tokens - prompt_part_len)
2024
2025
2026
2027
2028
2029
2030
2031
2032
2033
2034
2035
2036
            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

2037
2038
2039
                self.mrope_positions.cpu[:, dst_start:dst_end] = req.mrope_positions[
                    :, src_start:src_end
                ]
2040
2041
2042
2043
2044
2045
2046
                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

2047
                assert req.mrope_position_delta is not None
2048
                MRotaryEmbedding.get_next_input_positions_tensor(
2049
                    out=self.mrope_positions.np,
2050
2051
2052
2053
2054
                    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,
                )
2055
2056
2057

                mrope_pos_ptr += completion_part_len

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
2089
2090
2091
2092
2093
2094
2095
2096
2097
2098
2099
2100
2101
2102
2103
2104
    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

2105
2106
    def _calc_spec_decode_metadata(
        self,
2107
2108
2109
2110
2111
2112
2113
2114
2115
2116
2117
2118
2119
2120
2121
2122
        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
2123
2124
2125
2126

        # 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(
2127
2128
            num_sampled_tokens, cumsum_dtype=np.int32
        )
2129
        # Step 2. [0, 0, 0, 0, 103, 104, 104, 104, 206, 207, 207]
2130
        logits_indices = np.repeat(
2131
2132
            cu_num_scheduled_tokens - num_sampled_tokens, num_sampled_tokens
        )
2133
        # Step 3. [0, 1, 2, 3, 103, 104, 105, 106, 206, 207, 208]
2134
2135
2136
2137
2138
2139
        logits_indices += arange

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

        # Compute the draft logits indices.
2140
2141
2142
        # 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(
2143
2144
            num_draft_tokens, cumsum_dtype=np.int32
        )
2145
2146
        # [0, 0, 0, 5, 5, 9]
        target_logits_indices = np.repeat(
2147
2148
            cu_num_sampled_tokens - num_sampled_tokens, num_draft_tokens
        )
2149
2150
2151
2152
2153
        # [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(
2154
2155
            self.device, non_blocking=True
        )
2156
2157
2158
        cu_num_sampled_tokens = torch.from_numpy(cu_num_sampled_tokens).to(
            self.device, non_blocking=True
        )
2159
2160
2161
        logits_indices = torch.from_numpy(logits_indices).to(
            self.device, non_blocking=True
        )
2162
        target_logits_indices = torch.from_numpy(target_logits_indices).to(
2163
2164
            self.device, non_blocking=True
        )
2165
        bonus_logits_indices = torch.from_numpy(bonus_logits_indices).to(
2166
2167
            self.device, non_blocking=True
        )
2168

2169
2170
        # Compute the draft token ids.
        # draft_token_indices:      [  1,   2,   3, 105, 106, 208]
2171
        draft_token_ids = self.input_ids.gpu[logits_indices]
2172
2173
        draft_token_ids = draft_token_ids[target_logits_indices + 1]

2174
        return SpecDecodeMetadata(
2175
2176
2177
            draft_token_ids=draft_token_ids,
            num_draft_tokens=num_draft_tokens.tolist(),
            cu_num_draft_tokens=cu_num_draft_tokens,
2178
            cu_num_sampled_tokens=cu_num_sampled_tokens,
2179
2180
2181
2182
2183
            target_logits_indices=target_logits_indices,
            bonus_logits_indices=bonus_logits_indices,
            logits_indices=logits_indices,
        )

2184
2185
2186
2187
2188
2189
2190
    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
2191
        self.kv_sharing_fast_prefill_logits_indices[:num_logits].copy_(logits_indices)
2192
2193
2194
2195
2196
        # 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_(
2197
2198
            logits_indices[-1].item()
        )
2199
2200
2201
2202
2203
        # Dispatch for the decoder portion of the model.
        _, batch_desc = self.cudagraph_dispatcher.dispatch(
            num_logits, disable_full=True
        )
        num_logits_padded = batch_desc.num_tokens
2204
2205
2206
        logits_indices_padded = self.kv_sharing_fast_prefill_logits_indices[
            :num_logits_padded
        ]
2207
2208
        return logits_indices_padded

2209
    def _batch_mm_inputs_from_scheduler(
2210
2211
        self,
        scheduler_output: "SchedulerOutput",
2212
2213
2214
2215
2216
    ) -> tuple[
        list[str],
        list[MultiModalKwargsItem],
        list[tuple[str, PlaceholderRange]],
    ]:
2217
        """Batch multimodal inputs from scheduled encoder inputs.
2218
2219
2220

        Args:
            scheduler_output: The scheduler output containing scheduled encoder
2221
                inputs.
2222
2223

        Returns:
2224
            A tuple of (mm_hashes, mm_kwargs, mm_lora_refs) where:
2225
2226
            - mm_hashes: List of multimodal hashes for each item
            - mm_kwargs: List of multimodal kwargs for each item
2227
            - mm_lora_refs: List of (req_id, placeholder_range) for each item
2228
        """
2229
2230
        scheduled_encoder_inputs = scheduler_output.scheduled_encoder_inputs
        if not scheduled_encoder_inputs:
2231
            return [], [], []
2232
2233

        mm_hashes = list[str]()
2234
        mm_kwargs = list[MultiModalKwargsItem]()
2235
2236
2237
        # Multimodal LoRA reference info to map each multimodal item
        # back to its request & position
        mm_lora_refs = list[tuple[str, PlaceholderRange]]()
2238
2239
        for req_id, encoder_input_ids in scheduled_encoder_inputs.items():
            req_state = self.requests[req_id]
2240
2241

            for mm_input_id in encoder_input_ids:
2242
                mm_feature = req_state.mm_features[mm_input_id]
2243
2244
                if mm_feature.data is None:
                    continue
2245
2246

                mm_hashes.append(mm_feature.identifier)
2247
                mm_kwargs.append(mm_feature.data)
2248
                mm_lora_refs.append((req_id, mm_feature.mm_position))
2249

2250
        return mm_hashes, mm_kwargs, mm_lora_refs
2251

2252
2253
2254
    def _execute_mm_encoder(
        self, scheduler_output: "SchedulerOutput"
    ) -> list[torch.Tensor]:
2255
2256
2257
        mm_hashes, mm_kwargs, mm_lora_refs = self._batch_mm_inputs_from_scheduler(
            scheduler_output
        )
2258
2259

        if not mm_kwargs:
2260
            return []
2261

2262
2263
2264
2265
2266
2267
        should_time = bool(
            self.observability_config
            and self.observability_config.enable_mm_processor_stats
            and scheduler_output.scheduled_encoder_inputs
        )

2268
2269
2270
2271
2272
2273
2274
        # 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.
2275
        model = cast(SupportsMultiModal, self.model)
2276
2277
2278
2279
2280
2281
2282
2283
2284
2285
2286
2287
2288
2289
2290
2291
2292
2293
2294
2295
2296
2297
2298
2299
2300
2301
2302
2303
2304
2305
2306
2307
2308
2309
2310
2311
2312
2313
2314
2315
2316
2317
2318
2319
2320
2321
2322
2323
2324
2325
2326
2327
2328
2329
2330
2331
2332

        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]
                    pos_info.get_num_embeds
                )
                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,
                )

2333
        encoder_outputs: list[torch.Tensor] = []
2334
2335
        # Track the current index in mm_kwargs/mm_lora_refs to map groups to request IDs
        current_item_idx = 0
2336
        for modality, num_items, mm_kwargs_group in group_mm_kwargs_by_modality(
2337
2338
2339
            mm_kwargs,
            device=self.device,
            pin_memory=self.pin_memory,
2340
        ):
2341
            curr_group_outputs: MultiModalEmbeddings
2342
2343

            # EVS-related change.
2344
            # (ekhvedchenia): Temporary hack to limit peak memory usage when
2345
            # processing multimodal data. This solves the issue with scheduler
2346
2347
2348
2349
            # 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)
2350
2351
2352
2353
2354
2355
2356
            # 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
            ):
2357
                curr_group_outputs_lst = list[torch.Tensor]()
2358
2359
2360
2361
2362
2363
2364
2365
2366
2367
2368
                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,
                            )
2369
                        )
2370

2371
2372
2373
                        micro_batch_outputs = model.embed_multimodal(
                            **micro_batch_mm_inputs
                        )
2374

2375
                        curr_group_outputs_lst.extend(micro_batch_outputs)
2376
2377

                curr_group_outputs = curr_group_outputs_lst
2378
2379
2380
2381
2382
2383
2384
2385
            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.
2386
2387
2388
2389
2390

                with self.timed_encoder_operation(
                    should_time, mm_lora_refs, current_item_idx, num_items
                ):
                    curr_group_outputs = model.embed_multimodal(**mm_kwargs_group)
2391

2392
2393
            sanity_check_mm_encoder_outputs(
                curr_group_outputs,
2394
                expected_num_items=num_items,
2395
            )
2396
            encoder_outputs.extend(curr_group_outputs)
2397

2398
2399
            current_item_idx += num_items

2400
        # Cache the encoder outputs by mm_hash
2401
        for mm_hash, output in zip(mm_hashes, encoder_outputs):
2402
            self.encoder_cache[mm_hash] = output
2403
2404
            logger.debug("Finish execute for mm hash %s", mm_hash)
            self.maybe_save_ec_to_connector(self.encoder_cache, mm_hash)
2405

2406
2407
        return encoder_outputs

2408
    def _gather_mm_embeddings(
2409
2410
        self,
        scheduler_output: "SchedulerOutput",
2411
        shift_computed_tokens: int = 0,
2412
2413
2414
    ) -> tuple[list[torch.Tensor], torch.Tensor]:
        total_num_scheduled_tokens = scheduler_output.total_num_scheduled_tokens

2415
2416
2417
2418
2419
        # 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]

2420
        mm_embeds = list[torch.Tensor]()
2421
        is_mm_embed = is_mm_embed_buf.cpu
2422
2423
2424
        is_mm_embed[:total_num_scheduled_tokens] = False

        req_start_idx = 0
2425
        should_sync_mrope_positions = False
2426
        should_sync_xdrope_positions = False
2427

2428
        for req_id in self.input_batch.req_ids:
2429
2430
            mm_embeds_req: list[torch.Tensor] = []

2431
            num_scheduled_tokens = scheduler_output.num_scheduled_tokens[req_id]
2432
            req_state = self.requests[req_id]
2433
            num_computed_tokens = req_state.num_computed_tokens + shift_computed_tokens
2434

2435
2436
            for mm_feature in req_state.mm_features:
                pos_info = mm_feature.mm_position
2437
2438
                start_pos = pos_info.offset
                num_encoder_tokens = pos_info.length
2439
2440
2441
2442
2443
2444
2445
2446
2447
2448
2449
2450
2451
2452
2453
2454

                # 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,
2455
2456
                    num_encoder_tokens,
                )
2457
                assert start_idx < end_idx
2458
2459
2460
2461
2462
2463
2464
                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
2465

2466
                mm_hash = mm_feature.identifier
2467
                encoder_output = self.encoder_cache.get(mm_hash, None)
2468
                assert encoder_output is not None, f"Encoder cache miss for {mm_hash}."
2469
2470
2471

                if (is_embed := pos_info.is_embed) is not None:
                    is_embed = is_embed[start_idx:end_idx]
2472
2473
2474
                    mm_embeds_item = encoder_output[curr_embeds_start:curr_embeds_end]
                else:
                    mm_embeds_item = encoder_output[start_idx:end_idx]
2475

2476
                req_start_pos = req_start_idx + start_pos - num_computed_tokens
2477
2478
2479
                is_mm_embed[req_start_pos + start_idx : req_start_pos + end_idx] = (
                    True if is_embed is None else is_embed
                )
2480
2481
2482
                mm_embeds_req.append(mm_embeds_item)

            if self.is_multimodal_pruning_enabled and self.uses_mrope:
2483
                assert req_state.mrope_positions is not None
2484
2485
2486
2487
2488
2489
2490
                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,
2491
2492
                    )
                )
2493
2494
2495
2496
                req_state.mrope_positions.copy_(new_mrope_positions)
                req_state.mrope_position_delta = new_delta

            mm_embeds.extend(mm_embeds_req)
2497
2498
            req_start_idx += num_scheduled_tokens

2499
        is_mm_embed = is_mm_embed_buf.copy_to_gpu(total_num_scheduled_tokens)
2500
2501
2502

        if should_sync_mrope_positions:
            self._calc_mrope_positions(scheduler_output)
2503
            self.mrope_positions.copy_to_gpu(total_num_scheduled_tokens)
2504

2505
2506
2507
2508
        if should_sync_xdrope_positions:
            self._calc_xdrope_positions(scheduler_output)
            self.xdrope_positions.copy_to_gpu(total_num_scheduled_tokens)

2509
        return mm_embeds, is_mm_embed
2510

2511
    def get_model(self) -> nn.Module:
2512
        # get raw model out of the cudagraph wrapper.
2513
        if isinstance(self.model, (CUDAGraphWrapper, UBatchWrapper)):
2514
            return self.model.unwrap()
2515
2516
        return self.model

2517
2518
2519
2520
2521
2522
2523
2524
2525
2526
2527
2528
2529
2530
2531
    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")

        return supported_tasks

2532
2533
2534
2535
2536
    def get_supported_pooling_tasks(self) -> list[PoolingTask]:
        model = self.get_model()
        if not is_pooling_model(model):
            return []

2537
2538
        supported_tasks = list(model.pooler.get_supported_tasks())

2539
2540
2541
2542
        if "score" in supported_tasks:
            num_labels = getattr(self.model_config.hf_config, "num_labels", 0)
            if num_labels != 1:
                supported_tasks.remove("score")
2543
                logger.debug_once("Score API is only enabled for num_labels == 1.")
2544
2545

        return supported_tasks
2546

2547
2548
2549
2550
2551
2552
2553
2554
2555
2556
    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)

2557
    def sync_and_slice_intermediate_tensors(
2558
2559
        self,
        num_tokens: int,
2560
        intermediate_tensors: IntermediateTensors | None,
2561
2562
        sync_self: bool,
    ) -> IntermediateTensors:
2563
2564
2565
        assert self.intermediate_tensors is not None

        tp = self.vllm_config.parallel_config.tensor_parallel_size
2566
        is_rs = is_residual_scattered_for_sp(self.vllm_config, num_tokens)
2567
2568
2569
2570
2571
2572

        # 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():
2573
                is_scattered = k == "residual" and is_rs
2574
                copy_len = num_tokens // tp if is_scattered else num_tokens
2575
                self.intermediate_tensors[k][:copy_len].copy_(
2576
2577
2578
2579
2580
2581
2582
2583
2584
2585
2586
2587
2588
                    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:
2589
2590
2591
2592
2593
2594
2595
        """
        Step for the EPLB (Expert Parallelism Load Balancing) state.
        """
        if not self.parallel_config.enable_eplb:
            return

        assert self.eplb_state is not None
2596
2597
        model = self.get_model()
        assert is_mixture_of_experts(model)
2598
2599
2600
        self.eplb_state.step(
            is_dummy,
            is_profile,
2601
            log_stats=self.parallel_config.eplb_config.log_balancedness,
2602
2603
        )

2604
2605
2606
2607
2608
    def _pool(
        self,
        hidden_states: torch.Tensor,
        num_scheduled_tokens: int,
        num_scheduled_tokens_np: np.ndarray,
2609
2610
2611
2612
        kv_connector_output: KVConnectorOutput | None,
    ) -> ModelRunnerOutput | AsyncModelRunnerOutput:
        num_reqs = self.input_batch.num_reqs
        assert num_reqs == len(self.input_batch.pooling_params), (
2613
2614
            "Either all or none of the requests in a batch must be pooling request"
        )
2615

2616
        hidden_states = hidden_states[:num_scheduled_tokens]
2617
        seq_lens_cpu = self.seq_lens.cpu[:num_reqs]
2618

2619
        pooling_metadata = self.input_batch.get_pooling_metadata()
2620
        pooling_metadata.build_pooling_cursor(
2621
            num_scheduled_tokens_np, seq_lens_cpu, device=hidden_states.device
2622
        )
2623

2624
2625
        model = cast(VllmModelForPooling, self.model)
        raw_pooler_output: PoolerOutput = model.pooler(
2626
            hidden_states=hidden_states, pooling_metadata=pooling_metadata
2627
        )
2628
2629
2630
2631
2632
2633
2634
2635
2636
2637
2638
2639
2640
2641
2642
2643
2644
2645
2646
2647
2648
2649
2650
2651

        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,
            )

2652
        raw_pooler_output = json_map_leaves(
2653
            lambda x: None if x is None else x.to("cpu", non_blocking=True),
2654
2655
            raw_pooler_output,
        )
2656
2657
2658
2659
        model_runner_output.pooler_output = [
            out if include else None
            for out, include in zip(raw_pooler_output, finished_mask)
        ]
2660
2661
        self._sync_device()

2662
        return model_runner_output
2663

2664
    def _pad_for_sequence_parallelism(self, num_scheduled_tokens: int) -> int:
2665
2666
2667
        # Pad tokens to multiple of tensor_parallel_size when
        # enabled collective fusion for SP
        tp_size = self.vllm_config.parallel_config.tensor_parallel_size
2668
        if self.compilation_config.pass_config.enable_sp and tp_size > 1:
2669
2670
2671
            return round_up(num_scheduled_tokens, tp_size)
        return num_scheduled_tokens

Patrick von Platen's avatar
Patrick von Platen committed
2672
2673
2674
2675
2676
2677
2678
2679
2680
2681
2682
    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

2683
    def _preprocess(
2684
2685
        self,
        scheduler_output: "SchedulerOutput",
2686
        num_input_tokens: int,  # Padded
2687
        intermediate_tensors: IntermediateTensors | None = None,
2688
    ) -> tuple[
2689
2690
        torch.Tensor | None,
        torch.Tensor | None,
2691
        torch.Tensor,
2692
        IntermediateTensors | None,
2693
        dict[str, Any],
2694
        ECConnectorOutput | None,
2695
    ]:
2696
        num_scheduled_tokens = scheduler_output.total_num_scheduled_tokens
2697
        is_first_rank = get_pp_group().is_first_rank
2698
        is_encoder_decoder = self.model_config.is_encoder_decoder
2699

2700
2701
        # _prepare_inputs may reorder the batch, so we must gather multi
        # modal outputs after that to ensure the correct order
2702
2703
        ec_connector_output = None

2704
        if self.supports_mm_inputs and is_first_rank and not is_encoder_decoder:
2705
            # Run the multimodal encoder if any.
2706
2707
2708
2709
2710
2711
            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)
2712

2713
2714
2715
            # 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.
2716
            inputs_embeds_scheduled = self.model.embed_input_ids(
2717
2718
2719
                self.input_ids.gpu[:num_scheduled_tokens],
                multimodal_embeddings=mm_embeds,
                is_multimodal=is_mm_embed,
2720
            )
2721

2722
            # TODO(woosuk): Avoid the copy. Optimize.
2723
            self.inputs_embeds.gpu[:num_scheduled_tokens].copy_(inputs_embeds_scheduled)
2724

Patrick von Platen's avatar
Patrick von Platen committed
2725
            input_ids, inputs_embeds = self._prepare_mm_inputs(num_input_tokens)
2726
            model_kwargs = {
2727
                **self._init_model_kwargs(),
2728
2729
                **self._extract_mm_kwargs(scheduler_output),
            }
2730
        elif self.enable_prompt_embeds and is_first_rank:
2731
2732
2733
2734
2735
2736
2737
2738
2739
2740
2741
2742
            # 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).
2743
2744
2745
            token_ids_idx = (
                self.is_token_ids.gpu[:num_scheduled_tokens]
                .nonzero(as_tuple=False)
2746
                .squeeze(1)
2747
            )
2748
2749
2750
            # Some tokens ids may need to become embeds
            if token_ids_idx.numel() > 0:
                token_ids = self.input_ids.gpu[token_ids_idx]
2751
                tokens_to_embeds = self.model.embed_input_ids(input_ids=token_ids)
2752
2753
2754
                self.inputs_embeds.gpu[token_ids_idx] = tokens_to_embeds

            inputs_embeds = self.inputs_embeds.gpu[:num_input_tokens]
2755
            model_kwargs = self._init_model_kwargs()
2756
            input_ids = None
2757
        else:
2758
2759
2760
2761
            # 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.
2762
            input_ids = self.input_ids.gpu[:num_input_tokens]
2763
            inputs_embeds = None
2764
            model_kwargs = self._init_model_kwargs()
2765

2766
        if self.uses_mrope:
2767
            positions = self.mrope_positions.gpu[:, :num_input_tokens]
2768
2769
        elif self.uses_xdrope_dim > 0:
            positions = self.xdrope_positions.gpu[:, :num_input_tokens]
2770
        else:
2771
            positions = self.positions.gpu[:num_input_tokens]
2772

2773
        if is_first_rank:
2774
2775
            intermediate_tensors = None
        else:
2776
            assert intermediate_tensors is not None
2777
            intermediate_tensors = self.sync_and_slice_intermediate_tensors(
2778
2779
                num_input_tokens, intermediate_tensors, True
            )
2780

2781
        if is_encoder_decoder and scheduler_output.scheduled_encoder_inputs:
2782
2783
2784
2785
2786
2787
2788
            # 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})
2789

2790
2791
2792
2793
2794
2795
        return (
            input_ids,
            inputs_embeds,
            positions,
            intermediate_tensors,
            model_kwargs,
2796
            ec_connector_output,
2797
        )
2798

2799
    def _sample(
2800
        self,
2801
2802
        logits: torch.Tensor | None,
        spec_decode_metadata: SpecDecodeMetadata | None,
2803
    ) -> SamplerOutput:
2804
        # Sample the next token and get logprobs if needed.
2805
        sampling_metadata = self.input_batch.sampling_metadata
2806
2807
2808
        # 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()
2809
        if spec_decode_metadata is None:
2810
            return self.sampler(
2811
2812
2813
                logits=logits,
                sampling_metadata=sampling_metadata,
            )
2814

2815
2816
2817
2818
2819
2820
        # 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)

2821
        sampler_output = self.rejection_sampler(
2822
2823
            spec_decode_metadata,
            None,  # draft_probs
2824
            logits,
2825
2826
            sampling_metadata,
        )
2827
2828
2829
        return sampler_output

    def _bookkeeping_sync(
2830
2831
2832
        self,
        scheduler_output: "SchedulerOutput",
        sampler_output: SamplerOutput,
2833
        logits: torch.Tensor | None,
2834
2835
        hidden_states: torch.Tensor,
        num_scheduled_tokens: int,
2836
        spec_decode_metadata: SpecDecodeMetadata | None,
2837
    ) -> tuple[
2838
        dict[str, int],
2839
        LogprobsLists | None,
2840
        list[list[int]],
2841
        dict[str, LogprobsTensors | None],
2842
2843
2844
        list[str],
        dict[str, int],
        list[int],
2845
    ]:
2846
2847
2848
2849
        num_nans_in_logits = {}
        if envs.VLLM_COMPUTE_NANS_IN_LOGITS:
            num_nans_in_logits = self._get_nans_in_logits(logits)

2850
2851
2852
2853
        num_reqs = self.input_batch.num_reqs
        discard_sampled_tokens_req_indices = np.nonzero(
            self.discard_request_mask.np[:num_reqs]
        )[0]
2854
2855
2856
2857
        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)
2858

2859
2860
2861
        # 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()
2862
        req_id_to_index_output_copy = self.input_batch.req_id_to_index.copy()
2863
2864

        num_sampled_tokens = sampler_output.sampled_token_ids.shape[0]
2865
        sampled_token_ids = sampler_output.sampled_token_ids
2866
        logprobs_tensors = sampler_output.logprobs_tensors
2867
        invalid_req_indices = []
2868
        logprobs_lists = None
2869
2870
2871
2872
2873
2874
        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)
2875
2876
2877
                # 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()
2878
2879
2880

                if logprobs_tensors is not None:
                    logprobs_lists = logprobs_tensors.tolists()
2881
2882
            else:
                # Includes spec decode tokens.
2883
                valid_sampled_token_ids, logprobs_lists = RejectionSampler.parse_output(
2884
2885
                    sampled_token_ids,
                    self.input_batch.vocab_size,
2886
                    discard_sampled_tokens_req_indices,
2887
                    logprobs_tensors=logprobs_tensors,
2888
                )
2889
        else:
2890
            valid_sampled_token_ids = []
2891
            invalid_req_indices = discard_sampled_tokens_req_indices.tolist()
2892
2893
2894
2895
2896
            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.
2897
2898
2899
2900
            # 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
2901
2902
2903
2904
2905
            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
            }
2906

2907
2908
2909
2910
2911
        # 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.
2912
        req_ids = self.input_batch.req_ids
2913
2914
        for req_idx in range(num_sampled_tokens):
            if self.use_async_scheduling:
2915
                sampled_ids = [-1] if req_idx not in invalid_req_indices_set else None
2916
2917
            else:
                sampled_ids = valid_sampled_token_ids[req_idx]
2918

2919
            num_sampled_ids: int = len(sampled_ids) if sampled_ids else 0
2920

2921
            if not sampled_ids:
2922
2923
2924
                continue

            start_idx = self.input_batch.num_tokens_no_spec[req_idx]
2925
            end_idx = start_idx + num_sampled_ids
2926
2927
2928
2929
            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}"
2930
            )
2931

2932
2933
            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
2934
            self.input_batch.num_tokens_no_spec[req_idx] = end_idx
2935

2936
            req_id = req_ids[req_idx]
2937
2938
2939
            req_state = self.requests[req_id]
            req_state.output_token_ids.extend(sampled_ids)

2940
2941
2942
2943
2944
2945
        # Compute prompt logprobs if needed.
        prompt_logprobs_dict = self._get_prompt_logprobs_dict(
            hidden_states[:num_scheduled_tokens],
            scheduler_output.num_scheduled_tokens,
        )

2946
2947
2948
2949
2950
2951
2952
2953
2954
2955
        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,
        )

2956
2957
2958
2959
2960
2961
2962
2963
2964
2965
2966
2967
2968
2969
2970
    @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()

2971
2972
    def _model_forward(
        self,
2973
2974
2975
2976
        input_ids: torch.Tensor | None = None,
        positions: torch.Tensor | None = None,
        intermediate_tensors: IntermediateTensors | None = None,
        inputs_embeds: torch.Tensor | None = None,
2977
2978
2979
2980
2981
        **model_kwargs: dict[str, Any],
    ) -> Any:
        """Helper method to call the model forward pass.

        This method can be overridden by subclasses for model execution.
2982
        Motivation: We can inspect only this method versus
2983
2984
2985
2986
2987
2988
2989
2990
2991
2992
2993
2994
2995
2996
2997
2998
2999
3000
3001
3002
        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,
        )

3003
3004
3005
3006
3007
3008
3009
3010
3011
3012
3013
3014
3015
3016
3017
3018
3019
3020
3021
3022
3023
    @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
        )

3024
3025
3026
3027
3028
3029
3030
3031
3032
3033
3034
3035
3036
    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,
3037
        num_encoder_reqs: int = 0,
3038
    ) -> tuple[
3039
3040
        CUDAGraphMode,
        BatchDescriptor,
3041
        bool,
3042
3043
        torch.Tensor | None,
        CUDAGraphStat | None,
3044
    ]:
3045
3046
3047
3048
3049
3050
        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,
3051
        )
3052
3053
3054
3055
3056
        # 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
        )
3057
3058
3059
3060
3061
3062
3063

        has_lora = (
            len(self.input_batch.lora_id_to_lora_request) > 0
            if force_has_lora is None
            else force_has_lora
        )

3064
        num_tokens_padded = self._pad_for_sequence_parallelism(num_tokens)
3065
        dispatch_cudagraph = (
3066
            lambda num_tokens, disable_full: self.cudagraph_dispatcher.dispatch(
3067
3068
3069
                num_tokens=num_tokens,
                has_lora=has_lora,
                uniform_decode=uniform_decode,
3070
                disable_full=disable_full,
3071
3072
3073
3074
3075
            )
            if not force_eager
            else (CUDAGraphMode.NONE, BatchDescriptor(num_tokens_padded))
        )

3076
        cudagraph_mode, batch_descriptor = dispatch_cudagraph(
3077
            num_tokens_padded, use_cascade_attn or has_encoder_output
3078
        )
3079
        num_tokens_padded = batch_descriptor.num_tokens
3080
3081
3082
3083
3084
3085
3086
3087
3088
        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"
            )
3089
3090
3091

        # Extra coordination when running data-parallel since we need to coordinate
        # across ranks
3092
        should_ubatch, num_tokens_across_dp = False, None
3093
3094
3095
3096
3097
3098
3099
3100
3101
        if self.vllm_config.parallel_config.data_parallel_size > 1:
            # Disable DP padding when running eager to avoid excessive padding when
            # running prefills. This lets us set cudagraph_mode="NONE" on the prefiller
            # in a P/D setup and still use CUDA graphs (enabled by this padding) on the
            # decoder.
            allow_dp_padding = (
                self.compilation_config.cudagraph_mode != CUDAGraphMode.NONE
            )

3102
3103
3104
3105
3106
3107
3108
3109
3110
3111
3112
            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,
                    allow_dp_padding=allow_dp_padding,
                    num_tokens_padded=num_tokens_padded,
                    uniform_decode=uniform_decode,
                    num_scheduled_tokens_per_request=num_scheduled_tokens_np,
                    cudagraph_mode=cudagraph_mode.value,
                )
3113
3114
            )

3115
            # Extract DP-synced values
3116
3117
3118
            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())
3119
3120
3121
3122
3123
                # Re-dispatch with DP padding so we have the correct batch_descriptor
                cudagraph_mode, batch_descriptor = dispatch_cudagraph(
                    num_tokens_padded,
                    disable_full=synced_cudagraph_mode <= CUDAGraphMode.PIECEWISE.value,
                )
3124
3125
3126
3127
                # 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

3128
3129
3130
3131
3132
3133
3134
3135
3136
3137
3138
3139
        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,
3140
            should_ubatch,
3141
3142
3143
            num_tokens_across_dp,
            cudagraph_stats,
        )
3144

3145
3146
3147
3148
3149
3150
3151
3152
3153
3154
3155
3156
3157
3158
3159
3160
3161
3162
3163
3164
3165
3166
3167
3168
3169
3170
3171
3172
3173
3174
3175
3176
3177
3178
3179
3180
    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

3181
3182
3183
3184
3185
3186
3187
3188
3189
3190
3191
3192
3193
3194
3195
3196
3197
3198
3199
3200
3201
3202
3203
3204
3205
3206
3207
3208
3209
3210
3211
3212
3213
3214
3215
3216
3217
3218
3219
3220
3221
3222
3223
3224
3225
3226
3227
3228
3229
3230
3231
3232
3233
3234
3235
3236
3237
3238
3239
3240
3241
3242
3243
3244
3245
3246
3247
3248
3249
3250
3251
3252
3253
3254
    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

3255
3256
3257
3258
    @torch.inference_mode()
    def execute_model(
        self,
        scheduler_output: "SchedulerOutput",
3259
        intermediate_tensors: IntermediateTensors | None = None,
3260
    ) -> ModelRunnerOutput | AsyncModelRunnerOutput | IntermediateTensors | None:
3261
3262
3263
3264
3265
        if self.execute_model_state is not None:
            raise RuntimeError(
                "State error: sample_tokens() must be called "
                "after execute_model() returns None."
            )
3266

3267
3268
3269
3270
3271
3272
3273
        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.")

3274
3275
3276
3277
3278
        if scheduler_output.preempted_req_ids and has_kv_transfer_group():
            get_kv_transfer_group().handle_preemptions(
                scheduler_output.preempted_req_ids
            )

3279
        num_scheduled_tokens = scheduler_output.total_num_scheduled_tokens
3280
3281
3282
3283
3284
3285
        with (
            record_function_or_nullcontext("gpu_model_runner: preprocess"),
            self.synchronize_input_prep(),
        ):
            # Update persistent batch states.
            self._update_states(scheduler_output)
3286

3287
3288
            if has_ec_transfer() and get_ec_transfer().is_producer:
                with self.maybe_get_ec_connector_output(
3289
                    scheduler_output,
3290
3291
3292
3293
3294
3295
3296
3297
3298
3299
3300
3301
3302
3303
3304
3305
3306
3307
3308
3309
3310
3311
3312
3313
3314
3315
3316
3317
                    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"
3318
3319
                )

3320
3321
3322
3323
3324
3325
            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
3326

3327
3328
3329
3330
            logits_indices, spec_decode_metadata = self._prepare_inputs(
                scheduler_output,
                num_scheduled_tokens_np,
            )
3331

3332
3333
3334
3335
3336
            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(
3337
                    num_scheduled_tokens_np,
3338
3339
                    self.input_batch.num_computed_tokens_cpu[:num_reqs],
                    scheduler_output.num_common_prefix_blocks,
3340
3341
                )

3342
3343
3344
3345
3346
3347
3348
3349
3350
3351
3352
3353
3354
3355
            (
                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),
            )
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
            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,
            )

3384
3385
3386
3387
3388
3389
3390
3391
3392
3393
3394
            # 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)
            )
3395
3396
            pad_attn = cudagraph_mode == CUDAGraphMode.FULL

3397
3398
3399
3400
3401
3402
3403
3404
3405
3406
3407
3408
            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(),
                )

3409
3410
3411
            use_spec_decode = len(scheduler_output.scheduled_spec_decode_tokens) > 0
            ubatch_slices_attn = ubatch_slices_padded if pad_attn else ubatch_slices

3412
3413
3414
3415
3416
3417
3418
3419
3420
3421
3422
            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,
            )

3423
3424
3425
3426
3427
3428
3429
3430
3431
3432
3433
3434
            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,
3435
                    slot_mappings=slot_mappings_by_group,
3436
                )
3437
            )
3438

3439
3440
3441
3442
3443
3444
3445
3446
3447
            (
                input_ids,
                inputs_embeds,
                positions,
                intermediate_tensors,
                model_kwargs,
                ec_connector_output,
            ) = self._preprocess(
                scheduler_output, num_tokens_padded, intermediate_tensors
3448
            )
3449

3450
        # Set cudagraph mode to none if calc_kv_scales is true.
3451
3452
3453
        # KV scales calculation involves dynamic operations that are incompatible
        # with CUDA graph capture.
        if self.calculate_kv_scales:
3454
            cudagraph_mode = CUDAGraphMode.NONE
3455
3456
            # Mark KV scales as calculated after the first forward pass
            self.calculate_kv_scales = False
3457

3458
3459
3460
3461
3462
3463
3464
        # 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
        )

3465
3466
        # Run the model.
        # Use persistent buffers for CUDA graphs.
3467
3468
        with (
            set_forward_context(
3469
3470
                attn_metadata,
                self.vllm_config,
3471
                num_tokens=num_tokens_padded,
3472
                num_tokens_across_dp=num_tokens_across_dp,
3473
3474
                cudagraph_runtime_mode=cudagraph_mode,
                batch_descriptor=batch_desc,
3475
                ubatch_slices=ubatch_slices_padded,
3476
                slot_mapping=slot_mappings,
3477
                skip_compiled=has_encoder_input,
3478
            ),
3479
            record_function_or_nullcontext("gpu_model_runner: forward"),
3480
3481
            self.maybe_get_kv_connector_output(scheduler_output) as kv_connector_output,
        ):
3482
            model_output = self._model_forward(
3483
3484
3485
3486
3487
3488
3489
                input_ids=input_ids,
                positions=positions,
                intermediate_tensors=intermediate_tensors,
                inputs_embeds=inputs_embeds,
                **model_kwargs,
            )

3490
        with record_function_or_nullcontext("gpu_model_runner: postprocess"):
3491
            if self.use_aux_hidden_state_outputs:
3492
                # True when EAGLE 3 is used.
3493
3494
                hidden_states, aux_hidden_states = model_output
            else:
3495
                # Common case.
3496
3497
3498
                hidden_states = model_output
                aux_hidden_states = None

3499
3500
3501
3502
3503
            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)
3504
                    hidden_states.kv_connector_output = kv_connector_output
3505
                    self.kv_connector_output = kv_connector_output
3506
                    return hidden_states
3507

3508
                if self.is_pooling_model:
3509
                    # Return the pooling output.
3510
3511
3512
3513
3514
                    return self._pool(
                        hidden_states,
                        num_scheduled_tokens,
                        num_scheduled_tokens_np,
                        kv_connector_output,
3515
                    )
3516
3517

                sample_hidden_states = hidden_states[logits_indices]
3518
                logits = self.model.compute_logits(sample_hidden_states)
3519
3520
3521
3522
            else:
                # Rare case.
                assert not self.is_pooling_model

3523
                sample_hidden_states = hidden_states[logits_indices]
3524
                if not get_pp_group().is_last_rank:
3525
                    all_gather_tensors = {
3526
                        "residual": not is_residual_scattered_for_sp(
3527
                            self.vllm_config, num_tokens_padded
3528
                        )
3529
                    }
3530
                    get_pp_group().send_tensor_dict(
3531
3532
                        hidden_states.tensors,
                        all_gather_group=get_tp_group(),
3533
3534
                        all_gather_tensors=all_gather_tensors,
                    )
3535
3536
                    logits = None
                else:
3537
                    logits = self.model.compute_logits(sample_hidden_states)
3538

3539
                model_output_broadcast_data: dict[str, Any] = {}
3540
3541
3542
                if logits is not None:
                    model_output_broadcast_data["logits"] = logits.contiguous()

3543
                broadcasted = get_pp_group().broadcast_tensor_dict(
3544
3545
                    model_output_broadcast_data, src=len(get_pp_group().ranks) - 1
                )
3546
3547
                assert broadcasted is not None
                logits = broadcasted["logits"]
3548

3549
3550
3551
3552
3553
3554
3555
3556
        self.execute_model_state = ExecuteModelState(
            scheduler_output,
            logits,
            spec_decode_metadata,
            spec_decode_common_attn_metadata,
            hidden_states,
            sample_hidden_states,
            aux_hidden_states,
3557
            ec_connector_output,
3558
            cudagraph_stats,
3559
            slot_mappings,
3560
        )
3561
        self.kv_connector_output = kv_connector_output
3562
3563
3564
3565
3566
3567
        return None

    @torch.inference_mode
    def sample_tokens(
        self, grammar_output: "GrammarOutput | None"
    ) -> ModelRunnerOutput | AsyncModelRunnerOutput | IntermediateTensors:
3568
3569
3570
        kv_connector_output = self.kv_connector_output
        self.kv_connector_output = None

3571
3572
        if self.execute_model_state is None:
            # Nothing to do (PP non-final rank case), output isn't used.
3573
            if not kv_connector_output:
3574
                return None  # type: ignore[return-value]
3575
3576
3577
3578
3579
3580
3581
3582
3583

            # 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
3584
3585
3586
3587
3588
3589
3590
3591
3592
3593

        # Unpack ephemeral state.
        (
            scheduler_output,
            logits,
            spec_decode_metadata,
            spec_decode_common_attn_metadata,
            hidden_states,
            sample_hidden_states,
            aux_hidden_states,
3594
            ec_connector_output,
3595
            cudagraph_stats,
3596
            slot_mappings,
3597
3598
3599
3600
3601
3602
3603
3604
3605
        ) = 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
            )
3606

3607
        with record_function_or_nullcontext("gpu_model_runner: sample"):
3608
3609
            sampler_output = self._sample(logits, spec_decode_metadata)

3610
3611
3612
3613
        self._update_states_after_model_execute(
            sampler_output.sampled_token_ids, scheduler_output
        )

3614
3615
        self._draft_token_ids = None
        self._draft_token_req_ids = None
3616
3617
        self.input_batch.prev_sampled_token_ids = None

3618
        def propose_draft_token_ids(sampled_token_ids):
3619
            assert spec_decode_common_attn_metadata is not None
3620
            with record_function_or_nullcontext("gpu_model_runner: draft"):
3621
3622
3623
3624
3625
3626
3627
3628
3629
                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,
3630
                    slot_mappings,
3631
                )
3632
                self._copy_draft_token_ids_to_cpu(scheduler_output)
3633

3634
        spec_config = self.speculative_config
3635
3636
3637
3638
3639
        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
3640
            )
3641
3642
3643
3644
3645
            use_gpu_toks = (
                spec_config.use_eagle() or spec_config.uses_draft_model()
            ) and not spec_config.disable_padded_drafter_batch
            if use_gpu_toks:
                # EAGLE/DraftModel speculative decoding can use the GPU sampled tokens
3646
                # as inputs, and does not need to wait for bookkeeping to finish.
3647
                assert isinstance(self.drafter, EagleProposer | DraftModelProposer)
3648
3649
3650
3651
3652
3653
3654
3655
3656
3657
3658
3659
3660
                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,
                        )
3661
                    )
3662
3663
3664
3665
3666
3667
3668
3669
3670
3671
3672
                    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
3673

3674
        with record_function_or_nullcontext("gpu_model_runner: bookkeep"):
3675
3676
3677
3678
3679
3680
3681
3682
            (
                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,
3683
3684
3685
3686
3687
            ) = self._bookkeeping_sync(
                scheduler_output,
                sampler_output,
                logits,
                hidden_states,
3688
                scheduler_output.total_num_scheduled_tokens,
3689
                spec_decode_metadata,
3690
            )
3691

3692
        if propose_drafts_after_bookkeeping:
3693
3694
3695
            # 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)
3696

3697
        with record_function_or_nullcontext("gpu_model_runner: eplb"):
3698
            self.eplb_step()
3699

3700
        with record_function_or_nullcontext("gpu_model_runner: ModelRunnerOutput"):
3701
3702
3703
3704
3705
3706
3707
            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.")

3708
3709
3710
3711
3712
3713
3714
            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,
3715
3716
3717
                ec_connector_output=ec_connector_output
                if self.supports_mm_inputs
                else None,
3718
                num_nans_in_logits=num_nans_in_logits,
3719
                cudagraph_stats=cudagraph_stats,
3720
            )
3721

3722
3723
        if not self.use_async_scheduling:
            return output
3724

3725
3726
3727
3728
3729
3730
3731
3732
3733
        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,
3734
                vocab_size=self.input_batch.vocab_size,
3735
3736
3737
3738
3739
            )
        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
3740
            # any requests with sampling params that require output ids.
3741
3742
3743
3744
            self.input_batch.set_async_sampled_token_ids(
                async_output.sampled_token_ids_cpu,
                async_output.async_copy_ready_event,
            )
3745
3746
3747

        return async_output

3748
    def take_draft_token_ids(self) -> DraftTokenIds | None:
3749
        if not self.num_spec_tokens or not self._draft_token_req_ids:
3750
            return None
3751
        draft_token_ids, req_ids = self._get_draft_token_ids_cpu()
3752
        return DraftTokenIds(req_ids, draft_token_ids)
3753

3754
3755
3756
    def _copy_draft_token_ids_to_cpu(
        self, scheduler_output: "SchedulerOutput", zeros_only: bool = False
    ) -> None:
3757
3758
3759
3760
3761
3762
        # 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
        ):
3763
3764
3765
            return
        # We must also set the corresponding request ids.
        self._draft_token_req_ids = self.input_batch.req_ids.copy()
3766

3767
3768
3769
3770
3771
3772
3773
3774
3775
3776
3777
3778
3779
3780
3781
3782
3783
3784
3785
3786
        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()

3787
    def _get_draft_token_ids_cpu(self) -> tuple[list[list[int]], list[str]]:
3788
        if isinstance(self._draft_token_ids, list):
3789
3790
3791
3792
            return self._draft_token_ids, self.input_batch.req_ids
        req_ids = self._draft_token_req_ids
        if req_ids is None:
            return [], []
3793
3794
3795
        assert self.draft_token_ids_event is not None
        assert self.draft_token_ids_cpu is not None
        self.draft_token_ids_event.synchronize()
3796
        return self.draft_token_ids_cpu[: len(req_ids)].tolist(), req_ids
3797

3798
3799
3800
3801
3802
3803
3804
3805
3806
3807
3808
3809
3810
    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
3811
            assert counts_cpu is not None
3812
3813
3814
3815
3816
3817
3818
3819
            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
3820
3821
        sampled_count_event = self.valid_sampled_token_count_event
        if sampled_count_event is None or prev_sampled_token_ids is None:
3822
3823
3824
            return []

        counts_cpu = self.valid_sampled_token_count_cpu
3825
3826
        assert counts_cpu is not None
        sampled_count_event.synchronize()
3827
3828
        return counts_cpu[: prev_sampled_token_ids.shape[0]].tolist()

3829
3830
3831
    def propose_draft_token_ids(
        self,
        scheduler_output: "SchedulerOutput",
3832
        sampled_token_ids: torch.Tensor | list[list[int]],
3833
3834
3835
        sampling_metadata: SamplingMetadata,
        hidden_states: torch.Tensor,
        sample_hidden_states: torch.Tensor,
3836
3837
        aux_hidden_states: list[torch.Tensor] | None,
        spec_decode_metadata: SpecDecodeMetadata | None,
3838
        common_attn_metadata: CommonAttentionMetadata,
3839
        slot_mappings: dict[str, torch.Tensor] | list[dict[str, torch.Tensor]] | None,
3840
    ) -> list[list[int]] | torch.Tensor:
3841
        num_scheduled_tokens = scheduler_output.total_num_scheduled_tokens
3842
3843
3844
        spec_config = self.speculative_config
        assert spec_config is not None
        if spec_config.method == "ngram":
3845
            assert isinstance(sampled_token_ids, list)
3846
            assert isinstance(self.drafter, NgramProposer)
3847
            draft_token_ids = self.drafter.propose(
3848
                sampled_token_ids,
3849
3850
                self.input_batch.num_tokens_no_spec,
                self.input_batch.token_ids_cpu,
3851
                slot_mappings=slot_mappings,
3852
            )
3853
        elif spec_config.method == "suffix":
3854
3855
            assert isinstance(sampled_token_ids, list)
            assert isinstance(self.drafter, SuffixDecodingProposer)
3856
3857
3858
            draft_token_ids = self.drafter.propose(
                self.input_batch, sampled_token_ids, slot_mappings=slot_mappings
            )
3859
        elif spec_config.method == "medusa":
3860
            assert isinstance(sampled_token_ids, list)
3861
            assert isinstance(self.drafter, MedusaProposer)
3862

3863
3864
            if sample_hidden_states.shape[0] == len(sampled_token_ids):
                # The input to the target model does not include draft tokens.
3865
3866
3867
3868
                hidden_states = sample_hidden_states
            else:
                indices = []
                offset = 0
3869
3870
3871
                assert spec_decode_metadata is not None, (
                    "No spec decode metadata for medusa"
                )
3872
                for num_draft, tokens in zip(
3873
3874
                    spec_decode_metadata.num_draft_tokens, sampled_token_ids
                ):
3875
                    indices.append(offset + len(tokens) - 1)
3876
                    offset += num_draft + 1
3877
                indices = torch.tensor(indices, device=self.device)
3878
3879
                hidden_states = sample_hidden_states[indices]

3880
            draft_token_ids = self.drafter.propose(
3881
3882
                target_hidden_states=hidden_states,
                sampling_metadata=sampling_metadata,
3883
                slot_mappings=slot_mappings,
3884
            )
3885
3886
        elif spec_config.use_eagle() or spec_config.uses_draft_model():
            assert isinstance(self.drafter, EagleProposer | DraftModelProposer)
3887

3888
            if spec_config.disable_padded_drafter_batch:
3889
3890
3891
                # 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.
3892
3893
                assert isinstance(sampled_token_ids, list), (
                    "sampled_token_ids should be a python list when"
3894
                    "padded-batch is disabled."
3895
                )
3896
                next_token_ids = self.drafter.prepare_next_token_ids_cpu(
3897
3898
3899
3900
3901
                    sampled_token_ids,
                    self.requests,
                    self.input_batch,
                    scheduler_output.num_scheduled_tokens,
                )
3902
3903
3904
3905
3906
            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.
3907
3908
                assert isinstance(sampled_token_ids, torch.Tensor), (
                    "sampled_token_ids should be a torch.Tensor when"
3909
                    "padded-batch is enabled."
3910
3911
                )
                next_token_ids, valid_sampled_tokens_count = (
3912
3913
3914
3915
3916
                    self.drafter.prepare_next_token_ids_padded(
                        common_attn_metadata,
                        sampled_token_ids,
                        self.requests,
                        self.input_batch,
3917
                        self.discard_request_mask.gpu,
3918
                    )
3919
                )
3920
3921
3922
                self._copy_valid_sampled_token_count(
                    next_token_ids, valid_sampled_tokens_count
                )
Jiayi Yao's avatar
Jiayi Yao committed
3923

3924
            num_rejected_tokens_gpu = None
3925
            if spec_decode_metadata is None:
3926
                token_indices_to_sample = None
3927
                # input_ids can be None for multimodal models.
3928
                target_token_ids = self.input_ids.gpu[:num_scheduled_tokens]
3929
                target_positions = self._get_positions(num_scheduled_tokens)
3930
                if self.use_aux_hidden_state_outputs:
Wentao Ye's avatar
Wentao Ye committed
3931
                    assert aux_hidden_states is not None
3932
                    target_hidden_states = torch.cat(
3933
3934
                        [h[:num_scheduled_tokens] for h in aux_hidden_states], dim=-1
                    )
3935
3936
                else:
                    target_hidden_states = hidden_states[:num_scheduled_tokens]
3937
            else:
3938
                if spec_config.disable_padded_drafter_batch:
3939
                    token_indices_to_sample = None
3940
3941
3942
3943
3944
                    common_attn_metadata, token_indices = self.drafter.prepare_inputs(
                        common_attn_metadata,
                        sampled_token_ids,
                        spec_decode_metadata.num_draft_tokens,
                    )
3945
3946
3947
3948
3949
3950
3951
3952
3953
                    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]
3954
                else:
3955
3956
3957
3958
3959
3960
3961
3962
                    (
                        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,
3963
                    )
3964
3965
3966
3967
3968
3969
3970
3971
3972
3973
3974
                    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]
3975

3976
            if self.supports_mm_inputs:
3977
3978
3979
3980
3981
3982
                mm_embed_inputs = self._gather_mm_embeddings(
                    scheduler_output,
                    shift_computed_tokens=1,
                )
            else:
                mm_embed_inputs = None
3983

3984
            draft_token_ids = self.drafter.propose(
3985
3986
3987
3988
                target_token_ids=target_token_ids,
                target_positions=target_positions,
                target_hidden_states=target_hidden_states,
                next_token_ids=next_token_ids,
3989
                last_token_indices=token_indices_to_sample,
3990
                sampling_metadata=sampling_metadata,
3991
                common_attn_metadata=common_attn_metadata,
3992
                mm_embed_inputs=mm_embed_inputs,
3993
                num_rejected_tokens_gpu=num_rejected_tokens_gpu,
3994
                slot_mappings=slot_mappings,
3995
            )
3996

3997
        return draft_token_ids
3998

3999
4000
4001
    def update_config(self, overrides: dict[str, Any]) -> None:
        allowed_config_names = {"load_config", "model_config"}
        for config_name, config_overrides in overrides.items():
4002
4003
            assert config_name in allowed_config_names, (
                f"Config `{config_name}` not supported. "
4004
                f"Allowed configs: {allowed_config_names}"
4005
            )
4006
4007
4008
4009
            config = getattr(self, config_name)
            new_config = update_config(config, config_overrides)
            setattr(self, config_name, new_config)

4010
4011
4012
4013
4014
    def load_model(self, eep_scale_up: bool = False) -> None:
        """
        Args:
            eep_scale_up: the model loading is for elastic EP scale up.
        """
4015
4016
4017
4018
4019
        logger.info_once(
            "Starting to load model %s...",
            self.model_config.model,
            scope="global",
        )
4020
4021
4022
4023
4024
        global_expert_loads, old_global_expert_indices_per_model, rank_mapping = (
            EplbState.get_eep_state(self.parallel_config)
            if eep_scale_up
            else (None, None, None)
        )
4025

4026
4027
4028
4029
        if self.parallel_config.enable_eplb:
            self.eplb_state = EplbState(self.parallel_config, self.device)
            eplb_models = 0

4030
4031
4032
4033
4034
4035
4036
4037
4038
4039
        try:
            with DeviceMemoryProfiler() as m:
                time_before_load = time.perf_counter()
                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
4040
                    )
4041
4042
4043
4044
4045
4046
4047
4048
4049
4050
4051
4052
4053
4054
4055
                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
                    ):
                        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,
                        )
4056

4057
4058
4059
4060
4061
4062
4063
4064
4065
4066
4067
4068
4069
4070
4071
4072
4073
4074
4075
4076
4077
4078
                        global_expert_load = (
                            global_expert_loads[eplb_models]
                            if global_expert_loads
                            else None
                        )
                        old_global_expert_indices = (
                            old_global_expert_indices_per_model[eplb_models]
                            if old_global_expert_indices_per_model
                            else None
                        )
                        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,
                            global_expert_load,
                            old_global_expert_indices,
                            rank_mapping,
                        )
                        eplb_models += 1
4079

4080
4081
4082
4083
4084
4085
                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"
                        )
4086

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
                    # 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
4112
        logger.info_once(
4113
4114
            "Model loading took %s GiB memory and %.6f seconds",
            format_gib(self.model_memory_usage),
4115
            time_after_load - time_before_load,
4116
            scope="local",
4117
        )
4118
        prepare_communication_buffer_for_model(self.model)
4119
4120
4121
4122
        if (drafter := getattr(self, "drafter", None)) and (
            drafter_model := getattr(drafter, "model", None)
        ):
            prepare_communication_buffer_for_model(drafter_model)
4123
        mm_config = self.model_config.multimodal_config
4124
        self.is_multimodal_pruning_enabled = (
4125
            supports_multimodal_pruning(self.get_model())
4126
4127
            and mm_config is not None
            and mm_config.is_multimodal_pruning_enabled()
4128
        )
4129

4130
        if is_mixture_of_experts(self.model) and self.parallel_config.enable_eplb:
4131
4132
4133
4134
4135
4136
4137
4138
4139
4140
4141
            logger.info_once("EPLB is enabled for model %s.", self.model_config.model)
            global_expert_load = (
                global_expert_loads[eplb_models] if global_expert_loads else None
            )
            old_global_expert_indices = (
                old_global_expert_indices_per_model[eplb_models]
                if old_global_expert_indices_per_model
                else None
            )
            assert self.eplb_state is not None
            self.eplb_state.add_model(
4142
                self.model,
4143
                self.model_config,
4144
4145
4146
                global_expert_load,
                old_global_expert_indices,
                rank_mapping,
4147
            )
4148
4149
            if self.eplb_state.is_async:
                self.eplb_state.start_async_loop(rank_mapping=rank_mapping)
4150

4151
        if (
4152
4153
            self.vllm_config.compilation_config.mode
            == CompilationMode.STOCK_TORCH_COMPILE
4154
        ):
4155
            backend = self.vllm_config.compilation_config.init_backend(self.vllm_config)
4156
            compilation_counter.stock_torch_compile_count += 1
4157
            self.model.compile(fullgraph=True, backend=backend)
4158
            return
4159
        # for other compilation modes, cudagraph behavior is controlled by
4160
4161
4162
        # CudagraphWraper and CudagraphDispatcher of vllm.

        # wrap the model with full cudagraph wrapper if needed.
4163
4164
        cudagraph_mode = self.compilation_config.cudagraph_mode
        assert cudagraph_mode is not None
4165
4166
4167
4168
        if (
            cudagraph_mode.has_full_cudagraphs()
            and not self.parallel_config.use_ubatching
        ):
4169
4170
4171
            self.model = CUDAGraphWrapper(
                self.model, self.vllm_config, runtime_mode=CUDAGraphMode.FULL
            )
4172
        elif self.parallel_config.use_ubatching:
4173
            if cudagraph_mode.has_full_cudagraphs():
4174
4175
4176
                self.model = UBatchWrapper(
                    self.model, self.vllm_config, CUDAGraphMode.FULL, self.device
                )
4177
            else:
4178
4179
4180
                self.model = UBatchWrapper(
                    self.model, self.vllm_config, CUDAGraphMode.NONE, self.device
                )
4181

4182
    def _get_eagle3_aux_layers_from_config(self) -> tuple[int, ...] | None:
4183
4184
4185
4186
4187
4188
4189
4190
4191
4192
4193
4194
4195
4196
4197
4198
4199
4200
4201
4202
4203
4204
4205
        """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

4206
    def reload_weights(self) -> None:
4207
        assert getattr(self, "model", None) is not None, (
4208
            "Cannot reload weights before model is loaded."
4209
        )
4210
4211
        model_loader = get_model_loader(self.load_config)
        logger.info("Reloading weights inplace...")
4212
        model_loader.load_weights(self.get_model(), model_config=self.model_config)
4213

4214
4215
4216
4217
4218
    def save_tensorized_model(
        self,
        tensorizer_config: "TensorizerConfig",
    ) -> None:
        TensorizerLoader.save_model(
4219
            self.get_model(),
4220
            tensorizer_config=tensorizer_config,
4221
            model_config=self.model_config,
4222
4223
        )

4224
4225
4226
    def _get_prompt_logprobs_dict(
        self,
        hidden_states: torch.Tensor,
4227
        num_scheduled_tokens: dict[str, int],
4228
    ) -> dict[str, LogprobsTensors | None]:
4229
        num_prompt_logprobs_dict = self.num_prompt_logprobs
4230
4231
4232
        if not num_prompt_logprobs_dict:
            return {}

4233
        in_progress_dict = self.input_batch.in_progress_prompt_logprobs_cpu
4234
        prompt_logprobs_dict: dict[str, LogprobsTensors | None] = {}
4235
4236
4237
4238
4239

        # 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():
4240
4241
4242
4243
            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
4244
4245
4246

            # Get metadata for this request.
            request = self.requests[req_id]
4247
4248
4249
4250
            if request.prompt_token_ids is None:
                # Prompt logprobs is incompatible with prompt embeddings
                continue

4251
4252
            num_prompt_tokens = len(request.prompt_token_ids)
            prompt_token_ids = torch.tensor(request.prompt_token_ids).to(
4253
4254
                self.device, non_blocking=True
            )
4255

4256
4257
4258
4259
4260
4261
            # 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(
4262
4263
                    num_prompt_tokens - 1, num_prompt_logprobs + 1
                )
4264
4265
                in_progress_dict[req_id] = logprobs_tensors

4266
            # Determine number of logits to retrieve.
4267
4268
            start_idx = request.num_computed_tokens
            start_tok = start_idx + 1
4269
            num_remaining_tokens = num_prompt_tokens - start_tok
4270
            if num_tokens <= num_remaining_tokens:
4271
                # This is a chunk, more tokens remain.
4272
4273
4274
                # 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.
4275
4276
4277
4278
4279
                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)
4280
4281
4282
4283
4284
4285
4286
                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
4287
4288
4289
4290
4291

            # 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]
4292
            offset = self.query_start_loc.np[req_idx].item()
4293
            prompt_hidden_states = hidden_states[offset : offset + num_logits]
4294
            logits = self.model.compute_logits(prompt_hidden_states)
4295
4296
4297
4298

            # 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.
4299
            tgt_token_ids = prompt_token_ids[start_tok : start_tok + num_logits]
4300
4301

            # Compute prompt logprobs.
4302
4303
            logprobs = self.sampler.compute_logprobs(logits)
            token_ids, logprobs, ranks = self.sampler.gather_logprobs(
4304
4305
                logprobs, num_prompt_logprobs, tgt_token_ids
            )
4306
4307

            # Transfer GPU->CPU async.
4308
4309
            chunk_slice = slice(start_idx, start_idx + num_logits)
            logprobs_tensors.logprob_token_ids[chunk_slice].copy_(
4310
4311
4312
                token_ids, non_blocking=True
            )
            logprobs_tensors.logprobs[chunk_slice].copy_(logprobs, non_blocking=True)
4313
            logprobs_tensors.selected_token_ranks[chunk_slice].copy_(
4314
4315
                ranks, non_blocking=True
            )
4316
4317
4318
4319
4320

        # 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]
4321
            del in_progress_dict[req_id]
4322
4323

        # Must synchronize the non-blocking GPU->CPU transfers.
4324
        if prompt_logprobs_dict:
4325
            self._sync_device()
4326
4327
4328

        return prompt_logprobs_dict

4329
4330
    def _get_nans_in_logits(
        self,
4331
        logits: torch.Tensor | None,
4332
4333
4334
4335
4336
4337
4338
4339
4340
4341
4342
    ) -> 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])
4343
4344
4345
                    if num_nans_for_index is not None and req_index < logits.shape[0]
                    else 0
                )
4346
4347
4348
4349
            return num_nans_in_logits
        except IndexError:
            return {}

4350
    @contextmanager
4351
4352
4353
    def maybe_randomize_inputs(
        self, input_ids: torch.Tensor | None, inputs_embeds: torch.Tensor | None
    ):
4354
4355
4356
4357
        """
        Randomize input_ids if VLLM_RANDOMIZE_DP_DUMMY_INPUTS is set.
        This is to help balance expert-selection
         - during profile_run
4358
         - during DP rank dummy run
4359
        """
4360

4361
4362
4363
4364
        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
4365
        elif input_ids is not None:
4366
4367
4368
4369

            @functools.cache
            def rand_input_ids() -> torch.Tensor:
                return torch.randint_like(
4370
                    self.input_ids.gpu,
4371
4372
                    low=0,
                    high=self.model_config.get_vocab_size(),
4373
                )
4374

4375
            logger.debug_once("Randomizing dummy input_ids for DP Rank")
4376
            input_ids.copy_(rand_input_ids()[: input_ids.size(0)], non_blocking=True)
4377
4378
            yield
            input_ids.fill_(0)
4379
4380
4381
4382
4383
4384
4385
4386
4387
4388
4389
4390
4391
4392
4393
        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)
4394

4395
4396
4397
4398
4399
4400
    def _get_mm_dummy_batch(
        self,
        modality: str,
        max_items_per_batch: int,
    ) -> BatchedTensorInputs:
        """Dummy data for profiling and precompiling multimodal models."""
4401
4402
        assert self.mm_budget is not None

4403
4404
4405
        # 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,
4406
            mm_counts={modality: 1},
4407
            cache=self.mm_budget.cache,
4408
        )
4409
4410
4411
4412
4413
        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"
4414

4415
        dummy_mm_items = [dummy_mm_item] * max_items_per_batch
4416

4417
4418
4419
4420
4421
4422
4423
4424
        return next(
            mm_kwargs_group
            for _, _, mm_kwargs_group in group_mm_kwargs_by_modality(
                dummy_mm_items,
                device=self.device,
                pin_memory=self.pin_memory,
            )
        )
4425

4426
4427
4428
4429
    @torch.inference_mode()
    def _dummy_run(
        self,
        num_tokens: int,
4430
        cudagraph_runtime_mode: CUDAGraphMode | None = None,
4431
4432
        force_attention: bool = False,
        uniform_decode: bool = False,
4433
        allow_microbatching: bool = True,
4434
4435
        skip_eplb: bool = False,
        is_profile: bool = False,
4436
        create_mixed_batch: bool = False,
4437
        remove_lora: bool = True,
4438
        activate_lora: bool = False,
Rémi Delacourt's avatar
Rémi Delacourt committed
4439
        is_graph_capturing: bool = False,
4440
    ) -> tuple[torch.Tensor, torch.Tensor]:
4441
4442
4443
4444
4445
4446
4447
        """
        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.
4448
                - if not set will determine the cudagraph mode based on using
4449
                    the self.cudagraph_dispatcher.
4450
4451
4452
4453
                - CUDAGraphMode.NONE: No cudagraph, for warm up and profile run
                - CUDAGraphMode.PIECEWISE: Piecewise cudagraph.
                - CUDAGraphMode.FULL: Full cudagraph, attention metadata is
                    needed.
4454
            force_attention: If True, always create attention metadata. Used to
4455
4456
4457
4458
                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.
4459
4460
            create_mixed_batch: If True, create a mixed batch with both decode
                (1 token) and prefill (multiple tokens) requests.
4461
            remove_lora: If False, dummy LoRAs are not destroyed after the run
4462
            activate_lora: If False, dummy_run is performed without LoRAs.
4463
        """
4464
4465
        mm_config = self.vllm_config.model_config.multimodal_config
        if mm_config and mm_config.mm_encoder_only:
4466
4467
4468
4469
            # 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([])

4470
4471
4472
4473
        assert (
            cudagraph_runtime_mode is None
            or cudagraph_runtime_mode.valid_runtime_modes()
        )
4474

4475
        # If cudagraph_mode.decode_mode() == FULL and
4476
        # cudagraph_mode.separate_routine(). This means that we are using
4477
4478
4479
4480
4481
4482
4483
4484
4485
4486
4487
        # 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.
4488
        max_query_len = self.uniform_decode_query_len if uniform_decode else num_tokens
4489

4490
4491
4492
4493
4494
        # 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.
        assert num_tokens <= self.scheduler_config.max_num_batched_tokens
        max_num_reqs = self.scheduler_config.max_num_seqs
4495
4496
4497
4498
        if create_mixed_batch:
            assert not uniform_decode
            # Create mixed batch:
            # first half decode tokens, second half one prefill
4499
            num_decode_tokens = min(max_num_reqs - 1, num_tokens // 2)
4500
4501
4502
4503
            num_prefill_tokens = num_tokens - num_decode_tokens
            num_reqs = num_decode_tokens + 1

            # Create decode requests (1 token each) followed by prefill request
4504
            num_scheduled_tokens_list = [1] * num_decode_tokens + [num_prefill_tokens]
4505
4506
4507
            # Note: Overriding max_query_len to be the prefill tokens
            max_query_len = num_prefill_tokens
        elif uniform_decode:
4508
            assert not create_mixed_batch
4509
            num_reqs = min(max_num_reqs, cdiv(num_tokens, max_query_len))
4510
4511
            num_scheduled_tokens_list = [max_query_len] * num_reqs
            if num_tokens % max_query_len != 0:
4512
                num_scheduled_tokens_list[-1] = num_tokens % max_query_len
4513
4514
4515
4516
4517
4518
        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

4519
4520
        assert sum(num_scheduled_tokens_list) == num_tokens
        assert len(num_scheduled_tokens_list) == num_reqs
4521
        num_scheduled_tokens = np.array(num_scheduled_tokens_list, dtype=np.int32)
4522
4523
        num_tokens_unpadded = int(num_scheduled_tokens.sum())

4524
        num_sampled_tokens = np.ones(num_reqs, dtype=np.int32)
4525

4526
        _cudagraph_mode, batch_desc, should_ubatch, num_tokens_across_dp, _ = (
4527
4528
4529
4530
4531
4532
4533
4534
4535
4536
4537
4538
4539
4540
4541
4542
4543
4544
            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
                force_has_lora=activate_lora,
4545
4546
            )
        )
4547
4548
4549

        if cudagraph_runtime_mode is None:
            cudagraph_runtime_mode = _cudagraph_mode
4550
        else:
4551
4552
4553
4554
4555
4556
4557
4558
4559
            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
        )
4560
        ubatch_slices, ubatch_slices_padded = maybe_create_ubatch_slices(
4561
4562
4563
4564
4565
4566
4567
4568
4569
4570
            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,
4571
        )
4572

4573
        attn_metadata: PerLayerAttnMetadata | None = None
4574

4575
4576
4577
4578
4579
4580
4581
        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,
        )

4582
4583
        # If force_attention is True, we always capture attention. Otherwise,
        # it only happens for cudagraph_runtime_mode=FULL.
4584
        if force_attention or cudagraph_runtime_mode == CUDAGraphMode.FULL:
4585
4586
4587
4588
4589
4590
            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:
4591
                seq_lens = max_query_len  # type: ignore[assignment]
4592
            self.seq_lens.np[:num_reqs] = seq_lens
4593
4594
            self.seq_lens.np[num_reqs:] = 0
            self.seq_lens.copy_to_gpu()
4595

4596
4597
            cum_num_tokens, _ = self._get_cumsum_and_arange(num_scheduled_tokens)
            self.query_start_loc.np[1 : num_reqs + 1] = cum_num_tokens
4598
4599
            self.query_start_loc.copy_to_gpu()

4600
            pad_attn = cudagraph_runtime_mode == CUDAGraphMode.FULL
4601
            attn_metadata, _ = self._build_attention_metadata(
4602
4603
4604
                num_tokens=num_tokens_unpadded,
                num_reqs=num_reqs_padded,
                max_query_len=max_query_len,
4605
                ubatch_slices=ubatch_slices_padded if pad_attn else ubatch_slices,
4606
                for_cudagraph_capture=is_graph_capturing,
4607
                slot_mappings=slot_mappings_by_group,
4608
            )
4609

4610
        with self.maybe_dummy_run_with_lora(
4611
4612
4613
4614
4615
            self.lora_config,
            num_scheduled_tokens,
            num_sampled_tokens,
            activate_lora,
            remove_lora,
4616
        ):
4617
            # Make sure padding doesn't exceed max_num_tokens
4618
            assert num_tokens_padded <= self.max_num_tokens
4619
            model_kwargs = self._init_model_kwargs()
4620
            if self.supports_mm_inputs and not self.model_config.is_encoder_decoder:
Patrick von Platen's avatar
Patrick von Platen committed
4621
4622
                input_ids, inputs_embeds = self._prepare_mm_inputs(num_tokens_padded)

4623
                model_kwargs = {
4624
                    **model_kwargs,
4625
4626
                    **self._dummy_mm_kwargs(num_reqs),
                }
4627
4628
            elif self.enable_prompt_embeds:
                input_ids = None
4629
                inputs_embeds = self.inputs_embeds.gpu[:num_tokens_padded]
4630
                model_kwargs = self._init_model_kwargs()
4631
            else:
4632
                input_ids = self.input_ids.gpu[:num_tokens_padded]
4633
                inputs_embeds = None
4634

4635
            if self.uses_mrope:
4636
                positions = self.mrope_positions.gpu[:, :num_tokens_padded]
4637
            elif self.uses_xdrope_dim > 0:
4638
                positions = self.xdrope_positions.gpu[:, :num_tokens_padded]
4639
            else:
4640
                positions = self.positions.gpu[:num_tokens_padded]
4641
4642
4643
4644
4645
4646
4647
4648
4649

            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,
4650
4651
4652
                            device=self.device,
                        )
                    )
4653
4654

                intermediate_tensors = self.sync_and_slice_intermediate_tensors(
4655
                    num_tokens_padded, None, False
4656
                )
4657

4658
            if ubatch_slices_padded is not None:
4659
4660
4661
                # Adjust values to reflect a single ubatch.
                # TODO(sage,lucas): this is cruft that should be addressed in
                #  the padding refactor.
4662
                num_tokens_padded = ubatch_slices_padded[0].num_tokens
4663
                if num_tokens_across_dp is not None:
4664
                    num_tokens_across_dp[:] = num_tokens_padded
4665

4666
            with (
4667
                self.maybe_randomize_inputs(input_ids, inputs_embeds),
4668
                set_forward_context(
4669
4670
                    attn_metadata,
                    self.vllm_config,
4671
                    num_tokens=num_tokens_padded,
4672
4673
                    num_tokens_across_dp=num_tokens_across_dp,
                    cudagraph_runtime_mode=cudagraph_runtime_mode,
4674
                    batch_descriptor=batch_desc,
4675
                    ubatch_slices=ubatch_slices_padded,
4676
                    slot_mapping=slot_mappings,
4677
4678
                ),
            ):
4679
                outputs = self.model(
4680
4681
4682
4683
                    input_ids=input_ids,
                    positions=positions,
                    intermediate_tensors=intermediate_tensors,
                    inputs_embeds=inputs_embeds,
4684
                    **model_kwargs,
4685
                )
4686

4687
4688
4689
4690
            if self.use_aux_hidden_state_outputs:
                hidden_states, _ = outputs
            else:
                hidden_states = outputs
4691

4692
4693
4694
4695
4696
4697
            if self.speculative_config and (
                self.speculative_config.use_eagle()
                or self.speculative_config.uses_draft_model()
            ):
                assert isinstance(self.drafter, EagleProposer | DraftModelProposer)
                assert self.speculative_config is not None
4698
4699
4700
                # 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.
4701
                use_cudagraphs = (
4702
4703
4704
4705
4706
4707
4708
4709
4710
                    (
                        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
4711
4712
4713
4714
4715
4716
4717
4718
4719
4720
4721

                # 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
                if self.compilation_config.cudagraph_specialize_lora and activate_lora:
                    use_cudagraphs = False

                self.drafter.dummy_run(
                    num_tokens,
                    use_cudagraphs=use_cudagraphs,
Rémi Delacourt's avatar
Rémi Delacourt committed
4722
                    is_graph_capturing=is_graph_capturing,
4723
                    slot_mappings=slot_mappings,
4724
                )
4725

4726
4727
4728
4729
4730
4731
4732
4733
4734
4735
4736
        # 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()

4737
4738
4739
4740
4741
4742
4743
4744
4745
4746
        # 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)

4747
        logit_indices = np.cumsum(num_scheduled_tokens) - 1
4748
4749
4750
4751
        logit_indices_device = torch.from_numpy(logit_indices).to(
            self.device, non_blocking=True
        )
        return hidden_states, hidden_states[logit_indices_device]
4752
4753
4754
4755
4756
4757

    @torch.inference_mode()
    def _dummy_sampler_run(
        self,
        hidden_states: torch.Tensor,
    ) -> torch.Tensor:
4758
4759
4760
        # The dummy hidden states may contain special values,
        # like `inf` or `nan`.
        # To avoid breaking the sampler, we use a random tensor here instead.
4761

4762
4763
        mm_config = self.vllm_config.model_config.multimodal_config
        if mm_config and mm_config.mm_encoder_only:
4764
4765
4766
            # MM Encoder only model no need to run sampler.
            return torch.tensor([])

4767
        hidden_states = torch.rand_like(hidden_states)
4768

4769
        logits = self.model.compute_logits(hidden_states)
4770
4771
        num_reqs = logits.size(0)

4772
        dummy_tensors = lambda v: torch.full((num_reqs,), v, device=self.device)
4773
4774
4775
4776
4777
4778
4779
4780
4781
4782
4783
4784
4785
4786
4787

        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)],
4788
            spec_token_ids=[[] for _ in range(num_reqs)],
4789
4790
            allowed_token_ids_mask=None,
            bad_words_token_ids={},
4791
            logitsprocs=LogitsProcessors(),
4792
        )
4793
        try:
4794
4795
4796
            sampler_output = self.sampler(
                logits=logits, sampling_metadata=dummy_metadata
            )
4797
        except RuntimeError as e:
4798
            if "out of memory" in str(e):
4799
4800
4801
4802
                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 "
4803
4804
                    "initializing the engine."
                ) from e
4805
4806
            else:
                raise e
4807
        if self.speculative_config:
4808
4809
            draft_token_ids = [[0] for _ in range(num_reqs)]
            dummy_spec_decode_metadata = SpecDecodeMetadata.make_dummy(
4810
4811
                draft_token_ids, self.device
            )
4812
4813
4814
4815
4816
4817

            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
4818
4819
4820
4821
4822
            logits = torch.randn(
                num_tokens + num_reqs,
                logits.shape[-1],
                device=self.device,
                dtype=logits.dtype,
4823
            )
4824
4825
4826
            self.rejection_sampler(
                dummy_spec_decode_metadata,
                draft_probs,
4827
                logits,
4828
4829
                dummy_metadata,
            )
4830
        return sampler_output
4831

4832
    def _dummy_pooler_run_task(
4833
4834
        self,
        hidden_states: torch.Tensor,
4835
4836
        task: PoolingTask,
    ) -> PoolerOutput:
4837
4838
4839
4840
        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
4841
4842
4843
4844
        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
4845
4846
4847

        req_num_tokens = num_tokens // num_reqs

4848
        dummy_prompt_lens = torch.from_numpy(num_scheduled_tokens_np)
4849
4850
4851
        dummy_token_ids = torch.zeros(
            (num_reqs, req_num_tokens), dtype=torch.int32, device=self.device
        )
4852

4853
        model = cast(VllmModelForPooling, self.get_model())
4854
        dummy_pooling_params = PoolingParams(task=task)
4855
        dummy_pooling_params.verify(task=task, model_config=self.model_config)
4856
        to_update = model.pooler.get_pooling_updates(task)
4857
4858
        to_update.apply(dummy_pooling_params)

4859
        dummy_metadata = PoolingMetadata(
4860
4861
4862
            prompt_lens=dummy_prompt_lens,
            prompt_token_ids=dummy_token_ids,
            pooling_params=[dummy_pooling_params] * num_reqs,
4863
            pooling_states=[PoolingStates() for i in range(num_reqs)],
4864
        )
4865

4866
        dummy_metadata.build_pooling_cursor(
4867
            num_scheduled_tokens_np,
4868
4869
            seq_lens_cpu=dummy_prompt_lens,
            device=hidden_states.device,
4870
        )
4871

4872
        try:
4873
4874
4875
            return model.pooler(
                hidden_states=hidden_states, pooling_metadata=dummy_metadata
            )
4876
        except RuntimeError as e:
4877
            if "out of memory" in str(e):
4878
                raise RuntimeError(
4879
4880
4881
                    "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 "
4882
4883
                    "initializing the engine."
                ) from e
4884
4885
            else:
                raise e
4886
4887
4888
4889
4890
4891

    @torch.inference_mode()
    def _dummy_pooler_run(
        self,
        hidden_states: torch.Tensor,
    ) -> PoolerOutput:
4892
4893
        mm_config = self.vllm_config.model_config.multimodal_config
        if mm_config and mm_config.mm_encoder_only:
4894
4895
4896
            # MM Encoder only model not need to run pooler.
            return torch.tensor([])

4897
        # Find the task that has the largest output for subsequent steps
4898
4899
4900
        supported_pooling_tasks = self.get_supported_pooling_tasks()

        if not supported_pooling_tasks:
4901
4902
4903
4904
4905
4906
            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."
            )
4907

4908
        output_size = dict[PoolingTask, float]()
4909
        for task in supported_pooling_tasks:
4910
4911
            # Run a full batch with each task to ensure none of them OOMs
            output = self._dummy_pooler_run_task(hidden_states, task)
4912
            output_size[task] = sum(o.nbytes for o in output if o is not None)
4913
4914
4915
4916
            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)
4917

4918
    def profile_run(self) -> None:
4919
        # Profile with multimodal encoder & encoder cache.
4920
        if self.supports_mm_inputs:
4921
4922
            mm_config = self.model_config.multimodal_config
            if mm_config is not None and mm_config.skip_mm_profiling:
4923
                logger.info(
4924
                    "Skipping memory profiling for multimodal encoder and "
4925
4926
                    "encoder cache."
                )
4927
4928
4929
4930
4931
4932
4933
4934
            else:
                mm_budget = self.mm_budget
                assert mm_budget is not None

                if (encoder_budget := mm_budget.get_encoder_budget()) > 0:
                    # NOTE: Currently model is profiled with a single non-text
                    # modality with the max possible input tokens even when
                    # it supports multiple.
4935
                    dummy_modality = mm_budget.get_modality_with_max_tokens()
4936
4937
4938
                    max_mm_items_per_batch = mm_budget.max_items_per_batch_by_modality[
                        dummy_modality
                    ]
4939
4940
4941
4942
4943
4944
4945
4946
4947

                    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,
                    )
4948

4949
4950
4951
4952
4953
                    # Create dummy batch of multimodal inputs.
                    batched_dummy_mm_inputs = self._get_mm_dummy_batch(
                        dummy_modality,
                        max_mm_items_per_batch,
                    )
4954

4955
                    # Run multimodal encoder.
4956
                    dummy_encoder_outputs = self.model.embed_multimodal(
4957
4958
                        **batched_dummy_mm_inputs
                    )
4959

4960
4961
4962
4963
                    sanity_check_mm_encoder_outputs(
                        dummy_encoder_outputs,
                        expected_num_items=max_mm_items_per_batch,
                    )
4964
4965
                    for i, output in enumerate(dummy_encoder_outputs):
                        self.encoder_cache[f"tmp_{i}"] = output
4966

4967
        # Add `is_profile` here to pre-allocate communication buffers
4968
4969
4970
        hidden_states, last_hidden_states = self._dummy_run(
            self.max_num_tokens, is_profile=True
        )
4971
        if get_pp_group().is_last_rank:
4972
4973
4974
4975
            if self.is_pooling_model:
                output = self._dummy_pooler_run(hidden_states)
            else:
                output = self._dummy_sampler_run(last_hidden_states)
4976
        else:
4977
            output = None
4978
        self._sync_device()
4979
        del hidden_states, output
4980
        self.encoder_cache.clear()
4981
        gc.collect()
4982

4983
    def capture_model(self) -> int:
4984
        if self.compilation_config.cudagraph_mode == CUDAGraphMode.NONE:
4985
            logger.warning(
4986
                "Skipping CUDA graph capture. To turn on CUDA graph capture, "
4987
4988
                "ensure `cudagraph_mode` was not manually set to `NONE`"
            )
4989
            return 0
4990

4991
4992
        compilation_counter.num_gpu_runner_capture_triggers += 1

4993
4994
        start_time = time.perf_counter()

4995
4996
4997
4998
4999
5000
5001
5002
5003
5004
5005
5006
5007
5008
        @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()
5009
                    gc.collect()
5010

5011
5012
5013
        # 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.
5014
        set_cudagraph_capturing_enabled(True)
5015
        with freeze_gc(), graph_capture(device=self.device):
5016
            start_free_gpu_memory = torch.cuda.mem_get_info()[0]
5017

5018
5019
5020
5021
            for (
                runtime_mode,
                batch_descs,
            ) in self.cudagraph_dispatcher.get_capture_descs():
5022
                self._capture_cudagraphs(
5023
5024
                    batch_descriptors=batch_descs,
                    cudagraph_runtime_mode=runtime_mode,
5025
                )
5026

5027
5028
5029
            torch.cuda.synchronize()
            end_free_gpu_memory = torch.cuda.mem_get_info()[0]

5030
5031
5032
        # 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
5033
        # we may do lazy capturing in future that still allows capturing
5034
5035
        # after here.
        set_cudagraph_capturing_enabled(False)
5036

5037
5038
5039
5040
        # Lock workspace to prevent resizing during execution.
        # Max workspace sizes should have been captured during warmup/profiling.
        lock_workspace()

5041
5042
5043
5044
        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.
5045
        logger.info_once(
5046
5047
5048
            "Graph capturing finished in %.0f secs, took %.2f GiB",
            elapsed_time,
            cuda_graph_size / (1 << 30),
5049
            scope="local",
5050
        )
5051
        return cuda_graph_size
5052

5053
5054
    def _capture_cudagraphs(
        self,
5055
        batch_descriptors: list[BatchDescriptor],
5056
5057
5058
5059
5060
5061
        cudagraph_runtime_mode: CUDAGraphMode,
    ):
        assert (
            cudagraph_runtime_mode != CUDAGraphMode.NONE
            and cudagraph_runtime_mode.valid_runtime_modes()
        ), f"Invalid cudagraph runtime mode: {cudagraph_runtime_mode}"
5062

5063
5064
5065
5066
5067
5068
5069
5070
5071
5072
5073
5074
5075
5076
        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,
        )

5077
5078
        # Only rank 0 should print progress bar during capture
        if is_global_first_rank():
5079
5080
            batch_descriptors = tqdm(
                batch_descriptors,
5081
5082
5083
                disable=not self.load_config.use_tqdm_on_load,
                desc="Capturing CUDA graphs ({}, {})".format(
                    "decode" if uniform_decode else "mixed prefill-decode",
5084
5085
5086
                    cudagraph_runtime_mode.name,
                ),
            )
5087

5088
        # We skip EPLB here since we don't want to record dummy metrics
5089
5090
5091
5092
        for batch_desc in batch_descriptors:
            num_tokens = batch_desc.num_tokens
            activate_lora = batch_desc.has_lora

5093
            # We currently only capture ubatched graphs when its a FULL
5094
5095
5096
            # 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
5097
            allow_microbatching = (
5098
                self.parallel_config.use_ubatching
5099
5100
                and cudagraph_runtime_mode == CUDAGraphMode.FULL
                and uniform_decode
5101
5102
5103
5104
5105
                and check_ubatch_thresholds(
                    config=self.vllm_config.parallel_config,
                    num_tokens=num_tokens,
                    uniform_decode=uniform_decode,
                )
5106
            )
5107

5108
5109
            for _ in range(self.compilation_config.cudagraph_num_of_warmups):
                # Use CUDAGraphRuntimeStyle.NONE (default) for warmup.
5110
                # But be careful, warm up with `NONE` is orthogonal to
5111
5112
5113
                # 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.
5114
                dummy_run(
5115
5116
5117
                    num_tokens,
                    cudagraph_runtime_mode=CUDAGraphMode.NONE,
                    allow_microbatching=allow_microbatching,
5118
                    activate_lora=activate_lora,
5119
                )
5120
5121
5122

            # Capture run
            dummy_run(
5123
5124
5125
                num_tokens,
                cudagraph_runtime_mode=cudagraph_runtime_mode,
                allow_microbatching=allow_microbatching,
5126
                activate_lora=activate_lora,
Rémi Delacourt's avatar
Rémi Delacourt committed
5127
                is_graph_capturing=True,
5128
            )
5129
        self.maybe_remove_all_loras(self.lora_config)
5130

5131
5132
5133
5134
    def initialize_attn_backend(self, kv_cache_config: KVCacheConfig) -> None:
        """
        Initialize the attention backends and attention metadata builders.
        """
5135
        assert len(self.attn_groups) == 0, "Attention backends are already initialized"
5136

5137
5138
5139
5140
5141
5142
        class AttentionGroupKey(NamedTuple):
            attn_backend: type[AttentionBackend]
            kv_cache_spec: KVCacheSpec

        def get_attn_backends_for_group(
            kv_cache_group_spec: KVCacheGroupSpec,
5143
        ) -> tuple[dict[AttentionGroupKey, list[str]], set[type[AttentionBackend]]]:
5144
            layer_type = cast(type[Any], AttentionLayerBase)
5145
            layers = get_layers_from_vllm_config(
5146
                self.vllm_config, layer_type, kv_cache_group_spec.layer_names
5147
            )
5148
5149
            attn_backends = {}
            attn_backend_layers = defaultdict(list)
5150
            # Dedupe based on full class name; this is a bit safer than
5151
5152
5153
5154
            # 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.
5155
            for layer_name in kv_cache_group_spec.layer_names:
5156
                attn_backend = layers[layer_name].get_attn_backend()
5157
5158
5159
5160

                if layer_name in self.kv_sharing_fast_prefill_eligible_layers:
                    attn_backend = create_fast_prefill_custom_backend(
                        "FastPrefill",
5161
                        attn_backend,  # type: ignore[arg-type]
5162
5163
                    )

5164
5165
5166
                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):
5167
                    layer_kv_cache_spec = layer_kv_cache_spec.kv_cache_specs[layer_name]
5168
                key = (full_cls_name, layer_kv_cache_spec)
5169
5170
5171
                attn_backends[key] = AttentionGroupKey(
                    attn_backend, layer_kv_cache_spec
                )
5172
                attn_backend_layers[key].append(layer_name)
5173
5174
5175
5176
            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()),
            )
5177
5178

        def create_attn_groups(
5179
            attn_backends_map: dict[AttentionGroupKey, list[str]],
5180
            kv_cache_group_id: int,
5181
5182
        ) -> list[AttentionGroup]:
            attn_groups: list[AttentionGroup] = []
5183
            for (attn_backend, kv_cache_spec), layer_names in attn_backends_map.items():
5184
                attn_group = AttentionGroup(
5185
                    attn_backend,
5186
                    layer_names,
5187
                    kv_cache_spec,
5188
                    kv_cache_group_id,
5189
5190
                )

5191
5192
5193
                attn_groups.append(attn_group)
            return attn_groups

5194
        attention_backend_maps = []
5195
        attention_backend_list = []
5196
        for kv_cache_group_spec in kv_cache_config.kv_cache_groups:
5197
            attn_backends = get_attn_backends_for_group(kv_cache_group_spec)
5198
            attention_backend_maps.append(attn_backends[0])
5199
            attention_backend_list.append(attn_backends[1])
5200
5201

        # Resolve cudagraph_mode before actually initialize metadata_builders
5202
5203
5204
        self._check_and_update_cudagraph_mode(
            attention_backend_list, kv_cache_config.kv_cache_groups
        )
5205

5206
5207
5208
        # Check if attention backend supports PCP&DCP and related features.
        check_attention_cp_compatibility(self.vllm_config)

5209
5210
        for i, attn_backend_map in enumerate(attention_backend_maps):
            self.attn_groups.append(create_attn_groups(attn_backend_map, i))
5211

5212
5213
5214
5215
5216
5217
5218
5219
5220
5221
5222
5223
5224
5225
5226
    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
5227
5228
                    if not self.parallel_config.use_ubatching
                    else self.parallel_config.num_ubatches,
5229
                )
co63oc's avatar
co63oc committed
5230
        # Calculate reorder batch threshold (if needed)
5231
5232
        # Note (tdoublep): do this *after* constructing builders,
        # because some of them change the threshold at init time.
5233
5234
        self.calculate_reorder_batch_threshold()

5235
    def _check_and_update_cudagraph_mode(
5236
5237
5238
        self,
        attention_backends: list[set[type[AttentionBackend]]],
        kv_cache_groups: list[KVCacheGroupSpec],
5239
    ) -> None:
5240
        """
5241
        Resolve the cudagraph_mode when there are multiple attention
5242
        groups with potential conflicting CUDA graph support.
5243
5244
5245
        Then initialize the cudagraph_dispatcher based on the resolved
        cudagraph_mode.
        """
5246
        min_cg_support = AttentionCGSupport.ALWAYS
5247
        min_cg_backend_name = None
5248

5249
5250
5251
5252
5253
5254
5255
5256
5257
5258
5259
5260
        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__
5261
5262
        # Flexible resolve the cudagraph mode
        cudagraph_mode = self.compilation_config.cudagraph_mode
5263
        assert cudagraph_mode is not None
5264
        # check cudagraph for mixed batch is supported
5265
5266
5267
5268
5269
5270
        if (
            cudagraph_mode.mixed_mode() == CUDAGraphMode.FULL
            and min_cg_support != AttentionCGSupport.ALWAYS
        ):
            msg = (
                f"CUDAGraphMode.{cudagraph_mode.name} is not supported "
5271
                f"with {min_cg_backend_name} backend (support: "
5272
5273
                f"{min_cg_support})"
            )
5274
5275
            if min_cg_support == AttentionCGSupport.NEVER:
                # if not supported any full cudagraphs, just raise it.
5276
5277
                msg += (
                    "; please try cudagraph_mode=PIECEWISE, and "
5278
                    "make sure compilation mode is VLLM_COMPILE"
5279
                )
5280
5281
5282
5283
5284
                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"
5285
                cudagraph_mode = self.compilation_config.cudagraph_mode = (
5286
                    CUDAGraphMode.FULL_AND_PIECEWISE
5287
                )
5288
5289
            else:
                msg += "; setting cudagraph_mode=FULL_DECODE_ONLY"
5290
                cudagraph_mode = self.compilation_config.cudagraph_mode = (
5291
                    CUDAGraphMode.FULL_DECODE_ONLY
5292
                )
5293
5294
            logger.warning(msg)

5295
        # check that if we are doing decode full-cudagraphs it is supported
5296
5297
5298
5299
5300
5301
        if (
            cudagraph_mode.decode_mode() == CUDAGraphMode.FULL
            and min_cg_support == AttentionCGSupport.NEVER
        ):
            msg = (
                f"CUDAGraphMode.{cudagraph_mode.name} is not supported "
5302
                f"with {min_cg_backend_name} backend (support: "
5303
5304
                f"{min_cg_support})"
            )
5305
            if self.compilation_config.mode == CompilationMode.VLLM_COMPILE and (
5306
5307
5308
5309
5310
                self.compilation_config.splitting_ops_contain_attention()
                or self.compilation_config.use_inductor_graph_partition
            ):
                msg += (
                    "; setting cudagraph_mode=PIECEWISE because "
5311
                    "attention is compiled piecewise"
5312
5313
                )
                cudagraph_mode = self.compilation_config.cudagraph_mode = (
5314
                    CUDAGraphMode.PIECEWISE
5315
                )
5316
            else:
5317
5318
                msg += (
                    "; setting cudagraph_mode=NONE because "
5319
                    "attention is not compiled piecewise"
5320
5321
                )
                cudagraph_mode = self.compilation_config.cudagraph_mode = (
5322
                    CUDAGraphMode.NONE
5323
                )
5324
5325
            logger.warning(msg)

5326
5327
        # check that if we are doing spec-decode + decode full-cudagraphs it is
        # supported
5328
5329
5330
5331
5332
5333
5334
5335
        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 "
5336
                f"{min_cg_backend_name} (support: {min_cg_support})"
5337
            )
5338
5339
            if self.compilation_config.splitting_ops_contain_attention():
                msg += "; setting cudagraph_mode=PIECEWISE"
5340
                cudagraph_mode = self.compilation_config.cudagraph_mode = (
5341
                    CUDAGraphMode.PIECEWISE
5342
                )
5343
5344
            else:
                msg += "; setting cudagraph_mode=NONE"
5345
                cudagraph_mode = self.compilation_config.cudagraph_mode = (
5346
                    CUDAGraphMode.NONE
5347
                )
5348
5349
5350
5351
            logger.warning(msg)

        # double check that we can support full cudagraph if they are requested
        # even after automatic downgrades
5352
5353
5354
5355
5356
5357
        if (
            cudagraph_mode.has_full_cudagraphs()
            and min_cg_support == AttentionCGSupport.NEVER
        ):
            raise ValueError(
                f"CUDAGraphMode.{cudagraph_mode.name} is not "
5358
                f"supported with {min_cg_backend_name} backend ("
5359
5360
                f"support:{min_cg_support}) "
                "; please try cudagraph_mode=PIECEWISE, "
5361
                "and make sure compilation mode is VLLM_COMPILE"
5362
            )
5363

5364
5365
5366
5367
        # 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
5368
        # Will be removed in the near future when we have separate cudagraph capture
5369
5370
5371
5372
5373
5374
5375
5376
5377
        # 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
            )
5378
5379
5380
5381
            capture_sizes = self.compilation_config.cudagraph_capture_sizes
            self.cudagraph_batch_sizes = (
                capture_sizes if capture_sizes is not None else []
            )
5382

5383
5384
        # Trigger cudagraph dispatching keys initialization after
        # resolved cudagraph mode.
5385
        self.compilation_config.cudagraph_mode = cudagraph_mode
5386
        self.cudagraph_dispatcher.initialize_cudagraph_keys(
5387
            cudagraph_mode, self.uniform_decode_query_len
5388
        )
5389

5390
5391
5392
5393
5394
        # Initialize eagle's cudagraph dispatcher if using eagle spec decode.
        if self.speculative_config and self.speculative_config.use_eagle():
            assert isinstance(self.drafter, EagleProposer)
            self.drafter.initialize_cudagraph_keys(cudagraph_mode)

5395
5396
    def calculate_reorder_batch_threshold(self) -> None:
        """
5397
5398
5399
5400
        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.
5401
        """
5402
5403
        min_none_high = lambda a, b: a if b is None else b if a is None else min(a, b)

5404
        reorder_batch_thresholds: list[int | None] = [
5405
5406
5407
            group.get_metadata_builder().reorder_batch_threshold
            for group in self._attn_group_iterator()
        ]
5408
5409
5410
5411
5412
        # 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
5413
        self.reorder_batch_threshold = reduce(min_none_high, reorder_batch_thresholds)  # type: ignore[assignment]
5414

5415
5416
5417
    @staticmethod
    def select_common_block_size(
        kv_manager_block_size: int, attn_groups: list[AttentionGroup]
5418
5419
    ) -> int:
        """
5420
5421
5422
5423
5424
        Select a block size that is supported by all backends and is a factor of
        kv_manager_block_size.

        If kv_manager_block_size is supported by all backends, return it directly.
        Otherwise, return the max supported size.
5425
5426
5427
5428
5429
5430

        Args:
            kv_manager_block_size: Block size of KV cache
            attn_groups: List of attention groups

        Returns:
5431
            The selected block size
5432
5433

        Raises:
5434
            ValueError: If no valid block size found
5435
5436
        """

5437
5438
5439
5440
5441
5442
5443
5444
        def block_size_is_supported(
            backends: list[type[AttentionBackend]], block_size: int
        ) -> bool:
            """
            Check if the block size is supported by all backends.
            """
            for backend in backends:
                is_supported = False
5445
                for supported_size in backend.get_supported_kernel_block_sizes():
5446
5447
5448
5449
5450
5451
5452
5453
5454
5455
5456
5457
5458
5459
5460
5461
5462
5463
5464
5465
5466
5467
5468
5469
5470
5471
5472
5473
5474
5475
                    if isinstance(supported_size, int):
                        if block_size == supported_size:
                            is_supported = True
                    elif isinstance(supported_size, MultipleOf):
                        if block_size % supported_size.base == 0:
                            is_supported = True
                    else:
                        raise ValueError(f"Unknown supported size: {supported_size}")
                if not is_supported:
                    return False
            return True

        backends = [group.backend for group in attn_groups]

        # Case 1: if the block_size of kv cache manager is supported by all backends,
        # return it directly
        if block_size_is_supported(backends, kv_manager_block_size):
            return kv_manager_block_size

        # Case 2: otherwise, the block_size must be an `int`-format supported size of
        # at least one backend. Iterate over all `int`-format supported sizes in
        # descending order and return the first one that is supported by all backends.
        # Simple proof:
        # If the supported size b is in MultipleOf(x_i) format for all attention
        # backends i, and b a factor of kv_manager_block_size, then
        # kv_manager_block_size also satisfies MultipleOf(x_i) for all i. We will
        # return kv_manager_block_size in case 1.
        all_int_supported_sizes = set(
            supported_size
            for backend in backends
5476
            for supported_size in backend.get_supported_kernel_block_sizes()
5477
5478
            if isinstance(supported_size, int)
        )
5479

5480
5481
5482
5483
5484
5485
        for supported_size in sorted(all_int_supported_sizes, reverse=True):
            if kv_manager_block_size % supported_size != 0:
                continue
            if block_size_is_supported(backends, supported_size):
                return supported_size
        raise ValueError(f"No common block size for {kv_manager_block_size}. ")
5486

5487
5488
5489
    def may_reinitialize_input_batch(
        self, kv_cache_config: KVCacheConfig, kernel_block_sizes: list[int]
    ) -> None:
5490
5491
5492
5493
5494
5495
5496
        """
        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.
5497
            kernel_block_sizes: The kernel block sizes for each KV cache group.
5498
5499
5500
5501
        """
        block_sizes = [
            kv_cache_group.kv_cache_spec.block_size
            for kv_cache_group in kv_cache_config.kv_cache_groups
5502
            if not isinstance(kv_cache_group.kv_cache_spec, EncoderOnlyAttentionSpec)
5503
        ]
5504
5505
5506
5507
5508
5509
5510
5511
5512
5513
5514
5515
5516
5517
5518
5519
5520
5521
        max_num_blocks = []
        max_model_len = max(self.max_model_len, self.max_encoder_len)
        for i, kv_cache_group in enumerate(kv_cache_config.kv_cache_groups):
            if isinstance(kv_cache_group.kv_cache_spec, EncoderOnlyAttentionSpec):
                continue
            max_num_blocks_per_req = cdiv(
                max_model_len, block_sizes[i] * get_total_cp_world_size()
            )
            if isinstance(kv_cache_group.kv_cache_spec, MambaSpec):
                mamba_blocks_per_req = (
                    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_per_req = max(
                    max_num_blocks_per_req, mamba_blocks_per_req
                )
            max_num_blocks.append(max_num_blocks_per_req)
5522
5523
5524
5525

        if block_sizes != [self.cache_config.block_size] or kernel_block_sizes != [
            self.cache_config.block_size
        ]:
5526
5527
5528
            assert self.cache_config.cpu_offload_gb == 0, (
                "Cannot re-initialize the input batch when CPU weight "
                "offloading is enabled. See https://github.com/vllm-project/vllm/pull/18298 "  # noqa: E501
5529
5530
                "for more details."
            )
5531
5532
            self.input_batch = InputBatch(
                max_num_reqs=self.max_num_reqs,
5533
                max_model_len=max_model_len,
5534
5535
5536
5537
5538
                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,
5539
                kernel_block_sizes=kernel_block_sizes,
5540
                max_num_blocks_per_req=max_num_blocks,
5541
                is_spec_decode=bool(self.vllm_config.speculative_config),
5542
                logitsprocs=self.input_batch.logitsprocs,
5543
                logitsprocs_need_output_token_ids=self.input_batch.logitsprocs_need_output_token_ids,
5544
                is_pooling_model=self.is_pooling_model,
5545
5546
            )

5547
    def _allocate_kv_cache_tensors(
5548
5549
        self, kv_cache_config: KVCacheConfig
    ) -> dict[str, torch.Tensor]:
5550
        """
5551
5552
5553
        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.

5554
        Args:
5555
            kv_cache_config: The KV cache config
5556
        Returns:
5557
            dict[str, torch.Tensor]: A map between layer names to their
5558
            corresponding memory buffer for KV cache.
5559
        """
5560
5561
        kv_cache_raw_tensors: dict[str, torch.Tensor] = {}
        for kv_cache_tensor in kv_cache_config.kv_cache_tensors:
5562
5563
5564
            tensor = torch.zeros(
                kv_cache_tensor.size, dtype=torch.int8, device=self.device
            )
5565
5566
5567
5568
5569
            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:
5570
5571
5572
5573
            for layer_name in group.layer_names:
                if layer_name in self.runner_only_attn_layers:
                    continue
                layer_names.add(layer_name)
5574
5575
5576
        assert layer_names == set(kv_cache_raw_tensors.keys()), (
            "Some layers are not correctly initialized"
        )
5577
5578
        return kv_cache_raw_tensors

5579
5580
5581
    def _attn_group_iterator(self) -> Iterator[AttentionGroup]:
        return itertools.chain.from_iterable(self.attn_groups)

5582
    def _kv_cache_spec_attn_group_iterator(self) -> Iterator[AttentionGroup]:
5583
5584
        if not self.kv_cache_config.kv_cache_groups:
            return
5585
5586
        for attn_groups in self.attn_groups:
            yield from attn_groups
5587

5588
5589
5590
5591
5592
5593
5594
5595
5596
5597
5598
5599
5600
5601
5602
    def _prepare_kernel_block_sizes(self, kv_cache_config: KVCacheConfig) -> list[int]:
        """
        Generate kernel_block_sizes that matches each block_size.

        For attention backends that support virtual block splitting,
        use the supported block sizes from the backend.
        For other backends (like Mamba), use the same block size (no splitting).

        Args:
            kv_cache_config: The KV cache configuration.

        Returns:
            list[int]: List of kernel block sizes for each cache group.
        """
        kernel_block_sizes = []
5603
        for kv_cache_gid, kv_cache_group in enumerate(kv_cache_config.kv_cache_groups):
5604
5605
5606
5607
5608
5609
            kv_cache_spec = kv_cache_group.kv_cache_spec
            if isinstance(kv_cache_spec, UniformTypeKVCacheSpecs):
                # All layers in the UniformTypeKVCacheSpecs have the same type,
                # Pick an arbitrary one to dispatch.
                kv_cache_spec = next(iter(kv_cache_spec.kv_cache_specs.values()))
            if isinstance(kv_cache_spec, EncoderOnlyAttentionSpec):
5610
                continue
5611
            elif isinstance(kv_cache_spec, AttentionSpec):
5612
5613
5614
                # This is an attention backend that supports virtual
                # block splitting. Get the supported block sizes from
                # all backends in the group.
5615
                attn_groups = self.attn_groups[kv_cache_gid]
5616
                kv_manager_block_size = kv_cache_group.kv_cache_spec.block_size
5617
                selected_kernel_size = self.select_common_block_size(
5618
5619
5620
                    kv_manager_block_size, attn_groups
                )
                kernel_block_sizes.append(selected_kernel_size)
5621
            elif isinstance(kv_cache_spec, MambaSpec):
5622
5623
                # This is likely Mamba or other non-attention cache,
                # no splitting.
5624
                kernel_block_sizes.append(kv_cache_spec.block_size)
5625
5626
5627
5628
5629
5630
            else:
                raise NotImplementedError(
                    f"unknown kv cache spec {kv_cache_group.kv_cache_spec}"
                )
        return kernel_block_sizes

5631
5632
5633
5634
    def _reshape_kv_cache_tensors(
        self,
        kv_cache_config: KVCacheConfig,
        kv_cache_raw_tensors: dict[str, torch.Tensor],
5635
        kernel_block_sizes: list[int],
5636
    ) -> dict[str, torch.Tensor]:
5637
        """
5638
        Reshape the KV cache tensors to the desired shape and dtype.
5639

5640
        Args:
5641
5642
            kv_cache_config: The KV cache config
            kv_cache_raw_tensors: The KV cache buffer of each layer, with
5643
                correct size but uninitialized shape.
5644
            kernel_block_sizes: The kernel block sizes for each KV cache group.
5645
        Returns:
5646
            Dict[str, torch.Tensor]: A map between layer names to their
5647
5648
            corresponding memory buffer for KV cache.
        """
5649
        kv_caches: dict[str, torch.Tensor] = {}
5650
        has_attn, has_mamba = False, False
5651
5652
        for group in self._kv_cache_spec_attn_group_iterator():
            kv_cache_spec = group.kv_cache_spec
5653
            attn_backend = group.backend
5654
5655
5656
5657
            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]
5658
            for layer_name in group.layer_names:
5659
5660
                if layer_name in self.runner_only_attn_layers:
                    continue
5661
5662
                raw_tensor = kv_cache_raw_tensors[layer_name]
                assert raw_tensor.numel() % kv_cache_spec.page_size_bytes == 0
5663
                num_blocks = raw_tensor.numel() // kv_cache_spec.page_size_bytes
5664
                if isinstance(kv_cache_spec, AttentionSpec):
5665
                    has_attn = True
5666
5667
                    num_blocks_per_kv_block = (
                        kv_cache_spec.block_size // kernel_block_size
5668
5669
5670
                    )
                    kernel_num_blocks = num_blocks * num_blocks_per_kv_block

5671
                    kv_cache_shape = attn_backend.get_kv_cache_shape(
5672
                        kernel_num_blocks,
5673
                        kernel_block_size,
5674
5675
                        kv_cache_spec.num_kv_heads,
                        kv_cache_spec.head_size,
5676
5677
                        cache_dtype_str=self.cache_config.cache_dtype,
                    )
5678
                    dtype = kv_cache_spec.dtype
5679
                    try:
5680
                        kv_cache_stride_order = attn_backend.get_kv_cache_stride_order()
5681
                        assert len(kv_cache_stride_order) == len(kv_cache_shape)
5682
                    except (AttributeError, NotImplementedError):
5683
                        kv_cache_stride_order = tuple(range(len(kv_cache_shape)))
5684
5685
5686
5687
5688
                    # 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.
5689
5690
5691
                    kv_cache_shape = tuple(
                        kv_cache_shape[i] for i in kv_cache_stride_order
                    )
5692
5693
5694
5695
5696
                    # Maintain original KV shape view.
                    inv_order = [
                        kv_cache_stride_order.index(i)
                        for i in range(len(kv_cache_stride_order))
                    ]
5697
5698
5699
5700
5701
5702
                    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
5703
                elif isinstance(kv_cache_spec, MambaSpec):
5704
                    has_mamba = True
Chen Zhang's avatar
Chen Zhang committed
5705
5706
                    raw_tensor = kv_cache_raw_tensors[layer_name]
                    state_tensors = []
5707
                    storage_offset_bytes = 0
5708
                    for shape, dtype in zip(kv_cache_spec.shapes, kv_cache_spec.dtypes):
5709
5710
                        dtype_size = get_dtype_size(dtype)
                        num_element_per_page = (
5711
5712
                            kv_cache_spec.page_size_bytes // dtype_size
                        )
Chen Zhang's avatar
Chen Zhang committed
5713
                        target_shape = (num_blocks, *shape)
5714
5715
                        stride = torch.empty(target_shape).stride()
                        target_stride = (num_element_per_page, *stride[1:])
5716
                        assert storage_offset_bytes % dtype_size == 0
5717
5718
5719
5720
                        tensor = torch.as_strided(
                            raw_tensor.view(dtype),
                            size=target_shape,
                            stride=target_stride,
5721
                            storage_offset=storage_offset_bytes // dtype_size,
5722
                        )
Chen Zhang's avatar
Chen Zhang committed
5723
                        state_tensors.append(tensor)
5724
                        storage_offset_bytes += stride[0] * dtype_size
5725
5726

                    kv_caches[layer_name] = state_tensors
5727
                else:
5728
                    raise NotImplementedError
5729
5730

        if has_attn and has_mamba:
5731
            self._update_hybrid_attention_mamba_layout(kv_caches)
5732

5733
5734
        return kv_caches

5735
    def _update_hybrid_attention_mamba_layout(
5736
5737
        self, kv_caches: dict[str, torch.Tensor]
    ) -> None:
5738
        """
5739
5740
        Update the layout of attention layers from (2, num_blocks, ...) to
        (num_blocks, 2, ...).
5741
5742

        Args:
5743
            kv_caches: The KV cache buffer of each layer.
5744
5745
        """

5746
5747
        for group in self._kv_cache_spec_attn_group_iterator():
            kv_cache_spec = group.kv_cache_spec
5748
            for layer_name in group.layer_names:
5749
                kv_cache = kv_caches[layer_name]
5750
5751
5752
5753
                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 "
5754
                        f"a tensor of shape {kv_cache.shape}"
5755
                    )
5756
                    hidden_size = kv_cache.shape[2:].numel()
5757
5758
5759
5760
                    kv_cache.as_strided_(
                        size=kv_cache.shape,
                        stride=(hidden_size, 2 * hidden_size, *kv_cache.stride()[2:]),
                    )
5761

5762
    def initialize_kv_cache_tensors(
5763
        self, kv_cache_config: KVCacheConfig, kernel_block_sizes: list[int]
5764
    ) -> dict[str, torch.Tensor]:
5765
5766
5767
5768
5769
        """
        Initialize the memory buffer for KV cache.

        Args:
            kv_cache_config: The KV cache config
5770
5771
            kernel_block_sizes: The kernel block sizes for each KV cache group.

5772
        Returns:
5773
            Dict[str, torch.Tensor]: A map between layer names to their
5774
5775
            corresponding memory buffer for KV cache.
        """
5776
5777
5778
5779
5780
5781
5782
5783
5784
5785
5786
5787
5788
5789
5790
5791
5792
5793
5794
5795
5796
5797
5798
5799

        # 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
            )
5800

5801
        # Set up cross-layer KV cache sharing
5802
5803
        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)
5804
5805
            kv_caches[layer_name] = kv_caches[target_layer_name]

5806
5807
5808
5809
5810
5811
5812
5813
5814
        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,
        )
5815
5816
5817
        return kv_caches

    def maybe_add_kv_sharing_layers_to_kv_cache_groups(
5818
5819
        self, kv_cache_config: KVCacheConfig
    ) -> None:
5820
5821
5822
5823
5824
5825
5826
5827
5828
5829
5830
5831
5832
5833
5834
5835
5836
5837
        """
        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.
5838
            attn_layers = get_layers_from_vllm_config(self.vllm_config, Attention)
5839
5840
            for layer_name in reversed(attn_layers):
                if layer_name in self.shared_kv_cache_layers:
5841
                    self.kv_sharing_fast_prefill_eligible_layers.add(layer_name)
5842
5843
                else:
                    break
5844

5845
5846
5847
5848
5849
5850
5851
    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
        """
5852
        kv_cache_config = deepcopy(kv_cache_config)
5853
        self.kv_cache_config = kv_cache_config
5854
        self.may_add_encoder_only_layers_to_kv_cache_config()
5855
        self.maybe_add_kv_sharing_layers_to_kv_cache_groups(kv_cache_config)
5856
        self.initialize_attn_backend(kv_cache_config)
5857
5858
5859
5860
5861
5862
        # 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.
        kernel_block_sizes = self._prepare_kernel_block_sizes(kv_cache_config)
5863
5864
5865
5866

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

5867
        # Reinitialize need to after initialize_attn_backend
5868
5869
5870
5871
        self.may_reinitialize_input_batch(kv_cache_config, kernel_block_sizes)
        kv_caches = self.initialize_kv_cache_tensors(
            kv_cache_config, kernel_block_sizes
        )
5872

5873
5874
5875
5876
5877
        if self.speculative_config and (
            self.speculative_config.use_eagle()
            or self.speculative_config.uses_draft_model()
        ):
            assert isinstance(self.drafter, EagleProposer | DraftModelProposer)
5878
5879
5880
5881
            # 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
5882
        if has_kv_transfer_group():
5883
            kv_transfer_group = get_kv_transfer_group()
5884
5885
5886
5887
5888
5889
5890
            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)
5891
            kv_transfer_group.set_host_xfer_buffer_ops(copy_kv_blocks)
Robert Shaw's avatar
Robert Shaw committed
5892

5893
5894
5895
5896
5897
5898
5899
5900
5901
5902
5903
5904
5905
5906
5907
5908
5909
        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,
5910
            vllm_config=self.vllm_config,
5911
5912
        )

5913
5914
5915
5916
5917
    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
5918
        encoder_only_attn_specs: dict[AttentionSpec, list[str]] = defaultdict(list)
5919
5920
5921
        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:
5922
                attn_spec: AttentionSpec = EncoderOnlyAttentionSpec(
5923
5924
5925
                    block_size=block_size,
                    num_kv_heads=attn_module.num_kv_heads,
                    head_size=attn_module.head_size,
5926
5927
                    dtype=self.kv_cache_dtype,
                )
5928
5929
5930
                encoder_only_attn_specs[attn_spec].append(layer_name)
                self.runner_only_attn_layers.add(layer_name)
        if len(encoder_only_attn_specs) > 0:
5931
5932
5933
            assert len(encoder_only_attn_specs) == 1, (
                "Only support one encoder-only attention spec now"
            )
5934
5935
            spec, layer_names = encoder_only_attn_specs.popitem()
            self.kv_cache_config.kv_cache_groups.append(
5936
5937
                KVCacheGroupSpec(layer_names=layer_names, kv_cache_spec=spec)
            )
5938

5939
    def get_kv_cache_spec(self) -> dict[str, KVCacheSpec]:
5940
        """
5941
        Generates the KVCacheSpec by parsing the kv cache format from each
5942
5943
        Attention module in the static forward context.
        Returns:
5944
            KVCacheSpec: A dictionary mapping layer names to their KV cache
5945
5946
            format. Layers that do not need KV cache are not included.
        """
5947
5948
        if has_ec_transfer() and get_ec_transfer().is_producer:
            return {}
5949
        kv_cache_spec: dict[str, KVCacheSpec] = {}
5950
5951
        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
5952
        for layer_name, attn_module in attn_layers.items():
5953
5954
5955
5956
5957
5958
5959
5960
5961
5962
5963
5964
5965
5966
5967
            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
5968

5969
        return kv_cache_spec
5970

5971
    def _to_list(self, sampled_token_ids: torch.Tensor) -> list[list[int]]:
5972
5973
5974
5975
5976
5977
5978
5979
        # 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.
5980
        pinned = self.sampled_token_ids_pinned_cpu[: sampled_token_ids.shape[0]]
5981
5982
5983
        pinned.copy_(sampled_token_ids, non_blocking=True)
        self.transfer_event.record()
        self.transfer_event.synchronize()
5984
        return pinned.tolist()
5985
5986
5987
5988
5989
5990
5991
5992
5993
5994
5995
5996
5997
5998
5999
6000
6001
6002
6003
6004
6005
6006
6007
6008
6009
6010
6011
6012
6013
6014
6015
6016
6017
6018
6019
6020
6021
6022
6023
6024
6025
6026
6027
6028
6029
6030
6031
6032
6033
6034
6035
6036
6037
6038
6039
6040
6041
6042
6043
6044
6045
6046
6047
6048
6049
6050
6051
6052
6053
6054
6055
6056
6057
6058
6059
6060

    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}

        torch.cuda.synchronize()
        start_time = time.perf_counter()

        try:
            yield
        finally:
            torch.cuda.synchronize()
            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]
                    stats.encoder_forward_time += per_request_time
                    stats.num_encoder_calls += 1


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

    encoder_forward_time: float = 0.0
    """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 {
            "encoder_forward_time": self.encoder_forward_time,
            "num_encoder_calls": self.num_encoder_calls,
        }