gpu_model_runner.py 192 KB
Newer Older
1
# SPDX-License-Identifier: Apache-2.0
2
# SPDX-FileCopyrightText: Copyright contributors to the vLLM project
3

4
import gc
5
import itertools
6
import time
7
8
from collections import defaultdict
from collections.abc import Iterator
9
from contextlib import contextmanager
10
from copy import deepcopy
11
from typing import TYPE_CHECKING, Any, NamedTuple, Optional, Union, cast
12
13
14
15
16

import numpy as np
import torch
import torch.distributed
import torch.nn as nn
17
from tqdm import tqdm
18
from typing_extensions import TypeAlias
19

20
import vllm.envs as envs
21
from vllm.attention import Attention, AttentionType
22
from vllm.attention.backends.abstract import AttentionBackend
23
from vllm.attention.layers.chunked_local_attention import ChunkedLocalAttention
24
from vllm.compilation.counter import compilation_counter
25
26
from vllm.compilation.cuda_graph import CUDAGraphWrapper
from vllm.compilation.monitor import set_cudagraph_capturing_enabled
27
28
29
30
31
32
33
from vllm.config import (
    CompilationLevel,
    CUDAGraphMode,
    VllmConfig,
    get_layers_from_vllm_config,
    update_config,
)
34
from vllm.distributed.eplb.eplb_state import EplbState
35
from vllm.distributed.kv_transfer import get_kv_transfer_group, has_kv_transfer_group
36
from vllm.distributed.kv_transfer.kv_connector.utils import copy_kv_blocks
37
from vllm.distributed.parallel_state import (
38
39
40
41
42
43
    get_pp_group,
    get_tp_group,
    graph_capture,
    is_global_first_rank,
    prepare_communication_buffer_for_model,
)
44
from vllm.forward_context import BatchDescriptor, set_forward_context
45
from vllm.logger import init_logger
46
47
from vllm.model_executor.layers.attention_layer_base import AttentionLayerBase
from vllm.model_executor.layers.mamba.abstract import MambaBase
48
from vllm.model_executor.layers.rotary_embedding import MRotaryEmbedding
49
from vllm.model_executor.model_loader import TensorizerLoader, get_model_loader
50
from vllm.model_executor.models.deepseek_v2 import DeepseekV32IndexerCache
51
52
53
54
55
56
57
58
from vllm.model_executor.models.interfaces import (
    SupportsMultiModal,
    is_mixture_of_experts,
    supports_eagle3,
    supports_mrope,
    supports_multimodal_pruning,
    supports_transcription,
)
59
from vllm.model_executor.models.interfaces_base import (
60
61
62
63
    VllmModelForPooling,
    is_pooling_model,
    is_text_generation_model,
)
64
from vllm.multimodal import MULTIMODAL_REGISTRY
65
66
67
68
69
from vllm.multimodal.inputs import (
    BatchedTensorInputs,
    MultiModalKwargsItem,
    PlaceholderRange,
)
70
from vllm.multimodal.utils import group_mm_kwargs_by_modality
71
from vllm.pooling_params import PoolingParams
72
from vllm.sampling_params import SamplingType
73
from vllm.sequence import IntermediateTensors
74
from vllm.tasks import GenerationTask, PoolingTask, SupportedTask
75
76
77
78
79
80
81
82
83
84
85
86
from vllm.utils import (
    STR_DTYPE_TO_TORCH_DTYPE,
    DeviceMemoryProfiler,
    GiB_bytes,
    cdiv,
    check_use_alibi,
    get_dtype_size,
    is_pin_memory_available,
    length_from_prompt_token_ids_or_embeds,
    round_up,
    supports_dynamo,
)
87
from vllm.utils.jsontree import json_map_leaves
88
from vllm.v1.attention.backends.flash_attn import AttentionMetadata
89
from vllm.v1.attention.backends.gdn_attn import GDNAttentionMetadataBuilder
90
from vllm.v1.attention.backends.utils import (
91
92
93
    AttentionCGSupport,
    AttentionMetadataBuilder,
    CommonAttentionMetadata,
94
    create_fast_prefill_custom_backend,
95
96
97
    reorder_batch_to_split_decodes_and_prefills,
    split_attn_metadata,
)
98
from vllm.v1.cudagraph_dispatcher import CudagraphDispatcher
99
100
101
102
103
104
105
106
107
108
109
110
111
112
113
114
115
116
117
118
119
120
121
122
from vllm.v1.kv_cache_interface import (
    AttentionSpec,
    ChunkedLocalAttentionSpec,
    CrossAttentionSpec,
    EncoderOnlyAttentionSpec,
    FullAttentionSpec,
    KVCacheConfig,
    KVCacheGroupSpec,
    KVCacheSpec,
    MambaSpec,
    MLAAttentionSpec,
    SlidingWindowSpec,
    UniformTypeKVCacheSpecs,
)
from vllm.v1.outputs import (
    EMPTY_MODEL_RUNNER_OUTPUT,
    AsyncModelRunnerOutput,
    DraftTokenIds,
    LogprobsLists,
    LogprobsTensors,
    ModelRunnerOutput,
    PoolerOutput,
    SamplerOutput,
)
123
from vllm.v1.pool.metadata import PoolingMetadata
124
from vllm.v1.sample.logits_processor import LogitsProcessors, build_logitsprocs
125
from vllm.v1.sample.metadata import SamplingMetadata
126
from vllm.v1.sample.rejection_sampler import RejectionSampler
127
from vllm.v1.sample.sampler import Sampler
128
from vllm.v1.spec_decode.eagle import EagleProposer
129
from vllm.v1.spec_decode.medusa import MedusaProposer
130
from vllm.v1.spec_decode.metadata import SpecDecodeMetadata
131
from vllm.v1.spec_decode.ngram_proposer import NgramProposer
132
from vllm.v1.structured_output.utils import apply_grammar_bitmask
133
from vllm.v1.utils import CpuGpuBuffer, record_function_or_nullcontext
134
from vllm.v1.worker.dp_utils import coordinate_batch_across_dp
135
from vllm.v1.worker.gpu_input_batch import CachedRequestState, InputBatch
136
from vllm.v1.worker.gpu_ubatch_wrapper import UBatchWrapper
137
from vllm.v1.worker.kv_connector_model_runner_mixin import KVConnectorModelRunnerMixin
138
from vllm.v1.worker.lora_model_runner_mixin import LoRAModelRunnerMixin
139
140
141
142
143
from vllm.v1.worker.ubatch_utils import (
    UBatchSlice,
    UBatchSlices,
    check_ubatch_thresholds,
)
144
from vllm.v1.worker.utils import is_residual_scattered_for_sp
145

146
147
148
149
150
151
152
153
154
from .utils import (
    AttentionGroup,
    MultiModalBudget,
    add_kv_sharing_layers_to_kv_cache_groups,
    bind_kv_cache,
    gather_mm_placeholders,
    sanity_check_mm_encoder_outputs,
    scatter_mm_placeholders,
)
155

156
if TYPE_CHECKING:
157
    from vllm.model_executor.model_loader.tensorizer import TensorizerConfig
158
    from vllm.v1.core.sched.output import SchedulerOutput
159
160
161

logger = init_logger(__name__)

162
163
AttnMetadataDict: TypeAlias = dict[str, AttentionMetadata]
# list when ubatching is enabled
164
PerLayerAttnMetadata: TypeAlias = Union[list[AttnMetadataDict], AttnMetadataDict]
165

166

167
168
169
170
171
172
173
174
175
176
177
178
179
180
181
182
183
184
185
186
187
188
189
190
# Wrapper for ModelRunnerOutput to support overlapped execution.
class AsyncGPUModelRunnerOutput(AsyncModelRunnerOutput):
    def __init__(
        self,
        model_runner_output: ModelRunnerOutput,
        sampled_token_ids: torch.Tensor,
        invalid_req_indices: list[int],
        async_output_copy_stream: torch.cuda.Stream,
    ):
        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.
        self._async_copy_ready_event = torch.cuda.Event()

        # 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

        # 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)
            self._sampled_token_ids_cpu = self._sampled_token_ids.to(
191
192
                "cpu", non_blocking=True
            )
193
194
195
196
            self._async_copy_ready_event.record()

    def get_output(self) -> ModelRunnerOutput:
        """Copy the device tensors to the host and return a ModelRunnerOutput.
197

198
199
200
201
202
203
204
205
206
207
208
209
210
211
212
213
        This function blocks until the copy is finished.
        """
        self._async_copy_ready_event.synchronize()

        # Release the device tensor once the copy has completed
        del self._sampled_token_ids

        valid_sampled_token_ids = self._sampled_token_ids_cpu.tolist()
        for i in self._invalid_req_indices:
            valid_sampled_token_ids[i].clear()

        output = self._model_runner_output
        output.sampled_token_ids = valid_sampled_token_ids
        return output


214
class GPUModelRunner(LoRAModelRunnerMixin, KVConnectorModelRunnerMixin):
215
216
    def __init__(
        self,
217
        vllm_config: VllmConfig,
218
        device: torch.device,
219
    ):
220
221
222
        self.vllm_config = vllm_config
        self.model_config = vllm_config.model_config
        self.cache_config = vllm_config.cache_config
223
        self.compilation_config = vllm_config.compilation_config
224
225
226
227
228
229
        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
230

231
        from vllm.model_executor.models.utils import set_cpu_offload_max_bytes
232
233
234
235

        set_cpu_offload_max_bytes(int(self.cache_config.cpu_offload_gb * 1024**3))
        from vllm.model_executor.layers.batch_invariant import init_batch_invariance

236
        init_batch_invariance()
237

238
239
240
241
        model_config = self.model_config
        cache_config = self.cache_config
        scheduler_config = self.scheduler_config
        parallel_config = self.parallel_config
242
        self.device = device
243
244
245
246
247
        self.pin_memory = is_pin_memory_available()
        self.dtype = self.model_config.dtype
        if cache_config.cache_dtype == "auto":
            self.kv_cache_dtype = self.dtype
        else:
248
            self.kv_cache_dtype = STR_DTYPE_TO_TORCH_DTYPE[cache_config.cache_dtype]
249

250
        self.is_pooling_model = model_config.runner_type == "pooling"
251
        self.enable_prompt_embeds = model_config.enable_prompt_embeds
252
        self.is_multimodal_raw_input_only_model = (
253
254
            model_config.is_multimodal_raw_input_only_model
        )
255
256
        # This will be overridden in load_model()
        self.is_multimodal_pruning_enabled = False
257
        self.max_model_len = model_config.max_model_len
258
        self.dcp_world_size = self.parallel_config.decode_context_parallel_size
259
        self.max_num_tokens = scheduler_config.max_num_batched_tokens
260
        self.max_num_reqs = scheduler_config.max_num_seqs
261

262
263
264
265
266
        # 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 = (
267
268
269
            self.parallel_config.distributed_executor_backend == "external_launcher"
            and len(get_pp_group().ranks) > 0
        )
270

271
        # Model-related.
272
        self.num_query_heads = model_config.get_num_attention_heads(parallel_config)
273
        self.hidden_size = model_config.get_hidden_size()
274
        self.attention_chunk_size = model_config.attention_chunk_size
275
276
        # Only relevant for models using ALiBi (e.g, MPT)
        self.use_alibi = check_use_alibi(model_config)
277

278
        self.cascade_attn_enabled = not self.model_config.disable_cascade_attn
279

280
        # Multi-modal data support
281
        self.mm_registry = MULTIMODAL_REGISTRY
282
        self.uses_mrope = model_config.uses_mrope
283
        self.supports_mm_inputs = self.mm_registry.supports_multimodal_inputs(
284
285
            model_config
        )
286

287
288
289
        if self.model_config.is_encoder_decoder:
            # Maximum length of the encoder input, only for encoder-decoder
            # models.
290
            self.max_encoder_len = scheduler_config.max_num_encoder_input_tokens
291
292
293
        else:
            self.max_encoder_len = 0

294
        # Sampler
295
        self.sampler = Sampler(logprobs_mode=self.model_config.logprobs_mode)
296

297
298
299
300
301
302
303
        self.eplb_state: Optional[EplbState] = None
        """
        State of the expert parallelism load balancer.

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

304
        # Lazy initializations
305
        # self.model: nn.Module  # Set after load_model
306
        # Initialize in initialize_kv_cache
307
        self.kv_caches: list[torch.Tensor] = []
308
309
        # indexes: [kv_cache_group_id][attn_group]
        self.attn_groups: list[list[AttentionGroup]] = []
310
311
        # self.kv_cache_config: KVCacheConfig

312
313
        # mm_hash ->  encoder_output
        self.encoder_cache: dict[str, torch.Tensor] = {}
314

315
        self.use_aux_hidden_state_outputs = False
316
317
318
319
320
321
322
323
        # 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:
            if self.speculative_config.method == "ngram":
                self.drafter = NgramProposer(self.vllm_config)
            elif self.speculative_config.use_eagle():
324
                self.drafter = EagleProposer(self.vllm_config, self.device, self)  # type: ignore
325
326
327
328
                if self.speculative_config.method == "eagle3":
                    self.use_aux_hidden_state_outputs = True
            elif self.speculative_config.method == "medusa":
                self.drafter = MedusaProposer(
329
330
                    vllm_config=self.vllm_config, device=self.device
                )  # type: ignore
331
            else:
332
333
334
335
                raise ValueError(
                    "Unknown speculative decoding method: "
                    f"{self.speculative_config.method}"
                )
336
            self.rejection_sampler = RejectionSampler()
337

338
        # Request states.
339
        self.requests: dict[str, CachedRequestState] = {}
340
        self.comm_stream = torch.cuda.Stream()
341

342
343
344
345
346
347
348
349
350
        # 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.
351
352
        self.input_batch = InputBatch(
            max_num_reqs=self.max_num_reqs,
353
354
355
            # 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),
356
357
358
            max_num_batched_tokens=self.max_num_tokens,
            device=self.device,
            pin_memory=self.pin_memory,
359
            vocab_size=self.model_config.get_vocab_size(),
360
            block_sizes=[self.cache_config.block_size],
361
            is_spec_decode=bool(self.vllm_config.speculative_config),
362
            logitsprocs=build_logitsprocs(
363
364
365
                self.vllm_config,
                self.device,
                self.pin_memory,
366
                self.is_pooling_model,
367
368
                self.vllm_config.model_config.logits_processors,
            ),
369
            is_pooling_model=self.is_pooling_model,
370
        )
371

372
        self.use_async_scheduling = self.scheduler_config.async_scheduling
373
374
375
        self.async_output_copy_stream = (
            torch.cuda.Stream() if self.use_async_scheduling else None
        )
376

377
        # TODO(woosuk): Provide an option to tune the max cudagraph batch size.
378
379
380
        # The convention is different.
        # self.cudagraph_batch_sizes sorts in ascending order.
        # The batch sizes in the config are in descending order.
381
382
383
384
        if (
            self.compilation_config.cudagraph_capture_sizes
            and self.compilation_config.cudagraph_mode != CUDAGraphMode.NONE
        ):
385
            self.cudagraph_batch_sizes = list(
386
387
                reversed(self.compilation_config.cudagraph_capture_sizes)
            )
388

389
        # Cache the device properties.
390
        self._init_device_properties()
391

392
        # Persistent buffers for CUDA graphs.
393
394
395
396
397
        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
        )
398
        self.seq_lens = self._make_buffer(self.max_num_reqs, dtype=torch.int32)
399
400
401
        # 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.
402
403
404
405
406
407
408
        self.inputs_embeds = self._make_buffer(
            self.max_num_tokens, self.hidden_size, dtype=self.dtype, numpy=False
        )
        self.is_token_ids = self._make_buffer(self.max_num_tokens, dtype=torch.bool)
        self.discard_request_indices = self._make_buffer(
            self.max_num_reqs, dtype=torch.int64
        )
409
410
        self.num_discarded_requests = 0

411
412
413
414
415
416
        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
        )
417

418
419
        # Only relevant for multimodal models
        if self.supports_mm_inputs:
420
            self.is_mm_embed = self._make_buffer(self.max_num_tokens, dtype=torch.bool)
421

422
        # Only relevant for models using M-RoPE (e.g, Qwen2-VL)
423
        if self.uses_mrope:
Roger Wang's avatar
Roger Wang committed
424
425
426
427
            # 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
428
429
430
431
432
433

            # 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
434
            self.mrope_positions = self._make_buffer(
435
436
                (3, self.max_num_tokens + 1), dtype=torch.int64
            )
437

438
439
440
441
442
443
444
445
        # CUDA event to synchronize use of reused CPU tensors between steps
        # when async scheduling is enabled.
        self.prepare_inputs_event: Optional[torch.cuda.Event] = None
        if self.use_async_scheduling:
            self.prepare_inputs_event = torch.cuda.Event()
            # Start in a completed state.
            self.prepare_inputs_event.record(torch.cuda.default_stream())

446
447
        # None in the first PP rank. The rest are set after load_model.
        self.intermediate_tensors: Optional[IntermediateTensors] = None
448

449
        # OPTIMIZATION: Cache the tensors rather than creating them every step.
450
        # Keep in int64 to avoid overflow with long context
451
452
453
454
        self.arange_np = np.arange(
            max(self.max_num_reqs + 1, self.max_model_len, self.max_num_tokens),
            dtype=np.int64,
        )
455

456
457
458
459
460
        # 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] = {}
461
462
463
464
465
        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(
466
467
                self.max_num_tokens, dtype=torch.int32, device=self.device
            )
468

469
470
471
472
473
        self.uniform_decode_query_len = (
            1
            if not self.speculative_config
            else 1 + self.speculative_config.num_speculative_tokens
        )
474
475
476
477

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

478
479
480
481
482
483
484
485
486
        self.mm_budget = (
            MultiModalBudget(
                self.model_config,
                self.scheduler_config,
                self.mm_registry,
            )
            if self.supports_mm_inputs
            else None
        )
487

488
489
        self.reorder_batch_threshold: Optional[int] = None

490
491
492
493
494
        # 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()

495
        # Cached outputs.
496
        self._draft_token_ids: Optional[Union[list[list[int]], torch.Tensor]] = None
497
498
499
500
501
        self.transfer_event = torch.cuda.Event()
        self.sampled_token_ids_pinned_cpu = torch.empty(
            (self.max_model_len, 1),
            dtype=torch.int64,
            device="cpu",
502
503
            pin_memory=self.pin_memory,
        )
504

505
506
507
508
509
510
511
512
513
514
    def _get_positions(self, num_tokens: Any):
        if isinstance(num_tokens, int):
            if self.uses_mrope:
                return self.mrope_positions.gpu[:, :num_tokens]
            return self.positions.gpu[:num_tokens]
        else:
            if self.uses_mrope:
                return self.mrope_positions.gpu[:, num_tokens]
            return self.positions.gpu[num_tokens]

515
516
517
518
519
520
521
522
523
524
    def _make_buffer(
        self, *size: Union[int, torch.SymInt], dtype: torch.dtype, numpy: bool = True
    ) -> CpuGpuBuffer:
        return CpuGpuBuffer(
            *size,
            dtype=dtype,
            device=self.device,
            pin_memory=self.pin_memory,
            with_numpy=numpy,
        )
525

526
527
528
    def _init_model_kwargs(self, num_tokens: int):
        model_kwargs = dict[str, Any]()

529
        if not self.is_pooling_model:
530
531
            return model_kwargs

532
533
        num_reqs = self.input_batch.num_reqs
        pooling_params = self.input_batch.get_pooling_params()
534
535
536

        token_type_id_requests = dict[int, Any]()
        for i, param in enumerate(pooling_params):
537
538
539
540
541
            if (
                param.extra_kwargs is not None
                and (token_types := param.extra_kwargs.get("compressed_token_type_ids"))
                is not None
            ):
542
543
544
545
546
                token_type_id_requests[i] = token_types

        if len(token_type_id_requests) == 0:
            return model_kwargs

547
        seq_lens = self.seq_lens.gpu[:num_reqs]
548
549
550
551
552
553
554
555
        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(
556
557
            device=self.device
        )
558
559
        return model_kwargs

560
    def _may_reorder_batch(self, scheduler_output: "SchedulerOutput") -> None:
561
562
        """
        Update the order of requests in the batch based on the attention
563
        backend's needs. For example, some attention backends (namely MLA) may
564
565
566
567
568
569
        want to separate requests based on if the attention computation will be
        compute-bound or memory-bound.

        Args:
            scheduler_output: The scheduler output.
        """
570
571
572
573
574
575
576
577
        # 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

578
        if self.reorder_batch_threshold is not None:
579
580
581
            # NOTE(lucas): currently no backend supports the custom masking
            #  required for DCP with q_len > 1, so we assert here. Remove this
            #  assert once the custom mask is support is added to FA3.
582
            if self.dcp_world_size > 1:
583
                assert self.reorder_batch_threshold == 1, (
584
                    "DCP not support reorder_batch_threshold > 1 now."
585
                )
586
587
588
            reorder_batch_to_split_decodes_and_prefills(
                self.input_batch,
                scheduler_output,
589
590
                decode_threshold=self.reorder_batch_threshold,
            )
591

592
593
    # Note: used for model runner override.
    def _init_device_properties(self) -> None:
594
        """Initialize attributes from torch.cuda.get_device_properties"""
595
596
597
598
599
600
601
        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()

602
    def _update_states(self, scheduler_output: "SchedulerOutput") -> None:
603
604
605
606
607
608
        """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.

609
610
        The SamplingMetadata is updated and copied to the GPU if there is a
        new/resumed/paused/finished request in the batch.
611
612
        """
        # Remove finished requests from the cached states.
613
614
        for req_id in scheduler_output.finished_req_ids:
            self.requests.pop(req_id, None)
615
616
617
618
619
620
621
        # 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:
622
            self.input_batch.remove_request(req_id)
623
624

        # Free the cached encoder outputs.
625
626
        for mm_hash in scheduler_output.free_encoder_mm_hashes:
            self.encoder_cache.pop(mm_hash, None)
627

628
629
630
631
632
633
634
635
636
637
638
639
640
        # 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()
        unscheduled_req_ids = cached_req_ids - scheduled_req_ids
        # 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:
641
            self.input_batch.remove_request(req_id)
642

643
        reqs_to_add: list[CachedRequestState] = []
644
        # Add new requests to the cached states.
645
646
647
        for new_req_data in scheduler_output.scheduled_new_reqs:
            req_id = new_req_data.req_id
            sampling_params = new_req_data.sampling_params
648
            pooling_params = new_req_data.pooling_params
649

650
651
652
653
            if (
                sampling_params
                and sampling_params.sampling_type == SamplingType.RANDOM_SEED
            ):
654
655
656
657
658
                generator = torch.Generator(device=self.device)
                generator.manual_seed(sampling_params.seed)
            else:
                generator = None

659
660
            if self.is_pooling_model:
                assert pooling_params is not None
661
662
                task = pooling_params.task
                assert task is not None, "You did not set `task` in the API"
663

664
                model = cast(VllmModelForPooling, self.get_model())
665
                to_update = model.pooler.get_pooling_updates(task)
666
667
                to_update.apply(pooling_params)

668
            req_state = CachedRequestState(
669
                req_id=req_id,
670
                prompt_token_ids=new_req_data.prompt_token_ids,
671
                prompt_embeds=new_req_data.prompt_embeds,
672
                mm_features=new_req_data.mm_features,
673
                sampling_params=sampling_params,
674
                pooling_params=pooling_params,
675
                generator=generator,
676
677
                block_ids=new_req_data.block_ids,
                num_computed_tokens=new_req_data.num_computed_tokens,
678
                output_token_ids=[],
679
                lora_request=new_req_data.lora_request,
680
            )
681
682
            self.requests[req_id] = req_state

683
            # Only relevant for models using M-RoPE (e.g, Qwen2-VL)
684
            if self.uses_mrope:
685
                self._init_mrope_positions(req_state)
686

687
            reqs_to_add.append(req_state)
688

689
        # Update the states of the running/resumed requests.
690
        is_last_rank = get_pp_group().is_last_rank
691
692
        req_data = scheduler_output.scheduled_cached_reqs
        for i, req_id in enumerate(req_data.req_ids):
693
            req_state = self.requests[req_id]
694
695
696
            num_computed_tokens = req_data.num_computed_tokens[i]
            new_block_ids = req_data.new_block_ids[i]
            resumed_from_preemption = req_data.resumed_from_preemption[i]
697
            num_output_tokens = req_data.num_output_tokens[i]
698

699
            # Update the cached states.
700

701
            req_state.num_computed_tokens = num_computed_tokens
702
703
704
705
706
707
708
709

            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.
710
711
712
                num_new_tokens = (
                    num_computed_tokens + len(new_token_ids) - req_state.num_tokens
                )
713
714
715
716
                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:
717
                    req_state.output_token_ids.extend(new_token_ids[-num_new_tokens:])
718
719
720
721
722
723
724
            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:]

                req_index = self.input_batch.req_id_to_index.get(req_id)
                if req_index is not None:
725
726
727
728
729
                    old_end_idx = self.input_batch.num_tokens_no_spec[req_index]
                    end_idx = (
                        self.input_batch.num_prompt_tokens[req_index]
                        + num_output_tokens
                    )
730
731
                    self.input_batch.num_tokens[req_index] = end_idx
                    self.input_batch.num_tokens_no_spec[req_index] = end_idx
732
733
734
                    self.input_batch.is_token_ids[req_index, end_idx:old_end_idx] = (
                        False
                    )
735

736
            # Update the block IDs.
737
            if not resumed_from_preemption:
738
739
                if new_block_ids is not None:
                    # Append the new blocks to the existing block IDs.
740
                    for block_ids, new_ids in zip(req_state.block_ids, new_block_ids):
741
                        block_ids.extend(new_ids)
742
            else:
743
                assert new_block_ids is not None
744
745
                # The request is resumed from preemption.
                # Replace the existing block IDs with the new ones.
746
                req_state.block_ids = new_block_ids
747
748
749
750
751
752

            req_index = self.input_batch.req_id_to_index.get(req_id)
            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.
753
                reqs_to_add.append(req_state)
754
755
756
                continue

            # Update the persistent batch.
757
            self.input_batch.num_computed_tokens_cpu[req_index] = num_computed_tokens
758
            if new_block_ids is not None:
759
                self.input_batch.block_table.append_row(new_block_ids, req_index)
760
761
762
763
764
765
766

            # 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)
767
                self.input_batch.token_ids_cpu[
768
769
770
                    req_index, start_token_index:end_token_index
                ] = new_token_ids
                self.input_batch.num_tokens_no_spec[req_index] = end_token_index
771
                self.input_batch.num_tokens[req_index] = end_token_index
772

773
            # Add spec_token_ids to token_ids_cpu.
774
775
776
            spec_token_ids = scheduler_output.scheduled_spec_decode_tokens.get(
                req_id, ()
            )
777
778
779
780
781
            if spec_token_ids:
                num_spec_tokens = len(spec_token_ids)
                start_index = self.input_batch.num_tokens_no_spec[req_index]
                end_token_index = start_index + num_spec_tokens
                self.input_batch.token_ids_cpu[
782
783
                    req_index, start_index:end_token_index
                ] = spec_token_ids
784
785
786
                # NOTE(woosuk): `num_tokens` here may include spec tokens.
                self.input_batch.num_tokens[req_index] += num_spec_tokens

787
788
        # Add the new or resumed requests to the persistent batch.
        # The smaller empty indices are filled first.
789
790
        for request in reqs_to_add:
            self.input_batch.add_request(request)
791

792
793
794
795
796
797
        # 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()
798

799
    def _update_states_after_model_execute(
800
801
        self, output_token_ids: torch.Tensor
    ) -> None:
802
803
804
805
806
807
808
809
810
811
812
813
        """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.
        """
        if not self.model_config.is_hybrid or not self.speculative_config:
            return

        # Find the number of accepted tokens for each sequence.
814
815
816
817
818
819
820
821
822
823
824
825
826
827
828
829
830
831
832
833
        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()
        )
834
835
836
        for i, num_tokens in enumerate(num_accepted_tokens):
            self.input_batch.num_accepted_tokens_cpu[i] = num_tokens

837
838
839
840
841
842
    def _init_mrope_positions(self, req_state: CachedRequestState):
        image_grid_thw = []
        video_grid_thw = []
        second_per_grid_ts = []
        audio_feature_lengths = []
        use_audio_in_video = False
843
844
845
846
        for mm_feature in req_state.mm_features:
            mm_item = mm_feature.data
            if mm_item is None:
                continue
847
848
849
850
851
852
853
854
855
856
857
858
            mm_input = mm_item.get_data()
            if (t := mm_input.get("image_grid_thw")) is not None:
                image_grid_thw.append(t.tolist())
            if (t := mm_input.get("video_grid_thw")) is not None:
                video_grid_thw.append(t.tolist())
            if (t := mm_input.get("second_per_grid_ts")) is not None:
                second_per_grid_ts.append(t)
            if (t := mm_input.get("audio_feature_lengths")) is not None:
                audio_feature_lengths.append(t)
            if mm_input.get("use_audio_in_video") is True:
                use_audio_in_video = True

859
        if supports_mrope(self.model):
860
            req_state.mrope_positions, req_state.mrope_position_delta = (
861
862
863
864
865
866
867
868
869
                self.model.get_mrope_input_positions(
                    req_state.prompt_token_ids,
                    hf_config=self.model_config.hf_config,
                    image_grid_thw=image_grid_thw,
                    video_grid_thw=video_grid_thw,
                    second_per_grid_ts=second_per_grid_ts,
                    audio_feature_lengths=audio_feature_lengths,
                    use_audio_in_video=use_audio_in_video,
                )
870
            )
871
        else:
872
            req_state.mrope_positions, req_state.mrope_position_delta = (
873
874
875
876
877
878
879
880
881
                MRotaryEmbedding.get_input_positions_tensor(
                    req_state.prompt_token_ids,
                    hf_config=self.model_config.hf_config,
                    image_grid_thw=image_grid_thw,
                    video_grid_thw=video_grid_thw,
                    second_per_grid_ts=second_per_grid_ts,
                    audio_feature_lengths=audio_feature_lengths,
                    use_audio_in_video=use_audio_in_video,
                )
882
            )
883

884
    def _extract_mm_kwargs(
885
        self,
886
887
        scheduler_output: "SchedulerOutput",
    ) -> BatchedTensorInputs:
888
        if not scheduler_output or not self.is_multimodal_raw_input_only_model:
889
            return {}
890

891
892
        mm_kwargs = list[MultiModalKwargsItem]()
        for req in scheduler_output.scheduled_new_reqs:
893
894
895
            for feature in req.mm_features:
                if feature.data is not None:
                    mm_kwargs.append(feature.data)
896

897
        # Input all modalities at once
898
        model = cast(SupportsMultiModal, self.model)
899
900
        mm_kwargs_combined: BatchedTensorInputs = {}
        for _, _, mm_kwargs_group in group_mm_kwargs_by_modality(
901
902
903
904
            mm_kwargs,
            device=self.device,
            pin_memory=self.pin_memory,
            merge_by_field_config=model.merge_by_field_config,
905
906
        ):
            mm_kwargs_combined.update(mm_kwargs_group)
907

908
        return mm_kwargs_combined
909

910
    def _dummy_mm_kwargs(self, num_seqs: int) -> BatchedTensorInputs:
911
        if not self.is_multimodal_raw_input_only_model:
912
            return {}
913

914
915
916
917
918
        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)
919

920
921
922
923
924
925
926
927
928
929
930
931
932
933
934
935
936
937
938
939
    def _get_cumsum_and_arange(
        self,
        num_tokens: np.ndarray,
        cumsum_dtype: Optional[np.dtype] = None,
    ) -> 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

940
941
942
    def _prepare_input_ids(
        self, total_num_scheduled_tokens: int, cu_num_tokens: np.ndarray
    ) -> None:
943
        """Prepare the input IDs for the current batch.
944

945
946
947
948
949
950
951
        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)
952
953
954
            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)
955
956
957
958
959
960
961
962
963
964
965
966
967
968
969
970
971
972
            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
        flattened_indices = []
        prev_common_req_indices = []
        indices_match = True
        max_flattened_index = -1
        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.
                flattened_index = cu_num_tokens[cur_index].item() - 1
                flattened_indices.append(flattened_index)
973
                indices_match &= prev_index == flattened_index
974
975
976
977
978
979
                max_flattened_index = max(max_flattened_index, flattened_index)
        num_commmon_tokens = len(flattened_indices)
        if num_commmon_tokens < total_num_scheduled_tokens:
            # 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)
980
981
982
            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)
983
984
985
986
987
988
989
990
991
992
        if num_commmon_tokens == 0:
            # No requests in common with the previous iteration
            # So input_ids_cpu will have all the input ids.
            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_(
993
994
995
                self.input_batch.prev_sampled_token_ids[:num_commmon_tokens, 0],
                non_blocking=True,
            )
996
997
            if self.enable_prompt_embeds:
                self.is_token_ids.gpu[:num_commmon_tokens] = True
998
999
1000
            return
        # Upload the index tensors asynchronously
        # so the scatter can be non-blocking.
1001
1002
1003
        input_ids_index_tensor = torch.tensor(
            flattened_indices, dtype=torch.int64, pin_memory=self.pin_memory
        ).to(self.device, non_blocking=True)
1004
        prev_common_req_indices_tensor = torch.tensor(
1005
1006
            prev_common_req_indices, dtype=torch.int64, pin_memory=self.pin_memory
        ).to(self.device, non_blocking=True)
1007
1008
1009
1010
        self.input_ids.gpu.scatter_(
            dim=0,
            index=input_ids_index_tensor,
            src=self.input_batch.prev_sampled_token_ids[
1011
1012
1013
                prev_common_req_indices_tensor, 0
            ],
        )
1014

1015
1016
1017
1018
1019
1020
1021
1022
1023
1024
1025
1026
1027
1028
1029
1030
1031
1032
    def _get_encoder_seq_lens(
        self,
        scheduler_output: "SchedulerOutput",
        kv_cache_spec: KVCacheSpec,
        num_reqs: int,
    ) -> Optional[np.ndarray]:
        if not isinstance(kv_cache_spec, CrossAttentionSpec):
            return None

        # Build encoder_seq_lens array mapping request indices to
        # encoder lengths for inputs scheduled in this batch
        encoder_seq_lens = np.zeros(num_reqs, dtype=np.int32)
        for req_id in scheduler_output.scheduled_encoder_inputs:
            req_index = self.input_batch.req_id_to_index[req_id]
            encoder_seq_lens[req_index] = self.max_encoder_len

        return encoder_seq_lens

1033
    def _prepare_inputs(
1034
        self, scheduler_output: "SchedulerOutput"
1035
1036
1037
1038
1039
1040
1041
1042
1043
1044
1045
    ) -> tuple[
        PerLayerAttnMetadata,
        torch.Tensor,
        Optional[SpecDecodeMetadata],
        np.ndarray,
        Optional[CommonAttentionMetadata],
        int,
        Optional[UBatchSlices],
        Optional[torch.Tensor],
        bool,
    ]:
1046
1047
1048
        """
        :return: tuple[
            attn_metadata: layer-to-attention_metadata mapping,
1049
1050
1051
            logits_indices, spec_decode_metadata,
            num_scheduled_tokens, spec_decode_common_attn_metadata,
            max_num_scheduled_tokens, use_cascade_attn
1052
1053
        ]
        """
1054
1055
1056
1057
1058
1059
1060
        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.
1061
        self.input_batch.block_table.commit_block_table(num_reqs)
1062
1063

        # Get the number of scheduled tokens for each request.
1064
1065
1066
1067
        req_ids = self.input_batch.req_ids
        tokens = [scheduler_output.num_scheduled_tokens[i] for i in req_ids]
        num_scheduled_tokens = np.array(tokens, dtype=np.int32)
        max_num_scheduled_tokens = max(tokens)
1068
1069
1070

        # Get request indices.
        # E.g., [2, 5, 3] -> [0, 0, 1, 1, 1, 1, 1, 2, 2, 2]
1071
        req_indices = np.repeat(self.arange_np[:num_reqs], num_scheduled_tokens)
1072

1073
1074
        # cu_num_tokens: [2, 5, 3] -> [2, 7, 10]
        # arange: [0, 1, 0, 1, 2, 3, 4, 0, 1, 2]
1075
        cu_num_tokens, arange = self._get_cumsum_and_arange(num_scheduled_tokens)
1076
1077

        # Get positions.
1078
        positions_np = self.positions.np[:total_num_scheduled_tokens]
1079
1080
1081
1082
1083
        np.add(
            self.input_batch.num_computed_tokens_cpu[req_indices],
            arange,
            out=positions_np,
        )
1084

1085
1086
        # Calculate M-RoPE positions.
        # Only relevant for models using M-RoPE (e.g, Qwen2-VL)
1087
        if self.uses_mrope:
1088
1089
            self._calc_mrope_positions(scheduler_output)

1090
1091
1092
1093
        # 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.
1094
1095
1096
        token_indices = (
            positions_np + req_indices * self.input_batch.token_ids_cpu.shape[1]
        )
1097
        token_indices_tensor = torch.from_numpy(token_indices)
1098

1099
1100
1101
        # 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.
1102
1103
1104
1105
1106
1107
        torch.index_select(
            self.input_batch.token_ids_cpu_tensor.flatten(),
            0,
            token_indices_tensor,
            out=self.input_ids.cpu[:total_num_scheduled_tokens],
        )
1108
1109
1110
1111
1112
1113
        if self.enable_prompt_embeds:
            is_token_ids = self.input_batch.is_token_ids.flatten()
            torch.index_select(
                is_token_ids,
                0,
                token_indices_tensor,
1114
1115
                out=self.is_token_ids.cpu[:total_num_scheduled_tokens],
            )
1116
1117
1118
1119
1120
1121
1122
1123
1124
1125
1126
1127
1128
1129
1130
1131
1132
1133
1134
1135
1136
1137
1138
1139
1140
1141
1142
1143
1144
1145
1146
1147
1148

        # 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:
1149
1150
1151
                    self.inputs_embeds.cpu[
                        output_idx : output_idx + actual_num_sched
                    ].copy_(req_embeds[start_pos:actual_end])
1152
1153

                output_idx += num_sched
1154

1155
1156
        self.input_batch.block_table.compute_slot_mapping(req_indices, positions_np)
        self.input_batch.block_table.commit_slot_mapping(total_num_scheduled_tokens)
1157
1158

        # Prepare the attention metadata.
1159
        self.query_start_loc.np[0] = 0
1160
        self.query_start_loc.np[1 : num_reqs + 1] = cu_num_tokens
1161
1162
        # Note: pad query_start_loc to be non-decreasing, as kernels
        # like FlashAttention requires that
1163
        self.query_start_loc.np[num_reqs + 1 :].fill(cu_num_tokens[-1])
1164
        self.query_start_loc.copy_to_gpu()
1165
        query_start_loc = self.query_start_loc.gpu[: num_reqs + 1]
1166

1167
        num_tokens_unpadded = scheduler_output.total_num_scheduled_tokens
1168
        num_tokens_padded = self._get_num_input_tokens(num_tokens_unpadded)
1169
1170
1171
        uniform_decode = (
            max_num_scheduled_tokens == self.uniform_decode_query_len
        ) and (total_num_scheduled_tokens == num_reqs * max_num_scheduled_tokens)
1172
        ubatch_slices, num_tokens_across_dp = coordinate_batch_across_dp(
1173
1174
1175
            num_scheduled_tokens,
            num_tokens_unpadded,
            num_tokens_padded,
1176
1177
1178
            self.parallel_config,
            True,
            uniform_decode,
1179
        )
1180

1181
        self.seq_lens.np[:num_reqs] = (
1182
1183
            self.input_batch.num_computed_tokens_cpu[:num_reqs] + num_scheduled_tokens
        )
1184
        # Fill unused with 0 for full cuda graph mode.
1185
1186
1187
1188
        self.seq_lens.np[num_reqs:].fill(0)
        self.seq_lens.copy_to_gpu()
        seq_lens = self.seq_lens.gpu[:num_reqs]
        max_seq_len = self.seq_lens.np[:num_reqs].max().item()
1189

1190
        num_tokens = [self.requests[r].num_tokens for r in self.input_batch.req_ids]
1191
1192
1193
1194
1195
1196
1197
        num_tokens_np = np.array(num_tokens, dtype=np.int32)

        # Record the index of requests that should not be sampled,
        # so that we could clear the sampled tokens before returning
        discard_requests_mask = self.seq_lens.np[:num_reqs] < num_tokens_np
        discard_request_indices = np.nonzero(discard_requests_mask)[0]
        self.num_discarded_requests = len(discard_request_indices)
1198
1199
1200
        self.discard_request_indices.np[: self.num_discarded_requests] = (
            discard_request_indices
        )
1201
1202
1203

        self.discard_request_indices.copy_to_gpu(self.num_discarded_requests)

1204
        # Copy the tensors to the GPU.
1205
1206
        self._prepare_input_ids(total_num_scheduled_tokens, cu_num_tokens)

1207
        if self.uses_mrope:
1208
            # Only relevant for models using M-RoPE (e.g, Qwen2-VL)
1209
1210
            self.mrope_positions.gpu[:, :total_num_scheduled_tokens].copy_(
                self.mrope_positions.cpu[:, :total_num_scheduled_tokens],
1211
1212
                non_blocking=True,
            )
1213
1214
        else:
            # Common case (1D positions)
1215
            self.positions.copy_to_gpu(total_num_scheduled_tokens)
1216

1217
        use_spec_decode = len(scheduler_output.scheduled_spec_decode_tokens) > 0
1218
1219
1220
1221
1222
1223
1224
        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
1225
            num_draft_tokens = None
1226
1227
1228
1229
1230
1231
            spec_decode_metadata = None
        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)
1232
1233
1234
            # 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)
1235
1236
1237
1238
            for (
                req_id,
                draft_token_ids,
            ) in scheduler_output.scheduled_spec_decode_tokens.items():
1239
1240
                req_idx = self.input_batch.req_id_to_index[req_id]
                num_draft_tokens[req_idx] = len(draft_token_ids)
1241
1242
1243
1244
1245
1246
1247
1248
                num_decode_draft_tokens[req_idx] = (
                    len(draft_token_ids)
                    if (
                        self.input_batch.num_computed_tokens_cpu[req_idx]
                        >= self.input_batch.num_prompt_tokens[req_idx]
                    )
                    else -1
                )
1249
            spec_decode_metadata = self._calc_spec_decode_metadata(
1250
1251
                num_draft_tokens, cu_num_tokens
            )
1252
            logits_indices = spec_decode_metadata.logits_indices
1253
1254

            # For DECODE only cuda graph of some attention backends (e.g., GDN).
1255
            self.num_decode_draft_tokens.np[:num_reqs] = num_decode_draft_tokens
1256
1257
            self.num_decode_draft_tokens.np[num_reqs:].fill(-1)
            self.num_decode_draft_tokens.copy_to_gpu()
1258
1259
1260

        logits_indices_padded = None
        if self.cache_config.kv_sharing_fast_prefill:
1261
            logits_indices_padded = self._prepare_kv_sharing_fast_prefill(
1262
1263
                logits_indices
            )
1264

1265
1266
1267
        attn_metadata: PerLayerAttnMetadata = {}
        if ubatch_slices is not None:
            attn_metadata = [dict() for _ in range(len(ubatch_slices))]
1268
        use_cascade_attn = False
1269

1270
        # Used in the below loop.
1271
        query_start_loc_cpu = self.query_start_loc.cpu[: num_reqs + 1]
1272
        seq_lens_cpu = self.seq_lens.cpu[:num_reqs]
1273
1274
1275
        num_computed_tokens_cpu = self.input_batch.num_computed_tokens_cpu_tensor[
            :num_reqs
        ]
1276
        spec_decode_common_attn_metadata = None
1277
1278
        if use_spec_decode:
            self.num_accepted_tokens.np[:num_reqs] = (
1279
1280
                self.input_batch.num_accepted_tokens_cpu[:num_reqs]
            )
1281
1282
            self.num_accepted_tokens.np[num_reqs:].fill(1)
            self.num_accepted_tokens.copy_to_gpu()
1283

1284
1285
1286
        # Prepare the attention metadata for each KV cache group and make layers
        # in the same group share the same metadata.
        for kv_cache_group_id, kv_cache_group_spec in enumerate(
1287
1288
            self.kv_cache_config.kv_cache_groups
        ):
1289
            encoder_seq_lens = self._get_encoder_seq_lens(
1290
1291
                scheduler_output, kv_cache_group_spec.kv_cache_spec, num_reqs
            )
1292

1293
            if isinstance(kv_cache_group_spec.kv_cache_spec, EncoderOnlyAttentionSpec):
1294
1295
1296
1297
1298
                # Encoder-only layers do not have KV cache, so we need to
                # create a dummy block table and slot mapping for them.
                blk_table_tensor = torch.zeros(
                    (num_reqs, 1),
                    dtype=torch.int32,
1299
1300
1301
                    device=self.device,
                )
                slot_mapping = torch.zeros(
1302
                    (total_num_scheduled_tokens,),
1303
1304
1305
                    dtype=torch.int64,
                    device=self.device,
                )
1306
1307
1308
                num_common_prefix_blocks = 0
            else:
                blk_table = self.input_batch.block_table[kv_cache_group_id]
1309
                blk_table_tensor = blk_table.get_device_tensor(num_reqs)
1310
                slot_mapping = blk_table.slot_mapping.gpu[:total_num_scheduled_tokens]
1311
1312
1313

                # Fill unused with -1. Needed for reshape_and_cache in full cuda
                # graph mode.
1314
1315
1316
1317
                blk_table.slot_mapping.gpu[total_num_scheduled_tokens:].fill_(-1)
                num_common_prefix_blocks = scheduler_output.num_common_prefix_blocks[
                    kv_cache_group_id
                ]
1318

1319
            common_attn_metadata = CommonAttentionMetadata(
1320
1321
1322
1323
1324
                query_start_loc=query_start_loc,
                query_start_loc_cpu=query_start_loc_cpu,
                seq_lens=seq_lens,
                seq_lens_cpu=seq_lens_cpu,
                num_computed_tokens_cpu=num_computed_tokens_cpu,
1325
1326
1327
                num_reqs=num_reqs,
                num_actual_tokens=total_num_scheduled_tokens,
                max_query_len=max_num_scheduled_tokens,
1328
                max_seq_len=max_seq_len,
1329
1330
                block_table_tensor=blk_table_tensor,
                slot_mapping=slot_mapping,
1331
1332
                logits_indices_padded=logits_indices_padded,
                num_logits_indices=logits_indices.size(0),
1333
                causal=True,
1334
                encoder_seq_lens=encoder_seq_lens,
1335
1336
            )

1337
            if self.speculative_config and spec_decode_common_attn_metadata is None:
1338
                if isinstance(self.drafter, EagleProposer):
1339
1340
1341
1342
                    if (
                        self.drafter.attn_layer_names[0]
                        in kv_cache_group_spec.layer_names
                    ):
1343
1344
1345
                        spec_decode_common_attn_metadata = common_attn_metadata
                else:
                    spec_decode_common_attn_metadata = common_attn_metadata
1346

1347
1348
1349
            for attn_group in self.attn_groups[kv_cache_group_id]:
                # Prepare for cascade attention if enabled & beneficial.
                common_prefix_len = 0
1350
                builder = attn_group.get_metadata_builder()
1351
1352
1353
                if self.cascade_attn_enabled:
                    common_prefix_len = self._compute_cascade_attn_prefix_len(
                        num_scheduled_tokens,
1354
                        num_common_prefix_blocks,
1355
                        attn_group.kv_cache_spec,
1356
1357
                        builder,
                    )
1358

1359
                extra_attn_metadata_args = {}
1360
                if use_spec_decode and isinstance(builder, GDNAttentionMetadataBuilder):
1361
                    extra_attn_metadata_args = dict(
1362
1363
1364
1365
                        num_accepted_tokens=self.num_accepted_tokens.gpu[:num_reqs],
                        num_decode_draft_tokens_cpu=self.num_decode_draft_tokens.cpu[
                            :num_reqs
                        ],
1366
1367
                    )

1368
1369
                if ubatch_slices is not None:
                    common_attn_metadata_list = split_attn_metadata(
1370
1371
                        ubatch_slices, common_attn_metadata
                    )
1372
                    for ubid, common_attn_metadata in enumerate(
1373
1374
1375
1376
1377
1378
1379
1380
                        common_attn_metadata_list
                    ):
                        attn_metadata_i = attn_group.get_metadata_builder(
                            ubatch_id=ubid
                        ).build(
                            common_prefix_len=common_prefix_len,
                            common_attn_metadata=common_attn_metadata,
                        )
1381
1382
1383
1384
1385
1386
1387
1388
                        for layer_name in kv_cache_group_spec.layer_names:
                            assert type(attn_metadata) is list
                            attn_metadata[ubid][layer_name] = attn_metadata_i
                else:
                    assert isinstance(attn_metadata, dict)
                    attn_metadata_i = builder.build(
                        common_prefix_len=common_prefix_len,
                        common_attn_metadata=common_attn_metadata,
1389
1390
1391
                        **extra_attn_metadata_args,
                    )
                    use_cascade_attn |= getattr(attn_metadata_i, "use_cascade", False)
1392
1393
                    for layer_name in attn_group.layer_names:
                        attn_metadata[layer_name] = attn_metadata_i
1394

1395
1396
1397
1398
        # disable cascade attention when DBO
        if ubatch_slices is not None:
            use_cascade_attn = False

1399
1400
1401
1402
        # Hot-Swap lora model
        if self.lora_config:
            self.set_active_loras(self.input_batch, num_scheduled_tokens)

1403
1404
1405
1406
1407
1408
1409
1410
        return (
            attn_metadata,
            logits_indices,
            spec_decode_metadata,
            num_scheduled_tokens,
            spec_decode_common_attn_metadata,
            max_num_scheduled_tokens,
            ubatch_slices,
1411
            num_tokens_across_dp,
1412
1413
            use_cascade_attn,
        )
1414

1415
1416
1417
1418
    def _compute_cascade_attn_prefix_len(
        self,
        num_scheduled_tokens: np.ndarray,
        num_common_prefix_blocks: int,
1419
1420
        kv_cache_spec: KVCacheSpec,
        attn_metadata_builder: AttentionMetadataBuilder,
1421
1422
1423
1424
1425
1426
1427
1428
1429
1430
1431
1432
1433
1434
1435
1436
1437
1438
    ) -> 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.
        """
1439
        common_prefix_len = num_common_prefix_blocks * kv_cache_spec.block_size
1440
1441
1442
1443
1444
1445
1446
1447
1448
1449
1450
1451
1452
1453
1454
1455
1456
1457
1458
1459
1460
1461
1462
1463
1464
1465
1466
1467
1468
1469
1470
1471
1472
1473
1474
1475
1476
        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]
1477
        # Request 3's num_computed_tokens: 3 (i.e., [A, B, C])
1478
1479
1480
1481
1482
1483
1484
        # 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.
        num_reqs = len(num_scheduled_tokens)
        common_prefix_len = min(
1485
1486
            common_prefix_len, self.input_batch.num_computed_tokens_cpu[:num_reqs].min()
        )
1487
        # common_prefix_len should be a multiple of the block size.
1488
1489
1490
1491
1492
1493
1494
1495
1496
1497
1498
        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
        )
1499
1500
        assert isinstance(kv_cache_spec, AttentionSpec)
        use_cascade = attn_metadata_builder.use_cascade_attention(
1501
1502
1503
            common_prefix_len=common_prefix_len,
            query_lens=num_scheduled_tokens,
            num_query_heads=self.num_query_heads,
1504
            num_kv_heads=kv_cache_spec.num_kv_heads,
1505
            use_alibi=self.use_alibi,
1506
            use_sliding_window=use_sliding_window,
1507
            use_local_attention=use_local_attention,
1508
1509
1510
1511
            num_sms=self.num_sms,
        )
        return common_prefix_len if use_cascade else 0

1512
1513
    def _calc_mrope_positions(self, scheduler_output: "SchedulerOutput"):
        mrope_pos_ptr = 0
1514
        for index, req_id in enumerate(self.input_batch.req_ids):
1515
1516
1517
            req = self.requests[req_id]
            assert req.mrope_positions is not None

1518
1519
            num_computed_tokens = self.input_batch.num_computed_tokens_cpu[index]
            num_scheduled_tokens = scheduler_output.num_scheduled_tokens[req_id]
1520
            num_prompt_tokens = length_from_prompt_token_ids_or_embeds(
1521
1522
                req.prompt_token_ids, req.prompt_embeds
            )
1523
1524

            if num_computed_tokens + num_scheduled_tokens > num_prompt_tokens:
1525
1526
                prompt_part_len = max(0, num_prompt_tokens - num_computed_tokens)
                completion_part_len = max(0, num_scheduled_tokens - prompt_part_len)
1527
1528
1529
1530
1531
1532
1533
1534
1535
1536
1537
1538
1539
            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

1540
1541
1542
                self.mrope_positions.cpu[:, dst_start:dst_end] = req.mrope_positions[
                    :, src_start:src_end
                ]
1543
1544
1545
1546
1547
1548
1549
                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

1550
                MRotaryEmbedding.get_next_input_positions_tensor(
1551
                    out=self.mrope_positions.np,
1552
1553
1554
1555
1556
                    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,
                )
1557
1558
1559

                mrope_pos_ptr += completion_part_len

1560
1561
    def _calc_spec_decode_metadata(
        self,
1562
1563
1564
1565
1566
1567
1568
1569
1570
1571
1572
1573
1574
1575
1576
1577
        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
1578
1579
1580
1581

        # 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(
1582
1583
            num_sampled_tokens, cumsum_dtype=np.int32
        )
1584
        # Step 2. [0, 0, 0, 0, 103, 104, 104, 104, 206, 207, 207]
1585
        logits_indices = np.repeat(
1586
1587
            cu_num_scheduled_tokens - num_sampled_tokens, num_sampled_tokens
        )
1588
        # Step 3. [0, 1, 2, 3, 103, 104, 105, 106, 206, 207, 208]
1589
1590
1591
1592
1593
1594
        logits_indices += arange

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

        # Compute the draft logits indices.
1595
1596
1597
        # 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(
1598
1599
            num_draft_tokens, cumsum_dtype=np.int32
        )
1600
1601
        # [0, 0, 0, 5, 5, 9]
        target_logits_indices = np.repeat(
1602
1603
            cu_num_sampled_tokens - num_sampled_tokens, num_draft_tokens
        )
1604
1605
1606
1607
1608
        # [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(
1609
1610
1611
1612
1613
            self.device, non_blocking=True
        )
        logits_indices = torch.from_numpy(logits_indices).to(
            self.device, non_blocking=True
        )
1614
        target_logits_indices = torch.from_numpy(target_logits_indices).to(
1615
1616
            self.device, non_blocking=True
        )
1617
        bonus_logits_indices = torch.from_numpy(bonus_logits_indices).to(
1618
1619
            self.device, non_blocking=True
        )
1620

1621
1622
        # Compute the draft token ids.
        # draft_token_indices:      [  1,   2,   3, 105, 106, 208]
1623
        draft_token_ids = self.input_ids.gpu[logits_indices]
1624
1625
1626
1627
1628
1629
1630
1631
1632
1633
1634
1635
        draft_token_ids = draft_token_ids[target_logits_indices + 1]

        metadata = SpecDecodeMetadata(
            draft_token_ids=draft_token_ids,
            num_draft_tokens=num_draft_tokens.tolist(),
            cu_num_draft_tokens=cu_num_draft_tokens,
            target_logits_indices=target_logits_indices,
            bonus_logits_indices=bonus_logits_indices,
            logits_indices=logits_indices,
        )
        return metadata

1636
1637
1638
1639
1640
1641
1642
    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
1643
        self.kv_sharing_fast_prefill_logits_indices[:num_logits].copy_(logits_indices)
1644
1645
1646
1647
1648
        # 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_(
1649
1650
1651
1652
1653
1654
            logits_indices[-1].item()
        )
        if (
            self.compilation_config.cudagraph_mode != CUDAGraphMode.NONE
            and num_logits <= self.cudagraph_batch_sizes[-1]
        ):
1655
1656
1657
1658
1659
            # Use piecewise CUDA graphs.
            # Add padding to the batch size.
            num_logits_padded = self.vllm_config.pad_for_cudagraph(num_logits)
        else:
            num_logits_padded = num_logits
1660
1661
1662
        logits_indices_padded = self.kv_sharing_fast_prefill_logits_indices[
            :num_logits_padded
        ]
1663
1664
        return logits_indices_padded

1665
1666
1667
1668
1669
1670
1671
1672
    def _batch_mm_kwargs_from_scheduler(
        self,
        scheduler_output: "SchedulerOutput",
    ) -> tuple[list[MultiModalKwargsItem], list[tuple[str, PlaceholderRange]]]:
        """Batch multimodal kwargs from scheduled encoder inputs.

        Args:
            scheduler_output: The scheduler output containing scheduled encoder
1673
                inputs.
1674
1675
1676
1677
1678
1679

        Returns:
            A tuple of (mm_kwargs, req_ids_pos) where:
            - mm_kwargs: List of multimodal kwargs items to be batched
            - mm_hashes_pos: List of (mm_hash, position_info) tuples
        """
1680
1681
        scheduled_encoder_inputs = scheduler_output.scheduled_encoder_inputs
        if not scheduled_encoder_inputs:
1682
            return [], []
1683
        # Batch the multi-modal inputs.
1684
        mm_kwargs = list[MultiModalKwargsItem]()
1685
1686
        # list of tuple (mm_hash, position_info)
        mm_hashes_pos = list[tuple[str, PlaceholderRange]]()
1687
1688
        for req_id, encoder_input_ids in scheduled_encoder_inputs.items():
            req_state = self.requests[req_id]
1689
1690

            for mm_input_id in encoder_input_ids:
1691
1692
1693
1694
                mm_feature = req_state.mm_features[mm_input_id]
                mm_hash = mm_feature.identifier
                mm_kwargs.append(mm_feature.data)
                mm_hashes_pos.append((mm_hash, mm_feature.mm_position))
1695

1696
1697
1698
1699
1700
        return mm_kwargs, mm_hashes_pos

    def _execute_mm_encoder(self, scheduler_output: "SchedulerOutput"):
        # Batch the multi-modal inputs using the helper method.
        mm_kwargs, mm_hashes_pos = self._batch_mm_kwargs_from_scheduler(
1701
1702
            scheduler_output
        )
1703
1704
1705
1706

        if not mm_kwargs:
            return

1707
1708
1709
1710
1711
1712
1713
        # 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.
1714
        model = cast(SupportsMultiModal, self.model)
1715
        encoder_outputs = []
1716
        for modality, num_items, mm_kwargs_group in group_mm_kwargs_by_modality(
1717
1718
1719
1720
            mm_kwargs,
            device=self.device,
            pin_memory=self.pin_memory,
            merge_by_field_config=model.merge_by_field_config,
1721
        ):
1722
1723
1724
1725
1726
1727
1728
1729
1730
1731
1732
1733
            # (ekhvedchenia): Temporary hack to limit peak memory usage when
            # processing multimodal data.This solves the issue with scheduler
            # 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)
            curr_group_outputs = []

            if self.is_multimodal_pruning_enabled and modality == "video":
                micro_batch_size = 1
                for i in range(0, num_items, micro_batch_size):
                    micro_batch_mm_inputs = dict(
1734
1735
1736
                        (k, v[i : i + micro_batch_size])
                        for k, v in mm_kwargs_group.items()
                    )
1737
1738

                    micro_batch_outputs = model.get_multimodal_embeddings(
1739
1740
                        **micro_batch_mm_inputs
                    )
1741
1742
1743
1744
1745
1746
1747
1748
1749
1750

                    curr_group_outputs.extend(micro_batch_outputs)
            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.
1751
                curr_group_outputs = model.get_multimodal_embeddings(**mm_kwargs_group)
1752

1753
1754
            sanity_check_mm_encoder_outputs(
                curr_group_outputs,
1755
                expected_num_items=num_items,
1756
            )
1757
            encoder_outputs.extend(curr_group_outputs)
1758

1759
1760
1761
        # Cache the encoder outputs by mm_hash
        for (mm_hash, pos_info), output in zip(mm_hashes_pos, encoder_outputs):
            self.encoder_cache[mm_hash] = scatter_mm_placeholders(
1762
1763
1764
1765
1766
                output,
                is_embed=pos_info.is_embed,
            )

    def _gather_mm_embeddings(
1767
1768
        self,
        scheduler_output: "SchedulerOutput",
1769
        shift_computed_tokens: int = 0,
1770
1771
1772
1773
1774
1775
1776
1777
    ) -> tuple[list[torch.Tensor], torch.Tensor]:
        total_num_scheduled_tokens = scheduler_output.total_num_scheduled_tokens

        mm_embeds = list[torch.Tensor]()
        is_mm_embed = self.is_mm_embed.cpu
        is_mm_embed[:total_num_scheduled_tokens] = False

        req_start_idx = 0
1778
        should_sync_mrope_positions = False
1779

1780
        for req_id in self.input_batch.req_ids:
1781
1782
            mm_embeds_req: list[torch.Tensor] = []

1783
            num_scheduled_tokens = scheduler_output.num_scheduled_tokens[req_id]
1784
            req_state = self.requests[req_id]
1785
            num_computed_tokens = req_state.num_computed_tokens + shift_computed_tokens
1786

1787
1788
            for mm_feature in req_state.mm_features:
                pos_info = mm_feature.mm_position
1789
1790
                start_pos = pos_info.offset
                num_encoder_tokens = pos_info.length
1791
1792
1793
1794
1795
1796
1797
1798
1799
1800
1801
1802
1803
1804
1805
1806

                # 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,
1807
1808
                    num_encoder_tokens,
                )
1809
                assert start_idx < end_idx
1810

1811
                mm_hash = mm_feature.identifier
1812
                encoder_output = self.encoder_cache.get(mm_hash, None)
1813
                assert encoder_output is not None, f"Encoder cache miss for {mm_hash}."
1814
1815
1816
1817

                if (is_embed := pos_info.is_embed) is not None:
                    is_embed = is_embed[start_idx:end_idx]

1818
                req_start_pos = req_start_idx + start_pos - num_computed_tokens
1819
1820
1821
                is_mm_embed[req_start_pos + start_idx : req_start_pos + end_idx] = (
                    True if is_embed is None else is_embed
                )
1822

1823
1824
1825
1826
                mm_embeds_item = gather_mm_placeholders(
                    encoder_output[start_idx:end_idx],
                    is_embed=is_embed,
                )
1827
1828
1829
                mm_embeds_req.append(mm_embeds_item)

            if self.is_multimodal_pruning_enabled and self.uses_mrope:
1830
                assert req_state.mrope_positions is not None
1831
1832
1833
1834
1835
1836
1837
                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,
1838
1839
                    )
                )
1840
1841
1842
1843
                req_state.mrope_positions.copy_(new_mrope_positions)
                req_state.mrope_position_delta = new_delta

            mm_embeds.extend(mm_embeds_req)
1844
1845
1846
            req_start_idx += num_scheduled_tokens

        is_mm_embed = self.is_mm_embed.copy_to_gpu(total_num_scheduled_tokens)
1847
1848
1849

        if should_sync_mrope_positions:
            self._calc_mrope_positions(scheduler_output)
1850
            self.mrope_positions.copy_to_gpu(total_num_scheduled_tokens)
1851

1852
        return mm_embeds, is_mm_embed
1853

1854
1855
1856
1857
1858
1859
1860
1861
1862
1863
1864
1865
1866
1867
1868
1869
    def _extract_encoder_inputs(
        self,
        scheduler_output: "SchedulerOutput",
    ) -> dict[str, torch.Tensor]:
        """Extract encoder inputs for encoder-decoder models.

        This method extracts multimodal input features from scheduled encoder
        inputs and formats them for the encoder-decoder model forward pass.
        """
        # Batch the multi-modal inputs using the helper method.
        mm_kwargs, _ = self._batch_mm_kwargs_from_scheduler(scheduler_output)

        if not mm_kwargs:
            return {}

        # Group MM kwargs by modality and extract features
1870
        model = cast(SupportsMultiModal, self.model)
1871
1872
        encoder_features = {}
        for _, _, mm_kwargs_group in group_mm_kwargs_by_modality(
1873
1874
1875
1876
            mm_kwargs,
            device=self.device,
            pin_memory=self.pin_memory,
            merge_by_field_config=model.merge_by_field_config,
1877
1878
1879
1880
1881
1882
1883
1884
        ):
            # Add the grouped features to encoder_features dict
            # This allows the model to receive them as kwargs (e.g.,
            # input_features=...)
            encoder_features.update(mm_kwargs_group)

        return encoder_features

1885
    def get_model(self) -> nn.Module:
1886
        # get raw model out of the cudagraph wrapper.
1887
        if isinstance(self.model, (CUDAGraphWrapper, UBatchWrapper)):
1888
            return self.model.unwrap()
1889
1890
        return self.model

1891
1892
1893
1894
1895
1896
1897
1898
1899
1900
1901
1902
1903
1904
1905
    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

1906
1907
1908
1909
1910
    def get_supported_pooling_tasks(self) -> list[PoolingTask]:
        model = self.get_model()
        if not is_pooling_model(model):
            return []

1911
1912
        supported_tasks = list(model.pooler.get_supported_tasks())

1913
1914
1915
1916
        if (
            self.scheduler_config.chunked_prefill_enabled
            and "encode" in supported_tasks
        ):
1917
1918
            supported_tasks.remove("encode")

1919
1920
1921
1922
1923
1924
            logger.debug_once(
                "Chunked prefill is not supported with "
                "encode task which using ALL pooling. "
                "Please turn off chunked prefill by "
                "`--no-enable-chunked-prefill` before using it."
            )
1925
1926
1927
1928
1929

        if "score" in supported_tasks:
            num_labels = getattr(self.model_config.hf_config, "num_labels", 0)
            if num_labels != 1:
                supported_tasks.remove("score")
1930
                logger.debug_once("Score API is only enabled for num_labels == 1.")
1931
1932

        return supported_tasks
1933

1934
1935
1936
1937
1938
1939
1940
1941
1942
1943
    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)

1944
    def sync_and_slice_intermediate_tensors(
1945
1946
1947
1948
1949
        self,
        num_tokens: int,
        intermediate_tensors: IntermediateTensors,
        sync_self: bool,
    ) -> IntermediateTensors:
1950
1951
1952
        assert self.intermediate_tensors is not None

        tp = self.vllm_config.parallel_config.tensor_parallel_size
1953
        is_rs = is_residual_scattered_for_sp(self.vllm_config, num_tokens)
1954
1955
1956
1957
1958
1959

        # 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():
1960
                is_scattered = k == "residual" and is_rs
1961
                copy_len = num_tokens // tp if is_scattered else num_tokens
1962
                self.intermediate_tensors[k][:copy_len].copy_(
1963
1964
1965
1966
1967
1968
1969
1970
1971
1972
1973
1974
1975
                    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:
1976
1977
1978
1979
1980
1981
1982
        """
        Step for the EPLB (Expert Parallelism Load Balancing) state.
        """
        if not self.parallel_config.enable_eplb:
            return

        assert self.eplb_state is not None
1983
1984
        model = self.get_model()
        assert is_mixture_of_experts(model)
1985
        self.eplb_state.step(
1986
            model,
1987
1988
            is_dummy,
            is_profile,
1989
            log_stats=self.parallel_config.eplb_config.log_balancedness,
1990
1991
        )

1992
1993
1994
1995
    # This is where the second ubatch is adjusted to account for the padding.
    # Should be called after attention metadata creation. This just pads
    # the second ubatch slice out to the total number of tokens
    # (num_tokens + padding)
1996
1997
1998
1999
2000
2001
2002
    def pad_out_ubatch_slice(self, ubatch_slices: UBatchSlices, num_total_tokens: int):
        padded_second_ubatch_slice = slice(
            ubatch_slices[1].token_slice.start, num_total_tokens
        )
        ubatch_slices[1] = UBatchSlice(
            padded_second_ubatch_slice, padded_second_ubatch_slice
        )
2003

2004
2005
2006
2007
2008
2009
    def _pool(
        self,
        hidden_states: torch.Tensor,
        num_scheduled_tokens: int,
        num_scheduled_tokens_np: np.ndarray,
    ) -> ModelRunnerOutput:
2010
2011
2012
        assert self.input_batch.num_reqs == len(self.input_batch.pooling_params), (
            "Either all or none of the requests in a batch must be pooling request"
        )
2013

2014
        hidden_states = hidden_states[:num_scheduled_tokens]
2015
        pooling_metadata = self.input_batch.get_pooling_metadata()
2016
2017
2018
2019
        pooling_metadata.build_pooling_cursor(
            num_scheduled_tokens_np.tolist(), device=hidden_states.device
        )
        seq_lens_cpu = self.seq_lens.cpu[: self.input_batch.num_reqs]
2020

2021
2022
2023
2024
2025
2026
2027
2028
2029
2030
        model = cast(VllmModelForPooling, self.model)
        raw_pooler_output: PoolerOutput = model.pooler(
            hidden_states=hidden_states,
            pooling_metadata=pooling_metadata,
        )
        raw_pooler_output = json_map_leaves(
            lambda x: x.to("cpu", non_blocking=True),
            raw_pooler_output,
        )
        self._sync_device()
2031
2032
2033

        pooler_output: list[Optional[torch.Tensor]] = []
        for raw_output, seq_len, prompt_len in zip(
2034
2035
            raw_pooler_output, seq_lens_cpu, pooling_metadata.prompt_lens
        ):
2036
            output = raw_output if seq_len == prompt_len else None
2037
            pooler_output.append(output)
2038
2039
2040
2041
2042
2043
2044
2045
2046
2047

        return ModelRunnerOutput(
            req_ids=self.input_batch.req_ids,
            req_id_to_index=self.input_batch.req_id_to_index,
            sampled_token_ids=[],
            logprobs=None,
            prompt_logprobs_dict={},
            pooler_output=pooler_output,
        )

2048
    def _get_num_input_tokens(self, num_scheduled_tokens: int) -> int:
2049
2050
2051
2052
2053
2054
2055
        if (
            self.compilation_config.cudagraph_mode != CUDAGraphMode.NONE
            and not envs.VLLM_DISABLE_PAD_FOR_CUDAGRAPH
            and hasattr(self, "cudagraph_batch_sizes")
            and self.cudagraph_batch_sizes
            and num_scheduled_tokens <= self.cudagraph_batch_sizes[-1]
        ):
2056
2057
2058
2059
2060
2061
2062
2063
            # Use CUDA graphs.
            # Add padding to the batch size.
            return self.vllm_config.pad_for_cudagraph(num_scheduled_tokens)

        # Eager mode.
        # Pad tokens to multiple of tensor_parallel_size when
        # enabled collective fusion for SP
        tp_size = self.vllm_config.parallel_config.tensor_parallel_size
2064
2065
2066
2067
        if (
            self.compilation_config.pass_config.enable_sequence_parallelism
            and tp_size > 1
        ):
2068
2069
2070
            return round_up(num_scheduled_tokens, tp_size)
        return num_scheduled_tokens

2071
    def _preprocess(
2072
2073
        self,
        scheduler_output: "SchedulerOutput",
2074
        num_input_tokens: int,  # Padded
2075
        intermediate_tensors: Optional[IntermediateTensors] = None,
2076
2077
2078
2079
2080
2081
2082
2083
    ) -> tuple[
        int,
        Optional[torch.Tensor],
        Optional[torch.Tensor],
        torch.Tensor,
        Optional[IntermediateTensors],
        dict[str, Any],
    ]:
2084
        num_scheduled_tokens = scheduler_output.total_num_scheduled_tokens
2085

2086
2087
        # _prepare_inputs may reorder the batch, so we must gather multi
        # modal outputs after that to ensure the correct order
2088
2089
2090
2091
2092
        if (
            self.supports_mm_inputs
            and get_pp_group().is_first_rank
            and not self.model_config.is_encoder_decoder
        ):
2093
2094
            # Run the multimodal encoder if any.
            self._execute_mm_encoder(scheduler_output)
2095
            mm_embeds, is_mm_embed = self._gather_mm_embeddings(scheduler_output)
2096

2097
2098
2099
            # 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.
2100
            inputs_embeds_scheduled = self.model.get_input_embeddings(
2101
2102
2103
                self.input_ids.gpu[:num_scheduled_tokens],
                multimodal_embeddings=mm_embeds,
                is_multimodal=is_mm_embed,
2104
            )
2105

2106
            # TODO(woosuk): Avoid the copy. Optimize.
2107
            self.inputs_embeds.gpu[:num_scheduled_tokens].copy_(inputs_embeds_scheduled)
2108

2109
            input_ids = None
2110
            inputs_embeds = self.inputs_embeds.gpu[:num_input_tokens]
2111
2112
2113
2114
            model_kwargs = {
                **self._init_model_kwargs(num_scheduled_tokens),
                **self._extract_mm_kwargs(scheduler_output),
            }
2115
        elif self.enable_prompt_embeds and get_pp_group().is_first_rank:
2116
2117
2118
2119
2120
2121
2122
2123
2124
2125
2126
2127
            # 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).
2128
2129
2130
            token_ids_idx = (
                self.is_token_ids.gpu[:num_scheduled_tokens]
                .nonzero(as_tuple=False)
2131
                .squeeze(1)
2132
            )
2133
2134
2135
            # Some tokens ids may need to become embeds
            if token_ids_idx.numel() > 0:
                token_ids = self.input_ids.gpu[token_ids_idx]
2136
                tokens_to_embeds = self.model.get_input_embeddings(input_ids=token_ids)
2137
2138
2139
2140
2141
                self.inputs_embeds.gpu[token_ids_idx] = tokens_to_embeds

            inputs_embeds = self.inputs_embeds.gpu[:num_input_tokens]
            model_kwargs = self._init_model_kwargs(num_input_tokens)
            input_ids = None
2142
        else:
2143
2144
2145
2146
            # 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.
2147
            input_ids = self.input_ids.gpu[:num_input_tokens]
2148
            inputs_embeds = None
2149
            model_kwargs = self._init_model_kwargs(num_input_tokens)
2150
        if self.uses_mrope:
2151
            positions = self.mrope_positions.gpu[:, :num_input_tokens]
2152
        else:
2153
            positions = self.positions.gpu[:num_input_tokens]
2154

2155
2156
2157
        if get_pp_group().is_first_rank:
            intermediate_tensors = None
        else:
2158
            intermediate_tensors = self.sync_and_slice_intermediate_tensors(
2159
2160
                num_input_tokens, intermediate_tensors, True
            )
2161

2162
2163
2164
2165
        if (
            self.model_config.is_encoder_decoder
            and scheduler_output.scheduled_encoder_inputs
        ):
2166
2167
2168
            encoder_inputs = self._extract_encoder_inputs(scheduler_output)
            model_kwargs.update(encoder_inputs)

2169
2170
2171
2172
2173
2174
2175
2176
        return (
            num_scheduled_tokens,
            input_ids,
            inputs_embeds,
            positions,
            intermediate_tensors,
            model_kwargs,
        )
2177

2178
    def _sample(
2179
2180
2181
        self,
        logits: Optional[torch.Tensor],
        spec_decode_metadata: Optional[SpecDecodeMetadata],
2182
    ) -> SamplerOutput:
2183
        # Sample the next token and get logprobs if needed.
2184
        sampling_metadata = self.input_batch.sampling_metadata
2185
        if spec_decode_metadata is None:
2186
            sampler_output = self.sampler(
2187
2188
2189
2190
                logits=logits,
                sampling_metadata=sampling_metadata,
            )
        else:
2191
2192
2193
2194
            # When indexing with a tensor (bonus_logits_indices), PyTorch
            # creates a new tensor with separate storage from the original
            # logits tensor. This means any in-place operations on bonus_logits
            # won't affect the original logits tensor.
2195
            assert logits is not None
2196
            bonus_logits = logits[spec_decode_metadata.bonus_logits_indices]
2197
            sampler_output = self.sampler(
2198
                logits=bonus_logits,
2199
2200
2201
                sampling_metadata=sampling_metadata,
            )
            bonus_token_ids = sampler_output.sampled_token_ids
2202

2203
2204
2205
            # Just like `bonus_logits`, `target_logits` is a new tensor with
            # separate storage from the original `logits` tensor. Therefore,
            # it is safe to update `target_logits` in place.
2206
            target_logits = logits[spec_decode_metadata.target_logits_indices]
2207
            output_token_ids = self.rejection_sampler(
2208
                spec_decode_metadata,
2209
                None,  # draft_probs
2210
                target_logits,
2211
                bonus_token_ids,
2212
2213
                sampling_metadata,
            )
2214
            sampler_output.sampled_token_ids = output_token_ids
2215
            self._update_states_after_model_execute(output_token_ids)
2216

2217
2218
2219
        return sampler_output

    def _bookkeeping_sync(
2220
2221
2222
2223
2224
2225
        self,
        scheduler_output: "SchedulerOutput",
        sampler_output: SamplerOutput,
        logits: Optional[torch.Tensor],
        hidden_states: torch.Tensor,
        num_scheduled_tokens: int,
2226
    ) -> tuple[
2227
2228
2229
2230
2231
2232
2233
        dict[str, int],
        Optional[LogprobsLists],
        list[list[int]],
        dict[str, Optional[LogprobsTensors]],
        list[str],
        dict[str, int],
        list[int],
2234
    ]:
2235
2236
2237
2238
        num_nans_in_logits = {}
        if envs.VLLM_COMPUTE_NANS_IN_LOGITS:
            num_nans_in_logits = self._get_nans_in_logits(logits)

2239
2240
2241
        discard_sampled_tokens_req_indices = self.discard_request_indices.np[
            : self.num_discarded_requests
        ]
2242
2243
2244
2245
        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)
2246

2247
2248
2249
        # 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()
2250
        req_id_to_index_output_copy = self.input_batch.req_id_to_index.copy()
2251

2252
2253
        # NOTE: GPU -> CPU Sync happens here.
        # Move as many CPU operations as possible before this sync point.
2254
        logprobs_tensors = sampler_output.logprobs_tensors
2255
2256
2257
        logprobs_lists = (
            logprobs_tensors.tolists() if logprobs_tensors is not None else None
        )
2258
2259
2260

        # Compute prompt logprobs if needed.
        prompt_logprobs_dict = self._get_prompt_logprobs_dict(
2261
            hidden_states[:num_scheduled_tokens],
2262
            scheduler_output.num_scheduled_tokens,
2263
2264
        )

2265
        num_sampled_tokens = sampler_output.sampled_token_ids.shape[0]
2266
        sampled_token_ids = sampler_output.sampled_token_ids
2267
        invalid_req_indices = []
2268
2269
2270
2271
2272
2273
2274
2275
2276
2277
2278
2279
2280
2281
        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)
            else:
                # Includes spec decode tokens.
                valid_sampled_token_ids = self.rejection_sampler.parse_output(
                    sampled_token_ids,
                    self.input_batch.vocab_size,
                )
            # Mask out the sampled tokens that should not be sampled.
            for i in discard_sampled_tokens_req_indices:
2282
                valid_sampled_token_ids[int(i)].clear()
2283
        else:
2284
            valid_sampled_token_ids = []
2285
            invalid_req_indices = discard_sampled_tokens_req_indices.tolist()
2286
2287
2288
2289
2290
2291
            invalid_req_indices_set = set(invalid_req_indices)
            assert sampled_token_ids.shape[-1] == 1

            # 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.
2292
2293
            self.input_batch.prev_sampled_token_ids = sampled_token_ids
            self.input_batch.prev_sampled_token_ids_invalid_indices = (
2294
                invalid_req_indices_set
2295
            )
2296
2297
2298
2299
2300
            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
            }
2301

2302
2303
2304
2305
2306
        # 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.
2307
        req_ids = self.input_batch.req_ids
2308
2309
        for req_idx in range(num_sampled_tokens):
            if self.use_async_scheduling:
2310
                sampled_ids = [-1] if req_idx not in invalid_req_indices_set else None
2311
2312
            else:
                sampled_ids = valid_sampled_token_ids[req_idx]
2313
2314
2315
2316
2317
            if not sampled_ids:
                continue

            start_idx = self.input_batch.num_tokens_no_spec[req_idx]
            end_idx = start_idx + len(sampled_ids)
2318
2319
2320
            assert end_idx <= self.max_model_len + 1, (
                "Sampled token IDs exceed the max model length + 1. "
                f"Total number of tokens: {end_idx} > max_model_len + 1: "
2321
2322
                f"{self.max_model_len + 1}"
            )
2323
2324
2325
2326
2327
2328
2329
2330
2331
2332
2333
2334

            n_tokens_cache = len(sampled_ids)

            # Sampled token IDs exceed the max model length by 1. This is
            # legitimate as we can still sample 1 last token when the context
            # length equals the max model length. Note that we do not need to
            # cache this token ID as the sequence finishes after this step.
            # Additionally, the buffers token_ids_cpu and is_token_ids are of
            # size max model length only.
            if end_idx == self.max_model_len + 1:
                n_tokens_cache -= 1

2335
2336
2337
2338
2339
2340
            self.input_batch.token_ids_cpu[
                req_idx, start_idx : (start_idx + n_tokens_cache)
            ] = sampled_ids[:n_tokens_cache]
            self.input_batch.is_token_ids[
                req_idx, start_idx : (start_idx + n_tokens_cache)
            ] = True
2341

2342
2343
            self.input_batch.num_tokens_no_spec[req_idx] = end_idx
            self.input_batch.num_tokens[req_idx] = end_idx
2344

2345
            req_id = req_ids[req_idx]
2346
2347
2348
            req_state = self.requests[req_id]
            req_state.output_token_ids.extend(sampled_ids)

2349
2350
2351
2352
2353
2354
2355
2356
2357
2358
        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,
        )

2359
2360
2361
2362
2363
2364
2365
2366
2367
2368
2369
2370
2371
2372
2373
    @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()

2374
2375
2376
2377
2378
2379
2380
2381
2382
2383
2384
    def _model_forward(
        self,
        input_ids: Optional[torch.Tensor] = None,
        positions: Optional[torch.Tensor] = None,
        intermediate_tensors: Optional[IntermediateTensors] = None,
        inputs_embeds: Optional[torch.Tensor] = None,
        **model_kwargs: dict[str, Any],
    ) -> Any:
        """Helper method to call the model forward pass.

        This method can be overridden by subclasses for model execution.
2385
        Motivation: We can inspect only this method versus
2386
2387
2388
2389
2390
2391
2392
2393
2394
2395
2396
2397
2398
2399
2400
2401
2402
2403
2404
2405
        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,
        )

2406
2407
2408
2409
2410
2411
2412
    @torch.inference_mode()
    def execute_model(
        self,
        scheduler_output: "SchedulerOutput",
        intermediate_tensors: Optional[IntermediateTensors] = None,
    ) -> Union[ModelRunnerOutput, AsyncModelRunnerOutput, IntermediateTensors]:
        with record_function_or_nullcontext("Preprocess"):
2413
2414
2415
2416
2417
2418
2419
2420
2421
            with self.synchronize_input_prep():
                # Update persistent batch states.
                self._update_states(scheduler_output)

                if not scheduler_output.total_num_scheduled_tokens:
                    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(
2422
2423
                        scheduler_output, self.vllm_config
                    )
2424
2425
2426
2427
                if self.cache_config.kv_sharing_fast_prefill:
                    assert not self.input_batch.num_prompt_logprobs, (
                        "--kv-sharing-fast-prefill produces incorrect "
                        "logprobs for prompt tokens, tokens, please disable "
2428
2429
                        "it when the requests need prompt logprobs"
                    )
2430

2431
                # Prepare the decoder inputs.
2432
2433
2434
2435
2436
2437
2438
2439
                (
                    attn_metadata,
                    logits_indices,
                    spec_decode_metadata,
                    num_scheduled_tokens_np,
                    spec_decode_common_attn_metadata,
                    max_query_len,
                    ubatch_slices,
2440
                    num_tokens_across_dp,
2441
2442
                    use_cascade_attn,
                ) = self._prepare_inputs(scheduler_output)
2443

2444
2445
2446
2447
2448
2449
2450
2451
2452
2453
2454
            if ubatch_slices:
                assert num_tokens_across_dp is not None
                num_input_tokens = int(num_tokens_across_dp[0].item())
                self.pad_out_ubatch_slice(ubatch_slices, num_input_tokens)
            elif num_tokens_across_dp is not None:
                num_input_tokens = int(num_tokens_across_dp[0].item())
            else:
                num_input_tokens = self._get_num_input_tokens(
                    scheduler_output.total_num_scheduled_tokens
                )

2455
2456
2457
2458
2459
2460
2461
            (
                num_scheduled_tokens,
                input_ids,
                inputs_embeds,
                positions,
                intermediate_tensors,
                model_kwargs,
2462
            ) = self._preprocess(
2463
                scheduler_output, num_input_tokens, intermediate_tensors
2464
2465
2466
2467
2468
2469
2470
2471
2472
2473
2474
            )

            uniform_decode = (max_query_len == self.uniform_decode_query_len) and (
                num_scheduled_tokens == self.input_batch.num_reqs * max_query_len
            )
            batch_descriptor = BatchDescriptor(
                num_tokens=num_input_tokens, uniform_decode=uniform_decode
            )
            cudagraph_runtime_mode, batch_descriptor = (
                self.cudagraph_dispatcher.dispatch(batch_descriptor, use_cascade_attn)
            )
2475

2476
2477
        # Set cudagraph mode to none if calc_kv_scales is true.
        if attn_metadata is not None:
2478
2479
2480
2481
2482
            metadata_list = (
                attn_metadata.values()
                if isinstance(attn_metadata, dict)
                else [attn_metadata]
            )
2483
            if any(
2484
2485
                getattr(m, "enable_kv_scales_calculation", False) for m in metadata_list
            ):
2486
2487
                cudagraph_runtime_mode = CUDAGraphMode.NONE

2488
2489
        # Run the model.
        # Use persistent buffers for CUDA graphs.
2490
2491
        with (
            set_forward_context(
2492
2493
2494
2495
2496
2497
                attn_metadata,
                self.vllm_config,
                num_tokens=num_input_tokens,
                num_tokens_across_dp=num_tokens_across_dp,
                cudagraph_runtime_mode=cudagraph_runtime_mode,
                batch_descriptor=batch_descriptor,
2498
                ubatch_slices=ubatch_slices,
2499
2500
2501
2502
            ),
            record_function_or_nullcontext("Forward"),
            self.maybe_get_kv_connector_output(scheduler_output) as kv_connector_output,
        ):
2503
            model_output = self._model_forward(
2504
2505
2506
2507
2508
2509
2510
2511
2512
                input_ids=input_ids,
                positions=positions,
                intermediate_tensors=intermediate_tensors,
                inputs_embeds=inputs_embeds,
                **model_kwargs,
            )

        with record_function_or_nullcontext("Postprocess"):
            if self.use_aux_hidden_state_outputs:
2513
                # True when EAGLE 3 is used.
2514
2515
                hidden_states, aux_hidden_states = model_output
            else:
2516
                # Common case.
2517
2518
2519
                hidden_states = model_output
                aux_hidden_states = None

2520
2521
2522
2523
2524
            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)
2525
2526
                    hidden_states.kv_connector_output = kv_connector_output
                    return hidden_states
2527

2528
                if self.is_pooling_model:
2529
                    # Return the pooling output.
2530
2531
2532
                    output = self._pool(
                        hidden_states, num_scheduled_tokens, num_scheduled_tokens_np
                    )
2533
2534
                    output.kv_connector_output = kv_connector_output
                    return output
2535
2536

                sample_hidden_states = hidden_states[logits_indices]
2537
                logits = self.model.compute_logits(sample_hidden_states)
2538
2539
2540
2541
2542
            else:
                # Rare case.
                assert not self.is_pooling_model

                if not get_pp_group().is_last_rank:
2543
                    all_gather_tensors = {
2544
2545
2546
                        "residual": not is_residual_scattered_for_sp(
                            self.vllm_config, num_input_tokens
                        )
2547
                    }
2548
                    get_pp_group().send_tensor_dict(
2549
2550
                        hidden_states.tensors,
                        all_gather_group=get_tp_group(),
2551
2552
                        all_gather_tensors=all_gather_tensors,
                    )
2553
2554
2555
                    logits = None
                else:
                    sample_hidden_states = hidden_states[logits_indices]
2556
                    logits = self.model.compute_logits(sample_hidden_states)
2557
2558
2559
2560
2561

                model_output_broadcast_data = {}
                if logits is not None:
                    model_output_broadcast_data["logits"] = logits.contiguous()

2562
2563
2564
                model_output_broadcast_data = get_pp_group().broadcast_tensor_dict(
                    model_output_broadcast_data, src=len(get_pp_group().ranks) - 1
                )
2565
2566
2567
2568
2569
                assert model_output_broadcast_data is not None
                logits = model_output_broadcast_data["logits"]

            # Apply structured output bitmasks if present
            if scheduler_output.grammar_bitmask is not None:
2570
2571
2572
                apply_grammar_bitmask(
                    scheduler_output, self.input_batch, logits, self.device
                )
2573
2574
2575
2576

        with record_function_or_nullcontext("Sample"):
            sampler_output = self._sample(logits, spec_decode_metadata)

2577
2578
2579
2580
2581
2582
2583
2584
2585
2586
2587
2588
2589
2590
        def propose_draft_token_ids(sampled_token_ids):
            assert spec_decode_common_attn_metadata is not None
            with record_function_or_nullcontext("Draft"):
                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,
                )

2591
2592
2593
2594
2595
        use_padded_batch_for_eagle = (
            self.speculative_config
            and self.speculative_config.use_eagle()
            and not self.speculative_config.disable_padded_drafter_batch
        )
2596
2597
2598
        effective_drafter_max_model_len = self.max_model_len
        if effective_drafter_max_model_len is None:
            effective_drafter_max_model_len = self.model_config.max_model_len
2599
2600
2601
2602
2603
        if (
            self.speculative_config
            and self.speculative_config.draft_model_config is not None
            and self.speculative_config.draft_model_config.max_model_len is not None
        ):
2604
            effective_drafter_max_model_len = (
2605
2606
                self.speculative_config.draft_model_config.max_model_len
            )
2607
        input_fits_in_drafter = spec_decode_common_attn_metadata and (
2608
2609
2610
2611
            spec_decode_common_attn_metadata.max_seq_len
            + self.speculative_config.num_speculative_tokens
            <= effective_drafter_max_model_len
        )
2612
        if use_padded_batch_for_eagle and input_fits_in_drafter:
2613
2614
2615
2616
            # EAGLE speculative decoding can use the GPU sampled tokens
            # as inputs, and does not need to wait for bookkeeping to finish.
            propose_draft_token_ids(sampler_output.sampled_token_ids)

2617
2618
2619
2620
2621
2622
2623
2624
2625
        with record_function_or_nullcontext("Bookkeep"):
            (
                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,
2626
2627
2628
2629
2630
2631
2632
            ) = self._bookkeeping_sync(
                scheduler_output,
                sampler_output,
                logits,
                hidden_states,
                num_scheduled_tokens,
            )
2633

2634
2635
2636
2637
2638
        if (
            self.speculative_config
            and not use_padded_batch_for_eagle
            and input_fits_in_drafter
        ):
2639
2640
2641
            # 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)
2642

2643
2644
        with record_function_or_nullcontext("EPLB"):
            self.eplb_step()
2645

2646
2647
2648
        output = ModelRunnerOutput(
            req_ids=req_ids_output_copy,
            req_id_to_index=req_id_to_index_output_copy,
2649
2650
2651
2652
            sampled_token_ids=valid_sampled_token_ids,
            logprobs=logprobs_lists,
            prompt_logprobs_dict=prompt_logprobs_dict,
            pooler_output=[],
2653
            kv_connector_output=kv_connector_output,
2654
2655
2656
            num_nans_in_logits=num_nans_in_logits,
        )

2657
2658
2659
2660
2661
        if not self.use_async_scheduling:
            return output

        return AsyncGPUModelRunnerOutput(
            model_runner_output=output,
2662
            sampled_token_ids=sampler_output.sampled_token_ids,
2663
2664
2665
2666
            invalid_req_indices=invalid_req_indices,
            async_output_copy_stream=self.async_output_copy_stream,
        )

2667
2668
2669
2670
2671
2672
2673
2674
2675
2676
2677
    def take_draft_token_ids(self) -> Optional[DraftTokenIds]:
        if self._draft_token_ids is None:
            return None
        req_ids = self.input_batch.req_ids
        if isinstance(self._draft_token_ids, torch.Tensor):
            draft_token_ids = self._draft_token_ids.tolist()
        else:
            draft_token_ids = self._draft_token_ids
        self._draft_token_ids = None
        return DraftTokenIds(req_ids, draft_token_ids)

2678
2679
2680
    def propose_draft_token_ids(
        self,
        scheduler_output: "SchedulerOutput",
2681
        sampled_token_ids: Union[torch.Tensor, list[list[int]]],
2682
2683
2684
        sampling_metadata: SamplingMetadata,
        hidden_states: torch.Tensor,
        sample_hidden_states: torch.Tensor,
Wentao Ye's avatar
Wentao Ye committed
2685
        aux_hidden_states: Optional[list[torch.Tensor]],
2686
        spec_decode_metadata: Optional[SpecDecodeMetadata],
2687
        common_attn_metadata: CommonAttentionMetadata,
2688
    ) -> Union[list[list[int]], torch.Tensor]:
2689
2690
        num_scheduled_tokens = scheduler_output.total_num_scheduled_tokens
        if self.speculative_config.method == "ngram":
2691
            assert isinstance(sampled_token_ids, list)
2692
            assert isinstance(self.drafter, NgramProposer)
2693
            draft_token_ids = self.drafter.propose(
2694
2695
                sampled_token_ids,
                self.input_batch.req_ids,
2696
2697
                self.input_batch.num_tokens_no_spec,
                self.input_batch.token_ids_cpu,
2698
2699
                self.input_batch.spec_decode_unsupported_reqs,
            )
2700
        elif self.speculative_config.method == "medusa":
2701
            assert isinstance(sampled_token_ids, list)
2702
            assert isinstance(self.drafter, MedusaProposer)
2703

2704
2705
            if sample_hidden_states.shape[0] == len(sampled_token_ids):
                # The input to the target model does not include draft tokens.
2706
2707
2708
2709
                hidden_states = sample_hidden_states
            else:
                indices = []
                offset = 0
Wentao Ye's avatar
Wentao Ye committed
2710
                assert spec_decode_metadata is not None
2711
                for num_draft, tokens in zip(
2712
2713
                    spec_decode_metadata.num_draft_tokens, sampled_token_ids
                ):
2714
2715
                    indices.append(offset + len(tokens) - 1)
                    offset += num_draft + 1
2716
                indices = torch.tensor(indices, device=self.device)
2717
2718
                hidden_states = sample_hidden_states[indices]

2719
            draft_token_ids = self.drafter.propose(
2720
2721
2722
                target_hidden_states=hidden_states,
                sampling_metadata=sampling_metadata,
            )
2723
        elif self.speculative_config.use_eagle():
2724
            assert isinstance(self.drafter, EagleProposer)
2725
2726
2727
2728
2729

            if self.speculative_config.disable_padded_drafter_batch:
                # 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.
2730
2731
                assert isinstance(sampled_token_ids, list), (
                    "sampled_token_ids should be a python list when"
2732
                    "padded-batch is disabled."
2733
                )
2734
                next_token_ids = self.drafter.prepare_next_token_ids_cpu(
2735
2736
2737
2738
2739
                    sampled_token_ids,
                    self.requests,
                    self.input_batch,
                    scheduler_output.num_scheduled_tokens,
                )
2740
2741
2742
2743
2744
            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.
2745
2746
                assert isinstance(sampled_token_ids, torch.Tensor), (
                    "sampled_token_ids should be a torch.Tensor when"
2747
                    "padded-batch is enabled."
2748
2749
                )
                next_token_ids, valid_sampled_tokens_count = (
2750
2751
2752
2753
2754
2755
                    self.drafter.prepare_next_token_ids_padded(
                        common_attn_metadata,
                        sampled_token_ids,
                        self.requests,
                        self.input_batch,
                        self.discard_request_indices.gpu,
2756
                        self.num_discarded_requests,
2757
                    )
2758
                )
Jiayi Yao's avatar
Jiayi Yao committed
2759

2760
            if spec_decode_metadata is None:
2761
                token_indices_to_sample = None
2762
                # input_ids can be None for multimodal models.
2763
                target_token_ids = self.input_ids.gpu[:num_scheduled_tokens]
2764
                target_positions = self._get_positions(num_scheduled_tokens)
2765
                if self.use_aux_hidden_state_outputs:
Wentao Ye's avatar
Wentao Ye committed
2766
                    assert aux_hidden_states is not None
2767
                    target_hidden_states = torch.cat(
2768
2769
                        [h[:num_scheduled_tokens] for h in aux_hidden_states], dim=-1
                    )
2770
2771
                else:
                    target_hidden_states = hidden_states[:num_scheduled_tokens]
2772
            else:
2773
2774
                if self.speculative_config.disable_padded_drafter_batch:
                    token_indices_to_sample = None
2775
2776
2777
2778
2779
                    common_attn_metadata, token_indices = self.drafter.prepare_inputs(
                        common_attn_metadata,
                        sampled_token_ids,
                        spec_decode_metadata.num_draft_tokens,
                    )
2780
                else:
2781
                    common_attn_metadata, token_indices, token_indices_to_sample = (
2782
2783
2784
                        self.drafter.prepare_inputs_padded(
                            common_attn_metadata,
                            spec_decode_metadata,
2785
2786
2787
                            valid_sampled_tokens_count,
                        )
                    )
2788

2789
                target_token_ids = self.input_ids.gpu[token_indices]
2790
                target_positions = self._get_positions(token_indices)
2791
                if self.use_aux_hidden_state_outputs:
Wentao Ye's avatar
Wentao Ye committed
2792
                    assert aux_hidden_states is not None
2793
                    target_hidden_states = torch.cat(
2794
2795
                        [h[token_indices] for h in aux_hidden_states], dim=-1
                    )
2796
2797
                else:
                    target_hidden_states = hidden_states[token_indices]
2798

2799
            if self.supports_mm_inputs:
2800
2801
2802
2803
2804
2805
                mm_embed_inputs = self._gather_mm_embeddings(
                    scheduler_output,
                    shift_computed_tokens=1,
                )
            else:
                mm_embed_inputs = None
2806

2807
            draft_token_ids = self.drafter.propose(
2808
2809
2810
2811
                target_token_ids=target_token_ids,
                target_positions=target_positions,
                target_hidden_states=target_hidden_states,
                next_token_ids=next_token_ids,
2812
                last_token_indices=token_indices_to_sample,
2813
                sampling_metadata=sampling_metadata,
2814
                common_attn_metadata=common_attn_metadata,
2815
                mm_embed_inputs=mm_embed_inputs,
2816
            )
2817

2818
        return draft_token_ids
2819

2820
2821
2822
    def update_config(self, overrides: dict[str, Any]) -> None:
        allowed_config_names = {"load_config", "model_config"}
        for config_name, config_overrides in overrides.items():
2823
2824
            assert config_name in allowed_config_names, (
                f"Config `{config_name}` not supported. "
2825
                f"Allowed configs: {allowed_config_names}"
2826
            )
2827
2828
2829
2830
            config = getattr(self, config_name)
            new_config = update_config(config, config_overrides)
            setattr(self, config_name, new_config)

2831
2832
2833
2834
2835
    def load_model(self, eep_scale_up: bool = False) -> None:
        """
        Args:
            eep_scale_up: the model loading is for elastic EP scale up.
        """
2836
        logger.info("Starting to load model %s...", self.model_config.model)
2837
2838
        if eep_scale_up:
            from vllm.distributed.parallel_state import get_ep_group
2839
2840
2841
2842
2843

            num_local_physical_experts = torch.empty(1, dtype=torch.int32, device="cpu")
            torch.distributed.broadcast(
                num_local_physical_experts, group=get_ep_group().cpu_group, group_src=0
            )
2844
2845
            num_local_physical_experts = int(num_local_physical_experts.item())
            new_ep_size = get_ep_group().world_size
2846
            global_expert_load, old_global_expert_indices = EplbState.recv_state()
2847
            num_logical_experts = global_expert_load.shape[1]
2848
            self.parallel_config.eplb_config.num_redundant_experts = (
2849
2850
2851
2852
2853
2854
                num_local_physical_experts * new_ep_size - num_logical_experts
            )
            assert old_global_expert_indices.shape[1] % num_local_physical_experts == 0
            old_ep_size = (
                old_global_expert_indices.shape[1] // num_local_physical_experts
            )
2855
            rank_mapping = {
2856
                old_ep_rank: old_ep_rank for old_ep_rank in range(old_ep_size)
2857
2858
2859
2860
2861
2862
            }
        else:
            global_expert_load = None
            old_global_expert_indices = None
            rank_mapping = None

2863
        with DeviceMemoryProfiler() as m:
2864
            time_before_load = time.perf_counter()
2865
            model_loader = get_model_loader(self.load_config)
2866
2867
            logger.info("Loading model from scratch...")
            self.model = model_loader.load_model(
2868
2869
                vllm_config=self.vllm_config, model_config=self.model_config
            )
2870
            if self.lora_config:
2871
2872
2873
                self.model = self.load_lora_model(
                    self.model, self.vllm_config, self.device
                )
2874
2875
2876
            if hasattr(self, "drafter"):
                logger.info("Loading drafter model...")
                self.drafter.load_model(self.model)
2877
            if self.use_aux_hidden_state_outputs:
2878
                if not supports_eagle3(self.model):
2879
2880
                    raise RuntimeError(
                        "Model does not support EAGLE3 interface but "
2881
2882
                        "aux_hidden_state_outputs was requested"
                    )
2883
2884
2885
2886
2887
2888
2889
2890
2891
2892
2893
2894
2895

                # 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)
2896
            time_after_load = time.perf_counter()
2897
        self.model_memory_usage = m.consumed_memory
2898
2899
2900
2901
2902
        logger.info(
            "Model loading took %.4f GiB and %.6f seconds",
            self.model_memory_usage / GiB_bytes,
            time_after_load - time_before_load,
        )
2903
        prepare_communication_buffer_for_model(self.model)
2904

2905
2906
2907
2908
        self.is_multimodal_pruning_enabled = (
            supports_multimodal_pruning(self.model)
            and self.model_config.multimodal_config.is_multimodal_pruning_enabled()
        )
2909

2910
2911
        if is_mixture_of_experts(self.model) and self.parallel_config.enable_eplb:
            logger.info("EPLB is enabled for model %s.", self.model_config.model)
2912
2913
2914
2915
            self.eplb_state = EplbState.build(
                self.model,
                self.device,
                self.parallel_config,
2916
2917
2918
                global_expert_load,
                old_global_expert_indices,
                rank_mapping,
2919
2920
            )

2921
        if (
2922
2923
            self.vllm_config.compilation_config.level == CompilationLevel.DYNAMO_AS_IS
            and supports_dynamo()
2924
        ):
2925
            backend = self.vllm_config.compilation_config.init_backend(self.vllm_config)
2926
            compilation_counter.dynamo_as_is_count += 1
2927
            self.model.compile(fullgraph=True, backend=backend)
2928
2929
2930
2931
2932
            return
        # for other compilation levels, cudagraph behavior is controlled by
        # CudagraphWraper and CudagraphDispatcher of vllm.

        # wrap the model with full cudagraph wrapper if needed.
2933
2934
2935
2936
2937
2938
2939
        if (
            self.compilation_config.cudagraph_mode.has_full_cudagraphs()
            and not self.parallel_config.enable_dbo
        ):
            self.model = CUDAGraphWrapper(
                self.model, self.vllm_config, runtime_mode=CUDAGraphMode.FULL
            )
2940
2941
        elif self.parallel_config.enable_dbo:
            if self.compilation_config.cudagraph_mode.has_full_cudagraphs():
2942
2943
2944
                self.model = UBatchWrapper(
                    self.model, self.vllm_config, CUDAGraphMode.FULL, self.device
                )
2945
            else:
2946
2947
2948
                self.model = UBatchWrapper(
                    self.model, self.vllm_config, CUDAGraphMode.NONE, self.device
                )
2949

2950
2951
2952
2953
2954
2955
2956
2957
2958
2959
2960
2961
2962
2963
2964
2965
2966
2967
2968
2969
2970
2971
2972
2973
    def _get_eagle3_aux_layers_from_config(self) -> Optional[tuple[int, ...]]:
        """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

2974
    def reload_weights(self) -> None:
2975
        assert getattr(self, "model", None) is not None, (
2976
            "Cannot reload weights before model is loaded."
2977
        )
2978
2979
        model_loader = get_model_loader(self.load_config)
        logger.info("Reloading weights inplace...")
2980
        model_loader.load_weights(self.get_model(), model_config=self.model_config)
2981

2982
2983
2984
2985
2986
    def save_tensorized_model(
        self,
        tensorizer_config: "TensorizerConfig",
    ) -> None:
        TensorizerLoader.save_model(
2987
            self.get_model(),
2988
            tensorizer_config=tensorizer_config,
2989
            model_config=self.model_config,
2990
2991
        )

2992
2993
2994
    def _get_prompt_logprobs_dict(
        self,
        hidden_states: torch.Tensor,
2995
        num_scheduled_tokens: dict[str, int],
2996
    ) -> dict[str, Optional[LogprobsTensors]]:
2997
2998
2999
3000
        num_prompt_logprobs_dict = self.input_batch.num_prompt_logprobs
        if not num_prompt_logprobs_dict:
            return {}

3001
        in_progress_dict = self.input_batch.in_progress_prompt_logprobs_cpu
3002
        prompt_logprobs_dict: dict[str, Optional[LogprobsTensors]] = {}
3003
3004
3005
3006
3007

        # 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():
3008
            num_tokens = num_scheduled_tokens[req_id]
3009
3010
3011

            # Get metadata for this request.
            request = self.requests[req_id]
3012
3013
3014
3015
            if request.prompt_token_ids is None:
                # Prompt logprobs is incompatible with prompt embeddings
                continue

3016
3017
            num_prompt_tokens = len(request.prompt_token_ids)
            prompt_token_ids = torch.tensor(request.prompt_token_ids).to(
3018
3019
                self.device, non_blocking=True
            )
3020

3021
3022
3023
3024
3025
3026
            # 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(
3027
3028
                    num_prompt_tokens - 1, num_prompt_logprobs + 1
                )
3029
3030
                in_progress_dict[req_id] = logprobs_tensors

3031
            # Determine number of logits to retrieve.
3032
3033
            start_idx = request.num_computed_tokens
            start_tok = start_idx + 1
3034
            num_remaining_tokens = num_prompt_tokens - start_tok
3035
            if num_tokens <= num_remaining_tokens:
3036
                # This is a chunk, more tokens remain.
3037
3038
3039
                # 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.
3040
3041
3042
3043
3044
                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)
3045
3046
3047
3048
3049
3050
3051
                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
3052
3053
3054
3055
3056

            # 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]
3057
            offset = self.query_start_loc.np[req_idx].item()
3058
            prompt_hidden_states = hidden_states[offset : offset + num_logits]
3059
            logits = self.model.compute_logits(prompt_hidden_states)
3060
3061
3062
3063

            # 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.
3064
            tgt_token_ids = prompt_token_ids[start_tok : start_tok + num_logits]
3065
3066

            # Compute prompt logprobs.
3067
3068
            logprobs = self.sampler.compute_logprobs(logits)
            token_ids, logprobs, ranks = self.sampler.gather_logprobs(
3069
3070
                logprobs, num_prompt_logprobs, tgt_token_ids
            )
3071
3072

            # Transfer GPU->CPU async.
3073
3074
            chunk_slice = slice(start_idx, start_idx + num_logits)
            logprobs_tensors.logprob_token_ids[chunk_slice].copy_(
3075
3076
3077
                token_ids, non_blocking=True
            )
            logprobs_tensors.logprobs[chunk_slice].copy_(logprobs, non_blocking=True)
3078
            logprobs_tensors.selected_token_ranks[chunk_slice].copy_(
3079
3080
                ranks, non_blocking=True
            )
3081
3082
3083
3084
3085

        # 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]
3086
            del in_progress_dict[req_id]
3087
3088

        # Must synchronize the non-blocking GPU->CPU transfers.
3089
        if prompt_logprobs_dict:
3090
            self._sync_device()
3091
3092
3093

        return prompt_logprobs_dict

3094
3095
3096
3097
3098
3099
3100
3101
3102
3103
3104
3105
3106
3107
    def _get_nans_in_logits(
        self,
        logits: Optional[torch.Tensor],
    ) -> 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])
3108
3109
3110
                    if num_nans_for_index is not None and req_index < logits.shape[0]
                    else 0
                )
3111
3112
3113
3114
            return num_nans_in_logits
        except IndexError:
            return {}

3115
3116
3117
3118
3119
3120
    @contextmanager
    def maybe_randomize_inputs(self, input_ids: torch.Tensor):
        """
        Randomize input_ids if VLLM_RANDOMIZE_DP_DUMMY_INPUTS is set.
        This is to help balance expert-selection
         - during profile_run
3121
         - during DP rank dummy run
3122
3123
3124
3125
3126
3127
3128
3129
3130
3131
3132
        """
        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
        else:
            import functools

            @functools.cache
            def rand_input_ids() -> torch.Tensor:
                return torch.randint_like(
3133
                    self.input_ids.gpu,
3134
3135
                    low=0,
                    high=self.model_config.get_vocab_size(),
3136
3137
                    dtype=input_ids.dtype,
                )
3138

3139
            logger.debug_once("Randomizing dummy data for DP Rank")
3140
            input_ids.copy_(rand_input_ids()[: input_ids.size(0)], non_blocking=True)
3141
3142
3143
            yield
            input_ids.fill_(0)

3144
3145
3146
3147
3148
3149
    def _get_mm_dummy_batch(
        self,
        modality: str,
        max_items_per_batch: int,
    ) -> BatchedTensorInputs:
        """Dummy data for profiling and precompiling multimodal models."""
3150
3151
        assert self.mm_budget is not None

3152
3153
        dummy_decoder_data = self.mm_registry.get_decoder_dummy_data(
            model_config=self.model_config,
3154
            seq_len=self.max_model_len,
3155
            mm_counts={modality: 1},
3156
            cache=self.mm_budget.cache,
3157
3158
3159
3160
        )
        dummy_mm_data = dummy_decoder_data.multi_modal_data

        # Result in the maximum GPU consumption of the model
3161
3162
        dummy_mm_item = dummy_mm_data[modality][0]
        dummy_mm_items = [dummy_mm_item] * max_items_per_batch
3163

3164
        model = cast(SupportsMultiModal, self.model)
3165
3166
3167
3168
3169
3170
3171
3172
3173
        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,
                merge_by_field_config=model.merge_by_field_config,
            )
        )
3174

3175
3176
3177
3178
    @torch.inference_mode()
    def _dummy_run(
        self,
        num_tokens: int,
3179
        cudagraph_runtime_mode: Optional[CUDAGraphMode] = None,
3180
3181
        force_attention: bool = False,
        uniform_decode: bool = False,
3182
        allow_microbatching: bool = True,
3183
3184
        skip_eplb: bool = False,
        is_profile: bool = False,
3185
        create_mixed_batch: bool = False,
3186
        remove_lora: bool = True,
3187
    ) -> tuple[torch.Tensor, torch.Tensor]:
3188
3189
3190
3191
3192
3193
3194
        """
        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.
3195
                - if not set will determine the cudagraph mode based on using
3196
                    the self.cudagraph_dispatcher.
3197
3198
3199
3200
                - CUDAGraphMode.NONE: No cudagraph, for warm up and profile run
                - CUDAGraphMode.PIECEWISE: Piecewise cudagraph.
                - CUDAGraphMode.FULL: Full cudagraph, attention metadata is
                    needed.
3201
            force_attention: If True, always create attention metadata. Used to
3202
3203
3204
3205
                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.
3206
3207
            create_mixed_batch: If True, create a mixed batch with both decode
                (1 token) and prefill (multiple tokens) requests.
3208
            remove_lora: If False, dummy LoRAs are not destroyed after the run
3209
        """
3210
3211
3212
3213
        assert (
            cudagraph_runtime_mode is None
            or cudagraph_runtime_mode.valid_runtime_modes()
        )
3214

3215
        # If cudagraph_mode.decode_mode() == FULL and
3216
        # cudagraph_mode.separate_routine(). This means that we are using
3217
3218
3219
3220
3221
3222
3223
3224
3225
3226
3227
        # 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.
3228
        max_query_len = self.uniform_decode_query_len if uniform_decode else num_tokens
3229

3230
3231
3232
3233
3234
        # 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
3235
3236
3237
3238
        if create_mixed_batch:
            assert not uniform_decode
            # Create mixed batch:
            # first half decode tokens, second half one prefill
3239
            num_decode_tokens = min(max_num_reqs - 1, num_tokens // 2)
3240
3241
3242
3243
            num_prefill_tokens = num_tokens - num_decode_tokens
            num_reqs = num_decode_tokens + 1

            # Create decode requests (1 token each) followed by prefill request
3244
            num_scheduled_tokens_list = [1] * num_decode_tokens + [num_prefill_tokens]
3245
3246
3247
            # Note: Overriding max_query_len to be the prefill tokens
            max_query_len = num_prefill_tokens
        elif uniform_decode:
3248
            assert not create_mixed_batch
3249
            num_reqs = min(max_num_reqs, cdiv(num_tokens, max_query_len))
3250
3251
            num_scheduled_tokens_list = [max_query_len] * num_reqs
            if num_tokens % max_query_len != 0:
3252
                num_scheduled_tokens_list[-1] = num_tokens % max_query_len
3253
3254
3255
3256
3257
3258
        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

3259
3260
        assert sum(num_scheduled_tokens_list) == num_tokens
        assert len(num_scheduled_tokens_list) == num_reqs
3261
        num_scheduled_tokens = np.array(num_scheduled_tokens_list, dtype=np.int32)
3262
        total_num_scheduled_tokens = int(num_scheduled_tokens.sum())
3263

3264
3265
        # We currently only microbatch if the number of tokens is
        # over a certain threshold.
3266
3267
3268
3269
3270
3271
3272
3273
3274
3275
3276
        ubatch_slices, num_tokens_across_dp = coordinate_batch_across_dp(
            num_scheduled_tokens,
            total_num_scheduled_tokens,
            total_num_scheduled_tokens,
            self.vllm_config.parallel_config,
            allow_microbatching,
            uniform_decode,
        )
        num_tokens_after_padding = num_tokens
        if num_tokens_across_dp is not None:
            num_tokens_after_padding = int(num_tokens_across_dp[0])
3277
3278

        attn_metadata: Optional[PerLayerAttnMetadata] = None
3279
3280
3281

        # If force_attention is True, we always capture attention. Otherwise,
        # it only happens for cudagraph_runtime_mode=FULL.
3282
        if force_attention or cudagraph_runtime_mode == CUDAGraphMode.FULL:
3283
            attn_metadata = {}
3284
3285
            if ubatch_slices is not None:
                attn_metadata = [dict() for _ in range(len(ubatch_slices))]
3286

3287
3288
3289
3290
3291
3292
            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:
3293
                seq_lens = max_query_len
3294
            self.seq_lens.np[:num_reqs] = seq_lens
3295
3296
            self.seq_lens.np[num_reqs:] = 0
            self.seq_lens.copy_to_gpu()
3297

3298
3299
            cum_num_tokens, _ = self._get_cumsum_and_arange(num_scheduled_tokens)
            self.query_start_loc.np[1 : num_reqs + 1] = cum_num_tokens
3300
3301
            self.query_start_loc.copy_to_gpu()

3302
            for kv_cache_group_id, kv_cache_group_spec in enumerate(
3303
3304
                self.kv_cache_config.kv_cache_groups
            ):
3305
                common_attn_metadata = CommonAttentionMetadata(
3306
3307
                    query_start_loc=self.query_start_loc.gpu[: num_reqs + 1],
                    query_start_loc_cpu=self.query_start_loc.cpu[: num_reqs + 1],
3308
3309
                    seq_lens=self.seq_lens.gpu[:num_reqs],
                    seq_lens_cpu=self.seq_lens.cpu[:num_reqs],
3310
3311
3312
                    num_computed_tokens_cpu=self.input_batch.num_computed_tokens_cpu_tensor[
                        :num_reqs
                    ],
3313
3314
                    num_reqs=num_reqs,
                    num_actual_tokens=num_tokens,
3315
                    max_query_len=max_query_len,
3316
                    max_seq_len=self.max_model_len,
3317
3318
3319
                    block_table_tensor=self.input_batch.block_table[
                        kv_cache_group_id
                    ].get_device_tensor(num_reqs),
3320
                    slot_mapping=self.input_batch.block_table[
3321
3322
3323
3324
                        kv_cache_group_id
                    ].slot_mapping.gpu[:num_tokens],
                    causal=True,
                )
3325
                for attn_group in self.attn_groups[kv_cache_group_id]:
3326
3327
                    if ubatch_slices is not None:
                        common_attn_metadata_list = split_attn_metadata(
3328
3329
                            ubatch_slices, common_attn_metadata
                        )
3330
                        for ubid, common_attn_metadata in enumerate(
3331
3332
                            common_attn_metadata_list
                        ):
3333
                            assert common_attn_metadata.max_query_len == 1
3334
3335
3336
                            attn_metadata_i = attn_group.get_metadata_builder(
                                ubatch_id=ubid
                            ).build_for_cudagraph_capture(common_attn_metadata)
3337
                            for layer_name in attn_group.layer_names:
3338
                                assert type(attn_metadata) is list
3339
                                attn_metadata[ubid][layer_name] = attn_metadata_i
3340
3341
                    else:
                        assert type(attn_metadata) is dict
3342
3343
                        metadata_builder = attn_group.get_metadata_builder()
                        attn_metadata_i = metadata_builder.build_for_cudagraph_capture(
3344
3345
                            common_attn_metadata
                        )
3346
                        for layer_name in attn_group.layer_names:
3347
                            attn_metadata[layer_name] = attn_metadata_i
3348

3349
3350
3351
        with self.maybe_dummy_run_with_lora(
            self.lora_config, num_scheduled_tokens, remove_lora
        ):
3352
3353
3354
            # Make sure padding doesn't exceed max_num_tokens
            assert num_tokens_after_padding <= self.max_num_tokens
            model_kwargs = self._init_model_kwargs(num_tokens_after_padding)
3355
            if self.supports_mm_inputs and not self.model_config.is_encoder_decoder:
3356
                input_ids = None
3357
                inputs_embeds = self.inputs_embeds.gpu[:num_tokens_after_padding]
3358
                model_kwargs = {
3359
                    **model_kwargs,
3360
3361
                    **self._dummy_mm_kwargs(num_reqs),
                }
3362
3363
            elif self.enable_prompt_embeds:
                input_ids = None
3364
3365
                inputs_embeds = self.inputs_embeds.gpu[:num_tokens_after_padding]
                model_kwargs = self._init_model_kwargs(num_tokens_after_padding)
3366
            else:
3367
                input_ids = self.input_ids.gpu[:num_tokens_after_padding]
3368
                inputs_embeds = None
3369

3370
            if self.uses_mrope:
3371
                positions = self.mrope_positions.gpu[:, :num_tokens_after_padding]
3372
            else:
3373
                positions = self.positions.gpu[:num_tokens_after_padding]
3374
3375
3376
3377
3378
3379
3380
3381
3382

            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,
3383
3384
3385
                            device=self.device,
                        )
                    )
3386
3387

                intermediate_tensors = self.sync_and_slice_intermediate_tensors(
3388
                    num_tokens_after_padding, None, False
3389
                )
3390
3391

            # filter out the valid batch descriptor
3392
3393
3394
3395
3396
3397
3398
3399
3400
3401
            _cg_mode, batch_descriptor = (
                self.cudagraph_dispatcher.dispatch(
                    BatchDescriptor(
                        num_tokens=num_tokens_after_padding,
                        uniform_decode=uniform_decode,
                    )
                )
                if not is_profile
                else (CUDAGraphMode.NONE, None)
            )
3402
3403
3404
            if cudagraph_runtime_mode is not None:
                # we allow forcing NONE when the dispatcher disagrees to support
                # warm ups for cudagraph capture
3405
3406
3407
3408
                assert (
                    cudagraph_runtime_mode == CUDAGraphMode.NONE
                    or cudagraph_runtime_mode == _cg_mode
                ), (
3409
                    f"Cudagraph runtime mode mismatch at dummy_run. "
3410
3411
                    f"Expected {_cg_mode}, but got {cudagraph_runtime_mode}."
                )
3412
3413
            else:
                cudagraph_runtime_mode = _cg_mode
3414

3415
            if ubatch_slices is not None:
3416
3417
3418
3419
3420
3421
3422
                # Adjust values to reflect a single ubatch.
                # TODO(sage,lucas): this is cruft that should be addressed in
                #  the padding refactor.
                num_tokens_after_padding = ubatch_slices[0].num_tokens
                if num_tokens_across_dp is not None:
                    num_tokens_across_dp[:] = num_tokens_after_padding

3423
3424
3425
            with (
                self.maybe_randomize_inputs(input_ids),
                set_forward_context(
3426
3427
                    attn_metadata,
                    self.vllm_config,
3428
                    num_tokens=num_tokens_after_padding,
3429
3430
                    num_tokens_across_dp=num_tokens_across_dp,
                    cudagraph_runtime_mode=cudagraph_runtime_mode,
3431
                    batch_descriptor=batch_descriptor,
3432
3433
3434
                    ubatch_slices=ubatch_slices,
                ),
            ):
3435
                outputs = self.model(
3436
3437
3438
3439
                    input_ids=input_ids,
                    positions=positions,
                    intermediate_tensors=intermediate_tensors,
                    inputs_embeds=inputs_embeds,
3440
                    **model_kwargs,
3441
                )
3442

3443
3444
3445
3446
            if self.use_aux_hidden_state_outputs:
                hidden_states, _ = outputs
            else:
                hidden_states = outputs
3447

3448
            if self.speculative_config and self.speculative_config.use_eagle():
3449
3450
3451
                assert isinstance(self.drafter, EagleProposer)
                self.drafter.dummy_run(num_tokens)

3452
3453
3454
3455
3456
3457
3458
3459
3460
3461
        # 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)

3462
        logit_indices = np.cumsum(num_scheduled_tokens) - 1
3463
        return hidden_states, hidden_states[logit_indices]
3464
3465
3466
3467
3468
3469

    @torch.inference_mode()
    def _dummy_sampler_run(
        self,
        hidden_states: torch.Tensor,
    ) -> torch.Tensor:
3470
3471
3472
3473
        # The dummy hidden states may contain special values,
        # like `inf` or `nan`.
        # To avoid breaking the sampler, we use a random tensor here instead.
        hidden_states = torch.rand_like(hidden_states)
3474

3475
        logits = self.model.compute_logits(hidden_states)
3476
3477
        num_reqs = logits.size(0)

3478
        dummy_tensors = lambda v: torch.full((num_reqs,), v, device=self.device)
3479
3480
3481
3482
3483
3484
3485
3486
3487
3488
3489
3490
3491
3492
3493
3494
3495

        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)],
            allowed_token_ids_mask=None,
            bad_words_token_ids={},
3496
            logitsprocs=LogitsProcessors(),
3497
        )
3498
        try:
3499
3500
3501
            sampler_output = self.sampler(
                logits=logits, sampling_metadata=dummy_metadata
            )
3502
        except RuntimeError as e:
3503
            if "out of memory" in str(e):
3504
3505
3506
3507
                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 "
3508
3509
                    "initializing the engine."
                ) from e
3510
3511
            else:
                raise e
3512
        if self.speculative_config:
3513
3514
            draft_token_ids = [[0] for _ in range(num_reqs)]
            dummy_spec_decode_metadata = SpecDecodeMetadata.make_dummy(
3515
3516
                draft_token_ids, self.device
            )
3517
3518
3519
3520
3521
3522

            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
3523
3524
3525
            target_logits = torch.randn(
                num_tokens, logits.shape[-1], device=self.device, dtype=logits.dtype
            )
3526
3527
3528
            # NOTE(woosuk): Here, we should use int32 because the sampler uses
            # int32 for bonus_token_ids. If the dtype mismatches, re-compilation
            # will occur at runtime.
3529
3530
3531
            bonus_token_ids = torch.zeros(
                num_reqs, device=self.device, dtype=torch.int32
            )
3532
3533
3534
3535
3536
3537
3538
            self.rejection_sampler(
                dummy_spec_decode_metadata,
                draft_probs,
                target_logits,
                bonus_token_ids,
                dummy_metadata,
            )
3539
        return sampler_output
3540

3541
    def _dummy_pooler_run_task(
3542
3543
        self,
        hidden_states: torch.Tensor,
3544
3545
        task: PoolingTask,
    ) -> PoolerOutput:
3546
3547
3548
3549
3550
3551
3552
3553
3554
3555
3556
        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
        num_scheduled_tokens_list = [min_tokens_per_req] * num_reqs
        num_scheduled_tokens_list[-1] += num_tokens % num_reqs
        assert sum(num_scheduled_tokens_list) == num_tokens
        assert len(num_scheduled_tokens_list) == num_reqs

        req_num_tokens = num_tokens // num_reqs

3557
        dummy_prompt_lens = torch.tensor(
3558
3559
            num_scheduled_tokens_list,
            device="cpu",
3560
        )
3561
3562
3563
        dummy_token_ids = torch.zeros(
            (num_reqs, req_num_tokens), dtype=torch.int32, device=self.device
        )
3564

3565
        model = cast(VllmModelForPooling, self.get_model())
3566
        dummy_pooling_params = PoolingParams(task=task)
3567
        dummy_pooling_params.verify(task=task, model_config=self.model_config)
3568
        to_update = model.pooler.get_pooling_updates(task)
3569
3570
        to_update.apply(dummy_pooling_params)

3571
        dummy_metadata = PoolingMetadata(
3572
3573
3574
3575
            prompt_lens=dummy_prompt_lens,
            prompt_token_ids=dummy_token_ids,
            pooling_params=[dummy_pooling_params] * num_reqs,
        )
3576

3577
3578
3579
        dummy_metadata.build_pooling_cursor(
            num_scheduled_tokens_list, device=hidden_states.device
        )
3580

3581
        try:
3582
3583
3584
            return model.pooler(
                hidden_states=hidden_states, pooling_metadata=dummy_metadata
            )
3585
        except RuntimeError as e:
3586
            if "out of memory" in str(e):
3587
                raise RuntimeError(
3588
3589
3590
                    "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 "
3591
3592
                    "initializing the engine."
                ) from e
3593
3594
            else:
                raise e
3595
3596
3597
3598
3599
3600
3601
3602
3603
3604
3605

    @torch.inference_mode()
    def _dummy_pooler_run(
        self,
        hidden_states: torch.Tensor,
    ) -> PoolerOutput:
        # Find the task that has the largest output for subsequent steps
        output_size = dict[PoolingTask, float]()
        for task in self.get_supported_pooling_tasks():
            # Run a full batch with each task to ensure none of them OOMs
            output = self._dummy_pooler_run_task(hidden_states, task)
3606
            output_size[task] = sum(o.nbytes for o in output)
3607
3608
3609
3610
            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)
3611

3612
    def profile_run(self) -> None:
3613
        # Profile with multimodal encoder & encoder cache.
3614
        if self.supports_mm_inputs:
3615
            if self.model_config.multimodal_config.skip_mm_profiling:
3616
                logger.info(
3617
                    "Skipping memory profiling for multimodal encoder and "
3618
3619
                    "encoder cache."
                )
3620
3621
3622
3623
3624
3625
3626
3627
            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.
3628
                    dummy_modality = mm_budget.get_modality_with_max_tokens()
3629
3630
3631
                    max_mm_items_per_batch = mm_budget.max_items_per_batch_by_modality[
                        dummy_modality
                    ]
3632
3633
3634
3635
3636
3637
3638
3639
3640

                    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,
                    )
3641

3642
3643
3644
3645
3646
                    # Create dummy batch of multimodal inputs.
                    batched_dummy_mm_inputs = self._get_mm_dummy_batch(
                        dummy_modality,
                        max_mm_items_per_batch,
                    )
3647

3648
                    # Run multimodal encoder.
3649
3650
3651
                    dummy_encoder_outputs = self.model.get_multimodal_embeddings(
                        **batched_dummy_mm_inputs
                    )
3652

3653
3654
3655
3656
                    sanity_check_mm_encoder_outputs(
                        dummy_encoder_outputs,
                        expected_num_items=max_mm_items_per_batch,
                    )
3657

3658
3659
3660
3661
3662
3663
3664
3665
3666
3667
                    # NOTE: This happens when encoder cache needs to store
                    # the embeddings that encoder outputs are scattered onto.
                    # In this case we create dummy embeddings of size
                    # (encode_budget, hidden_size) and scatter encoder
                    # output into it.
                    encoder_output_shape = dummy_encoder_outputs[0].shape
                    if encoder_output_shape[0] < encoder_budget:
                        expanded_outputs = []
                        for output in dummy_encoder_outputs:
                            expanded = output.new_zeros(
3668
3669
                                (encoder_budget, encoder_output_shape[-1])
                            )
3670
3671
3672
3673
3674
3675
                            num_tokens = output.shape[0]
                            expanded[:num_tokens].copy_(output)
                            expanded_outputs.append(expanded)

                        dummy_encoder_outputs = expanded_outputs

3676
                    # Cache the dummy encoder outputs.
3677
                    self.encoder_cache["tmp"] = dict(enumerate(dummy_encoder_outputs))
3678

3679
        # Add `is_profile` here to pre-allocate communication buffers
3680
3681
3682
        hidden_states, last_hidden_states = self._dummy_run(
            self.max_num_tokens, is_profile=True
        )
3683
        if get_pp_group().is_last_rank:
3684
3685
3686
3687
            if self.is_pooling_model:
                output = self._dummy_pooler_run(hidden_states)
            else:
                output = self._dummy_sampler_run(last_hidden_states)
3688
        else:
3689
            output = None
3690
        self._sync_device()
3691
        del hidden_states, output
3692
        self.encoder_cache.clear()
3693
        gc.collect()
3694

3695
    def capture_model(self) -> int:
3696
        if self.compilation_config.cudagraph_mode == CUDAGraphMode.NONE:
3697
            logger.warning(
3698
                "Skipping CUDA graph capture. To turn on CUDA graph capture, "
3699
3700
                "ensure `cudagraph_mode` was not manually set to `NONE`"
            )
3701
            return 0
3702
3703
        else:
            self.initialize_cudagraph_capture()
3704

3705
3706
        compilation_counter.num_gpu_runner_capture_triggers += 1

3707
3708
        start_time = time.perf_counter()

3709
3710
3711
3712
3713
3714
3715
3716
3717
3718
3719
3720
3721
3722
        @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()
3723
                    gc.collect()
3724

3725
3726
3727
        # 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.
3728
        set_cudagraph_capturing_enabled(True)
3729
        with freeze_gc(), graph_capture(device=self.device):
3730
            start_free_gpu_memory = torch.cuda.mem_get_info()[0]
3731
            cudagraph_mode = self.compilation_config.cudagraph_mode
3732
            assert cudagraph_mode is not None
3733
3734
3735
3736
3737
3738
3739
            if cudagraph_mode.mixed_mode() != CUDAGraphMode.NONE:
                cudagraph_runtime_mode = cudagraph_mode.mixed_mode()

                compilation_cases = list(reversed(self.cudagraph_batch_sizes))
                self._capture_cudagraphs(
                    compilation_cases,
                    cudagraph_runtime_mode=cudagraph_runtime_mode,
3740
3741
                    uniform_decode=False,
                )
3742

3743
3744
            # Capture full cudagraph for uniform decode batches if we
            # don't already have full mixed prefill-decode cudagraphs.
3745
3746
3747
3748
3749
3750
3751
            if (
                cudagraph_mode.decode_mode() == CUDAGraphMode.FULL
                and cudagraph_mode.separate_routine()
            ):
                max_num_tokens = (
                    self.scheduler_config.max_num_seqs * self.uniform_decode_query_len
                )
3752
                decode_cudagraph_batch_sizes = [
3753
3754
3755
                    x
                    for x in self.cudagraph_batch_sizes
                    if x <= max_num_tokens and x >= self.uniform_decode_query_len
3756
                ]
3757
                compilation_cases_decode = list(reversed(decode_cudagraph_batch_sizes))
3758
3759
3760
                self._capture_cudagraphs(
                    compilation_cases=compilation_cases_decode,
                    cudagraph_runtime_mode=CUDAGraphMode.FULL,
3761
3762
                    uniform_decode=True,
                )
3763

3764
3765
3766
            torch.cuda.synchronize()
            end_free_gpu_memory = torch.cuda.mem_get_info()[0]

3767
3768
3769
        # 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
3770
        # we may do lazy capturing in future that still allows capturing
3771
3772
        # after here.
        set_cudagraph_capturing_enabled(False)
3773
3774
3775
3776
3777

        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.
3778
3779
3780
3781
3782
        logger.info(
            "Graph capturing finished in %.0f secs, took %.2f GiB",
            elapsed_time,
            cuda_graph_size / (1 << 30),
        )
3783
        return cuda_graph_size
3784

3785
3786
3787
3788
3789
3790
3791
3792
3793
3794
    def _capture_cudagraphs(
        self,
        compilation_cases: list[int],
        cudagraph_runtime_mode: CUDAGraphMode,
        uniform_decode: bool,
    ):
        assert (
            cudagraph_runtime_mode != CUDAGraphMode.NONE
            and cudagraph_runtime_mode.valid_runtime_modes()
        ), f"Invalid cudagraph runtime mode: {cudagraph_runtime_mode}"
3795
3796
3797
3798
3799
3800
3801
3802

        # Only rank 0 should print progress bar during capture
        if is_global_first_rank():
            compilation_cases = tqdm(
                compilation_cases,
                disable=not self.load_config.use_tqdm_on_load,
                desc="Capturing CUDA graphs ({}, {})".format(
                    "decode" if uniform_decode else "mixed prefill-decode",
3803
3804
3805
                    cudagraph_runtime_mode.name,
                ),
            )
3806

3807
3808
3809
        # We skip EPLB here since we don't want to record dummy metrics
        for num_tokens in compilation_cases:
            # We currently only capture ubatched graphs when its a FULL
3810
3811
3812
            # 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
3813
3814
3815
3816
            allow_microbatching = (
                self.parallel_config.enable_dbo
                and cudagraph_runtime_mode == CUDAGraphMode.FULL
                and uniform_decode
3817
3818
3819
3820
3821
                and check_ubatch_thresholds(
                    config=self.vllm_config.parallel_config,
                    num_tokens=num_tokens,
                    uniform_decode=uniform_decode,
                )
3822
            )
3823

3824
3825
3826
3827
3828
3829
            for _ in range(self.compilation_config.cudagraph_num_of_warmups):
                # Use CUDAGraphRuntimeStyle.NONE (default) for warmup.
                # But be careful, warm up with `NONE`is orthogonal to
                # 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.
3830
3831
3832
3833
3834
3835
3836
3837
3838
3839
3840
3841
3842
3843
3844
3845
3846
3847
                force_attention = cudagraph_runtime_mode == CUDAGraphMode.FULL
                self._dummy_run(
                    num_tokens,
                    cudagraph_runtime_mode=CUDAGraphMode.NONE,
                    force_attention=force_attention,
                    uniform_decode=uniform_decode,
                    allow_microbatching=allow_microbatching,
                    skip_eplb=True,
                    remove_lora=False,
                )
            self._dummy_run(
                num_tokens,
                cudagraph_runtime_mode=cudagraph_runtime_mode,
                uniform_decode=uniform_decode,
                allow_microbatching=allow_microbatching,
                skip_eplb=True,
                remove_lora=False,
            )
3848
        self.maybe_remove_all_loras(self.lora_config)
3849

3850
3851
3852
3853
    def initialize_attn_backend(self, kv_cache_config: KVCacheConfig) -> None:
        """
        Initialize the attention backends and attention metadata builders.
        """
3854
        assert len(self.attn_groups) == 0, "Attention backends are already initialized"
3855

3856
3857
3858
3859
3860
3861
3862
3863
        class AttentionGroupKey(NamedTuple):
            attn_backend: type[AttentionBackend]
            kv_cache_spec: KVCacheSpec

        def get_attn_backends_for_group(
            kv_cache_group_spec: KVCacheGroupSpec,
        ) -> dict[AttentionGroupKey, list[str]]:
            layers = get_layers_from_vllm_config(
3864
3865
                self.vllm_config, AttentionLayerBase, kv_cache_group_spec.layer_names
            )
3866
3867
            attn_backends = {}
            attn_backend_layers = defaultdict(list)
3868
            # Dedupe based on full class name; this is a bit safer than
3869
3870
3871
3872
            # 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.
3873
            for layer_name in kv_cache_group_spec.layer_names:
3874
                attn_backend = layers[layer_name].get_attn_backend()
3875
3876
3877
3878
3879
3880
3881

                if layer_name in self.kv_sharing_fast_prefill_eligible_layers:
                    attn_backend = create_fast_prefill_custom_backend(
                        "FastPrefill",
                        attn_backend,
                    )

3882
3883
3884
                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):
3885
                    layer_kv_cache_spec = layer_kv_cache_spec.kv_cache_specs[layer_name]
3886
                key = (full_cls_name, layer_kv_cache_spec)
3887
3888
3889
                attn_backends[key] = AttentionGroupKey(
                    attn_backend, layer_kv_cache_spec
                )
3890
                attn_backend_layers[key].append(layer_name)
3891
            return {attn_backends[k]: v for k, v in attn_backend_layers.items()}
3892
3893

        def create_attn_groups(
3894
            attn_backends_map: dict[AttentionGroupKey, list[str]],
3895
3896
        ) -> list[AttentionGroup]:
            attn_groups: list[AttentionGroup] = []
3897
            for (attn_backend, kv_cache_spec), layer_names in attn_backends_map.items():
3898
3899
                attn_group = AttentionGroup.create_with_metadata_builders(
                    attn_backend,
3900
                    layer_names,
3901
                    kv_cache_spec,
3902
3903
                    self.vllm_config,
                    self.device,
3904
                    num_metadata_builders=1
3905
3906
                    if not self.parallel_config.enable_dbo
                    else 2,
3907
3908
                )

3909
3910
3911
3912
                attn_groups.append(attn_group)
            return attn_groups

        for kv_cache_group_spec in kv_cache_config.kv_cache_groups:
3913
3914
            attn_backends = get_attn_backends_for_group(kv_cache_group_spec)
            self.attn_groups.append(create_attn_groups(attn_backends))
3915

co63oc's avatar
co63oc committed
3916
        # Calculate reorder batch threshold (if needed)
3917
3918
        self.calculate_reorder_batch_threshold()

3919
    def initialize_cudagraph_capture(self) -> None:
3920
        """
3921
        Resolve the cudagraph_mode when there are multiple attention
3922
3923
3924
3925
        backends with potential conflicting CUDA graph support.
        Then initialize the cudagraph_dispatcher based on the resolved
        cudagraph_mode.
        """
3926
3927
3928
3929
        min_cg_support = AttentionCGSupport.ALWAYS
        min_cg_builder_name = None

        for attn_group in self._attn_group_iterator():
3930
            builder = attn_group.get_metadata_builder()
3931
3932
3933
3934
3935
3936
            if builder.cudagraph_support.value < min_cg_support.value:
                min_cg_support = builder.cudagraph_support
                min_cg_builder_name = builder.__class__.__name__
        # Flexible resolve the cudagraph mode
        cudagraph_mode = self.compilation_config.cudagraph_mode
        # check cudagraph for mixed batch is supported
3937
3938
3939
3940
3941
3942
3943
3944
3945
        if (
            cudagraph_mode.mixed_mode() == CUDAGraphMode.FULL
            and min_cg_support != AttentionCGSupport.ALWAYS
        ):
            msg = (
                f"CUDAGraphMode.{cudagraph_mode.name} is not supported "
                f"with {min_cg_builder_name} backend (support: "
                f"{min_cg_support})"
            )
3946
3947
            if min_cg_support == AttentionCGSupport.NEVER:
                # if not supported any full cudagraphs, just raise it.
3948
3949
                msg += (
                    "; please try cudagraph_mode=PIECEWISE, and "
3950
                    "make sure compilation level is piecewise"
3951
                )
3952
3953
3954
3955
3956
                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"
3957
                cudagraph_mode = self.compilation_config.cudagraph_mode = (
3958
                    CUDAGraphMode.FULL_AND_PIECEWISE
3959
                )
3960
3961
            else:
                msg += "; setting cudagraph_mode=FULL_DECODE_ONLY"
3962
                cudagraph_mode = self.compilation_config.cudagraph_mode = (
3963
                    CUDAGraphMode.FULL_DECODE_ONLY
3964
                )
3965
3966
            logger.warning(msg)

3967
        # check that if we are doing decode full-cudagraphs it is supported
3968
3969
3970
3971
3972
3973
3974
3975
3976
3977
3978
3979
3980
3981
3982
        if (
            cudagraph_mode.decode_mode() == CUDAGraphMode.FULL
            and min_cg_support == AttentionCGSupport.NEVER
        ):
            msg = (
                f"CUDAGraphMode.{cudagraph_mode.name} is not supported "
                f"with {min_cg_builder_name} backend (support: "
                f"{min_cg_support})"
            )
            if self.compilation_config.level == CompilationLevel.PIECEWISE and (
                self.compilation_config.splitting_ops_contain_attention()
                or self.compilation_config.use_inductor_graph_partition
            ):
                msg += (
                    "; setting cudagraph_mode=PIECEWISE because "
3983
                    "attention is compiled piecewise"
3984
3985
                )
                cudagraph_mode = self.compilation_config.cudagraph_mode = (
3986
                    CUDAGraphMode.PIECEWISE
3987
                )
3988
            else:
3989
3990
                msg += (
                    "; setting cudagraph_mode=NONE because "
3991
                    "attention is not compiled piecewise"
3992
3993
                )
                cudagraph_mode = self.compilation_config.cudagraph_mode = (
3994
                    CUDAGraphMode.NONE
3995
                )
3996
3997
            logger.warning(msg)

3998
3999
        # check that if we are doing spec-decode + decode full-cudagraphs it is
        # supported
4000
4001
4002
4003
4004
4005
4006
4007
4008
4009
        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 "
                f"{min_cg_builder_name} (support: {min_cg_support})"
            )
4010
4011
            if self.compilation_config.splitting_ops_contain_attention():
                msg += "; setting cudagraph_mode=PIECEWISE"
4012
                cudagraph_mode = self.compilation_config.cudagraph_mode = (
4013
                    CUDAGraphMode.PIECEWISE
4014
                )
4015
4016
            else:
                msg += "; setting cudagraph_mode=NONE"
4017
                cudagraph_mode = self.compilation_config.cudagraph_mode = (
4018
                    CUDAGraphMode.NONE
4019
                )
4020
4021
4022
4023
            logger.warning(msg)

        # double check that we can support full cudagraph if they are requested
        # even after automatic downgrades
4024
4025
4026
4027
4028
4029
4030
4031
4032
4033
4034
        if (
            cudagraph_mode.has_full_cudagraphs()
            and min_cg_support == AttentionCGSupport.NEVER
        ):
            raise ValueError(
                f"CUDAGraphMode.{cudagraph_mode.name} is not "
                f"supported with {min_cg_builder_name} backend ("
                f"support:{min_cg_support}) "
                "; please try cudagraph_mode=PIECEWISE, "
                "and make sure compilation level is piecewise"
            )
4035
4036
4037
4038

        # Trigger cudagraph dispatching keys initialization here (after
        # initializing attn backends).
        self.cudagraph_dispatcher.initialize_cudagraph_keys(
4039
4040
            self.compilation_config.cudagraph_mode, self.uniform_decode_query_len
        )
4041

4042
4043
4044
4045
4046
    def calculate_reorder_batch_threshold(self) -> None:
        """
        Check that if any backends reorder batches; that the reordering
        is compatible (e.g., decode threshold is the same)
        """
4047
        for group in self._attn_group_iterator():
4048
            attn_metadata_builder_i = group.get_metadata_builder()
4049

4050
4051
            # check that if any backends reorder batches; that the reordering
            # is compatible (e.g., decode threshold is the same)
4052
            reorder_batch_threshold_i = attn_metadata_builder_i.reorder_batch_threshold
4053
4054
            if reorder_batch_threshold_i is not None:
                if self.reorder_batch_threshold is not None:
4055
                    if reorder_batch_threshold_i != self.reorder_batch_threshold:
4056
4057
4058
4059
                        raise ValueError(
                            f"Attention backend reorders decodes with "
                            f"threshold {reorder_batch_threshold_i} but other "
                            f"backend uses threshold "
4060
4061
                            f"{self.reorder_batch_threshold}"
                        )
4062
4063
4064
                else:
                    self.reorder_batch_threshold = reorder_batch_threshold_i

4065
    def may_reinitialize_input_batch(self, kv_cache_config: KVCacheConfig) -> None:
4066
4067
4068
4069
4070
4071
4072
4073
4074
4075
4076
4077
4078
4079
4080
4081
        """
        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.
        """
        block_sizes = [
            kv_cache_group.kv_cache_spec.block_size
            for kv_cache_group in kv_cache_config.kv_cache_groups
        ]
        if block_sizes != [self.cache_config.block_size]:
            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
4082
4083
                "for more details."
            )
4084
4085
            self.input_batch = InputBatch(
                max_num_reqs=self.max_num_reqs,
4086
                max_model_len=max(self.max_model_len, self.max_encoder_len),
4087
4088
4089
4090
4091
                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,
4092
                is_spec_decode=bool(self.vllm_config.speculative_config),
4093
4094
                logitsprocs=self.input_batch.logitsprocs,
                is_pooling_model=self.is_pooling_model,
4095
4096
                num_speculative_tokens=(
                    self.vllm_config.speculative_config.num_speculative_tokens
4097
4098
4099
                    if self.vllm_config.speculative_config
                    else 0
                ),
4100
4101
            )

4102
    def _allocate_kv_cache_tensors(
4103
4104
        self, kv_cache_config: KVCacheConfig
    ) -> dict[str, torch.Tensor]:
4105
        """
4106
4107
4108
        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.

4109
        Args:
4110
            kv_cache_config: The KV cache config
4111
        Returns:
4112
            dict[str, torch.Tensor]: A map between layer names to their
4113
            corresponding memory buffer for KV cache.
4114
        """
4115
4116
        kv_cache_raw_tensors: dict[str, torch.Tensor] = {}
        for kv_cache_tensor in kv_cache_config.kv_cache_tensors:
4117
4118
4119
            tensor = torch.zeros(
                kv_cache_tensor.size, dtype=torch.int8, device=self.device
            )
4120
4121
4122
4123
4124
            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:
4125
4126
4127
4128
            for layer_name in group.layer_names:
                if layer_name in self.runner_only_attn_layers:
                    continue
                layer_names.add(layer_name)
4129
4130
4131
        assert layer_names == set(kv_cache_raw_tensors.keys()), (
            "Some layers are not correctly initialized"
        )
4132
4133
        return kv_cache_raw_tensors

4134
4135
4136
    def _attn_group_iterator(self) -> Iterator[AttentionGroup]:
        return itertools.chain.from_iterable(self.attn_groups)

4137
    def _kv_cache_spec_attn_group_iterator(self) -> Iterator[AttentionGroup]:
4138
4139
        if not self.kv_cache_config.kv_cache_groups:
            return
4140
4141
        for attn_groups in self.attn_groups:
            yield from attn_groups
4142

4143
4144
4145
4146
4147
    def _reshape_kv_cache_tensors(
        self,
        kv_cache_config: KVCacheConfig,
        kv_cache_raw_tensors: dict[str, torch.Tensor],
    ) -> dict[str, torch.Tensor]:
4148
        """
4149
        Reshape the KV cache tensors to the desired shape and dtype.
4150

4151
        Args:
4152
4153
            kv_cache_config: The KV cache config
            kv_cache_raw_tensors: The KV cache buffer of each layer, with
4154
                correct size but uninitialized shape.
4155
        Returns:
4156
            Dict[str, torch.Tensor]: A map between layer names to their
4157
4158
            corresponding memory buffer for KV cache.
        """
4159
        kv_caches: dict[str, torch.Tensor] = {}
4160
        has_attn, has_mamba = False, False
4161
4162
        for group in self._kv_cache_spec_attn_group_iterator():
            kv_cache_spec = group.kv_cache_spec
4163
4164
            attn_backend = group.backend
            for layer_name in group.layer_names:
4165
4166
                if layer_name in self.runner_only_attn_layers:
                    continue
4167
4168
                raw_tensor = kv_cache_raw_tensors[layer_name]
                assert raw_tensor.numel() % kv_cache_spec.page_size_bytes == 0
4169
                num_blocks = raw_tensor.numel() // kv_cache_spec.page_size_bytes
4170
                if isinstance(kv_cache_spec, AttentionSpec):
4171
                    has_attn = True
4172
                    kv_cache_shape = attn_backend.get_kv_cache_shape(
4173
4174
4175
4176
                        num_blocks,
                        kv_cache_spec.block_size,
                        kv_cache_spec.num_kv_heads,
                        kv_cache_spec.head_size,
4177
4178
                        cache_dtype_str=self.cache_config.cache_dtype,
                    )
4179
                    dtype = kv_cache_spec.dtype
4180
                    try:
4181
4182
                        kv_cache_stride_order = attn_backend.get_kv_cache_stride_order()
                        assert len(kv_cache_stride_order) == len(kv_cache_shape)
4183
                    except (AttributeError, NotImplementedError):
4184
                        kv_cache_stride_order = tuple(range(len(kv_cache_shape)))
4185
4186
4187
4188
4189
                    # 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.
4190
4191
4192
                    kv_cache_shape = tuple(
                        kv_cache_shape[i] for i in kv_cache_stride_order
                    )
4193
4194
4195
4196
4197
                    # Maintain original KV shape view.
                    inv_order = [
                        kv_cache_stride_order.index(i)
                        for i in range(len(kv_cache_stride_order))
                    ]
4198
4199
4200
4201
4202
4203
                    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
4204
                elif isinstance(kv_cache_spec, MambaSpec):
4205
                    has_mamba = True
Chen Zhang's avatar
Chen Zhang committed
4206
4207
                    raw_tensor = kv_cache_raw_tensors[layer_name]
                    state_tensors = []
4208
                    storage_offset_bytes = 0
4209
                    for shape, dtype in zip(kv_cache_spec.shapes, kv_cache_spec.dtypes):
4210
4211
                        dtype_size = get_dtype_size(dtype)
                        num_element_per_page = (
4212
4213
                            kv_cache_spec.page_size_bytes // dtype_size
                        )
Chen Zhang's avatar
Chen Zhang committed
4214
                        target_shape = (num_blocks, *shape)
4215
4216
                        stride = torch.empty(target_shape).stride()
                        target_stride = (num_element_per_page, *stride[1:])
4217
                        assert storage_offset_bytes % dtype_size == 0
4218
4219
4220
4221
                        tensor = torch.as_strided(
                            raw_tensor.view(dtype),
                            size=target_shape,
                            stride=target_stride,
4222
                            storage_offset=storage_offset_bytes // dtype_size,
4223
                        )
Chen Zhang's avatar
Chen Zhang committed
4224
                        state_tensors.append(tensor)
4225
                        storage_offset_bytes += stride[0] * dtype_size
4226
4227

                    kv_caches[layer_name] = state_tensors
4228
                else:
4229
                    raise NotImplementedError
4230
4231

        if has_attn and has_mamba:
4232
            self._update_hybrid_attention_mamba_layout(kv_caches)
4233

4234
4235
        return kv_caches

4236
    def _update_hybrid_attention_mamba_layout(
4237
4238
        self, kv_caches: dict[str, torch.Tensor]
    ) -> None:
4239
        """
4240
4241
        Update the layout of attention layers from (2, num_blocks, ...) to
        (num_blocks, 2, ...).
4242
4243

        Args:
4244
            kv_caches: The KV cache buffer of each layer.
4245
4246
        """

4247
4248
        for group in self._kv_cache_spec_attn_group_iterator():
            kv_cache_spec = group.kv_cache_spec
4249
            for layer_name in group.layer_names:
4250
                kv_cache = kv_caches[layer_name]
4251
4252
4253
4254
                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 "
4255
                        f"a tensor of shape {kv_cache.shape}"
4256
                    )
4257
                    hidden_size = kv_cache.shape[2:].numel()
4258
4259
4260
4261
                    kv_cache.as_strided_(
                        size=kv_cache.shape,
                        stride=(hidden_size, 2 * hidden_size, *kv_cache.stride()[2:]),
                    )
4262

4263
    def initialize_kv_cache_tensors(
4264
4265
        self, kv_cache_config: KVCacheConfig
    ) -> dict[str, torch.Tensor]:
4266
4267
4268
4269
4270
4271
        """
        Initialize the memory buffer for KV cache.

        Args:
            kv_cache_config: The KV cache config
        Returns:
4272
            Dict[str, torch.Tensor]: A map between layer names to their
4273
4274
4275
4276
4277
            corresponding memory buffer for KV cache.
        """
        # 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
4278
4279
4280
        kv_caches = self._reshape_kv_cache_tensors(
            kv_cache_config, kv_cache_raw_tensors
        )
4281

4282
        # Set up cross-layer KV cache sharing
4283
4284
        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)
4285
4286
            kv_caches[layer_name] = kv_caches[target_layer_name]

4287
4288
4289
4290
4291
4292
4293
4294
4295
        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,
        )
4296
4297
4298
        return kv_caches

    def maybe_add_kv_sharing_layers_to_kv_cache_groups(
4299
4300
        self, kv_cache_config: KVCacheConfig
    ) -> None:
4301
4302
4303
4304
4305
4306
4307
4308
4309
4310
4311
4312
4313
4314
4315
4316
4317
4318
        """
        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.
4319
            attn_layers = get_layers_from_vllm_config(self.vllm_config, Attention)
4320
4321
            for layer_name in reversed(attn_layers):
                if layer_name in self.shared_kv_cache_layers:
4322
                    self.kv_sharing_fast_prefill_eligible_layers.add(layer_name)
4323
4324
                else:
                    break
4325

4326
4327
4328
4329
4330
4331
4332
    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
        """
4333
        kv_cache_config = deepcopy(kv_cache_config)
4334
4335
        self.kv_cache_config = kv_cache_config
        self.may_reinitialize_input_batch(kv_cache_config)
4336
        self.may_add_encoder_only_layers_to_kv_cache_config()
4337
        self.maybe_add_kv_sharing_layers_to_kv_cache_groups(kv_cache_config)
4338
4339
4340
        self.initialize_attn_backend(kv_cache_config)
        kv_caches = self.initialize_kv_cache_tensors(kv_cache_config)

4341
4342
4343
4344
4345
4346
        if self.speculative_config and self.speculative_config.use_eagle():
            assert isinstance(self.drafter, EagleProposer)
            # 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
4347
        if has_kv_transfer_group():
4348
4349
4350
            kv_transfer_group = get_kv_transfer_group()
            kv_transfer_group.register_kv_caches(kv_caches)
            kv_transfer_group.set_host_xfer_buffer_ops(copy_kv_blocks)
Robert Shaw's avatar
Robert Shaw committed
4351

4352
        if self.dcp_world_size > 1:
4353
            layer_names = self.attn_groups[0][0].layer_names
4354
4355
4356
            layers = get_layers_from_vllm_config(
                self.vllm_config, AttentionLayerBase, layer_names
            )
4357
4358
4359
4360
4361
            for layer in layers.values():
                assert layer.impl.need_to_return_lse_for_decode, (
                    "DCP requires attention impls to return"
                    " the softmax lse for decode, but the impl "
                    f"{layer.impl.__class__.__name__} "
4362
4363
                    "does not return the softmax lse for decode."
                )
4364

4365
4366
4367
4368
4369
    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
4370
        encoder_only_attn_specs: dict[AttentionSpec, list[str]] = defaultdict(list)
4371
4372
4373
        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:
4374
                attn_spec: AttentionSpec = EncoderOnlyAttentionSpec(
4375
4376
4377
                    block_size=block_size,
                    num_kv_heads=attn_module.num_kv_heads,
                    head_size=attn_module.head_size,
4378
4379
                    dtype=self.kv_cache_dtype,
                )
4380
4381
4382
                encoder_only_attn_specs[attn_spec].append(layer_name)
                self.runner_only_attn_layers.add(layer_name)
        if len(encoder_only_attn_specs) > 0:
4383
4384
4385
            assert len(encoder_only_attn_specs) == 1, (
                "Only support one encoder-only attention spec now"
            )
4386
4387
            spec, layer_names = encoder_only_attn_specs.popitem()
            self.kv_cache_config.kv_cache_groups.append(
4388
4389
                KVCacheGroupSpec(layer_names=layer_names, kv_cache_spec=spec)
            )
4390

4391
    def get_kv_cache_spec(self) -> dict[str, KVCacheSpec]:
4392
        """
4393
        Generates the KVCacheSpec by parsing the kv cache format from each
4394
4395
        Attention module in the static forward context.
        Returns:
4396
            KVCacheSpec: A dictionary mapping layer names to their KV cache
4397
4398
4399
4400
            format. Layers that do not need KV cache are not included.
        """

        block_size = self.vllm_config.cache_config.block_size
4401
        use_mla = self.vllm_config.model_config.use_mla
4402
        cache_dtype_str = self.vllm_config.cache_config.cache_dtype
4403
        kv_cache_spec: dict[str, KVCacheSpec] = {}
Chen Zhang's avatar
Chen Zhang committed
4404
4405
        attn_layers = get_layers_from_vllm_config(self.vllm_config, Attention)
        for layer_name, attn_module in attn_layers.items():
4406
            if (kv_tgt_layer := attn_module.kv_sharing_target_layer_name) is not None:
4407
4408
4409
4410
4411
4412
4413
4414
4415
4416
                # 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

4417
4418
            # TODO(lucas): move the attention specs into the model layers like
            # the attention backends
4419
            if attn_module.attn_type == AttentionType.DECODER:
4420
                if attn_module.sliding_window is not None:
4421
                    assert not use_mla, "MLA is not supported for slidingwindow"
4422
4423
4424
4425
4426
                    kv_cache_spec[layer_name] = SlidingWindowSpec(
                        block_size=block_size,
                        num_kv_heads=attn_module.num_kv_heads,
                        head_size=attn_module.head_size,
                        dtype=self.kv_cache_dtype,
4427
4428
                        sliding_window=attn_module.sliding_window,
                    )
4429
4430
4431
4432
4433
4434
                elif use_mla:
                    kv_cache_spec[layer_name] = MLAAttentionSpec(
                        block_size=block_size,
                        num_kv_heads=attn_module.num_kv_heads,
                        head_size=attn_module.head_size,
                        dtype=self.kv_cache_dtype,
4435
4436
4437
4438
4439
                        cache_dtype_str=cache_dtype_str,
                    )
                elif self.attention_chunk_size is not None and isinstance(
                    attn_module, ChunkedLocalAttention
                ):
4440
                    kv_cache_spec[layer_name] = ChunkedLocalAttentionSpec(
4441
4442
4443
4444
                        block_size=block_size,
                        num_kv_heads=attn_module.num_kv_heads,
                        head_size=attn_module.head_size,
                        dtype=self.kv_cache_dtype,
4445
4446
                        attention_chunk_size=self.attention_chunk_size,
                    )
4447
4448
4449
4450
4451
                else:
                    kv_cache_spec[layer_name] = FullAttentionSpec(
                        block_size=block_size,
                        num_kv_heads=attn_module.num_kv_heads,
                        head_size=attn_module.head_size,
4452
4453
                        dtype=self.kv_cache_dtype,
                    )
4454
4455
4456
4457
4458
            elif attn_module.attn_type == AttentionType.ENCODER_DECODER:
                kv_cache_spec[layer_name] = CrossAttentionSpec(
                    block_size=block_size,
                    num_kv_heads=attn_module.num_kv_heads,
                    head_size=attn_module.head_size,
4459
4460
4461
4462
4463
4464
                    dtype=self.kv_cache_dtype,
                )
            elif attn_module.attn_type in (
                AttentionType.ENCODER,
                AttentionType.ENCODER_ONLY,
            ):
4465
4466
4467
                # encoder-only attention does not need KV cache.
                continue
            else:
4468
                raise ValueError(f"Unknown attention type: {attn_module.attn_type}")
4469

4470
        mamba_layers = get_layers_from_vllm_config(self.vllm_config, MambaBase)
Chen Zhang's avatar
Chen Zhang committed
4471
        if len(mamba_layers) > 0:
4472
4473
4474
4475
4476
            if (
                self.vllm_config.speculative_config is not None
                and self.vllm_config.model_config.hf_config.model_type
                not in ["qwen3_next"]
            ):
Chen Zhang's avatar
Chen Zhang committed
4477
                raise NotImplementedError(
4478
4479
                    "Mamba with speculative decoding is not supported yet."
                )
4480
            mamba_block_size = self.vllm_config.cache_config.mamba_block_size
4481
            page_size_padded = self.vllm_config.cache_config.mamba_page_size_padded
4482

Chen Zhang's avatar
Chen Zhang committed
4483
4484
4485
            for layer_name, mamba_module in mamba_layers.items():
                kv_cache_spec[layer_name] = MambaSpec(
                    shapes=mamba_module.get_state_shape(),
4486
                    dtypes=mamba_module.get_state_dtype(),
4487
                    block_size=mamba_block_size,
4488
                    page_size_padded=page_size_padded,
4489
4490
4491
                    mamba_type=mamba_module.mamba_type,
                    num_speculative_blocks=(
                        self.speculative_config.num_speculative_tokens
4492
4493
4494
                        if self.speculative_config
                        else 0
                    ),
4495
                )
4496
        ds_indexer_layers = get_layers_from_vllm_config(
4497
4498
            self.vllm_config, DeepseekV32IndexerCache
        )
4499
4500
        for layer_name, ds_indexer_module in ds_indexer_layers.items():
            kv_cache_spec[layer_name] = ds_indexer_module.get_kv_cache_spec()
4501

4502
        return kv_cache_spec
4503
4504
4505
4506
4507
4508
4509
4510
4511
4512

    def _to_list(self, sampled_token_ids: torch.Tensor) -> list[list[int]]:
        # 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.
4513
        pinned = self.sampled_token_ids_pinned_cpu[: sampled_token_ids.shape[0]]
4514
4515
4516
4517
        pinned.copy_(sampled_token_ids, non_blocking=True)
        self.transfer_event.record()
        self.transfer_event.synchronize()
        return pinned.tolist()