gpu_model_runner.py 199 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, MultipleOf
23
from vllm.attention.layer import MLAAttention
24
from vllm.attention.layers.chunked_local_attention import ChunkedLocalAttention
25
from vllm.compilation.counter import compilation_counter
26
27
from vllm.compilation.cuda_graph import CUDAGraphWrapper
from vllm.compilation.monitor import set_cudagraph_capturing_enabled
28
29
30
31
32
33
34
from vllm.config import (
    CompilationLevel,
    CUDAGraphMode,
    VllmConfig,
    get_layers_from_vllm_config,
    update_config,
)
35
from vllm.distributed.eplb.eplb_state import EplbState
36
from vllm.distributed.kv_transfer import get_kv_transfer_group, has_kv_transfer_group
37
from vllm.distributed.kv_transfer.kv_connector.utils import copy_kv_blocks
38
from vllm.distributed.parallel_state import (
39
40
41
42
43
44
    get_pp_group,
    get_tp_group,
    graph_capture,
    is_global_first_rank,
    prepare_communication_buffer_for_model,
)
45
from vllm.forward_context import BatchDescriptor, set_forward_context
46
from vllm.logger import init_logger
47
48
from vllm.model_executor.layers.attention_layer_base import AttentionLayerBase
from vllm.model_executor.layers.mamba.abstract import MambaBase
49
from vllm.model_executor.layers.rotary_embedding import MRotaryEmbedding
50
from vllm.model_executor.model_loader import TensorizerLoader, get_model_loader
51
from vllm.model_executor.models.deepseek_v2 import DeepseekV32IndexerCache
52
53
54
55
56
57
58
59
from vllm.model_executor.models.interfaces import (
    SupportsMultiModal,
    is_mixture_of_experts,
    supports_eagle3,
    supports_mrope,
    supports_multimodal_pruning,
    supports_transcription,
)
60
from vllm.model_executor.models.interfaces_base import (
61
62
63
64
    VllmModelForPooling,
    is_pooling_model,
    is_text_generation_model,
)
65
from vllm.multimodal import MULTIMODAL_REGISTRY
66
67
68
69
70
from vllm.multimodal.inputs import (
    BatchedTensorInputs,
    MultiModalKwargsItem,
    PlaceholderRange,
)
71
from vllm.multimodal.utils import group_mm_kwargs_by_modality
72
from vllm.pooling_params import PoolingParams
73
from vllm.sampling_params import SamplingType
74
from vllm.sequence import IntermediateTensors
75
from vllm.tasks import GenerationTask, PoolingTask, SupportedTask
76
77
78
79
80
81
82
83
84
85
86
87
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,
)
88
from vllm.utils.jsontree import json_map_leaves
89
from vllm.v1.attention.backends.flash_attn import AttentionMetadata
90
from vllm.v1.attention.backends.gdn_attn import GDNAttentionMetadataBuilder
91
from vllm.v1.attention.backends.utils import (
92
93
94
    AttentionCGSupport,
    AttentionMetadataBuilder,
    CommonAttentionMetadata,
95
    create_fast_prefill_custom_backend,
96
97
98
    reorder_batch_to_split_decodes_and_prefills,
    split_attn_metadata,
)
99
from vllm.v1.cudagraph_dispatcher import CudagraphDispatcher
100
101
102
103
104
105
106
107
108
109
110
111
112
113
114
115
116
117
118
119
120
121
122
123
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,
)
124
from vllm.v1.pool.metadata import PoolingMetadata
125
from vllm.v1.sample.logits_processor import LogitsProcessors, build_logitsprocs
126
from vllm.v1.sample.metadata import SamplingMetadata
127
from vllm.v1.sample.rejection_sampler import RejectionSampler
128
from vllm.v1.sample.sampler import Sampler
129
from vllm.v1.spec_decode.eagle import EagleProposer
130
from vllm.v1.spec_decode.medusa import MedusaProposer
131
from vllm.v1.spec_decode.metadata import SpecDecodeMetadata
132
from vllm.v1.spec_decode.ngram_proposer import NgramProposer
133
from vllm.v1.structured_output.utils import apply_grammar_bitmask
134
from vllm.v1.utils import CpuGpuBuffer, record_function_or_nullcontext
135
from vllm.v1.worker.dp_utils import coordinate_batch_across_dp
136
from vllm.v1.worker.gpu_input_batch import CachedRequestState, InputBatch
137
from vllm.v1.worker.gpu_ubatch_wrapper import UBatchWrapper
138
from vllm.v1.worker.kv_connector_model_runner_mixin import KVConnectorModelRunnerMixin
139
from vllm.v1.worker.lora_model_runner_mixin import LoRAModelRunnerMixin
140
141
142
143
144
from vllm.v1.worker.ubatch_utils import (
    UBatchSlice,
    UBatchSlices,
    check_ubatch_thresholds,
)
145
from vllm.v1.worker.utils import is_residual_scattered_for_sp
146

147
148
149
150
151
152
153
154
155
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,
)
156

157
if TYPE_CHECKING:
158
    from vllm.model_executor.model_loader.tensorizer import TensorizerConfig
159
    from vllm.v1.core.sched.output import SchedulerOutput
160
161
162

logger = init_logger(__name__)

163
164
AttnMetadataDict: TypeAlias = dict[str, AttentionMetadata]
# list when ubatching is enabled
165
PerLayerAttnMetadata: TypeAlias = Union[list[AttnMetadataDict], AttnMetadataDict]
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
191
# 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(
192
193
                "cpu", non_blocking=True
            )
194
195
196
197
            self._async_copy_ready_event.record()

    def get_output(self) -> ModelRunnerOutput:
        """Copy the device tensors to the host and return a ModelRunnerOutput.
198

199
200
201
202
203
204
205
206
207
208
209
210
211
212
213
214
        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


215
class GPUModelRunner(LoRAModelRunnerMixin, KVConnectorModelRunnerMixin):
216
217
    def __init__(
        self,
218
        vllm_config: VllmConfig,
219
        device: torch.device,
220
    ):
221
222
223
        self.vllm_config = vllm_config
        self.model_config = vllm_config.model_config
        self.cache_config = vllm_config.cache_config
224
        self.compilation_config = vllm_config.compilation_config
225
226
227
228
229
230
        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
231

232
        from vllm.model_executor.models.utils import set_cpu_offload_max_bytes
233
234
235
236

        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

237
        init_batch_invariance()
238

239
240
241
242
        model_config = self.model_config
        cache_config = self.cache_config
        scheduler_config = self.scheduler_config
        parallel_config = self.parallel_config
243
        self.device = device
244
245
246
247
248
        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:
249
            self.kv_cache_dtype = STR_DTYPE_TO_TORCH_DTYPE[cache_config.cache_dtype]
250

251
        self.is_pooling_model = model_config.runner_type == "pooling"
252
        self.enable_prompt_embeds = model_config.enable_prompt_embeds
253
        self.is_multimodal_raw_input_only_model = (
254
255
            model_config.is_multimodal_raw_input_only_model
        )
256
257
        # This will be overridden in load_model()
        self.is_multimodal_pruning_enabled = False
258
        self.max_model_len = model_config.max_model_len
259
        self.dcp_world_size = self.parallel_config.decode_context_parallel_size
260
        self.max_num_tokens = scheduler_config.max_num_batched_tokens
261
        self.max_num_reqs = scheduler_config.max_num_seqs
262

263
264
265
266
267
        # 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 = (
268
269
270
            self.parallel_config.distributed_executor_backend == "external_launcher"
            and len(get_pp_group().ranks) > 0
        )
271

272
        # Model-related.
273
        self.num_query_heads = model_config.get_num_attention_heads(parallel_config)
274
        self.hidden_size = model_config.get_hidden_size()
275
        self.attention_chunk_size = model_config.attention_chunk_size
276
277
        # Only relevant for models using ALiBi (e.g, MPT)
        self.use_alibi = check_use_alibi(model_config)
278

279
        self.cascade_attn_enabled = not self.model_config.disable_cascade_attn
280

281
        # Multi-modal data support
282
        self.mm_registry = MULTIMODAL_REGISTRY
283
        self.uses_mrope = model_config.uses_mrope
284
        self.supports_mm_inputs = self.mm_registry.supports_multimodal_inputs(
285
286
            model_config
        )
287

288
289
290
        if self.model_config.is_encoder_decoder:
            # Maximum length of the encoder input, only for encoder-decoder
            # models.
291
            self.max_encoder_len = scheduler_config.max_num_encoder_input_tokens
292
293
294
        else:
            self.max_encoder_len = 0

295
        # Sampler
296
        self.sampler = Sampler(logprobs_mode=self.model_config.logprobs_mode)
297

298
299
300
301
302
303
304
        self.eplb_state: Optional[EplbState] = None
        """
        State of the expert parallelism load balancer.

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

305
        # Lazy initializations
306
        # self.model: nn.Module  # Set after load_model
307
        # Initialize in initialize_kv_cache
308
        self.kv_caches: list[torch.Tensor] = []
309
310
        # indexes: [kv_cache_group_id][attn_group]
        self.attn_groups: list[list[AttentionGroup]] = []
311
312
        # self.kv_cache_config: KVCacheConfig

313
314
        # mm_hash ->  encoder_output
        self.encoder_cache: dict[str, torch.Tensor] = {}
315

316
        self.use_aux_hidden_state_outputs = False
317
318
319
320
321
322
323
324
        # 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():
325
                self.drafter = EagleProposer(self.vllm_config, self.device, self)  # type: ignore
326
327
328
329
                if self.speculative_config.method == "eagle3":
                    self.use_aux_hidden_state_outputs = True
            elif self.speculative_config.method == "medusa":
                self.drafter = MedusaProposer(
330
331
                    vllm_config=self.vllm_config, device=self.device
                )  # type: ignore
332
            else:
333
334
335
336
                raise ValueError(
                    "Unknown speculative decoding method: "
                    f"{self.speculative_config.method}"
                )
337
            self.rejection_sampler = RejectionSampler()
338

339
        # Request states.
340
        self.requests: dict[str, CachedRequestState] = {}
341
        self.comm_stream = torch.cuda.Stream()
342

343
344
345
346
347
348
349
350
351
        # 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.
352
353
        self.input_batch = InputBatch(
            max_num_reqs=self.max_num_reqs,
354
355
356
            # 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),
357
358
359
            max_num_batched_tokens=self.max_num_tokens,
            device=self.device,
            pin_memory=self.pin_memory,
360
            vocab_size=self.model_config.get_vocab_size(),
361
            block_sizes=[self.cache_config.block_size],
362
            kernel_block_sizes=[self.cache_config.block_size],
363
            is_spec_decode=bool(self.vllm_config.speculative_config),
364
            logitsprocs=build_logitsprocs(
365
366
367
                self.vllm_config,
                self.device,
                self.pin_memory,
368
                self.is_pooling_model,
369
370
                self.vllm_config.model_config.logits_processors,
            ),
371
            is_pooling_model=self.is_pooling_model,
372
        )
373

374
        self.use_async_scheduling = self.scheduler_config.async_scheduling
375
376
377
        self.async_output_copy_stream = (
            torch.cuda.Stream() if self.use_async_scheduling else None
        )
378

379
        # TODO(woosuk): Provide an option to tune the max cudagraph batch size.
380
381
382
        # The convention is different.
        # self.cudagraph_batch_sizes sorts in ascending order.
        # The batch sizes in the config are in descending order.
383
384
385
386
        if (
            self.compilation_config.cudagraph_capture_sizes
            and self.compilation_config.cudagraph_mode != CUDAGraphMode.NONE
        ):
387
            self.cudagraph_batch_sizes = list(
388
389
                reversed(self.compilation_config.cudagraph_capture_sizes)
            )
390

391
        # Cache the device properties.
392
        self._init_device_properties()
393

394
        # Persistent buffers for CUDA graphs.
395
396
397
398
399
        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
        )
400
        self.seq_lens = self._make_buffer(self.max_num_reqs, dtype=torch.int32)
401
402
403
404
        if self.dcp_world_size > 1:
            self.dcp_local_seq_lens = self._make_buffer(
                self.max_num_reqs, dtype=torch.int32
            )
405
406
407
        # 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.
408
409
410
411
412
413
414
        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
        )
415
416
        self.num_discarded_requests = 0

417
418
419
420
421
422
        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
        )
423

424
425
        # Only relevant for multimodal models
        if self.supports_mm_inputs:
426
            self.is_mm_embed = self._make_buffer(self.max_num_tokens, dtype=torch.bool)
427

428
        # Only relevant for models using M-RoPE (e.g, Qwen2-VL)
429
        if self.uses_mrope:
Roger Wang's avatar
Roger Wang committed
430
431
432
433
            # 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
434
435
436
437
438
439

            # 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
440
            self.mrope_positions = self._make_buffer(
441
442
                (3, self.max_num_tokens + 1), dtype=torch.int64
            )
443

444
445
446
447
448
449
450
451
        # 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())

452
453
        # None in the first PP rank. The rest are set after load_model.
        self.intermediate_tensors: Optional[IntermediateTensors] = None
454

455
        # OPTIMIZATION: Cache the tensors rather than creating them every step.
456
        # Keep in int64 to avoid overflow with long context
457
458
459
460
        self.arange_np = np.arange(
            max(self.max_num_reqs + 1, self.max_model_len, self.max_num_tokens),
            dtype=np.int64,
        )
461

462
463
464
465
466
        # 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] = {}
467
468
469
470
471
        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(
472
473
                self.max_num_tokens, dtype=torch.int32, device=self.device
            )
474

475
476
477
478
479
        self.uniform_decode_query_len = (
            1
            if not self.speculative_config
            else 1 + self.speculative_config.num_speculative_tokens
        )
480
481
482
483

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

484
485
486
487
488
489
490
491
492
        self.mm_budget = (
            MultiModalBudget(
                self.model_config,
                self.scheduler_config,
                self.mm_registry,
            )
            if self.supports_mm_inputs
            else None
        )
493

494
495
        self.reorder_batch_threshold: Optional[int] = None

496
497
498
499
500
        # 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()

501
        # Cached outputs.
502
        self._draft_token_ids: Optional[Union[list[list[int]], torch.Tensor]] = None
503
504
505
506
507
        self.transfer_event = torch.cuda.Event()
        self.sampled_token_ids_pinned_cpu = torch.empty(
            (self.max_model_len, 1),
            dtype=torch.int64,
            device="cpu",
508
509
            pin_memory=self.pin_memory,
        )
510

511
512
513
514
    def reset_mm_cache(self) -> None:
        if self.mm_budget:
            self.mm_budget.reset_cache()

515
516
517
518
519
520
521
522
523
524
    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]

525
526
527
528
529
530
531
532
533
534
    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,
        )
535

536
537
538
    def _init_model_kwargs(self, num_tokens: int):
        model_kwargs = dict[str, Any]()

539
        if not self.is_pooling_model:
540
541
            return model_kwargs

542
543
        num_reqs = self.input_batch.num_reqs
        pooling_params = self.input_batch.get_pooling_params()
544
545
546

        token_type_id_requests = dict[int, Any]()
        for i, param in enumerate(pooling_params):
547
548
549
550
551
            if (
                param.extra_kwargs is not None
                and (token_types := param.extra_kwargs.get("compressed_token_type_ids"))
                is not None
            ):
552
553
554
555
556
                token_type_id_requests[i] = token_types

        if len(token_type_id_requests) == 0:
            return model_kwargs

557
        seq_lens = self.seq_lens.gpu[:num_reqs]
558
559
560
561
562
563
564
565
        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(
566
567
            device=self.device
        )
568
569
        return model_kwargs

570
    def _may_reorder_batch(self, scheduler_output: "SchedulerOutput") -> None:
571
572
        """
        Update the order of requests in the batch based on the attention
573
        backend's needs. For example, some attention backends (namely MLA) may
574
575
576
577
578
579
        want to separate requests based on if the attention computation will be
        compute-bound or memory-bound.

        Args:
            scheduler_output: The scheduler output.
        """
580
581
582
583
584
585
586
587
        # 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

588
        if self.reorder_batch_threshold is not None:
589
590
591
            # 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.
592
593
594
595
            if (
                self.dcp_world_size > 1
                and envs.VLLM_ATTENTION_BACKEND != "FLASH_ATTN_MLA"
            ):
596
                assert self.reorder_batch_threshold == 1, (
597
                    "DCP not support reorder_batch_threshold > 1 now."
598
                )
599
600
601
            reorder_batch_to_split_decodes_and_prefills(
                self.input_batch,
                scheduler_output,
602
603
                decode_threshold=self.reorder_batch_threshold,
            )
604

605
606
    # Note: used for model runner override.
    def _init_device_properties(self) -> None:
607
        """Initialize attributes from torch.cuda.get_device_properties"""
608
609
610
611
612
613
614
        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()

615
    def _update_states(self, scheduler_output: "SchedulerOutput") -> None:
616
617
618
619
620
621
        """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.

622
623
        The SamplingMetadata is updated and copied to the GPU if there is a
        new/resumed/paused/finished request in the batch.
624
625
        """
        # Remove finished requests from the cached states.
626
627
        for req_id in scheduler_output.finished_req_ids:
            self.requests.pop(req_id, None)
628
629
630
631
632
633
634
        # 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:
635
            self.input_batch.remove_request(req_id)
636
637

        # Free the cached encoder outputs.
638
639
        for mm_hash in scheduler_output.free_encoder_mm_hashes:
            self.encoder_cache.pop(mm_hash, None)
640

641
642
643
644
645
646
647
648
649
650
651
652
653
        # 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:
654
            self.input_batch.remove_request(req_id)
655

656
        reqs_to_add: list[CachedRequestState] = []
657
        # Add new requests to the cached states.
658
659
660
        for new_req_data in scheduler_output.scheduled_new_reqs:
            req_id = new_req_data.req_id
            sampling_params = new_req_data.sampling_params
661
            pooling_params = new_req_data.pooling_params
662

663
664
665
666
            if (
                sampling_params
                and sampling_params.sampling_type == SamplingType.RANDOM_SEED
            ):
667
668
669
670
671
                generator = torch.Generator(device=self.device)
                generator.manual_seed(sampling_params.seed)
            else:
                generator = None

672
673
            if self.is_pooling_model:
                assert pooling_params is not None
674
675
                task = pooling_params.task
                assert task is not None, "You did not set `task` in the API"
676

677
                model = cast(VllmModelForPooling, self.get_model())
678
                to_update = model.pooler.get_pooling_updates(task)
679
680
                to_update.apply(pooling_params)

681
            req_state = CachedRequestState(
682
                req_id=req_id,
683
                prompt_token_ids=new_req_data.prompt_token_ids,
684
                prompt_embeds=new_req_data.prompt_embeds,
685
                mm_features=new_req_data.mm_features,
686
                sampling_params=sampling_params,
687
                pooling_params=pooling_params,
688
                generator=generator,
689
690
                block_ids=new_req_data.block_ids,
                num_computed_tokens=new_req_data.num_computed_tokens,
691
                output_token_ids=[],
692
                lora_request=new_req_data.lora_request,
693
            )
694
695
            self.requests[req_id] = req_state

696
            # Only relevant for models using M-RoPE (e.g, Qwen2-VL)
697
            if self.uses_mrope:
698
                self._init_mrope_positions(req_state)
699

700
            reqs_to_add.append(req_state)
701

702
        # Update the states of the running/resumed requests.
703
        is_last_rank = get_pp_group().is_last_rank
704
705
        req_data = scheduler_output.scheduled_cached_reqs
        for i, req_id in enumerate(req_data.req_ids):
706
            req_state = self.requests[req_id]
707
708
709
            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]
710
            num_output_tokens = req_data.num_output_tokens[i]
711

712
            # Update the cached states.
713

714
            req_state.num_computed_tokens = num_computed_tokens
715
            req_index = self.input_batch.req_id_to_index.get(req_id)
716
717
718
719
720
721
722
723

            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.
724
725
726
                num_new_tokens = (
                    num_computed_tokens + len(new_token_ids) - req_state.num_tokens
                )
727
728
729
730
                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:
731
                    req_state.output_token_ids.extend(new_token_ids[-num_new_tokens:])
732
733
734
735
736
            elif num_output_tokens < len(req_state.output_token_ids):
                # Some output tokens were discarded due to a sync-KV-load
                # failure. Align the cached state.
                del req_state.output_token_ids[num_output_tokens:]
                if req_index is not None:
737
738
739
740
                    end_idx = (
                        self.input_batch.num_prompt_tokens[req_index]
                        + num_output_tokens
                    )
741
742
                    self.input_batch.num_tokens[req_index] = end_idx
                    self.input_batch.num_tokens_no_spec[req_index] = end_idx
743

744
            # Update the block IDs.
745
            if not resumed_from_preemption:
746
747
                if new_block_ids is not None:
                    # Append the new blocks to the existing block IDs.
748
                    for block_ids, new_ids in zip(req_state.block_ids, new_block_ids):
749
                        block_ids.extend(new_ids)
750
            else:
751
                assert req_index is None
752
                assert new_block_ids is not None
753
754
                # The request is resumed from preemption.
                # Replace the existing block IDs with the new ones.
755
                req_state.block_ids = new_block_ids
756
757
758
759
760

            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.
761
                reqs_to_add.append(req_state)
762
763
764
                continue

            # Update the persistent batch.
765
            self.input_batch.num_computed_tokens_cpu[req_index] = num_computed_tokens
766
            if new_block_ids is not None:
767
                self.input_batch.block_table.append_row(new_block_ids, req_index)
768
769
770
771
772
773
774

            # 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)
775
                self.input_batch.token_ids_cpu[
776
777
778
                    req_index, start_token_index:end_token_index
                ] = new_token_ids
                self.input_batch.num_tokens_no_spec[req_index] = end_token_index
779
                self.input_batch.num_tokens[req_index] = end_token_index
780

781
            # Add spec_token_ids to token_ids_cpu.
782
783
784
            spec_token_ids = scheduler_output.scheduled_spec_decode_tokens.get(
                req_id, ()
            )
785
786
787
788
789
            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[
790
791
                    req_index, start_index:end_token_index
                ] = spec_token_ids
792
793
                # NOTE(woosuk): `num_tokens` here may include spec tokens.
                self.input_batch.num_tokens[req_index] += num_spec_tokens
794
                self.input_batch.spec_token_ids[req_index] = spec_token_ids
795

796
797
        # Add the new or resumed requests to the persistent batch.
        # The smaller empty indices are filled first.
798
799
        for request in reqs_to_add:
            self.input_batch.add_request(request)
800

801
802
803
804
805
806
        # 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()
807

808
    def _update_states_after_model_execute(
809
810
        self, output_token_ids: torch.Tensor
    ) -> None:
811
812
813
814
815
816
817
818
819
820
821
822
        """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.
823
824
825
826
827
828
829
830
831
832
833
834
835
836
837
838
839
840
841
842
        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()
        )
843
844
845
        for i, num_tokens in enumerate(num_accepted_tokens):
            self.input_batch.num_accepted_tokens_cpu[i] = num_tokens

846
847
848
849
850
851
    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
852
853
854
855
        for mm_feature in req_state.mm_features:
            mm_item = mm_feature.data
            if mm_item is None:
                continue
856
857
858
859
860
861
862
863
864
865
866
867
            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

868
        if supports_mrope(self.model):
869
            req_state.mrope_positions, req_state.mrope_position_delta = (
870
871
872
873
874
875
876
877
878
                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,
                )
879
            )
880
        else:
881
            req_state.mrope_positions, req_state.mrope_position_delta = (
882
883
884
885
886
887
888
889
890
                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,
                )
891
            )
892

893
    def _extract_mm_kwargs(
894
        self,
895
896
        scheduler_output: "SchedulerOutput",
    ) -> BatchedTensorInputs:
897
        if not scheduler_output or not self.is_multimodal_raw_input_only_model:
898
            return {}
899

900
901
        mm_kwargs = list[MultiModalKwargsItem]()
        for req in scheduler_output.scheduled_new_reqs:
902
903
904
            for feature in req.mm_features:
                if feature.data is not None:
                    mm_kwargs.append(feature.data)
905

906
        # Input all modalities at once
907
        model = cast(SupportsMultiModal, self.model)
908
909
        mm_kwargs_combined: BatchedTensorInputs = {}
        for _, _, mm_kwargs_group in group_mm_kwargs_by_modality(
910
911
912
913
            mm_kwargs,
            device=self.device,
            pin_memory=self.pin_memory,
            merge_by_field_config=model.merge_by_field_config,
914
915
        ):
            mm_kwargs_combined.update(mm_kwargs_group)
916

917
        return mm_kwargs_combined
918

919
    def _dummy_mm_kwargs(self, num_seqs: int) -> BatchedTensorInputs:
920
        if not self.is_multimodal_raw_input_only_model:
921
            return {}
922

923
924
925
926
927
        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)
928

929
930
931
932
933
934
935
936
937
938
939
940
941
942
943
944
945
946
947
948
    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

949
950
951
    def _prepare_input_ids(
        self, total_num_scheduled_tokens: int, cu_num_tokens: np.ndarray
    ) -> None:
952
        """Prepare the input IDs for the current batch.
953

954
955
956
957
958
959
960
        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)
961
962
963
            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)
964
965
966
967
968
969
970
971
972
973
974
975
976
977
978
979
980
981
            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)
982
                indices_match &= prev_index == flattened_index
983
984
985
986
987
988
                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)
989
990
991
            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)
992
993
994
995
996
997
998
999
1000
1001
        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_(
1002
1003
1004
                self.input_batch.prev_sampled_token_ids[:num_commmon_tokens, 0],
                non_blocking=True,
            )
1005
1006
            if self.enable_prompt_embeds:
                self.is_token_ids.gpu[:num_commmon_tokens] = True
1007
1008
1009
            return
        # Upload the index tensors asynchronously
        # so the scatter can be non-blocking.
1010
1011
1012
        input_ids_index_tensor = torch.tensor(
            flattened_indices, dtype=torch.int64, pin_memory=self.pin_memory
        ).to(self.device, non_blocking=True)
1013
        prev_common_req_indices_tensor = torch.tensor(
1014
1015
            prev_common_req_indices, dtype=torch.int64, pin_memory=self.pin_memory
        ).to(self.device, non_blocking=True)
1016
1017
1018
1019
        self.input_ids.gpu.scatter_(
            dim=0,
            index=input_ids_index_tensor,
            src=self.input_batch.prev_sampled_token_ids[
1020
1021
1022
                prev_common_req_indices_tensor, 0
            ],
        )
1023

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

1042
    def _prepare_inputs(
1043
        self, scheduler_output: "SchedulerOutput"
1044
1045
1046
1047
1048
1049
1050
1051
1052
1053
1054
    ) -> tuple[
        PerLayerAttnMetadata,
        torch.Tensor,
        Optional[SpecDecodeMetadata],
        np.ndarray,
        Optional[CommonAttentionMetadata],
        int,
        Optional[UBatchSlices],
        Optional[torch.Tensor],
        bool,
    ]:
1055
1056
1057
        """
        :return: tuple[
            attn_metadata: layer-to-attention_metadata mapping,
1058
1059
1060
            logits_indices, spec_decode_metadata,
            num_scheduled_tokens, spec_decode_common_attn_metadata,
            max_num_scheduled_tokens, use_cascade_attn
1061
1062
        ]
        """
1063
1064
1065
1066
1067
1068
1069
        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.
1070
        self.input_batch.block_table.commit_block_table(num_reqs)
1071
1072

        # Get the number of scheduled tokens for each request.
1073
1074
1075
1076
        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)
1077
1078
1079

        # Get request indices.
        # E.g., [2, 5, 3] -> [0, 0, 1, 1, 1, 1, 1, 2, 2, 2]
1080
        req_indices = np.repeat(self.arange_np[:num_reqs], num_scheduled_tokens)
1081

1082
1083
        # cu_num_tokens: [2, 5, 3] -> [2, 7, 10]
        # arange: [0, 1, 0, 1, 2, 3, 4, 0, 1, 2]
1084
        cu_num_tokens, arange = self._get_cumsum_and_arange(num_scheduled_tokens)
1085
1086

        # Get positions.
1087
        positions_np = self.positions.np[:total_num_scheduled_tokens]
1088
1089
1090
1091
1092
        np.add(
            self.input_batch.num_computed_tokens_cpu[req_indices],
            arange,
            out=positions_np,
        )
1093

1094
1095
        # Calculate M-RoPE positions.
        # Only relevant for models using M-RoPE (e.g, Qwen2-VL)
1096
        if self.uses_mrope:
1097
1098
            self._calc_mrope_positions(scheduler_output)

1099
1100
1101
1102
        # 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.
1103
1104
1105
        token_indices = (
            positions_np + req_indices * self.input_batch.token_ids_cpu.shape[1]
        )
1106
        token_indices_tensor = torch.from_numpy(token_indices)
1107

1108
1109
1110
        # 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.
1111
1112
1113
1114
1115
1116
        torch.index_select(
            self.input_batch.token_ids_cpu_tensor.flatten(),
            0,
            token_indices_tensor,
            out=self.input_ids.cpu[:total_num_scheduled_tokens],
        )
1117
1118
1119
1120
1121
1122
        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,
1123
1124
                out=self.is_token_ids.cpu[:total_num_scheduled_tokens],
            )
1125
1126
1127
1128
1129
1130
1131
1132
1133
1134
1135
1136
1137
1138
1139
1140
1141
1142
1143
1144
1145
1146
1147
1148
1149
1150
1151
1152
1153
1154
1155
1156
1157

        # 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:
1158
1159
1160
                    self.inputs_embeds.cpu[
                        output_idx : output_idx + actual_num_sched
                    ].copy_(req_embeds[start_pos:actual_end])
1161
1162

                output_idx += num_sched
1163

1164
1165
        self.input_batch.block_table.compute_slot_mapping(req_indices, positions_np)
        self.input_batch.block_table.commit_slot_mapping(total_num_scheduled_tokens)
1166
1167

        # Prepare the attention metadata.
1168
        self.query_start_loc.np[0] = 0
1169
        self.query_start_loc.np[1 : num_reqs + 1] = cu_num_tokens
1170
1171
        # Note: pad query_start_loc to be non-decreasing, as kernels
        # like FlashAttention requires that
1172
        self.query_start_loc.np[num_reqs + 1 :].fill(cu_num_tokens[-1])
1173
        self.query_start_loc.copy_to_gpu()
1174
        query_start_loc = self.query_start_loc.gpu[: num_reqs + 1]
1175

1176
        num_tokens_unpadded = scheduler_output.total_num_scheduled_tokens
1177
        num_tokens_padded = self._get_num_input_tokens(num_tokens_unpadded)
1178
1179
1180
        uniform_decode = (
            max_num_scheduled_tokens == self.uniform_decode_query_len
        ) and (total_num_scheduled_tokens == num_reqs * max_num_scheduled_tokens)
1181
1182
1183
1184
1185
1186
1187

        # Disable DP padding when running eager to avoid excessive padding when
        # running prefills. This lets us set enforce_eager on the prefiller in
        # a P/D setup and still use CUDA graphs (enabled by this padding) on the
        # decoder.
        allow_dp_padding = self.compilation_config.cudagraph_mode != CUDAGraphMode.NONE

1188
        ubatch_slices, num_tokens_across_dp = coordinate_batch_across_dp(
1189
1190
1191
1192
1193
1194
1195
            num_tokens_unpadded=num_tokens_unpadded,
            parallel_config=self.parallel_config,
            allow_microbatching=True,
            allow_dp_padding=allow_dp_padding,
            num_tokens_padded=num_tokens_padded,
            uniform_decode=uniform_decode,
            num_scheduled_tokens_per_request=num_scheduled_tokens,
1196
        )
1197

1198
        self.seq_lens.np[:num_reqs] = (
1199
1200
            self.input_batch.num_computed_tokens_cpu[:num_reqs] + num_scheduled_tokens
        )
1201
        # Fill unused with 0 for full cuda graph mode.
1202
1203
1204
1205
        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()
1206

1207
        num_tokens = [self.requests[r].num_tokens for r in self.input_batch.req_ids]
1208
1209
1210
1211
1212
1213
1214
        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)
1215
1216
1217
        self.discard_request_indices.np[: self.num_discarded_requests] = (
            discard_request_indices
        )
1218
1219
1220

        self.discard_request_indices.copy_to_gpu(self.num_discarded_requests)

1221
        # Copy the tensors to the GPU.
1222
1223
        self._prepare_input_ids(total_num_scheduled_tokens, cu_num_tokens)

1224
        if self.uses_mrope:
1225
            # Only relevant for models using M-RoPE (e.g, Qwen2-VL)
1226
1227
            self.mrope_positions.gpu[:, :total_num_scheduled_tokens].copy_(
                self.mrope_positions.cpu[:, :total_num_scheduled_tokens],
1228
1229
                non_blocking=True,
            )
1230
1231
        else:
            # Common case (1D positions)
1232
            self.positions.copy_to_gpu(total_num_scheduled_tokens)
1233

1234
        use_spec_decode = len(scheduler_output.scheduled_spec_decode_tokens) > 0
1235
1236
1237
1238
1239
1240
1241
        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
1242
            num_draft_tokens = None
1243
1244
1245
1246
1247
1248
            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)
1249
1250
1251
            # 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)
1252
1253
1254
1255
            for (
                req_id,
                draft_token_ids,
            ) in scheduler_output.scheduled_spec_decode_tokens.items():
1256
1257
                req_idx = self.input_batch.req_id_to_index[req_id]
                num_draft_tokens[req_idx] = len(draft_token_ids)
1258
1259
1260
1261
1262
1263
1264
1265
                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
                )
1266
            spec_decode_metadata = self._calc_spec_decode_metadata(
1267
1268
                num_draft_tokens, cu_num_tokens
            )
1269
            logits_indices = spec_decode_metadata.logits_indices
1270
1271

            # For DECODE only cuda graph of some attention backends (e.g., GDN).
1272
            self.num_decode_draft_tokens.np[:num_reqs] = num_decode_draft_tokens
1273
1274
            self.num_decode_draft_tokens.np[num_reqs:].fill(-1)
            self.num_decode_draft_tokens.copy_to_gpu()
1275
1276
1277

        logits_indices_padded = None
        if self.cache_config.kv_sharing_fast_prefill:
1278
            logits_indices_padded = self._prepare_kv_sharing_fast_prefill(
1279
1280
                logits_indices
            )
1281

1282
1283
1284
        attn_metadata: PerLayerAttnMetadata = {}
        if ubatch_slices is not None:
            attn_metadata = [dict() for _ in range(len(ubatch_slices))]
1285
        use_cascade_attn = False
1286

1287
        # Used in the below loop.
1288
        query_start_loc_cpu = self.query_start_loc.cpu[: num_reqs + 1]
1289
        seq_lens_cpu = self.seq_lens.cpu[:num_reqs]
1290
1291
1292
        num_computed_tokens_cpu = self.input_batch.num_computed_tokens_cpu_tensor[
            :num_reqs
        ]
1293
        spec_decode_common_attn_metadata = None
1294
1295
        if use_spec_decode:
            self.num_accepted_tokens.np[:num_reqs] = (
1296
1297
                self.input_batch.num_accepted_tokens_cpu[:num_reqs]
            )
1298
1299
            self.num_accepted_tokens.np[num_reqs:].fill(1)
            self.num_accepted_tokens.copy_to_gpu()
1300

1301
1302
1303
        # 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(
1304
1305
            self.kv_cache_config.kv_cache_groups
        ):
1306
            encoder_seq_lens = self._get_encoder_seq_lens(
1307
1308
                scheduler_output, kv_cache_group_spec.kv_cache_spec, num_reqs
            )
1309

1310
            if isinstance(kv_cache_group_spec.kv_cache_spec, EncoderOnlyAttentionSpec):
1311
1312
1313
1314
1315
                # 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,
1316
1317
1318
                    device=self.device,
                )
                slot_mapping = torch.zeros(
1319
                    (total_num_scheduled_tokens,),
1320
1321
1322
                    dtype=torch.int64,
                    device=self.device,
                )
1323
1324
1325
                num_common_prefix_blocks = 0
            else:
                blk_table = self.input_batch.block_table[kv_cache_group_id]
1326
                blk_table_tensor = blk_table.get_device_tensor(num_reqs)
1327
                slot_mapping = blk_table.slot_mapping.gpu[:total_num_scheduled_tokens]
1328
1329
1330

                # Fill unused with -1. Needed for reshape_and_cache in full cuda
                # graph mode.
1331
1332
1333
1334
                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
                ]
1335

1336
            common_attn_metadata = CommonAttentionMetadata(
1337
1338
1339
1340
1341
                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,
1342
1343
1344
                num_reqs=num_reqs,
                num_actual_tokens=total_num_scheduled_tokens,
                max_query_len=max_num_scheduled_tokens,
1345
                max_seq_len=max_seq_len,
1346
1347
                block_table_tensor=blk_table_tensor,
                slot_mapping=slot_mapping,
1348
1349
                logits_indices_padded=logits_indices_padded,
                num_logits_indices=logits_indices.size(0),
1350
                causal=True,
1351
                encoder_seq_lens=encoder_seq_lens,
1352
1353
1354
                dcp_local_seq_lens=self.dcp_local_seq_lens.gpu[:num_reqs]
                if self.dcp_world_size > 1
                else None,
1355
1356
            )

1357
            if self.speculative_config and spec_decode_common_attn_metadata is None:
1358
                if isinstance(self.drafter, EagleProposer):
1359
1360
1361
1362
                    if (
                        self.drafter.attn_layer_names[0]
                        in kv_cache_group_spec.layer_names
                    ):
1363
1364
1365
                        spec_decode_common_attn_metadata = common_attn_metadata
                else:
                    spec_decode_common_attn_metadata = common_attn_metadata
1366

1367
1368
1369
            for attn_group in self.attn_groups[kv_cache_group_id]:
                # Prepare for cascade attention if enabled & beneficial.
                common_prefix_len = 0
1370
                builder = attn_group.get_metadata_builder()
1371
1372
1373
                if self.cascade_attn_enabled:
                    common_prefix_len = self._compute_cascade_attn_prefix_len(
                        num_scheduled_tokens,
1374
                        num_common_prefix_blocks,
1375
                        attn_group.kv_cache_spec,
1376
1377
                        builder,
                    )
1378

1379
                extra_attn_metadata_args = {}
1380
                if use_spec_decode and isinstance(builder, GDNAttentionMetadataBuilder):
1381
                    extra_attn_metadata_args = dict(
1382
1383
1384
1385
                        num_accepted_tokens=self.num_accepted_tokens.gpu[:num_reqs],
                        num_decode_draft_tokens_cpu=self.num_decode_draft_tokens.cpu[
                            :num_reqs
                        ],
1386
1387
                    )

1388
1389
                if ubatch_slices is not None:
                    common_attn_metadata_list = split_attn_metadata(
1390
1391
                        ubatch_slices, common_attn_metadata
                    )
1392
                    for ubid, common_attn_metadata in enumerate(
1393
1394
1395
1396
1397
1398
1399
1400
                        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,
                        )
1401
1402
1403
1404
1405
1406
1407
1408
                        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,
1409
1410
1411
                        **extra_attn_metadata_args,
                    )
                    use_cascade_attn |= getattr(attn_metadata_i, "use_cascade", False)
1412
1413
                    for layer_name in attn_group.layer_names:
                        attn_metadata[layer_name] = attn_metadata_i
1414

1415
1416
1417
1418
        # disable cascade attention when DBO
        if ubatch_slices is not None:
            use_cascade_attn = False

1419
1420
1421
1422
        # Hot-Swap lora model
        if self.lora_config:
            self.set_active_loras(self.input_batch, num_scheduled_tokens)

1423
1424
1425
1426
1427
1428
1429
1430
        return (
            attn_metadata,
            logits_indices,
            spec_decode_metadata,
            num_scheduled_tokens,
            spec_decode_common_attn_metadata,
            max_num_scheduled_tokens,
            ubatch_slices,
1431
            num_tokens_across_dp,
1432
1433
            use_cascade_attn,
        )
1434

1435
1436
1437
1438
    def _compute_cascade_attn_prefix_len(
        self,
        num_scheduled_tokens: np.ndarray,
        num_common_prefix_blocks: int,
1439
1440
        kv_cache_spec: KVCacheSpec,
        attn_metadata_builder: AttentionMetadataBuilder,
1441
1442
1443
1444
1445
1446
1447
1448
1449
1450
1451
1452
1453
1454
1455
1456
1457
1458
    ) -> 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.
        """
1459
        common_prefix_len = num_common_prefix_blocks * kv_cache_spec.block_size
1460
1461
1462
1463
1464
1465
1466
1467
1468
1469
1470
1471
1472
1473
1474
1475
1476
1477
1478
1479
1480
1481
1482
1483
1484
1485
1486
1487
1488
1489
1490
1491
1492
1493
1494
1495
1496
        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]
1497
        # Request 3's num_computed_tokens: 3 (i.e., [A, B, C])
1498
1499
1500
1501
1502
1503
1504
        # 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(
1505
1506
            common_prefix_len, self.input_batch.num_computed_tokens_cpu[:num_reqs].min()
        )
1507
        # common_prefix_len should be a multiple of the block size.
1508
1509
1510
1511
1512
1513
1514
1515
1516
1517
1518
        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
        )
1519
1520
        assert isinstance(kv_cache_spec, AttentionSpec)
        use_cascade = attn_metadata_builder.use_cascade_attention(
1521
1522
1523
            common_prefix_len=common_prefix_len,
            query_lens=num_scheduled_tokens,
            num_query_heads=self.num_query_heads,
1524
            num_kv_heads=kv_cache_spec.num_kv_heads,
1525
            use_alibi=self.use_alibi,
1526
            use_sliding_window=use_sliding_window,
1527
            use_local_attention=use_local_attention,
1528
1529
1530
1531
            num_sms=self.num_sms,
        )
        return common_prefix_len if use_cascade else 0

1532
1533
    def _calc_mrope_positions(self, scheduler_output: "SchedulerOutput"):
        mrope_pos_ptr = 0
1534
        for index, req_id in enumerate(self.input_batch.req_ids):
1535
1536
1537
            req = self.requests[req_id]
            assert req.mrope_positions is not None

1538
1539
            num_computed_tokens = self.input_batch.num_computed_tokens_cpu[index]
            num_scheduled_tokens = scheduler_output.num_scheduled_tokens[req_id]
1540
            num_prompt_tokens = length_from_prompt_token_ids_or_embeds(
1541
1542
                req.prompt_token_ids, req.prompt_embeds
            )
1543
1544

            if num_computed_tokens + num_scheduled_tokens > num_prompt_tokens:
1545
1546
                prompt_part_len = max(0, num_prompt_tokens - num_computed_tokens)
                completion_part_len = max(0, num_scheduled_tokens - prompt_part_len)
1547
1548
1549
1550
1551
1552
1553
1554
1555
1556
1557
1558
1559
            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

1560
1561
1562
                self.mrope_positions.cpu[:, dst_start:dst_end] = req.mrope_positions[
                    :, src_start:src_end
                ]
1563
1564
1565
1566
1567
1568
1569
                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

1570
                MRotaryEmbedding.get_next_input_positions_tensor(
1571
                    out=self.mrope_positions.np,
1572
1573
1574
1575
1576
                    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,
                )
1577
1578
1579

                mrope_pos_ptr += completion_part_len

1580
1581
    def _calc_spec_decode_metadata(
        self,
1582
1583
1584
1585
1586
1587
1588
1589
1590
1591
1592
1593
1594
1595
1596
1597
        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
1598
1599
1600
1601

        # 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(
1602
1603
            num_sampled_tokens, cumsum_dtype=np.int32
        )
1604
        # Step 2. [0, 0, 0, 0, 103, 104, 104, 104, 206, 207, 207]
1605
        logits_indices = np.repeat(
1606
1607
            cu_num_scheduled_tokens - num_sampled_tokens, num_sampled_tokens
        )
1608
        # Step 3. [0, 1, 2, 3, 103, 104, 105, 106, 206, 207, 208]
1609
1610
1611
1612
1613
1614
        logits_indices += arange

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

        # Compute the draft logits indices.
1615
1616
1617
        # 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(
1618
1619
            num_draft_tokens, cumsum_dtype=np.int32
        )
1620
1621
        # [0, 0, 0, 5, 5, 9]
        target_logits_indices = np.repeat(
1622
1623
            cu_num_sampled_tokens - num_sampled_tokens, num_draft_tokens
        )
1624
1625
1626
1627
1628
        # [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(
1629
1630
1631
1632
1633
            self.device, non_blocking=True
        )
        logits_indices = torch.from_numpy(logits_indices).to(
            self.device, non_blocking=True
        )
1634
        target_logits_indices = torch.from_numpy(target_logits_indices).to(
1635
1636
            self.device, non_blocking=True
        )
1637
        bonus_logits_indices = torch.from_numpy(bonus_logits_indices).to(
1638
1639
            self.device, non_blocking=True
        )
1640

1641
1642
        # Compute the draft token ids.
        # draft_token_indices:      [  1,   2,   3, 105, 106, 208]
1643
        draft_token_ids = self.input_ids.gpu[logits_indices]
1644
1645
1646
1647
1648
1649
1650
1651
1652
1653
1654
1655
        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

1656
1657
1658
1659
1660
1661
1662
    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
1663
        self.kv_sharing_fast_prefill_logits_indices[:num_logits].copy_(logits_indices)
1664
1665
1666
1667
1668
        # 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_(
1669
1670
1671
1672
1673
1674
            logits_indices[-1].item()
        )
        if (
            self.compilation_config.cudagraph_mode != CUDAGraphMode.NONE
            and num_logits <= self.cudagraph_batch_sizes[-1]
        ):
1675
1676
1677
1678
1679
            # 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
1680
1681
1682
        logits_indices_padded = self.kv_sharing_fast_prefill_logits_indices[
            :num_logits_padded
        ]
1683
1684
        return logits_indices_padded

1685
1686
1687
1688
1689
1690
1691
1692
    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
1693
                inputs.
1694
1695
1696
1697
1698
1699

        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
        """
1700
1701
        scheduled_encoder_inputs = scheduler_output.scheduled_encoder_inputs
        if not scheduled_encoder_inputs:
1702
            return [], []
1703
        # Batch the multi-modal inputs.
1704
        mm_kwargs = list[MultiModalKwargsItem]()
1705
1706
        # list of tuple (mm_hash, position_info)
        mm_hashes_pos = list[tuple[str, PlaceholderRange]]()
1707
1708
        for req_id, encoder_input_ids in scheduled_encoder_inputs.items():
            req_state = self.requests[req_id]
1709
1710

            for mm_input_id in encoder_input_ids:
1711
1712
1713
1714
                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))
1715

1716
1717
1718
1719
1720
        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(
1721
1722
            scheduler_output
        )
1723
1724
1725
1726

        if not mm_kwargs:
            return

1727
1728
1729
1730
1731
1732
1733
        # 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.
1734
        model = cast(SupportsMultiModal, self.model)
1735
        encoder_outputs = []
1736
        for modality, num_items, mm_kwargs_group in group_mm_kwargs_by_modality(
1737
1738
1739
1740
            mm_kwargs,
            device=self.device,
            pin_memory=self.pin_memory,
            merge_by_field_config=model.merge_by_field_config,
1741
        ):
1742
1743
1744
1745
1746
1747
1748
1749
1750
1751
1752
1753
            # (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(
1754
1755
1756
                        (k, v[i : i + micro_batch_size])
                        for k, v in mm_kwargs_group.items()
                    )
1757
1758

                    micro_batch_outputs = model.get_multimodal_embeddings(
1759
1760
                        **micro_batch_mm_inputs
                    )
1761
1762
1763
1764
1765
1766
1767
1768
1769
1770

                    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.
1771
                curr_group_outputs = model.get_multimodal_embeddings(**mm_kwargs_group)
1772

1773
1774
            sanity_check_mm_encoder_outputs(
                curr_group_outputs,
1775
                expected_num_items=num_items,
1776
            )
1777
            encoder_outputs.extend(curr_group_outputs)
1778

1779
1780
1781
        # 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(
1782
1783
1784
1785
1786
                output,
                is_embed=pos_info.is_embed,
            )

    def _gather_mm_embeddings(
1787
1788
        self,
        scheduler_output: "SchedulerOutput",
1789
        shift_computed_tokens: int = 0,
1790
1791
1792
1793
1794
1795
1796
1797
    ) -> 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
1798
        should_sync_mrope_positions = False
1799

1800
        for req_id in self.input_batch.req_ids:
1801
1802
            mm_embeds_req: list[torch.Tensor] = []

1803
            num_scheduled_tokens = scheduler_output.num_scheduled_tokens[req_id]
1804
            req_state = self.requests[req_id]
1805
            num_computed_tokens = req_state.num_computed_tokens + shift_computed_tokens
1806

1807
1808
            for mm_feature in req_state.mm_features:
                pos_info = mm_feature.mm_position
1809
1810
                start_pos = pos_info.offset
                num_encoder_tokens = pos_info.length
1811
1812
1813
1814
1815
1816
1817
1818
1819
1820
1821
1822
1823
1824
1825
1826

                # 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,
1827
1828
                    num_encoder_tokens,
                )
1829
                assert start_idx < end_idx
1830

1831
                mm_hash = mm_feature.identifier
1832
                encoder_output = self.encoder_cache.get(mm_hash, None)
1833
                assert encoder_output is not None, f"Encoder cache miss for {mm_hash}."
1834
1835
1836
1837

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

1838
                req_start_pos = req_start_idx + start_pos - num_computed_tokens
1839
1840
1841
                is_mm_embed[req_start_pos + start_idx : req_start_pos + end_idx] = (
                    True if is_embed is None else is_embed
                )
1842

1843
1844
1845
1846
                mm_embeds_item = gather_mm_placeholders(
                    encoder_output[start_idx:end_idx],
                    is_embed=is_embed,
                )
1847
1848
1849
                mm_embeds_req.append(mm_embeds_item)

            if self.is_multimodal_pruning_enabled and self.uses_mrope:
1850
                assert req_state.mrope_positions is not None
1851
1852
1853
1854
1855
1856
1857
                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,
1858
1859
                    )
                )
1860
1861
1862
1863
                req_state.mrope_positions.copy_(new_mrope_positions)
                req_state.mrope_position_delta = new_delta

            mm_embeds.extend(mm_embeds_req)
1864
1865
1866
            req_start_idx += num_scheduled_tokens

        is_mm_embed = self.is_mm_embed.copy_to_gpu(total_num_scheduled_tokens)
1867
1868
1869

        if should_sync_mrope_positions:
            self._calc_mrope_positions(scheduler_output)
1870
            self.mrope_positions.copy_to_gpu(total_num_scheduled_tokens)
1871

1872
        return mm_embeds, is_mm_embed
1873

1874
1875
1876
1877
1878
1879
1880
1881
1882
1883
1884
1885
1886
1887
1888
1889
    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
1890
        model = cast(SupportsMultiModal, self.model)
1891
1892
        encoder_features = {}
        for _, _, mm_kwargs_group in group_mm_kwargs_by_modality(
1893
1894
1895
1896
            mm_kwargs,
            device=self.device,
            pin_memory=self.pin_memory,
            merge_by_field_config=model.merge_by_field_config,
1897
1898
1899
1900
1901
1902
1903
1904
        ):
            # 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

1905
    def get_model(self) -> nn.Module:
1906
        # get raw model out of the cudagraph wrapper.
1907
        if isinstance(self.model, (CUDAGraphWrapper, UBatchWrapper)):
1908
            return self.model.unwrap()
1909
1910
        return self.model

1911
1912
1913
1914
1915
1916
1917
1918
1919
1920
1921
1922
1923
1924
1925
    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

1926
1927
1928
1929
1930
    def get_supported_pooling_tasks(self) -> list[PoolingTask]:
        model = self.get_model()
        if not is_pooling_model(model):
            return []

1931
1932
        supported_tasks = list(model.pooler.get_supported_tasks())

1933
1934
1935
1936
        if (
            self.scheduler_config.chunked_prefill_enabled
            and "encode" in supported_tasks
        ):
1937
1938
            supported_tasks.remove("encode")

1939
1940
1941
1942
1943
1944
            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."
            )
1945
1946
1947
1948
1949

        if "score" in supported_tasks:
            num_labels = getattr(self.model_config.hf_config, "num_labels", 0)
            if num_labels != 1:
                supported_tasks.remove("score")
1950
                logger.debug_once("Score API is only enabled for num_labels == 1.")
1951
1952

        return supported_tasks
1953

1954
1955
1956
1957
1958
1959
1960
1961
1962
1963
    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)

1964
    def sync_and_slice_intermediate_tensors(
1965
1966
1967
1968
1969
        self,
        num_tokens: int,
        intermediate_tensors: IntermediateTensors,
        sync_self: bool,
    ) -> IntermediateTensors:
1970
1971
1972
        assert self.intermediate_tensors is not None

        tp = self.vllm_config.parallel_config.tensor_parallel_size
1973
        is_rs = is_residual_scattered_for_sp(self.vllm_config, num_tokens)
1974
1975
1976
1977
1978
1979

        # 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():
1980
                is_scattered = k == "residual" and is_rs
1981
                copy_len = num_tokens // tp if is_scattered else num_tokens
1982
                self.intermediate_tensors[k][:copy_len].copy_(
1983
1984
1985
1986
1987
1988
1989
1990
1991
1992
1993
1994
1995
                    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:
1996
1997
1998
1999
2000
2001
2002
        """
        Step for the EPLB (Expert Parallelism Load Balancing) state.
        """
        if not self.parallel_config.enable_eplb:
            return

        assert self.eplb_state is not None
2003
2004
        model = self.get_model()
        assert is_mixture_of_experts(model)
2005
        self.eplb_state.step(
2006
            model,
2007
2008
            is_dummy,
            is_profile,
2009
            log_stats=self.parallel_config.eplb_config.log_balancedness,
2010
2011
        )

2012
2013
2014
2015
    # 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)
2016
2017
    @staticmethod
    def pad_out_ubatch_slice(ubatch_slices: UBatchSlices, num_total_tokens: int):
2018
2019
2020
2021
2022
2023
        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
        )
2024

2025
2026
2027
2028
2029
2030
    def _pool(
        self,
        hidden_states: torch.Tensor,
        num_scheduled_tokens: int,
        num_scheduled_tokens_np: np.ndarray,
    ) -> ModelRunnerOutput:
2031
2032
2033
        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"
        )
2034

2035
        hidden_states = hidden_states[:num_scheduled_tokens]
2036
        pooling_metadata = self.input_batch.get_pooling_metadata()
2037
2038
2039
2040
        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]
2041

2042
2043
2044
2045
2046
2047
2048
2049
2050
2051
        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()
2052
2053
2054

        pooler_output: list[Optional[torch.Tensor]] = []
        for raw_output, seq_len, prompt_len in zip(
2055
2056
            raw_pooler_output, seq_lens_cpu, pooling_metadata.prompt_lens
        ):
2057
            output = raw_output if seq_len == prompt_len else None
2058
            pooler_output.append(output)
2059
2060
2061
2062
2063
2064
2065
2066
2067
2068

        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,
        )

2069
    def _get_num_input_tokens(self, num_scheduled_tokens: int) -> int:
2070
2071
2072
2073
2074
2075
2076
        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]
        ):
2077
2078
2079
2080
2081
2082
2083
2084
            # 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
2085
2086
2087
2088
        if (
            self.compilation_config.pass_config.enable_sequence_parallelism
            and tp_size > 1
        ):
2089
2090
2091
            return round_up(num_scheduled_tokens, tp_size)
        return num_scheduled_tokens

2092
    def _preprocess(
2093
2094
        self,
        scheduler_output: "SchedulerOutput",
2095
        num_input_tokens: int,  # Padded
2096
        intermediate_tensors: Optional[IntermediateTensors] = None,
2097
2098
2099
2100
2101
2102
2103
2104
    ) -> tuple[
        int,
        Optional[torch.Tensor],
        Optional[torch.Tensor],
        torch.Tensor,
        Optional[IntermediateTensors],
        dict[str, Any],
    ]:
2105
        num_scheduled_tokens = scheduler_output.total_num_scheduled_tokens
2106
        is_first_rank = get_pp_group().is_first_rank
2107

2108
2109
        # _prepare_inputs may reorder the batch, so we must gather multi
        # modal outputs after that to ensure the correct order
2110
2111
        if (
            self.supports_mm_inputs
2112
            and is_first_rank
2113
2114
            and not self.model_config.is_encoder_decoder
        ):
2115
2116
            # Run the multimodal encoder if any.
            self._execute_mm_encoder(scheduler_output)
2117
            mm_embeds, is_mm_embed = self._gather_mm_embeddings(scheduler_output)
2118

2119
2120
2121
            # 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.
2122
            inputs_embeds_scheduled = self.model.get_input_embeddings(
2123
2124
2125
                self.input_ids.gpu[:num_scheduled_tokens],
                multimodal_embeddings=mm_embeds,
                is_multimodal=is_mm_embed,
2126
            )
2127

2128
            # TODO(woosuk): Avoid the copy. Optimize.
2129
            self.inputs_embeds.gpu[:num_scheduled_tokens].copy_(inputs_embeds_scheduled)
2130

2131
            input_ids = None
2132
            inputs_embeds = self.inputs_embeds.gpu[:num_input_tokens]
2133
2134
2135
2136
            model_kwargs = {
                **self._init_model_kwargs(num_scheduled_tokens),
                **self._extract_mm_kwargs(scheduler_output),
            }
2137
        elif self.enable_prompt_embeds and is_first_rank:
2138
2139
2140
2141
2142
2143
2144
2145
2146
2147
2148
2149
            # 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).
2150
2151
2152
            token_ids_idx = (
                self.is_token_ids.gpu[:num_scheduled_tokens]
                .nonzero(as_tuple=False)
2153
                .squeeze(1)
2154
            )
2155
2156
2157
            # Some tokens ids may need to become embeds
            if token_ids_idx.numel() > 0:
                token_ids = self.input_ids.gpu[token_ids_idx]
2158
                tokens_to_embeds = self.model.get_input_embeddings(input_ids=token_ids)
2159
2160
2161
2162
2163
                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
2164
        else:
2165
2166
2167
2168
            # 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.
2169
            input_ids = self.input_ids.gpu[:num_input_tokens]
2170
            inputs_embeds = None
2171
            model_kwargs = self._init_model_kwargs(num_input_tokens)
2172
        if self.uses_mrope:
2173
            positions = self.mrope_positions.gpu[:, :num_input_tokens]
2174
        else:
2175
            positions = self.positions.gpu[:num_input_tokens]
2176

2177
        if is_first_rank:
2178
2179
            intermediate_tensors = None
        else:
2180
            intermediate_tensors = self.sync_and_slice_intermediate_tensors(
2181
2182
                num_input_tokens, intermediate_tensors, True
            )
2183

2184
2185
2186
2187
        if (
            self.model_config.is_encoder_decoder
            and scheduler_output.scheduled_encoder_inputs
        ):
2188
2189
2190
            encoder_inputs = self._extract_encoder_inputs(scheduler_output)
            model_kwargs.update(encoder_inputs)

2191
2192
2193
2194
2195
2196
2197
2198
        return (
            num_scheduled_tokens,
            input_ids,
            inputs_embeds,
            positions,
            intermediate_tensors,
            model_kwargs,
        )
2199

2200
    def _sample(
2201
2202
2203
        self,
        logits: Optional[torch.Tensor],
        spec_decode_metadata: Optional[SpecDecodeMetadata],
2204
    ) -> SamplerOutput:
2205
        # Sample the next token and get logprobs if needed.
2206
        sampling_metadata = self.input_batch.sampling_metadata
2207
        if spec_decode_metadata is None:
2208
            return self.sampler(
2209
2210
2211
                logits=logits,
                sampling_metadata=sampling_metadata,
            )
2212

2213
2214
2215
2216
2217
2218
2219
2220
2221
2222
2223
2224
2225
2226
2227
2228
2229
2230
2231
2232
2233
2234
2235
2236
2237
2238
        # 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.
        assert logits is not None
        bonus_logits = logits[spec_decode_metadata.bonus_logits_indices]
        sampler_output = self.sampler(
            logits=bonus_logits,
            sampling_metadata=sampling_metadata,
            predict_bonus_token=True,
        )
        bonus_token_ids = sampler_output.sampled_token_ids

        # 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.
        target_logits = logits[spec_decode_metadata.target_logits_indices]
        output_token_ids = self.rejection_sampler(
            spec_decode_metadata,
            None,  # draft_probs
            target_logits,
            bonus_token_ids,
            sampling_metadata,
        )
        sampler_output.sampled_token_ids = output_token_ids
        self._update_states_after_model_execute(output_token_ids)
2239
2240
2241
        return sampler_output

    def _bookkeeping_sync(
2242
2243
2244
2245
2246
2247
        self,
        scheduler_output: "SchedulerOutput",
        sampler_output: SamplerOutput,
        logits: Optional[torch.Tensor],
        hidden_states: torch.Tensor,
        num_scheduled_tokens: int,
2248
    ) -> tuple[
2249
2250
2251
2252
2253
2254
2255
        dict[str, int],
        Optional[LogprobsLists],
        list[list[int]],
        dict[str, Optional[LogprobsTensors]],
        list[str],
        dict[str, int],
        list[int],
2256
    ]:
2257
2258
2259
2260
        num_nans_in_logits = {}
        if envs.VLLM_COMPUTE_NANS_IN_LOGITS:
            num_nans_in_logits = self._get_nans_in_logits(logits)

2261
2262
2263
        discard_sampled_tokens_req_indices = self.discard_request_indices.np[
            : self.num_discarded_requests
        ]
2264
2265
2266
2267
        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)
2268

2269
2270
2271
        # 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()
2272
        req_id_to_index_output_copy = self.input_batch.req_id_to_index.copy()
2273

2274
2275
        # NOTE: GPU -> CPU Sync happens here.
        # Move as many CPU operations as possible before this sync point.
2276
        logprobs_tensors = sampler_output.logprobs_tensors
2277
2278
2279
        logprobs_lists = (
            logprobs_tensors.tolists() if logprobs_tensors is not None else None
        )
2280
2281
2282

        # Compute prompt logprobs if needed.
        prompt_logprobs_dict = self._get_prompt_logprobs_dict(
2283
            hidden_states[:num_scheduled_tokens],
2284
            scheduler_output.num_scheduled_tokens,
2285
2286
        )

2287
        num_sampled_tokens = sampler_output.sampled_token_ids.shape[0]
2288
        sampled_token_ids = sampler_output.sampled_token_ids
2289
        invalid_req_indices = []
2290
2291
2292
2293
2294
2295
2296
2297
2298
2299
2300
2301
2302
2303
        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:
2304
                valid_sampled_token_ids[int(i)].clear()
2305
        else:
2306
            valid_sampled_token_ids = []
2307
            invalid_req_indices = discard_sampled_tokens_req_indices.tolist()
2308
2309
2310
2311
2312
2313
            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.
2314
            self.input_batch.prev_sampled_token_ids = sampled_token_ids
2315
2316
2317
2318
2319
            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
            }
2320

2321
2322
2323
2324
2325
        # 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.
2326
        req_ids = self.input_batch.req_ids
2327
2328
        for req_idx in range(num_sampled_tokens):
            if self.use_async_scheduling:
2329
                sampled_ids = [-1] if req_idx not in invalid_req_indices_set else None
2330
2331
            else:
                sampled_ids = valid_sampled_token_ids[req_idx]
2332
2333
2334
2335
2336
            if not sampled_ids:
                continue

            start_idx = self.input_batch.num_tokens_no_spec[req_idx]
            end_idx = start_idx + len(sampled_ids)
2337
2338
2339
2340
            assert end_idx <= self.max_model_len, (
                "Sampled token IDs exceed the max model length. "
                f"Total number of tokens: {end_idx} > max_model_len: "
                f"{self.max_model_len}"
2341
            )
2342

2343
2344
            self.input_batch.token_ids_cpu[req_idx, start_idx:end_idx] = sampled_ids
            self.input_batch.is_token_ids[req_idx, start_idx:end_idx] = True
2345
2346
            self.input_batch.num_tokens_no_spec[req_idx] = end_idx
            self.input_batch.num_tokens[req_idx] = end_idx
2347

2348
            req_id = req_ids[req_idx]
2349
2350
2351
            req_state = self.requests[req_id]
            req_state.output_token_ids.extend(sampled_ids)

2352
2353
2354
2355
2356
2357
2358
2359
2360
2361
        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,
        )

2362
2363
2364
2365
2366
2367
2368
2369
2370
2371
2372
2373
2374
2375
2376
    @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()

2377
2378
2379
2380
2381
2382
2383
2384
2385
2386
2387
    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.
2388
        Motivation: We can inspect only this method versus
2389
2390
2391
2392
2393
2394
2395
2396
2397
2398
2399
2400
2401
2402
2403
2404
2405
2406
2407
2408
        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,
        )

2409
2410
2411
2412
2413
2414
2415
    @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"):
2416
2417
2418
2419
2420
2421
2422
2423
2424
            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(
2425
2426
                        scheduler_output, self.vllm_config
                    )
2427
2428
2429
2430
                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 "
2431
2432
                        "it when the requests need prompt logprobs"
                    )
2433

2434
                # Prepare the decoder inputs.
2435
2436
2437
2438
2439
2440
2441
2442
                (
                    attn_metadata,
                    logits_indices,
                    spec_decode_metadata,
                    num_scheduled_tokens_np,
                    spec_decode_common_attn_metadata,
                    max_query_len,
                    ubatch_slices,
2443
                    num_tokens_across_dp,
2444
2445
                    use_cascade_attn,
                ) = self._prepare_inputs(scheduler_output)
2446

2447
            dp_rank = self.parallel_config.data_parallel_rank
2448
2449
            if ubatch_slices:
                assert num_tokens_across_dp is not None
2450
                num_input_tokens = int(num_tokens_across_dp[dp_rank].item())
2451
2452
                self.pad_out_ubatch_slice(ubatch_slices, num_input_tokens)
            elif num_tokens_across_dp is not None:
2453
                num_input_tokens = int(num_tokens_across_dp[dp_rank].item())
2454
2455
2456
2457
2458
            else:
                num_input_tokens = self._get_num_input_tokens(
                    scheduler_output.total_num_scheduled_tokens
                )

2459
2460
2461
2462
2463
2464
2465
            (
                num_scheduled_tokens,
                input_ids,
                inputs_embeds,
                positions,
                intermediate_tensors,
                model_kwargs,
2466
            ) = self._preprocess(
2467
                scheduler_output, num_input_tokens, intermediate_tensors
2468
2469
2470
2471
2472
2473
2474
2475
2476
2477
2478
            )

            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)
            )
2479

2480
2481
        # Set cudagraph mode to none if calc_kv_scales is true.
        if attn_metadata is not None:
2482
2483
2484
2485
2486
            metadata_list = (
                attn_metadata.values()
                if isinstance(attn_metadata, dict)
                else [attn_metadata]
            )
2487
            if any(
2488
2489
                getattr(m, "enable_kv_scales_calculation", False) for m in metadata_list
            ):
2490
2491
                cudagraph_runtime_mode = CUDAGraphMode.NONE

2492
2493
        # Run the model.
        # Use persistent buffers for CUDA graphs.
2494
2495
        with (
            set_forward_context(
2496
2497
2498
2499
2500
2501
                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,
2502
                ubatch_slices=ubatch_slices,
2503
2504
2505
2506
            ),
            record_function_or_nullcontext("Forward"),
            self.maybe_get_kv_connector_output(scheduler_output) as kv_connector_output,
        ):
2507
            model_output = self._model_forward(
2508
2509
2510
2511
2512
2513
2514
2515
2516
                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:
2517
                # True when EAGLE 3 is used.
2518
2519
                hidden_states, aux_hidden_states = model_output
            else:
2520
                # Common case.
2521
2522
2523
                hidden_states = model_output
                aux_hidden_states = None

2524
2525
2526
2527
2528
            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)
2529
2530
                    hidden_states.kv_connector_output = kv_connector_output
                    return hidden_states
2531

2532
                if self.is_pooling_model:
2533
                    # Return the pooling output.
2534
2535
2536
                    output = self._pool(
                        hidden_states, num_scheduled_tokens, num_scheduled_tokens_np
                    )
2537
2538
                    output.kv_connector_output = kv_connector_output
                    return output
2539
2540

                sample_hidden_states = hidden_states[logits_indices]
2541
                logits = self.model.compute_logits(sample_hidden_states)
2542
2543
2544
2545
2546
            else:
                # Rare case.
                assert not self.is_pooling_model

                if not get_pp_group().is_last_rank:
2547
                    all_gather_tensors = {
2548
2549
2550
                        "residual": not is_residual_scattered_for_sp(
                            self.vllm_config, num_input_tokens
                        )
2551
                    }
2552
                    get_pp_group().send_tensor_dict(
2553
2554
                        hidden_states.tensors,
                        all_gather_group=get_tp_group(),
2555
2556
                        all_gather_tensors=all_gather_tensors,
                    )
2557
2558
2559
                    logits = None
                else:
                    sample_hidden_states = hidden_states[logits_indices]
2560
                    logits = self.model.compute_logits(sample_hidden_states)
2561
2562
2563
2564
2565

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

2566
2567
2568
                model_output_broadcast_data = get_pp_group().broadcast_tensor_dict(
                    model_output_broadcast_data, src=len(get_pp_group().ranks) - 1
                )
2569
2570
2571
2572
2573
                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:
2574
2575
2576
                apply_grammar_bitmask(
                    scheduler_output, self.input_batch, logits, self.device
                )
2577
2578
2579
2580

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

2581
2582
2583
2584
2585
2586
2587
2588
2589
2590
2591
2592
2593
2594
        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,
                )

2595
2596
2597
2598
2599
        use_padded_batch_for_eagle = (
            self.speculative_config
            and self.speculative_config.use_eagle()
            and not self.speculative_config.disable_padded_drafter_batch
        )
2600
2601
2602
        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
2603
2604
2605
2606
2607
        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
        ):
2608
            effective_drafter_max_model_len = (
2609
2610
                self.speculative_config.draft_model_config.max_model_len
            )
2611
        input_fits_in_drafter = spec_decode_common_attn_metadata and (
2612
2613
2614
2615
            spec_decode_common_attn_metadata.max_seq_len
            + self.speculative_config.num_speculative_tokens
            <= effective_drafter_max_model_len
        )
2616
        if use_padded_batch_for_eagle and input_fits_in_drafter:
2617
2618
2619
2620
            # 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)

2621
2622
2623
2624
2625
2626
2627
2628
2629
        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,
2630
2631
2632
2633
2634
2635
2636
            ) = self._bookkeeping_sync(
                scheduler_output,
                sampler_output,
                logits,
                hidden_states,
                num_scheduled_tokens,
            )
2637

2638
2639
2640
2641
2642
        if (
            self.speculative_config
            and not use_padded_batch_for_eagle
            and input_fits_in_drafter
        ):
2643
2644
2645
            # 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)
2646

2647
2648
        with record_function_or_nullcontext("EPLB"):
            self.eplb_step()
2649

2650
2651
2652
        output = ModelRunnerOutput(
            req_ids=req_ids_output_copy,
            req_id_to_index=req_id_to_index_output_copy,
2653
2654
2655
2656
            sampled_token_ids=valid_sampled_token_ids,
            logprobs=logprobs_lists,
            prompt_logprobs_dict=prompt_logprobs_dict,
            pooler_output=[],
2657
            kv_connector_output=kv_connector_output,
2658
2659
2660
            num_nans_in_logits=num_nans_in_logits,
        )

2661
2662
2663
2664
2665
        if not self.use_async_scheduling:
            return output

        return AsyncGPUModelRunnerOutput(
            model_runner_output=output,
2666
            sampled_token_ids=sampler_output.sampled_token_ids,
2667
2668
2669
2670
            invalid_req_indices=invalid_req_indices,
            async_output_copy_stream=self.async_output_copy_stream,
        )

2671
2672
2673
2674
2675
2676
2677
2678
2679
2680
2681
    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)

2682
2683
2684
    def propose_draft_token_ids(
        self,
        scheduler_output: "SchedulerOutput",
2685
        sampled_token_ids: Union[torch.Tensor, list[list[int]]],
2686
2687
2688
        sampling_metadata: SamplingMetadata,
        hidden_states: torch.Tensor,
        sample_hidden_states: torch.Tensor,
Wentao Ye's avatar
Wentao Ye committed
2689
        aux_hidden_states: Optional[list[torch.Tensor]],
2690
        spec_decode_metadata: Optional[SpecDecodeMetadata],
2691
        common_attn_metadata: CommonAttentionMetadata,
2692
    ) -> Union[list[list[int]], torch.Tensor]:
2693
2694
        num_scheduled_tokens = scheduler_output.total_num_scheduled_tokens
        if self.speculative_config.method == "ngram":
2695
            assert isinstance(sampled_token_ids, list)
2696
            assert isinstance(self.drafter, NgramProposer)
2697
            draft_token_ids = self.drafter.propose(
2698
2699
                sampled_token_ids,
                self.input_batch.req_ids,
2700
2701
                self.input_batch.num_tokens_no_spec,
                self.input_batch.token_ids_cpu,
2702
2703
                self.input_batch.spec_decode_unsupported_reqs,
            )
2704
        elif self.speculative_config.method == "medusa":
2705
            assert isinstance(sampled_token_ids, list)
2706
            assert isinstance(self.drafter, MedusaProposer)
2707

2708
2709
            if sample_hidden_states.shape[0] == len(sampled_token_ids):
                # The input to the target model does not include draft tokens.
2710
2711
2712
2713
                hidden_states = sample_hidden_states
            else:
                indices = []
                offset = 0
Wentao Ye's avatar
Wentao Ye committed
2714
                assert spec_decode_metadata is not None
2715
                for num_draft, tokens in zip(
2716
2717
                    spec_decode_metadata.num_draft_tokens, sampled_token_ids
                ):
2718
2719
                    indices.append(offset + len(tokens) - 1)
                    offset += num_draft + 1
2720
                indices = torch.tensor(indices, device=self.device)
2721
2722
                hidden_states = sample_hidden_states[indices]

2723
            draft_token_ids = self.drafter.propose(
2724
2725
2726
                target_hidden_states=hidden_states,
                sampling_metadata=sampling_metadata,
            )
2727
        elif self.speculative_config.use_eagle():
2728
            assert isinstance(self.drafter, EagleProposer)
2729
2730
2731
2732
2733

            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.
2734
2735
                assert isinstance(sampled_token_ids, list), (
                    "sampled_token_ids should be a python list when"
2736
                    "padded-batch is disabled."
2737
                )
2738
                next_token_ids = self.drafter.prepare_next_token_ids_cpu(
2739
2740
2741
2742
2743
                    sampled_token_ids,
                    self.requests,
                    self.input_batch,
                    scheduler_output.num_scheduled_tokens,
                )
2744
2745
2746
2747
2748
            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.
2749
2750
                assert isinstance(sampled_token_ids, torch.Tensor), (
                    "sampled_token_ids should be a torch.Tensor when"
2751
                    "padded-batch is enabled."
2752
2753
                )
                next_token_ids, valid_sampled_tokens_count = (
2754
2755
2756
2757
2758
2759
                    self.drafter.prepare_next_token_ids_padded(
                        common_attn_metadata,
                        sampled_token_ids,
                        self.requests,
                        self.input_batch,
                        self.discard_request_indices.gpu,
2760
                        self.num_discarded_requests,
2761
                    )
2762
                )
Jiayi Yao's avatar
Jiayi Yao committed
2763

2764
            if spec_decode_metadata is None:
2765
                token_indices_to_sample = None
2766
                # input_ids can be None for multimodal models.
2767
                target_token_ids = self.input_ids.gpu[:num_scheduled_tokens]
2768
                target_positions = self._get_positions(num_scheduled_tokens)
2769
                if self.use_aux_hidden_state_outputs:
Wentao Ye's avatar
Wentao Ye committed
2770
                    assert aux_hidden_states is not None
2771
                    target_hidden_states = torch.cat(
2772
2773
                        [h[:num_scheduled_tokens] for h in aux_hidden_states], dim=-1
                    )
2774
2775
                else:
                    target_hidden_states = hidden_states[:num_scheduled_tokens]
2776
            else:
2777
2778
                if self.speculative_config.disable_padded_drafter_batch:
                    token_indices_to_sample = None
2779
2780
2781
2782
2783
                    common_attn_metadata, token_indices = self.drafter.prepare_inputs(
                        common_attn_metadata,
                        sampled_token_ids,
                        spec_decode_metadata.num_draft_tokens,
                    )
2784
                else:
2785
                    common_attn_metadata, token_indices, token_indices_to_sample = (
2786
2787
2788
                        self.drafter.prepare_inputs_padded(
                            common_attn_metadata,
                            spec_decode_metadata,
2789
2790
2791
                            valid_sampled_tokens_count,
                        )
                    )
2792

2793
                target_token_ids = self.input_ids.gpu[token_indices]
2794
                target_positions = self._get_positions(token_indices)
2795
                if self.use_aux_hidden_state_outputs:
Wentao Ye's avatar
Wentao Ye committed
2796
                    assert aux_hidden_states is not None
2797
                    target_hidden_states = torch.cat(
2798
2799
                        [h[token_indices] for h in aux_hidden_states], dim=-1
                    )
2800
2801
                else:
                    target_hidden_states = hidden_states[token_indices]
2802

2803
            if self.supports_mm_inputs:
2804
2805
2806
2807
2808
2809
                mm_embed_inputs = self._gather_mm_embeddings(
                    scheduler_output,
                    shift_computed_tokens=1,
                )
            else:
                mm_embed_inputs = None
2810

2811
            draft_token_ids = self.drafter.propose(
2812
2813
2814
2815
                target_token_ids=target_token_ids,
                target_positions=target_positions,
                target_hidden_states=target_hidden_states,
                next_token_ids=next_token_ids,
2816
                last_token_indices=token_indices_to_sample,
2817
                sampling_metadata=sampling_metadata,
2818
                common_attn_metadata=common_attn_metadata,
2819
                mm_embed_inputs=mm_embed_inputs,
2820
            )
2821

2822
        return draft_token_ids
2823

2824
2825
2826
    def update_config(self, overrides: dict[str, Any]) -> None:
        allowed_config_names = {"load_config", "model_config"}
        for config_name, config_overrides in overrides.items():
2827
2828
            assert config_name in allowed_config_names, (
                f"Config `{config_name}` not supported. "
2829
                f"Allowed configs: {allowed_config_names}"
2830
            )
2831
2832
2833
2834
            config = getattr(self, config_name)
            new_config = update_config(config, config_overrides)
            setattr(self, config_name, new_config)

2835
2836
2837
2838
2839
    def load_model(self, eep_scale_up: bool = False) -> None:
        """
        Args:
            eep_scale_up: the model loading is for elastic EP scale up.
        """
2840
        logger.info("Starting to load model %s...", self.model_config.model)
2841
2842
        if eep_scale_up:
            from vllm.distributed.parallel_state import get_ep_group
2843
2844
2845
2846
2847

            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
            )
2848
2849
            num_local_physical_experts = int(num_local_physical_experts.item())
            new_ep_size = get_ep_group().world_size
2850
            global_expert_load, old_global_expert_indices = EplbState.recv_state()
2851
            num_logical_experts = global_expert_load.shape[1]
2852
            self.parallel_config.eplb_config.num_redundant_experts = (
2853
2854
2855
2856
2857
2858
                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
            )
2859
            rank_mapping = {
2860
                old_ep_rank: old_ep_rank for old_ep_rank in range(old_ep_size)
2861
2862
2863
2864
2865
2866
            }
        else:
            global_expert_load = None
            old_global_expert_indices = None
            rank_mapping = None

2867
        with DeviceMemoryProfiler() as m:
2868
            time_before_load = time.perf_counter()
2869
            model_loader = get_model_loader(self.load_config)
2870
2871
            logger.info("Loading model from scratch...")
            self.model = model_loader.load_model(
2872
2873
                vllm_config=self.vllm_config, model_config=self.model_config
            )
2874
            if self.lora_config:
2875
2876
2877
                self.model = self.load_lora_model(
                    self.model, self.vllm_config, self.device
                )
2878
2879
2880
            if hasattr(self, "drafter"):
                logger.info("Loading drafter model...")
                self.drafter.load_model(self.model)
2881
            if self.use_aux_hidden_state_outputs:
2882
                if not supports_eagle3(self.model):
2883
2884
                    raise RuntimeError(
                        "Model does not support EAGLE3 interface but "
2885
2886
                        "aux_hidden_state_outputs was requested"
                    )
2887
2888
2889
2890
2891
2892
2893
2894
2895
2896
2897
2898
2899

                # 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)
2900
            time_after_load = time.perf_counter()
2901
        self.model_memory_usage = m.consumed_memory
2902
2903
2904
2905
2906
        logger.info(
            "Model loading took %.4f GiB and %.6f seconds",
            self.model_memory_usage / GiB_bytes,
            time_after_load - time_before_load,
        )
2907
        prepare_communication_buffer_for_model(self.model)
2908

2909
2910
2911
2912
        self.is_multimodal_pruning_enabled = (
            supports_multimodal_pruning(self.model)
            and self.model_config.multimodal_config.is_multimodal_pruning_enabled()
        )
2913

2914
2915
        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)
2916
2917
2918
2919
            self.eplb_state = EplbState.build(
                self.model,
                self.device,
                self.parallel_config,
2920
2921
2922
                global_expert_load,
                old_global_expert_indices,
                rank_mapping,
2923
2924
            )

2925
        if (
2926
2927
            self.vllm_config.compilation_config.level == CompilationLevel.DYNAMO_AS_IS
            and supports_dynamo()
2928
        ):
2929
            backend = self.vllm_config.compilation_config.init_backend(self.vllm_config)
2930
            compilation_counter.dynamo_as_is_count += 1
2931
            self.model.compile(fullgraph=True, backend=backend)
2932
2933
2934
2935
2936
            return
        # for other compilation levels, cudagraph behavior is controlled by
        # CudagraphWraper and CudagraphDispatcher of vllm.

        # wrap the model with full cudagraph wrapper if needed.
2937
2938
2939
2940
2941
2942
2943
        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
            )
2944
2945
        elif self.parallel_config.enable_dbo:
            if self.compilation_config.cudagraph_mode.has_full_cudagraphs():
2946
2947
2948
                self.model = UBatchWrapper(
                    self.model, self.vllm_config, CUDAGraphMode.FULL, self.device
                )
2949
            else:
2950
2951
2952
                self.model = UBatchWrapper(
                    self.model, self.vllm_config, CUDAGraphMode.NONE, self.device
                )
2953

2954
2955
2956
2957
2958
2959
2960
2961
2962
2963
2964
2965
2966
2967
2968
2969
2970
2971
2972
2973
2974
2975
2976
2977
    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

2978
    def reload_weights(self) -> None:
2979
        assert getattr(self, "model", None) is not None, (
2980
            "Cannot reload weights before model is loaded."
2981
        )
2982
2983
        model_loader = get_model_loader(self.load_config)
        logger.info("Reloading weights inplace...")
2984
        model_loader.load_weights(self.get_model(), model_config=self.model_config)
2985

2986
2987
2988
2989
2990
    def save_tensorized_model(
        self,
        tensorizer_config: "TensorizerConfig",
    ) -> None:
        TensorizerLoader.save_model(
2991
            self.get_model(),
2992
            tensorizer_config=tensorizer_config,
2993
            model_config=self.model_config,
2994
2995
        )

2996
2997
2998
    def _get_prompt_logprobs_dict(
        self,
        hidden_states: torch.Tensor,
2999
        num_scheduled_tokens: dict[str, int],
3000
    ) -> dict[str, Optional[LogprobsTensors]]:
3001
3002
3003
3004
        num_prompt_logprobs_dict = self.input_batch.num_prompt_logprobs
        if not num_prompt_logprobs_dict:
            return {}

3005
        in_progress_dict = self.input_batch.in_progress_prompt_logprobs_cpu
3006
        prompt_logprobs_dict: dict[str, Optional[LogprobsTensors]] = {}
3007
3008
3009
3010
3011

        # 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():
3012
            num_tokens = num_scheduled_tokens[req_id]
3013
3014
3015

            # Get metadata for this request.
            request = self.requests[req_id]
3016
3017
3018
3019
            if request.prompt_token_ids is None:
                # Prompt logprobs is incompatible with prompt embeddings
                continue

3020
3021
            num_prompt_tokens = len(request.prompt_token_ids)
            prompt_token_ids = torch.tensor(request.prompt_token_ids).to(
3022
3023
                self.device, non_blocking=True
            )
3024

3025
3026
3027
3028
3029
3030
            # 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(
3031
3032
                    num_prompt_tokens - 1, num_prompt_logprobs + 1
                )
3033
3034
                in_progress_dict[req_id] = logprobs_tensors

3035
            # Determine number of logits to retrieve.
3036
3037
            start_idx = request.num_computed_tokens
            start_tok = start_idx + 1
3038
            num_remaining_tokens = num_prompt_tokens - start_tok
3039
            if num_tokens <= num_remaining_tokens:
3040
                # This is a chunk, more tokens remain.
3041
3042
3043
                # 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.
3044
3045
3046
3047
3048
                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)
3049
3050
3051
3052
3053
3054
3055
                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
3056
3057
3058
3059
3060

            # 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]
3061
            offset = self.query_start_loc.np[req_idx].item()
3062
            prompt_hidden_states = hidden_states[offset : offset + num_logits]
3063
            logits = self.model.compute_logits(prompt_hidden_states)
3064
3065
3066
3067

            # 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.
3068
            tgt_token_ids = prompt_token_ids[start_tok : start_tok + num_logits]
3069
3070

            # Compute prompt logprobs.
3071
3072
            logprobs = self.sampler.compute_logprobs(logits)
            token_ids, logprobs, ranks = self.sampler.gather_logprobs(
3073
3074
                logprobs, num_prompt_logprobs, tgt_token_ids
            )
3075
3076

            # Transfer GPU->CPU async.
3077
3078
            chunk_slice = slice(start_idx, start_idx + num_logits)
            logprobs_tensors.logprob_token_ids[chunk_slice].copy_(
3079
3080
3081
                token_ids, non_blocking=True
            )
            logprobs_tensors.logprobs[chunk_slice].copy_(logprobs, non_blocking=True)
3082
            logprobs_tensors.selected_token_ranks[chunk_slice].copy_(
3083
3084
                ranks, non_blocking=True
            )
3085
3086
3087
3088
3089

        # 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]
3090
            del in_progress_dict[req_id]
3091
3092

        # Must synchronize the non-blocking GPU->CPU transfers.
3093
        if prompt_logprobs_dict:
3094
            self._sync_device()
3095
3096
3097

        return prompt_logprobs_dict

3098
3099
3100
3101
3102
3103
3104
3105
3106
3107
3108
3109
3110
3111
    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])
3112
3113
3114
                    if num_nans_for_index is not None and req_index < logits.shape[0]
                    else 0
                )
3115
3116
3117
3118
            return num_nans_in_logits
        except IndexError:
            return {}

3119
3120
3121
3122
3123
3124
    @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
3125
         - during DP rank dummy run
3126
3127
3128
3129
3130
3131
3132
3133
3134
3135
3136
        """
        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(
3137
                    self.input_ids.gpu,
3138
3139
                    low=0,
                    high=self.model_config.get_vocab_size(),
3140
3141
                    dtype=input_ids.dtype,
                )
3142

3143
            logger.debug_once("Randomizing dummy data for DP Rank")
3144
            input_ids.copy_(rand_input_ids()[: input_ids.size(0)], non_blocking=True)
3145
3146
3147
            yield
            input_ids.fill_(0)

3148
3149
3150
3151
3152
3153
    def _get_mm_dummy_batch(
        self,
        modality: str,
        max_items_per_batch: int,
    ) -> BatchedTensorInputs:
        """Dummy data for profiling and precompiling multimodal models."""
3154
3155
        assert self.mm_budget is not None

3156
3157
        dummy_decoder_data = self.mm_registry.get_decoder_dummy_data(
            model_config=self.model_config,
3158
            seq_len=self.max_model_len,
3159
            mm_counts={modality: 1},
3160
            cache=self.mm_budget.cache,
3161
3162
3163
3164
        )
        dummy_mm_data = dummy_decoder_data.multi_modal_data

        # Result in the maximum GPU consumption of the model
3165
3166
        dummy_mm_item = dummy_mm_data[modality][0]
        dummy_mm_items = [dummy_mm_item] * max_items_per_batch
3167

3168
        model = cast(SupportsMultiModal, self.model)
3169
3170
3171
3172
3173
3174
3175
3176
3177
        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,
            )
        )
3178

3179
3180
3181
3182
    @torch.inference_mode()
    def _dummy_run(
        self,
        num_tokens: int,
3183
        cudagraph_runtime_mode: Optional[CUDAGraphMode] = None,
3184
3185
        force_attention: bool = False,
        uniform_decode: bool = False,
3186
        allow_microbatching: bool = True,
3187
3188
        skip_eplb: bool = False,
        is_profile: bool = False,
3189
        create_mixed_batch: bool = False,
3190
        remove_lora: bool = True,
3191
    ) -> tuple[torch.Tensor, torch.Tensor]:
3192
3193
3194
3195
3196
3197
3198
        """
        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.
3199
                - if not set will determine the cudagraph mode based on using
3200
                    the self.cudagraph_dispatcher.
3201
3202
3203
3204
                - CUDAGraphMode.NONE: No cudagraph, for warm up and profile run
                - CUDAGraphMode.PIECEWISE: Piecewise cudagraph.
                - CUDAGraphMode.FULL: Full cudagraph, attention metadata is
                    needed.
3205
            force_attention: If True, always create attention metadata. Used to
3206
3207
3208
3209
                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.
3210
3211
            create_mixed_batch: If True, create a mixed batch with both decode
                (1 token) and prefill (multiple tokens) requests.
3212
            remove_lora: If False, dummy LoRAs are not destroyed after the run
3213
        """
3214
3215
3216
3217
        assert (
            cudagraph_runtime_mode is None
            or cudagraph_runtime_mode.valid_runtime_modes()
        )
3218

3219
        # If cudagraph_mode.decode_mode() == FULL and
3220
        # cudagraph_mode.separate_routine(). This means that we are using
3221
3222
3223
3224
3225
3226
3227
3228
3229
3230
3231
        # 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.
3232
        max_query_len = self.uniform_decode_query_len if uniform_decode else num_tokens
3233

3234
3235
3236
3237
3238
        # 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
3239
3240
3241
3242
        if create_mixed_batch:
            assert not uniform_decode
            # Create mixed batch:
            # first half decode tokens, second half one prefill
3243
            num_decode_tokens = min(max_num_reqs - 1, num_tokens // 2)
3244
3245
3246
3247
            num_prefill_tokens = num_tokens - num_decode_tokens
            num_reqs = num_decode_tokens + 1

            # Create decode requests (1 token each) followed by prefill request
3248
            num_scheduled_tokens_list = [1] * num_decode_tokens + [num_prefill_tokens]
3249
3250
3251
            # Note: Overriding max_query_len to be the prefill tokens
            max_query_len = num_prefill_tokens
        elif uniform_decode:
3252
            assert not create_mixed_batch
3253
            num_reqs = min(max_num_reqs, cdiv(num_tokens, max_query_len))
3254
3255
            num_scheduled_tokens_list = [max_query_len] * num_reqs
            if num_tokens % max_query_len != 0:
3256
                num_scheduled_tokens_list[-1] = num_tokens % max_query_len
3257
3258
3259
3260
3261
3262
        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

3263
3264
        assert sum(num_scheduled_tokens_list) == num_tokens
        assert len(num_scheduled_tokens_list) == num_reqs
3265
        num_scheduled_tokens = np.array(num_scheduled_tokens_list, dtype=np.int32)
3266
        total_num_scheduled_tokens = int(num_scheduled_tokens.sum())
3267

3268
3269
3270
        # Disable DP padding when running eager
        allow_dp_padding = self.compilation_config.cudagraph_mode != CUDAGraphMode.NONE

3271
3272
        # We currently only microbatch if the number of tokens is
        # over a certain threshold.
3273
        ubatch_slices, num_tokens_across_dp = coordinate_batch_across_dp(
3274
3275
3276
3277
3278
3279
3280
            num_tokens_unpadded=total_num_scheduled_tokens,
            parallel_config=self.vllm_config.parallel_config,
            allow_microbatching=allow_microbatching,
            allow_dp_padding=allow_dp_padding,
            num_tokens_padded=total_num_scheduled_tokens,
            uniform_decode=uniform_decode,
            num_scheduled_tokens_per_request=num_scheduled_tokens,
3281
3282
3283
        )
        num_tokens_after_padding = num_tokens
        if num_tokens_across_dp is not None:
3284
3285
            dp_rank = self.parallel_config.data_parallel_rank
            num_tokens_after_padding = int(num_tokens_across_dp[dp_rank])
3286
3287

        attn_metadata: Optional[PerLayerAttnMetadata] = None
3288
3289
3290

        # If force_attention is True, we always capture attention. Otherwise,
        # it only happens for cudagraph_runtime_mode=FULL.
3291
        if force_attention or cudagraph_runtime_mode == CUDAGraphMode.FULL:
3292
            attn_metadata = {}
3293
3294
            if ubatch_slices is not None:
                attn_metadata = [dict() for _ in range(len(ubatch_slices))]
3295

3296
3297
3298
3299
3300
3301
            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:
3302
                seq_lens = max_query_len
3303
            self.seq_lens.np[:num_reqs] = seq_lens
3304
3305
            self.seq_lens.np[num_reqs:] = 0
            self.seq_lens.copy_to_gpu()
3306

3307
3308
            cum_num_tokens, _ = self._get_cumsum_and_arange(num_scheduled_tokens)
            self.query_start_loc.np[1 : num_reqs + 1] = cum_num_tokens
3309
3310
            self.query_start_loc.copy_to_gpu()

3311
            for kv_cache_group_id, kv_cache_group_spec in enumerate(
3312
3313
                self.kv_cache_config.kv_cache_groups
            ):
3314
                common_attn_metadata = CommonAttentionMetadata(
3315
3316
                    query_start_loc=self.query_start_loc.gpu[: num_reqs + 1],
                    query_start_loc_cpu=self.query_start_loc.cpu[: num_reqs + 1],
3317
3318
                    seq_lens=self.seq_lens.gpu[:num_reqs],
                    seq_lens_cpu=self.seq_lens.cpu[:num_reqs],
3319
3320
3321
                    num_computed_tokens_cpu=self.input_batch.num_computed_tokens_cpu_tensor[
                        :num_reqs
                    ],
3322
3323
                    num_reqs=num_reqs,
                    num_actual_tokens=num_tokens,
3324
                    max_query_len=max_query_len,
3325
                    max_seq_len=self.max_model_len,
3326
3327
3328
                    block_table_tensor=self.input_batch.block_table[
                        kv_cache_group_id
                    ].get_device_tensor(num_reqs),
3329
                    slot_mapping=self.input_batch.block_table[
3330
3331
3332
                        kv_cache_group_id
                    ].slot_mapping.gpu[:num_tokens],
                    causal=True,
3333
3334
3335
                    dcp_local_seq_lens=self.dcp_local_seq_lens.gpu[:num_reqs]
                    if self.dcp_world_size > 1
                    else None,
3336
                )
3337
                for attn_group in self.attn_groups[kv_cache_group_id]:
3338
3339
                    if ubatch_slices is not None:
                        common_attn_metadata_list = split_attn_metadata(
3340
3341
                            ubatch_slices, common_attn_metadata
                        )
3342
                        for ubid, common_attn_metadata in enumerate(
3343
3344
                            common_attn_metadata_list
                        ):
3345
                            assert common_attn_metadata.max_query_len == 1
3346
3347
3348
                            attn_metadata_i = attn_group.get_metadata_builder(
                                ubatch_id=ubid
                            ).build_for_cudagraph_capture(common_attn_metadata)
3349
                            for layer_name in attn_group.layer_names:
3350
                                assert type(attn_metadata) is list
3351
                                attn_metadata[ubid][layer_name] = attn_metadata_i
3352
3353
                    else:
                        assert type(attn_metadata) is dict
3354
3355
                        metadata_builder = attn_group.get_metadata_builder()
                        attn_metadata_i = metadata_builder.build_for_cudagraph_capture(
3356
3357
                            common_attn_metadata
                        )
3358
                        for layer_name in attn_group.layer_names:
3359
                            attn_metadata[layer_name] = attn_metadata_i
3360

3361
3362
3363
        with self.maybe_dummy_run_with_lora(
            self.lora_config, num_scheduled_tokens, remove_lora
        ):
3364
3365
3366
            # 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)
3367
            if self.supports_mm_inputs and not self.model_config.is_encoder_decoder:
3368
                input_ids = None
3369
                inputs_embeds = self.inputs_embeds.gpu[:num_tokens_after_padding]
3370
                model_kwargs = {
3371
                    **model_kwargs,
3372
3373
                    **self._dummy_mm_kwargs(num_reqs),
                }
3374
3375
            elif self.enable_prompt_embeds:
                input_ids = None
3376
3377
                inputs_embeds = self.inputs_embeds.gpu[:num_tokens_after_padding]
                model_kwargs = self._init_model_kwargs(num_tokens_after_padding)
3378
            else:
3379
                input_ids = self.input_ids.gpu[:num_tokens_after_padding]
3380
                inputs_embeds = None
3381

3382
            if self.uses_mrope:
3383
                positions = self.mrope_positions.gpu[:, :num_tokens_after_padding]
3384
            else:
3385
                positions = self.positions.gpu[:num_tokens_after_padding]
3386
3387
3388
3389
3390
3391
3392
3393
3394

            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,
3395
3396
3397
                            device=self.device,
                        )
                    )
3398
3399

                intermediate_tensors = self.sync_and_slice_intermediate_tensors(
3400
                    num_tokens_after_padding, None, False
3401
                )
3402
3403

            # filter out the valid batch descriptor
3404
3405
3406
3407
3408
3409
3410
3411
3412
3413
            _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)
            )
3414
3415
3416
            if cudagraph_runtime_mode is not None:
                # we allow forcing NONE when the dispatcher disagrees to support
                # warm ups for cudagraph capture
3417
3418
3419
3420
                assert (
                    cudagraph_runtime_mode == CUDAGraphMode.NONE
                    or cudagraph_runtime_mode == _cg_mode
                ), (
3421
                    f"Cudagraph runtime mode mismatch at dummy_run. "
3422
3423
                    f"Expected {_cg_mode}, but got {cudagraph_runtime_mode}."
                )
3424
3425
            else:
                cudagraph_runtime_mode = _cg_mode
3426

3427
            if ubatch_slices is not None:
3428
3429
3430
3431
3432
3433
3434
                # 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

3435
3436
3437
            with (
                self.maybe_randomize_inputs(input_ids),
                set_forward_context(
3438
3439
                    attn_metadata,
                    self.vllm_config,
3440
                    num_tokens=num_tokens_after_padding,
3441
3442
                    num_tokens_across_dp=num_tokens_across_dp,
                    cudagraph_runtime_mode=cudagraph_runtime_mode,
3443
                    batch_descriptor=batch_descriptor,
3444
3445
3446
                    ubatch_slices=ubatch_slices,
                ),
            ):
3447
                outputs = self.model(
3448
3449
3450
3451
                    input_ids=input_ids,
                    positions=positions,
                    intermediate_tensors=intermediate_tensors,
                    inputs_embeds=inputs_embeds,
3452
                    **model_kwargs,
3453
                )
3454

3455
3456
3457
3458
            if self.use_aux_hidden_state_outputs:
                hidden_states, _ = outputs
            else:
                hidden_states = outputs
3459

3460
            if self.speculative_config and self.speculative_config.use_eagle():
3461
                assert isinstance(self.drafter, EagleProposer)
3462
3463
                use_cudagraphs = cudagraph_runtime_mode == CUDAGraphMode.PIECEWISE
                self.drafter.dummy_run(num_tokens, use_cudagraphs=use_cudagraphs)
3464

3465
3466
3467
3468
3469
3470
3471
3472
3473
3474
        # 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)

3475
        logit_indices = np.cumsum(num_scheduled_tokens) - 1
3476
        return hidden_states, hidden_states[logit_indices]
3477
3478
3479
3480
3481
3482

    @torch.inference_mode()
    def _dummy_sampler_run(
        self,
        hidden_states: torch.Tensor,
    ) -> torch.Tensor:
3483
3484
3485
3486
        # 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)
3487

3488
        logits = self.model.compute_logits(hidden_states)
3489
3490
        num_reqs = logits.size(0)

3491
        dummy_tensors = lambda v: torch.full((num_reqs,), v, device=self.device)
3492
3493
3494
3495
3496
3497
3498
3499
3500
3501
3502
3503
3504
3505
3506

        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)],
3507
            spec_token_ids=[[] for _ in range(num_reqs)],
3508
3509
            allowed_token_ids_mask=None,
            bad_words_token_ids={},
3510
            logitsprocs=LogitsProcessors(),
3511
        )
3512
        try:
3513
3514
3515
            sampler_output = self.sampler(
                logits=logits, sampling_metadata=dummy_metadata
            )
3516
        except RuntimeError as e:
3517
            if "out of memory" in str(e):
3518
3519
3520
3521
                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 "
3522
3523
                    "initializing the engine."
                ) from e
3524
3525
            else:
                raise e
3526
        if self.speculative_config:
3527
3528
            draft_token_ids = [[0] for _ in range(num_reqs)]
            dummy_spec_decode_metadata = SpecDecodeMetadata.make_dummy(
3529
3530
                draft_token_ids, self.device
            )
3531
3532
3533
3534
3535
3536

            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
3537
3538
3539
            target_logits = torch.randn(
                num_tokens, logits.shape[-1], device=self.device, dtype=logits.dtype
            )
3540
3541
3542
            # 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.
3543
3544
3545
            bonus_token_ids = torch.zeros(
                num_reqs, device=self.device, dtype=torch.int32
            )
3546
3547
3548
3549
3550
3551
3552
            self.rejection_sampler(
                dummy_spec_decode_metadata,
                draft_probs,
                target_logits,
                bonus_token_ids,
                dummy_metadata,
            )
3553
        return sampler_output
3554

3555
    def _dummy_pooler_run_task(
3556
3557
        self,
        hidden_states: torch.Tensor,
3558
3559
        task: PoolingTask,
    ) -> PoolerOutput:
3560
3561
3562
3563
3564
3565
3566
3567
3568
3569
3570
        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

3571
        dummy_prompt_lens = torch.tensor(
3572
3573
            num_scheduled_tokens_list,
            device="cpu",
3574
        )
3575
3576
3577
        dummy_token_ids = torch.zeros(
            (num_reqs, req_num_tokens), dtype=torch.int32, device=self.device
        )
3578

3579
        model = cast(VllmModelForPooling, self.get_model())
3580
        dummy_pooling_params = PoolingParams(task=task)
3581
        dummy_pooling_params.verify(task=task, model_config=self.model_config)
3582
        to_update = model.pooler.get_pooling_updates(task)
3583
3584
        to_update.apply(dummy_pooling_params)

3585
        dummy_metadata = PoolingMetadata(
3586
3587
3588
3589
            prompt_lens=dummy_prompt_lens,
            prompt_token_ids=dummy_token_ids,
            pooling_params=[dummy_pooling_params] * num_reqs,
        )
3590

3591
3592
3593
        dummy_metadata.build_pooling_cursor(
            num_scheduled_tokens_list, device=hidden_states.device
        )
3594

3595
        try:
3596
3597
3598
            return model.pooler(
                hidden_states=hidden_states, pooling_metadata=dummy_metadata
            )
3599
        except RuntimeError as e:
3600
            if "out of memory" in str(e):
3601
                raise RuntimeError(
3602
3603
3604
                    "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 "
3605
3606
                    "initializing the engine."
                ) from e
3607
3608
            else:
                raise e
3609
3610
3611
3612
3613
3614
3615
3616
3617
3618
3619

    @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)
3620
            output_size[task] = sum(o.nbytes for o in output)
3621
3622
3623
3624
            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)
3625

3626
    def profile_run(self) -> None:
3627
        # Profile with multimodal encoder & encoder cache.
3628
        if self.supports_mm_inputs:
3629
            if self.model_config.multimodal_config.skip_mm_profiling:
3630
                logger.info(
3631
                    "Skipping memory profiling for multimodal encoder and "
3632
3633
                    "encoder cache."
                )
3634
3635
3636
3637
3638
3639
3640
3641
            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.
3642
                    dummy_modality = mm_budget.get_modality_with_max_tokens()
3643
3644
3645
                    max_mm_items_per_batch = mm_budget.max_items_per_batch_by_modality[
                        dummy_modality
                    ]
3646
3647
3648
3649
3650
3651
3652
3653
3654

                    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,
                    )
3655

3656
3657
3658
3659
3660
                    # Create dummy batch of multimodal inputs.
                    batched_dummy_mm_inputs = self._get_mm_dummy_batch(
                        dummy_modality,
                        max_mm_items_per_batch,
                    )
3661

3662
                    # Run multimodal encoder.
3663
3664
3665
                    dummy_encoder_outputs = self.model.get_multimodal_embeddings(
                        **batched_dummy_mm_inputs
                    )
3666

3667
3668
3669
3670
                    sanity_check_mm_encoder_outputs(
                        dummy_encoder_outputs,
                        expected_num_items=max_mm_items_per_batch,
                    )
3671

3672
3673
3674
3675
3676
3677
3678
3679
3680
3681
                    # 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(
3682
3683
                                (encoder_budget, encoder_output_shape[-1])
                            )
3684
3685
3686
3687
3688
3689
                            num_tokens = output.shape[0]
                            expanded[:num_tokens].copy_(output)
                            expanded_outputs.append(expanded)

                        dummy_encoder_outputs = expanded_outputs

3690
                    # Cache the dummy encoder outputs.
3691
                    self.encoder_cache["tmp"] = dict(enumerate(dummy_encoder_outputs))
3692

3693
        # Add `is_profile` here to pre-allocate communication buffers
3694
3695
3696
        hidden_states, last_hidden_states = self._dummy_run(
            self.max_num_tokens, is_profile=True
        )
3697
        if get_pp_group().is_last_rank:
3698
3699
3700
3701
            if self.is_pooling_model:
                output = self._dummy_pooler_run(hidden_states)
            else:
                output = self._dummy_sampler_run(last_hidden_states)
3702
        else:
3703
            output = None
3704
        self._sync_device()
3705
        del hidden_states, output
3706
        self.encoder_cache.clear()
3707
        gc.collect()
3708

3709
    def capture_model(self) -> int:
3710
        if self.compilation_config.cudagraph_mode == CUDAGraphMode.NONE:
3711
            logger.warning(
3712
                "Skipping CUDA graph capture. To turn on CUDA graph capture, "
3713
3714
                "ensure `cudagraph_mode` was not manually set to `NONE`"
            )
3715
            return 0
3716
3717
        else:
            self.initialize_cudagraph_capture()
3718

3719
3720
        compilation_counter.num_gpu_runner_capture_triggers += 1

3721
3722
        start_time = time.perf_counter()

3723
3724
3725
3726
3727
3728
3729
3730
3731
3732
3733
3734
3735
3736
        @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()
3737
                    gc.collect()
3738

3739
3740
3741
        # 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.
3742
        set_cudagraph_capturing_enabled(True)
3743
        with freeze_gc(), graph_capture(device=self.device):
3744
            start_free_gpu_memory = torch.cuda.mem_get_info()[0]
3745
            cudagraph_mode = self.compilation_config.cudagraph_mode
3746
            assert cudagraph_mode is not None
3747
3748
3749
3750
3751
3752
3753
            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,
3754
3755
                    uniform_decode=False,
                )
3756

3757
3758
            # Capture full cudagraph for uniform decode batches if we
            # don't already have full mixed prefill-decode cudagraphs.
3759
3760
3761
3762
3763
3764
3765
            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
                )
3766
                decode_cudagraph_batch_sizes = [
3767
3768
                    x
                    for x in self.cudagraph_batch_sizes
3769
                    if max_num_tokens >= x >= self.uniform_decode_query_len
3770
                ]
3771
                compilation_cases_decode = list(reversed(decode_cudagraph_batch_sizes))
3772
3773
3774
                self._capture_cudagraphs(
                    compilation_cases=compilation_cases_decode,
                    cudagraph_runtime_mode=CUDAGraphMode.FULL,
3775
3776
                    uniform_decode=True,
                )
3777

3778
3779
3780
            torch.cuda.synchronize()
            end_free_gpu_memory = torch.cuda.mem_get_info()[0]

3781
3782
3783
        # 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
3784
        # we may do lazy capturing in future that still allows capturing
3785
3786
        # after here.
        set_cudagraph_capturing_enabled(False)
3787
3788
3789
3790
3791

        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.
3792
3793
3794
3795
3796
        logger.info(
            "Graph capturing finished in %.0f secs, took %.2f GiB",
            elapsed_time,
            cuda_graph_size / (1 << 30),
        )
3797
        return cuda_graph_size
3798

3799
3800
3801
3802
3803
3804
3805
3806
3807
3808
    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}"
3809
3810
3811
3812
3813
3814
3815
3816

        # 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",
3817
3818
3819
                    cudagraph_runtime_mode.name,
                ),
            )
3820

3821
3822
3823
        # 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
3824
3825
3826
            # 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
3827
3828
3829
3830
            allow_microbatching = (
                self.parallel_config.enable_dbo
                and cudagraph_runtime_mode == CUDAGraphMode.FULL
                and uniform_decode
3831
3832
3833
3834
3835
                and check_ubatch_thresholds(
                    config=self.vllm_config.parallel_config,
                    num_tokens=num_tokens,
                    uniform_decode=uniform_decode,
                )
3836
            )
3837

3838
3839
3840
3841
3842
3843
            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.
3844
3845
3846
3847
3848
3849
3850
3851
3852
3853
3854
3855
3856
3857
3858
3859
3860
3861
                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,
            )
3862
        self.maybe_remove_all_loras(self.lora_config)
3863

3864
3865
3866
3867
    def initialize_attn_backend(self, kv_cache_config: KVCacheConfig) -> None:
        """
        Initialize the attention backends and attention metadata builders.
        """
3868
        assert len(self.attn_groups) == 0, "Attention backends are already initialized"
3869

3870
3871
3872
3873
3874
3875
3876
3877
        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(
3878
3879
                self.vllm_config, AttentionLayerBase, kv_cache_group_spec.layer_names
            )
3880
3881
            attn_backends = {}
            attn_backend_layers = defaultdict(list)
3882
            # Dedupe based on full class name; this is a bit safer than
3883
3884
3885
3886
            # 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.
3887
            for layer_name in kv_cache_group_spec.layer_names:
3888
                attn_backend = layers[layer_name].get_attn_backend()
3889
3890
3891
3892
3893
3894
3895

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

3896
3897
3898
                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):
3899
                    layer_kv_cache_spec = layer_kv_cache_spec.kv_cache_specs[layer_name]
3900
                key = (full_cls_name, layer_kv_cache_spec)
3901
3902
3903
                attn_backends[key] = AttentionGroupKey(
                    attn_backend, layer_kv_cache_spec
                )
3904
                attn_backend_layers[key].append(layer_name)
3905
            return {attn_backends[k]: v for k, v in attn_backend_layers.items()}
3906
3907

        def create_attn_groups(
3908
            attn_backends_map: dict[AttentionGroupKey, list[str]],
3909
3910
        ) -> list[AttentionGroup]:
            attn_groups: list[AttentionGroup] = []
3911
            for (attn_backend, kv_cache_spec), layer_names in attn_backends_map.items():
3912
3913
                attn_group = AttentionGroup.create_with_metadata_builders(
                    attn_backend,
3914
                    layer_names,
3915
                    kv_cache_spec,
3916
3917
                    self.vllm_config,
                    self.device,
3918
                    num_metadata_builders=1
3919
3920
                    if not self.parallel_config.enable_dbo
                    else 2,
3921
3922
                )

3923
3924
3925
3926
                attn_groups.append(attn_group)
            return attn_groups

        for kv_cache_group_spec in kv_cache_config.kv_cache_groups:
3927
3928
            attn_backends = get_attn_backends_for_group(kv_cache_group_spec)
            self.attn_groups.append(create_attn_groups(attn_backends))
3929

co63oc's avatar
co63oc committed
3930
        # Calculate reorder batch threshold (if needed)
3931
3932
        self.calculate_reorder_batch_threshold()

3933
    def initialize_cudagraph_capture(self) -> None:
3934
        """
3935
        Resolve the cudagraph_mode when there are multiple attention
3936
3937
3938
3939
        backends with potential conflicting CUDA graph support.
        Then initialize the cudagraph_dispatcher based on the resolved
        cudagraph_mode.
        """
3940
3941
3942
3943
        min_cg_support = AttentionCGSupport.ALWAYS
        min_cg_builder_name = None

        for attn_group in self._attn_group_iterator():
3944
            builder = attn_group.get_metadata_builder()
3945
3946
3947
3948
3949
3950
            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
3951
3952
3953
3954
3955
3956
3957
3958
3959
        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})"
            )
3960
3961
            if min_cg_support == AttentionCGSupport.NEVER:
                # if not supported any full cudagraphs, just raise it.
3962
3963
                msg += (
                    "; please try cudagraph_mode=PIECEWISE, and "
3964
                    "make sure compilation level is piecewise"
3965
                )
3966
3967
3968
3969
3970
                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"
3971
                cudagraph_mode = self.compilation_config.cudagraph_mode = (
3972
                    CUDAGraphMode.FULL_AND_PIECEWISE
3973
                )
3974
3975
            else:
                msg += "; setting cudagraph_mode=FULL_DECODE_ONLY"
3976
                cudagraph_mode = self.compilation_config.cudagraph_mode = (
3977
                    CUDAGraphMode.FULL_DECODE_ONLY
3978
                )
3979
3980
            logger.warning(msg)

3981
        # check that if we are doing decode full-cudagraphs it is supported
3982
3983
3984
3985
3986
3987
3988
3989
3990
3991
3992
3993
3994
3995
3996
        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 "
3997
                    "attention is compiled piecewise"
3998
3999
                )
                cudagraph_mode = self.compilation_config.cudagraph_mode = (
4000
                    CUDAGraphMode.PIECEWISE
4001
                )
4002
            else:
4003
4004
                msg += (
                    "; setting cudagraph_mode=NONE because "
4005
                    "attention is not compiled piecewise"
4006
4007
                )
                cudagraph_mode = self.compilation_config.cudagraph_mode = (
4008
                    CUDAGraphMode.NONE
4009
                )
4010
4011
            logger.warning(msg)

4012
4013
        # check that if we are doing spec-decode + decode full-cudagraphs it is
        # supported
4014
4015
4016
4017
4018
4019
4020
4021
4022
4023
        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})"
            )
4024
4025
            if self.compilation_config.splitting_ops_contain_attention():
                msg += "; setting cudagraph_mode=PIECEWISE"
4026
                cudagraph_mode = self.compilation_config.cudagraph_mode = (
4027
                    CUDAGraphMode.PIECEWISE
4028
                )
4029
4030
            else:
                msg += "; setting cudagraph_mode=NONE"
4031
                cudagraph_mode = self.compilation_config.cudagraph_mode = (
4032
                    CUDAGraphMode.NONE
4033
                )
4034
4035
4036
4037
            logger.warning(msg)

        # double check that we can support full cudagraph if they are requested
        # even after automatic downgrades
4038
4039
4040
4041
4042
4043
4044
4045
4046
4047
4048
        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"
            )
4049
4050
4051
4052

        # Trigger cudagraph dispatching keys initialization here (after
        # initializing attn backends).
        self.cudagraph_dispatcher.initialize_cudagraph_keys(
4053
4054
            self.compilation_config.cudagraph_mode, self.uniform_decode_query_len
        )
4055

4056
4057
4058
4059
4060
    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)
        """
4061
        for group in self._attn_group_iterator():
4062
            attn_metadata_builder_i = group.get_metadata_builder()
4063

4064
4065
            # check that if any backends reorder batches; that the reordering
            # is compatible (e.g., decode threshold is the same)
4066
            reorder_batch_threshold_i = attn_metadata_builder_i.reorder_batch_threshold
4067
4068
            if reorder_batch_threshold_i is not None:
                if self.reorder_batch_threshold is not None:
4069
                    if reorder_batch_threshold_i != self.reorder_batch_threshold:
4070
4071
4072
4073
                        raise ValueError(
                            f"Attention backend reorders decodes with "
                            f"threshold {reorder_batch_threshold_i} but other "
                            f"backend uses threshold "
4074
4075
                            f"{self.reorder_batch_threshold}"
                        )
4076
4077
4078
                else:
                    self.reorder_batch_threshold = reorder_batch_threshold_i

4079
4080
4081
4082
4083
4084
4085
4086
4087
4088
4089
4090
4091
4092
4093
4094
4095
4096
4097
4098
4099
4100
4101
4102
4103
4104
4105
4106
4107
4108
4109
4110
4111
4112
4113
4114
4115
4116
4117
4118
4119
4120
4121
4122
4123
4124
4125
4126
4127
4128
4129
4130
4131
4132
4133
4134
4135
4136
4137
4138
4139
4140
4141
4142
4143
4144
4145
4146
4147
4148
4149
4150
4151
4152
4153
4154
4155
4156
4157
4158
    def _find_compatible_block_sizes(
        self,
        kv_manager_block_size: int,
        backend_cls: type[AttentionBackend],
        return_all: bool = False,
    ) -> list[int]:
        """
        Find compatible block sizes for a backend.

        Args:
            kv_manager_block_size: Physical block size of KV cache
            backend_cls: Attention backend class
            return_all: Return all compatible sizes if True, max size if False

        Returns:
            Compatible block size(s) based on return_all parameter

        Raises:
            ValueError: If no compatible block size found
        """
        supported_block_size = backend_cls.get_supported_kernel_block_size()
        compatible_sizes = []

        for block_size in supported_block_size:
            if isinstance(block_size, int):
                if kv_manager_block_size % block_size == 0:
                    compatible_sizes.append(block_size)
            elif (
                isinstance(block_size, MultipleOf)
                and kv_manager_block_size % block_size.base == 0
            ):
                compatible_sizes.append(kv_manager_block_size)

        if not compatible_sizes:
            raise ValueError(f"No compatible block size for {kv_manager_block_size}")

        return compatible_sizes if return_all else [max(compatible_sizes)]

    def _select_common_block_size(
        self, kv_manager_block_size: int, attn_groups: list[AttentionGroup]
    ) -> int:
        """
        Select common block size for all backends.

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

        Returns:
            Block size supported by all backends,
            prioritizing cache_config.block_size

        Raises:
            ValueError: If no common block size found
        """
        all_backend_supports = []

        for attn_group in attn_groups:
            compatible_sizes = self._find_compatible_block_sizes(
                kv_manager_block_size, attn_group.backend, return_all=True
            )
            supported_sizes = sorted(list(set(compatible_sizes)), reverse=True)
            all_backend_supports.append(set(supported_sizes))

        common_supported_sizes = set.intersection(*all_backend_supports)

        if not common_supported_sizes:
            error_msg = f"No common block size for {kv_manager_block_size}. "
            for i, attn_group in enumerate(attn_groups):
                supported = all_backend_supports[i]
                error_msg += (
                    f"Backend {attn_group.backend} supports: {sorted(supported)}. "
                )
            raise ValueError(error_msg)

        if self.cache_config.block_size in common_supported_sizes:
            return self.cache_config.block_size

        return max(common_supported_sizes)

4159
    def may_reinitialize_input_batch(self, kv_cache_config: KVCacheConfig) -> None:
4160
4161
4162
4163
4164
4165
4166
4167
4168
4169
4170
        """
        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
4171
            if not isinstance(kv_cache_group.kv_cache_spec, EncoderOnlyAttentionSpec)
4172
        ]
4173
4174
4175
4176
4177
4178
4179

        # Generate kernel_block_sizes that matches each block_size
        kernel_block_sizes = self._prepare_kernel_block_sizes(kv_cache_config)

        if block_sizes != [self.cache_config.block_size] or kernel_block_sizes != [
            self.cache_config.block_size
        ]:
4180
4181
4182
            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
4183
4184
                "for more details."
            )
4185
4186
            self.input_batch = InputBatch(
                max_num_reqs=self.max_num_reqs,
4187
                max_model_len=max(self.max_model_len, self.max_encoder_len),
4188
4189
4190
4191
4192
                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,
4193
                kernel_block_sizes=kernel_block_sizes,
4194
                is_spec_decode=bool(self.vllm_config.speculative_config),
4195
4196
                logitsprocs=self.input_batch.logitsprocs,
                is_pooling_model=self.is_pooling_model,
4197
4198
                num_speculative_tokens=(
                    self.vllm_config.speculative_config.num_speculative_tokens
4199
4200
4201
                    if self.vllm_config.speculative_config
                    else 0
                ),
4202
4203
            )

4204
    def _allocate_kv_cache_tensors(
4205
4206
        self, kv_cache_config: KVCacheConfig
    ) -> dict[str, torch.Tensor]:
4207
        """
4208
4209
4210
        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.

4211
        Args:
4212
            kv_cache_config: The KV cache config
4213
        Returns:
4214
            dict[str, torch.Tensor]: A map between layer names to their
4215
            corresponding memory buffer for KV cache.
4216
        """
4217
4218
        kv_cache_raw_tensors: dict[str, torch.Tensor] = {}
        for kv_cache_tensor in kv_cache_config.kv_cache_tensors:
4219
4220
4221
            tensor = torch.zeros(
                kv_cache_tensor.size, dtype=torch.int8, device=self.device
            )
4222
4223
4224
4225
4226
            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:
4227
4228
4229
4230
            for layer_name in group.layer_names:
                if layer_name in self.runner_only_attn_layers:
                    continue
                layer_names.add(layer_name)
4231
4232
4233
        assert layer_names == set(kv_cache_raw_tensors.keys()), (
            "Some layers are not correctly initialized"
        )
4234
4235
        return kv_cache_raw_tensors

4236
4237
4238
    def _attn_group_iterator(self) -> Iterator[AttentionGroup]:
        return itertools.chain.from_iterable(self.attn_groups)

4239
    def _kv_cache_spec_attn_group_iterator(self) -> Iterator[AttentionGroup]:
4240
4241
        if not self.kv_cache_config.kv_cache_groups:
            return
4242
4243
        for attn_groups in self.attn_groups:
            yield from attn_groups
4244

4245
4246
4247
4248
4249
4250
4251
4252
4253
4254
4255
4256
4257
4258
4259
4260
4261
4262
    def _prepare_kernel_block_sizes(self, kv_cache_config: KVCacheConfig) -> list[int]:
        """
        Generate kernel_block_sizes that matches each block_size.

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

        Args:
            kv_cache_config: The KV cache configuration.

        Returns:
            list[int]: List of kernel block sizes for each cache group.
        """
        kernel_block_sizes = []
        for kv_cache_group_id, kv_cache_group in enumerate(
            kv_cache_config.kv_cache_groups
        ):
4263
4264
4265
4266
4267
4268
            kv_cache_spec = kv_cache_group.kv_cache_spec
            if isinstance(kv_cache_spec, UniformTypeKVCacheSpecs):
                # All layers in the UniformTypeKVCacheSpecs have the same type,
                # Pick an arbitrary one to dispatch.
                kv_cache_spec = next(iter(kv_cache_spec.kv_cache_specs.values()))
            if isinstance(kv_cache_spec, EncoderOnlyAttentionSpec):
4269
                continue
4270
            elif isinstance(kv_cache_spec, AttentionSpec):
4271
4272
4273
4274
4275
4276
4277
4278
4279
                # This is an attention backend that supports virtual
                # block splitting. Get the supported block sizes from
                # all backends in the group.
                attn_groups = self.attn_groups[kv_cache_group_id]
                kv_manager_block_size = kv_cache_group.kv_cache_spec.block_size
                selected_kernel_size = self._select_common_block_size(
                    kv_manager_block_size, attn_groups
                )
                kernel_block_sizes.append(selected_kernel_size)
4280
            elif isinstance(kv_cache_spec, MambaSpec):
4281
4282
                # This is likely Mamba or other non-attention cache,
                # no splitting.
4283
                kernel_block_sizes.append(kv_cache_spec.block_size)
4284
4285
4286
4287
4288
4289
            else:
                raise NotImplementedError(
                    f"unknown kv cache spec {kv_cache_group.kv_cache_spec}"
                )
        return kernel_block_sizes

4290
4291
4292
4293
4294
    def _reshape_kv_cache_tensors(
        self,
        kv_cache_config: KVCacheConfig,
        kv_cache_raw_tensors: dict[str, torch.Tensor],
    ) -> dict[str, torch.Tensor]:
4295
        """
4296
        Reshape the KV cache tensors to the desired shape and dtype.
4297

4298
        Args:
4299
4300
            kv_cache_config: The KV cache config
            kv_cache_raw_tensors: The KV cache buffer of each layer, with
4301
                correct size but uninitialized shape.
4302
        Returns:
4303
            Dict[str, torch.Tensor]: A map between layer names to their
4304
4305
            corresponding memory buffer for KV cache.
        """
4306
        kv_caches: dict[str, torch.Tensor] = {}
4307
        has_attn, has_mamba = False, False
4308
4309
        for group in self._kv_cache_spec_attn_group_iterator():
            kv_cache_spec = group.kv_cache_spec
4310
4311
            attn_backend = group.backend
            for layer_name in group.layer_names:
4312
4313
                if layer_name in self.runner_only_attn_layers:
                    continue
4314
4315
                raw_tensor = kv_cache_raw_tensors[layer_name]
                assert raw_tensor.numel() % kv_cache_spec.page_size_bytes == 0
4316
                num_blocks = raw_tensor.numel() // kv_cache_spec.page_size_bytes
4317
                if isinstance(kv_cache_spec, AttentionSpec):
4318
                    has_attn = True
4319
4320
4321
4322
4323
4324
4325
4326
                    kv_manager_block_size = kv_cache_spec.block_size
                    kernel_size_list = self._find_compatible_block_sizes(
                        kv_manager_block_size, attn_backend, return_all=False
                    )
                    kernel_size = kernel_size_list[0]
                    num_blocks_per_kv_block = kv_manager_block_size // kernel_size
                    kernel_num_blocks = num_blocks * num_blocks_per_kv_block

4327
                    kv_cache_shape = attn_backend.get_kv_cache_shape(
4328
4329
                        kernel_num_blocks,
                        kernel_size,
4330
4331
                        kv_cache_spec.num_kv_heads,
                        kv_cache_spec.head_size,
4332
4333
                        cache_dtype_str=self.cache_config.cache_dtype,
                    )
4334
                    dtype = kv_cache_spec.dtype
4335
                    try:
4336
                        kv_cache_stride_order = attn_backend.get_kv_cache_stride_order()  # noqa: E501
4337
                        assert len(kv_cache_stride_order) == len(kv_cache_shape)
4338
                    except (AttributeError, NotImplementedError):
4339
                        kv_cache_stride_order = tuple(range(len(kv_cache_shape)))
4340
4341
4342
4343
4344
                    # 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.
4345
4346
4347
                    kv_cache_shape = tuple(
                        kv_cache_shape[i] for i in kv_cache_stride_order
                    )
4348
4349
4350
4351
4352
                    # Maintain original KV shape view.
                    inv_order = [
                        kv_cache_stride_order.index(i)
                        for i in range(len(kv_cache_stride_order))
                    ]
4353
4354
4355
4356
4357
4358
                    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
4359
                elif isinstance(kv_cache_spec, MambaSpec):
4360
                    has_mamba = True
Chen Zhang's avatar
Chen Zhang committed
4361
4362
                    raw_tensor = kv_cache_raw_tensors[layer_name]
                    state_tensors = []
4363
                    storage_offset_bytes = 0
4364
                    for shape, dtype in zip(kv_cache_spec.shapes, kv_cache_spec.dtypes):
4365
4366
                        dtype_size = get_dtype_size(dtype)
                        num_element_per_page = (
4367
4368
                            kv_cache_spec.page_size_bytes // dtype_size
                        )
Chen Zhang's avatar
Chen Zhang committed
4369
                        target_shape = (num_blocks, *shape)
4370
4371
                        stride = torch.empty(target_shape).stride()
                        target_stride = (num_element_per_page, *stride[1:])
4372
                        assert storage_offset_bytes % dtype_size == 0
4373
4374
4375
4376
                        tensor = torch.as_strided(
                            raw_tensor.view(dtype),
                            size=target_shape,
                            stride=target_stride,
4377
                            storage_offset=storage_offset_bytes // dtype_size,
4378
                        )
Chen Zhang's avatar
Chen Zhang committed
4379
                        state_tensors.append(tensor)
4380
                        storage_offset_bytes += stride[0] * dtype_size
4381
4382

                    kv_caches[layer_name] = state_tensors
4383
                else:
4384
                    raise NotImplementedError
4385
4386

        if has_attn and has_mamba:
4387
            self._update_hybrid_attention_mamba_layout(kv_caches)
4388

4389
4390
        return kv_caches

4391
    def _update_hybrid_attention_mamba_layout(
4392
4393
        self, kv_caches: dict[str, torch.Tensor]
    ) -> None:
4394
        """
4395
4396
        Update the layout of attention layers from (2, num_blocks, ...) to
        (num_blocks, 2, ...).
4397
4398

        Args:
4399
            kv_caches: The KV cache buffer of each layer.
4400
4401
        """

4402
4403
        for group in self._kv_cache_spec_attn_group_iterator():
            kv_cache_spec = group.kv_cache_spec
4404
            for layer_name in group.layer_names:
4405
                kv_cache = kv_caches[layer_name]
4406
4407
4408
4409
                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 "
4410
                        f"a tensor of shape {kv_cache.shape}"
4411
                    )
4412
                    hidden_size = kv_cache.shape[2:].numel()
4413
4414
4415
4416
                    kv_cache.as_strided_(
                        size=kv_cache.shape,
                        stride=(hidden_size, 2 * hidden_size, *kv_cache.stride()[2:]),
                    )
4417

4418
    def initialize_kv_cache_tensors(
4419
4420
        self, kv_cache_config: KVCacheConfig
    ) -> dict[str, torch.Tensor]:
4421
4422
4423
4424
4425
4426
        """
        Initialize the memory buffer for KV cache.

        Args:
            kv_cache_config: The KV cache config
        Returns:
4427
            Dict[str, torch.Tensor]: A map between layer names to their
4428
4429
4430
4431
4432
            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
4433
4434
4435
        kv_caches = self._reshape_kv_cache_tensors(
            kv_cache_config, kv_cache_raw_tensors
        )
4436

4437
        # Set up cross-layer KV cache sharing
4438
4439
        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)
4440
4441
            kv_caches[layer_name] = kv_caches[target_layer_name]

4442
4443
4444
4445
4446
4447
4448
4449
4450
        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,
        )
4451
4452
4453
        return kv_caches

    def maybe_add_kv_sharing_layers_to_kv_cache_groups(
4454
4455
        self, kv_cache_config: KVCacheConfig
    ) -> None:
4456
4457
4458
4459
4460
4461
4462
4463
4464
4465
4466
4467
4468
4469
4470
4471
4472
4473
        """
        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.
4474
            attn_layers = get_layers_from_vllm_config(self.vllm_config, Attention)
4475
4476
            for layer_name in reversed(attn_layers):
                if layer_name in self.shared_kv_cache_layers:
4477
                    self.kv_sharing_fast_prefill_eligible_layers.add(layer_name)
4478
4479
                else:
                    break
4480

4481
4482
4483
4484
4485
4486
4487
    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
        """
4488
        kv_cache_config = deepcopy(kv_cache_config)
4489
        self.kv_cache_config = kv_cache_config
4490
        self.may_add_encoder_only_layers_to_kv_cache_config()
4491
        self.maybe_add_kv_sharing_layers_to_kv_cache_groups(kv_cache_config)
4492
        self.initialize_attn_backend(kv_cache_config)
4493
4494
        # Reinitialize need to after initialize_attn_backend
        self.may_reinitialize_input_batch(kv_cache_config)
4495
4496
        kv_caches = self.initialize_kv_cache_tensors(kv_cache_config)

4497
4498
4499
4500
4501
4502
        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
4503
        if has_kv_transfer_group():
4504
4505
4506
            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
4507

4508
        if self.dcp_world_size > 1:
4509
            layer_names = self.attn_groups[0][0].layer_names
4510
4511
4512
            layers = get_layers_from_vllm_config(
                self.vllm_config, AttentionLayerBase, layer_names
            )
4513
4514
4515
4516
4517
            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__} "
4518
4519
                    "does not return the softmax lse for decode."
                )
4520

4521
4522
4523
4524
4525
    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
4526
        encoder_only_attn_specs: dict[AttentionSpec, list[str]] = defaultdict(list)
4527
4528
4529
        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:
4530
                attn_spec: AttentionSpec = EncoderOnlyAttentionSpec(
4531
4532
4533
                    block_size=block_size,
                    num_kv_heads=attn_module.num_kv_heads,
                    head_size=attn_module.head_size,
4534
4535
                    dtype=self.kv_cache_dtype,
                )
4536
4537
4538
                encoder_only_attn_specs[attn_spec].append(layer_name)
                self.runner_only_attn_layers.add(layer_name)
        if len(encoder_only_attn_specs) > 0:
4539
4540
4541
            assert len(encoder_only_attn_specs) == 1, (
                "Only support one encoder-only attention spec now"
            )
4542
4543
            spec, layer_names = encoder_only_attn_specs.popitem()
            self.kv_cache_config.kv_cache_groups.append(
4544
4545
                KVCacheGroupSpec(layer_names=layer_names, kv_cache_spec=spec)
            )
4546

4547
    def get_kv_cache_spec(self) -> dict[str, KVCacheSpec]:
4548
        """
4549
        Generates the KVCacheSpec by parsing the kv cache format from each
4550
4551
        Attention module in the static forward context.
        Returns:
4552
            KVCacheSpec: A dictionary mapping layer names to their KV cache
4553
4554
4555
4556
            format. Layers that do not need KV cache are not included.
        """

        block_size = self.vllm_config.cache_config.block_size
4557
        use_mla = self.vllm_config.model_config.use_mla
4558
        cache_dtype_str = self.vllm_config.cache_config.cache_dtype
4559
        kv_cache_spec: dict[str, KVCacheSpec] = {}
4560
        attn_layers = get_layers_from_vllm_config(self.vllm_config, AttentionLayerBase)
Chen Zhang's avatar
Chen Zhang committed
4561
        for layer_name, attn_module in attn_layers.items():
4562
4563
4564
4565
4566
4567
4568
4569
4570
4571
4572
4573
4574
            if isinstance(attn_module, Attention):
                if (
                    kv_tgt_layer := attn_module.kv_sharing_target_layer_name
                ) is not None:
                    # 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
4575

4576
4577
4578
4579
4580
4581
4582
4583
4584
4585
4586
4587
4588
4589
4590
4591
4592
4593
4594
4595
4596
4597
4598
4599
4600
4601
4602
4603
4604
4605
4606
                # TODO(lucas): move the attention specs into the model layers like
                # the attention backends
                if attn_module.attn_type == AttentionType.DECODER:
                    if attn_module.sliding_window is not None:
                        assert not use_mla, "MLA is not supported for slidingwindow"
                        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,
                            sliding_window=attn_module.sliding_window,
                        )
                    elif self.attention_chunk_size is not None and isinstance(
                        attn_module, ChunkedLocalAttention
                    ):
                        kv_cache_spec[layer_name] = ChunkedLocalAttentionSpec(
                            block_size=block_size,
                            num_kv_heads=attn_module.num_kv_heads,
                            head_size=attn_module.head_size,
                            dtype=self.kv_cache_dtype,
                            attention_chunk_size=self.attention_chunk_size,
                        )
                    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,
                            dtype=self.kv_cache_dtype,
                        )
                elif attn_module.attn_type == AttentionType.ENCODER_DECODER:
                    kv_cache_spec[layer_name] = CrossAttentionSpec(
4607
4608
4609
4610
                        block_size=block_size,
                        num_kv_heads=attn_module.num_kv_heads,
                        head_size=attn_module.head_size,
                        dtype=self.kv_cache_dtype,
4611
                    )
4612
4613
4614
                elif attn_module.attn_type in (
                    AttentionType.ENCODER,
                    AttentionType.ENCODER_ONLY,
4615
                ):
4616
4617
                    # encoder-only attention does not need KV cache.
                    continue
4618
                else:
4619
4620
4621
4622
                    raise ValueError(f"Unknown attention type: {attn_module.attn_type}")

            elif isinstance(attn_module, MLAAttention):
                kv_cache_spec[layer_name] = MLAAttentionSpec(
4623
                    block_size=block_size,
4624
                    num_kv_heads=1,
4625
                    head_size=attn_module.head_size,
4626
                    dtype=self.kv_cache_dtype,
4627
                    cache_dtype_str=cache_dtype_str,
4628
                )
4629

4630
4631
4632
4633
4634
4635
4636
4637
4638
4639
4640
            elif isinstance(attn_module, MambaBase):
                if (
                    self.vllm_config.speculative_config is not None
                    and self.vllm_config.model_config.hf_config.model_type
                    not in ["qwen3_next"]
                ):
                    raise NotImplementedError(
                        "Mamba with speculative decoding is not supported yet."
                    )
                mamba_block_size = self.vllm_config.cache_config.mamba_block_size
                page_size_padded = self.vllm_config.cache_config.mamba_page_size_padded
Chen Zhang's avatar
Chen Zhang committed
4641
                kv_cache_spec[layer_name] = MambaSpec(
4642
4643
                    shapes=attn_module.get_state_shape(),
                    dtypes=attn_module.get_state_dtype(),
4644
                    block_size=mamba_block_size,
4645
                    page_size_padded=page_size_padded,
4646
                    mamba_type=attn_module.mamba_type,
4647
4648
                    num_speculative_blocks=(
                        self.speculative_config.num_speculative_tokens
4649
4650
4651
                        if self.speculative_config
                        else 0
                    ),
4652
                )
4653

4654
        ds_indexer_layers = get_layers_from_vllm_config(
4655
4656
            self.vllm_config, DeepseekV32IndexerCache
        )
4657
4658
        for layer_name, ds_indexer_module in ds_indexer_layers.items():
            kv_cache_spec[layer_name] = ds_indexer_module.get_kv_cache_spec()
4659

4660
        return kv_cache_spec
4661
4662
4663
4664
4665
4666
4667
4668
4669
4670

    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.
4671
        pinned = self.sampled_token_ids_pinned_cpu[: sampled_token_ids.shape[0]]
4672
4673
4674
4675
        pinned.copy_(sampled_token_ids, non_blocking=True)
        self.transfer_event.record()
        self.transfer_event.synchronize()
        return pinned.tolist()