gpu_model_runner.py 197 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 itertools import product
12
from typing import TYPE_CHECKING, Any, NamedTuple, TypeAlias, cast
13
14
15
16
17

import numpy as np
import torch
import torch.distributed
import torch.nn as nn
18
from tqdm import tqdm
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.compilation.counter import compilation_counter
24
25
from vllm.compilation.cuda_graph import CUDAGraphWrapper
from vllm.compilation.monitor import set_cudagraph_capturing_enabled
26
from vllm.config import (
27
    CompilationMode,
28
29
30
31
32
    CUDAGraphMode,
    VllmConfig,
    get_layers_from_vllm_config,
    update_config,
)
33
from vllm.distributed.eplb.eplb_state import EplbState
34
from vllm.distributed.kv_transfer import get_kv_transfer_group, has_kv_transfer_group
35
from vllm.distributed.kv_transfer.kv_connector.utils import copy_kv_blocks
36
from vllm.distributed.parallel_state import (
37
38
39
40
41
42
    get_pp_group,
    get_tp_group,
    graph_capture,
    is_global_first_rank,
    prepare_communication_buffer_for_model,
)
43
from vllm.forward_context import BatchDescriptor, set_forward_context
44
from vllm.logger import init_logger
45
from vllm.model_executor.layers.attention_layer_base import AttentionLayerBase
46
from vllm.model_executor.layers.rotary_embedding import MRotaryEmbedding
47
from vllm.model_executor.model_loader import TensorizerLoader, get_model_loader
48
49
50
51
52
53
54
55
from vllm.model_executor.models.interfaces import (
    SupportsMultiModal,
    is_mixture_of_experts,
    supports_eagle3,
    supports_mrope,
    supports_multimodal_pruning,
    supports_transcription,
)
56
from vllm.model_executor.models.interfaces_base import (
57
58
59
60
    VllmModelForPooling,
    is_pooling_model,
    is_text_generation_model,
)
61
from vllm.multimodal import MULTIMODAL_REGISTRY
62
63
64
65
66
from vllm.multimodal.inputs import (
    BatchedTensorInputs,
    MultiModalKwargsItem,
    PlaceholderRange,
)
67
from vllm.multimodal.utils import group_mm_kwargs_by_modality
68
from vllm.pooling_params import PoolingParams
69
from vllm.sampling_params import SamplingType
70
from vllm.sequence import IntermediateTensors
71
from vllm.tasks import GenerationTask, PoolingTask, SupportedTask
72
73
74
75
76
77
78
from vllm.utils import (
    cdiv,
    check_use_alibi,
    is_pin_memory_available,
    length_from_prompt_token_ids_or_embeds,
    round_up,
)
79
from vllm.utils.jsontree import json_map_leaves
80
81
from vllm.utils.mem_constants import GiB_bytes
from vllm.utils.mem_utils import DeviceMemoryProfiler
82
83
84
85
86
from vllm.utils.torch_utils import (
    get_dtype_size,
    kv_cache_dtype_str_to_dtype,
    supports_dynamo,
)
87
from vllm.v1.attention.backends.flash_attn import AttentionMetadata
88
from vllm.v1.attention.backends.gdn_attn import GDNAttentionMetadataBuilder
89
from vllm.v1.attention.backends.utils import (
90
91
92
    AttentionCGSupport,
    AttentionMetadataBuilder,
    CommonAttentionMetadata,
93
    create_fast_prefill_custom_backend,
94
95
96
    reorder_batch_to_split_decodes_and_prefills,
    split_attn_metadata,
)
97
from vllm.v1.cudagraph_dispatcher import CudagraphDispatcher
98
99
100
101
102
103
104
105
106
107
108
109
110
111
112
113
114
115
116
117
118
119
120
from vllm.v1.kv_cache_interface import (
    AttentionSpec,
    ChunkedLocalAttentionSpec,
    CrossAttentionSpec,
    EncoderOnlyAttentionSpec,
    FullAttentionSpec,
    KVCacheConfig,
    KVCacheGroupSpec,
    KVCacheSpec,
    MambaSpec,
    SlidingWindowSpec,
    UniformTypeKVCacheSpecs,
)
from vllm.v1.outputs import (
    EMPTY_MODEL_RUNNER_OUTPUT,
    AsyncModelRunnerOutput,
    DraftTokenIds,
    LogprobsLists,
    LogprobsTensors,
    ModelRunnerOutput,
    PoolerOutput,
    SamplerOutput,
)
121
from vllm.v1.pool.metadata import PoolingMetadata
122
from vllm.v1.sample.logits_processor import LogitsProcessors, build_logitsprocs
123
from vllm.v1.sample.metadata import SamplingMetadata
124
from vllm.v1.sample.rejection_sampler import RejectionSampler
125
from vllm.v1.sample.sampler import Sampler
126
from vllm.v1.spec_decode.eagle import EagleProposer
127
from vllm.v1.spec_decode.medusa import MedusaProposer
128
from vllm.v1.spec_decode.metadata import SpecDecodeMetadata
129
from vllm.v1.spec_decode.ngram_proposer import NgramProposer
130
from vllm.v1.structured_output.utils import apply_grammar_bitmask
131
from vllm.v1.utils import CpuGpuBuffer, record_function_or_nullcontext
132
from vllm.v1.worker.dp_utils import coordinate_batch_across_dp
133
from vllm.v1.worker.gpu_input_batch import CachedRequestState, InputBatch
134
from vllm.v1.worker.gpu_ubatch_wrapper import UBatchWrapper
135
from vllm.v1.worker.kv_connector_model_runner_mixin import KVConnectorModelRunnerMixin
136
from vllm.v1.worker.lora_model_runner_mixin import LoRAModelRunnerMixin
137
138
139
140
141
from vllm.v1.worker.ubatch_utils import (
    UBatchSlice,
    UBatchSlices,
    check_ubatch_thresholds,
)
142
from vllm.v1.worker.utils import is_residual_scattered_for_sp
143

144
145
146
147
148
149
150
151
152
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,
)
153

154
if TYPE_CHECKING:
155
    from vllm.model_executor.model_loader.tensorizer import TensorizerConfig
156
    from vllm.v1.core.sched.output import SchedulerOutput
157
158
159

logger = init_logger(__name__)

160
161
AttnMetadataDict: TypeAlias = dict[str, AttentionMetadata]
# list when ubatching is enabled
162
PerLayerAttnMetadata: TypeAlias = list[AttnMetadataDict] | AttnMetadataDict
163

164

165
166
167
168
169
170
171
172
173
174
175
176
177
# 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.
178
        self.async_copy_ready_event = torch.cuda.Event()
179
180
181
182
183
184
185
186
187

        # 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)
188
            self.sampled_token_ids_cpu = self._sampled_token_ids.to(
189
190
                "cpu", non_blocking=True
            )
191
            self.async_copy_ready_event.record()
192
193
194

    def get_output(self) -> ModelRunnerOutput:
        """Copy the device tensors to the host and return a ModelRunnerOutput.
195

196
197
        This function blocks until the copy is finished.
        """
198
        self.async_copy_ready_event.synchronize()
199
200
201
202

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

203
        valid_sampled_token_ids = self.sampled_token_ids_cpu.tolist()
204
205
206
207
208
209
210
211
        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


212
class GPUModelRunner(LoRAModelRunnerMixin, KVConnectorModelRunnerMixin):
213
214
    def __init__(
        self,
215
        vllm_config: VllmConfig,
216
        device: torch.device,
217
    ):
218
219
220
        self.vllm_config = vllm_config
        self.model_config = vllm_config.model_config
        self.cache_config = vllm_config.cache_config
221
        self.compilation_config = vllm_config.compilation_config
222
223
224
225
226
227
        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
228

229
        from vllm.model_executor.models.utils import set_cpu_offload_max_bytes
230
231

        set_cpu_offload_max_bytes(int(self.cache_config.cpu_offload_gb * 1024**3))
232

233
234
235
236
        model_config = self.model_config
        cache_config = self.cache_config
        scheduler_config = self.scheduler_config
        parallel_config = self.parallel_config
237
        self.device = device
238
239
        self.pin_memory = is_pin_memory_available()
        self.dtype = self.model_config.dtype
240
241
242
        self.kv_cache_dtype = kv_cache_dtype_str_to_dtype(
            cache_config.cache_dtype, self.model_config
        )
243

244
        self.is_pooling_model = model_config.runner_type == "pooling"
245
        self.enable_prompt_embeds = model_config.enable_prompt_embeds
246
        self.is_multimodal_raw_input_only_model = (
247
248
            model_config.is_multimodal_raw_input_only_model
        )
249
250
        # This will be overridden in load_model()
        self.is_multimodal_pruning_enabled = False
251
        self.max_model_len = model_config.max_model_len
252
        self.dcp_world_size = self.parallel_config.decode_context_parallel_size
253
        self.max_num_tokens = scheduler_config.max_num_batched_tokens
254
        self.max_num_reqs = scheduler_config.max_num_seqs
255

256
257
258
259
260
        # 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 = (
261
262
263
            self.parallel_config.distributed_executor_backend == "external_launcher"
            and len(get_pp_group().ranks) > 0
        )
264

265
        # Model-related.
266
        self.num_query_heads = model_config.get_num_attention_heads(parallel_config)
267
        self.hidden_size = model_config.get_hidden_size()
268
        self.attention_chunk_size = model_config.attention_chunk_size
269
270
        # Only relevant for models using ALiBi (e.g, MPT)
        self.use_alibi = check_use_alibi(model_config)
271

272
        self.cascade_attn_enabled = not self.model_config.disable_cascade_attn
273

274
        # Multi-modal data support
275
        self.mm_registry = MULTIMODAL_REGISTRY
276
        self.uses_mrope = model_config.uses_mrope
277
        self.supports_mm_inputs = self.mm_registry.supports_multimodal_inputs(
278
279
            model_config
        )
280

281
282
283
        if self.model_config.is_encoder_decoder:
            # Maximum length of the encoder input, only for encoder-decoder
            # models.
284
            self.max_encoder_len = scheduler_config.max_num_encoder_input_tokens
285
286
287
        else:
            self.max_encoder_len = 0

288
        # Sampler
289
        self.sampler = Sampler(logprobs_mode=self.model_config.logprobs_mode)
290

291
        self.eplb_state: EplbState | None = None
292
293
294
295
296
297
        """
        State of the expert parallelism load balancer.

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

298
        # Lazy initializations
299
        # self.model: nn.Module  # Set after load_model
300
        # Initialize in initialize_kv_cache
301
        self.kv_caches: list[torch.Tensor] = []
302
303
        # indexes: [kv_cache_group_id][attn_group]
        self.attn_groups: list[list[AttentionGroup]] = []
304
305
        # self.kv_cache_config: KVCacheConfig

306
307
        # mm_hash ->  encoder_output
        self.encoder_cache: dict[str, torch.Tensor] = {}
308

309
        self.use_aux_hidden_state_outputs = False
310
311
312
313
314
315
316
317
        # 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():
318
                self.drafter = EagleProposer(self.vllm_config, self.device, self)  # type: ignore
319
320
321
322
                if self.speculative_config.method == "eagle3":
                    self.use_aux_hidden_state_outputs = True
            elif self.speculative_config.method == "medusa":
                self.drafter = MedusaProposer(
323
324
                    vllm_config=self.vllm_config, device=self.device
                )  # type: ignore
325
            else:
326
327
328
329
                raise ValueError(
                    "Unknown speculative decoding method: "
                    f"{self.speculative_config.method}"
                )
330
            self.rejection_sampler = RejectionSampler()
331

332
        # Request states.
333
        self.requests: dict[str, CachedRequestState] = {}
334
        self.comm_stream = torch.cuda.Stream()
335

336
337
338
339
340
341
342
343
344
        # 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.
345
        custom_logitsprocs = model_config.logits_processors
346
347
        self.input_batch = InputBatch(
            max_num_reqs=self.max_num_reqs,
348
349
350
            # 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),
351
352
353
            max_num_batched_tokens=self.max_num_tokens,
            device=self.device,
            pin_memory=self.pin_memory,
354
            vocab_size=self.model_config.get_vocab_size(),
355
            block_sizes=[self.cache_config.block_size],
356
            kernel_block_sizes=[self.cache_config.block_size],
357
            is_spec_decode=bool(self.vllm_config.speculative_config),
358
            logitsprocs=build_logitsprocs(
359
360
361
                self.vllm_config,
                self.device,
                self.pin_memory,
362
                self.is_pooling_model,
363
                custom_logitsprocs,
364
            ),
365
366
367
            # We currently don't know whether a particular custom logits processor
            # uses output token ids so we set this conservatively.
            logitsprocs_need_output_token_ids=bool(custom_logitsprocs),
368
            is_pooling_model=self.is_pooling_model,
369
        )
370

371
        self.use_async_scheduling = self.scheduler_config.async_scheduling
372
373
374
375
376
377
378
379
380
        # Separate cuda stream for overlapping transfer of sampled token ids from
        # GPU to CPU when async scheduling is enabled.
        self.async_output_copy_stream: torch.cuda.Stream | None = None
        # cuda event to synchronize use of reused CPU tensors between steps
        # when async scheduling is enabled.
        self.prepare_inputs_event: torch.cuda.Event | None = None
        if self.use_async_scheduling:
            self.async_output_copy_stream = torch.cuda.Stream()
            self.prepare_inputs_event = torch.cuda.Event()
381

382
        # TODO(woosuk): Provide an option to tune the max cudagraph batch size.
383
384
385
        # The convention is different.
        # self.cudagraph_batch_sizes sorts in ascending order.
        # The batch sizes in the config are in descending order.
386
387
388
389
        if (
            self.compilation_config.cudagraph_capture_sizes
            and self.compilation_config.cudagraph_mode != CUDAGraphMode.NONE
        ):
390
            self.cudagraph_batch_sizes = list(
391
392
                reversed(self.compilation_config.cudagraph_capture_sizes)
            )
393

394
        # Cache the device properties.
395
        self._init_device_properties()
396

397
        # Persistent buffers for CUDA graphs.
398
399
400
401
402
        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
        )
403
        self.seq_lens = self._make_buffer(self.max_num_reqs, dtype=torch.int32)
404
405
406
407
        if self.dcp_world_size > 1:
            self.dcp_local_seq_lens = self._make_buffer(
                self.max_num_reqs, dtype=torch.int32
            )
408
409
410
        # 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.
411
412
413
414
415
416
417
        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
        )
418
419
        self.num_discarded_requests = 0

420
421
422
423
424
425
        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
        )
426

427
428
        # Only relevant for multimodal models
        if self.supports_mm_inputs:
429
            self.is_mm_embed = self._make_buffer(self.max_num_tokens, dtype=torch.bool)
430

431
        # Only relevant for models using M-RoPE (e.g, Qwen2-VL)
432
        if self.uses_mrope:
Roger Wang's avatar
Roger Wang committed
433
434
435
436
            # 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
437
438
439
440
441
442

            # 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
443
            self.mrope_positions = self._make_buffer(
444
445
                (3, self.max_num_tokens + 1), dtype=torch.int64
            )
446

447
        # None in the first PP rank. The rest are set after load_model.
448
        self.intermediate_tensors: IntermediateTensors | None = None
449

450
        # OPTIMIZATION: Cache the tensors rather than creating them every step.
451
        # Keep in int64 to avoid overflow with long context
452
453
454
455
        self.arange_np = np.arange(
            max(self.max_num_reqs + 1, self.max_model_len, self.max_num_tokens),
            dtype=np.int64,
        )
456

457
458
459
460
461
        # 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] = {}
462
463
464
465
466
        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(
467
468
                self.max_num_tokens, dtype=torch.int32, device=self.device
            )
469

470
471
472
473
474
        self.uniform_decode_query_len = (
            1
            if not self.speculative_config
            else 1 + self.speculative_config.num_speculative_tokens
        )
475
476
477
478

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

479
480
481
482
483
484
485
486
487
        self.mm_budget = (
            MultiModalBudget(
                self.model_config,
                self.scheduler_config,
                self.mm_registry,
            )
            if self.supports_mm_inputs
            else None
        )
488

489
        self.reorder_batch_threshold: int | None = None
490

491
492
493
494
495
        # 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()

496
        # Cached outputs.
497
        self._draft_token_ids: list[list[int]] | torch.Tensor | None = None
498
499
500
501
502
        self.transfer_event = torch.cuda.Event()
        self.sampled_token_ids_pinned_cpu = torch.empty(
            (self.max_model_len, 1),
            dtype=torch.int64,
            device="cpu",
503
504
            pin_memory=self.pin_memory,
        )
505

506
507
508
509
    def reset_mm_cache(self) -> None:
        if self.mm_budget:
            self.mm_budget.reset_cache()

510
511
512
513
514
515
516
517
518
519
    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]

520
    def _make_buffer(
521
        self, *size: int | torch.SymInt, dtype: torch.dtype, numpy: bool = True
522
523
524
525
526
527
528
529
    ) -> CpuGpuBuffer:
        return CpuGpuBuffer(
            *size,
            dtype=dtype,
            device=self.device,
            pin_memory=self.pin_memory,
            with_numpy=numpy,
        )
530

531
532
533
    def _init_model_kwargs(self, num_tokens: int):
        model_kwargs = dict[str, Any]()

534
        if not self.is_pooling_model:
535
536
            return model_kwargs

537
538
        num_reqs = self.input_batch.num_reqs
        pooling_params = self.input_batch.get_pooling_params()
539
540
541

        token_type_id_requests = dict[int, Any]()
        for i, param in enumerate(pooling_params):
542
543
544
545
546
            if (
                param.extra_kwargs is not None
                and (token_types := param.extra_kwargs.get("compressed_token_type_ids"))
                is not None
            ):
547
548
549
550
551
                token_type_id_requests[i] = token_types

        if len(token_type_id_requests) == 0:
            return model_kwargs

552
        seq_lens = self.seq_lens.gpu[:num_reqs]
553
554
555
556
557
558
559
560
        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(
561
562
            device=self.device
        )
563
564
        return model_kwargs

565
    def _may_reorder_batch(self, scheduler_output: "SchedulerOutput") -> None:
566
567
        """
        Update the order of requests in the batch based on the attention
568
        backend's needs. For example, some attention backends (namely MLA) may
569
570
571
572
573
574
        want to separate requests based on if the attention computation will be
        compute-bound or memory-bound.

        Args:
            scheduler_output: The scheduler output.
        """
575
576
577
578
579
580
581
582
        # 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

583
        if self.reorder_batch_threshold is not None:
584
585
586
            # 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.
587
588
589
590
            if (
                self.dcp_world_size > 1
                and envs.VLLM_ATTENTION_BACKEND != "FLASH_ATTN_MLA"
            ):
591
                assert self.reorder_batch_threshold == 1, (
592
                    "DCP not support reorder_batch_threshold > 1 now."
593
                )
594
595
596
            reorder_batch_to_split_decodes_and_prefills(
                self.input_batch,
                scheduler_output,
597
598
                decode_threshold=self.reorder_batch_threshold,
            )
599

600
601
    # Note: used for model runner override.
    def _init_device_properties(self) -> None:
602
        """Initialize attributes from torch.cuda.get_device_properties"""
603
604
605
606
607
608
609
        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()

610
    def _update_states(self, scheduler_output: "SchedulerOutput") -> None:
611
612
613
614
615
616
        """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.

617
618
        The SamplingMetadata is updated and copied to the GPU if there is a
        new/resumed/paused/finished request in the batch.
619
620
        """
        # Remove finished requests from the cached states.
621
622
        for req_id in scheduler_output.finished_req_ids:
            self.requests.pop(req_id, None)
623
624
625
626
627
628
629
        # 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:
630
            self.input_batch.remove_request(req_id)
631
632

        # Free the cached encoder outputs.
633
634
        for mm_hash in scheduler_output.free_encoder_mm_hashes:
            self.encoder_cache.pop(mm_hash, None)
635

636
637
638
639
640
641
642
643
644
645
646
647
648
        # 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:
649
            self.input_batch.remove_request(req_id)
650

651
        reqs_to_add: list[CachedRequestState] = []
652
        # Add new requests to the cached states.
653
654
655
        for new_req_data in scheduler_output.scheduled_new_reqs:
            req_id = new_req_data.req_id
            sampling_params = new_req_data.sampling_params
656
            pooling_params = new_req_data.pooling_params
657

658
659
660
661
            if (
                sampling_params
                and sampling_params.sampling_type == SamplingType.RANDOM_SEED
            ):
662
663
664
665
666
                generator = torch.Generator(device=self.device)
                generator.manual_seed(sampling_params.seed)
            else:
                generator = None

667
668
            if self.is_pooling_model:
                assert pooling_params is not None
669
670
                task = pooling_params.task
                assert task is not None, "You did not set `task` in the API"
671

672
                model = cast(VllmModelForPooling, self.get_model())
673
                to_update = model.pooler.get_pooling_updates(task)
674
675
                to_update.apply(pooling_params)

676
            req_state = CachedRequestState(
677
                req_id=req_id,
678
                prompt_token_ids=new_req_data.prompt_token_ids,
679
                prompt_embeds=new_req_data.prompt_embeds,
680
                mm_features=new_req_data.mm_features,
681
                sampling_params=sampling_params,
682
                pooling_params=pooling_params,
683
                generator=generator,
684
685
                block_ids=new_req_data.block_ids,
                num_computed_tokens=new_req_data.num_computed_tokens,
686
                output_token_ids=[],
687
                lora_request=new_req_data.lora_request,
688
            )
689
690
            self.requests[req_id] = req_state

691
            # Only relevant for models using M-RoPE (e.g, Qwen2-VL)
692
            if self.uses_mrope:
693
                self._init_mrope_positions(req_state)
694

695
            reqs_to_add.append(req_state)
696

697
        # Update the states of the running/resumed requests.
698
        is_last_rank = get_pp_group().is_last_rank
699
700
        req_data = scheduler_output.scheduled_cached_reqs
        for i, req_id in enumerate(req_data.req_ids):
701
            req_state = self.requests[req_id]
702
703
704
            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]
705
            num_output_tokens = req_data.num_output_tokens[i]
706

707
            # Update the cached states.
708

709
            req_state.num_computed_tokens = num_computed_tokens
710
            req_index = self.input_batch.req_id_to_index.get(req_id)
711
712
713
714
715
716
717
718

            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.
719
720
721
                num_new_tokens = (
                    num_computed_tokens + len(new_token_ids) - req_state.num_tokens
                )
722
723
724
725
                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:
726
                    req_state.output_token_ids.extend(new_token_ids[-num_new_tokens:])
727
728
729
730
731
            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:
732
733
734
735
                    end_idx = (
                        self.input_batch.num_prompt_tokens[req_index]
                        + num_output_tokens
                    )
736
737
                    self.input_batch.num_tokens[req_index] = end_idx
                    self.input_batch.num_tokens_no_spec[req_index] = end_idx
738

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

752
753
754
755
756
757
                if self.use_async_scheduling and num_output_tokens > 0:
                    # We must recover the output token ids for resumed requests in the
                    # async scheduling case, so that correct input_ids are obtained.
                    resumed_token_ids = req_data.resumed_req_token_ids[i]
                    assert resumed_token_ids is not None
                    req_state.output_token_ids = resumed_token_ids[-num_output_tokens:]
758
759
760
761
            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.
762
                reqs_to_add.append(req_state)
763
764
765
                continue

            # Update the persistent batch.
766
            self.input_batch.num_computed_tokens_cpu[req_index] = num_computed_tokens
767
            if new_block_ids is not None:
768
                self.input_batch.block_table.append_row(new_block_ids, req_index)
769
770
771
772
773
774
775

            # 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)
776
                self.input_batch.token_ids_cpu[
777
778
779
                    req_index, start_token_index:end_token_index
                ] = new_token_ids
                self.input_batch.num_tokens_no_spec[req_index] = end_token_index
780
                self.input_batch.num_tokens[req_index] = end_token_index
781

782
            # Add spec_token_ids to token_ids_cpu.
783
            spec_token_ids = scheduler_output.scheduled_spec_decode_tokens.get(
784
                req_id, []
785
            )
786
787
788
789
790
            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[
791
792
                    req_index, start_index:end_token_index
                ] = spec_token_ids
793
794
                # NOTE(woosuk): `num_tokens` here may include spec tokens.
                self.input_batch.num_tokens[req_index] += num_spec_tokens
795
796
797
798
799
800
801

            # When speculative decoding is used with structured output,
            # the scheduler can drop draft tokens that do not
            # conform to the schema. This can result in
            # scheduler_output.scheduled_spec_decode_tokens being empty,
            # even when speculative decoding is enabled.
            self.input_batch.spec_token_ids[req_index] = spec_token_ids
802

803
804
        # Add the new or resumed requests to the persistent batch.
        # The smaller empty indices are filled first.
805
806
        for request in reqs_to_add:
            self.input_batch.add_request(request)
807

808
809
810
811
812
813
        # 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()
814

815
    def _update_states_after_model_execute(
816
817
        self, output_token_ids: torch.Tensor
    ) -> None:
818
819
820
821
822
823
824
825
826
827
828
829
        """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.
830
831
832
833
834
835
836
837
838
839
840
841
842
843
844
845
846
847
848
849
        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()
        )
850
851
852
        for i, num_tokens in enumerate(num_accepted_tokens):
            self.input_batch.num_accepted_tokens_cpu[i] = num_tokens

853
854
855
856
857
858
    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
859
860
861
862
        for mm_feature in req_state.mm_features:
            mm_item = mm_feature.data
            if mm_item is None:
                continue
863
864
865
866
867
868
869
870
871
872
873
874
            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

875
876
877
878
879
880
881
882
883
884
885
        assert supports_mrope(self.get_model()), "M-RoPE support is not implemented."

        req_state.mrope_positions, req_state.mrope_position_delta = (
            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,
886
            )
887
        )
888

889
    def _extract_mm_kwargs(
890
        self,
891
892
        scheduler_output: "SchedulerOutput",
    ) -> BatchedTensorInputs:
893
        if not scheduler_output or not self.is_multimodal_raw_input_only_model:
894
            return {}
895

896
897
        mm_kwargs = list[MultiModalKwargsItem]()
        for req in scheduler_output.scheduled_new_reqs:
898
899
900
            for feature in req.mm_features:
                if feature.data is not None:
                    mm_kwargs.append(feature.data)
901

902
        # Input all modalities at once
903
        model = cast(SupportsMultiModal, self.model)
904
905
        mm_kwargs_combined: BatchedTensorInputs = {}
        for _, _, mm_kwargs_group in group_mm_kwargs_by_modality(
906
907
908
909
            mm_kwargs,
            device=self.device,
            pin_memory=self.pin_memory,
            merge_by_field_config=model.merge_by_field_config,
910
911
        ):
            mm_kwargs_combined.update(mm_kwargs_group)
912

913
        return mm_kwargs_combined
914

915
    def _dummy_mm_kwargs(self, num_seqs: int) -> BatchedTensorInputs:
916
        if not self.is_multimodal_raw_input_only_model:
917
            return {}
918

919
920
921
922
923
        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)
924

925
926
927
    def _get_cumsum_and_arange(
        self,
        num_tokens: np.ndarray,
928
        cumsum_dtype: np.dtype | None = None,
929
930
931
932
933
934
935
936
937
938
939
940
941
942
943
944
    ) -> 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

945
946
947
    def _prepare_input_ids(
        self, total_num_scheduled_tokens: int, cu_num_tokens: np.ndarray
    ) -> None:
948
        """Prepare the input IDs for the current batch.
949

950
951
952
953
954
955
956
        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)
957
958
959
            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)
960
961
962
963
964
965
966
967
968
969
970
971
972
973
974
975
976
977
            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)
978
                indices_match &= prev_index == flattened_index
979
980
981
982
983
984
                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)
985
986
987
            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)
988
989
        if num_commmon_tokens == 0:
            # No requests in common with the previous iteration
990
            # So input_ids.cpu will have all the input ids.
991
992
993
994
995
996
997
            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_(
998
999
1000
                self.input_batch.prev_sampled_token_ids[:num_commmon_tokens, 0],
                non_blocking=True,
            )
1001
1002
            if self.enable_prompt_embeds:
                self.is_token_ids.gpu[:num_commmon_tokens] = True
1003
            return
1004
        # Upload the index tensors asynchronously so the scatter can be non-blocking.
1005
1006
1007
        input_ids_index_tensor = torch.tensor(
            flattened_indices, dtype=torch.int64, pin_memory=self.pin_memory
        ).to(self.device, non_blocking=True)
1008
        prev_common_req_indices_tensor = torch.tensor(
1009
1010
            prev_common_req_indices, dtype=torch.int64, pin_memory=self.pin_memory
        ).to(self.device, non_blocking=True)
1011
1012
1013
1014
        self.input_ids.gpu.scatter_(
            dim=0,
            index=input_ids_index_tensor,
            src=self.input_batch.prev_sampled_token_ids[
1015
1016
1017
                prev_common_req_indices_tensor, 0
            ],
        )
1018

1019
1020
1021
1022
1023
    def _get_encoder_seq_lens(
        self,
        scheduler_output: "SchedulerOutput",
        kv_cache_spec: KVCacheSpec,
        num_reqs: int,
1024
    ) -> np.ndarray | None:
1025
1026
1027
1028
1029
1030
1031
1032
1033
1034
1035
1036
        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

1037
    def _prepare_inputs(
1038
        self, scheduler_output: "SchedulerOutput"
1039
1040
1041
    ) -> tuple[
        PerLayerAttnMetadata,
        torch.Tensor,
1042
        SpecDecodeMetadata | None,
1043
        np.ndarray,
1044
        CommonAttentionMetadata | None,
1045
        int,
1046
1047
        UBatchSlices | None,
        torch.Tensor | None,
1048
1049
        bool,
    ]:
1050
1051
1052
        """
        :return: tuple[
            attn_metadata: layer-to-attention_metadata mapping,
1053
1054
1055
            logits_indices, spec_decode_metadata,
            num_scheduled_tokens, spec_decode_common_attn_metadata,
            max_num_scheduled_tokens, use_cascade_attn
1056
1057
        ]
        """
1058
1059
1060
1061
1062
1063
1064
        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.
1065
        self.input_batch.block_table.commit_block_table(num_reqs)
1066
1067

        # Get the number of scheduled tokens for each request.
1068
1069
1070
1071
        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)
1072
1073
1074

        # Get request indices.
        # E.g., [2, 5, 3] -> [0, 0, 1, 1, 1, 1, 1, 2, 2, 2]
1075
        req_indices = np.repeat(self.arange_np[:num_reqs], num_scheduled_tokens)
1076

1077
1078
        # cu_num_tokens: [2, 5, 3] -> [2, 7, 10]
        # arange: [0, 1, 0, 1, 2, 3, 4, 0, 1, 2]
1079
        cu_num_tokens, arange = self._get_cumsum_and_arange(num_scheduled_tokens)
1080
1081

        # Get positions.
1082
        positions_np = self.positions.np[:total_num_scheduled_tokens]
1083
1084
1085
1086
1087
        np.add(
            self.input_batch.num_computed_tokens_cpu[req_indices],
            arange,
            out=positions_np,
        )
1088

1089
1090
        # Calculate M-RoPE positions.
        # Only relevant for models using M-RoPE (e.g, Qwen2-VL)
1091
        if self.uses_mrope:
1092
1093
            self._calc_mrope_positions(scheduler_output)

1094
1095
1096
1097
        # 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.
1098
1099
1100
        token_indices = (
            positions_np + req_indices * self.input_batch.token_ids_cpu.shape[1]
        )
1101
        token_indices_tensor = torch.from_numpy(token_indices)
1102

1103
1104
1105
        # 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.
1106
1107
1108
1109
1110
1111
        torch.index_select(
            self.input_batch.token_ids_cpu_tensor.flatten(),
            0,
            token_indices_tensor,
            out=self.input_ids.cpu[:total_num_scheduled_tokens],
        )
1112
1113
1114
1115
1116
1117
        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,
1118
1119
                out=self.is_token_ids.cpu[:total_num_scheduled_tokens],
            )
1120
1121
1122
1123
1124
1125
1126
1127
1128
1129
1130
1131
1132
1133
1134
1135
1136
1137
1138
1139
1140
1141
1142
1143
1144
1145
1146
1147
1148
1149
1150
1151
1152

        # 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:
1153
1154
1155
                    self.inputs_embeds.cpu[
                        output_idx : output_idx + actual_num_sched
                    ].copy_(req_embeds[start_pos:actual_end])
1156
1157

                output_idx += num_sched
1158

1159
1160
        self.input_batch.block_table.compute_slot_mapping(req_indices, positions_np)
        self.input_batch.block_table.commit_slot_mapping(total_num_scheduled_tokens)
1161
1162

        # Prepare the attention metadata.
1163
        self.query_start_loc.np[0] = 0
1164
        self.query_start_loc.np[1 : num_reqs + 1] = cu_num_tokens
1165
1166
        # Note: pad query_start_loc to be non-decreasing, as kernels
        # like FlashAttention requires that
1167
        self.query_start_loc.np[num_reqs + 1 :].fill(cu_num_tokens[-1])
1168
        self.query_start_loc.copy_to_gpu()
1169
        query_start_loc = self.query_start_loc.gpu[: num_reqs + 1]
1170

1171
        num_tokens_unpadded = scheduler_output.total_num_scheduled_tokens
1172
        num_tokens_padded = self._get_num_input_tokens(num_tokens_unpadded)
1173
1174
1175
        uniform_decode = (
            max_num_scheduled_tokens == self.uniform_decode_query_len
        ) and (total_num_scheduled_tokens == num_reqs * max_num_scheduled_tokens)
1176
1177
1178
1179
1180
1181
1182

        # 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

1183
        ubatch_slices, num_tokens_across_dp = coordinate_batch_across_dp(
1184
1185
1186
1187
1188
1189
1190
            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,
1191
        )
1192

1193
        self.seq_lens.np[:num_reqs] = (
1194
1195
            self.input_batch.num_computed_tokens_cpu[:num_reqs] + num_scheduled_tokens
        )
1196
        # Fill unused with 0 for full cuda graph mode.
1197
1198
1199
1200
        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()
1201

1202
        num_tokens = [self.requests[r].num_tokens for r in self.input_batch.req_ids]
1203
1204
1205
1206
1207
1208
1209
        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)
1210
1211
1212
        self.discard_request_indices.np[: self.num_discarded_requests] = (
            discard_request_indices
        )
1213
1214
1215

        self.discard_request_indices.copy_to_gpu(self.num_discarded_requests)

1216
        # Copy the tensors to the GPU.
1217
1218
        self._prepare_input_ids(total_num_scheduled_tokens, cu_num_tokens)

1219
        if self.uses_mrope:
1220
            # Only relevant for models using M-RoPE (e.g, Qwen2-VL)
1221
1222
            self.mrope_positions.gpu[:, :total_num_scheduled_tokens].copy_(
                self.mrope_positions.cpu[:, :total_num_scheduled_tokens],
1223
1224
                non_blocking=True,
            )
1225
1226
        else:
            # Common case (1D positions)
1227
            self.positions.copy_to_gpu(total_num_scheduled_tokens)
1228

1229
        use_spec_decode = len(scheduler_output.scheduled_spec_decode_tokens) > 0
1230
1231
1232
1233
1234
1235
1236
        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
1237
            num_draft_tokens = None
1238
1239
1240
1241
1242
1243
            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)
1244
1245
1246
            # 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)
1247
1248
1249
1250
            for (
                req_id,
                draft_token_ids,
            ) in scheduler_output.scheduled_spec_decode_tokens.items():
1251
1252
                req_idx = self.input_batch.req_id_to_index[req_id]
                num_draft_tokens[req_idx] = len(draft_token_ids)
1253
1254
1255
1256
1257
1258
1259
1260
                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
                )
1261
            spec_decode_metadata = self._calc_spec_decode_metadata(
1262
1263
                num_draft_tokens, cu_num_tokens
            )
1264
            logits_indices = spec_decode_metadata.logits_indices
1265
1266

            # For DECODE only cuda graph of some attention backends (e.g., GDN).
1267
            self.num_decode_draft_tokens.np[:num_reqs] = num_decode_draft_tokens
1268
1269
            self.num_decode_draft_tokens.np[num_reqs:].fill(-1)
            self.num_decode_draft_tokens.copy_to_gpu()
1270
1271
1272

        logits_indices_padded = None
        if self.cache_config.kv_sharing_fast_prefill:
1273
            logits_indices_padded = self._prepare_kv_sharing_fast_prefill(
1274
1275
                logits_indices
            )
1276

1277
1278
1279
        attn_metadata: PerLayerAttnMetadata = {}
        if ubatch_slices is not None:
            attn_metadata = [dict() for _ in range(len(ubatch_slices))]
1280
        use_cascade_attn = False
1281

1282
        # Used in the below loop.
1283
        query_start_loc_cpu = self.query_start_loc.cpu[: num_reqs + 1]
1284
        seq_lens_cpu = self.seq_lens.cpu[:num_reqs]
1285
1286
1287
        num_computed_tokens_cpu = self.input_batch.num_computed_tokens_cpu_tensor[
            :num_reqs
        ]
1288
        spec_decode_common_attn_metadata = None
1289
1290
        if use_spec_decode:
            self.num_accepted_tokens.np[:num_reqs] = (
1291
1292
                self.input_batch.num_accepted_tokens_cpu[:num_reqs]
            )
1293
1294
            self.num_accepted_tokens.np[num_reqs:].fill(1)
            self.num_accepted_tokens.copy_to_gpu()
1295

1296
1297
1298
        # 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(
1299
1300
            self.kv_cache_config.kv_cache_groups
        ):
1301
            encoder_seq_lens = self._get_encoder_seq_lens(
1302
1303
                scheduler_output, kv_cache_group_spec.kv_cache_spec, num_reqs
            )
1304

1305
            if isinstance(kv_cache_group_spec.kv_cache_spec, EncoderOnlyAttentionSpec):
1306
1307
1308
1309
1310
                # 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,
1311
1312
1313
                    device=self.device,
                )
                slot_mapping = torch.zeros(
1314
                    (total_num_scheduled_tokens,),
1315
1316
1317
                    dtype=torch.int64,
                    device=self.device,
                )
1318
1319
1320
                num_common_prefix_blocks = 0
            else:
                blk_table = self.input_batch.block_table[kv_cache_group_id]
1321
                blk_table_tensor = blk_table.get_device_tensor(num_reqs)
1322
                slot_mapping = blk_table.slot_mapping.gpu[:total_num_scheduled_tokens]
1323
1324
1325

                # Fill unused with -1. Needed for reshape_and_cache in full cuda
                # graph mode.
1326
1327
1328
1329
                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
                ]
1330

1331
            common_attn_metadata = CommonAttentionMetadata(
1332
1333
1334
1335
1336
                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,
1337
1338
1339
                num_reqs=num_reqs,
                num_actual_tokens=total_num_scheduled_tokens,
                max_query_len=max_num_scheduled_tokens,
1340
                max_seq_len=max_seq_len,
1341
1342
                block_table_tensor=blk_table_tensor,
                slot_mapping=slot_mapping,
1343
1344
                logits_indices_padded=logits_indices_padded,
                num_logits_indices=logits_indices.size(0),
1345
                causal=True,
1346
                encoder_seq_lens=encoder_seq_lens,
1347
1348
1349
                dcp_local_seq_lens=self.dcp_local_seq_lens.gpu[:num_reqs]
                if self.dcp_world_size > 1
                else None,
1350
1351
            )

1352
            if self.speculative_config and spec_decode_common_attn_metadata is None:
1353
                if isinstance(self.drafter, EagleProposer):
1354
1355
1356
1357
                    if (
                        self.drafter.attn_layer_names[0]
                        in kv_cache_group_spec.layer_names
                    ):
1358
1359
1360
                        spec_decode_common_attn_metadata = common_attn_metadata
                else:
                    spec_decode_common_attn_metadata = common_attn_metadata
1361

1362
1363
1364
            for attn_group in self.attn_groups[kv_cache_group_id]:
                # Prepare for cascade attention if enabled & beneficial.
                common_prefix_len = 0
1365
                builder = attn_group.get_metadata_builder()
1366
1367
1368
                if self.cascade_attn_enabled:
                    common_prefix_len = self._compute_cascade_attn_prefix_len(
                        num_scheduled_tokens,
1369
                        num_common_prefix_blocks,
1370
                        attn_group.kv_cache_spec,
1371
1372
                        builder,
                    )
1373

1374
                extra_attn_metadata_args = {}
1375
                if use_spec_decode and isinstance(builder, GDNAttentionMetadataBuilder):
1376
                    extra_attn_metadata_args = dict(
1377
1378
1379
1380
                        num_accepted_tokens=self.num_accepted_tokens.gpu[:num_reqs],
                        num_decode_draft_tokens_cpu=self.num_decode_draft_tokens.cpu[
                            :num_reqs
                        ],
1381
1382
                    )

1383
1384
                if ubatch_slices is not None:
                    common_attn_metadata_list = split_attn_metadata(
1385
1386
                        ubatch_slices, common_attn_metadata
                    )
1387
                    for ubid, common_attn_metadata in enumerate(
1388
1389
1390
1391
1392
1393
1394
1395
                        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,
                        )
1396
1397
1398
1399
1400
1401
1402
1403
                        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,
1404
1405
1406
                        **extra_attn_metadata_args,
                    )
                    use_cascade_attn |= getattr(attn_metadata_i, "use_cascade", False)
1407
1408
                    for layer_name in attn_group.layer_names:
                        attn_metadata[layer_name] = attn_metadata_i
1409

1410
1411
1412
1413
        # disable cascade attention when DBO
        if ubatch_slices is not None:
            use_cascade_attn = False

1414
1415
1416
1417
        # Hot-Swap lora model
        if self.lora_config:
            self.set_active_loras(self.input_batch, num_scheduled_tokens)

1418
1419
1420
1421
1422
1423
1424
1425
        return (
            attn_metadata,
            logits_indices,
            spec_decode_metadata,
            num_scheduled_tokens,
            spec_decode_common_attn_metadata,
            max_num_scheduled_tokens,
            ubatch_slices,
1426
            num_tokens_across_dp,
1427
1428
            use_cascade_attn,
        )
1429

1430
1431
1432
1433
    def _compute_cascade_attn_prefix_len(
        self,
        num_scheduled_tokens: np.ndarray,
        num_common_prefix_blocks: int,
1434
1435
        kv_cache_spec: KVCacheSpec,
        attn_metadata_builder: AttentionMetadataBuilder,
1436
1437
1438
1439
1440
1441
1442
1443
1444
1445
1446
1447
1448
1449
1450
1451
1452
1453
    ) -> 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.
        """
1454
        common_prefix_len = num_common_prefix_blocks * kv_cache_spec.block_size
1455
1456
1457
1458
1459
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
        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]
1492
        # Request 3's num_computed_tokens: 3 (i.e., [A, B, C])
1493
1494
1495
1496
1497
1498
1499
        # 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(
1500
1501
            common_prefix_len, self.input_batch.num_computed_tokens_cpu[:num_reqs].min()
        )
1502
        # common_prefix_len should be a multiple of the block size.
1503
1504
1505
1506
1507
1508
1509
1510
1511
1512
1513
        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
        )
1514
1515
        assert isinstance(kv_cache_spec, AttentionSpec)
        use_cascade = attn_metadata_builder.use_cascade_attention(
1516
1517
1518
            common_prefix_len=common_prefix_len,
            query_lens=num_scheduled_tokens,
            num_query_heads=self.num_query_heads,
1519
            num_kv_heads=kv_cache_spec.num_kv_heads,
1520
            use_alibi=self.use_alibi,
1521
            use_sliding_window=use_sliding_window,
1522
            use_local_attention=use_local_attention,
1523
            num_sms=self.num_sms,
1524
            dcp_world_size=self.dcp_world_size,
1525
1526
1527
        )
        return common_prefix_len if use_cascade else 0

1528
1529
    def _calc_mrope_positions(self, scheduler_output: "SchedulerOutput"):
        mrope_pos_ptr = 0
1530
        for index, req_id in enumerate(self.input_batch.req_ids):
1531
1532
1533
            req = self.requests[req_id]
            assert req.mrope_positions is not None

1534
1535
            num_computed_tokens = self.input_batch.num_computed_tokens_cpu[index]
            num_scheduled_tokens = scheduler_output.num_scheduled_tokens[req_id]
1536
            num_prompt_tokens = length_from_prompt_token_ids_or_embeds(
1537
1538
                req.prompt_token_ids, req.prompt_embeds
            )
1539
1540

            if num_computed_tokens + num_scheduled_tokens > num_prompt_tokens:
1541
1542
                prompt_part_len = max(0, num_prompt_tokens - num_computed_tokens)
                completion_part_len = max(0, num_scheduled_tokens - prompt_part_len)
1543
1544
1545
1546
1547
1548
1549
1550
1551
1552
1553
1554
1555
            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

1556
1557
1558
                self.mrope_positions.cpu[:, dst_start:dst_end] = req.mrope_positions[
                    :, src_start:src_end
                ]
1559
1560
1561
1562
1563
1564
1565
                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

1566
                MRotaryEmbedding.get_next_input_positions_tensor(
1567
                    out=self.mrope_positions.np,
1568
1569
1570
1571
1572
                    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,
                )
1573
1574
1575

                mrope_pos_ptr += completion_part_len

1576
1577
    def _calc_spec_decode_metadata(
        self,
1578
1579
1580
1581
1582
1583
1584
1585
1586
1587
1588
1589
1590
1591
1592
1593
        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
1594
1595
1596
1597

        # 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(
1598
1599
            num_sampled_tokens, cumsum_dtype=np.int32
        )
1600
        # Step 2. [0, 0, 0, 0, 103, 104, 104, 104, 206, 207, 207]
1601
        logits_indices = np.repeat(
1602
1603
            cu_num_scheduled_tokens - num_sampled_tokens, num_sampled_tokens
        )
1604
        # Step 3. [0, 1, 2, 3, 103, 104, 105, 106, 206, 207, 208]
1605
1606
1607
1608
1609
1610
        logits_indices += arange

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

        # Compute the draft logits indices.
1611
1612
1613
        # 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(
1614
1615
            num_draft_tokens, cumsum_dtype=np.int32
        )
1616
1617
        # [0, 0, 0, 5, 5, 9]
        target_logits_indices = np.repeat(
1618
1619
            cu_num_sampled_tokens - num_sampled_tokens, num_draft_tokens
        )
1620
1621
1622
1623
1624
        # [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(
1625
1626
1627
1628
1629
            self.device, non_blocking=True
        )
        logits_indices = torch.from_numpy(logits_indices).to(
            self.device, non_blocking=True
        )
1630
        target_logits_indices = torch.from_numpy(target_logits_indices).to(
1631
1632
            self.device, non_blocking=True
        )
1633
        bonus_logits_indices = torch.from_numpy(bonus_logits_indices).to(
1634
1635
            self.device, non_blocking=True
        )
1636

1637
1638
        # Compute the draft token ids.
        # draft_token_indices:      [  1,   2,   3, 105, 106, 208]
1639
        draft_token_ids = self.input_ids.gpu[logits_indices]
1640
1641
1642
1643
1644
1645
1646
1647
1648
1649
1650
1651
        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

1652
1653
1654
1655
1656
1657
1658
    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
1659
        self.kv_sharing_fast_prefill_logits_indices[:num_logits].copy_(logits_indices)
1660
1661
1662
1663
1664
        # 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_(
1665
1666
1667
1668
1669
1670
            logits_indices[-1].item()
        )
        if (
            self.compilation_config.cudagraph_mode != CUDAGraphMode.NONE
            and num_logits <= self.cudagraph_batch_sizes[-1]
        ):
1671
1672
1673
1674
1675
            # 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
1676
1677
1678
        logits_indices_padded = self.kv_sharing_fast_prefill_logits_indices[
            :num_logits_padded
        ]
1679
1680
        return logits_indices_padded

1681
1682
1683
1684
1685
1686
1687
1688
    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
1689
                inputs.
1690
1691
1692
1693
1694
1695

        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
        """
1696
1697
        scheduled_encoder_inputs = scheduler_output.scheduled_encoder_inputs
        if not scheduled_encoder_inputs:
1698
            return [], []
1699
        # Batch the multi-modal inputs.
1700
        mm_kwargs = list[MultiModalKwargsItem]()
1701
1702
        # list of tuple (mm_hash, position_info)
        mm_hashes_pos = list[tuple[str, PlaceholderRange]]()
1703
1704
        for req_id, encoder_input_ids in scheduled_encoder_inputs.items():
            req_state = self.requests[req_id]
1705
1706

            for mm_input_id in encoder_input_ids:
1707
1708
1709
1710
                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))
1711

1712
1713
1714
1715
1716
        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(
1717
1718
            scheduler_output
        )
1719
1720
1721
1722

        if not mm_kwargs:
            return

1723
1724
1725
1726
1727
1728
1729
        # 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.
1730
        model = cast(SupportsMultiModal, self.model)
1731
        encoder_outputs = []
1732
        for modality, num_items, mm_kwargs_group in group_mm_kwargs_by_modality(
1733
1734
1735
1736
            mm_kwargs,
            device=self.device,
            pin_memory=self.pin_memory,
            merge_by_field_config=model.merge_by_field_config,
1737
        ):
1738
1739
1740
1741
1742
1743
1744
1745
1746
1747
1748
1749
            # (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(
1750
1751
1752
                        (k, v[i : i + micro_batch_size])
                        for k, v in mm_kwargs_group.items()
                    )
1753
1754

                    micro_batch_outputs = model.get_multimodal_embeddings(
1755
1756
                        **micro_batch_mm_inputs
                    )
1757
1758
1759
1760
1761
1762
1763
1764
1765
1766

                    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.
1767
                curr_group_outputs = model.get_multimodal_embeddings(**mm_kwargs_group)
1768

1769
1770
            sanity_check_mm_encoder_outputs(
                curr_group_outputs,
1771
                expected_num_items=num_items,
1772
            )
1773
            encoder_outputs.extend(curr_group_outputs)
1774

1775
1776
1777
        # 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(
1778
1779
1780
1781
1782
                output,
                is_embed=pos_info.is_embed,
            )

    def _gather_mm_embeddings(
1783
1784
        self,
        scheduler_output: "SchedulerOutput",
1785
        shift_computed_tokens: int = 0,
1786
1787
1788
1789
1790
1791
1792
1793
    ) -> 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
1794
        should_sync_mrope_positions = False
1795

1796
        for req_id in self.input_batch.req_ids:
1797
1798
            mm_embeds_req: list[torch.Tensor] = []

1799
            num_scheduled_tokens = scheduler_output.num_scheduled_tokens[req_id]
1800
            req_state = self.requests[req_id]
1801
            num_computed_tokens = req_state.num_computed_tokens + shift_computed_tokens
1802

1803
1804
            for mm_feature in req_state.mm_features:
                pos_info = mm_feature.mm_position
1805
1806
                start_pos = pos_info.offset
                num_encoder_tokens = pos_info.length
1807
1808
1809
1810
1811
1812
1813
1814
1815
1816
1817
1818
1819
1820
1821
1822

                # 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,
1823
1824
                    num_encoder_tokens,
                )
1825
                assert start_idx < end_idx
1826

1827
                mm_hash = mm_feature.identifier
1828
                encoder_output = self.encoder_cache.get(mm_hash, None)
1829
                assert encoder_output is not None, f"Encoder cache miss for {mm_hash}."
1830
1831
1832
1833

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

1834
                req_start_pos = req_start_idx + start_pos - num_computed_tokens
1835
1836
1837
                is_mm_embed[req_start_pos + start_idx : req_start_pos + end_idx] = (
                    True if is_embed is None else is_embed
                )
1838

1839
1840
1841
1842
                mm_embeds_item = gather_mm_placeholders(
                    encoder_output[start_idx:end_idx],
                    is_embed=is_embed,
                )
1843
1844
1845
                mm_embeds_req.append(mm_embeds_item)

            if self.is_multimodal_pruning_enabled and self.uses_mrope:
1846
                assert req_state.mrope_positions is not None
1847
1848
1849
1850
1851
1852
1853
                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,
1854
1855
                    )
                )
1856
1857
1858
1859
                req_state.mrope_positions.copy_(new_mrope_positions)
                req_state.mrope_position_delta = new_delta

            mm_embeds.extend(mm_embeds_req)
1860
1861
1862
            req_start_idx += num_scheduled_tokens

        is_mm_embed = self.is_mm_embed.copy_to_gpu(total_num_scheduled_tokens)
1863
1864
1865

        if should_sync_mrope_positions:
            self._calc_mrope_positions(scheduler_output)
1866
            self.mrope_positions.copy_to_gpu(total_num_scheduled_tokens)
1867

1868
        return mm_embeds, is_mm_embed
1869

1870
1871
1872
1873
1874
1875
1876
1877
1878
1879
1880
1881
1882
1883
1884
1885
    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
1886
        model = cast(SupportsMultiModal, self.model)
1887
1888
        encoder_features = {}
        for _, _, mm_kwargs_group in group_mm_kwargs_by_modality(
1889
1890
1891
1892
            mm_kwargs,
            device=self.device,
            pin_memory=self.pin_memory,
            merge_by_field_config=model.merge_by_field_config,
1893
1894
1895
1896
1897
1898
1899
1900
        ):
            # 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

1901
    def get_model(self) -> nn.Module:
1902
        # get raw model out of the cudagraph wrapper.
1903
        if isinstance(self.model, (CUDAGraphWrapper, UBatchWrapper)):
1904
            return self.model.unwrap()
1905
1906
        return self.model

1907
1908
1909
1910
1911
1912
1913
1914
1915
1916
1917
1918
1919
1920
1921
    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

1922
1923
1924
1925
1926
    def get_supported_pooling_tasks(self) -> list[PoolingTask]:
        model = self.get_model()
        if not is_pooling_model(model):
            return []

1927
1928
        supported_tasks = list(model.pooler.get_supported_tasks())

1929
1930
1931
1932
1933
        if self.scheduler_config.chunked_prefill_enabled:
            if "token_embed" in supported_tasks:
                supported_tasks.remove("token_embed")
            if "token_classify" in supported_tasks:
                supported_tasks.remove("token_classify")
1934

1935
1936
            logger.debug_once(
                "Chunked prefill is not supported with "
1937
1938
                "token_embed and token_classify tasks "
                "which using ALL pooling. "
1939
1940
1941
                "Please turn off chunked prefill by "
                "`--no-enable-chunked-prefill` before using it."
            )
1942
1943
1944
1945
1946

        if "score" in supported_tasks:
            num_labels = getattr(self.model_config.hf_config, "num_labels", 0)
            if num_labels != 1:
                supported_tasks.remove("score")
1947
                logger.debug_once("Score API is only enabled for num_labels == 1.")
1948
1949

        return supported_tasks
1950

1951
1952
1953
1954
1955
1956
1957
1958
1959
1960
    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)

1961
    def sync_and_slice_intermediate_tensors(
1962
1963
1964
1965
1966
        self,
        num_tokens: int,
        intermediate_tensors: IntermediateTensors,
        sync_self: bool,
    ) -> IntermediateTensors:
1967
1968
1969
        assert self.intermediate_tensors is not None

        tp = self.vllm_config.parallel_config.tensor_parallel_size
1970
        is_rs = is_residual_scattered_for_sp(self.vllm_config, num_tokens)
1971
1972
1973
1974
1975
1976

        # 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():
1977
                is_scattered = k == "residual" and is_rs
1978
                copy_len = num_tokens // tp if is_scattered else num_tokens
1979
                self.intermediate_tensors[k][:copy_len].copy_(
1980
1981
1982
1983
1984
1985
1986
1987
1988
1989
1990
1991
1992
                    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:
1993
1994
1995
1996
1997
1998
1999
        """
        Step for the EPLB (Expert Parallelism Load Balancing) state.
        """
        if not self.parallel_config.enable_eplb:
            return

        assert self.eplb_state is not None
2000
2001
        model = self.get_model()
        assert is_mixture_of_experts(model)
2002
        self.eplb_state.step(
2003
            model,
2004
2005
            is_dummy,
            is_profile,
2006
            log_stats=self.parallel_config.eplb_config.log_balancedness,
2007
2008
        )

2009
2010
2011
2012
    # 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)
2013
2014
    @staticmethod
    def pad_out_ubatch_slice(ubatch_slices: UBatchSlices, num_total_tokens: int):
2015
2016
2017
2018
2019
2020
        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
        )
2021

2022
2023
2024
2025
2026
2027
    def _pool(
        self,
        hidden_states: torch.Tensor,
        num_scheduled_tokens: int,
        num_scheduled_tokens_np: np.ndarray,
    ) -> ModelRunnerOutput:
2028
2029
2030
        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"
        )
2031

2032
        hidden_states = hidden_states[:num_scheduled_tokens]
2033
        pooling_metadata = self.input_batch.get_pooling_metadata()
2034
2035
2036
2037
        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]
2038

2039
2040
2041
2042
2043
2044
2045
2046
2047
2048
        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()
2049

2050
        pooler_output: list[torch.Tensor | None] = []
2051
        for raw_output, seq_len, prompt_len in zip(
2052
2053
            raw_pooler_output, seq_lens_cpu, pooling_metadata.prompt_lens
        ):
2054
            output = raw_output if seq_len == prompt_len else None
2055
            pooler_output.append(output)
2056
2057
2058
2059
2060
2061
2062
2063
2064
2065

        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,
        )

2066
    def _get_num_input_tokens(self, num_scheduled_tokens: int) -> int:
2067
2068
2069
2070
2071
2072
        if (
            self.compilation_config.cudagraph_mode != CUDAGraphMode.NONE
            and hasattr(self, "cudagraph_batch_sizes")
            and self.cudagraph_batch_sizes
            and num_scheduled_tokens <= self.cudagraph_batch_sizes[-1]
        ):
2073
2074
2075
2076
2077
2078
2079
2080
            # 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
2081
2082
2083
2084
        if (
            self.compilation_config.pass_config.enable_sequence_parallelism
            and tp_size > 1
        ):
2085
2086
2087
            return round_up(num_scheduled_tokens, tp_size)
        return num_scheduled_tokens

2088
    def _preprocess(
2089
2090
        self,
        scheduler_output: "SchedulerOutput",
2091
        num_input_tokens: int,  # Padded
2092
        intermediate_tensors: IntermediateTensors | None = None,
2093
2094
    ) -> tuple[
        int,
2095
2096
        torch.Tensor | None,
        torch.Tensor | None,
2097
        torch.Tensor,
2098
        IntermediateTensors | None,
2099
2100
        dict[str, Any],
    ]:
2101
        num_scheduled_tokens = scheduler_output.total_num_scheduled_tokens
2102
        is_first_rank = get_pp_group().is_first_rank
2103

2104
2105
        # _prepare_inputs may reorder the batch, so we must gather multi
        # modal outputs after that to ensure the correct order
2106
2107
        if (
            self.supports_mm_inputs
2108
            and is_first_rank
2109
2110
            and not self.model_config.is_encoder_decoder
        ):
2111
2112
            # Run the multimodal encoder if any.
            self._execute_mm_encoder(scheduler_output)
2113
            mm_embeds, is_mm_embed = self._gather_mm_embeddings(scheduler_output)
2114

2115
2116
2117
            # 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.
2118
            inputs_embeds_scheduled = self.model.get_input_embeddings(
2119
2120
2121
                self.input_ids.gpu[:num_scheduled_tokens],
                multimodal_embeddings=mm_embeds,
                is_multimodal=is_mm_embed,
2122
            )
2123

2124
            # TODO(woosuk): Avoid the copy. Optimize.
2125
            self.inputs_embeds.gpu[:num_scheduled_tokens].copy_(inputs_embeds_scheduled)
2126

2127
            input_ids = None
2128
            inputs_embeds = self.inputs_embeds.gpu[:num_input_tokens]
2129
2130
2131
2132
            model_kwargs = {
                **self._init_model_kwargs(num_scheduled_tokens),
                **self._extract_mm_kwargs(scheduler_output),
            }
2133
        elif self.enable_prompt_embeds and is_first_rank:
2134
2135
2136
2137
2138
2139
2140
2141
2142
2143
2144
2145
            # 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).
2146
2147
2148
            token_ids_idx = (
                self.is_token_ids.gpu[:num_scheduled_tokens]
                .nonzero(as_tuple=False)
2149
                .squeeze(1)
2150
            )
2151
2152
2153
            # Some tokens ids may need to become embeds
            if token_ids_idx.numel() > 0:
                token_ids = self.input_ids.gpu[token_ids_idx]
2154
                tokens_to_embeds = self.model.get_input_embeddings(input_ids=token_ids)
2155
2156
2157
2158
2159
                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
2160
        else:
2161
2162
2163
2164
            # 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.
2165
            input_ids = self.input_ids.gpu[:num_input_tokens]
2166
            inputs_embeds = None
2167
            model_kwargs = self._init_model_kwargs(num_input_tokens)
2168
        if self.uses_mrope:
2169
            positions = self.mrope_positions.gpu[:, :num_input_tokens]
2170
        else:
2171
            positions = self.positions.gpu[:num_input_tokens]
2172

2173
        if is_first_rank:
2174
2175
            intermediate_tensors = None
        else:
2176
            intermediate_tensors = self.sync_and_slice_intermediate_tensors(
2177
2178
                num_input_tokens, intermediate_tensors, True
            )
2179

2180
2181
2182
2183
        if (
            self.model_config.is_encoder_decoder
            and scheduler_output.scheduled_encoder_inputs
        ):
2184
2185
2186
            encoder_inputs = self._extract_encoder_inputs(scheduler_output)
            model_kwargs.update(encoder_inputs)

2187
2188
2189
2190
2191
2192
2193
2194
        return (
            num_scheduled_tokens,
            input_ids,
            inputs_embeds,
            positions,
            intermediate_tensors,
            model_kwargs,
        )
2195

2196
    def _sample(
2197
        self,
2198
2199
        logits: torch.Tensor | None,
        spec_decode_metadata: SpecDecodeMetadata | None,
2200
    ) -> SamplerOutput:
2201
        # Sample the next token and get logprobs if needed.
2202
        sampling_metadata = self.input_batch.sampling_metadata
2203
        if spec_decode_metadata is None:
2204
2205
2206
            # Update output token ids with tokens sampled in last step
            # if async scheduling and required by current sampling params.
            self.input_batch.update_async_output_token_ids()
2207
            return self.sampler(
2208
2209
2210
                logits=logits,
                sampling_metadata=sampling_metadata,
            )
2211

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
        # 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)
2238
2239
2240
        return sampler_output

    def _bookkeeping_sync(
2241
2242
2243
        self,
        scheduler_output: "SchedulerOutput",
        sampler_output: SamplerOutput,
2244
        logits: torch.Tensor | None,
2245
2246
        hidden_states: torch.Tensor,
        num_scheduled_tokens: int,
2247
    ) -> tuple[
2248
        dict[str, int],
2249
        LogprobsLists | None,
2250
        list[list[int]],
2251
        dict[str, LogprobsTensors | None],
2252
2253
2254
        list[str],
        dict[str, int],
        list[int],
2255
    ]:
2256
2257
2258
2259
        num_nans_in_logits = {}
        if envs.VLLM_COMPUTE_NANS_IN_LOGITS:
            num_nans_in_logits = self._get_nans_in_logits(logits)

2260
2261
2262
        discard_sampled_tokens_req_indices = self.discard_request_indices.np[
            : self.num_discarded_requests
        ]
2263
2264
2265
2266
        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)
2267

2268
2269
2270
        # 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()
2271
        req_id_to_index_output_copy = self.input_batch.req_id_to_index.copy()
2272

2273
2274
        # NOTE: GPU -> CPU Sync happens here.
        # Move as many CPU operations as possible before this sync point.
2275
        logprobs_tensors = sampler_output.logprobs_tensors
2276
2277
2278
        logprobs_lists = (
            logprobs_tensors.tolists() if logprobs_tensors is not None else None
        )
2279
2280
2281

        # Compute prompt logprobs if needed.
        prompt_logprobs_dict = self._get_prompt_logprobs_dict(
2282
            hidden_states[:num_scheduled_tokens],
2283
            scheduler_output.num_scheduled_tokens,
2284
2285
        )

2286
        num_sampled_tokens = sampler_output.sampled_token_ids.shape[0]
2287
        sampled_token_ids = sampler_output.sampled_token_ids
2288
        invalid_req_indices = []
2289
2290
2291
2292
2293
2294
2295
2296
2297
2298
2299
2300
2301
2302
        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:
2303
                valid_sampled_token_ids[int(i)].clear()
2304
        else:
2305
            valid_sampled_token_ids = []
2306
            invalid_req_indices = discard_sampled_tokens_req_indices.tolist()
2307
2308
2309
2310
2311
2312
            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.
2313
            self.input_batch.prev_sampled_token_ids = sampled_token_ids
2314
2315
2316
2317
2318
            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
            }
2319

2320
2321
2322
2323
2324
        # 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.
2325
        req_ids = self.input_batch.req_ids
2326
2327
        for req_idx in range(num_sampled_tokens):
            if self.use_async_scheduling:
2328
                sampled_ids = [-1] if req_idx not in invalid_req_indices_set else None
2329
2330
            else:
                sampled_ids = valid_sampled_token_ids[req_idx]
2331
2332
2333
2334
2335
            if not sampled_ids:
                continue

            start_idx = self.input_batch.num_tokens_no_spec[req_idx]
            end_idx = start_idx + len(sampled_ids)
2336
2337
2338
2339
            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}"
2340
            )
2341

2342
2343
            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
2344
2345
            self.input_batch.num_tokens_no_spec[req_idx] = end_idx
            self.input_batch.num_tokens[req_idx] = end_idx
2346

2347
            req_id = req_ids[req_idx]
2348
2349
2350
            req_state = self.requests[req_id]
            req_state.output_token_ids.extend(sampled_ids)

2351
2352
2353
2354
2355
2356
2357
2358
2359
2360
        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,
        )

2361
2362
2363
2364
2365
2366
2367
2368
2369
2370
2371
2372
2373
2374
2375
    @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()

2376
2377
    def _model_forward(
        self,
2378
2379
2380
2381
        input_ids: torch.Tensor | None = None,
        positions: torch.Tensor | None = None,
        intermediate_tensors: IntermediateTensors | None = None,
        inputs_embeds: torch.Tensor | None = None,
2382
2383
2384
2385
2386
        **model_kwargs: dict[str, Any],
    ) -> Any:
        """Helper method to call the model forward pass.

        This method can be overridden by subclasses for model execution.
2387
        Motivation: We can inspect only this method versus
2388
2389
2390
2391
2392
2393
2394
2395
2396
2397
2398
2399
2400
2401
2402
2403
2404
2405
2406
2407
        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,
        )

2408
2409
2410
2411
    @torch.inference_mode()
    def execute_model(
        self,
        scheduler_output: "SchedulerOutput",
2412
2413
        intermediate_tensors: IntermediateTensors | None = None,
    ) -> ModelRunnerOutput | AsyncModelRunnerOutput | IntermediateTensors:
2414
        with record_function_or_nullcontext("Preprocess"):
2415
2416
2417
2418
2419
2420
2421
2422
2423
            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(
2424
2425
                        scheduler_output, self.vllm_config
                    )
2426
2427
2428
2429
                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 "
2430
2431
                        "it when the requests need prompt logprobs"
                    )
2432

2433
                # Prepare the decoder inputs.
2434
2435
2436
2437
2438
2439
2440
2441
                (
                    attn_metadata,
                    logits_indices,
                    spec_decode_metadata,
                    num_scheduled_tokens_np,
                    spec_decode_common_attn_metadata,
                    max_query_len,
                    ubatch_slices,
2442
                    num_tokens_across_dp,
2443
2444
                    use_cascade_attn,
                ) = self._prepare_inputs(scheduler_output)
2445

2446
            dp_rank = self.parallel_config.data_parallel_rank
2447
2448
            if ubatch_slices:
                assert num_tokens_across_dp is not None
2449
                num_input_tokens = int(num_tokens_across_dp[dp_rank].item())
2450
2451
                self.pad_out_ubatch_slice(ubatch_slices, num_input_tokens)
            elif num_tokens_across_dp is not None:
2452
                num_input_tokens = int(num_tokens_across_dp[dp_rank].item())
2453
2454
2455
2456
2457
            else:
                num_input_tokens = self._get_num_input_tokens(
                    scheduler_output.total_num_scheduled_tokens
                )

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

            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(
2473
2474
2475
                num_tokens=num_input_tokens,
                uniform_decode=uniform_decode,
                has_lora=len(self.input_batch.lora_id_to_lora_request) > 0,
2476
2477
2478
2479
            )
            cudagraph_runtime_mode, batch_descriptor = (
                self.cudagraph_dispatcher.dispatch(batch_descriptor, use_cascade_attn)
            )
2480

2481
2482
        # Set cudagraph mode to none if calc_kv_scales is true.
        if attn_metadata is not None:
2483
2484
2485
2486
2487
            metadata_list = (
                attn_metadata.values()
                if isinstance(attn_metadata, dict)
                else [attn_metadata]
            )
2488
            if any(
2489
2490
                getattr(m, "enable_kv_scales_calculation", False) for m in metadata_list
            ):
2491
2492
                cudagraph_runtime_mode = CUDAGraphMode.NONE

2493
2494
        # Run the model.
        # Use persistent buffers for CUDA graphs.
2495
2496
        with (
            set_forward_context(
2497
2498
2499
2500
2501
2502
                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,
2503
                ubatch_slices=ubatch_slices,
2504
2505
2506
2507
            ),
            record_function_or_nullcontext("Forward"),
            self.maybe_get_kv_connector_output(scheduler_output) as kv_connector_output,
        ):
2508
            model_output = self._model_forward(
2509
2510
2511
2512
2513
2514
2515
2516
2517
                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:
2518
                # True when EAGLE 3 is used.
2519
2520
                hidden_states, aux_hidden_states = model_output
            else:
2521
                # Common case.
2522
2523
2524
                hidden_states = model_output
                aux_hidden_states = None

2525
2526
2527
2528
2529
            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)
2530
2531
                    hidden_states.kv_connector_output = kv_connector_output
                    return hidden_states
2532

2533
                if self.is_pooling_model:
2534
                    # Return the pooling output.
2535
2536
2537
                    output = self._pool(
                        hidden_states, num_scheduled_tokens, num_scheduled_tokens_np
                    )
2538
2539
                    output.kv_connector_output = kv_connector_output
                    return output
2540
2541

                sample_hidden_states = hidden_states[logits_indices]
2542
                logits = self.model.compute_logits(sample_hidden_states)
2543
2544
2545
2546
2547
            else:
                # Rare case.
                assert not self.is_pooling_model

                if not get_pp_group().is_last_rank:
2548
                    all_gather_tensors = {
2549
2550
2551
                        "residual": not is_residual_scattered_for_sp(
                            self.vllm_config, num_input_tokens
                        )
2552
                    }
2553
                    get_pp_group().send_tensor_dict(
2554
2555
                        hidden_states.tensors,
                        all_gather_group=get_tp_group(),
2556
2557
                        all_gather_tensors=all_gather_tensors,
                    )
2558
2559
2560
                    logits = None
                else:
                    sample_hidden_states = hidden_states[logits_indices]
2561
                    logits = self.model.compute_logits(sample_hidden_states)
2562
2563
2564
2565
2566

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

2567
2568
2569
                model_output_broadcast_data = get_pp_group().broadcast_tensor_dict(
                    model_output_broadcast_data, src=len(get_pp_group().ranks) - 1
                )
2570
2571
2572
2573
                assert model_output_broadcast_data is not None
                logits = model_output_broadcast_data["logits"]

            # Apply structured output bitmasks if present
2574
2575
            if scheduler_output.structured_output_request_ids:
                apply_grammar_bitmask(scheduler_output, self.input_batch, logits)
2576
2577
2578
2579

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

2580
2581
2582
2583
2584
2585
2586
2587
2588
2589
2590
2591
2592
2593
        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,
                )

2594
2595
2596
2597
2598
        use_padded_batch_for_eagle = (
            self.speculative_config
            and self.speculative_config.use_eagle()
            and not self.speculative_config.disable_padded_drafter_batch
        )
2599
2600
2601
        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
2602
2603
2604
2605
2606
        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
        ):
2607
            effective_drafter_max_model_len = (
2608
2609
                self.speculative_config.draft_model_config.max_model_len
            )
2610
        input_fits_in_drafter = spec_decode_common_attn_metadata and (
2611
2612
2613
2614
            spec_decode_common_attn_metadata.max_seq_len
            + self.speculative_config.num_speculative_tokens
            <= effective_drafter_max_model_len
        )
2615
        if use_padded_batch_for_eagle and input_fits_in_drafter:
2616
2617
2618
2619
            # 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)

2620
2621
2622
2623
2624
2625
2626
2627
2628
        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,
2629
2630
2631
2632
2633
2634
2635
            ) = self._bookkeeping_sync(
                scheduler_output,
                sampler_output,
                logits,
                hidden_states,
                num_scheduled_tokens,
            )
2636

2637
2638
2639
2640
2641
        if (
            self.speculative_config
            and not use_padded_batch_for_eagle
            and input_fits_in_drafter
        ):
2642
2643
2644
            # 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)
2645

2646
2647
        with record_function_or_nullcontext("EPLB"):
            self.eplb_step()
2648

2649
2650
2651
        output = ModelRunnerOutput(
            req_ids=req_ids_output_copy,
            req_id_to_index=req_id_to_index_output_copy,
2652
2653
2654
2655
            sampled_token_ids=valid_sampled_token_ids,
            logprobs=logprobs_lists,
            prompt_logprobs_dict=prompt_logprobs_dict,
            pooler_output=[],
2656
            kv_connector_output=kv_connector_output,
2657
2658
2659
            num_nans_in_logits=num_nans_in_logits,
        )

2660
2661
2662
        if not self.use_async_scheduling:
            return output

2663
        async_output = AsyncGPUModelRunnerOutput(
2664
            model_runner_output=output,
2665
            sampled_token_ids=sampler_output.sampled_token_ids,
2666
2667
2668
2669
            invalid_req_indices=invalid_req_indices,
            async_output_copy_stream=self.async_output_copy_stream,
        )

2670
2671
2672
2673
2674
2675
2676
2677
2678
        # Save ref of sampled_token_ids CPU tensor if the batch contains
        # any requests with sampling params that that require output ids.
        self.input_batch.set_async_sampled_token_ids(
            async_output.sampled_token_ids_cpu,
            async_output.async_copy_ready_event,
        )

        return async_output

2679
    def take_draft_token_ids(self) -> DraftTokenIds | None:
2680
2681
2682
2683
2684
2685
2686
2687
2688
2689
        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)

2690
2691
2692
    def propose_draft_token_ids(
        self,
        scheduler_output: "SchedulerOutput",
2693
        sampled_token_ids: torch.Tensor | list[list[int]],
2694
2695
2696
        sampling_metadata: SamplingMetadata,
        hidden_states: torch.Tensor,
        sample_hidden_states: torch.Tensor,
2697
2698
        aux_hidden_states: list[torch.Tensor] | None,
        spec_decode_metadata: SpecDecodeMetadata | None,
2699
        common_attn_metadata: CommonAttentionMetadata,
2700
    ) -> list[list[int]] | torch.Tensor:
2701
2702
        num_scheduled_tokens = scheduler_output.total_num_scheduled_tokens
        if self.speculative_config.method == "ngram":
2703
            assert isinstance(sampled_token_ids, list)
2704
            assert isinstance(self.drafter, NgramProposer)
2705
            draft_token_ids = self.drafter.propose(
2706
2707
                sampled_token_ids,
                self.input_batch.req_ids,
2708
2709
                self.input_batch.num_tokens_no_spec,
                self.input_batch.token_ids_cpu,
2710
2711
                self.input_batch.spec_decode_unsupported_reqs,
            )
2712
        elif self.speculative_config.method == "medusa":
2713
            assert isinstance(sampled_token_ids, list)
2714
            assert isinstance(self.drafter, MedusaProposer)
2715

2716
2717
            if sample_hidden_states.shape[0] == len(sampled_token_ids):
                # The input to the target model does not include draft tokens.
2718
2719
2720
2721
                hidden_states = sample_hidden_states
            else:
                indices = []
                offset = 0
Wentao Ye's avatar
Wentao Ye committed
2722
                assert spec_decode_metadata is not None
2723
                for num_draft, tokens in zip(
2724
2725
                    spec_decode_metadata.num_draft_tokens, sampled_token_ids
                ):
2726
2727
                    indices.append(offset + len(tokens) - 1)
                    offset += num_draft + 1
2728
                indices = torch.tensor(indices, device=self.device)
2729
2730
                hidden_states = sample_hidden_states[indices]

2731
            draft_token_ids = self.drafter.propose(
2732
2733
2734
                target_hidden_states=hidden_states,
                sampling_metadata=sampling_metadata,
            )
2735
        elif self.speculative_config.use_eagle():
2736
            assert isinstance(self.drafter, EagleProposer)
2737
2738
2739
2740
2741

            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.
2742
2743
                assert isinstance(sampled_token_ids, list), (
                    "sampled_token_ids should be a python list when"
2744
                    "padded-batch is disabled."
2745
                )
2746
                next_token_ids = self.drafter.prepare_next_token_ids_cpu(
2747
2748
2749
2750
2751
                    sampled_token_ids,
                    self.requests,
                    self.input_batch,
                    scheduler_output.num_scheduled_tokens,
                )
2752
2753
2754
2755
2756
            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.
2757
2758
                assert isinstance(sampled_token_ids, torch.Tensor), (
                    "sampled_token_ids should be a torch.Tensor when"
2759
                    "padded-batch is enabled."
2760
2761
                )
                next_token_ids, valid_sampled_tokens_count = (
2762
2763
2764
2765
2766
2767
                    self.drafter.prepare_next_token_ids_padded(
                        common_attn_metadata,
                        sampled_token_ids,
                        self.requests,
                        self.input_batch,
                        self.discard_request_indices.gpu,
2768
                        self.num_discarded_requests,
2769
                    )
2770
                )
Jiayi Yao's avatar
Jiayi Yao committed
2771

2772
            if spec_decode_metadata is None:
2773
                token_indices_to_sample = None
2774
                # input_ids can be None for multimodal models.
2775
                target_token_ids = self.input_ids.gpu[:num_scheduled_tokens]
2776
                target_positions = self._get_positions(num_scheduled_tokens)
2777
                if self.use_aux_hidden_state_outputs:
Wentao Ye's avatar
Wentao Ye committed
2778
                    assert aux_hidden_states is not None
2779
                    target_hidden_states = torch.cat(
2780
2781
                        [h[:num_scheduled_tokens] for h in aux_hidden_states], dim=-1
                    )
2782
2783
                else:
                    target_hidden_states = hidden_states[:num_scheduled_tokens]
2784
            else:
2785
2786
                if self.speculative_config.disable_padded_drafter_batch:
                    token_indices_to_sample = None
2787
2788
2789
2790
2791
                    common_attn_metadata, token_indices = self.drafter.prepare_inputs(
                        common_attn_metadata,
                        sampled_token_ids,
                        spec_decode_metadata.num_draft_tokens,
                    )
2792
                else:
2793
                    common_attn_metadata, token_indices, token_indices_to_sample = (
2794
2795
2796
                        self.drafter.prepare_inputs_padded(
                            common_attn_metadata,
                            spec_decode_metadata,
2797
2798
2799
                            valid_sampled_tokens_count,
                        )
                    )
2800

2801
                target_token_ids = self.input_ids.gpu[token_indices]
2802
                target_positions = self._get_positions(token_indices)
2803
                if self.use_aux_hidden_state_outputs:
Wentao Ye's avatar
Wentao Ye committed
2804
                    assert aux_hidden_states is not None
2805
                    target_hidden_states = torch.cat(
2806
2807
                        [h[token_indices] for h in aux_hidden_states], dim=-1
                    )
2808
2809
                else:
                    target_hidden_states = hidden_states[token_indices]
2810

2811
            if self.supports_mm_inputs:
2812
2813
2814
2815
2816
2817
                mm_embed_inputs = self._gather_mm_embeddings(
                    scheduler_output,
                    shift_computed_tokens=1,
                )
            else:
                mm_embed_inputs = None
2818

2819
            draft_token_ids = self.drafter.propose(
2820
2821
2822
2823
                target_token_ids=target_token_ids,
                target_positions=target_positions,
                target_hidden_states=target_hidden_states,
                next_token_ids=next_token_ids,
2824
                last_token_indices=token_indices_to_sample,
2825
                sampling_metadata=sampling_metadata,
2826
                common_attn_metadata=common_attn_metadata,
2827
                mm_embed_inputs=mm_embed_inputs,
2828
            )
2829

2830
        return draft_token_ids
2831

2832
2833
2834
    def update_config(self, overrides: dict[str, Any]) -> None:
        allowed_config_names = {"load_config", "model_config"}
        for config_name, config_overrides in overrides.items():
2835
2836
            assert config_name in allowed_config_names, (
                f"Config `{config_name}` not supported. "
2837
                f"Allowed configs: {allowed_config_names}"
2838
            )
2839
2840
2841
2842
            config = getattr(self, config_name)
            new_config = update_config(config, config_overrides)
            setattr(self, config_name, new_config)

2843
2844
2845
2846
2847
    def load_model(self, eep_scale_up: bool = False) -> None:
        """
        Args:
            eep_scale_up: the model loading is for elastic EP scale up.
        """
2848
        logger.info("Starting to load model %s...", self.model_config.model)
2849
2850
        if eep_scale_up:
            from vllm.distributed.parallel_state import get_ep_group
2851
2852
2853
2854
2855

            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
            )
2856
2857
            num_local_physical_experts = int(num_local_physical_experts.item())
            new_ep_size = get_ep_group().world_size
2858
            global_expert_load, old_global_expert_indices = EplbState.recv_state()
2859
            num_logical_experts = global_expert_load.shape[1]
2860
            self.parallel_config.eplb_config.num_redundant_experts = (
2861
2862
2863
2864
2865
2866
                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
            )
2867
            rank_mapping = {
2868
                old_ep_rank: old_ep_rank for old_ep_rank in range(old_ep_size)
2869
2870
2871
2872
2873
2874
            }
        else:
            global_expert_load = None
            old_global_expert_indices = None
            rank_mapping = None

2875
        with DeviceMemoryProfiler() as m:
2876
            time_before_load = time.perf_counter()
2877
            model_loader = get_model_loader(self.load_config)
2878
            self.model = model_loader.load_model(
2879
2880
                vllm_config=self.vllm_config, model_config=self.model_config
            )
2881
            if self.lora_config:
2882
2883
2884
                self.model = self.load_lora_model(
                    self.model, self.vllm_config, self.device
                )
2885
2886
2887
            if hasattr(self, "drafter"):
                logger.info("Loading drafter model...")
                self.drafter.load_model(self.model)
2888
            if self.use_aux_hidden_state_outputs:
2889
                if not supports_eagle3(self.get_model()):
2890
2891
                    raise RuntimeError(
                        "Model does not support EAGLE3 interface but "
2892
2893
                        "aux_hidden_state_outputs was requested"
                    )
2894
2895
2896
2897
2898
2899
2900
2901
2902
2903
2904
2905
2906

                # 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)
2907
            time_after_load = time.perf_counter()
2908
        self.model_memory_usage = m.consumed_memory
2909
2910
2911
2912
2913
        logger.info(
            "Model loading took %.4f GiB and %.6f seconds",
            self.model_memory_usage / GiB_bytes,
            time_after_load - time_before_load,
        )
2914
        prepare_communication_buffer_for_model(self.model)
2915

2916
        self.is_multimodal_pruning_enabled = (
2917
            supports_multimodal_pruning(self.get_model())
2918
2919
            and self.model_config.multimodal_config.is_multimodal_pruning_enabled()
        )
2920

2921
2922
        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)
2923
2924
2925
2926
            self.eplb_state = EplbState.build(
                self.model,
                self.device,
                self.parallel_config,
2927
2928
2929
                global_expert_load,
                old_global_expert_indices,
                rank_mapping,
2930
2931
            )

2932
        if (
2933
2934
            self.vllm_config.compilation_config.mode
            == CompilationMode.STOCK_TORCH_COMPILE
2935
            and supports_dynamo()
2936
        ):
2937
            backend = self.vllm_config.compilation_config.init_backend(self.vllm_config)
2938
            compilation_counter.stock_torch_compile_count += 1
2939
            self.model.compile(fullgraph=True, backend=backend)
2940
            return
2941
        # for other compilation modes, cudagraph behavior is controlled by
2942
2943
2944
        # CudagraphWraper and CudagraphDispatcher of vllm.

        # wrap the model with full cudagraph wrapper if needed.
2945
2946
2947
2948
2949
2950
2951
        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
            )
2952
2953
        elif self.parallel_config.enable_dbo:
            if self.compilation_config.cudagraph_mode.has_full_cudagraphs():
2954
2955
2956
                self.model = UBatchWrapper(
                    self.model, self.vllm_config, CUDAGraphMode.FULL, self.device
                )
2957
            else:
2958
2959
2960
                self.model = UBatchWrapper(
                    self.model, self.vllm_config, CUDAGraphMode.NONE, self.device
                )
2961

2962
    def _get_eagle3_aux_layers_from_config(self) -> tuple[int, ...] | None:
2963
2964
2965
2966
2967
2968
2969
2970
2971
2972
2973
2974
2975
2976
2977
2978
2979
2980
2981
2982
2983
2984
2985
        """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

2986
    def reload_weights(self) -> None:
2987
        assert getattr(self, "model", None) is not None, (
2988
            "Cannot reload weights before model is loaded."
2989
        )
2990
2991
        model_loader = get_model_loader(self.load_config)
        logger.info("Reloading weights inplace...")
2992
        model_loader.load_weights(self.get_model(), model_config=self.model_config)
2993

2994
2995
2996
2997
2998
    def save_tensorized_model(
        self,
        tensorizer_config: "TensorizerConfig",
    ) -> None:
        TensorizerLoader.save_model(
2999
            self.get_model(),
3000
            tensorizer_config=tensorizer_config,
3001
            model_config=self.model_config,
3002
3003
        )

3004
3005
3006
    def _get_prompt_logprobs_dict(
        self,
        hidden_states: torch.Tensor,
3007
        num_scheduled_tokens: dict[str, int],
3008
    ) -> dict[str, LogprobsTensors | None]:
3009
3010
3011
3012
        num_prompt_logprobs_dict = self.input_batch.num_prompt_logprobs
        if not num_prompt_logprobs_dict:
            return {}

3013
        in_progress_dict = self.input_batch.in_progress_prompt_logprobs_cpu
3014
        prompt_logprobs_dict: dict[str, LogprobsTensors | None] = {}
3015
3016
3017
3018
3019

        # 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():
3020
            num_tokens = num_scheduled_tokens[req_id]
3021
3022
3023

            # Get metadata for this request.
            request = self.requests[req_id]
3024
3025
3026
3027
            if request.prompt_token_ids is None:
                # Prompt logprobs is incompatible with prompt embeddings
                continue

3028
3029
            num_prompt_tokens = len(request.prompt_token_ids)
            prompt_token_ids = torch.tensor(request.prompt_token_ids).to(
3030
3031
                self.device, non_blocking=True
            )
3032

3033
3034
3035
3036
3037
3038
            # 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(
3039
3040
                    num_prompt_tokens - 1, num_prompt_logprobs + 1
                )
3041
3042
                in_progress_dict[req_id] = logprobs_tensors

3043
            # Determine number of logits to retrieve.
3044
3045
            start_idx = request.num_computed_tokens
            start_tok = start_idx + 1
3046
            num_remaining_tokens = num_prompt_tokens - start_tok
3047
            if num_tokens <= num_remaining_tokens:
3048
                # This is a chunk, more tokens remain.
3049
3050
3051
                # 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.
3052
3053
3054
3055
3056
                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)
3057
3058
3059
3060
3061
3062
3063
                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
3064
3065
3066
3067
3068

            # 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]
3069
            offset = self.query_start_loc.np[req_idx].item()
3070
            prompt_hidden_states = hidden_states[offset : offset + num_logits]
3071
            logits = self.model.compute_logits(prompt_hidden_states)
3072
3073
3074
3075

            # 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.
3076
            tgt_token_ids = prompt_token_ids[start_tok : start_tok + num_logits]
3077
3078

            # Compute prompt logprobs.
3079
3080
            logprobs = self.sampler.compute_logprobs(logits)
            token_ids, logprobs, ranks = self.sampler.gather_logprobs(
3081
3082
                logprobs, num_prompt_logprobs, tgt_token_ids
            )
3083
3084

            # Transfer GPU->CPU async.
3085
3086
            chunk_slice = slice(start_idx, start_idx + num_logits)
            logprobs_tensors.logprob_token_ids[chunk_slice].copy_(
3087
3088
3089
                token_ids, non_blocking=True
            )
            logprobs_tensors.logprobs[chunk_slice].copy_(logprobs, non_blocking=True)
3090
            logprobs_tensors.selected_token_ranks[chunk_slice].copy_(
3091
3092
                ranks, non_blocking=True
            )
3093
3094
3095
3096
3097

        # 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]
3098
            del in_progress_dict[req_id]
3099
3100

        # Must synchronize the non-blocking GPU->CPU transfers.
3101
        if prompt_logprobs_dict:
3102
            self._sync_device()
3103
3104
3105

        return prompt_logprobs_dict

3106
3107
    def _get_nans_in_logits(
        self,
3108
        logits: torch.Tensor | None,
3109
3110
3111
3112
3113
3114
3115
3116
3117
3118
3119
    ) -> 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])
3120
3121
3122
                    if num_nans_for_index is not None and req_index < logits.shape[0]
                    else 0
                )
3123
3124
3125
3126
            return num_nans_in_logits
        except IndexError:
            return {}

3127
3128
3129
3130
3131
3132
    @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
3133
         - during DP rank dummy run
3134
3135
3136
3137
3138
3139
3140
3141
3142
3143
3144
        """
        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(
3145
                    self.input_ids.gpu,
3146
3147
                    low=0,
                    high=self.model_config.get_vocab_size(),
3148
3149
                    dtype=input_ids.dtype,
                )
3150

3151
            logger.debug_once("Randomizing dummy data for DP Rank")
3152
            input_ids.copy_(rand_input_ids()[: input_ids.size(0)], non_blocking=True)
3153
3154
3155
            yield
            input_ids.fill_(0)

3156
3157
3158
3159
3160
3161
    def _get_mm_dummy_batch(
        self,
        modality: str,
        max_items_per_batch: int,
    ) -> BatchedTensorInputs:
        """Dummy data for profiling and precompiling multimodal models."""
3162
3163
        assert self.mm_budget is not None

3164
3165
        dummy_decoder_data = self.mm_registry.get_decoder_dummy_data(
            model_config=self.model_config,
3166
            seq_len=self.max_model_len,
3167
            mm_counts={modality: 1},
3168
            cache=self.mm_budget.cache,
3169
3170
3171
3172
        )
        dummy_mm_data = dummy_decoder_data.multi_modal_data

        # Result in the maximum GPU consumption of the model
3173
3174
        dummy_mm_item = dummy_mm_data[modality][0]
        dummy_mm_items = [dummy_mm_item] * max_items_per_batch
3175

3176
        model = cast(SupportsMultiModal, self.model)
3177
3178
3179
3180
3181
3182
3183
3184
3185
        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,
            )
        )
3186

3187
3188
3189
3190
    @torch.inference_mode()
    def _dummy_run(
        self,
        num_tokens: int,
3191
        cudagraph_runtime_mode: CUDAGraphMode | None = None,
3192
3193
        force_attention: bool = False,
        uniform_decode: bool = False,
3194
        allow_microbatching: bool = True,
3195
3196
        skip_eplb: bool = False,
        is_profile: bool = False,
3197
        create_mixed_batch: bool = False,
3198
        remove_lora: bool = True,
3199
        activate_lora: bool = False,
3200
    ) -> tuple[torch.Tensor, torch.Tensor]:
3201
3202
3203
3204
3205
3206
3207
        """
        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.
3208
                - if not set will determine the cudagraph mode based on using
3209
                    the self.cudagraph_dispatcher.
3210
3211
3212
3213
                - CUDAGraphMode.NONE: No cudagraph, for warm up and profile run
                - CUDAGraphMode.PIECEWISE: Piecewise cudagraph.
                - CUDAGraphMode.FULL: Full cudagraph, attention metadata is
                    needed.
3214
            force_attention: If True, always create attention metadata. Used to
3215
3216
3217
3218
                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.
3219
3220
            create_mixed_batch: If True, create a mixed batch with both decode
                (1 token) and prefill (multiple tokens) requests.
3221
            remove_lora: If False, dummy LoRAs are not destroyed after the run
3222
            activate_lora: If False, dummy_run is performed without LoRAs.
3223
        """
3224
3225
3226
3227
        assert (
            cudagraph_runtime_mode is None
            or cudagraph_runtime_mode.valid_runtime_modes()
        )
3228

3229
        # If cudagraph_mode.decode_mode() == FULL and
3230
        # cudagraph_mode.separate_routine(). This means that we are using
3231
3232
3233
3234
3235
3236
3237
3238
3239
3240
3241
        # 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.
3242
        max_query_len = self.uniform_decode_query_len if uniform_decode else num_tokens
3243

3244
3245
3246
3247
3248
        # 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
3249
3250
3251
3252
        if create_mixed_batch:
            assert not uniform_decode
            # Create mixed batch:
            # first half decode tokens, second half one prefill
3253
            num_decode_tokens = min(max_num_reqs - 1, num_tokens // 2)
3254
3255
3256
3257
            num_prefill_tokens = num_tokens - num_decode_tokens
            num_reqs = num_decode_tokens + 1

            # Create decode requests (1 token each) followed by prefill request
3258
            num_scheduled_tokens_list = [1] * num_decode_tokens + [num_prefill_tokens]
3259
3260
3261
            # Note: Overriding max_query_len to be the prefill tokens
            max_query_len = num_prefill_tokens
        elif uniform_decode:
3262
            assert not create_mixed_batch
3263
            num_reqs = min(max_num_reqs, cdiv(num_tokens, max_query_len))
3264
3265
            num_scheduled_tokens_list = [max_query_len] * num_reqs
            if num_tokens % max_query_len != 0:
3266
                num_scheduled_tokens_list[-1] = num_tokens % max_query_len
3267
3268
3269
3270
3271
3272
        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

3273
3274
        assert sum(num_scheduled_tokens_list) == num_tokens
        assert len(num_scheduled_tokens_list) == num_reqs
3275
        num_scheduled_tokens = np.array(num_scheduled_tokens_list, dtype=np.int32)
3276
        total_num_scheduled_tokens = int(num_scheduled_tokens.sum())
3277

3278
3279
3280
        # Disable DP padding when running eager
        allow_dp_padding = self.compilation_config.cudagraph_mode != CUDAGraphMode.NONE

3281
3282
        # We currently only microbatch if the number of tokens is
        # over a certain threshold.
3283
        ubatch_slices, num_tokens_across_dp = coordinate_batch_across_dp(
3284
3285
3286
3287
3288
3289
3290
            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,
3291
3292
3293
        )
        num_tokens_after_padding = num_tokens
        if num_tokens_across_dp is not None:
3294
3295
            dp_rank = self.parallel_config.data_parallel_rank
            num_tokens_after_padding = int(num_tokens_across_dp[dp_rank])
3296

3297
        attn_metadata: PerLayerAttnMetadata | None = None
3298
3299
3300

        # If force_attention is True, we always capture attention. Otherwise,
        # it only happens for cudagraph_runtime_mode=FULL.
3301
        if force_attention or cudagraph_runtime_mode == CUDAGraphMode.FULL:
3302
            attn_metadata = {}
3303
3304
            if ubatch_slices is not None:
                attn_metadata = [dict() for _ in range(len(ubatch_slices))]
3305

3306
3307
3308
3309
3310
3311
            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:
3312
                seq_lens = max_query_len
3313
            self.seq_lens.np[:num_reqs] = seq_lens
3314
3315
            self.seq_lens.np[num_reqs:] = 0
            self.seq_lens.copy_to_gpu()
3316

3317
3318
            cum_num_tokens, _ = self._get_cumsum_and_arange(num_scheduled_tokens)
            self.query_start_loc.np[1 : num_reqs + 1] = cum_num_tokens
3319
3320
            self.query_start_loc.copy_to_gpu()

3321
            for kv_cache_group_id, kv_cache_group_spec in enumerate(
3322
3323
                self.kv_cache_config.kv_cache_groups
            ):
3324
                common_attn_metadata = CommonAttentionMetadata(
3325
3326
                    query_start_loc=self.query_start_loc.gpu[: num_reqs + 1],
                    query_start_loc_cpu=self.query_start_loc.cpu[: num_reqs + 1],
3327
3328
                    seq_lens=self.seq_lens.gpu[:num_reqs],
                    seq_lens_cpu=self.seq_lens.cpu[:num_reqs],
3329
3330
3331
                    num_computed_tokens_cpu=self.input_batch.num_computed_tokens_cpu_tensor[
                        :num_reqs
                    ],
3332
3333
                    num_reqs=num_reqs,
                    num_actual_tokens=num_tokens,
3334
                    max_query_len=max_query_len,
3335
                    max_seq_len=self.max_model_len,
3336
3337
3338
                    block_table_tensor=self.input_batch.block_table[
                        kv_cache_group_id
                    ].get_device_tensor(num_reqs),
3339
                    slot_mapping=self.input_batch.block_table[
3340
3341
3342
                        kv_cache_group_id
                    ].slot_mapping.gpu[:num_tokens],
                    causal=True,
3343
3344
3345
                    dcp_local_seq_lens=self.dcp_local_seq_lens.gpu[:num_reqs]
                    if self.dcp_world_size > 1
                    else None,
3346
                )
3347
                for attn_group in self.attn_groups[kv_cache_group_id]:
3348
3349
                    if ubatch_slices is not None:
                        common_attn_metadata_list = split_attn_metadata(
3350
3351
                            ubatch_slices, common_attn_metadata
                        )
3352
                        for ubid, common_attn_metadata in enumerate(
3353
3354
                            common_attn_metadata_list
                        ):
3355
                            assert common_attn_metadata.max_query_len == 1
3356
3357
3358
                            attn_metadata_i = attn_group.get_metadata_builder(
                                ubatch_id=ubid
                            ).build_for_cudagraph_capture(common_attn_metadata)
3359
                            for layer_name in attn_group.layer_names:
3360
                                assert type(attn_metadata) is list
3361
                                attn_metadata[ubid][layer_name] = attn_metadata_i
3362
3363
                    else:
                        assert type(attn_metadata) is dict
3364
3365
                        metadata_builder = attn_group.get_metadata_builder()
                        attn_metadata_i = metadata_builder.build_for_cudagraph_capture(
3366
3367
                            common_attn_metadata
                        )
3368
                        for layer_name in attn_group.layer_names:
3369
                            attn_metadata[layer_name] = attn_metadata_i
3370

3371
        with self.maybe_dummy_run_with_lora(
3372
            self.lora_config, num_scheduled_tokens, activate_lora, remove_lora
3373
        ):
3374
3375
3376
            # 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)
3377
            if self.supports_mm_inputs and not self.model_config.is_encoder_decoder:
3378
                input_ids = None
3379
                inputs_embeds = self.inputs_embeds.gpu[:num_tokens_after_padding]
3380
                model_kwargs = {
3381
                    **model_kwargs,
3382
3383
                    **self._dummy_mm_kwargs(num_reqs),
                }
3384
3385
            elif self.enable_prompt_embeds:
                input_ids = None
3386
3387
                inputs_embeds = self.inputs_embeds.gpu[:num_tokens_after_padding]
                model_kwargs = self._init_model_kwargs(num_tokens_after_padding)
3388
            else:
3389
                input_ids = self.input_ids.gpu[:num_tokens_after_padding]
3390
                inputs_embeds = None
3391

3392
            if self.uses_mrope:
3393
                positions = self.mrope_positions.gpu[:, :num_tokens_after_padding]
3394
            else:
3395
                positions = self.positions.gpu[:num_tokens_after_padding]
3396
3397
3398
3399
3400
3401
3402
3403
3404

            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,
3405
3406
3407
                            device=self.device,
                        )
                    )
3408
3409

                intermediate_tensors = self.sync_and_slice_intermediate_tensors(
3410
                    num_tokens_after_padding, None, False
3411
                )
3412
3413

            # filter out the valid batch descriptor
3414
3415
3416
3417
3418
            _cg_mode, batch_descriptor = (
                self.cudagraph_dispatcher.dispatch(
                    BatchDescriptor(
                        num_tokens=num_tokens_after_padding,
                        uniform_decode=uniform_decode,
3419
                        has_lora=activate_lora and self.lora_config is not None,
3420
3421
3422
3423
3424
                    )
                )
                if not is_profile
                else (CUDAGraphMode.NONE, None)
            )
3425
3426
3427
            if cudagraph_runtime_mode is not None:
                # we allow forcing NONE when the dispatcher disagrees to support
                # warm ups for cudagraph capture
3428
3429
3430
3431
                assert (
                    cudagraph_runtime_mode == CUDAGraphMode.NONE
                    or cudagraph_runtime_mode == _cg_mode
                ), (
3432
                    f"Cudagraph runtime mode mismatch at dummy_run. "
3433
3434
                    f"Expected {_cg_mode}, but got {cudagraph_runtime_mode}."
                )
3435
3436
            else:
                cudagraph_runtime_mode = _cg_mode
3437

3438
            if ubatch_slices is not None:
3439
3440
3441
3442
3443
3444
3445
                # 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

3446
3447
3448
            with (
                self.maybe_randomize_inputs(input_ids),
                set_forward_context(
3449
3450
                    attn_metadata,
                    self.vllm_config,
3451
                    num_tokens=num_tokens_after_padding,
3452
3453
                    num_tokens_across_dp=num_tokens_across_dp,
                    cudagraph_runtime_mode=cudagraph_runtime_mode,
3454
                    batch_descriptor=batch_descriptor,
3455
3456
3457
                    ubatch_slices=ubatch_slices,
                ),
            ):
3458
                outputs = self.model(
3459
3460
3461
3462
                    input_ids=input_ids,
                    positions=positions,
                    intermediate_tensors=intermediate_tensors,
                    inputs_embeds=inputs_embeds,
3463
                    **model_kwargs,
3464
                )
3465

3466
3467
3468
3469
            if self.use_aux_hidden_state_outputs:
                hidden_states, _ = outputs
            else:
                hidden_states = outputs
3470

3471
            if self.speculative_config and self.speculative_config.use_eagle():
3472
                assert isinstance(self.drafter, EagleProposer)
3473
3474
                use_cudagraphs = cudagraph_runtime_mode == CUDAGraphMode.PIECEWISE
                self.drafter.dummy_run(num_tokens, use_cudagraphs=use_cudagraphs)
3475

3476
3477
3478
3479
3480
3481
3482
3483
3484
3485
        # 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)

3486
        logit_indices = np.cumsum(num_scheduled_tokens) - 1
3487
        return hidden_states, hidden_states[logit_indices]
3488
3489
3490
3491
3492
3493

    @torch.inference_mode()
    def _dummy_sampler_run(
        self,
        hidden_states: torch.Tensor,
    ) -> torch.Tensor:
3494
3495
3496
3497
        # 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)
3498

3499
        logits = self.model.compute_logits(hidden_states)
3500
3501
        num_reqs = logits.size(0)

3502
        dummy_tensors = lambda v: torch.full((num_reqs,), v, device=self.device)
3503
3504
3505
3506
3507
3508
3509
3510
3511
3512
3513
3514
3515
3516
3517

        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)],
3518
            spec_token_ids=[[] for _ in range(num_reqs)],
3519
3520
            allowed_token_ids_mask=None,
            bad_words_token_ids={},
3521
            logitsprocs=LogitsProcessors(),
3522
        )
3523
        try:
3524
3525
3526
            sampler_output = self.sampler(
                logits=logits, sampling_metadata=dummy_metadata
            )
3527
        except RuntimeError as e:
3528
            if "out of memory" in str(e):
3529
3530
3531
3532
                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 "
3533
3534
                    "initializing the engine."
                ) from e
3535
3536
            else:
                raise e
3537
        if self.speculative_config:
3538
3539
            draft_token_ids = [[0] for _ in range(num_reqs)]
            dummy_spec_decode_metadata = SpecDecodeMetadata.make_dummy(
3540
3541
                draft_token_ids, self.device
            )
3542
3543
3544
3545
3546
3547

            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
3548
3549
3550
            target_logits = torch.randn(
                num_tokens, logits.shape[-1], device=self.device, dtype=logits.dtype
            )
3551
3552
3553
            # 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.
3554
3555
3556
            bonus_token_ids = torch.zeros(
                num_reqs, device=self.device, dtype=torch.int32
            )
3557
3558
3559
3560
3561
3562
3563
            self.rejection_sampler(
                dummy_spec_decode_metadata,
                draft_probs,
                target_logits,
                bonus_token_ids,
                dummy_metadata,
            )
3564
        return sampler_output
3565

3566
    def _dummy_pooler_run_task(
3567
3568
        self,
        hidden_states: torch.Tensor,
3569
3570
        task: PoolingTask,
    ) -> PoolerOutput:
3571
3572
3573
3574
3575
3576
3577
3578
3579
3580
3581
        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

3582
        dummy_prompt_lens = torch.tensor(
3583
3584
            num_scheduled_tokens_list,
            device="cpu",
3585
        )
3586
3587
3588
        dummy_token_ids = torch.zeros(
            (num_reqs, req_num_tokens), dtype=torch.int32, device=self.device
        )
3589

3590
        model = cast(VllmModelForPooling, self.get_model())
3591
        dummy_pooling_params = PoolingParams(task=task)
3592
        dummy_pooling_params.verify(task=task, model_config=self.model_config)
3593
        to_update = model.pooler.get_pooling_updates(task)
3594
3595
        to_update.apply(dummy_pooling_params)

3596
        dummy_metadata = PoolingMetadata(
3597
3598
3599
3600
            prompt_lens=dummy_prompt_lens,
            prompt_token_ids=dummy_token_ids,
            pooling_params=[dummy_pooling_params] * num_reqs,
        )
3601

3602
3603
3604
        dummy_metadata.build_pooling_cursor(
            num_scheduled_tokens_list, device=hidden_states.device
        )
3605

3606
        try:
3607
3608
3609
            return model.pooler(
                hidden_states=hidden_states, pooling_metadata=dummy_metadata
            )
3610
        except RuntimeError as e:
3611
            if "out of memory" in str(e):
3612
                raise RuntimeError(
3613
3614
3615
                    "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 "
3616
3617
                    "initializing the engine."
                ) from e
3618
3619
            else:
                raise e
3620
3621
3622
3623
3624
3625
3626

    @torch.inference_mode()
    def _dummy_pooler_run(
        self,
        hidden_states: torch.Tensor,
    ) -> PoolerOutput:
        # Find the task that has the largest output for subsequent steps
3627
3628
3629
3630
3631
3632
3633
3634
3635
3636
3637
3638
3639
3640
3641
3642
3643
3644
3645
3646
        supported_pooling_tasks = self.get_supported_pooling_tasks()

        if not supported_pooling_tasks:
            if self.scheduler_config.chunked_prefill_enabled:
                raise RuntimeError(
                    f"Model {self.model_config.model} does not support "
                    "any pooling tasks with chunked prefill enabled. "
                    "Please add --no-enable-chunked-prefill to your "
                    "config or CLI args. See "
                    "https://docs.vllm.ai/en/latest/models/pooling_models.html "
                    "to learn more."
                )
            else:
                raise RuntimeError(
                    f"Model {self.model_config.model} does not support "
                    "any pooling tasks. See "
                    "https://docs.vllm.ai/en/latest/models/pooling_models.html "
                    "to learn more."
                )

3647
        output_size = dict[PoolingTask, float]()
3648
        for task in supported_pooling_tasks:
3649
3650
            # Run a full batch with each task to ensure none of them OOMs
            output = self._dummy_pooler_run_task(hidden_states, task)
3651
            output_size[task] = sum(o.nbytes for o in output)
3652
3653
3654
3655
            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)
3656

3657
    def profile_run(self) -> None:
3658
        # Profile with multimodal encoder & encoder cache.
3659
        if self.supports_mm_inputs:
3660
            if self.model_config.multimodal_config.skip_mm_profiling:
3661
                logger.info(
3662
                    "Skipping memory profiling for multimodal encoder and "
3663
3664
                    "encoder cache."
                )
3665
3666
3667
3668
3669
3670
3671
3672
            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.
3673
                    dummy_modality = mm_budget.get_modality_with_max_tokens()
3674
3675
3676
                    max_mm_items_per_batch = mm_budget.max_items_per_batch_by_modality[
                        dummy_modality
                    ]
3677
3678
3679
3680
3681
3682
3683
3684
3685

                    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,
                    )
3686

3687
3688
3689
3690
3691
                    # Create dummy batch of multimodal inputs.
                    batched_dummy_mm_inputs = self._get_mm_dummy_batch(
                        dummy_modality,
                        max_mm_items_per_batch,
                    )
3692

3693
                    # Run multimodal encoder.
3694
3695
3696
                    dummy_encoder_outputs = self.model.get_multimodal_embeddings(
                        **batched_dummy_mm_inputs
                    )
3697

3698
3699
3700
3701
                    sanity_check_mm_encoder_outputs(
                        dummy_encoder_outputs,
                        expected_num_items=max_mm_items_per_batch,
                    )
3702

3703
3704
3705
3706
3707
3708
3709
3710
3711
3712
                    # 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(
3713
3714
                                (encoder_budget, encoder_output_shape[-1])
                            )
3715
3716
3717
3718
3719
3720
                            num_tokens = output.shape[0]
                            expanded[:num_tokens].copy_(output)
                            expanded_outputs.append(expanded)

                        dummy_encoder_outputs = expanded_outputs

3721
                    # Cache the dummy encoder outputs.
3722
                    self.encoder_cache["tmp"] = dict(enumerate(dummy_encoder_outputs))
3723

3724
        # Add `is_profile` here to pre-allocate communication buffers
3725
3726
3727
        hidden_states, last_hidden_states = self._dummy_run(
            self.max_num_tokens, is_profile=True
        )
3728
        if get_pp_group().is_last_rank:
3729
3730
3731
3732
            if self.is_pooling_model:
                output = self._dummy_pooler_run(hidden_states)
            else:
                output = self._dummy_sampler_run(last_hidden_states)
3733
        else:
3734
            output = None
3735
        self._sync_device()
3736
        del hidden_states, output
3737
        self.encoder_cache.clear()
3738
        gc.collect()
3739

3740
    def capture_model(self) -> int:
3741
        if self.compilation_config.cudagraph_mode == CUDAGraphMode.NONE:
3742
            logger.warning(
3743
                "Skipping CUDA graph capture. To turn on CUDA graph capture, "
3744
3745
                "ensure `cudagraph_mode` was not manually set to `NONE`"
            )
3746
            return 0
3747
3748
        else:
            self.initialize_cudagraph_capture()
3749

3750
3751
        compilation_counter.num_gpu_runner_capture_triggers += 1

3752
3753
        start_time = time.perf_counter()

3754
3755
3756
3757
3758
3759
3760
3761
3762
3763
3764
3765
3766
3767
        @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()
3768
                    gc.collect()
3769

3770
3771
3772
        # 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.
3773
        set_cudagraph_capturing_enabled(True)
3774
        with freeze_gc(), graph_capture(device=self.device):
3775
            start_free_gpu_memory = torch.cuda.mem_get_info()[0]
3776
            cudagraph_mode = self.compilation_config.cudagraph_mode
3777
            assert cudagraph_mode is not None
3778
3779
3780
3781
3782
3783
3784
3785
3786

            if self.lora_config:
                if self.compilation_config.cudagraph_specialize_lora:
                    lora_cases = [True, False]
                else:
                    lora_cases = [True]
            else:
                lora_cases = [False]

3787
3788
3789
            if cudagraph_mode.mixed_mode() != CUDAGraphMode.NONE:
                cudagraph_runtime_mode = cudagraph_mode.mixed_mode()

3790
3791
3792
                compilation_cases = list(
                    product(reversed(self.cudagraph_batch_sizes), lora_cases)
                )
3793
3794
3795
                self._capture_cudagraphs(
                    compilation_cases,
                    cudagraph_runtime_mode=cudagraph_runtime_mode,
3796
3797
                    uniform_decode=False,
                )
3798

3799
3800
            # Capture full cudagraph for uniform decode batches if we
            # don't already have full mixed prefill-decode cudagraphs.
3801
3802
3803
3804
3805
3806
3807
            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
                )
3808
                decode_cudagraph_batch_sizes = [
3809
3810
                    x
                    for x in self.cudagraph_batch_sizes
3811
                    if max_num_tokens >= x >= self.uniform_decode_query_len
3812
                ]
3813
3814
3815
                compilation_cases_decode = list(
                    product(reversed(decode_cudagraph_batch_sizes), lora_cases)
                )
3816
3817
3818
                self._capture_cudagraphs(
                    compilation_cases=compilation_cases_decode,
                    cudagraph_runtime_mode=CUDAGraphMode.FULL,
3819
3820
                    uniform_decode=True,
                )
3821

3822
3823
3824
            torch.cuda.synchronize()
            end_free_gpu_memory = torch.cuda.mem_get_info()[0]

3825
3826
3827
        # 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
3828
        # we may do lazy capturing in future that still allows capturing
3829
3830
        # after here.
        set_cudagraph_capturing_enabled(False)
3831
3832
3833
3834
3835

        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.
3836
3837
3838
3839
3840
        logger.info(
            "Graph capturing finished in %.0f secs, took %.2f GiB",
            elapsed_time,
            cuda_graph_size / (1 << 30),
        )
3841
        return cuda_graph_size
3842

3843
3844
    def _capture_cudagraphs(
        self,
3845
        compilation_cases: list[tuple[int, bool]],
3846
3847
3848
3849
3850
3851
3852
        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}"
3853
3854
3855
3856
3857
3858
3859
3860

        # 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",
3861
3862
3863
                    cudagraph_runtime_mode.name,
                ),
            )
3864

3865
        # We skip EPLB here since we don't want to record dummy metrics
3866
        for num_tokens, activate_lora in compilation_cases:
3867
            # We currently only capture ubatched graphs when its a FULL
3868
3869
3870
            # 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
3871
3872
3873
3874
            allow_microbatching = (
                self.parallel_config.enable_dbo
                and cudagraph_runtime_mode == CUDAGraphMode.FULL
                and uniform_decode
3875
3876
3877
3878
3879
                and check_ubatch_thresholds(
                    config=self.vllm_config.parallel_config,
                    num_tokens=num_tokens,
                    uniform_decode=uniform_decode,
                )
3880
            )
3881

3882
3883
3884
3885
3886
3887
            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.
3888
3889
3890
3891
3892
3893
3894
3895
3896
                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,
3897
                    activate_lora=activate_lora,
3898
3899
3900
3901
3902
3903
3904
3905
                )
            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,
3906
                activate_lora=activate_lora,
3907
            )
3908
        self.maybe_remove_all_loras(self.lora_config)
3909

3910
3911
3912
3913
    def initialize_attn_backend(self, kv_cache_config: KVCacheConfig) -> None:
        """
        Initialize the attention backends and attention metadata builders.
        """
3914
        assert len(self.attn_groups) == 0, "Attention backends are already initialized"
3915

3916
3917
3918
3919
3920
3921
3922
3923
        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(
3924
3925
                self.vllm_config, AttentionLayerBase, kv_cache_group_spec.layer_names
            )
3926
3927
            attn_backends = {}
            attn_backend_layers = defaultdict(list)
3928
            # Dedupe based on full class name; this is a bit safer than
3929
3930
3931
3932
            # 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.
3933
            for layer_name in kv_cache_group_spec.layer_names:
3934
                attn_backend = layers[layer_name].get_attn_backend()
3935
3936
3937
3938
3939
3940
3941

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

3942
3943
3944
                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):
3945
                    layer_kv_cache_spec = layer_kv_cache_spec.kv_cache_specs[layer_name]
3946
                key = (full_cls_name, layer_kv_cache_spec)
3947
3948
3949
                attn_backends[key] = AttentionGroupKey(
                    attn_backend, layer_kv_cache_spec
                )
3950
                attn_backend_layers[key].append(layer_name)
3951
            return {attn_backends[k]: v for k, v in attn_backend_layers.items()}
3952
3953

        def create_attn_groups(
3954
            attn_backends_map: dict[AttentionGroupKey, list[str]],
3955
3956
        ) -> list[AttentionGroup]:
            attn_groups: list[AttentionGroup] = []
3957
            for (attn_backend, kv_cache_spec), layer_names in attn_backends_map.items():
3958
3959
                attn_group = AttentionGroup.create_with_metadata_builders(
                    attn_backend,
3960
                    layer_names,
3961
                    kv_cache_spec,
3962
3963
                    self.vllm_config,
                    self.device,
3964
                    num_metadata_builders=1
3965
3966
                    if not self.parallel_config.enable_dbo
                    else 2,
3967
3968
                )

3969
3970
3971
3972
                attn_groups.append(attn_group)
            return attn_groups

        for kv_cache_group_spec in kv_cache_config.kv_cache_groups:
3973
3974
            attn_backends = get_attn_backends_for_group(kv_cache_group_spec)
            self.attn_groups.append(create_attn_groups(attn_backends))
3975

co63oc's avatar
co63oc committed
3976
        # Calculate reorder batch threshold (if needed)
3977
3978
        self.calculate_reorder_batch_threshold()

3979
    def initialize_cudagraph_capture(self) -> None:
3980
        """
3981
        Resolve the cudagraph_mode when there are multiple attention
3982
3983
3984
3985
        backends with potential conflicting CUDA graph support.
        Then initialize the cudagraph_dispatcher based on the resolved
        cudagraph_mode.
        """
3986
3987
3988
3989
        min_cg_support = AttentionCGSupport.ALWAYS
        min_cg_builder_name = None

        for attn_group in self._attn_group_iterator():
3990
            builder = attn_group.get_metadata_builder()
3991
3992
3993
3994
3995
3996
            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
3997
3998
3999
4000
4001
4002
4003
4004
4005
        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})"
            )
4006
4007
            if min_cg_support == AttentionCGSupport.NEVER:
                # if not supported any full cudagraphs, just raise it.
4008
4009
                msg += (
                    "; please try cudagraph_mode=PIECEWISE, and "
4010
                    "make sure compilation mode is VLLM_COMPILE"
4011
                )
4012
4013
4014
4015
4016
                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"
4017
                cudagraph_mode = self.compilation_config.cudagraph_mode = (
4018
                    CUDAGraphMode.FULL_AND_PIECEWISE
4019
                )
4020
4021
            else:
                msg += "; setting cudagraph_mode=FULL_DECODE_ONLY"
4022
                cudagraph_mode = self.compilation_config.cudagraph_mode = (
4023
                    CUDAGraphMode.FULL_DECODE_ONLY
4024
                )
4025
4026
            logger.warning(msg)

4027
        # check that if we are doing decode full-cudagraphs it is supported
4028
4029
4030
4031
4032
4033
4034
4035
4036
        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})"
            )
4037
            if self.compilation_config.mode == CompilationMode.VLLM_COMPILE and (
4038
4039
4040
4041
4042
                self.compilation_config.splitting_ops_contain_attention()
                or self.compilation_config.use_inductor_graph_partition
            ):
                msg += (
                    "; setting cudagraph_mode=PIECEWISE because "
4043
                    "attention is compiled piecewise"
4044
4045
                )
                cudagraph_mode = self.compilation_config.cudagraph_mode = (
4046
                    CUDAGraphMode.PIECEWISE
4047
                )
4048
            else:
4049
4050
                msg += (
                    "; setting cudagraph_mode=NONE because "
4051
                    "attention is not compiled piecewise"
4052
4053
                )
                cudagraph_mode = self.compilation_config.cudagraph_mode = (
4054
                    CUDAGraphMode.NONE
4055
                )
4056
4057
            logger.warning(msg)

4058
4059
        # check that if we are doing spec-decode + decode full-cudagraphs it is
        # supported
4060
4061
4062
4063
4064
4065
4066
4067
4068
4069
        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})"
            )
4070
4071
            if self.compilation_config.splitting_ops_contain_attention():
                msg += "; setting cudagraph_mode=PIECEWISE"
4072
                cudagraph_mode = self.compilation_config.cudagraph_mode = (
4073
                    CUDAGraphMode.PIECEWISE
4074
                )
4075
4076
            else:
                msg += "; setting cudagraph_mode=NONE"
4077
                cudagraph_mode = self.compilation_config.cudagraph_mode = (
4078
                    CUDAGraphMode.NONE
4079
                )
4080
4081
4082
4083
            logger.warning(msg)

        # double check that we can support full cudagraph if they are requested
        # even after automatic downgrades
4084
4085
4086
4087
4088
4089
4090
4091
4092
        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, "
4093
                "and make sure compilation mode is VLLM_COMPILE"
4094
            )
4095
4096
4097
4098

        # Trigger cudagraph dispatching keys initialization here (after
        # initializing attn backends).
        self.cudagraph_dispatcher.initialize_cudagraph_keys(
4099
4100
            self.compilation_config.cudagraph_mode, self.uniform_decode_query_len
        )
4101

4102
4103
4104
4105
4106
    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)
        """
4107
        for group in self._attn_group_iterator():
4108
            attn_metadata_builder_i = group.get_metadata_builder()
4109

4110
4111
            # check that if any backends reorder batches; that the reordering
            # is compatible (e.g., decode threshold is the same)
4112
            reorder_batch_threshold_i = attn_metadata_builder_i.reorder_batch_threshold
4113
4114
            if reorder_batch_threshold_i is not None:
                if self.reorder_batch_threshold is not None:
4115
                    if reorder_batch_threshold_i != self.reorder_batch_threshold:
4116
4117
4118
4119
                        raise ValueError(
                            f"Attention backend reorders decodes with "
                            f"threshold {reorder_batch_threshold_i} but other "
                            f"backend uses threshold "
4120
4121
                            f"{self.reorder_batch_threshold}"
                        )
4122
4123
4124
                else:
                    self.reorder_batch_threshold = reorder_batch_threshold_i

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
4159
4160
4161
4162
4163
4164
4165
4166
4167
4168
4169
4170
4171
4172
4173
4174
4175
4176
4177
4178
4179
4180
4181
4182
4183
4184
4185
4186
4187
4188
4189
4190
4191
4192
4193
4194
4195
4196
4197
4198
4199
4200
4201
4202
4203
4204
    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)

4205
    def may_reinitialize_input_batch(self, kv_cache_config: KVCacheConfig) -> None:
4206
4207
4208
4209
4210
4211
4212
4213
4214
4215
4216
        """
        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
4217
            if not isinstance(kv_cache_group.kv_cache_spec, EncoderOnlyAttentionSpec)
4218
        ]
4219
4220
4221
4222
4223
4224
4225

        # 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
        ]:
4226
4227
4228
            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
4229
4230
                "for more details."
            )
4231
4232
            self.input_batch = InputBatch(
                max_num_reqs=self.max_num_reqs,
4233
                max_model_len=max(self.max_model_len, self.max_encoder_len),
4234
4235
4236
4237
4238
                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,
4239
                kernel_block_sizes=kernel_block_sizes,
4240
                is_spec_decode=bool(self.vllm_config.speculative_config),
4241
                logitsprocs=self.input_batch.logitsprocs,
4242
                logitsprocs_need_output_token_ids=self.input_batch.logitsprocs_need_output_token_ids,
4243
                is_pooling_model=self.is_pooling_model,
4244
4245
                num_speculative_tokens=(
                    self.vllm_config.speculative_config.num_speculative_tokens
4246
4247
4248
                    if self.vllm_config.speculative_config
                    else 0
                ),
4249
4250
            )

4251
    def _allocate_kv_cache_tensors(
4252
4253
        self, kv_cache_config: KVCacheConfig
    ) -> dict[str, torch.Tensor]:
4254
        """
4255
4256
4257
        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.

4258
        Args:
4259
            kv_cache_config: The KV cache config
4260
        Returns:
4261
            dict[str, torch.Tensor]: A map between layer names to their
4262
            corresponding memory buffer for KV cache.
4263
        """
4264
4265
        kv_cache_raw_tensors: dict[str, torch.Tensor] = {}
        for kv_cache_tensor in kv_cache_config.kv_cache_tensors:
4266
4267
4268
            tensor = torch.zeros(
                kv_cache_tensor.size, dtype=torch.int8, device=self.device
            )
4269
4270
4271
4272
4273
            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:
4274
4275
4276
4277
            for layer_name in group.layer_names:
                if layer_name in self.runner_only_attn_layers:
                    continue
                layer_names.add(layer_name)
4278
4279
4280
        assert layer_names == set(kv_cache_raw_tensors.keys()), (
            "Some layers are not correctly initialized"
        )
4281
4282
        return kv_cache_raw_tensors

4283
4284
4285
    def _attn_group_iterator(self) -> Iterator[AttentionGroup]:
        return itertools.chain.from_iterable(self.attn_groups)

4286
    def _kv_cache_spec_attn_group_iterator(self) -> Iterator[AttentionGroup]:
4287
4288
        if not self.kv_cache_config.kv_cache_groups:
            return
4289
4290
        for attn_groups in self.attn_groups:
            yield from attn_groups
4291

4292
4293
4294
4295
4296
4297
4298
4299
4300
4301
4302
4303
4304
4305
4306
4307
4308
4309
    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
        ):
4310
4311
4312
4313
4314
4315
            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):
4316
                continue
4317
            elif isinstance(kv_cache_spec, AttentionSpec):
4318
4319
4320
4321
4322
4323
4324
4325
4326
                # 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)
4327
            elif isinstance(kv_cache_spec, MambaSpec):
4328
4329
                # This is likely Mamba or other non-attention cache,
                # no splitting.
4330
                kernel_block_sizes.append(kv_cache_spec.block_size)
4331
4332
4333
4334
4335
4336
            else:
                raise NotImplementedError(
                    f"unknown kv cache spec {kv_cache_group.kv_cache_spec}"
                )
        return kernel_block_sizes

4337
4338
4339
4340
4341
    def _reshape_kv_cache_tensors(
        self,
        kv_cache_config: KVCacheConfig,
        kv_cache_raw_tensors: dict[str, torch.Tensor],
    ) -> dict[str, torch.Tensor]:
4342
        """
4343
        Reshape the KV cache tensors to the desired shape and dtype.
4344

4345
        Args:
4346
4347
            kv_cache_config: The KV cache config
            kv_cache_raw_tensors: The KV cache buffer of each layer, with
4348
                correct size but uninitialized shape.
4349
        Returns:
4350
            Dict[str, torch.Tensor]: A map between layer names to their
4351
4352
            corresponding memory buffer for KV cache.
        """
4353
        kv_caches: dict[str, torch.Tensor] = {}
4354
        has_attn, has_mamba = False, False
4355
4356
        for group in self._kv_cache_spec_attn_group_iterator():
            kv_cache_spec = group.kv_cache_spec
4357
4358
            attn_backend = group.backend
            for layer_name in group.layer_names:
4359
4360
                if layer_name in self.runner_only_attn_layers:
                    continue
4361
4362
                raw_tensor = kv_cache_raw_tensors[layer_name]
                assert raw_tensor.numel() % kv_cache_spec.page_size_bytes == 0
4363
                num_blocks = raw_tensor.numel() // kv_cache_spec.page_size_bytes
4364
                if isinstance(kv_cache_spec, AttentionSpec):
4365
                    has_attn = True
4366
4367
4368
4369
4370
4371
4372
4373
                    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

4374
                    kv_cache_shape = attn_backend.get_kv_cache_shape(
4375
4376
                        kernel_num_blocks,
                        kernel_size,
4377
4378
                        kv_cache_spec.num_kv_heads,
                        kv_cache_spec.head_size,
4379
4380
                        cache_dtype_str=self.cache_config.cache_dtype,
                    )
4381
                    dtype = kv_cache_spec.dtype
4382
                    try:
4383
                        kv_cache_stride_order = attn_backend.get_kv_cache_stride_order()  # noqa: E501
4384
                        assert len(kv_cache_stride_order) == len(kv_cache_shape)
4385
                    except (AttributeError, NotImplementedError):
4386
                        kv_cache_stride_order = tuple(range(len(kv_cache_shape)))
4387
4388
4389
4390
4391
                    # 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.
4392
4393
4394
                    kv_cache_shape = tuple(
                        kv_cache_shape[i] for i in kv_cache_stride_order
                    )
4395
4396
4397
4398
4399
                    # Maintain original KV shape view.
                    inv_order = [
                        kv_cache_stride_order.index(i)
                        for i in range(len(kv_cache_stride_order))
                    ]
4400
4401
4402
4403
4404
4405
                    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
4406
                elif isinstance(kv_cache_spec, MambaSpec):
4407
                    has_mamba = True
Chen Zhang's avatar
Chen Zhang committed
4408
4409
                    raw_tensor = kv_cache_raw_tensors[layer_name]
                    state_tensors = []
4410
                    storage_offset_bytes = 0
4411
                    for shape, dtype in zip(kv_cache_spec.shapes, kv_cache_spec.dtypes):
4412
4413
                        dtype_size = get_dtype_size(dtype)
                        num_element_per_page = (
4414
4415
                            kv_cache_spec.page_size_bytes // dtype_size
                        )
Chen Zhang's avatar
Chen Zhang committed
4416
                        target_shape = (num_blocks, *shape)
4417
4418
                        stride = torch.empty(target_shape).stride()
                        target_stride = (num_element_per_page, *stride[1:])
4419
                        assert storage_offset_bytes % dtype_size == 0
4420
4421
4422
4423
                        tensor = torch.as_strided(
                            raw_tensor.view(dtype),
                            size=target_shape,
                            stride=target_stride,
4424
                            storage_offset=storage_offset_bytes // dtype_size,
4425
                        )
Chen Zhang's avatar
Chen Zhang committed
4426
                        state_tensors.append(tensor)
4427
                        storage_offset_bytes += stride[0] * dtype_size
4428
4429

                    kv_caches[layer_name] = state_tensors
4430
                else:
4431
                    raise NotImplementedError
4432
4433

        if has_attn and has_mamba:
4434
            self._update_hybrid_attention_mamba_layout(kv_caches)
4435

4436
4437
        return kv_caches

4438
    def _update_hybrid_attention_mamba_layout(
4439
4440
        self, kv_caches: dict[str, torch.Tensor]
    ) -> None:
4441
        """
4442
4443
        Update the layout of attention layers from (2, num_blocks, ...) to
        (num_blocks, 2, ...).
4444
4445

        Args:
4446
            kv_caches: The KV cache buffer of each layer.
4447
4448
        """

4449
4450
        for group in self._kv_cache_spec_attn_group_iterator():
            kv_cache_spec = group.kv_cache_spec
4451
            for layer_name in group.layer_names:
4452
                kv_cache = kv_caches[layer_name]
4453
4454
4455
4456
                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 "
4457
                        f"a tensor of shape {kv_cache.shape}"
4458
                    )
4459
                    hidden_size = kv_cache.shape[2:].numel()
4460
4461
4462
4463
                    kv_cache.as_strided_(
                        size=kv_cache.shape,
                        stride=(hidden_size, 2 * hidden_size, *kv_cache.stride()[2:]),
                    )
4464

4465
    def initialize_kv_cache_tensors(
4466
4467
        self, kv_cache_config: KVCacheConfig
    ) -> dict[str, torch.Tensor]:
4468
4469
4470
4471
4472
4473
        """
        Initialize the memory buffer for KV cache.

        Args:
            kv_cache_config: The KV cache config
        Returns:
4474
            Dict[str, torch.Tensor]: A map between layer names to their
4475
4476
4477
4478
4479
            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
4480
4481
4482
        kv_caches = self._reshape_kv_cache_tensors(
            kv_cache_config, kv_cache_raw_tensors
        )
4483

4484
        # Set up cross-layer KV cache sharing
4485
4486
        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)
4487
4488
            kv_caches[layer_name] = kv_caches[target_layer_name]

4489
4490
4491
4492
4493
4494
4495
4496
4497
        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,
        )
4498
4499
4500
        return kv_caches

    def maybe_add_kv_sharing_layers_to_kv_cache_groups(
4501
4502
        self, kv_cache_config: KVCacheConfig
    ) -> None:
4503
4504
4505
4506
4507
4508
4509
4510
4511
4512
4513
4514
4515
4516
4517
4518
4519
4520
        """
        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.
4521
            attn_layers = get_layers_from_vllm_config(self.vllm_config, Attention)
4522
4523
            for layer_name in reversed(attn_layers):
                if layer_name in self.shared_kv_cache_layers:
4524
                    self.kv_sharing_fast_prefill_eligible_layers.add(layer_name)
4525
4526
                else:
                    break
4527

4528
4529
4530
4531
4532
4533
4534
    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
        """
4535
        kv_cache_config = deepcopy(kv_cache_config)
4536
        self.kv_cache_config = kv_cache_config
4537
        self.may_add_encoder_only_layers_to_kv_cache_config()
4538
        self.maybe_add_kv_sharing_layers_to_kv_cache_groups(kv_cache_config)
4539
        self.initialize_attn_backend(kv_cache_config)
4540
4541
        # Reinitialize need to after initialize_attn_backend
        self.may_reinitialize_input_batch(kv_cache_config)
4542
4543
        kv_caches = self.initialize_kv_cache_tensors(kv_cache_config)

4544
4545
4546
4547
4548
4549
        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
4550
        if has_kv_transfer_group():
4551
4552
4553
            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
4554

4555
        if self.dcp_world_size > 1:
4556
            layer_names = self.attn_groups[0][0].layer_names
4557
4558
4559
            layers = get_layers_from_vllm_config(
                self.vllm_config, AttentionLayerBase, layer_names
            )
4560
4561
4562
4563
4564
            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__} "
4565
4566
                    "does not return the softmax lse for decode."
                )
4567

4568
4569
4570
4571
4572
    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
4573
        encoder_only_attn_specs: dict[AttentionSpec, list[str]] = defaultdict(list)
4574
4575
4576
        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:
4577
                attn_spec: AttentionSpec = EncoderOnlyAttentionSpec(
4578
4579
4580
                    block_size=block_size,
                    num_kv_heads=attn_module.num_kv_heads,
                    head_size=attn_module.head_size,
4581
4582
                    dtype=self.kv_cache_dtype,
                )
4583
4584
4585
                encoder_only_attn_specs[attn_spec].append(layer_name)
                self.runner_only_attn_layers.add(layer_name)
        if len(encoder_only_attn_specs) > 0:
4586
4587
4588
            assert len(encoder_only_attn_specs) == 1, (
                "Only support one encoder-only attention spec now"
            )
4589
4590
            spec, layer_names = encoder_only_attn_specs.popitem()
            self.kv_cache_config.kv_cache_groups.append(
4591
4592
                KVCacheGroupSpec(layer_names=layer_names, kv_cache_spec=spec)
            )
4593

4594
    def get_kv_cache_spec(self) -> dict[str, KVCacheSpec]:
4595
        """
4596
        Generates the KVCacheSpec by parsing the kv cache format from each
4597
4598
        Attention module in the static forward context.
        Returns:
4599
            KVCacheSpec: A dictionary mapping layer names to their KV cache
4600
4601
4602
            format. Layers that do not need KV cache are not included.
        """

4603
        kv_cache_spec: dict[str, KVCacheSpec] = {}
4604
        attn_layers = get_layers_from_vllm_config(self.vllm_config, AttentionLayerBase)
Chen Zhang's avatar
Chen Zhang committed
4605
        for layer_name, attn_module in attn_layers.items():
4606
4607
4608
4609
4610
4611
4612
4613
4614
4615
4616
4617
4618
4619
4620
            if isinstance(attn_module, Attention) and (
                kv_tgt_layer := attn_module.kv_sharing_target_layer_name
            ):
                # The layer doesn't need its own KV cache and will use that of
                # the target layer. We skip creating a KVCacheSpec for it, so
                # that KV cache management logic will act as this layer does
                # not exist, and doesn't allocate KV cache for the layer. This
                # enables the memory saving of cross-layer kv sharing, allowing
                # a given amount of memory to accommodate longer context lengths
                # or enable more requests to be processed simultaneously.
                self.shared_kv_cache_layers[layer_name] = kv_tgt_layer
                continue
            # Skip modules that don't need KV cache (eg encoder-only attention)
            if spec := attn_module.get_kv_cache_spec(self.vllm_config):
                kv_cache_spec[layer_name] = spec
4621

4622
        return kv_cache_spec
4623
4624
4625
4626
4627
4628
4629
4630
4631
4632

    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.
4633
        pinned = self.sampled_token_ids_pinned_cpu[: sampled_token_ids.shape[0]]
4634
4635
4636
4637
        pinned.copy_(sampled_token_ids, non_blocking=True)
        self.transfer_event.record()
        self.transfer_event.synchronize()
        return pinned.tolist()