gpu_model_runner.py 203 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 functools import reduce
12
from itertools import product
13
from typing import TYPE_CHECKING, Any, NamedTuple, TypeAlias, cast
14
15
16
17
18

import numpy as np
import torch
import torch.distributed
import torch.nn as nn
19
from tqdm import tqdm
20

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

143
144
145
146
147
148
149
150
151
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,
)
152

153
if TYPE_CHECKING:
154
    from vllm.model_executor.model_loader.tensorizer import TensorizerConfig
155
    from vllm.v1.core.sched.output import GrammarOutput, SchedulerOutput
156
157
158

logger = init_logger(__name__)

159
160
AttnMetadataDict: TypeAlias = dict[str, AttentionMetadata]
# list when ubatching is enabled
161
PerLayerAttnMetadata: TypeAlias = list[AttnMetadataDict] | AttnMetadataDict
162

163

164
165
166
167
168
169
# Wrapper for ModelRunnerOutput to support overlapped execution.
class AsyncGPUModelRunnerOutput(AsyncModelRunnerOutput):
    def __init__(
        self,
        model_runner_output: ModelRunnerOutput,
        sampled_token_ids: torch.Tensor,
170
        logprobs_tensors: torch.Tensor | None,
171
172
173
174
175
176
177
        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

        # 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
183
        self._logprobs_tensors = logprobs_tensors
184
185
186
187
188

        # 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)
189
            self.sampled_token_ids_cpu = self._sampled_token_ids.to(
190
191
                "cpu", non_blocking=True
            )
192
193
194
195
196
            self._logprobs_tensors_cpu = (
                self._logprobs_tensors.to_cpu_nonblocking()
                if self._logprobs_tensors
                else None
            )
197
            self.async_copy_ready_event.record()
198
199
200

    def get_output(self) -> ModelRunnerOutput:
        """Copy the device tensors to the host and return a ModelRunnerOutput.
201

202
203
        This function blocks until the copy is finished.
        """
204
        self.async_copy_ready_event.synchronize()
205

206
207
        # Release the device tensors once the copy has completed.
        del self._logprobs_tensors
208
209
        del self._sampled_token_ids

210
        valid_sampled_token_ids = self.sampled_token_ids_cpu.tolist()
211
212
213
214
215
        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
216
217
218
219
        if self._logprobs_tensors_cpu:
            # NOTE(nick): this will need to be updated to use cu_num_accepted_tokens
            # for async sched + spec decode + logprobs compatibility.
            output.logprobs = self._logprobs_tensors_cpu.tolists()
220
221
222
        return output


223
224
225
226
227
228
229
230
231
232
233
234
235
236
class ExecuteModelState(NamedTuple):
    """Ephemeral cached state transferred between execute_model() and
    sample_tokens(), after execute_model() returns None."""

    scheduler_output: "SchedulerOutput"
    logits: torch.Tensor
    spec_decode_metadata: SpecDecodeMetadata | None
    spec_decode_common_attn_metadata: CommonAttentionMetadata | None
    hidden_states: torch.Tensor
    sample_hidden_states: torch.Tensor
    aux_hidden_states: list[torch.Tensor] | None
    kv_connector_output: KVConnectorOutput | None


237
class GPUModelRunner(LoRAModelRunnerMixin, KVConnectorModelRunnerMixin):
238
239
    def __init__(
        self,
240
        vllm_config: VllmConfig,
241
        device: torch.device,
242
    ):
243
244
245
        self.vllm_config = vllm_config
        self.model_config = vllm_config.model_config
        self.cache_config = vllm_config.cache_config
246
        self.compilation_config = vllm_config.compilation_config
247
248
249
250
251
252
        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
253

254
        from vllm.model_executor.models.utils import set_cpu_offload_max_bytes
255
256

        set_cpu_offload_max_bytes(int(self.cache_config.cpu_offload_gb * 1024**3))
257

258
259
260
261
        model_config = self.model_config
        cache_config = self.cache_config
        scheduler_config = self.scheduler_config
        parallel_config = self.parallel_config
262
        self.device = device
263
264
        self.pin_memory = is_pin_memory_available()
        self.dtype = self.model_config.dtype
265
266
267
        self.kv_cache_dtype = kv_cache_dtype_str_to_dtype(
            cache_config.cache_dtype, self.model_config
        )
268

269
        self.is_pooling_model = model_config.runner_type == "pooling"
270
        self.enable_prompt_embeds = model_config.enable_prompt_embeds
271
        self.is_multimodal_raw_input_only_model = (
272
273
            model_config.is_multimodal_raw_input_only_model
        )
274
275
        # This will be overridden in load_model()
        self.is_multimodal_pruning_enabled = False
276
        self.max_model_len = model_config.max_model_len
277
        self.dcp_world_size = self.parallel_config.decode_context_parallel_size
278
        self.max_num_tokens = scheduler_config.max_num_batched_tokens
279
        self.max_num_reqs = scheduler_config.max_num_seqs
280

281
282
283
284
285
        # 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 = (
286
287
288
            self.parallel_config.distributed_executor_backend == "external_launcher"
            and len(get_pp_group().ranks) > 0
        )
289

290
        # Model-related.
291
        self.num_query_heads = model_config.get_num_attention_heads(parallel_config)
292
        self.hidden_size = model_config.get_hidden_size()
293
        self.attention_chunk_size = model_config.attention_chunk_size
294
        # Only relevant for models using ALiBi (e.g, MPT)
295
        self.use_alibi = model_config.uses_alibi
296

297
        self.cascade_attn_enabled = not self.model_config.disable_cascade_attn
298

299
        # Multi-modal data support
300
        self.mm_registry = MULTIMODAL_REGISTRY
301
        self.uses_mrope = model_config.uses_mrope
302
        self.supports_mm_inputs = self.mm_registry.supports_multimodal_inputs(
303
304
            model_config
        )
305

306
307
308
        if self.model_config.is_encoder_decoder:
            # Maximum length of the encoder input, only for encoder-decoder
            # models.
309
            self.max_encoder_len = scheduler_config.max_num_encoder_input_tokens
310
311
312
        else:
            self.max_encoder_len = 0

313
        # Sampler
314
        self.sampler = Sampler(logprobs_mode=self.model_config.logprobs_mode)
315

316
        self.eplb_state: EplbState | None = None
317
318
319
320
321
322
        """
        State of the expert parallelism load balancer.

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

323
        # Lazy initializations
324
        # self.model: nn.Module  # Set after load_model
325
        # Initialize in initialize_kv_cache
326
        self.kv_caches: list[torch.Tensor] = []
327
328
        # indexes: [kv_cache_group_id][attn_group]
        self.attn_groups: list[list[AttentionGroup]] = []
329
330
        # self.kv_cache_config: KVCacheConfig

331
332
        # mm_hash ->  encoder_output
        self.encoder_cache: dict[str, torch.Tensor] = {}
333

334
        self.use_aux_hidden_state_outputs = False
335
336
337
338
339
        # 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:
340
341
342
            self.drafter: (
                NgramProposer | SuffixDecodingProposer | EagleProposer | MedusaProposer
            )
343
344
            if self.speculative_config.method == "ngram":
                self.drafter = NgramProposer(self.vllm_config)
345
346
            elif self.speculative_config.method == "suffix":
                self.drafter = SuffixDecodingProposer(self.vllm_config)
347
            elif self.speculative_config.use_eagle():
348
                self.drafter = EagleProposer(self.vllm_config, self.device, self)
349
350
351
352
                if self.speculative_config.method == "eagle3":
                    self.use_aux_hidden_state_outputs = True
            elif self.speculative_config.method == "medusa":
                self.drafter = MedusaProposer(
353
                    vllm_config=self.vllm_config, device=self.device
354
                )
355
            else:
356
357
358
359
                raise ValueError(
                    "Unknown speculative decoding method: "
                    f"{self.speculative_config.method}"
                )
360
            self.rejection_sampler = RejectionSampler(self.sampler)
361

362
        # Request states.
363
        self.requests: dict[str, CachedRequestState] = {}
364
        self.comm_stream = torch.cuda.Stream()
365

366
367
368
369
370
371
372
373
374
        # 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.
375
        custom_logitsprocs = model_config.logits_processors
376
377
        self.input_batch = InputBatch(
            max_num_reqs=self.max_num_reqs,
378
379
380
            # 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),
381
382
383
            max_num_batched_tokens=self.max_num_tokens,
            device=self.device,
            pin_memory=self.pin_memory,
384
            vocab_size=self.model_config.get_vocab_size(),
385
            block_sizes=[self.cache_config.block_size],
386
            kernel_block_sizes=[self.cache_config.block_size],
387
            is_spec_decode=bool(self.vllm_config.speculative_config),
388
            logitsprocs=build_logitsprocs(
389
390
391
                self.vllm_config,
                self.device,
                self.pin_memory,
392
                self.is_pooling_model,
393
                custom_logitsprocs,
394
            ),
395
396
397
            # 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),
398
            is_pooling_model=self.is_pooling_model,
399
        )
400

401
        self.use_async_scheduling = self.scheduler_config.async_scheduling
402
403
404
405
406
407
408
409
410
        # 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()
411

412
        # self.cudagraph_batch_sizes sorts in ascending order.
413
414
415
416
        if (
            self.compilation_config.cudagraph_capture_sizes
            and self.compilation_config.cudagraph_mode != CUDAGraphMode.NONE
        ):
417
418
            self.cudagraph_batch_sizes = sorted(
                self.compilation_config.cudagraph_capture_sizes
419
            )
420

421
        # Cache the device properties.
422
        self._init_device_properties()
423

424
        # Persistent buffers for CUDA graphs.
425
426
427
428
429
        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
        )
430
        self.seq_lens = self._make_buffer(self.max_num_reqs, dtype=torch.int32)
431
432
433
434
        if self.dcp_world_size > 1:
            self.dcp_local_seq_lens = self._make_buffer(
                self.max_num_reqs, dtype=torch.int32
            )
435
436
437
        # 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.
438
439
440
441
442
443
444
        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
        )
445
446
        self.num_discarded_requests = 0

447
448
449
450
451
452
        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
        )
453

454
455
        # Only relevant for multimodal models
        if self.supports_mm_inputs:
456
            self.is_mm_embed = self._make_buffer(self.max_num_tokens, dtype=torch.bool)
457

458
        # Only relevant for models using M-RoPE (e.g, Qwen2-VL)
459
        if self.uses_mrope:
Roger Wang's avatar
Roger Wang committed
460
461
462
463
            # 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
464
465
466
467
468
469

            # 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
470
            self.mrope_positions = self._make_buffer(
471
472
                (3, self.max_num_tokens + 1), dtype=torch.int64
            )
473

474
        # None in the first PP rank. The rest are set after load_model.
475
        self.intermediate_tensors: IntermediateTensors | None = None
476

477
        # OPTIMIZATION: Cache the tensors rather than creating them every step.
478
        # Keep in int64 to avoid overflow with long context
479
480
481
482
        self.arange_np = np.arange(
            max(self.max_num_reqs + 1, self.max_model_len, self.max_num_tokens),
            dtype=np.int64,
        )
483

484
485
486
487
488
        # 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] = {}
489
490
491
492
493
        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(
494
495
                self.max_num_tokens, dtype=torch.int32, device=self.device
            )
496

497
498
499
500
501
        self.uniform_decode_query_len = (
            1
            if not self.speculative_config
            else 1 + self.speculative_config.num_speculative_tokens
        )
502
503
504
505

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

506
507
508
509
510
511
512
513
514
        self.mm_budget = (
            MultiModalBudget(
                self.model_config,
                self.scheduler_config,
                self.mm_registry,
            )
            if self.supports_mm_inputs
            else None
        )
515

516
        self.reorder_batch_threshold: int | None = None
517

518
519
520
521
522
        # 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()

523
        # Cached outputs.
524
        self._draft_token_ids: list[list[int]] | torch.Tensor | None = None
525
526
        self.transfer_event = torch.cuda.Event()
        self.sampled_token_ids_pinned_cpu = torch.empty(
527
            (self.max_num_reqs, 1),
528
529
            dtype=torch.int64,
            device="cpu",
530
531
            pin_memory=self.pin_memory,
        )
532

533
534
535
        # Ephemeral state transferred between execute_model() and sample_tokens().
        self.execute_model_state: ExecuteModelState | None = None

536
537
538
539
    def reset_mm_cache(self) -> None:
        if self.mm_budget:
            self.mm_budget.reset_cache()

540
541
542
543
544
545
546
547
548
549
    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]

550
    def _make_buffer(
551
        self, *size: int | torch.SymInt, dtype: torch.dtype, numpy: bool = True
552
553
554
555
556
557
558
559
    ) -> CpuGpuBuffer:
        return CpuGpuBuffer(
            *size,
            dtype=dtype,
            device=self.device,
            pin_memory=self.pin_memory,
            with_numpy=numpy,
        )
560

561
562
563
    def _init_model_kwargs(self, num_tokens: int):
        model_kwargs = dict[str, Any]()

564
        if not self.is_pooling_model:
565
566
            return model_kwargs

567
568
        num_reqs = self.input_batch.num_reqs
        pooling_params = self.input_batch.get_pooling_params()
569
570
571

        token_type_id_requests = dict[int, Any]()
        for i, param in enumerate(pooling_params):
572
573
574
575
576
            if (
                param.extra_kwargs is not None
                and (token_types := param.extra_kwargs.get("compressed_token_type_ids"))
                is not None
            ):
577
578
579
580
581
                token_type_id_requests[i] = token_types

        if len(token_type_id_requests) == 0:
            return model_kwargs

582
        seq_lens = self.seq_lens.gpu[:num_reqs]
583
584
585
586
587
588
589
590
        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(
591
592
            device=self.device
        )
593
594
        return model_kwargs

595
    def _may_reorder_batch(self, scheduler_output: "SchedulerOutput") -> None:
596
597
        """
        Update the order of requests in the batch based on the attention
598
        backend's needs. For example, some attention backends (namely MLA) may
599
600
601
602
603
604
        want to separate requests based on if the attention computation will be
        compute-bound or memory-bound.

        Args:
            scheduler_output: The scheduler output.
        """
605
606
607
608
609
610
611
612
        # 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

613
        if self.reorder_batch_threshold is not None:
614
615
616
            # 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.
617
618
619
620
            if (
                self.dcp_world_size > 1
                and envs.VLLM_ATTENTION_BACKEND != "FLASH_ATTN_MLA"
            ):
621
                assert self.reorder_batch_threshold == 1, (
622
                    "DCP not support reorder_batch_threshold > 1 now."
623
                )
624
625
626
            reorder_batch_to_split_decodes_and_prefills(
                self.input_batch,
                scheduler_output,
627
628
                decode_threshold=self.reorder_batch_threshold,
            )
629

630
631
    # Note: used for model runner override.
    def _init_device_properties(self) -> None:
632
        """Initialize attributes from torch.cuda.get_device_properties"""
633
634
635
636
637
638
639
        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()

640
    def _update_states(self, scheduler_output: "SchedulerOutput") -> None:
641
642
643
644
645
646
        """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.

647
648
        The SamplingMetadata is updated and copied to the GPU if there is a
        new/resumed/paused/finished request in the batch.
649
650
        """
        # Remove finished requests from the cached states.
651
652
        for req_id in scheduler_output.finished_req_ids:
            self.requests.pop(req_id, None)
653
654
655
656
657
658
659
        # 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:
660
            self.input_batch.remove_request(req_id)
661
662

        # Free the cached encoder outputs.
663
664
        for mm_hash in scheduler_output.free_encoder_mm_hashes:
            self.encoder_cache.pop(mm_hash, None)
665

666
667
668
669
670
671
672
673
674
675
676
677
678
        # 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:
679
            self.input_batch.remove_request(req_id)
680

681
        reqs_to_add: list[CachedRequestState] = []
682
        # Add new requests to the cached states.
683
684
685
        for new_req_data in scheduler_output.scheduled_new_reqs:
            req_id = new_req_data.req_id
            sampling_params = new_req_data.sampling_params
686
            pooling_params = new_req_data.pooling_params
687

688
689
690
691
            if (
                sampling_params
                and sampling_params.sampling_type == SamplingType.RANDOM_SEED
            ):
692
693
694
695
696
                generator = torch.Generator(device=self.device)
                generator.manual_seed(sampling_params.seed)
            else:
                generator = None

697
698
            if self.is_pooling_model:
                assert pooling_params is not None
699
700
                task = pooling_params.task
                assert task is not None, "You did not set `task` in the API"
701

702
                model = cast(VllmModelForPooling, self.get_model())
703
                to_update = model.pooler.get_pooling_updates(task)
704
705
                to_update.apply(pooling_params)

706
            req_state = CachedRequestState(
707
                req_id=req_id,
708
                prompt_token_ids=new_req_data.prompt_token_ids,
709
                prompt_embeds=new_req_data.prompt_embeds,
710
                mm_features=new_req_data.mm_features,
711
                sampling_params=sampling_params,
712
                pooling_params=pooling_params,
713
                generator=generator,
714
715
                block_ids=new_req_data.block_ids,
                num_computed_tokens=new_req_data.num_computed_tokens,
716
                output_token_ids=[],
717
                lora_request=new_req_data.lora_request,
718
            )
719
720
            self.requests[req_id] = req_state

721
            # Only relevant for models using M-RoPE (e.g, Qwen2-VL)
722
            if self.uses_mrope:
723
                self._init_mrope_positions(req_state)
724

725
            reqs_to_add.append(req_state)
726

727
        # Update the states of the running/resumed requests.
728
        is_last_rank = get_pp_group().is_last_rank
729
730
        req_data = scheduler_output.scheduled_cached_reqs
        for i, req_id in enumerate(req_data.req_ids):
731
            req_state = self.requests[req_id]
732
733
            num_computed_tokens = req_data.num_computed_tokens[i]
            new_block_ids = req_data.new_block_ids[i]
734
            resumed_from_preemption = req_id in req_data.resumed_req_ids
735
            num_output_tokens = req_data.num_output_tokens[i]
736

737
            # Update the cached states.
738

739
            req_state.num_computed_tokens = num_computed_tokens
740
            req_index = self.input_batch.req_id_to_index.get(req_id)
741
742
743
744
745
746
747
748

            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.
749
750
751
                num_new_tokens = (
                    num_computed_tokens + len(new_token_ids) - req_state.num_tokens
                )
752
753
754
755
                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:
756
                    req_state.output_token_ids.extend(new_token_ids[-num_new_tokens:])
757
758
759
760
761
            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:
762
763
764
765
                    end_idx = (
                        self.input_batch.num_prompt_tokens[req_index]
                        + num_output_tokens
                    )
766
767
                    self.input_batch.num_tokens[req_index] = end_idx
                    self.input_batch.num_tokens_no_spec[req_index] = end_idx
768

769
            # Update the block IDs.
770
            if not resumed_from_preemption:
771
772
                if new_block_ids is not None:
                    # Append the new blocks to the existing block IDs.
773
                    for block_ids, new_ids in zip(req_state.block_ids, new_block_ids):
774
                        block_ids.extend(new_ids)
775
            else:
776
                assert req_index is None
777
                assert new_block_ids is not None
778
779
                # The request is resumed from preemption.
                # Replace the existing block IDs with the new ones.
780
                req_state.block_ids = new_block_ids
781
782
783
784
785

            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.
786
787
788
789
790
791
792

                if self.use_async_scheduling and num_output_tokens > 0:
                    # We must recover the output token ids for resumed requests in the
                    # async scheduling case, so that correct input_ids are obtained.
                    resumed_token_ids = req_data.all_token_ids[req_id]
                    req_state.output_token_ids = resumed_token_ids[-num_output_tokens:]

793
                reqs_to_add.append(req_state)
794
795
796
                continue

            # Update the persistent batch.
797
            self.input_batch.num_computed_tokens_cpu[req_index] = num_computed_tokens
798
            if new_block_ids is not None:
799
                self.input_batch.block_table.append_row(new_block_ids, req_index)
800
801
802
803
804
805
806

            # 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)
807
                self.input_batch.token_ids_cpu[
808
809
810
                    req_index, start_token_index:end_token_index
                ] = new_token_ids
                self.input_batch.num_tokens_no_spec[req_index] = end_token_index
811
                self.input_batch.num_tokens[req_index] = end_token_index
812

813
            # Add spec_token_ids to token_ids_cpu.
814
            spec_token_ids = scheduler_output.scheduled_spec_decode_tokens.get(
815
                req_id, []
816
            )
817
818
819
820
821
            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[
822
823
                    req_index, start_index:end_token_index
                ] = spec_token_ids
824
825
                # NOTE(woosuk): `num_tokens` here may include spec tokens.
                self.input_batch.num_tokens[req_index] += num_spec_tokens
826
827
828
829
830
831
832

            # 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
833

834
835
        # Add the new or resumed requests to the persistent batch.
        # The smaller empty indices are filled first.
836
837
        for request in reqs_to_add:
            self.input_batch.add_request(request)
838

839
840
841
842
843
844
        # 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()
845

846
    def _update_states_after_model_execute(
847
848
        self, output_token_ids: torch.Tensor
    ) -> None:
849
850
851
852
853
854
855
856
857
858
859
860
        """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.
861
862
863
864
865
866
867
868
869
870
871
872
873
874
875
876
877
878
879
880
        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()
        )
881
882
883
        for i, num_tokens in enumerate(num_accepted_tokens):
            self.input_batch.num_accepted_tokens_cpu[i] = num_tokens

884
885
886
887
888
889
    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
890
891
892
893
        for mm_feature in req_state.mm_features:
            mm_item = mm_feature.data
            if mm_item is None:
                continue
894
895
896
897
898
899
900
901
902
903
904
905
            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

906
907
908
909
910
911
912
913
914
915
916
        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,
917
            )
918
        )
919

920
    def _extract_mm_kwargs(
921
        self,
922
923
        scheduler_output: "SchedulerOutput",
    ) -> BatchedTensorInputs:
924
        if not scheduler_output or not self.is_multimodal_raw_input_only_model:
925
            return {}
926

927
928
        mm_kwargs = list[MultiModalKwargsItem]()
        for req in scheduler_output.scheduled_new_reqs:
929
930
931
            for feature in req.mm_features:
                if feature.data is not None:
                    mm_kwargs.append(feature.data)
932

933
        # Input all modalities at once
934
        model = cast(SupportsMultiModal, self.model)
935
936
        mm_kwargs_combined: BatchedTensorInputs = {}
        for _, _, mm_kwargs_group in group_mm_kwargs_by_modality(
937
938
939
940
            mm_kwargs,
            device=self.device,
            pin_memory=self.pin_memory,
            merge_by_field_config=model.merge_by_field_config,
941
942
        ):
            mm_kwargs_combined.update(mm_kwargs_group)
943

944
        return mm_kwargs_combined
945

946
    def _dummy_mm_kwargs(self, num_seqs: int) -> BatchedTensorInputs:
947
        if not self.is_multimodal_raw_input_only_model:
948
            return {}
949

950
951
952
953
954
        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)
955

956
957
958
    def _get_cumsum_and_arange(
        self,
        num_tokens: np.ndarray,
959
        cumsum_dtype: np.dtype | None = None,
960
961
962
963
964
965
966
967
968
969
970
971
972
973
974
975
    ) -> 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

976
977
978
    def _prepare_input_ids(
        self, total_num_scheduled_tokens: int, cu_num_tokens: np.ndarray
    ) -> None:
979
        """Prepare the input IDs for the current batch.
980

981
982
983
984
985
986
987
        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)
988
989
990
            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)
991
992
993
994
995
996
997
998
999
1000
1001
1002
1003
1004
1005
1006
1007
1008
            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)
1009
                indices_match &= prev_index == flattened_index
1010
1011
1012
1013
1014
1015
                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)
1016
1017
1018
            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)
1019
1020
        if num_commmon_tokens == 0:
            # No requests in common with the previous iteration
1021
            # So input_ids.cpu will have all the input ids.
1022
1023
1024
1025
1026
1027
1028
            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_(
1029
1030
1031
                self.input_batch.prev_sampled_token_ids[:num_commmon_tokens, 0],
                non_blocking=True,
            )
1032
1033
            if self.enable_prompt_embeds:
                self.is_token_ids.gpu[:num_commmon_tokens] = True
1034
            return
1035
        # Upload the index tensors asynchronously so the scatter can be non-blocking.
1036
1037
1038
        input_ids_index_tensor = torch.tensor(
            flattened_indices, dtype=torch.int64, pin_memory=self.pin_memory
        ).to(self.device, non_blocking=True)
1039
        prev_common_req_indices_tensor = torch.tensor(
1040
1041
            prev_common_req_indices, dtype=torch.int64, pin_memory=self.pin_memory
        ).to(self.device, non_blocking=True)
1042
1043
1044
1045
        self.input_ids.gpu.scatter_(
            dim=0,
            index=input_ids_index_tensor,
            src=self.input_batch.prev_sampled_token_ids[
1046
1047
1048
                prev_common_req_indices_tensor, 0
            ],
        )
1049

1050
1051
1052
1053
1054
    def _get_encoder_seq_lens(
        self,
        scheduler_output: "SchedulerOutput",
        kv_cache_spec: KVCacheSpec,
        num_reqs: int,
1055
    ) -> np.ndarray | None:
1056
1057
1058
1059
1060
1061
1062
1063
1064
1065
1066
1067
        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

1068
    def _prepare_inputs(
1069
        self, scheduler_output: "SchedulerOutput"
1070
1071
1072
    ) -> tuple[
        PerLayerAttnMetadata,
        torch.Tensor,
1073
        SpecDecodeMetadata | None,
1074
        np.ndarray,
1075
        CommonAttentionMetadata | None,
1076
        int,
1077
1078
        UBatchSlices | None,
        torch.Tensor | None,
1079
1080
        bool,
    ]:
1081
1082
1083
        """
        :return: tuple[
            attn_metadata: layer-to-attention_metadata mapping,
1084
1085
1086
            logits_indices, spec_decode_metadata,
            num_scheduled_tokens, spec_decode_common_attn_metadata,
            max_num_scheduled_tokens, use_cascade_attn
1087
1088
        ]
        """
1089
1090
1091
1092
1093
1094
1095
        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.
1096
        self.input_batch.block_table.commit_block_table(num_reqs)
1097
1098

        # Get the number of scheduled tokens for each request.
1099
1100
1101
1102
        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)
1103
1104
1105

        # Get request indices.
        # E.g., [2, 5, 3] -> [0, 0, 1, 1, 1, 1, 1, 2, 2, 2]
1106
        req_indices = np.repeat(self.arange_np[:num_reqs], num_scheduled_tokens)
1107

1108
1109
        # cu_num_tokens: [2, 5, 3] -> [2, 7, 10]
        # arange: [0, 1, 0, 1, 2, 3, 4, 0, 1, 2]
1110
        cu_num_tokens, arange = self._get_cumsum_and_arange(num_scheduled_tokens)
1111
1112

        # Get positions.
1113
        positions_np = self.positions.np[:total_num_scheduled_tokens]
1114
1115
1116
1117
1118
        np.add(
            self.input_batch.num_computed_tokens_cpu[req_indices],
            arange,
            out=positions_np,
        )
1119

1120
1121
        # Calculate M-RoPE positions.
        # Only relevant for models using M-RoPE (e.g, Qwen2-VL)
1122
        if self.uses_mrope:
1123
1124
            self._calc_mrope_positions(scheduler_output)

1125
1126
1127
1128
        # 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.
1129
1130
1131
        token_indices = (
            positions_np + req_indices * self.input_batch.token_ids_cpu.shape[1]
        )
1132
        token_indices_tensor = torch.from_numpy(token_indices)
1133

1134
1135
1136
        # 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.
1137
1138
1139
1140
1141
1142
        torch.index_select(
            self.input_batch.token_ids_cpu_tensor.flatten(),
            0,
            token_indices_tensor,
            out=self.input_ids.cpu[:total_num_scheduled_tokens],
        )
1143
        if self.enable_prompt_embeds:
1144
            is_token_ids = self.input_batch.is_token_ids_tensor.flatten()
1145
1146
1147
1148
            torch.index_select(
                is_token_ids,
                0,
                token_indices_tensor,
1149
1150
                out=self.is_token_ids.cpu[:total_num_scheduled_tokens],
            )
1151
1152
1153
1154
1155
1156
1157
1158
1159
1160
1161
1162
1163
1164
1165
1166
1167
1168
1169
1170
1171
1172
1173
1174
1175
1176
1177
1178
1179
1180
1181
1182
1183

        # 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:
1184
1185
1186
                    self.inputs_embeds.cpu[
                        output_idx : output_idx + actual_num_sched
                    ].copy_(req_embeds[start_pos:actual_end])
1187
1188

                output_idx += num_sched
1189

1190
1191
        self.input_batch.block_table.compute_slot_mapping(req_indices, positions_np)
        self.input_batch.block_table.commit_slot_mapping(total_num_scheduled_tokens)
1192
1193

        # Prepare the attention metadata.
1194
        self.query_start_loc.np[0] = 0
1195
        self.query_start_loc.np[1 : num_reqs + 1] = cu_num_tokens
1196
1197
        # Note: pad query_start_loc to be non-decreasing, as kernels
        # like FlashAttention requires that
1198
        self.query_start_loc.np[num_reqs + 1 :].fill(cu_num_tokens[-1])
1199
        self.query_start_loc.copy_to_gpu()
1200
        query_start_loc = self.query_start_loc.gpu[: num_reqs + 1]
1201

1202
        num_tokens_unpadded = scheduler_output.total_num_scheduled_tokens
1203
        num_tokens_padded = self._get_num_input_tokens(num_tokens_unpadded)
1204
1205
1206
        uniform_decode = (
            max_num_scheduled_tokens == self.uniform_decode_query_len
        ) and (total_num_scheduled_tokens == num_reqs * max_num_scheduled_tokens)
1207
1208
1209
1210
1211
1212
1213

        # 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

1214
        ubatch_slices, num_tokens_across_dp = coordinate_batch_across_dp(
1215
1216
1217
1218
1219
1220
1221
            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,
1222
        )
1223

1224
        self.seq_lens.np[:num_reqs] = (
1225
1226
            self.input_batch.num_computed_tokens_cpu[:num_reqs] + num_scheduled_tokens
        )
1227
        # Fill unused with 0 for full cuda graph mode.
1228
1229
1230
1231
        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()
1232

1233
        num_tokens = [self.requests[r].num_tokens for r in self.input_batch.req_ids]
1234
1235
1236
1237
1238
1239
1240
        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)
1241
1242
1243
        self.discard_request_indices.np[: self.num_discarded_requests] = (
            discard_request_indices
        )
1244
1245
1246

        self.discard_request_indices.copy_to_gpu(self.num_discarded_requests)

1247
        # Copy the tensors to the GPU.
1248
1249
        self._prepare_input_ids(total_num_scheduled_tokens, cu_num_tokens)

1250
        if self.uses_mrope:
1251
            # Only relevant for models using M-RoPE (e.g, Qwen2-VL)
1252
1253
            self.mrope_positions.gpu[:, :total_num_scheduled_tokens].copy_(
                self.mrope_positions.cpu[:, :total_num_scheduled_tokens],
1254
1255
                non_blocking=True,
            )
1256
1257
        else:
            # Common case (1D positions)
1258
            self.positions.copy_to_gpu(total_num_scheduled_tokens)
1259

1260
        use_spec_decode = len(scheduler_output.scheduled_spec_decode_tokens) > 0
1261
1262
1263
1264
1265
1266
1267
        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
1268
            num_draft_tokens = None
1269
1270
1271
1272
1273
1274
            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)
1275
1276
1277
            # 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)
1278
1279
1280
1281
            for (
                req_id,
                draft_token_ids,
            ) in scheduler_output.scheduled_spec_decode_tokens.items():
1282
1283
                req_idx = self.input_batch.req_id_to_index[req_id]
                num_draft_tokens[req_idx] = len(draft_token_ids)
1284
1285
1286
1287
1288
1289
1290
1291
                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
                )
1292
            spec_decode_metadata = self._calc_spec_decode_metadata(
1293
1294
                num_draft_tokens, cu_num_tokens
            )
1295
            logits_indices = spec_decode_metadata.logits_indices
1296
1297

            # For DECODE only cuda graph of some attention backends (e.g., GDN).
1298
            self.num_decode_draft_tokens.np[:num_reqs] = num_decode_draft_tokens
1299
1300
            self.num_decode_draft_tokens.np[num_reqs:].fill(-1)
            self.num_decode_draft_tokens.copy_to_gpu()
1301
1302
1303

        logits_indices_padded = None
        if self.cache_config.kv_sharing_fast_prefill:
1304
            logits_indices_padded = self._prepare_kv_sharing_fast_prefill(
1305
1306
                logits_indices
            )
1307

1308
1309
1310
        attn_metadata: PerLayerAttnMetadata = {}
        if ubatch_slices is not None:
            attn_metadata = [dict() for _ in range(len(ubatch_slices))]
1311
        use_cascade_attn = False
1312

1313
        # Used in the below loop.
1314
        query_start_loc_cpu = self.query_start_loc.cpu[: num_reqs + 1]
1315
        seq_lens_cpu = self.seq_lens.cpu[:num_reqs]
1316
1317
1318
        num_computed_tokens_cpu = self.input_batch.num_computed_tokens_cpu_tensor[
            :num_reqs
        ]
1319
        spec_decode_common_attn_metadata = None
1320
1321
        if use_spec_decode:
            self.num_accepted_tokens.np[:num_reqs] = (
1322
1323
                self.input_batch.num_accepted_tokens_cpu[:num_reqs]
            )
1324
1325
            self.num_accepted_tokens.np[num_reqs:].fill(1)
            self.num_accepted_tokens.copy_to_gpu()
1326

1327
1328
1329
        # 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(
1330
1331
            self.kv_cache_config.kv_cache_groups
        ):
1332
            encoder_seq_lens = self._get_encoder_seq_lens(
1333
1334
                scheduler_output, kv_cache_group_spec.kv_cache_spec, num_reqs
            )
1335

1336
            if isinstance(kv_cache_group_spec.kv_cache_spec, EncoderOnlyAttentionSpec):
1337
1338
1339
1340
1341
                # 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,
1342
1343
1344
                    device=self.device,
                )
                slot_mapping = torch.zeros(
1345
                    (total_num_scheduled_tokens,),
1346
1347
1348
                    dtype=torch.int64,
                    device=self.device,
                )
1349
1350
1351
                num_common_prefix_blocks = 0
            else:
                blk_table = self.input_batch.block_table[kv_cache_group_id]
1352
                blk_table_tensor = blk_table.get_device_tensor(num_reqs)
1353
                slot_mapping = blk_table.slot_mapping.gpu[:total_num_scheduled_tokens]
1354
1355
1356

                # Fill unused with -1. Needed for reshape_and_cache in full cuda
                # graph mode.
1357
1358
1359
1360
                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
                ]
1361

1362
            common_attn_metadata = CommonAttentionMetadata(
1363
1364
1365
1366
1367
                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,
1368
1369
1370
                num_reqs=num_reqs,
                num_actual_tokens=total_num_scheduled_tokens,
                max_query_len=max_num_scheduled_tokens,
1371
                max_seq_len=max_seq_len,
1372
1373
                block_table_tensor=blk_table_tensor,
                slot_mapping=slot_mapping,
1374
1375
                logits_indices_padded=logits_indices_padded,
                num_logits_indices=logits_indices.size(0),
1376
                causal=True,
1377
                encoder_seq_lens=encoder_seq_lens,
1378
1379
1380
                dcp_local_seq_lens=self.dcp_local_seq_lens.gpu[:num_reqs]
                if self.dcp_world_size > 1
                else None,
1381
1382
            )

1383
            if self.speculative_config and spec_decode_common_attn_metadata is None:
1384
                if isinstance(self.drafter, EagleProposer):
1385
1386
1387
1388
                    if (
                        self.drafter.attn_layer_names[0]
                        in kv_cache_group_spec.layer_names
                    ):
1389
1390
1391
                        spec_decode_common_attn_metadata = common_attn_metadata
                else:
                    spec_decode_common_attn_metadata = common_attn_metadata
1392

1393
1394
1395
            for attn_group in self.attn_groups[kv_cache_group_id]:
                # Prepare for cascade attention if enabled & beneficial.
                common_prefix_len = 0
1396
                builder = attn_group.get_metadata_builder()
1397
1398
1399
                if self.cascade_attn_enabled:
                    common_prefix_len = self._compute_cascade_attn_prefix_len(
                        num_scheduled_tokens,
1400
                        num_common_prefix_blocks,
1401
                        attn_group.kv_cache_spec,
1402
1403
                        builder,
                    )
1404

1405
                extra_attn_metadata_args = {}
1406
                if use_spec_decode and isinstance(builder, GDNAttentionMetadataBuilder):
1407
                    extra_attn_metadata_args = dict(
1408
1409
1410
1411
                        num_accepted_tokens=self.num_accepted_tokens.gpu[:num_reqs],
                        num_decode_draft_tokens_cpu=self.num_decode_draft_tokens.cpu[
                            :num_reqs
                        ],
1412
1413
                    )

1414
1415
                if ubatch_slices is not None:
                    common_attn_metadata_list = split_attn_metadata(
1416
1417
                        ubatch_slices, common_attn_metadata
                    )
1418
                    for ubid, common_attn_metadata in enumerate(
1419
1420
1421
1422
1423
1424
1425
1426
                        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,
                        )
1427
1428
1429
1430
1431
1432
1433
1434
                        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,
1435
1436
1437
                        **extra_attn_metadata_args,
                    )
                    use_cascade_attn |= getattr(attn_metadata_i, "use_cascade", False)
1438
1439
                    for layer_name in attn_group.layer_names:
                        attn_metadata[layer_name] = attn_metadata_i
1440

1441
1442
1443
1444
        # disable cascade attention when DBO
        if ubatch_slices is not None:
            use_cascade_attn = False

1445
1446
1447
1448
        # Hot-Swap lora model
        if self.lora_config:
            self.set_active_loras(self.input_batch, num_scheduled_tokens)

1449
1450
1451
1452
1453
1454
1455
1456
        return (
            attn_metadata,
            logits_indices,
            spec_decode_metadata,
            num_scheduled_tokens,
            spec_decode_common_attn_metadata,
            max_num_scheduled_tokens,
            ubatch_slices,
1457
            num_tokens_across_dp,
1458
1459
            use_cascade_attn,
        )
1460

1461
1462
1463
1464
    def _compute_cascade_attn_prefix_len(
        self,
        num_scheduled_tokens: np.ndarray,
        num_common_prefix_blocks: int,
1465
1466
        kv_cache_spec: KVCacheSpec,
        attn_metadata_builder: AttentionMetadataBuilder,
1467
1468
1469
1470
1471
1472
1473
1474
1475
1476
1477
1478
1479
1480
1481
1482
1483
1484
    ) -> 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.
        """
1485
        common_prefix_len = num_common_prefix_blocks * kv_cache_spec.block_size
1486
1487
1488
1489
1490
1491
1492
1493
1494
1495
1496
1497
1498
1499
1500
1501
1502
1503
1504
1505
1506
1507
1508
1509
1510
1511
1512
1513
1514
1515
1516
1517
1518
1519
1520
1521
1522
        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]
1523
        # Request 3's num_computed_tokens: 3 (i.e., [A, B, C])
1524
1525
1526
1527
1528
1529
1530
        # 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(
1531
1532
            common_prefix_len, self.input_batch.num_computed_tokens_cpu[:num_reqs].min()
        )
1533
        # common_prefix_len should be a multiple of the block size.
1534
1535
1536
1537
1538
1539
1540
1541
1542
1543
1544
        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
        )
1545
1546
        assert isinstance(kv_cache_spec, AttentionSpec)
        use_cascade = attn_metadata_builder.use_cascade_attention(
1547
1548
1549
            common_prefix_len=common_prefix_len,
            query_lens=num_scheduled_tokens,
            num_query_heads=self.num_query_heads,
1550
            num_kv_heads=kv_cache_spec.num_kv_heads,
1551
            use_alibi=self.use_alibi,
1552
            use_sliding_window=use_sliding_window,
1553
            use_local_attention=use_local_attention,
1554
            num_sms=self.num_sms,
1555
            dcp_world_size=self.dcp_world_size,
1556
1557
1558
        )
        return common_prefix_len if use_cascade else 0

1559
1560
    def _calc_mrope_positions(self, scheduler_output: "SchedulerOutput"):
        mrope_pos_ptr = 0
1561
        for index, req_id in enumerate(self.input_batch.req_ids):
1562
1563
1564
            req = self.requests[req_id]
            assert req.mrope_positions is not None

1565
1566
            num_computed_tokens = self.input_batch.num_computed_tokens_cpu[index]
            num_scheduled_tokens = scheduler_output.num_scheduled_tokens[req_id]
1567
            num_prompt_tokens = length_from_prompt_token_ids_or_embeds(
1568
1569
                req.prompt_token_ids, req.prompt_embeds
            )
1570
1571

            if num_computed_tokens + num_scheduled_tokens > num_prompt_tokens:
1572
1573
                prompt_part_len = max(0, num_prompt_tokens - num_computed_tokens)
                completion_part_len = max(0, num_scheduled_tokens - prompt_part_len)
1574
1575
1576
1577
1578
1579
1580
1581
1582
1583
1584
1585
1586
            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

1587
1588
1589
                self.mrope_positions.cpu[:, dst_start:dst_end] = req.mrope_positions[
                    :, src_start:src_end
                ]
1590
1591
1592
1593
1594
1595
1596
                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

1597
                MRotaryEmbedding.get_next_input_positions_tensor(
1598
                    out=self.mrope_positions.np,
1599
1600
1601
1602
1603
                    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,
                )
1604
1605
1606

                mrope_pos_ptr += completion_part_len

1607
1608
    def _calc_spec_decode_metadata(
        self,
1609
1610
1611
1612
1613
1614
1615
1616
1617
1618
1619
1620
1621
1622
1623
1624
        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
1625
1626
1627
1628

        # 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(
1629
1630
            num_sampled_tokens, cumsum_dtype=np.int32
        )
1631
        # Step 2. [0, 0, 0, 0, 103, 104, 104, 104, 206, 207, 207]
1632
        logits_indices = np.repeat(
1633
1634
            cu_num_scheduled_tokens - num_sampled_tokens, num_sampled_tokens
        )
1635
        # Step 3. [0, 1, 2, 3, 103, 104, 105, 106, 206, 207, 208]
1636
1637
1638
1639
1640
1641
        logits_indices += arange

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

        # Compute the draft logits indices.
1642
1643
1644
        # 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(
1645
1646
            num_draft_tokens, cumsum_dtype=np.int32
        )
1647
1648
        # [0, 0, 0, 5, 5, 9]
        target_logits_indices = np.repeat(
1649
1650
            cu_num_sampled_tokens - num_sampled_tokens, num_draft_tokens
        )
1651
1652
1653
1654
1655
        # [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(
1656
1657
            self.device, non_blocking=True
        )
1658
1659
1660
        cu_num_sampled_tokens = torch.from_numpy(cu_num_sampled_tokens).to(
            self.device, non_blocking=True
        )
1661
1662
1663
        logits_indices = torch.from_numpy(logits_indices).to(
            self.device, non_blocking=True
        )
1664
        target_logits_indices = torch.from_numpy(target_logits_indices).to(
1665
1666
            self.device, non_blocking=True
        )
1667
        bonus_logits_indices = torch.from_numpy(bonus_logits_indices).to(
1668
1669
            self.device, non_blocking=True
        )
1670

1671
1672
        # Compute the draft token ids.
        # draft_token_indices:      [  1,   2,   3, 105, 106, 208]
1673
        draft_token_ids = self.input_ids.gpu[logits_indices]
1674
1675
        draft_token_ids = draft_token_ids[target_logits_indices + 1]

1676
        return SpecDecodeMetadata(
1677
1678
1679
            draft_token_ids=draft_token_ids,
            num_draft_tokens=num_draft_tokens.tolist(),
            cu_num_draft_tokens=cu_num_draft_tokens,
1680
            cu_num_sampled_tokens=cu_num_sampled_tokens,
1681
1682
1683
1684
1685
            target_logits_indices=target_logits_indices,
            bonus_logits_indices=bonus_logits_indices,
            logits_indices=logits_indices,
        )

1686
1687
1688
1689
1690
1691
1692
    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
1693
        self.kv_sharing_fast_prefill_logits_indices[:num_logits].copy_(logits_indices)
1694
1695
1696
1697
1698
        # 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_(
1699
1700
1701
1702
1703
1704
            logits_indices[-1].item()
        )
        if (
            self.compilation_config.cudagraph_mode != CUDAGraphMode.NONE
            and num_logits <= self.cudagraph_batch_sizes[-1]
        ):
1705
1706
1707
1708
1709
            # 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
1710
1711
1712
        logits_indices_padded = self.kv_sharing_fast_prefill_logits_indices[
            :num_logits_padded
        ]
1713
1714
        return logits_indices_padded

1715
1716
1717
1718
1719
1720
1721
1722
    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
1723
                inputs.
1724
1725
1726
1727
1728
1729

        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
        """
1730
1731
        scheduled_encoder_inputs = scheduler_output.scheduled_encoder_inputs
        if not scheduled_encoder_inputs:
1732
            return [], []
1733
        # Batch the multi-modal inputs.
1734
        mm_kwargs = list[MultiModalKwargsItem]()
1735
1736
        # list of tuple (mm_hash, position_info)
        mm_hashes_pos = list[tuple[str, PlaceholderRange]]()
1737
1738
        for req_id, encoder_input_ids in scheduled_encoder_inputs.items():
            req_state = self.requests[req_id]
1739
1740

            for mm_input_id in encoder_input_ids:
1741
1742
1743
1744
                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))
1745

1746
1747
1748
1749
1750
        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(
1751
1752
            scheduler_output
        )
1753
1754
1755
1756

        if not mm_kwargs:
            return

1757
1758
1759
1760
1761
1762
1763
        # 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.
1764
        model = cast(SupportsMultiModal, self.model)
1765
        encoder_outputs = []
1766
        for modality, num_items, mm_kwargs_group in group_mm_kwargs_by_modality(
1767
1768
1769
1770
            mm_kwargs,
            device=self.device,
            pin_memory=self.pin_memory,
            merge_by_field_config=model.merge_by_field_config,
1771
        ):
1772
1773
1774
            curr_group_outputs = []

            # EVS-related change.
1775
            # (ekhvedchenia): Temporary hack to limit peak memory usage when
1776
            # processing multimodal data. This solves the issue with scheduler
1777
1778
1779
1780
            # 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)
1781
1782
1783
1784
1785
1786
1787
1788
1789
1790
1791
1792
1793
1794
1795
1796
1797
            # TODO(ywang96): Fix memory profiling to take EVS into account and
            # remove this hack.
            if (
                self.is_multimodal_pruning_enabled
                and modality == "video"
                and num_items > 1
            ):
                for video_mm_kwargs_item in filter(
                    lambda item: item.modality == "video", mm_kwargs
                ):
                    _, _, micro_batch_mm_inputs = next(
                        group_mm_kwargs_by_modality(
                            [video_mm_kwargs_item],
                            device=self.device,
                            pin_memory=self.pin_memory,
                            merge_by_field_config=model.merge_by_field_config,
                        )
1798
                    )
1799
1800

                    micro_batch_outputs = model.get_multimodal_embeddings(
1801
1802
                        **micro_batch_mm_inputs
                    )
1803
1804
1805
1806
1807
1808
1809
1810
1811
1812

                    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.
1813
                curr_group_outputs = model.get_multimodal_embeddings(**mm_kwargs_group)
1814

1815
1816
            sanity_check_mm_encoder_outputs(
                curr_group_outputs,
1817
                expected_num_items=num_items,
1818
            )
1819
            encoder_outputs.extend(curr_group_outputs)
1820

1821
1822
1823
        # 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(
1824
1825
1826
1827
1828
                output,
                is_embed=pos_info.is_embed,
            )

    def _gather_mm_embeddings(
1829
1830
        self,
        scheduler_output: "SchedulerOutput",
1831
        shift_computed_tokens: int = 0,
1832
1833
1834
1835
1836
1837
1838
1839
    ) -> 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
1840
        should_sync_mrope_positions = False
1841

1842
        for req_id in self.input_batch.req_ids:
1843
1844
            mm_embeds_req: list[torch.Tensor] = []

1845
            num_scheduled_tokens = scheduler_output.num_scheduled_tokens[req_id]
1846
            req_state = self.requests[req_id]
1847
            num_computed_tokens = req_state.num_computed_tokens + shift_computed_tokens
1848

1849
1850
            for mm_feature in req_state.mm_features:
                pos_info = mm_feature.mm_position
1851
1852
                start_pos = pos_info.offset
                num_encoder_tokens = pos_info.length
1853
1854
1855
1856
1857
1858
1859
1860
1861
1862
1863
1864
1865
1866
1867
1868

                # 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,
1869
1870
                    num_encoder_tokens,
                )
1871
                assert start_idx < end_idx
1872

1873
                mm_hash = mm_feature.identifier
1874
                encoder_output = self.encoder_cache.get(mm_hash, None)
1875
                assert encoder_output is not None, f"Encoder cache miss for {mm_hash}."
1876
1877
1878
1879

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

1880
                req_start_pos = req_start_idx + start_pos - num_computed_tokens
1881
1882
1883
                is_mm_embed[req_start_pos + start_idx : req_start_pos + end_idx] = (
                    True if is_embed is None else is_embed
                )
1884

1885
1886
1887
1888
                mm_embeds_item = gather_mm_placeholders(
                    encoder_output[start_idx:end_idx],
                    is_embed=is_embed,
                )
1889
1890
1891
                mm_embeds_req.append(mm_embeds_item)

            if self.is_multimodal_pruning_enabled and self.uses_mrope:
1892
                assert req_state.mrope_positions is not None
1893
1894
1895
1896
1897
1898
1899
                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,
1900
1901
                    )
                )
1902
1903
1904
1905
                req_state.mrope_positions.copy_(new_mrope_positions)
                req_state.mrope_position_delta = new_delta

            mm_embeds.extend(mm_embeds_req)
1906
1907
1908
            req_start_idx += num_scheduled_tokens

        is_mm_embed = self.is_mm_embed.copy_to_gpu(total_num_scheduled_tokens)
1909
1910
1911

        if should_sync_mrope_positions:
            self._calc_mrope_positions(scheduler_output)
1912
            self.mrope_positions.copy_to_gpu(total_num_scheduled_tokens)
1913

1914
        return mm_embeds, is_mm_embed
1915

1916
1917
1918
1919
1920
1921
1922
1923
1924
1925
1926
1927
1928
1929
1930
1931
    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
1932
        model = cast(SupportsMultiModal, self.model)
1933
1934
        encoder_features = {}
        for _, _, mm_kwargs_group in group_mm_kwargs_by_modality(
1935
1936
1937
1938
            mm_kwargs,
            device=self.device,
            pin_memory=self.pin_memory,
            merge_by_field_config=model.merge_by_field_config,
1939
1940
1941
1942
1943
1944
1945
1946
        ):
            # 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

1947
    def get_model(self) -> nn.Module:
1948
        # get raw model out of the cudagraph wrapper.
1949
        if isinstance(self.model, (CUDAGraphWrapper, UBatchWrapper)):
1950
            return self.model.unwrap()
1951
1952
        return self.model

1953
1954
1955
1956
1957
1958
1959
1960
1961
1962
1963
1964
1965
1966
1967
    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

1968
1969
1970
1971
1972
    def get_supported_pooling_tasks(self) -> list[PoolingTask]:
        model = self.get_model()
        if not is_pooling_model(model):
            return []

1973
1974
        supported_tasks = list(model.pooler.get_supported_tasks())

1975
1976
1977
1978
1979
        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")
1980

1981
1982
            logger.debug_once(
                "Chunked prefill is not supported with "
1983
1984
                "token_embed and token_classify tasks "
                "which using ALL pooling. "
1985
1986
1987
                "Please turn off chunked prefill by "
                "`--no-enable-chunked-prefill` before using it."
            )
1988
1989
1990
1991
1992

        if "score" in supported_tasks:
            num_labels = getattr(self.model_config.hf_config, "num_labels", 0)
            if num_labels != 1:
                supported_tasks.remove("score")
1993
                logger.debug_once("Score API is only enabled for num_labels == 1.")
1994
1995

        return supported_tasks
1996

1997
1998
1999
2000
2001
2002
2003
2004
2005
2006
    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)

2007
    def sync_and_slice_intermediate_tensors(
2008
2009
2010
2011
2012
        self,
        num_tokens: int,
        intermediate_tensors: IntermediateTensors,
        sync_self: bool,
    ) -> IntermediateTensors:
2013
2014
2015
        assert self.intermediate_tensors is not None

        tp = self.vllm_config.parallel_config.tensor_parallel_size
2016
        is_rs = is_residual_scattered_for_sp(self.vllm_config, num_tokens)
2017
2018
2019
2020
2021
2022

        # 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():
2023
                is_scattered = k == "residual" and is_rs
2024
                copy_len = num_tokens // tp if is_scattered else num_tokens
2025
                self.intermediate_tensors[k][:copy_len].copy_(
2026
2027
2028
2029
2030
2031
2032
2033
2034
2035
2036
2037
2038
                    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:
2039
2040
2041
2042
2043
2044
2045
        """
        Step for the EPLB (Expert Parallelism Load Balancing) state.
        """
        if not self.parallel_config.enable_eplb:
            return

        assert self.eplb_state is not None
2046
2047
        model = self.get_model()
        assert is_mixture_of_experts(model)
2048
2049
2050
        self.eplb_state.step(
            is_dummy,
            is_profile,
2051
            log_stats=self.parallel_config.eplb_config.log_balancedness,
2052
2053
        )

2054
2055
2056
2057
    # 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)
2058
2059
    @staticmethod
    def pad_out_ubatch_slice(ubatch_slices: UBatchSlices, num_total_tokens: int):
2060
2061
2062
2063
2064
2065
        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
        )
2066

2067
2068
2069
2070
2071
2072
    def _pool(
        self,
        hidden_states: torch.Tensor,
        num_scheduled_tokens: int,
        num_scheduled_tokens_np: np.ndarray,
    ) -> ModelRunnerOutput:
2073
2074
2075
        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"
        )
2076

2077
        hidden_states = hidden_states[:num_scheduled_tokens]
2078
        pooling_metadata = self.input_batch.get_pooling_metadata()
2079
2080
2081
2082
        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]
2083

2084
2085
2086
2087
2088
2089
2090
2091
2092
2093
        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()
2094

2095
        pooler_output: list[torch.Tensor | None] = []
2096
        for raw_output, seq_len, prompt_len in zip(
2097
2098
            raw_pooler_output, seq_lens_cpu, pooling_metadata.prompt_lens
        ):
2099
            output = raw_output if seq_len == prompt_len else None
2100
            pooler_output.append(output)
2101
2102
2103
2104
2105
2106
2107
2108
2109
2110

        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,
        )

2111
    def _get_num_input_tokens(self, num_scheduled_tokens: int) -> int:
2112
2113
2114
2115
2116
2117
        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]
        ):
2118
2119
2120
2121
2122
2123
2124
2125
            # 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
2126
2127
2128
2129
        if (
            self.compilation_config.pass_config.enable_sequence_parallelism
            and tp_size > 1
        ):
2130
2131
2132
            return round_up(num_scheduled_tokens, tp_size)
        return num_scheduled_tokens

2133
    def _preprocess(
2134
2135
        self,
        scheduler_output: "SchedulerOutput",
2136
        num_input_tokens: int,  # Padded
2137
        intermediate_tensors: IntermediateTensors | None = None,
2138
    ) -> tuple[
2139
2140
        torch.Tensor | None,
        torch.Tensor | None,
2141
        torch.Tensor,
2142
        IntermediateTensors | None,
2143
2144
        dict[str, Any],
    ]:
2145
        num_scheduled_tokens = scheduler_output.total_num_scheduled_tokens
2146
        is_first_rank = get_pp_group().is_first_rank
2147

2148
2149
        # _prepare_inputs may reorder the batch, so we must gather multi
        # modal outputs after that to ensure the correct order
2150
2151
        if (
            self.supports_mm_inputs
2152
            and is_first_rank
2153
2154
            and not self.model_config.is_encoder_decoder
        ):
2155
2156
            # Run the multimodal encoder if any.
            self._execute_mm_encoder(scheduler_output)
2157
            mm_embeds, is_mm_embed = self._gather_mm_embeddings(scheduler_output)
2158

2159
2160
2161
            # 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.
2162
            inputs_embeds_scheduled = self.model.get_input_embeddings(
2163
2164
2165
                self.input_ids.gpu[:num_scheduled_tokens],
                multimodal_embeddings=mm_embeds,
                is_multimodal=is_mm_embed,
2166
            )
2167

2168
            # TODO(woosuk): Avoid the copy. Optimize.
2169
            self.inputs_embeds.gpu[:num_scheduled_tokens].copy_(inputs_embeds_scheduled)
2170

2171
            input_ids = None
2172
            inputs_embeds = self.inputs_embeds.gpu[:num_input_tokens]
2173
2174
2175
2176
            model_kwargs = {
                **self._init_model_kwargs(num_scheduled_tokens),
                **self._extract_mm_kwargs(scheduler_output),
            }
2177
        elif self.enable_prompt_embeds and is_first_rank:
2178
2179
2180
2181
2182
2183
2184
2185
2186
2187
2188
2189
            # 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).
2190
2191
2192
            token_ids_idx = (
                self.is_token_ids.gpu[:num_scheduled_tokens]
                .nonzero(as_tuple=False)
2193
                .squeeze(1)
2194
            )
2195
2196
2197
            # Some tokens ids may need to become embeds
            if token_ids_idx.numel() > 0:
                token_ids = self.input_ids.gpu[token_ids_idx]
2198
                tokens_to_embeds = self.model.get_input_embeddings(input_ids=token_ids)
2199
2200
2201
2202
2203
                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
2204
        else:
2205
2206
2207
2208
            # 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.
2209
            input_ids = self.input_ids.gpu[:num_input_tokens]
2210
            inputs_embeds = None
2211
            model_kwargs = self._init_model_kwargs(num_input_tokens)
2212
        if self.uses_mrope:
2213
            positions = self.mrope_positions.gpu[:, :num_input_tokens]
2214
        else:
2215
            positions = self.positions.gpu[:num_input_tokens]
2216

2217
        if is_first_rank:
2218
2219
            intermediate_tensors = None
        else:
2220
            intermediate_tensors = self.sync_and_slice_intermediate_tensors(
2221
2222
                num_input_tokens, intermediate_tensors, True
            )
2223

2224
2225
2226
2227
        if (
            self.model_config.is_encoder_decoder
            and scheduler_output.scheduled_encoder_inputs
        ):
2228
2229
2230
            encoder_inputs = self._extract_encoder_inputs(scheduler_output)
            model_kwargs.update(encoder_inputs)

2231
2232
2233
2234
2235
2236
2237
        return (
            input_ids,
            inputs_embeds,
            positions,
            intermediate_tensors,
            model_kwargs,
        )
2238

2239
    def _sample(
2240
        self,
2241
2242
        logits: torch.Tensor | None,
        spec_decode_metadata: SpecDecodeMetadata | None,
2243
    ) -> SamplerOutput:
2244
        # Sample the next token and get logprobs if needed.
2245
        sampling_metadata = self.input_batch.sampling_metadata
2246
        if spec_decode_metadata is None:
2247
2248
2249
            # 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()
2250
            return self.sampler(
2251
2252
2253
                logits=logits,
                sampling_metadata=sampling_metadata,
            )
2254

2255
        sampler_output = self.rejection_sampler(
2256
2257
            spec_decode_metadata,
            None,  # draft_probs
2258
            logits,
2259
2260
            sampling_metadata,
        )
2261
        self._update_states_after_model_execute(sampler_output.sampled_token_ids)
2262
2263
2264
        return sampler_output

    def _bookkeeping_sync(
2265
2266
2267
        self,
        scheduler_output: "SchedulerOutput",
        sampler_output: SamplerOutput,
2268
        logits: torch.Tensor | None,
2269
2270
        hidden_states: torch.Tensor,
        num_scheduled_tokens: int,
2271
        spec_decode_metadata: SpecDecodeMetadata | None,
2272
    ) -> tuple[
2273
        dict[str, int],
2274
        LogprobsLists | None,
2275
        list[list[int]],
2276
        dict[str, LogprobsTensors | None],
2277
2278
2279
        list[str],
        dict[str, int],
        list[int],
2280
    ]:
2281
2282
2283
2284
        num_nans_in_logits = {}
        if envs.VLLM_COMPUTE_NANS_IN_LOGITS:
            num_nans_in_logits = self._get_nans_in_logits(logits)

2285
2286
2287
        discard_sampled_tokens_req_indices = self.discard_request_indices.np[
            : self.num_discarded_requests
        ]
2288
2289
2290
2291
        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)
2292

2293
2294
2295
        # 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()
2296
        req_id_to_index_output_copy = self.input_batch.req_id_to_index.copy()
2297
2298

        num_sampled_tokens = sampler_output.sampled_token_ids.shape[0]
2299
        sampled_token_ids = sampler_output.sampled_token_ids
2300
        invalid_req_indices = []
2301
2302
2303
2304
2305
2306
2307
2308
2309
2310
2311
2312
2313
2314
        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:
2315
                valid_sampled_token_ids[int(i)].clear()
2316
        else:
2317
            valid_sampled_token_ids = []
2318
            invalid_req_indices = discard_sampled_tokens_req_indices.tolist()
2319
2320
2321
2322
2323
2324
            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.
2325
            self.input_batch.prev_sampled_token_ids = sampled_token_ids
2326
2327
2328
2329
2330
            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
            }
2331

2332
2333
2334
2335
2336
        # 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.
2337
        req_ids = self.input_batch.req_ids
2338
2339
2340
2341
        logprobs_tensors = sampler_output.logprobs_tensors
        cu_num_accepted_tokens = (
            [0] if spec_decode_metadata and logprobs_tensors else None
        )
2342
2343
        for req_idx in range(num_sampled_tokens):
            if self.use_async_scheduling:
2344
                sampled_ids = [-1] if req_idx not in invalid_req_indices_set else None
2345
2346
            else:
                sampled_ids = valid_sampled_token_ids[req_idx]
2347
2348
2349
2350
2351
2352
2353
2354

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

            if cu_num_accepted_tokens is not None:
                cu_num_accepted_tokens.append(
                    cu_num_accepted_tokens[-1] + num_sampled_ids
                )

2355
2356
2357
2358
            if not sampled_ids:
                continue

            start_idx = self.input_batch.num_tokens_no_spec[req_idx]
2359
            end_idx = start_idx + num_sampled_ids
2360
2361
2362
2363
            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}"
2364
            )
2365

2366
2367
            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
2368
2369
            self.input_batch.num_tokens_no_spec[req_idx] = end_idx
            self.input_batch.num_tokens[req_idx] = end_idx
2370

2371
            req_id = req_ids[req_idx]
2372
2373
2374
            req_state = self.requests[req_id]
            req_state.output_token_ids.extend(sampled_ids)

2375
2376
        logprobs_lists = (
            logprobs_tensors.tolists(cu_num_accepted_tokens)
2377
            if not self.use_async_scheduling and logprobs_tensors is not None
2378
2379
2380
2381
2382
2383
2384
2385
2386
            else None
        )

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

2387
2388
2389
2390
2391
2392
2393
2394
2395
2396
        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,
        )

2397
2398
2399
2400
2401
2402
2403
2404
2405
2406
2407
2408
2409
2410
2411
    @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()

2412
2413
    def _model_forward(
        self,
2414
2415
2416
2417
        input_ids: torch.Tensor | None = None,
        positions: torch.Tensor | None = None,
        intermediate_tensors: IntermediateTensors | None = None,
        inputs_embeds: torch.Tensor | None = None,
2418
2419
2420
2421
2422
        **model_kwargs: dict[str, Any],
    ) -> Any:
        """Helper method to call the model forward pass.

        This method can be overridden by subclasses for model execution.
2423
        Motivation: We can inspect only this method versus
2424
2425
2426
2427
2428
2429
2430
2431
2432
2433
2434
2435
2436
2437
2438
2439
2440
2441
2442
2443
        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,
        )

2444
2445
2446
2447
    @torch.inference_mode()
    def execute_model(
        self,
        scheduler_output: "SchedulerOutput",
2448
        intermediate_tensors: IntermediateTensors | None = None,
2449
2450
2451
2452
2453
2454
2455
    ) -> ModelRunnerOutput | IntermediateTensors | None:
        if self.execute_model_state is not None:
            raise RuntimeError(
                "State error: sample_tokens() must be called "
                "after execute_model() returns None."
            )
        num_scheduled_tokens = scheduler_output.total_num_scheduled_tokens
2456
        with record_function_or_nullcontext("Preprocess"):
2457
2458
2459
2460
            with self.synchronize_input_prep():
                # Update persistent batch states.
                self._update_states(scheduler_output)

2461
                if not num_scheduled_tokens:
2462
2463
2464
2465
                    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(
2466
2467
                        scheduler_output, self.vllm_config
                    )
2468
2469
2470
2471
                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 "
2472
2473
                        "it when the requests need prompt logprobs"
                    )
2474

2475
                # Prepare the decoder inputs.
2476
2477
2478
2479
2480
2481
2482
2483
                (
                    attn_metadata,
                    logits_indices,
                    spec_decode_metadata,
                    num_scheduled_tokens_np,
                    spec_decode_common_attn_metadata,
                    max_query_len,
                    ubatch_slices,
2484
                    num_tokens_across_dp,
2485
2486
                    use_cascade_attn,
                ) = self._prepare_inputs(scheduler_output)
2487

2488
            dp_rank = self.parallel_config.data_parallel_rank
2489
2490
            if ubatch_slices:
                assert num_tokens_across_dp is not None
2491
                num_input_tokens = int(num_tokens_across_dp[dp_rank].item())
2492
2493
                self.pad_out_ubatch_slice(ubatch_slices, num_input_tokens)
            elif num_tokens_across_dp is not None:
2494
                num_input_tokens = int(num_tokens_across_dp[dp_rank].item())
2495
2496
2497
2498
2499
            else:
                num_input_tokens = self._get_num_input_tokens(
                    scheduler_output.total_num_scheduled_tokens
                )

2500
2501
2502
2503
2504
2505
            (
                input_ids,
                inputs_embeds,
                positions,
                intermediate_tensors,
                model_kwargs,
2506
            ) = self._preprocess(
2507
                scheduler_output, num_input_tokens, intermediate_tensors
2508
2509
2510
2511
2512
2513
            )

            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(
2514
2515
2516
                num_tokens=num_input_tokens,
                uniform_decode=uniform_decode,
                has_lora=len(self.input_batch.lora_id_to_lora_request) > 0,
2517
2518
2519
2520
            )
            cudagraph_runtime_mode, batch_descriptor = (
                self.cudagraph_dispatcher.dispatch(batch_descriptor, use_cascade_attn)
            )
2521

2522
2523
        # Set cudagraph mode to none if calc_kv_scales is true.
        if attn_metadata is not None:
2524
2525
2526
2527
2528
            metadata_list = (
                attn_metadata.values()
                if isinstance(attn_metadata, dict)
                else [attn_metadata]
            )
2529
            if any(
2530
2531
                getattr(m, "enable_kv_scales_calculation", False) for m in metadata_list
            ):
2532
2533
                cudagraph_runtime_mode = CUDAGraphMode.NONE

2534
2535
        # Run the model.
        # Use persistent buffers for CUDA graphs.
2536
2537
        with (
            set_forward_context(
2538
2539
2540
2541
2542
2543
                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,
2544
                ubatch_slices=ubatch_slices,
2545
2546
2547
2548
            ),
            record_function_or_nullcontext("Forward"),
            self.maybe_get_kv_connector_output(scheduler_output) as kv_connector_output,
        ):
2549
            model_output = self._model_forward(
2550
2551
2552
2553
2554
2555
2556
2557
2558
                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:
2559
                # True when EAGLE 3 is used.
2560
2561
                hidden_states, aux_hidden_states = model_output
            else:
2562
                # Common case.
2563
2564
2565
                hidden_states = model_output
                aux_hidden_states = None

2566
2567
2568
2569
2570
            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)
2571
2572
                    hidden_states.kv_connector_output = kv_connector_output
                    return hidden_states
2573

2574
                if self.is_pooling_model:
2575
                    # Return the pooling output.
2576
2577
2578
                    output = self._pool(
                        hidden_states, num_scheduled_tokens, num_scheduled_tokens_np
                    )
2579
2580
                    output.kv_connector_output = kv_connector_output
                    return output
2581
2582

                sample_hidden_states = hidden_states[logits_indices]
2583
                logits = self.model.compute_logits(sample_hidden_states)
2584
2585
2586
2587
            else:
                # Rare case.
                assert not self.is_pooling_model

2588
                sample_hidden_states = hidden_states[logits_indices]
2589
                if not get_pp_group().is_last_rank:
2590
                    all_gather_tensors = {
2591
2592
2593
                        "residual": not is_residual_scattered_for_sp(
                            self.vllm_config, num_input_tokens
                        )
2594
                    }
2595
                    get_pp_group().send_tensor_dict(
2596
2597
                        hidden_states.tensors,
                        all_gather_group=get_tp_group(),
2598
2599
                        all_gather_tensors=all_gather_tensors,
                    )
2600
2601
                    logits = None
                else:
2602
                    logits = self.model.compute_logits(sample_hidden_states)
2603
2604
2605
2606
2607

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

2608
2609
2610
                model_output_broadcast_data = get_pp_group().broadcast_tensor_dict(
                    model_output_broadcast_data, src=len(get_pp_group().ranks) - 1
                )
2611
2612
2613
                assert model_output_broadcast_data is not None
                logits = model_output_broadcast_data["logits"]

2614
2615
2616
2617
2618
2619
2620
2621
2622
2623
2624
2625
2626
2627
2628
2629
2630
2631
2632
2633
2634
2635
2636
2637
2638
2639
2640
2641
2642
2643
2644
2645
2646
2647
2648
2649
2650
2651
2652
        self.execute_model_state = ExecuteModelState(
            scheduler_output,
            logits,
            spec_decode_metadata,
            spec_decode_common_attn_metadata,
            hidden_states,
            sample_hidden_states,
            aux_hidden_states,
            kv_connector_output,
        )
        return None

    @torch.inference_mode
    def sample_tokens(
        self, grammar_output: "GrammarOutput | None"
    ) -> ModelRunnerOutput | AsyncModelRunnerOutput | IntermediateTensors:
        if self.execute_model_state is None:
            # Nothing to do (PP non-final rank case), output isn't used.
            return None  # noqa

        # Unpack ephemeral state.
        (
            scheduler_output,
            logits,
            spec_decode_metadata,
            spec_decode_common_attn_metadata,
            hidden_states,
            sample_hidden_states,
            aux_hidden_states,
            kv_connector_output,
        ) = self.execute_model_state
        # Clear ephemeral state.
        self.execute_model_state = None

        # Apply structured output bitmasks if present.
        if grammar_output is not None:
            apply_grammar_bitmask(
                scheduler_output, grammar_output, self.input_batch, logits
            )
2653
2654
2655
2656

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

2657
2658
2659
2660
2661
2662
2663
2664
2665
2666
2667
2668
2669
2670
        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,
                )

2671
2672
2673
2674
2675
        use_padded_batch_for_eagle = (
            self.speculative_config
            and self.speculative_config.use_eagle()
            and not self.speculative_config.disable_padded_drafter_batch
        )
2676
2677
2678
        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
2679
2680
2681
2682
2683
        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
        ):
2684
            effective_drafter_max_model_len = (
2685
2686
                self.speculative_config.draft_model_config.max_model_len
            )
2687
        input_fits_in_drafter = spec_decode_common_attn_metadata and (
2688
2689
2690
2691
            spec_decode_common_attn_metadata.max_seq_len
            + self.speculative_config.num_speculative_tokens
            <= effective_drafter_max_model_len
        )
2692
        if use_padded_batch_for_eagle and input_fits_in_drafter:
2693
2694
2695
2696
            # 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)

2697
2698
2699
2700
2701
2702
2703
2704
2705
        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,
2706
2707
2708
2709
2710
            ) = self._bookkeeping_sync(
                scheduler_output,
                sampler_output,
                logits,
                hidden_states,
2711
                scheduler_output.total_num_scheduled_tokens,
2712
                spec_decode_metadata,
2713
            )
2714

2715
2716
2717
2718
2719
        if (
            self.speculative_config
            and not use_padded_batch_for_eagle
            and input_fits_in_drafter
        ):
2720
2721
2722
            # 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)
2723

2724
2725
        with record_function_or_nullcontext("EPLB"):
            self.eplb_step()
2726

2727
2728
2729
        output = ModelRunnerOutput(
            req_ids=req_ids_output_copy,
            req_id_to_index=req_id_to_index_output_copy,
2730
2731
2732
2733
            sampled_token_ids=valid_sampled_token_ids,
            logprobs=logprobs_lists,
            prompt_logprobs_dict=prompt_logprobs_dict,
            pooler_output=[],
2734
            kv_connector_output=kv_connector_output,
2735
2736
2737
            num_nans_in_logits=num_nans_in_logits,
        )

2738
2739
2740
        if not self.use_async_scheduling:
            return output

2741
        async_output = AsyncGPUModelRunnerOutput(
2742
            model_runner_output=output,
2743
            sampled_token_ids=sampler_output.sampled_token_ids,
2744
            logprobs_tensors=sampler_output.logprobs_tensors,
2745
2746
2747
2748
            invalid_req_indices=invalid_req_indices,
            async_output_copy_stream=self.async_output_copy_stream,
        )

2749
2750
2751
2752
2753
2754
2755
2756
2757
        # 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

2758
    def take_draft_token_ids(self) -> DraftTokenIds | None:
2759
2760
2761
2762
2763
2764
2765
2766
2767
2768
        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)

2769
2770
2771
    def propose_draft_token_ids(
        self,
        scheduler_output: "SchedulerOutput",
2772
        sampled_token_ids: torch.Tensor | list[list[int]],
2773
2774
2775
        sampling_metadata: SamplingMetadata,
        hidden_states: torch.Tensor,
        sample_hidden_states: torch.Tensor,
2776
2777
        aux_hidden_states: list[torch.Tensor] | None,
        spec_decode_metadata: SpecDecodeMetadata | None,
2778
        common_attn_metadata: CommonAttentionMetadata,
2779
    ) -> list[list[int]] | torch.Tensor:
2780
2781
        num_scheduled_tokens = scheduler_output.total_num_scheduled_tokens
        if self.speculative_config.method == "ngram":
2782
            assert isinstance(sampled_token_ids, list)
2783
            assert isinstance(self.drafter, NgramProposer)
2784
            draft_token_ids = self.drafter.propose(
2785
2786
                sampled_token_ids,
                self.input_batch.req_ids,
2787
2788
                self.input_batch.num_tokens_no_spec,
                self.input_batch.token_ids_cpu,
2789
2790
                self.input_batch.spec_decode_unsupported_reqs,
            )
2791
2792
2793
2794
        elif self.speculative_config.method == "suffix":
            assert isinstance(sampled_token_ids, list)
            assert isinstance(self.drafter, SuffixDecodingProposer)
            draft_token_ids = self.drafter.propose(self.input_batch, sampled_token_ids)
2795
        elif self.speculative_config.method == "medusa":
2796
            assert isinstance(sampled_token_ids, list)
2797
            assert isinstance(self.drafter, MedusaProposer)
2798

2799
2800
            if sample_hidden_states.shape[0] == len(sampled_token_ids):
                # The input to the target model does not include draft tokens.
2801
2802
2803
2804
                hidden_states = sample_hidden_states
            else:
                indices = []
                offset = 0
2805
2806
2807
                assert spec_decode_metadata is not None, (
                    "No spec decode metadata for medusa"
                )
2808
                for num_draft, tokens in zip(
2809
2810
                    spec_decode_metadata.num_draft_tokens, sampled_token_ids
                ):
2811
2812
                    indices.append(offset + len(tokens) - 1)
                    offset += num_draft + 1
2813
                indices = torch.tensor(indices, device=self.device)
2814
2815
                hidden_states = sample_hidden_states[indices]

2816
            draft_token_ids = self.drafter.propose(
2817
2818
2819
                target_hidden_states=hidden_states,
                sampling_metadata=sampling_metadata,
            )
2820
        elif self.speculative_config.use_eagle():
2821
            assert isinstance(self.drafter, EagleProposer)
2822
2823
2824
2825
2826

            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.
2827
2828
                assert isinstance(sampled_token_ids, list), (
                    "sampled_token_ids should be a python list when"
2829
                    "padded-batch is disabled."
2830
                )
2831
                next_token_ids = self.drafter.prepare_next_token_ids_cpu(
2832
2833
2834
2835
2836
                    sampled_token_ids,
                    self.requests,
                    self.input_batch,
                    scheduler_output.num_scheduled_tokens,
                )
2837
2838
2839
2840
2841
            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.
2842
2843
                assert isinstance(sampled_token_ids, torch.Tensor), (
                    "sampled_token_ids should be a torch.Tensor when"
2844
                    "padded-batch is enabled."
2845
2846
                )
                next_token_ids, valid_sampled_tokens_count = (
2847
2848
2849
2850
2851
2852
                    self.drafter.prepare_next_token_ids_padded(
                        common_attn_metadata,
                        sampled_token_ids,
                        self.requests,
                        self.input_batch,
                        self.discard_request_indices.gpu,
2853
                        self.num_discarded_requests,
2854
                    )
2855
                )
Jiayi Yao's avatar
Jiayi Yao committed
2856

2857
            if spec_decode_metadata is None:
2858
                token_indices_to_sample = None
2859
                # input_ids can be None for multimodal models.
2860
                target_token_ids = self.input_ids.gpu[:num_scheduled_tokens]
2861
                target_positions = self._get_positions(num_scheduled_tokens)
2862
                if self.use_aux_hidden_state_outputs:
Wentao Ye's avatar
Wentao Ye committed
2863
                    assert aux_hidden_states is not None
2864
                    target_hidden_states = torch.cat(
2865
2866
                        [h[:num_scheduled_tokens] for h in aux_hidden_states], dim=-1
                    )
2867
2868
                else:
                    target_hidden_states = hidden_states[:num_scheduled_tokens]
2869
            else:
2870
2871
                if self.speculative_config.disable_padded_drafter_batch:
                    token_indices_to_sample = None
2872
2873
2874
2875
2876
                    common_attn_metadata, token_indices = self.drafter.prepare_inputs(
                        common_attn_metadata,
                        sampled_token_ids,
                        spec_decode_metadata.num_draft_tokens,
                    )
2877
                else:
2878
                    common_attn_metadata, token_indices, token_indices_to_sample = (
2879
2880
2881
                        self.drafter.prepare_inputs_padded(
                            common_attn_metadata,
                            spec_decode_metadata,
2882
2883
2884
                            valid_sampled_tokens_count,
                        )
                    )
2885

2886
                target_token_ids = self.input_ids.gpu[token_indices]
2887
                target_positions = self._get_positions(token_indices)
2888
                if self.use_aux_hidden_state_outputs:
Wentao Ye's avatar
Wentao Ye committed
2889
                    assert aux_hidden_states is not None
2890
                    target_hidden_states = torch.cat(
2891
2892
                        [h[token_indices] for h in aux_hidden_states], dim=-1
                    )
2893
2894
                else:
                    target_hidden_states = hidden_states[token_indices]
2895

2896
            if self.supports_mm_inputs:
2897
2898
2899
2900
2901
2902
                mm_embed_inputs = self._gather_mm_embeddings(
                    scheduler_output,
                    shift_computed_tokens=1,
                )
            else:
                mm_embed_inputs = None
2903

2904
            draft_token_ids = self.drafter.propose(
2905
2906
2907
2908
                target_token_ids=target_token_ids,
                target_positions=target_positions,
                target_hidden_states=target_hidden_states,
                next_token_ids=next_token_ids,
2909
                last_token_indices=token_indices_to_sample,
2910
                sampling_metadata=sampling_metadata,
2911
                common_attn_metadata=common_attn_metadata,
2912
                mm_embed_inputs=mm_embed_inputs,
2913
            )
2914

2915
        return draft_token_ids
2916

2917
2918
2919
    def update_config(self, overrides: dict[str, Any]) -> None:
        allowed_config_names = {"load_config", "model_config"}
        for config_name, config_overrides in overrides.items():
2920
2921
            assert config_name in allowed_config_names, (
                f"Config `{config_name}` not supported. "
2922
                f"Allowed configs: {allowed_config_names}"
2923
            )
2924
2925
2926
2927
            config = getattr(self, config_name)
            new_config = update_config(config, config_overrides)
            setattr(self, config_name, new_config)

2928
2929
2930
2931
2932
    def load_model(self, eep_scale_up: bool = False) -> None:
        """
        Args:
            eep_scale_up: the model loading is for elastic EP scale up.
        """
2933
2934
2935
2936
2937
        logger.info_once(
            "Starting to load model %s...",
            self.model_config.model,
            scope="global",
        )
2938
2939
2940
2941
2942
        global_expert_loads, old_global_expert_indices_per_model, rank_mapping = (
            EplbState.get_eep_state(self.parallel_config)
            if eep_scale_up
            else (None, None, None)
        )
2943

2944
2945
2946
        if self.parallel_config.enable_eplb:
            self.eplb_state = EplbState(self.parallel_config, self.device)
            eplb_models = 0
2947
        with DeviceMemoryProfiler() as m:
2948
            time_before_load = time.perf_counter()
2949
            model_loader = get_model_loader(self.load_config)
2950
            self.model = model_loader.load_model(
2951
2952
                vllm_config=self.vllm_config, model_config=self.model_config
            )
2953
            if self.lora_config:
2954
2955
2956
                self.model = self.load_lora_model(
                    self.model, self.vllm_config, self.device
                )
2957
            if hasattr(self, "drafter"):
2958
                logger.info_once("Loading drafter model...")
2959
                self.drafter.load_model(self.model)
2960
2961
2962
2963
2964
2965
2966
2967
2968
2969
2970
2971
2972
2973
2974
2975
2976
2977
2978
2979
2980
2981
2982
2983
2984
2985
2986
2987
2988
2989
2990
                if (
                    hasattr(self.drafter, "model")
                    and is_mixture_of_experts(self.drafter.model)
                    and self.parallel_config.enable_eplb
                ):
                    logger.info_once(
                        "EPLB is enabled for drafter model %s.",
                        self.vllm_config.speculative_config.draft_model_config.model,
                    )

                    global_expert_load = (
                        global_expert_loads[eplb_models]
                        if global_expert_loads
                        else None
                    )
                    old_global_expert_indices = (
                        old_global_expert_indices_per_model[eplb_models]
                        if old_global_expert_indices_per_model
                        else None
                    )
                    if self.eplb_state is None:
                        self.eplb_state = EplbState(self.parallel_config, self.device)
                    self.eplb_state.add_model(
                        self.drafter.model,
                        self.vllm_config.speculative_config.draft_model_config,
                        global_expert_load,
                        old_global_expert_indices,
                        rank_mapping,
                    )
                    eplb_models += 1

2991
            if self.use_aux_hidden_state_outputs:
2992
                if not supports_eagle3(self.get_model()):
2993
2994
                    raise RuntimeError(
                        "Model does not support EAGLE3 interface but "
2995
2996
                        "aux_hidden_state_outputs was requested"
                    )
2997
2998
2999
3000
3001
3002
3003
3004
3005
3006
3007
3008
3009

                # 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)
3010
            time_after_load = time.perf_counter()
3011
        self.model_memory_usage = m.consumed_memory
3012
        logger.info_once(
3013
3014
3015
            "Model loading took %.4f GiB and %.6f seconds",
            self.model_memory_usage / GiB_bytes,
            time_after_load - time_before_load,
3016
            scope="local",
3017
        )
3018
        prepare_communication_buffer_for_model(self.model)
3019
        self.is_multimodal_pruning_enabled = (
3020
            supports_multimodal_pruning(self.get_model())
3021
3022
            and self.model_config.multimodal_config.is_multimodal_pruning_enabled()
        )
3023

3024
        if is_mixture_of_experts(self.model) and self.parallel_config.enable_eplb:
3025
3026
3027
3028
3029
3030
3031
3032
3033
3034
3035
            logger.info_once("EPLB is enabled for model %s.", self.model_config.model)
            global_expert_load = (
                global_expert_loads[eplb_models] if global_expert_loads else None
            )
            old_global_expert_indices = (
                old_global_expert_indices_per_model[eplb_models]
                if old_global_expert_indices_per_model
                else None
            )
            assert self.eplb_state is not None
            self.eplb_state.add_model(
3036
                self.model,
3037
                self.model_config,
3038
3039
3040
                global_expert_load,
                old_global_expert_indices,
                rank_mapping,
3041
3042
            )

3043
        if (
3044
3045
            self.vllm_config.compilation_config.mode
            == CompilationMode.STOCK_TORCH_COMPILE
3046
            and supports_dynamo()
3047
        ):
3048
            backend = self.vllm_config.compilation_config.init_backend(self.vllm_config)
3049
            compilation_counter.stock_torch_compile_count += 1
3050
            self.model.compile(fullgraph=True, backend=backend)
3051
            return
3052
        # for other compilation modes, cudagraph behavior is controlled by
3053
3054
3055
        # CudagraphWraper and CudagraphDispatcher of vllm.

        # wrap the model with full cudagraph wrapper if needed.
3056
3057
3058
3059
3060
3061
3062
        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
            )
3063
3064
        elif self.parallel_config.enable_dbo:
            if self.compilation_config.cudagraph_mode.has_full_cudagraphs():
3065
3066
3067
                self.model = UBatchWrapper(
                    self.model, self.vllm_config, CUDAGraphMode.FULL, self.device
                )
3068
            else:
3069
3070
3071
                self.model = UBatchWrapper(
                    self.model, self.vllm_config, CUDAGraphMode.NONE, self.device
                )
3072

3073
    def _get_eagle3_aux_layers_from_config(self) -> tuple[int, ...] | None:
3074
3075
3076
3077
3078
3079
3080
3081
3082
3083
3084
3085
3086
3087
3088
3089
3090
3091
3092
3093
3094
3095
3096
        """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

3097
    def reload_weights(self) -> None:
3098
        assert getattr(self, "model", None) is not None, (
3099
            "Cannot reload weights before model is loaded."
3100
        )
3101
3102
        model_loader = get_model_loader(self.load_config)
        logger.info("Reloading weights inplace...")
3103
        model_loader.load_weights(self.get_model(), model_config=self.model_config)
3104

3105
3106
3107
3108
3109
    def save_tensorized_model(
        self,
        tensorizer_config: "TensorizerConfig",
    ) -> None:
        TensorizerLoader.save_model(
3110
            self.get_model(),
3111
            tensorizer_config=tensorizer_config,
3112
            model_config=self.model_config,
3113
3114
        )

3115
3116
3117
    def _get_prompt_logprobs_dict(
        self,
        hidden_states: torch.Tensor,
3118
        num_scheduled_tokens: dict[str, int],
3119
    ) -> dict[str, LogprobsTensors | None]:
3120
3121
3122
3123
        num_prompt_logprobs_dict = self.input_batch.num_prompt_logprobs
        if not num_prompt_logprobs_dict:
            return {}

3124
        in_progress_dict = self.input_batch.in_progress_prompt_logprobs_cpu
3125
        prompt_logprobs_dict: dict[str, LogprobsTensors | None] = {}
3126
3127
3128
3129
3130

        # 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():
3131
            num_tokens = num_scheduled_tokens[req_id]
3132
3133
3134

            # Get metadata for this request.
            request = self.requests[req_id]
3135
3136
3137
3138
            if request.prompt_token_ids is None:
                # Prompt logprobs is incompatible with prompt embeddings
                continue

3139
3140
            num_prompt_tokens = len(request.prompt_token_ids)
            prompt_token_ids = torch.tensor(request.prompt_token_ids).to(
3141
3142
                self.device, non_blocking=True
            )
3143

3144
3145
3146
3147
3148
3149
            # 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(
3150
3151
                    num_prompt_tokens - 1, num_prompt_logprobs + 1
                )
3152
3153
                in_progress_dict[req_id] = logprobs_tensors

3154
            # Determine number of logits to retrieve.
3155
3156
            start_idx = request.num_computed_tokens
            start_tok = start_idx + 1
3157
            num_remaining_tokens = num_prompt_tokens - start_tok
3158
            if num_tokens <= num_remaining_tokens:
3159
                # This is a chunk, more tokens remain.
3160
3161
3162
                # 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.
3163
3164
3165
3166
3167
                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)
3168
3169
3170
3171
3172
3173
3174
                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
3175
3176
3177
3178
3179

            # 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]
3180
            offset = self.query_start_loc.np[req_idx].item()
3181
            prompt_hidden_states = hidden_states[offset : offset + num_logits]
3182
            logits = self.model.compute_logits(prompt_hidden_states)
3183
3184
3185
3186

            # 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.
3187
            tgt_token_ids = prompt_token_ids[start_tok : start_tok + num_logits]
3188
3189

            # Compute prompt logprobs.
3190
3191
            logprobs = self.sampler.compute_logprobs(logits)
            token_ids, logprobs, ranks = self.sampler.gather_logprobs(
3192
3193
                logprobs, num_prompt_logprobs, tgt_token_ids
            )
3194
3195

            # Transfer GPU->CPU async.
3196
3197
            chunk_slice = slice(start_idx, start_idx + num_logits)
            logprobs_tensors.logprob_token_ids[chunk_slice].copy_(
3198
3199
3200
                token_ids, non_blocking=True
            )
            logprobs_tensors.logprobs[chunk_slice].copy_(logprobs, non_blocking=True)
3201
            logprobs_tensors.selected_token_ranks[chunk_slice].copy_(
3202
3203
                ranks, non_blocking=True
            )
3204
3205
3206
3207
3208

        # 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]
3209
            del in_progress_dict[req_id]
3210
3211

        # Must synchronize the non-blocking GPU->CPU transfers.
3212
        if prompt_logprobs_dict:
3213
            self._sync_device()
3214
3215
3216

        return prompt_logprobs_dict

3217
3218
    def _get_nans_in_logits(
        self,
3219
        logits: torch.Tensor | None,
3220
3221
3222
3223
3224
3225
3226
3227
3228
3229
3230
    ) -> 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])
3231
3232
3233
                    if num_nans_for_index is not None and req_index < logits.shape[0]
                    else 0
                )
3234
3235
3236
3237
            return num_nans_in_logits
        except IndexError:
            return {}

3238
3239
3240
3241
3242
3243
    @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
3244
         - during DP rank dummy run
3245
3246
3247
3248
3249
3250
3251
3252
3253
3254
3255
        """
        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(
3256
                    self.input_ids.gpu,
3257
3258
                    low=0,
                    high=self.model_config.get_vocab_size(),
3259
3260
                    dtype=input_ids.dtype,
                )
3261

3262
            logger.debug_once("Randomizing dummy data for DP Rank")
3263
            input_ids.copy_(rand_input_ids()[: input_ids.size(0)], non_blocking=True)
3264
3265
3266
            yield
            input_ids.fill_(0)

3267
3268
3269
3270
3271
3272
    def _get_mm_dummy_batch(
        self,
        modality: str,
        max_items_per_batch: int,
    ) -> BatchedTensorInputs:
        """Dummy data for profiling and precompiling multimodal models."""
3273
3274
        assert self.mm_budget is not None

3275
3276
        dummy_decoder_data = self.mm_registry.get_decoder_dummy_data(
            model_config=self.model_config,
3277
            seq_len=self.max_model_len,
3278
            mm_counts={modality: 1},
3279
            cache=self.mm_budget.cache,
3280
3281
3282
3283
        )
        dummy_mm_data = dummy_decoder_data.multi_modal_data

        # Result in the maximum GPU consumption of the model
3284
3285
        dummy_mm_item = dummy_mm_data[modality][0]
        dummy_mm_items = [dummy_mm_item] * max_items_per_batch
3286

3287
        model = cast(SupportsMultiModal, self.model)
3288
3289
3290
3291
3292
3293
3294
3295
3296
        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,
            )
        )
3297

3298
3299
3300
3301
    @torch.inference_mode()
    def _dummy_run(
        self,
        num_tokens: int,
3302
        cudagraph_runtime_mode: CUDAGraphMode | None = None,
3303
3304
        force_attention: bool = False,
        uniform_decode: bool = False,
3305
        allow_microbatching: bool = True,
3306
3307
        skip_eplb: bool = False,
        is_profile: bool = False,
3308
        create_mixed_batch: bool = False,
3309
        remove_lora: bool = True,
3310
        activate_lora: bool = False,
3311
    ) -> tuple[torch.Tensor, torch.Tensor]:
3312
3313
3314
3315
3316
3317
3318
        """
        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.
3319
                - if not set will determine the cudagraph mode based on using
3320
                    the self.cudagraph_dispatcher.
3321
3322
3323
3324
                - CUDAGraphMode.NONE: No cudagraph, for warm up and profile run
                - CUDAGraphMode.PIECEWISE: Piecewise cudagraph.
                - CUDAGraphMode.FULL: Full cudagraph, attention metadata is
                    needed.
3325
            force_attention: If True, always create attention metadata. Used to
3326
3327
3328
3329
                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.
3330
3331
            create_mixed_batch: If True, create a mixed batch with both decode
                (1 token) and prefill (multiple tokens) requests.
3332
            remove_lora: If False, dummy LoRAs are not destroyed after the run
3333
            activate_lora: If False, dummy_run is performed without LoRAs.
3334
        """
3335
3336
3337
3338
        assert (
            cudagraph_runtime_mode is None
            or cudagraph_runtime_mode.valid_runtime_modes()
        )
3339

3340
        # If cudagraph_mode.decode_mode() == FULL and
3341
        # cudagraph_mode.separate_routine(). This means that we are using
3342
3343
3344
3345
3346
3347
3348
3349
3350
3351
3352
        # 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.
3353
        max_query_len = self.uniform_decode_query_len if uniform_decode else num_tokens
3354

3355
3356
3357
3358
3359
        # 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
3360
3361
3362
3363
        if create_mixed_batch:
            assert not uniform_decode
            # Create mixed batch:
            # first half decode tokens, second half one prefill
3364
            num_decode_tokens = min(max_num_reqs - 1, num_tokens // 2)
3365
3366
3367
3368
            num_prefill_tokens = num_tokens - num_decode_tokens
            num_reqs = num_decode_tokens + 1

            # Create decode requests (1 token each) followed by prefill request
3369
            num_scheduled_tokens_list = [1] * num_decode_tokens + [num_prefill_tokens]
3370
3371
3372
            # Note: Overriding max_query_len to be the prefill tokens
            max_query_len = num_prefill_tokens
        elif uniform_decode:
3373
            assert not create_mixed_batch
3374
            num_reqs = min(max_num_reqs, cdiv(num_tokens, max_query_len))
3375
3376
            num_scheduled_tokens_list = [max_query_len] * num_reqs
            if num_tokens % max_query_len != 0:
3377
                num_scheduled_tokens_list[-1] = num_tokens % max_query_len
3378
3379
3380
3381
3382
3383
        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

3384
3385
        assert sum(num_scheduled_tokens_list) == num_tokens
        assert len(num_scheduled_tokens_list) == num_reqs
3386
        num_scheduled_tokens = np.array(num_scheduled_tokens_list, dtype=np.int32)
3387
        total_num_scheduled_tokens = int(num_scheduled_tokens.sum())
3388

3389
3390
3391
        # Disable DP padding when running eager
        allow_dp_padding = self.compilation_config.cudagraph_mode != CUDAGraphMode.NONE

3392
3393
        # We currently only microbatch if the number of tokens is
        # over a certain threshold.
3394
        ubatch_slices, num_tokens_across_dp = coordinate_batch_across_dp(
3395
3396
3397
3398
3399
3400
3401
            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,
3402
3403
3404
        )
        num_tokens_after_padding = num_tokens
        if num_tokens_across_dp is not None:
3405
3406
            dp_rank = self.parallel_config.data_parallel_rank
            num_tokens_after_padding = int(num_tokens_across_dp[dp_rank])
3407

3408
        attn_metadata: PerLayerAttnMetadata | None = None
3409
3410
3411

        # If force_attention is True, we always capture attention. Otherwise,
        # it only happens for cudagraph_runtime_mode=FULL.
3412
        if force_attention or cudagraph_runtime_mode == CUDAGraphMode.FULL:
3413
            attn_metadata = {}
3414
3415
            if ubatch_slices is not None:
                attn_metadata = [dict() for _ in range(len(ubatch_slices))]
3416

3417
3418
3419
3420
3421
3422
            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:
3423
                seq_lens = max_query_len
3424
            self.seq_lens.np[:num_reqs] = seq_lens
3425
3426
            self.seq_lens.np[num_reqs:] = 0
            self.seq_lens.copy_to_gpu()
3427

3428
3429
            cum_num_tokens, _ = self._get_cumsum_and_arange(num_scheduled_tokens)
            self.query_start_loc.np[1 : num_reqs + 1] = cum_num_tokens
3430
3431
            self.query_start_loc.copy_to_gpu()

3432
            for kv_cache_group_id, kv_cache_group_spec in enumerate(
3433
3434
                self.kv_cache_config.kv_cache_groups
            ):
3435
                common_attn_metadata = CommonAttentionMetadata(
3436
3437
                    query_start_loc=self.query_start_loc.gpu[: num_reqs + 1],
                    query_start_loc_cpu=self.query_start_loc.cpu[: num_reqs + 1],
3438
3439
                    seq_lens=self.seq_lens.gpu[:num_reqs],
                    seq_lens_cpu=self.seq_lens.cpu[:num_reqs],
3440
3441
3442
                    num_computed_tokens_cpu=self.input_batch.num_computed_tokens_cpu_tensor[
                        :num_reqs
                    ],
3443
3444
                    num_reqs=num_reqs,
                    num_actual_tokens=num_tokens,
3445
                    max_query_len=max_query_len,
3446
                    max_seq_len=self.max_model_len,
3447
3448
3449
                    block_table_tensor=self.input_batch.block_table[
                        kv_cache_group_id
                    ].get_device_tensor(num_reqs),
3450
                    slot_mapping=self.input_batch.block_table[
3451
3452
3453
                        kv_cache_group_id
                    ].slot_mapping.gpu[:num_tokens],
                    causal=True,
3454
3455
3456
                    dcp_local_seq_lens=self.dcp_local_seq_lens.gpu[:num_reqs]
                    if self.dcp_world_size > 1
                    else None,
3457
                )
3458
                for attn_group in self.attn_groups[kv_cache_group_id]:
3459
3460
                    if ubatch_slices is not None:
                        common_attn_metadata_list = split_attn_metadata(
3461
3462
                            ubatch_slices, common_attn_metadata
                        )
3463
                        for ubid, common_attn_metadata in enumerate(
3464
3465
                            common_attn_metadata_list
                        ):
3466
                            assert common_attn_metadata.max_query_len == 1
3467
3468
3469
                            attn_metadata_i = attn_group.get_metadata_builder(
                                ubatch_id=ubid
                            ).build_for_cudagraph_capture(common_attn_metadata)
3470
                            for layer_name in attn_group.layer_names:
3471
                                assert type(attn_metadata) is list
3472
                                attn_metadata[ubid][layer_name] = attn_metadata_i
3473
3474
                    else:
                        assert type(attn_metadata) is dict
3475
3476
                        metadata_builder = attn_group.get_metadata_builder()
                        attn_metadata_i = metadata_builder.build_for_cudagraph_capture(
3477
3478
                            common_attn_metadata
                        )
3479
                        for layer_name in attn_group.layer_names:
3480
                            attn_metadata[layer_name] = attn_metadata_i
3481

3482
        with self.maybe_dummy_run_with_lora(
3483
            self.lora_config, num_scheduled_tokens, activate_lora, remove_lora
3484
        ):
3485
3486
3487
            # 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)
3488
            if self.supports_mm_inputs and not self.model_config.is_encoder_decoder:
3489
                input_ids = None
3490
                inputs_embeds = self.inputs_embeds.gpu[:num_tokens_after_padding]
3491
                model_kwargs = {
3492
                    **model_kwargs,
3493
3494
                    **self._dummy_mm_kwargs(num_reqs),
                }
3495
3496
            elif self.enable_prompt_embeds:
                input_ids = None
3497
3498
                inputs_embeds = self.inputs_embeds.gpu[:num_tokens_after_padding]
                model_kwargs = self._init_model_kwargs(num_tokens_after_padding)
3499
            else:
3500
                input_ids = self.input_ids.gpu[:num_tokens_after_padding]
3501
                inputs_embeds = None
3502

3503
            if self.uses_mrope:
3504
                positions = self.mrope_positions.gpu[:, :num_tokens_after_padding]
3505
            else:
3506
                positions = self.positions.gpu[:num_tokens_after_padding]
3507
3508
3509
3510
3511
3512
3513
3514
3515

            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,
3516
3517
3518
                            device=self.device,
                        )
                    )
3519
3520

                intermediate_tensors = self.sync_and_slice_intermediate_tensors(
3521
                    num_tokens_after_padding, None, False
3522
                )
3523
3524

            # filter out the valid batch descriptor
3525
3526
3527
3528
3529
            _cg_mode, batch_descriptor = (
                self.cudagraph_dispatcher.dispatch(
                    BatchDescriptor(
                        num_tokens=num_tokens_after_padding,
                        uniform_decode=uniform_decode,
3530
                        has_lora=activate_lora and self.lora_config is not None,
3531
3532
3533
3534
3535
                    )
                )
                if not is_profile
                else (CUDAGraphMode.NONE, None)
            )
3536
3537
3538
            if cudagraph_runtime_mode is not None:
                # we allow forcing NONE when the dispatcher disagrees to support
                # warm ups for cudagraph capture
3539
3540
3541
3542
                assert (
                    cudagraph_runtime_mode == CUDAGraphMode.NONE
                    or cudagraph_runtime_mode == _cg_mode
                ), (
3543
                    f"Cudagraph runtime mode mismatch at dummy_run. "
3544
3545
                    f"Expected {_cg_mode}, but got {cudagraph_runtime_mode}."
                )
3546
3547
            else:
                cudagraph_runtime_mode = _cg_mode
3548

3549
            if ubatch_slices is not None:
3550
3551
3552
3553
3554
3555
3556
                # 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

3557
3558
3559
            with (
                self.maybe_randomize_inputs(input_ids),
                set_forward_context(
3560
3561
                    attn_metadata,
                    self.vllm_config,
3562
                    num_tokens=num_tokens_after_padding,
3563
3564
                    num_tokens_across_dp=num_tokens_across_dp,
                    cudagraph_runtime_mode=cudagraph_runtime_mode,
3565
                    batch_descriptor=batch_descriptor,
3566
3567
3568
                    ubatch_slices=ubatch_slices,
                ),
            ):
3569
                outputs = self.model(
3570
3571
3572
3573
                    input_ids=input_ids,
                    positions=positions,
                    intermediate_tensors=intermediate_tensors,
                    inputs_embeds=inputs_embeds,
3574
                    **model_kwargs,
3575
                )
3576

3577
3578
3579
3580
            if self.use_aux_hidden_state_outputs:
                hidden_states, _ = outputs
            else:
                hidden_states = outputs
3581

3582
            if self.speculative_config and self.speculative_config.use_eagle():
3583
                assert isinstance(self.drafter, EagleProposer)
3584
3585
3586
3587
                use_cudagraphs = (
                    cudagraph_runtime_mode == CUDAGraphMode.PIECEWISE
                    and not self.speculative_config.enforce_eager
                )
3588
                self.drafter.dummy_run(num_tokens, use_cudagraphs=use_cudagraphs)
3589

3590
3591
3592
3593
3594
3595
3596
3597
3598
3599
        # 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)

3600
        logit_indices = np.cumsum(num_scheduled_tokens) - 1
3601
3602
3603
3604
        logit_indices_device = torch.from_numpy(logit_indices).to(
            self.device, non_blocking=True
        )
        return hidden_states, hidden_states[logit_indices_device]
3605
3606
3607
3608
3609
3610

    @torch.inference_mode()
    def _dummy_sampler_run(
        self,
        hidden_states: torch.Tensor,
    ) -> torch.Tensor:
3611
3612
3613
3614
        # 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)
3615

3616
        logits = self.model.compute_logits(hidden_states)
3617
3618
        num_reqs = logits.size(0)

3619
        dummy_tensors = lambda v: torch.full((num_reqs,), v, device=self.device)
3620
3621
3622
3623
3624
3625
3626
3627
3628
3629
3630
3631
3632
3633
3634

        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)],
3635
            spec_token_ids=[[] for _ in range(num_reqs)],
3636
3637
            allowed_token_ids_mask=None,
            bad_words_token_ids={},
3638
            logitsprocs=LogitsProcessors(),
3639
        )
3640
        try:
3641
3642
3643
            sampler_output = self.sampler(
                logits=logits, sampling_metadata=dummy_metadata
            )
3644
        except RuntimeError as e:
3645
            if "out of memory" in str(e):
3646
3647
3648
3649
                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 "
3650
3651
                    "initializing the engine."
                ) from e
3652
3653
            else:
                raise e
3654
        if self.speculative_config:
3655
3656
            draft_token_ids = [[0] for _ in range(num_reqs)]
            dummy_spec_decode_metadata = SpecDecodeMetadata.make_dummy(
3657
3658
                draft_token_ids, self.device
            )
3659
3660
3661
3662
3663
3664

            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
3665
3666
3667
3668
3669
            logits = torch.randn(
                num_tokens + num_reqs,
                logits.shape[-1],
                device=self.device,
                dtype=logits.dtype,
3670
            )
3671
3672
3673
            self.rejection_sampler(
                dummy_spec_decode_metadata,
                draft_probs,
3674
                logits,
3675
3676
                dummy_metadata,
            )
3677
        return sampler_output
3678

3679
    def _dummy_pooler_run_task(
3680
3681
        self,
        hidden_states: torch.Tensor,
3682
3683
        task: PoolingTask,
    ) -> PoolerOutput:
3684
3685
3686
3687
3688
3689
3690
3691
3692
3693
3694
        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

3695
        dummy_prompt_lens = torch.tensor(
3696
3697
            num_scheduled_tokens_list,
            device="cpu",
3698
        )
3699
3700
3701
        dummy_token_ids = torch.zeros(
            (num_reqs, req_num_tokens), dtype=torch.int32, device=self.device
        )
3702

3703
        model = cast(VllmModelForPooling, self.get_model())
3704
        dummy_pooling_params = PoolingParams(task=task)
3705
        dummy_pooling_params.verify(task=task, model_config=self.model_config)
3706
        to_update = model.pooler.get_pooling_updates(task)
3707
3708
        to_update.apply(dummy_pooling_params)

3709
        dummy_metadata = PoolingMetadata(
3710
3711
3712
3713
            prompt_lens=dummy_prompt_lens,
            prompt_token_ids=dummy_token_ids,
            pooling_params=[dummy_pooling_params] * num_reqs,
        )
3714

3715
3716
3717
        dummy_metadata.build_pooling_cursor(
            num_scheduled_tokens_list, device=hidden_states.device
        )
3718

3719
        try:
3720
3721
3722
            return model.pooler(
                hidden_states=hidden_states, pooling_metadata=dummy_metadata
            )
3723
        except RuntimeError as e:
3724
            if "out of memory" in str(e):
3725
                raise RuntimeError(
3726
3727
3728
                    "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 "
3729
3730
                    "initializing the engine."
                ) from e
3731
3732
            else:
                raise e
3733
3734
3735
3736
3737
3738
3739

    @torch.inference_mode()
    def _dummy_pooler_run(
        self,
        hidden_states: torch.Tensor,
    ) -> PoolerOutput:
        # Find the task that has the largest output for subsequent steps
3740
3741
3742
3743
3744
3745
3746
3747
3748
3749
3750
3751
3752
3753
3754
3755
3756
3757
3758
3759
        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."
                )

3760
        output_size = dict[PoolingTask, float]()
3761
        for task in supported_pooling_tasks:
3762
3763
            # Run a full batch with each task to ensure none of them OOMs
            output = self._dummy_pooler_run_task(hidden_states, task)
3764
            output_size[task] = sum(o.nbytes for o in output)
3765
3766
3767
3768
            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)
3769

3770
    def profile_run(self) -> None:
3771
        # Profile with multimodal encoder & encoder cache.
3772
        if self.supports_mm_inputs:
3773
            if self.model_config.multimodal_config.skip_mm_profiling:
3774
                logger.info(
3775
                    "Skipping memory profiling for multimodal encoder and "
3776
3777
                    "encoder cache."
                )
3778
3779
3780
3781
3782
3783
3784
3785
            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.
3786
                    dummy_modality = mm_budget.get_modality_with_max_tokens()
3787
3788
3789
                    max_mm_items_per_batch = mm_budget.max_items_per_batch_by_modality[
                        dummy_modality
                    ]
3790
3791
3792
3793
3794
3795
3796
3797
3798

                    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,
                    )
3799

3800
3801
3802
3803
3804
                    # Create dummy batch of multimodal inputs.
                    batched_dummy_mm_inputs = self._get_mm_dummy_batch(
                        dummy_modality,
                        max_mm_items_per_batch,
                    )
3805

3806
                    # Run multimodal encoder.
3807
3808
3809
                    dummy_encoder_outputs = self.model.get_multimodal_embeddings(
                        **batched_dummy_mm_inputs
                    )
3810

3811
3812
3813
3814
                    sanity_check_mm_encoder_outputs(
                        dummy_encoder_outputs,
                        expected_num_items=max_mm_items_per_batch,
                    )
3815

3816
3817
3818
3819
3820
3821
3822
3823
3824
3825
                    # 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(
3826
3827
                                (encoder_budget, encoder_output_shape[-1])
                            )
3828
3829
3830
3831
3832
3833
                            num_tokens = output.shape[0]
                            expanded[:num_tokens].copy_(output)
                            expanded_outputs.append(expanded)

                        dummy_encoder_outputs = expanded_outputs

3834
                    # Cache the dummy encoder outputs.
3835
                    self.encoder_cache["tmp"] = dict(enumerate(dummy_encoder_outputs))
3836

3837
        # Add `is_profile` here to pre-allocate communication buffers
3838
3839
3840
        hidden_states, last_hidden_states = self._dummy_run(
            self.max_num_tokens, is_profile=True
        )
3841
        if get_pp_group().is_last_rank:
3842
3843
3844
3845
            if self.is_pooling_model:
                output = self._dummy_pooler_run(hidden_states)
            else:
                output = self._dummy_sampler_run(last_hidden_states)
3846
        else:
3847
            output = None
3848
        self._sync_device()
3849
        del hidden_states, output
3850
        self.encoder_cache.clear()
3851
        gc.collect()
3852

3853
    def capture_model(self) -> int:
3854
        if self.compilation_config.cudagraph_mode == CUDAGraphMode.NONE:
3855
            logger.warning(
3856
                "Skipping CUDA graph capture. To turn on CUDA graph capture, "
3857
3858
                "ensure `cudagraph_mode` was not manually set to `NONE`"
            )
3859
            return 0
3860

3861
3862
        compilation_counter.num_gpu_runner_capture_triggers += 1

3863
3864
        start_time = time.perf_counter()

3865
3866
3867
3868
3869
3870
3871
3872
3873
3874
3875
3876
3877
3878
        @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()
3879
                    gc.collect()
3880

3881
3882
3883
        # 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.
3884
        set_cudagraph_capturing_enabled(True)
3885
        with freeze_gc(), graph_capture(device=self.device):
3886
            start_free_gpu_memory = torch.cuda.mem_get_info()[0]
3887
            cudagraph_mode = self.compilation_config.cudagraph_mode
3888
            assert cudagraph_mode is not None
3889
3890
3891
3892
3893
3894
3895
3896
3897

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

3898
3899
            if cudagraph_mode.mixed_mode() != CUDAGraphMode.NONE:
                cudagraph_runtime_mode = cudagraph_mode.mixed_mode()
3900
                # make sure we capture the largest batch size first
3901
3902
3903
                compilation_cases = list(
                    product(reversed(self.cudagraph_batch_sizes), lora_cases)
                )
3904
3905
3906
                self._capture_cudagraphs(
                    compilation_cases,
                    cudagraph_runtime_mode=cudagraph_runtime_mode,
3907
3908
                    uniform_decode=False,
                )
3909

3910
3911
            # Capture full cudagraph for uniform decode batches if we
            # don't already have full mixed prefill-decode cudagraphs.
3912
3913
3914
3915
3916
3917
3918
            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
                )
3919
                decode_cudagraph_batch_sizes = [
3920
3921
                    x
                    for x in self.cudagraph_batch_sizes
3922
                    if max_num_tokens >= x >= self.uniform_decode_query_len
3923
                ]
3924
3925
3926
                compilation_cases_decode = list(
                    product(reversed(decode_cudagraph_batch_sizes), lora_cases)
                )
3927
3928
3929
                self._capture_cudagraphs(
                    compilation_cases=compilation_cases_decode,
                    cudagraph_runtime_mode=CUDAGraphMode.FULL,
3930
3931
                    uniform_decode=True,
                )
3932

3933
3934
3935
            torch.cuda.synchronize()
            end_free_gpu_memory = torch.cuda.mem_get_info()[0]

3936
3937
3938
        # 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
3939
        # we may do lazy capturing in future that still allows capturing
3940
3941
        # after here.
        set_cudagraph_capturing_enabled(False)
3942
3943
3944
3945
3946

        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.
3947
        logger.info_once(
3948
3949
3950
            "Graph capturing finished in %.0f secs, took %.2f GiB",
            elapsed_time,
            cuda_graph_size / (1 << 30),
3951
            scope="local",
3952
        )
3953
        return cuda_graph_size
3954

3955
3956
    def _capture_cudagraphs(
        self,
3957
        compilation_cases: list[tuple[int, bool]],
3958
3959
3960
3961
3962
3963
3964
        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}"
3965
3966
3967
3968
3969
3970
3971
3972

        # 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",
3973
3974
3975
                    cudagraph_runtime_mode.name,
                ),
            )
3976

3977
        # We skip EPLB here since we don't want to record dummy metrics
3978
        for num_tokens, activate_lora in compilation_cases:
3979
            # We currently only capture ubatched graphs when its a FULL
3980
3981
3982
            # 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
3983
3984
3985
3986
            allow_microbatching = (
                self.parallel_config.enable_dbo
                and cudagraph_runtime_mode == CUDAGraphMode.FULL
                and uniform_decode
3987
3988
3989
3990
3991
                and check_ubatch_thresholds(
                    config=self.vllm_config.parallel_config,
                    num_tokens=num_tokens,
                    uniform_decode=uniform_decode,
                )
3992
            )
3993

3994
3995
3996
3997
3998
3999
            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.
4000
4001
4002
4003
4004
4005
4006
4007
4008
                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,
4009
                    activate_lora=activate_lora,
4010
4011
4012
4013
4014
4015
4016
4017
                )
            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,
4018
                activate_lora=activate_lora,
4019
            )
4020
        self.maybe_remove_all_loras(self.lora_config)
4021

4022
4023
4024
4025
    def initialize_attn_backend(self, kv_cache_config: KVCacheConfig) -> None:
        """
        Initialize the attention backends and attention metadata builders.
        """
4026
        assert len(self.attn_groups) == 0, "Attention backends are already initialized"
4027

4028
4029
4030
4031
4032
4033
        class AttentionGroupKey(NamedTuple):
            attn_backend: type[AttentionBackend]
            kv_cache_spec: KVCacheSpec

        def get_attn_backends_for_group(
            kv_cache_group_spec: KVCacheGroupSpec,
4034
        ) -> tuple[dict[AttentionGroupKey, list[str]], set[type[AttentionBackend]]]:
4035
            layers = get_layers_from_vllm_config(
4036
4037
                self.vllm_config, AttentionLayerBase, kv_cache_group_spec.layer_names
            )
4038
4039
            attn_backends = {}
            attn_backend_layers = defaultdict(list)
4040
            # Dedupe based on full class name; this is a bit safer than
4041
4042
4043
4044
            # 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.
4045
            for layer_name in kv_cache_group_spec.layer_names:
4046
                attn_backend = layers[layer_name].get_attn_backend()
4047
4048
4049
4050
4051
4052
4053

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

4054
4055
4056
                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):
4057
                    layer_kv_cache_spec = layer_kv_cache_spec.kv_cache_specs[layer_name]
4058
                key = (full_cls_name, layer_kv_cache_spec)
4059
4060
4061
                attn_backends[key] = AttentionGroupKey(
                    attn_backend, layer_kv_cache_spec
                )
4062
                attn_backend_layers[key].append(layer_name)
4063
4064
4065
4066
            return (
                {attn_backends[k]: v for k, v in attn_backend_layers.items()},
                set(group_key.attn_backend for group_key in attn_backends.values()),
            )
4067
4068

        def create_attn_groups(
4069
            attn_backends_map: dict[AttentionGroupKey, list[str]],
4070
            kv_cache_group_id: int,
4071
4072
        ) -> list[AttentionGroup]:
            attn_groups: list[AttentionGroup] = []
4073
            for (attn_backend, kv_cache_spec), layer_names in attn_backends_map.items():
4074
                attn_group = AttentionGroup(
4075
                    attn_backend,
4076
                    layer_names,
4077
                    kv_cache_spec,
4078
                    kv_cache_group_id,
4079
4080
                )

4081
4082
4083
                attn_groups.append(attn_group)
            return attn_groups

4084
4085
        attention_backend_maps = []
        attention_backend_set: set[type[AttentionBackend]] = set()
4086
        for kv_cache_group_spec in kv_cache_config.kv_cache_groups:
4087
            attn_backends = get_attn_backends_for_group(kv_cache_group_spec)
4088
4089
4090
4091
4092
4093
            attention_backend_maps.append(attn_backends[0])
            attention_backend_set.update(attn_backends[1])

        # Resolve cudagraph_mode before actually initialize metadata_builders
        self._check_and_update_cudagraph_mode(attention_backend_set)

4094
4095
        for i, attn_backend_map in enumerate(attention_backend_maps):
            self.attn_groups.append(create_attn_groups(attn_backend_map, i))
4096

4097
4098
4099
4100
4101
4102
4103
4104
4105
4106
4107
4108
4109
4110
4111
4112
4113
4114
    def initialize_metadata_builders(
        self, kv_cache_config: KVCacheConfig, kernel_block_sizes: list[int]
    ) -> None:
        """
        Create the metadata builders for all KV cache groups and attn groups.
        """
        for kv_cache_group_id in range(len(kv_cache_config.kv_cache_groups)):
            for attn_group in self.attn_groups[kv_cache_group_id]:
                attn_group.create_metadata_builders(
                    self.vllm_config,
                    self.device,
                    kernel_block_sizes[kv_cache_group_id]
                    if kv_cache_group_id < len(kernel_block_sizes)
                    else None,
                    num_metadata_builders=1
                    if not self.parallel_config.enable_dbo
                    else 2,
                )
co63oc's avatar
co63oc committed
4115
        # Calculate reorder batch threshold (if needed)
4116
4117
        # Note (tdoublep): do this *after* constructing builders,
        # because some of them change the threshold at init time.
4118
4119
        self.calculate_reorder_batch_threshold()

4120
4121
4122
    def _check_and_update_cudagraph_mode(
        self, attention_backends: set[type[AttentionBackend]]
    ) -> None:
4123
        """
4124
        Resolve the cudagraph_mode when there are multiple attention
4125
4126
4127
4128
        backends with potential conflicting CUDA graph support.
        Then initialize the cudagraph_dispatcher based on the resolved
        cudagraph_mode.
        """
4129
        min_cg_support = AttentionCGSupport.ALWAYS
4130
        min_cg_backend_name = None
4131

4132
4133
4134
4135
4136
        for attn_backend in attention_backends:
            builder_cls = attn_backend.get_builder_cls()
            if builder_cls.cudagraph_support.value < min_cg_support.value:
                min_cg_support = builder_cls.cudagraph_support
                min_cg_backend_name = attn_backend.__name__
4137
4138
4139
        # Flexible resolve the cudagraph mode
        cudagraph_mode = self.compilation_config.cudagraph_mode
        # check cudagraph for mixed batch is supported
4140
4141
4142
4143
4144
4145
        if (
            cudagraph_mode.mixed_mode() == CUDAGraphMode.FULL
            and min_cg_support != AttentionCGSupport.ALWAYS
        ):
            msg = (
                f"CUDAGraphMode.{cudagraph_mode.name} is not supported "
4146
                f"with {min_cg_backend_name} backend (support: "
4147
4148
                f"{min_cg_support})"
            )
4149
4150
            if min_cg_support == AttentionCGSupport.NEVER:
                # if not supported any full cudagraphs, just raise it.
4151
4152
                msg += (
                    "; please try cudagraph_mode=PIECEWISE, and "
4153
                    "make sure compilation mode is VLLM_COMPILE"
4154
                )
4155
4156
4157
4158
4159
                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"
4160
                cudagraph_mode = self.compilation_config.cudagraph_mode = (
4161
                    CUDAGraphMode.FULL_AND_PIECEWISE
4162
                )
4163
4164
            else:
                msg += "; setting cudagraph_mode=FULL_DECODE_ONLY"
4165
                cudagraph_mode = self.compilation_config.cudagraph_mode = (
4166
                    CUDAGraphMode.FULL_DECODE_ONLY
4167
                )
4168
4169
            logger.warning(msg)

4170
        # check that if we are doing decode full-cudagraphs it is supported
4171
4172
4173
4174
4175
4176
        if (
            cudagraph_mode.decode_mode() == CUDAGraphMode.FULL
            and min_cg_support == AttentionCGSupport.NEVER
        ):
            msg = (
                f"CUDAGraphMode.{cudagraph_mode.name} is not supported "
4177
                f"with {min_cg_backend_name} backend (support: "
4178
4179
                f"{min_cg_support})"
            )
4180
            if self.compilation_config.mode == CompilationMode.VLLM_COMPILE and (
4181
4182
4183
4184
4185
                self.compilation_config.splitting_ops_contain_attention()
                or self.compilation_config.use_inductor_graph_partition
            ):
                msg += (
                    "; setting cudagraph_mode=PIECEWISE because "
4186
                    "attention is compiled piecewise"
4187
4188
                )
                cudagraph_mode = self.compilation_config.cudagraph_mode = (
4189
                    CUDAGraphMode.PIECEWISE
4190
                )
4191
            else:
4192
4193
                msg += (
                    "; setting cudagraph_mode=NONE because "
4194
                    "attention is not compiled piecewise"
4195
4196
                )
                cudagraph_mode = self.compilation_config.cudagraph_mode = (
4197
                    CUDAGraphMode.NONE
4198
                )
4199
4200
            logger.warning(msg)

4201
4202
        # check that if we are doing spec-decode + decode full-cudagraphs it is
        # supported
4203
4204
4205
4206
4207
4208
4209
4210
        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 "
4211
                f"{min_cg_backend_name} (support: {min_cg_support})"
4212
            )
4213
4214
            if self.compilation_config.splitting_ops_contain_attention():
                msg += "; setting cudagraph_mode=PIECEWISE"
4215
                cudagraph_mode = self.compilation_config.cudagraph_mode = (
4216
                    CUDAGraphMode.PIECEWISE
4217
                )
4218
4219
            else:
                msg += "; setting cudagraph_mode=NONE"
4220
                cudagraph_mode = self.compilation_config.cudagraph_mode = (
4221
                    CUDAGraphMode.NONE
4222
                )
4223
4224
4225
4226
            logger.warning(msg)

        # double check that we can support full cudagraph if they are requested
        # even after automatic downgrades
4227
4228
4229
4230
4231
4232
        if (
            cudagraph_mode.has_full_cudagraphs()
            and min_cg_support == AttentionCGSupport.NEVER
        ):
            raise ValueError(
                f"CUDAGraphMode.{cudagraph_mode.name} is not "
4233
                f"supported with {min_cg_backend_name} backend ("
4234
4235
                f"support:{min_cg_support}) "
                "; please try cudagraph_mode=PIECEWISE, "
4236
                "and make sure compilation mode is VLLM_COMPILE"
4237
            )
4238

4239
4240
        # Trigger cudagraph dispatching keys initialization after
        # resolved cudagraph mode.
4241
        self.cudagraph_dispatcher.initialize_cudagraph_keys(
4242
4243
            self.compilation_config.cudagraph_mode, self.uniform_decode_query_len
        )
4244

4245
4246
    def calculate_reorder_batch_threshold(self) -> None:
        """
4247
4248
4249
4250
        Choose the minimum reorder batch threshold from all attention groups.
        Backends should be able to support lower threshold then what they request
        just may have a performance penalty due to that backend treating decodes
        as prefills.
4251
        """
4252
4253
4254
4255
4256
4257
        min_none_high = lambda a, b: a if b is None else b if a is None else min(a, b)

        reorder_batch_thresholds = [
            group.get_metadata_builder().reorder_batch_threshold
            for group in self._attn_group_iterator()
        ]
4258
4259
4260
4261
4262
        # If there are no attention groups (attention-free model) or no backend
        # reports a threshold, leave reordering disabled.
        if len(reorder_batch_thresholds) == 0:
            self.reorder_batch_threshold = None
            return
4263
        self.reorder_batch_threshold = reduce(min_none_high, reorder_batch_thresholds)
4264

4265
4266
4267
    @staticmethod
    def select_common_block_size(
        kv_manager_block_size: int, attn_groups: list[AttentionGroup]
4268
4269
    ) -> int:
        """
4270
4271
4272
4273
4274
        Select a block size that is supported by all backends and is a factor of
        kv_manager_block_size.

        If kv_manager_block_size is supported by all backends, return it directly.
        Otherwise, return the max supported size.
4275
4276
4277
4278
4279
4280

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

        Returns:
4281
            The selected block size
4282
4283

        Raises:
4284
            ValueError: If no valid block size found
4285
4286
        """

4287
4288
4289
4290
4291
4292
4293
4294
4295
4296
4297
4298
4299
4300
4301
4302
4303
4304
4305
4306
4307
4308
4309
4310
4311
4312
4313
4314
4315
4316
4317
4318
4319
4320
4321
4322
4323
4324
4325
4326
4327
4328
        def block_size_is_supported(
            backends: list[type[AttentionBackend]], block_size: int
        ) -> bool:
            """
            Check if the block size is supported by all backends.
            """
            for backend in backends:
                is_supported = False
                for supported_size in backend.get_supported_kernel_block_size():
                    if isinstance(supported_size, int):
                        if block_size == supported_size:
                            is_supported = True
                    elif isinstance(supported_size, MultipleOf):
                        if block_size % supported_size.base == 0:
                            is_supported = True
                    else:
                        raise ValueError(f"Unknown supported size: {supported_size}")
                if not is_supported:
                    return False
            return True

        backends = [group.backend for group in attn_groups]

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

        # Case 2: otherwise, the block_size must be an `int`-format supported size of
        # at least one backend. Iterate over all `int`-format supported sizes in
        # descending order and return the first one that is supported by all backends.
        # Simple proof:
        # If the supported size b is in MultipleOf(x_i) format for all attention
        # backends i, and b a factor of kv_manager_block_size, then
        # kv_manager_block_size also satisfies MultipleOf(x_i) for all i. We will
        # return kv_manager_block_size in case 1.
        all_int_supported_sizes = set(
            supported_size
            for backend in backends
            for supported_size in backend.get_supported_kernel_block_size()
            if isinstance(supported_size, int)
        )
4329

4330
4331
4332
4333
4334
4335
        for supported_size in sorted(all_int_supported_sizes, reverse=True):
            if kv_manager_block_size % supported_size != 0:
                continue
            if block_size_is_supported(backends, supported_size):
                return supported_size
        raise ValueError(f"No common block size for {kv_manager_block_size}. ")
4336

4337
4338
4339
    def may_reinitialize_input_batch(
        self, kv_cache_config: KVCacheConfig, kernel_block_sizes: list[int]
    ) -> None:
4340
4341
4342
4343
4344
4345
4346
        """
        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.
4347
            kernel_block_sizes: The kernel block sizes for each KV cache group.
4348
4349
4350
4351
        """
        block_sizes = [
            kv_cache_group.kv_cache_spec.block_size
            for kv_cache_group in kv_cache_config.kv_cache_groups
4352
            if not isinstance(kv_cache_group.kv_cache_spec, EncoderOnlyAttentionSpec)
4353
        ]
4354
4355
4356
4357

        if block_sizes != [self.cache_config.block_size] or kernel_block_sizes != [
            self.cache_config.block_size
        ]:
4358
4359
4360
            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
4361
4362
                "for more details."
            )
4363
4364
            self.input_batch = InputBatch(
                max_num_reqs=self.max_num_reqs,
4365
                max_model_len=max(self.max_model_len, self.max_encoder_len),
4366
4367
4368
4369
4370
                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,
4371
                kernel_block_sizes=kernel_block_sizes,
4372
                is_spec_decode=bool(self.vllm_config.speculative_config),
4373
                logitsprocs=self.input_batch.logitsprocs,
4374
                logitsprocs_need_output_token_ids=self.input_batch.logitsprocs_need_output_token_ids,
4375
                is_pooling_model=self.is_pooling_model,
4376
4377
                num_speculative_tokens=(
                    self.vllm_config.speculative_config.num_speculative_tokens
4378
4379
4380
                    if self.vllm_config.speculative_config
                    else 0
                ),
4381
4382
            )

4383
    def _allocate_kv_cache_tensors(
4384
4385
        self, kv_cache_config: KVCacheConfig
    ) -> dict[str, torch.Tensor]:
4386
        """
4387
4388
4389
        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.

4390
        Args:
4391
            kv_cache_config: The KV cache config
4392
        Returns:
4393
            dict[str, torch.Tensor]: A map between layer names to their
4394
            corresponding memory buffer for KV cache.
4395
        """
4396
4397
        kv_cache_raw_tensors: dict[str, torch.Tensor] = {}
        for kv_cache_tensor in kv_cache_config.kv_cache_tensors:
4398
4399
4400
            tensor = torch.zeros(
                kv_cache_tensor.size, dtype=torch.int8, device=self.device
            )
4401
4402
4403
4404
4405
            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:
4406
4407
4408
4409
            for layer_name in group.layer_names:
                if layer_name in self.runner_only_attn_layers:
                    continue
                layer_names.add(layer_name)
4410
4411
4412
        assert layer_names == set(kv_cache_raw_tensors.keys()), (
            "Some layers are not correctly initialized"
        )
4413
4414
        return kv_cache_raw_tensors

4415
4416
4417
    def _attn_group_iterator(self) -> Iterator[AttentionGroup]:
        return itertools.chain.from_iterable(self.attn_groups)

4418
    def _kv_cache_spec_attn_group_iterator(self) -> Iterator[AttentionGroup]:
4419
4420
        if not self.kv_cache_config.kv_cache_groups:
            return
4421
4422
        for attn_groups in self.attn_groups:
            yield from attn_groups
4423

4424
4425
4426
4427
4428
4429
4430
4431
4432
4433
4434
4435
4436
4437
4438
4439
4440
4441
    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
        ):
4442
4443
4444
4445
4446
4447
            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):
4448
                continue
4449
            elif isinstance(kv_cache_spec, AttentionSpec):
4450
4451
4452
4453
4454
                # 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
4455
                selected_kernel_size = self.select_common_block_size(
4456
4457
4458
                    kv_manager_block_size, attn_groups
                )
                kernel_block_sizes.append(selected_kernel_size)
4459
            elif isinstance(kv_cache_spec, MambaSpec):
4460
4461
                # This is likely Mamba or other non-attention cache,
                # no splitting.
4462
                kernel_block_sizes.append(kv_cache_spec.block_size)
4463
4464
4465
4466
4467
4468
            else:
                raise NotImplementedError(
                    f"unknown kv cache spec {kv_cache_group.kv_cache_spec}"
                )
        return kernel_block_sizes

4469
4470
4471
4472
    def _reshape_kv_cache_tensors(
        self,
        kv_cache_config: KVCacheConfig,
        kv_cache_raw_tensors: dict[str, torch.Tensor],
4473
        kernel_block_sizes: list[int],
4474
    ) -> dict[str, torch.Tensor]:
4475
        """
4476
        Reshape the KV cache tensors to the desired shape and dtype.
4477

4478
        Args:
4479
4480
            kv_cache_config: The KV cache config
            kv_cache_raw_tensors: The KV cache buffer of each layer, with
4481
                correct size but uninitialized shape.
4482
            kernel_block_sizes: The kernel block sizes for each KV cache group.
4483
        Returns:
4484
            Dict[str, torch.Tensor]: A map between layer names to their
4485
4486
            corresponding memory buffer for KV cache.
        """
4487
        kv_caches: dict[str, torch.Tensor] = {}
4488
        has_attn, has_mamba = False, False
4489
4490
        for group in self._kv_cache_spec_attn_group_iterator():
            kv_cache_spec = group.kv_cache_spec
4491
            attn_backend = group.backend
4492
4493
4494
4495
            if group.kv_cache_group_id == len(kernel_block_sizes):
                # There may be a last group for layers without kv cache.
                continue
            kernel_block_size = kernel_block_sizes[group.kv_cache_group_id]
4496
            for layer_name in group.layer_names:
4497
4498
                if layer_name in self.runner_only_attn_layers:
                    continue
4499
4500
                raw_tensor = kv_cache_raw_tensors[layer_name]
                assert raw_tensor.numel() % kv_cache_spec.page_size_bytes == 0
4501
                num_blocks = raw_tensor.numel() // kv_cache_spec.page_size_bytes
4502
                if isinstance(kv_cache_spec, AttentionSpec):
4503
                    has_attn = True
4504
4505
                    num_blocks_per_kv_block = (
                        kv_cache_spec.block_size // kernel_block_size
4506
4507
4508
                    )
                    kernel_num_blocks = num_blocks * num_blocks_per_kv_block

4509
                    kv_cache_shape = attn_backend.get_kv_cache_shape(
4510
                        kernel_num_blocks,
4511
                        kernel_block_size,
4512
4513
                        kv_cache_spec.num_kv_heads,
                        kv_cache_spec.head_size,
4514
4515
                        cache_dtype_str=self.cache_config.cache_dtype,
                    )
4516
                    dtype = kv_cache_spec.dtype
4517
                    try:
4518
                        kv_cache_stride_order = attn_backend.get_kv_cache_stride_order()
4519
                        assert len(kv_cache_stride_order) == len(kv_cache_shape)
4520
                    except (AttributeError, NotImplementedError):
4521
                        kv_cache_stride_order = tuple(range(len(kv_cache_shape)))
4522
4523
4524
4525
4526
                    # 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.
4527
4528
4529
                    kv_cache_shape = tuple(
                        kv_cache_shape[i] for i in kv_cache_stride_order
                    )
4530
4531
4532
4533
4534
                    # Maintain original KV shape view.
                    inv_order = [
                        kv_cache_stride_order.index(i)
                        for i in range(len(kv_cache_stride_order))
                    ]
4535
4536
4537
4538
4539
4540
                    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
4541
                elif isinstance(kv_cache_spec, MambaSpec):
4542
                    has_mamba = True
Chen Zhang's avatar
Chen Zhang committed
4543
4544
                    raw_tensor = kv_cache_raw_tensors[layer_name]
                    state_tensors = []
4545
                    storage_offset_bytes = 0
4546
                    for shape, dtype in zip(kv_cache_spec.shapes, kv_cache_spec.dtypes):
4547
4548
                        dtype_size = get_dtype_size(dtype)
                        num_element_per_page = (
4549
4550
                            kv_cache_spec.page_size_bytes // dtype_size
                        )
Chen Zhang's avatar
Chen Zhang committed
4551
                        target_shape = (num_blocks, *shape)
4552
4553
                        stride = torch.empty(target_shape).stride()
                        target_stride = (num_element_per_page, *stride[1:])
4554
                        assert storage_offset_bytes % dtype_size == 0
4555
4556
4557
4558
                        tensor = torch.as_strided(
                            raw_tensor.view(dtype),
                            size=target_shape,
                            stride=target_stride,
4559
                            storage_offset=storage_offset_bytes // dtype_size,
4560
                        )
Chen Zhang's avatar
Chen Zhang committed
4561
                        state_tensors.append(tensor)
4562
                        storage_offset_bytes += stride[0] * dtype_size
4563
4564

                    kv_caches[layer_name] = state_tensors
4565
                else:
4566
                    raise NotImplementedError
4567
4568

        if has_attn and has_mamba:
4569
            self._update_hybrid_attention_mamba_layout(kv_caches)
4570

4571
4572
        return kv_caches

4573
    def _update_hybrid_attention_mamba_layout(
4574
4575
        self, kv_caches: dict[str, torch.Tensor]
    ) -> None:
4576
        """
4577
4578
        Update the layout of attention layers from (2, num_blocks, ...) to
        (num_blocks, 2, ...).
4579
4580

        Args:
4581
            kv_caches: The KV cache buffer of each layer.
4582
4583
        """

4584
4585
        for group in self._kv_cache_spec_attn_group_iterator():
            kv_cache_spec = group.kv_cache_spec
4586
            for layer_name in group.layer_names:
4587
                kv_cache = kv_caches[layer_name]
4588
4589
4590
4591
                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 "
4592
                        f"a tensor of shape {kv_cache.shape}"
4593
                    )
4594
                    hidden_size = kv_cache.shape[2:].numel()
4595
4596
4597
4598
                    kv_cache.as_strided_(
                        size=kv_cache.shape,
                        stride=(hidden_size, 2 * hidden_size, *kv_cache.stride()[2:]),
                    )
4599

4600
    def initialize_kv_cache_tensors(
4601
        self, kv_cache_config: KVCacheConfig, kernel_block_sizes: list[int]
4602
    ) -> dict[str, torch.Tensor]:
4603
4604
4605
4606
4607
        """
        Initialize the memory buffer for KV cache.

        Args:
            kv_cache_config: The KV cache config
4608
4609
            kernel_block_sizes: The kernel block sizes for each KV cache group.

4610
        Returns:
4611
            Dict[str, torch.Tensor]: A map between layer names to their
4612
4613
4614
4615
4616
            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
4617
        kv_caches = self._reshape_kv_cache_tensors(
4618
            kv_cache_config, kv_cache_raw_tensors, kernel_block_sizes
4619
        )
4620

4621
        # Set up cross-layer KV cache sharing
4622
4623
        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)
4624
4625
            kv_caches[layer_name] = kv_caches[target_layer_name]

4626
4627
4628
4629
4630
4631
4632
4633
4634
        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,
        )
4635
4636
4637
        return kv_caches

    def maybe_add_kv_sharing_layers_to_kv_cache_groups(
4638
4639
        self, kv_cache_config: KVCacheConfig
    ) -> None:
4640
4641
4642
4643
4644
4645
4646
4647
4648
4649
4650
4651
4652
4653
4654
4655
4656
4657
        """
        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.
4658
            attn_layers = get_layers_from_vllm_config(self.vllm_config, Attention)
4659
4660
            for layer_name in reversed(attn_layers):
                if layer_name in self.shared_kv_cache_layers:
4661
                    self.kv_sharing_fast_prefill_eligible_layers.add(layer_name)
4662
4663
                else:
                    break
4664

4665
4666
4667
4668
4669
4670
4671
    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
        """
4672
        kv_cache_config = deepcopy(kv_cache_config)
4673
        self.kv_cache_config = kv_cache_config
4674
        self.may_add_encoder_only_layers_to_kv_cache_config()
4675
        self.maybe_add_kv_sharing_layers_to_kv_cache_groups(kv_cache_config)
4676
        self.initialize_attn_backend(kv_cache_config)
4677
4678
4679
4680
4681
4682
        # The kernel block size for all KV cache groups. For example, if
        # kv_cache_manager uses block_size 256 for a given group, but the attention
        # backends for that group only supports block_size 64, we will return
        # kernel_block_size 64 and split the 256-token-block to 4 blocks with 64
        # tokens each.
        kernel_block_sizes = self._prepare_kernel_block_sizes(kv_cache_config)
4683
4684
4685
4686

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

4687
        # Reinitialize need to after initialize_attn_backend
4688
4689
4690
4691
        self.may_reinitialize_input_batch(kv_cache_config, kernel_block_sizes)
        kv_caches = self.initialize_kv_cache_tensors(
            kv_cache_config, kernel_block_sizes
        )
4692

4693
4694
4695
4696
4697
4698
        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
4699
        if has_kv_transfer_group():
4700
4701
4702
            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
4703

4704
        if self.dcp_world_size > 1:
4705
            layers = get_layers_from_vllm_config(self.vllm_config, AttentionLayerBase)
4706
4707
4708
4709
4710
            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__} "
4711
4712
                    "does not return the softmax lse for decode."
                )
4713

4714
4715
4716
4717
4718
    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
4719
        encoder_only_attn_specs: dict[AttentionSpec, list[str]] = defaultdict(list)
4720
4721
4722
        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:
4723
                attn_spec: AttentionSpec = EncoderOnlyAttentionSpec(
4724
4725
4726
                    block_size=block_size,
                    num_kv_heads=attn_module.num_kv_heads,
                    head_size=attn_module.head_size,
4727
4728
                    dtype=self.kv_cache_dtype,
                )
4729
4730
4731
                encoder_only_attn_specs[attn_spec].append(layer_name)
                self.runner_only_attn_layers.add(layer_name)
        if len(encoder_only_attn_specs) > 0:
4732
4733
4734
            assert len(encoder_only_attn_specs) == 1, (
                "Only support one encoder-only attention spec now"
            )
4735
4736
            spec, layer_names = encoder_only_attn_specs.popitem()
            self.kv_cache_config.kv_cache_groups.append(
4737
4738
                KVCacheGroupSpec(layer_names=layer_names, kv_cache_spec=spec)
            )
4739

4740
    def get_kv_cache_spec(self) -> dict[str, KVCacheSpec]:
4741
        """
4742
        Generates the KVCacheSpec by parsing the kv cache format from each
4743
4744
        Attention module in the static forward context.
        Returns:
4745
            KVCacheSpec: A dictionary mapping layer names to their KV cache
4746
4747
4748
            format. Layers that do not need KV cache are not included.
        """

4749
        kv_cache_spec: dict[str, KVCacheSpec] = {}
4750
        attn_layers = get_layers_from_vllm_config(self.vllm_config, AttentionLayerBase)
Chen Zhang's avatar
Chen Zhang committed
4751
        for layer_name, attn_module in attn_layers.items():
4752
4753
4754
4755
4756
4757
4758
4759
4760
4761
4762
4763
4764
4765
4766
            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
4767

4768
        return kv_cache_spec
4769
4770
4771
4772
4773
4774
4775
4776
4777
4778

    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.
4779
        pinned = self.sampled_token_ids_pinned_cpu[: sampled_token_ids.shape[0]]
4780
4781
4782
4783
        pinned.copy_(sampled_token_ids, non_blocking=True)
        self.transfer_event.record()
        self.transfer_event.synchronize()
        return pinned.tolist()