gpu_model_runner.py 206 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
    get_dcp_group,
39
40
41
42
43
44
    get_pp_group,
    get_tp_group,
    graph_capture,
    is_global_first_rank,
    prepare_communication_buffer_for_model,
)
45
from vllm.forward_context import BatchDescriptor, set_forward_context
46
from vllm.logger import init_logger
47
from vllm.model_executor.layers.attention_layer_base import AttentionLayerBase
48
from vllm.model_executor.layers.rotary_embedding import MRotaryEmbedding
49
from vllm.model_executor.model_loader import TensorizerLoader, get_model_loader
50
51
52
53
54
55
56
57
from vllm.model_executor.models.interfaces import (
    SupportsMultiModal,
    is_mixture_of_experts,
    supports_eagle3,
    supports_mrope,
    supports_multimodal_pruning,
    supports_transcription,
)
58
from vllm.model_executor.models.interfaces_base import (
59
60
61
62
    VllmModelForPooling,
    is_pooling_model,
    is_text_generation_model,
)
63
from vllm.multimodal import MULTIMODAL_REGISTRY
64
65
66
67
68
from vllm.multimodal.inputs import (
    BatchedTensorInputs,
    MultiModalKwargsItem,
    PlaceholderRange,
)
69
from vllm.multimodal.utils import group_mm_kwargs_by_modality
70
from vllm.pooling_params import PoolingParams
71
from vllm.sampling_params import SamplingType
72
from vllm.sequence import IntermediateTensors
73
from vllm.tasks import GenerationTask, PoolingTask, SupportedTask
74
from vllm.utils import length_from_prompt_token_ids_or_embeds
75
from vllm.utils.jsontree import json_map_leaves
76
from vllm.utils.math_utils import cdiv, round_up
77
78
from vllm.utils.mem_constants import GiB_bytes
from vllm.utils.mem_utils import DeviceMemoryProfiler
79
from vllm.utils.platform_utils import is_pin_memory_available
80
81
82
83
84
from vllm.utils.torch_utils import (
    get_dtype_size,
    kv_cache_dtype_str_to_dtype,
    supports_dynamo,
)
85
from vllm.v1.attention.backends.flash_attn import AttentionMetadata
86
from vllm.v1.attention.backends.gdn_attn import GDNAttentionMetadataBuilder
87
from vllm.v1.attention.backends.utils import (
88
89
90
    AttentionCGSupport,
    AttentionMetadataBuilder,
    CommonAttentionMetadata,
91
    create_fast_prefill_custom_backend,
92
    get_dcp_local_seq_lens,
93
94
95
    reorder_batch_to_split_decodes_and_prefills,
    split_attn_metadata,
)
96
from vllm.v1.cudagraph_dispatcher import CudagraphDispatcher
97
98
99
100
101
102
103
104
105
106
107
108
109
110
111
112
113
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,
114
    KVConnectorOutput,
115
116
117
118
119
120
    LogprobsLists,
    LogprobsTensors,
    ModelRunnerOutput,
    PoolerOutput,
    SamplerOutput,
)
121
from vllm.v1.pool.metadata import PoolingMetadata
122
from vllm.v1.sample.logits_processor import LogitsProcessors, build_logitsprocs
123
from vllm.v1.sample.metadata import SamplingMetadata
124
from vllm.v1.sample.rejection_sampler import RejectionSampler
125
from vllm.v1.sample.sampler import Sampler
126
from vllm.v1.spec_decode.eagle import EagleProposer
127
from vllm.v1.spec_decode.medusa import MedusaProposer
128
from vllm.v1.spec_decode.metadata import SpecDecodeMetadata
129
from vllm.v1.spec_decode.ngram_proposer import NgramProposer
130
from vllm.v1.spec_decode.suffix_decoding import SuffixDecodingProposer
131
from vllm.v1.structured_output.utils import apply_grammar_bitmask
132
from vllm.v1.utils import CpuGpuBuffer, record_function_or_nullcontext
133
from vllm.v1.worker.dp_utils import coordinate_batch_across_dp
134
from vllm.v1.worker.gpu_input_batch import CachedRequestState, InputBatch
135
from vllm.v1.worker.gpu_ubatch_wrapper import UBatchWrapper
136
from vllm.v1.worker.kv_connector_model_runner_mixin import KVConnectorModelRunnerMixin
137
from vllm.v1.worker.lora_model_runner_mixin import LoRAModelRunnerMixin
138
139
140
141
142
from vllm.v1.worker.ubatch_utils import (
    UBatchSlice,
    UBatchSlices,
    check_ubatch_thresholds,
)
143
from vllm.v1.worker.utils import is_residual_scattered_for_sp
144

145
146
147
148
149
150
151
152
153
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,
)
154

155
if TYPE_CHECKING:
156
    from vllm.model_executor.model_loader.tensorizer import TensorizerConfig
157
    from vllm.v1.core.sched.output import GrammarOutput, SchedulerOutput
158
159
160

logger = init_logger(__name__)

161
162
AttnMetadataDict: TypeAlias = dict[str, AttentionMetadata]
# list when ubatching is enabled
163
PerLayerAttnMetadata: TypeAlias = list[AttnMetadataDict] | AttnMetadataDict
164

165

166
167
168
169
170
171
# Wrapper for ModelRunnerOutput to support overlapped execution.
class AsyncGPUModelRunnerOutput(AsyncModelRunnerOutput):
    def __init__(
        self,
        model_runner_output: ModelRunnerOutput,
        sampled_token_ids: torch.Tensor,
172
        logprobs_tensors: torch.Tensor | None,
173
174
175
176
177
178
179
        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.
180
        self.async_copy_ready_event = torch.cuda.Event()
181
182
183
184

        # 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
185
        self._logprobs_tensors = logprobs_tensors
186
187
188
189
190

        # 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)
191
            self.sampled_token_ids_cpu = self._sampled_token_ids.to(
192
193
                "cpu", non_blocking=True
            )
194
195
196
197
198
            self._logprobs_tensors_cpu = (
                self._logprobs_tensors.to_cpu_nonblocking()
                if self._logprobs_tensors
                else None
            )
199
            self.async_copy_ready_event.record()
200
201
202

    def get_output(self) -> ModelRunnerOutput:
        """Copy the device tensors to the host and return a ModelRunnerOutput.
203

204
205
        This function blocks until the copy is finished.
        """
206
        self.async_copy_ready_event.synchronize()
207

208
209
        # Release the device tensors once the copy has completed.
        del self._logprobs_tensors
210
211
        del self._sampled_token_ids

212
        valid_sampled_token_ids = self.sampled_token_ids_cpu.tolist()
213
214
215
216
217
        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
218
219
220
221
        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()
222
223
224
        return output


225
226
227
228
229
230
231
232
233
234
235
236
237
238
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


239
class GPUModelRunner(LoRAModelRunnerMixin, KVConnectorModelRunnerMixin):
240
241
    def __init__(
        self,
242
        vllm_config: VllmConfig,
243
        device: torch.device,
244
    ):
245
246
247
        self.vllm_config = vllm_config
        self.model_config = vllm_config.model_config
        self.cache_config = vllm_config.cache_config
248
        self.compilation_config = vllm_config.compilation_config
249
250
251
252
253
254
        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
255

256
        from vllm.model_executor.models.utils import set_cpu_offload_max_bytes
257
258

        set_cpu_offload_max_bytes(int(self.cache_config.cpu_offload_gb * 1024**3))
259

260
261
262
263
        model_config = self.model_config
        cache_config = self.cache_config
        scheduler_config = self.scheduler_config
        parallel_config = self.parallel_config
264
        self.device = device
265
266
        self.pin_memory = is_pin_memory_available()
        self.dtype = self.model_config.dtype
267
268
269
        self.kv_cache_dtype = kv_cache_dtype_str_to_dtype(
            cache_config.cache_dtype, self.model_config
        )
270

271
        self.is_pooling_model = model_config.runner_type == "pooling"
272
        self.enable_prompt_embeds = model_config.enable_prompt_embeds
273
        self.is_multimodal_raw_input_only_model = (
274
275
            model_config.is_multimodal_raw_input_only_model
        )
276
277
        # This will be overridden in load_model()
        self.is_multimodal_pruning_enabled = False
278
        self.max_model_len = model_config.max_model_len
279
        self.dcp_world_size = self.parallel_config.decode_context_parallel_size
280
        self.dcp_rank = 0 if self.dcp_world_size <= 1 else get_dcp_group().rank_in_group
281
        self.max_num_tokens = scheduler_config.max_num_batched_tokens
282
        self.max_num_reqs = scheduler_config.max_num_seqs
283

284
285
286
287
288
        # 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 = (
289
290
291
            self.parallel_config.distributed_executor_backend == "external_launcher"
            and len(get_pp_group().ranks) > 0
        )
292

293
        # Model-related.
294
        self.num_query_heads = model_config.get_num_attention_heads(parallel_config)
295
        self.hidden_size = model_config.get_hidden_size()
296
        self.attention_chunk_size = model_config.attention_chunk_size
297
        # Only relevant for models using ALiBi (e.g, MPT)
298
        self.use_alibi = model_config.uses_alibi
299

300
        self.cascade_attn_enabled = not self.model_config.disable_cascade_attn
301

302
        # Multi-modal data support
303
        self.mm_registry = MULTIMODAL_REGISTRY
304
        self.uses_mrope = model_config.uses_mrope
305
        self.supports_mm_inputs = self.mm_registry.supports_multimodal_inputs(
306
307
            model_config
        )
308

309
310
311
        if self.model_config.is_encoder_decoder:
            # Maximum length of the encoder input, only for encoder-decoder
            # models.
312
            self.max_encoder_len = scheduler_config.max_num_encoder_input_tokens
313
314
315
        else:
            self.max_encoder_len = 0

316
        # Sampler
317
        self.sampler = Sampler(logprobs_mode=self.model_config.logprobs_mode)
318

319
        self.eplb_state: EplbState | None = None
320
321
322
323
324
325
        """
        State of the expert parallelism load balancer.

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

326
        # Lazy initializations
327
        # self.model: nn.Module  # Set after load_model
328
        # Initialize in initialize_kv_cache
329
        self.kv_caches: list[torch.Tensor] = []
330
331
        # indexes: [kv_cache_group_id][attn_group]
        self.attn_groups: list[list[AttentionGroup]] = []
332
333
        # self.kv_cache_config: KVCacheConfig

334
335
        # mm_hash ->  encoder_output
        self.encoder_cache: dict[str, torch.Tensor] = {}
336

337
        self.use_aux_hidden_state_outputs = False
338
339
340
341
342
        # 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:
343
344
345
            self.drafter: (
                NgramProposer | SuffixDecodingProposer | EagleProposer | MedusaProposer
            )
346
347
            if self.speculative_config.method == "ngram":
                self.drafter = NgramProposer(self.vllm_config)
348
349
            elif self.speculative_config.method == "suffix":
                self.drafter = SuffixDecodingProposer(self.vllm_config)
350
            elif self.speculative_config.use_eagle():
351
                self.drafter = EagleProposer(self.vllm_config, self.device, self)
352
353
354
355
                if self.speculative_config.method == "eagle3":
                    self.use_aux_hidden_state_outputs = True
            elif self.speculative_config.method == "medusa":
                self.drafter = MedusaProposer(
356
                    vllm_config=self.vllm_config, device=self.device
357
                )
358
            else:
359
360
361
362
                raise ValueError(
                    "Unknown speculative decoding method: "
                    f"{self.speculative_config.method}"
                )
363
            self.rejection_sampler = RejectionSampler(self.sampler)
364

365
        # Request states.
366
        self.requests: dict[str, CachedRequestState] = {}
367
        self.comm_stream = torch.cuda.Stream()
368

369
370
371
372
373
374
375
376
377
        # 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.
378
        custom_logitsprocs = model_config.logits_processors
379
380
        self.input_batch = InputBatch(
            max_num_reqs=self.max_num_reqs,
381
382
383
            # 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),
384
385
386
            max_num_batched_tokens=self.max_num_tokens,
            device=self.device,
            pin_memory=self.pin_memory,
387
            vocab_size=self.model_config.get_vocab_size(),
388
            block_sizes=[self.cache_config.block_size],
389
            kernel_block_sizes=[self.cache_config.block_size],
390
            is_spec_decode=bool(self.vllm_config.speculative_config),
391
            logitsprocs=build_logitsprocs(
392
393
394
                self.vllm_config,
                self.device,
                self.pin_memory,
395
                self.is_pooling_model,
396
                custom_logitsprocs,
397
            ),
398
399
400
            # 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),
401
            is_pooling_model=self.is_pooling_model,
402
            dcp_kv_cache_interleave_size=self.parallel_config.dcp_kv_cache_interleave_size,
403
        )
404

405
        self.use_async_scheduling = self.scheduler_config.async_scheduling
406
407
408
409
410
411
412
413
414
        # 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()
415

416
        # self.cudagraph_batch_sizes sorts in ascending order.
417
418
419
420
        if (
            self.compilation_config.cudagraph_capture_sizes
            and self.compilation_config.cudagraph_mode != CUDAGraphMode.NONE
        ):
421
422
            self.cudagraph_batch_sizes = sorted(
                self.compilation_config.cudagraph_capture_sizes
423
            )
424

425
        # Cache the device properties.
426
        self._init_device_properties()
427

428
        # Persistent buffers for CUDA graphs.
429
430
431
432
433
        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
        )
434
        self.seq_lens = self._make_buffer(self.max_num_reqs, dtype=torch.int32)
435
436
437
438
        if self.dcp_world_size > 1:
            self.dcp_local_seq_lens = self._make_buffer(
                self.max_num_reqs, dtype=torch.int32
            )
439
440
441
        # 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.
442
443
444
445
446
447
448
        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
        )
449
450
        self.num_discarded_requests = 0

451
452
453
454
455
456
        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
        )
457

458
459
        # Only relevant for multimodal models
        if self.supports_mm_inputs:
460
            self.is_mm_embed = self._make_buffer(self.max_num_tokens, dtype=torch.bool)
461

462
        # Only relevant for models using M-RoPE (e.g, Qwen2-VL)
463
        if self.uses_mrope:
Roger Wang's avatar
Roger Wang committed
464
465
466
467
            # 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
468
469
470
471
472
473

            # 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
474
            self.mrope_positions = self._make_buffer(
475
476
                (3, self.max_num_tokens + 1), dtype=torch.int64
            )
477

478
        # None in the first PP rank. The rest are set after load_model.
479
        self.intermediate_tensors: IntermediateTensors | None = None
480

481
        # OPTIMIZATION: Cache the tensors rather than creating them every step.
482
        # Keep in int64 to avoid overflow with long context
483
484
485
486
        self.arange_np = np.arange(
            max(self.max_num_reqs + 1, self.max_model_len, self.max_num_tokens),
            dtype=np.int64,
        )
487

488
489
490
491
492
        # 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] = {}
493
494
495
496
497
        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(
498
499
                self.max_num_tokens, dtype=torch.int32, device=self.device
            )
500

501
502
503
504
505
        self.uniform_decode_query_len = (
            1
            if not self.speculative_config
            else 1 + self.speculative_config.num_speculative_tokens
        )
506
507
508
509

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

510
511
512
513
514
515
516
517
518
        self.mm_budget = (
            MultiModalBudget(
                self.model_config,
                self.scheduler_config,
                self.mm_registry,
            )
            if self.supports_mm_inputs
            else None
        )
519

520
        self.reorder_batch_threshold: int | None = None
521

522
523
524
525
526
        # 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()

527
        # Cached outputs.
528
        self._draft_token_ids: list[list[int]] | torch.Tensor | None = None
529
530
        self.transfer_event = torch.cuda.Event()
        self.sampled_token_ids_pinned_cpu = torch.empty(
531
            (self.max_num_reqs, 1),
532
533
            dtype=torch.int64,
            device="cpu",
534
535
            pin_memory=self.pin_memory,
        )
536

537
538
539
        # Ephemeral state transferred between execute_model() and sample_tokens().
        self.execute_model_state: ExecuteModelState | None = None

540
541
542
543
    def reset_mm_cache(self) -> None:
        if self.mm_budget:
            self.mm_budget.reset_cache()

544
545
546
547
548
549
550
551
552
553
    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]

554
    def _make_buffer(
555
        self, *size: int | torch.SymInt, dtype: torch.dtype, numpy: bool = True
556
557
558
559
560
561
562
563
    ) -> CpuGpuBuffer:
        return CpuGpuBuffer(
            *size,
            dtype=dtype,
            device=self.device,
            pin_memory=self.pin_memory,
            with_numpy=numpy,
        )
564

565
566
567
    def _init_model_kwargs(self, num_tokens: int):
        model_kwargs = dict[str, Any]()

568
        if not self.is_pooling_model:
569
570
            return model_kwargs

571
572
        num_reqs = self.input_batch.num_reqs
        pooling_params = self.input_batch.get_pooling_params()
573
574
575

        token_type_id_requests = dict[int, Any]()
        for i, param in enumerate(pooling_params):
576
577
578
579
580
            if (
                param.extra_kwargs is not None
                and (token_types := param.extra_kwargs.get("compressed_token_type_ids"))
                is not None
            ):
581
582
583
584
585
                token_type_id_requests[i] = token_types

        if len(token_type_id_requests) == 0:
            return model_kwargs

586
        seq_lens = self.seq_lens.gpu[:num_reqs]
587
588
589
590
591
592
593
594
        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(
595
596
            device=self.device
        )
597
598
        return model_kwargs

599
    def _may_reorder_batch(self, scheduler_output: "SchedulerOutput") -> None:
600
601
        """
        Update the order of requests in the batch based on the attention
602
        backend's needs. For example, some attention backends (namely MLA) may
603
604
605
606
607
608
        want to separate requests based on if the attention computation will be
        compute-bound or memory-bound.

        Args:
            scheduler_output: The scheduler output.
        """
609
610
611
612
613
614
615
616
        # 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

617
        if self.reorder_batch_threshold is not None:
618
619
620
            # 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.
621
622
623
624
            if (
                self.dcp_world_size > 1
                and envs.VLLM_ATTENTION_BACKEND != "FLASH_ATTN_MLA"
            ):
625
                assert self.reorder_batch_threshold == 1, (
626
                    "DCP not support reorder_batch_threshold > 1 now."
627
                )
628
629
630
            reorder_batch_to_split_decodes_and_prefills(
                self.input_batch,
                scheduler_output,
631
632
                decode_threshold=self.reorder_batch_threshold,
            )
633

634
635
    # Note: used for model runner override.
    def _init_device_properties(self) -> None:
636
        """Initialize attributes from torch.cuda.get_device_properties"""
637
638
639
640
641
642
643
        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()

644
    def _update_states(self, scheduler_output: "SchedulerOutput") -> None:
645
646
647
648
649
650
        """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.

651
652
        The SamplingMetadata is updated and copied to the GPU if there is a
        new/resumed/paused/finished request in the batch.
653
654
        """
        # Remove finished requests from the cached states.
655
656
        for req_id in scheduler_output.finished_req_ids:
            self.requests.pop(req_id, None)
657
658
659
660
661
662
663
        # 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:
664
            self.input_batch.remove_request(req_id)
665
666

        # Free the cached encoder outputs.
667
668
        for mm_hash in scheduler_output.free_encoder_mm_hashes:
            self.encoder_cache.pop(mm_hash, None)
669

670
671
672
673
674
675
676
677
678
679
680
681
682
        # 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:
683
            self.input_batch.remove_request(req_id)
684

685
        reqs_to_add: list[CachedRequestState] = []
686
        # Add new requests to the cached states.
687
688
689
        for new_req_data in scheduler_output.scheduled_new_reqs:
            req_id = new_req_data.req_id
            sampling_params = new_req_data.sampling_params
690
            pooling_params = new_req_data.pooling_params
691

692
693
694
695
            if (
                sampling_params
                and sampling_params.sampling_type == SamplingType.RANDOM_SEED
            ):
696
697
698
699
700
                generator = torch.Generator(device=self.device)
                generator.manual_seed(sampling_params.seed)
            else:
                generator = None

701
702
            if self.is_pooling_model:
                assert pooling_params is not None
703
704
                task = pooling_params.task
                assert task is not None, "You did not set `task` in the API"
705

706
                model = cast(VllmModelForPooling, self.get_model())
707
                to_update = model.pooler.get_pooling_updates(task)
708
709
                to_update.apply(pooling_params)

710
            req_state = CachedRequestState(
711
                req_id=req_id,
712
                prompt_token_ids=new_req_data.prompt_token_ids,
713
                prompt_embeds=new_req_data.prompt_embeds,
714
                mm_features=new_req_data.mm_features,
715
                sampling_params=sampling_params,
716
                pooling_params=pooling_params,
717
                generator=generator,
718
719
                block_ids=new_req_data.block_ids,
                num_computed_tokens=new_req_data.num_computed_tokens,
720
                output_token_ids=[],
721
                lora_request=new_req_data.lora_request,
722
            )
723
724
            self.requests[req_id] = req_state

725
            # Only relevant for models using M-RoPE (e.g, Qwen2-VL)
726
            if self.uses_mrope:
727
                self._init_mrope_positions(req_state)
728

729
            reqs_to_add.append(req_state)
730

731
        # Update the states of the running/resumed requests.
732
        is_last_rank = get_pp_group().is_last_rank
733
734
        req_data = scheduler_output.scheduled_cached_reqs
        for i, req_id in enumerate(req_data.req_ids):
735
            req_state = self.requests[req_id]
736
737
            num_computed_tokens = req_data.num_computed_tokens[i]
            new_block_ids = req_data.new_block_ids[i]
738
            resumed_from_preemption = req_id in req_data.resumed_req_ids
739
            num_output_tokens = req_data.num_output_tokens[i]
740

741
            # Update the cached states.
742

743
            req_state.num_computed_tokens = num_computed_tokens
744
            req_index = self.input_batch.req_id_to_index.get(req_id)
745
746
747
748
749
750
751
752

            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.
753
754
755
                num_new_tokens = (
                    num_computed_tokens + len(new_token_ids) - req_state.num_tokens
                )
756
757
758
759
                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:
760
                    req_state.output_token_ids.extend(new_token_ids[-num_new_tokens:])
761
762
763
764
765
            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:
766
767
768
769
                    end_idx = (
                        self.input_batch.num_prompt_tokens[req_index]
                        + num_output_tokens
                    )
770
771
                    self.input_batch.num_tokens[req_index] = end_idx
                    self.input_batch.num_tokens_no_spec[req_index] = end_idx
772

773
            # Update the block IDs.
774
            if not resumed_from_preemption:
775
776
                if new_block_ids is not None:
                    # Append the new blocks to the existing block IDs.
777
                    for block_ids, new_ids in zip(req_state.block_ids, new_block_ids):
778
                        block_ids.extend(new_ids)
779
            else:
780
                assert req_index is None
781
                assert new_block_ids is not None
782
783
                # The request is resumed from preemption.
                # Replace the existing block IDs with the new ones.
784
                req_state.block_ids = new_block_ids
785
786
787
788
789

            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.
790
791
792
793
794
795
796

                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:]

797
                reqs_to_add.append(req_state)
798
799
800
                continue

            # Update the persistent batch.
801
            self.input_batch.num_computed_tokens_cpu[req_index] = num_computed_tokens
802
            if new_block_ids is not None:
803
                self.input_batch.block_table.append_row(new_block_ids, req_index)
804
805
806
807
808
809
810

            # 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)
811
                self.input_batch.token_ids_cpu[
812
813
814
                    req_index, start_token_index:end_token_index
                ] = new_token_ids
                self.input_batch.num_tokens_no_spec[req_index] = end_token_index
815
                self.input_batch.num_tokens[req_index] = end_token_index
816

817
            # Add spec_token_ids to token_ids_cpu.
818
            spec_token_ids = scheduler_output.scheduled_spec_decode_tokens.get(
819
                req_id, []
820
            )
821
822
823
824
825
            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[
826
827
                    req_index, start_index:end_token_index
                ] = spec_token_ids
828
829
                # NOTE(woosuk): `num_tokens` here may include spec tokens.
                self.input_batch.num_tokens[req_index] += num_spec_tokens
830
831
832
833
834
835
836

            # 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
837

838
839
        # Add the new or resumed requests to the persistent batch.
        # The smaller empty indices are filled first.
840
841
        for request in reqs_to_add:
            self.input_batch.add_request(request)
842

843
844
845
846
847
848
        # 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()
849

850
    def _update_states_after_model_execute(
851
852
        self, output_token_ids: torch.Tensor
    ) -> None:
853
854
855
856
857
858
859
860
861
862
863
864
        """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.
865
866
867
868
869
870
871
872
873
874
875
876
877
878
879
880
881
882
883
884
        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()
        )
885
886
887
        for i, num_tokens in enumerate(num_accepted_tokens):
            self.input_batch.num_accepted_tokens_cpu[i] = num_tokens

888
889
890
891
892
893
    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
894
895
896
897
        for mm_feature in req_state.mm_features:
            mm_item = mm_feature.data
            if mm_item is None:
                continue
898
899
900
901
902
903
904
905
906
907
908
909
            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

910
911
912
913
914
915
916
917
918
919
920
        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,
921
            )
922
        )
923

924
    def _extract_mm_kwargs(
925
        self,
926
927
        scheduler_output: "SchedulerOutput",
    ) -> BatchedTensorInputs:
928
        if not scheduler_output or not self.is_multimodal_raw_input_only_model:
929
            return {}
930

931
932
        mm_kwargs = list[MultiModalKwargsItem]()
        for req in scheduler_output.scheduled_new_reqs:
933
934
935
            for feature in req.mm_features:
                if feature.data is not None:
                    mm_kwargs.append(feature.data)
936

937
        # Input all modalities at once
938
        model = cast(SupportsMultiModal, self.model)
939
940
        mm_kwargs_combined: BatchedTensorInputs = {}
        for _, _, mm_kwargs_group in group_mm_kwargs_by_modality(
941
942
943
944
            mm_kwargs,
            device=self.device,
            pin_memory=self.pin_memory,
            merge_by_field_config=model.merge_by_field_config,
945
            multimodal_cpu_fields=model.multimodal_cpu_fields,
946
947
        ):
            mm_kwargs_combined.update(mm_kwargs_group)
948

949
        return mm_kwargs_combined
950

951
    def _dummy_mm_kwargs(self, num_seqs: int) -> BatchedTensorInputs:
952
        if not self.is_multimodal_raw_input_only_model:
953
            return {}
954

955
956
957
958
959
        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)
960

961
962
963
    def _get_cumsum_and_arange(
        self,
        num_tokens: np.ndarray,
964
        cumsum_dtype: np.dtype | None = None,
965
966
967
968
969
970
971
972
973
974
975
976
977
978
979
980
    ) -> 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

981
982
983
    def _prepare_input_ids(
        self, total_num_scheduled_tokens: int, cu_num_tokens: np.ndarray
    ) -> None:
984
        """Prepare the input IDs for the current batch.
985

986
987
988
989
990
991
992
        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)
993
994
995
            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)
996
997
998
999
1000
1001
1002
1003
1004
1005
1006
1007
1008
1009
1010
1011
1012
1013
            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)
1014
                indices_match &= prev_index == flattened_index
1015
1016
1017
1018
1019
1020
                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)
1021
1022
1023
            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)
1024
1025
        if num_commmon_tokens == 0:
            # No requests in common with the previous iteration
1026
            # So input_ids.cpu will have all the input ids.
1027
1028
1029
1030
1031
1032
1033
            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_(
1034
1035
1036
                self.input_batch.prev_sampled_token_ids[:num_commmon_tokens, 0],
                non_blocking=True,
            )
1037
1038
            if self.enable_prompt_embeds:
                self.is_token_ids.gpu[:num_commmon_tokens] = True
1039
            return
1040
        # Upload the index tensors asynchronously so the scatter can be non-blocking.
1041
1042
1043
        input_ids_index_tensor = torch.tensor(
            flattened_indices, dtype=torch.int64, pin_memory=self.pin_memory
        ).to(self.device, non_blocking=True)
1044
        prev_common_req_indices_tensor = torch.tensor(
1045
1046
            prev_common_req_indices, dtype=torch.int64, pin_memory=self.pin_memory
        ).to(self.device, non_blocking=True)
1047
1048
1049
1050
        self.input_ids.gpu.scatter_(
            dim=0,
            index=input_ids_index_tensor,
            src=self.input_batch.prev_sampled_token_ids[
1051
1052
1053
                prev_common_req_indices_tensor, 0
            ],
        )
1054

1055
1056
    def _get_encoder_seq_lens(
        self,
1057
        scheduled_encoder_inputs: dict[str, list[int]],
1058
1059
        kv_cache_spec: KVCacheSpec,
        num_reqs: int,
1060
    ) -> np.ndarray | None:
1061
1062
1063
1064
1065
1066
        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)
1067
        for req_id in scheduled_encoder_inputs:
1068
1069
1070
1071
1072
            req_index = self.input_batch.req_id_to_index[req_id]
            encoder_seq_lens[req_index] = self.max_encoder_len

        return encoder_seq_lens

1073
    def _prepare_inputs(
1074
1075
1076
1077
        self,
        scheduler_output: "SchedulerOutput",
        num_scheduled_tokens: np.ndarray,
        max_num_scheduled_tokens: int,
1078
1079
    ) -> tuple[
        torch.Tensor,
1080
1081
1082
        SpecDecodeMetadata | None,
        UBatchSlices | None,
        torch.Tensor | None,
1083
    ]:
1084
1085
        """
        :return: tuple[
1086
            logits_indices, spec_decode_metadata,
1087
            ubatch_slices, num_tokens_across_dp,
1088
1089
        ]
        """
1090
1091
1092
1093
1094
1095
1096
        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.
1097
        self.input_batch.block_table.commit_block_table(num_reqs)
1098
1099
1100

        # Get request indices.
        # E.g., [2, 5, 3] -> [0, 0, 1, 1, 1, 1, 1, 2, 2, 2]
1101
        req_indices = np.repeat(self.arange_np[:num_reqs], num_scheduled_tokens)
1102

1103
1104
        # cu_num_tokens: [2, 5, 3] -> [2, 7, 10]
        # arange: [0, 1, 0, 1, 2, 3, 4, 0, 1, 2]
1105
        cu_num_tokens, arange = self._get_cumsum_and_arange(num_scheduled_tokens)
1106
1107

        # Get positions.
1108
        positions_np = self.positions.np[:total_num_scheduled_tokens]
1109
1110
1111
1112
1113
        np.add(
            self.input_batch.num_computed_tokens_cpu[req_indices],
            arange,
            out=positions_np,
        )
1114

1115
1116
        # Calculate M-RoPE positions.
        # Only relevant for models using M-RoPE (e.g, Qwen2-VL)
1117
        if self.uses_mrope:
1118
1119
            self._calc_mrope_positions(scheduler_output)

1120
1121
1122
1123
        # 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.
1124
1125
1126
        token_indices = (
            positions_np + req_indices * self.input_batch.token_ids_cpu.shape[1]
        )
1127
        token_indices_tensor = torch.from_numpy(token_indices)
1128

1129
1130
1131
        # 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.
1132
1133
1134
1135
1136
1137
        torch.index_select(
            self.input_batch.token_ids_cpu_tensor.flatten(),
            0,
            token_indices_tensor,
            out=self.input_ids.cpu[:total_num_scheduled_tokens],
        )
1138
        if self.enable_prompt_embeds:
1139
            is_token_ids = self.input_batch.is_token_ids_tensor.flatten()
1140
1141
1142
1143
            torch.index_select(
                is_token_ids,
                0,
                token_indices_tensor,
1144
1145
                out=self.is_token_ids.cpu[:total_num_scheduled_tokens],
            )
1146
1147
1148
1149
1150
1151
1152
1153
1154
1155
1156
1157
1158
1159
1160
1161
1162
1163
1164
1165
1166
1167
1168
1169
1170
1171
1172
1173
1174
1175
1176
1177
1178

        # 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:
1179
1180
1181
                    self.inputs_embeds.cpu[
                        output_idx : output_idx + actual_num_sched
                    ].copy_(req_embeds[start_pos:actual_end])
1182
1183

                output_idx += num_sched
1184

1185
1186
        self.input_batch.block_table.compute_slot_mapping(req_indices, positions_np)
        self.input_batch.block_table.commit_slot_mapping(total_num_scheduled_tokens)
1187
1188

        # Prepare the attention metadata.
1189
        self.query_start_loc.np[0] = 0
1190
        self.query_start_loc.np[1 : num_reqs + 1] = cu_num_tokens
1191
1192
        # Note: pad query_start_loc to be non-decreasing, as kernels
        # like FlashAttention requires that
1193
        self.query_start_loc.np[num_reqs + 1 :].fill(cu_num_tokens[-1])
1194
        self.query_start_loc.copy_to_gpu()
1195
        query_start_loc = self.query_start_loc.gpu[: num_reqs + 1]
1196

1197
        num_tokens_unpadded = scheduler_output.total_num_scheduled_tokens
1198
        num_tokens_padded = self._get_num_input_tokens(num_tokens_unpadded)
1199
1200
1201
        uniform_decode = (
            max_num_scheduled_tokens == self.uniform_decode_query_len
        ) and (total_num_scheduled_tokens == num_reqs * max_num_scheduled_tokens)
1202
1203
1204
1205
1206
1207
1208

        # 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

1209
        ubatch_slices, num_tokens_across_dp = coordinate_batch_across_dp(
1210
1211
1212
1213
1214
1215
1216
            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,
1217
        )
1218

1219
        self.seq_lens.np[:num_reqs] = (
1220
1221
            self.input_batch.num_computed_tokens_cpu[:num_reqs] + num_scheduled_tokens
        )
1222
        # Fill unused with 0 for full cuda graph mode.
1223
1224
        self.seq_lens.np[num_reqs:].fill(0)
        self.seq_lens.copy_to_gpu()
1225

1226
        num_tokens = [self.requests[r].num_tokens for r in self.input_batch.req_ids]
1227
1228
1229
1230
1231
1232
1233
        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)
1234
1235
1236
        self.discard_request_indices.np[: self.num_discarded_requests] = (
            discard_request_indices
        )
1237
1238
1239

        self.discard_request_indices.copy_to_gpu(self.num_discarded_requests)

1240
        # Copy the tensors to the GPU.
1241
1242
        self._prepare_input_ids(total_num_scheduled_tokens, cu_num_tokens)

1243
        if self.uses_mrope:
1244
            # Only relevant for models using M-RoPE (e.g, Qwen2-VL)
1245
1246
            self.mrope_positions.gpu[:, :total_num_scheduled_tokens].copy_(
                self.mrope_positions.cpu[:, :total_num_scheduled_tokens],
1247
1248
                non_blocking=True,
            )
1249
1250
        else:
            # Common case (1D positions)
1251
            self.positions.copy_to_gpu(total_num_scheduled_tokens)
1252

1253
        use_spec_decode = len(scheduler_output.scheduled_spec_decode_tokens) > 0
1254
1255
1256
1257
1258
1259
1260
        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
1261
            num_draft_tokens = None
1262
            spec_decode_metadata = None
1263
            num_sampled_tokens = np.ones(num_reqs, dtype=np.int32)
1264
1265
1266
1267
1268
        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)
1269
1270
1271
            # 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)
1272
1273
1274
1275
            for (
                req_id,
                draft_token_ids,
            ) in scheduler_output.scheduled_spec_decode_tokens.items():
1276
1277
                req_idx = self.input_batch.req_id_to_index[req_id]
                num_draft_tokens[req_idx] = len(draft_token_ids)
1278
1279
1280
1281
1282
1283
1284
1285
                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
                )
1286
            spec_decode_metadata = self._calc_spec_decode_metadata(
1287
1288
                num_draft_tokens, cu_num_tokens
            )
1289
            logits_indices = spec_decode_metadata.logits_indices
1290
            num_sampled_tokens = num_draft_tokens + 1
1291
            # For DECODE only cuda graph of some attention backends (e.g., GDN).
1292
            self.num_decode_draft_tokens.np[:num_reqs] = num_decode_draft_tokens
1293
1294
            self.num_decode_draft_tokens.np[num_reqs:].fill(-1)
            self.num_decode_draft_tokens.copy_to_gpu()
1295

1296
1297
1298
1299
1300
        # Hot-Swap lora model
        if self.lora_config:
            assert (
                np.sum(num_sampled_tokens)
                <= self.vllm_config.scheduler_config.max_num_batched_tokens
1301
            )
1302
1303
1304
1305
1306
1307
1308
1309
1310
1311
1312
1313
1314
1315
1316
1317
1318
1319
1320
1321
1322
1323
1324
1325
1326
1327
1328
1329
1330
1331
1332
1333
1334
1335
            self.set_active_loras(
                self.input_batch, num_scheduled_tokens, num_sampled_tokens
            )

        return (
            logits_indices,
            spec_decode_metadata,
            ubatch_slices,
            num_tokens_across_dp,
        )

    def _build_attention_metadata(
        self,
        total_num_scheduled_tokens: int,
        max_num_scheduled_tokens: int,
        num_reqs: int,
        ubatch_slices: UBatchSlices | None = None,
        logits_indices: torch.Tensor | None = None,
        use_spec_decode: bool = False,
        for_cudagraph_capture: bool = False,
        scheduled_encoder_inputs: dict[str, list[int]] | None = None,
        cascade_attn_prefix_lens: list[list[int]] | None = None,
    ) -> tuple[PerLayerAttnMetadata, CommonAttentionMetadata | None]:
        """
        :return: tuple[attn_metadata, spec_decode_common_attn_metadata]
        """
        logits_indices_padded = None
        num_logits_indices = 0
        if logits_indices is not None:
            num_logits_indices = logits_indices.size(0)
            if self.cache_config.kv_sharing_fast_prefill:
                logits_indices_padded = self._prepare_kv_sharing_fast_prefill(
                    logits_indices
                )
1336

1337
1338
1339
1340
1341
1342
1343
1344
1345
1346
        # update seq_lens of decode reqs under DCP.
        if self.dcp_world_size > 1:
            self.dcp_local_seq_lens.cpu[:num_reqs] = get_dcp_local_seq_lens(
                self.seq_lens.cpu[:num_reqs],
                self.dcp_world_size,
                self.dcp_rank,
                self.parallel_config.dcp_kv_cache_interleave_size,
            )
            self.dcp_local_seq_lens.copy_to_gpu(num_reqs)

1347
1348
1349
        attn_metadata: PerLayerAttnMetadata = {}
        if ubatch_slices is not None:
            attn_metadata = [dict() for _ in range(len(ubatch_slices))]
1350

1351
1352
        # Used in the below loop
        query_start_loc = self.query_start_loc.gpu[: num_reqs + 1]
1353
        query_start_loc_cpu = self.query_start_loc.cpu[: num_reqs + 1]
1354
        seq_lens = self.seq_lens.gpu[:num_reqs]
1355
        seq_lens_cpu = self.seq_lens.cpu[:num_reqs]
1356
1357
1358
        num_computed_tokens_cpu = self.input_batch.num_computed_tokens_cpu_tensor[
            :num_reqs
        ]
1359
1360
1361
        dcp_local_seq_lens = (
            self.dcp_local_seq_lens.gpu[:num_reqs] if self.dcp_world_size > 1 else None
        )
1362
        spec_decode_common_attn_metadata = None
1363
1364
1365
1366
1367
1368
1369
1370
1371

        if for_cudagraph_capture:
            # For some attention backends (e.g. FA) with sliding window models we need
            # to make sure the backend see a max_seq_len that is larger to the sliding
            # window size when capturing to make sure the correct kernel is selected.
            max_seq_len = self.max_model_len
        else:
            max_seq_len = self.seq_lens.np[:num_reqs].max().item()

1372
1373
        if use_spec_decode:
            self.num_accepted_tokens.np[:num_reqs] = (
1374
1375
                self.input_batch.num_accepted_tokens_cpu[:num_reqs]
            )
1376
1377
            self.num_accepted_tokens.np[num_reqs:].fill(1)
            self.num_accepted_tokens.copy_to_gpu()
1378

1379
1380
        # Prepare the attention metadata for each KV cache group and make layers
        # in the same group share the same metadata.
1381
        for kv_cache_gid, kv_cache_group in enumerate(
1382
1383
            self.kv_cache_config.kv_cache_groups
        ):
1384
            encoder_seq_lens = self._get_encoder_seq_lens(
1385
1386
1387
                scheduled_encoder_inputs or {},
                kv_cache_group.kv_cache_spec,
                num_reqs,
1388
            )
1389

1390
            if isinstance(kv_cache_group.kv_cache_spec, EncoderOnlyAttentionSpec):
1391
1392
1393
1394
1395
                # 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,
1396
1397
1398
                    device=self.device,
                )
                slot_mapping = torch.zeros(
1399
                    (total_num_scheduled_tokens,),
1400
1401
1402
                    dtype=torch.int64,
                    device=self.device,
                )
1403
            else:
1404
                blk_table = self.input_batch.block_table[kv_cache_gid]
1405
                blk_table_tensor = blk_table.get_device_tensor(num_reqs)
1406
                slot_mapping = blk_table.slot_mapping.gpu[:total_num_scheduled_tokens]
1407
1408
1409

                # Fill unused with -1. Needed for reshape_and_cache in full cuda
                # graph mode.
1410
                blk_table.slot_mapping.gpu[total_num_scheduled_tokens:].fill_(-1)
1411

1412
            common_attn_metadata = CommonAttentionMetadata(
1413
1414
1415
1416
1417
                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,
1418
1419
1420
                num_reqs=num_reqs,
                num_actual_tokens=total_num_scheduled_tokens,
                max_query_len=max_num_scheduled_tokens,
1421
                max_seq_len=max_seq_len,
1422
1423
                block_table_tensor=blk_table_tensor,
                slot_mapping=slot_mapping,
1424
                logits_indices_padded=logits_indices_padded,
1425
                num_logits_indices=num_logits_indices,
1426
                causal=True,
1427
                encoder_seq_lens=encoder_seq_lens,
1428
                dcp_local_seq_lens=dcp_local_seq_lens,
1429
1430
            )

1431
            if self.speculative_config and spec_decode_common_attn_metadata is None:
1432
                if isinstance(self.drafter, EagleProposer):
1433
                    if self.drafter.attn_layer_names[0] in kv_cache_group.layer_names:
1434
1435
1436
                        spec_decode_common_attn_metadata = common_attn_metadata
                else:
                    spec_decode_common_attn_metadata = common_attn_metadata
1437

1438
1439
1440
1441
1442
1443
            for attn_gid, attn_group in enumerate(self.attn_groups[kv_cache_gid]):
                cascade_attn_prefix_len = (
                    cascade_attn_prefix_lens[kv_cache_gid][attn_gid]
                    if cascade_attn_prefix_lens
                    else 0
                )
1444
                builder = attn_group.get_metadata_builder()
1445

1446
                extra_attn_metadata_args = {}
1447
                if use_spec_decode and isinstance(builder, GDNAttentionMetadataBuilder):
1448
                    extra_attn_metadata_args = dict(
1449
1450
1451
1452
                        num_accepted_tokens=self.num_accepted_tokens.gpu[:num_reqs],
                        num_decode_draft_tokens_cpu=self.num_decode_draft_tokens.cpu[
                            :num_reqs
                        ],
1453
1454
                    )

1455
1456
                if ubatch_slices is not None:
                    common_attn_metadata_list = split_attn_metadata(
1457
1458
                        ubatch_slices, common_attn_metadata
                    )
1459
                    for ubid, common_attn_metadata in enumerate(
1460
1461
                        common_attn_metadata_list
                    ):
1462
1463
1464
1465
1466
1467
1468
1469
1470
1471
1472
                        builder = attn_group.get_metadata_builder(ubatch_id=ubid)
                        if for_cudagraph_capture:
                            attn_metadata_i = builder.build_for_cudagraph_capture(
                                common_attn_metadata
                            )
                        else:
                            attn_metadata_i = builder.build(
                                common_prefix_len=cascade_attn_prefix_len,
                                common_attn_metadata=common_attn_metadata,
                            )
                        for layer_name in kv_cache_group.layer_names:
1473
1474
1475
1476
                            assert type(attn_metadata) is list
                            attn_metadata[ubid][layer_name] = attn_metadata_i
                else:
                    assert isinstance(attn_metadata, dict)
1477
1478
1479
1480
1481
1482
1483
1484
1485
1486
                    if for_cudagraph_capture:
                        attn_metadata_i = builder.build_for_cudagraph_capture(
                            common_attn_metadata
                        )
                    else:
                        attn_metadata_i = builder.build(
                            common_prefix_len=cascade_attn_prefix_len,
                            common_attn_metadata=common_attn_metadata,
                            **extra_attn_metadata_args,
                        )
1487
1488
                    for layer_name in attn_group.layer_names:
                        attn_metadata[layer_name] = attn_metadata_i
1489

1490
        return attn_metadata, spec_decode_common_attn_metadata
1491

1492
1493
1494
1495
1496
1497
1498
1499
1500
1501
    def _compute_cascade_attn_prefix_lens(
        self,
        num_scheduled_tokens: np.ndarray,
        num_common_prefix_blocks: list[int],
    ) -> list[list[int]] | None:
        """
        :return: Optional[cascade_attn_prefix_lens]
            cascade_attn_prefix_lens is 2D: ``[kv_cache_group_id][attn_group_idx]``,
            None if we should not use cascade attention
        """
1502

1503
1504
1505
1506
1507
1508
1509
1510
1511
1512
1513
1514
1515
1516
1517
1518
1519
1520
1521
1522
1523
1524
        use_cascade_attn = False
        num_kv_cache_groups = len(self.kv_cache_config.kv_cache_groups)
        cascade_attn_prefix_lens: list[list[int]] = [
            [] for _ in range(num_kv_cache_groups)
        ]

        for kv_cache_gid in range(num_kv_cache_groups):
            for attn_group in self.attn_groups[kv_cache_gid]:
                if isinstance(attn_group.kv_cache_spec, EncoderOnlyAttentionSpec):
                    cascade_attn_prefix_len = 0
                else:
                    # 0 if cascade attention should not be used
                    cascade_attn_prefix_len = self._compute_cascade_attn_prefix_len(
                        num_scheduled_tokens,
                        num_common_prefix_blocks[kv_cache_gid],
                        attn_group.kv_cache_spec,
                        attn_group.get_metadata_builder(),
                    )
                cascade_attn_prefix_lens[kv_cache_gid].append(cascade_attn_prefix_len)
                use_cascade_attn |= cascade_attn_prefix_len > 0

        return cascade_attn_prefix_lens if use_cascade_attn else None
1525

1526
1527
1528
1529
    def _compute_cascade_attn_prefix_len(
        self,
        num_scheduled_tokens: np.ndarray,
        num_common_prefix_blocks: int,
1530
1531
        kv_cache_spec: KVCacheSpec,
        attn_metadata_builder: AttentionMetadataBuilder,
1532
1533
1534
1535
1536
1537
1538
1539
1540
1541
1542
1543
1544
1545
1546
1547
1548
1549
    ) -> 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.
        """
1550

1551
        common_prefix_len = num_common_prefix_blocks * kv_cache_spec.block_size
1552
1553
1554
1555
1556
1557
1558
1559
1560
1561
1562
1563
1564
1565
1566
1567
1568
1569
1570
1571
1572
1573
1574
1575
1576
1577
1578
1579
1580
1581
1582
1583
1584
1585
1586
1587
1588
        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]
1589
        # Request 3's num_computed_tokens: 3 (i.e., [A, B, C])
1590
1591
1592
1593
1594
1595
1596
        # 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(
1597
1598
            common_prefix_len, self.input_batch.num_computed_tokens_cpu[:num_reqs].min()
        )
1599
        # common_prefix_len should be a multiple of the block size.
1600
1601
1602
1603
1604
1605
1606
1607
1608
1609
1610
        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
        )
1611
1612
        assert isinstance(kv_cache_spec, AttentionSpec)
        use_cascade = attn_metadata_builder.use_cascade_attention(
1613
1614
1615
            common_prefix_len=common_prefix_len,
            query_lens=num_scheduled_tokens,
            num_query_heads=self.num_query_heads,
1616
            num_kv_heads=kv_cache_spec.num_kv_heads,
1617
            use_alibi=self.use_alibi,
1618
            use_sliding_window=use_sliding_window,
1619
            use_local_attention=use_local_attention,
1620
            num_sms=self.num_sms,
1621
            dcp_world_size=self.dcp_world_size,
1622
1623
1624
        )
        return common_prefix_len if use_cascade else 0

1625
1626
    def _calc_mrope_positions(self, scheduler_output: "SchedulerOutput"):
        mrope_pos_ptr = 0
1627
        for index, req_id in enumerate(self.input_batch.req_ids):
1628
1629
1630
            req = self.requests[req_id]
            assert req.mrope_positions is not None

1631
1632
            num_computed_tokens = self.input_batch.num_computed_tokens_cpu[index]
            num_scheduled_tokens = scheduler_output.num_scheduled_tokens[req_id]
1633
            num_prompt_tokens = length_from_prompt_token_ids_or_embeds(
1634
1635
                req.prompt_token_ids, req.prompt_embeds
            )
1636
1637

            if num_computed_tokens + num_scheduled_tokens > num_prompt_tokens:
1638
1639
                prompt_part_len = max(0, num_prompt_tokens - num_computed_tokens)
                completion_part_len = max(0, num_scheduled_tokens - prompt_part_len)
1640
1641
1642
1643
1644
1645
1646
1647
1648
1649
1650
1651
1652
            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

1653
1654
1655
                self.mrope_positions.cpu[:, dst_start:dst_end] = req.mrope_positions[
                    :, src_start:src_end
                ]
1656
1657
1658
1659
1660
1661
1662
                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

1663
                MRotaryEmbedding.get_next_input_positions_tensor(
1664
                    out=self.mrope_positions.np,
1665
1666
1667
1668
1669
                    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,
                )
1670
1671
1672

                mrope_pos_ptr += completion_part_len

1673
1674
    def _calc_spec_decode_metadata(
        self,
1675
1676
1677
1678
1679
1680
1681
1682
1683
1684
1685
1686
1687
1688
1689
1690
        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
1691
1692
1693
1694

        # 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(
1695
1696
            num_sampled_tokens, cumsum_dtype=np.int32
        )
1697
        # Step 2. [0, 0, 0, 0, 103, 104, 104, 104, 206, 207, 207]
1698
        logits_indices = np.repeat(
1699
1700
            cu_num_scheduled_tokens - num_sampled_tokens, num_sampled_tokens
        )
1701
        # Step 3. [0, 1, 2, 3, 103, 104, 105, 106, 206, 207, 208]
1702
1703
1704
1705
1706
1707
        logits_indices += arange

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

        # Compute the draft logits indices.
1708
1709
1710
        # 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(
1711
1712
            num_draft_tokens, cumsum_dtype=np.int32
        )
1713
1714
        # [0, 0, 0, 5, 5, 9]
        target_logits_indices = np.repeat(
1715
1716
            cu_num_sampled_tokens - num_sampled_tokens, num_draft_tokens
        )
1717
1718
1719
1720
1721
        # [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(
1722
1723
            self.device, non_blocking=True
        )
1724
1725
1726
        cu_num_sampled_tokens = torch.from_numpy(cu_num_sampled_tokens).to(
            self.device, non_blocking=True
        )
1727
1728
1729
        logits_indices = torch.from_numpy(logits_indices).to(
            self.device, non_blocking=True
        )
1730
        target_logits_indices = torch.from_numpy(target_logits_indices).to(
1731
1732
            self.device, non_blocking=True
        )
1733
        bonus_logits_indices = torch.from_numpy(bonus_logits_indices).to(
1734
1735
            self.device, non_blocking=True
        )
1736

1737
1738
        # Compute the draft token ids.
        # draft_token_indices:      [  1,   2,   3, 105, 106, 208]
1739
        draft_token_ids = self.input_ids.gpu[logits_indices]
1740
1741
        draft_token_ids = draft_token_ids[target_logits_indices + 1]

1742
        return SpecDecodeMetadata(
1743
1744
1745
            draft_token_ids=draft_token_ids,
            num_draft_tokens=num_draft_tokens.tolist(),
            cu_num_draft_tokens=cu_num_draft_tokens,
1746
            cu_num_sampled_tokens=cu_num_sampled_tokens,
1747
1748
1749
1750
1751
            target_logits_indices=target_logits_indices,
            bonus_logits_indices=bonus_logits_indices,
            logits_indices=logits_indices,
        )

1752
1753
1754
1755
1756
1757
1758
    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
1759
        self.kv_sharing_fast_prefill_logits_indices[:num_logits].copy_(logits_indices)
1760
1761
1762
1763
1764
        # 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_(
1765
1766
1767
1768
1769
1770
            logits_indices[-1].item()
        )
        if (
            self.compilation_config.cudagraph_mode != CUDAGraphMode.NONE
            and num_logits <= self.cudagraph_batch_sizes[-1]
        ):
1771
1772
1773
1774
1775
            # 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
1776
1777
1778
        logits_indices_padded = self.kv_sharing_fast_prefill_logits_indices[
            :num_logits_padded
        ]
1779
1780
        return logits_indices_padded

1781
1782
1783
1784
1785
1786
1787
1788
    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
1789
                inputs.
1790
1791
1792
1793
1794
1795

        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
        """
1796
1797
        scheduled_encoder_inputs = scheduler_output.scheduled_encoder_inputs
        if not scheduled_encoder_inputs:
1798
            return [], []
1799
        # Batch the multi-modal inputs.
1800
        mm_kwargs = list[MultiModalKwargsItem]()
1801
1802
        # list of tuple (mm_hash, position_info)
        mm_hashes_pos = list[tuple[str, PlaceholderRange]]()
1803
1804
        for req_id, encoder_input_ids in scheduled_encoder_inputs.items():
            req_state = self.requests[req_id]
1805
1806

            for mm_input_id in encoder_input_ids:
1807
1808
1809
1810
                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))
1811

1812
1813
1814
1815
1816
        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(
1817
1818
            scheduler_output
        )
1819
1820
1821
1822

        if not mm_kwargs:
            return

1823
1824
1825
1826
1827
1828
1829
        # 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.
1830
        model = cast(SupportsMultiModal, self.model)
1831
        encoder_outputs = []
1832
        for modality, num_items, mm_kwargs_group in group_mm_kwargs_by_modality(
1833
1834
1835
1836
            mm_kwargs,
            device=self.device,
            pin_memory=self.pin_memory,
            merge_by_field_config=model.merge_by_field_config,
1837
            multimodal_cpu_fields=model.multimodal_cpu_fields,
1838
        ):
1839
1840
1841
            curr_group_outputs = []

            # EVS-related change.
1842
            # (ekhvedchenia): Temporary hack to limit peak memory usage when
1843
            # processing multimodal data. This solves the issue with scheduler
1844
1845
1846
1847
            # 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)
1848
1849
1850
1851
1852
1853
1854
1855
1856
1857
1858
1859
1860
1861
1862
1863
            # 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,
1864
                            multimodal_cpu_fields=model.multimodal_cpu_fields,
1865
                        )
1866
                    )
1867
1868

                    micro_batch_outputs = model.get_multimodal_embeddings(
1869
1870
                        **micro_batch_mm_inputs
                    )
1871
1872
1873
1874
1875
1876
1877
1878
1879
1880

                    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.
1881
                curr_group_outputs = model.get_multimodal_embeddings(**mm_kwargs_group)
1882

1883
1884
            sanity_check_mm_encoder_outputs(
                curr_group_outputs,
1885
                expected_num_items=num_items,
1886
            )
1887
            encoder_outputs.extend(curr_group_outputs)
1888

1889
1890
1891
        # 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(
1892
1893
1894
1895
1896
                output,
                is_embed=pos_info.is_embed,
            )

    def _gather_mm_embeddings(
1897
1898
        self,
        scheduler_output: "SchedulerOutput",
1899
        shift_computed_tokens: int = 0,
1900
1901
1902
1903
1904
1905
1906
1907
    ) -> 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
1908
        should_sync_mrope_positions = False
1909

1910
        for req_id in self.input_batch.req_ids:
1911
1912
            mm_embeds_req: list[torch.Tensor] = []

1913
            num_scheduled_tokens = scheduler_output.num_scheduled_tokens[req_id]
1914
            req_state = self.requests[req_id]
1915
            num_computed_tokens = req_state.num_computed_tokens + shift_computed_tokens
1916

1917
1918
            for mm_feature in req_state.mm_features:
                pos_info = mm_feature.mm_position
1919
1920
                start_pos = pos_info.offset
                num_encoder_tokens = pos_info.length
1921
1922
1923
1924
1925
1926
1927
1928
1929
1930
1931
1932
1933
1934
1935
1936

                # 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,
1937
1938
                    num_encoder_tokens,
                )
1939
                assert start_idx < end_idx
1940

1941
                mm_hash = mm_feature.identifier
1942
                encoder_output = self.encoder_cache.get(mm_hash, None)
1943
                assert encoder_output is not None, f"Encoder cache miss for {mm_hash}."
1944
1945
1946
1947

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

1948
                req_start_pos = req_start_idx + start_pos - num_computed_tokens
1949
1950
1951
                is_mm_embed[req_start_pos + start_idx : req_start_pos + end_idx] = (
                    True if is_embed is None else is_embed
                )
1952

1953
1954
1955
1956
                mm_embeds_item = gather_mm_placeholders(
                    encoder_output[start_idx:end_idx],
                    is_embed=is_embed,
                )
1957
1958
1959
                mm_embeds_req.append(mm_embeds_item)

            if self.is_multimodal_pruning_enabled and self.uses_mrope:
1960
                assert req_state.mrope_positions is not None
1961
1962
1963
1964
1965
1966
1967
                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,
1968
1969
                    )
                )
1970
1971
1972
1973
                req_state.mrope_positions.copy_(new_mrope_positions)
                req_state.mrope_position_delta = new_delta

            mm_embeds.extend(mm_embeds_req)
1974
1975
1976
            req_start_idx += num_scheduled_tokens

        is_mm_embed = self.is_mm_embed.copy_to_gpu(total_num_scheduled_tokens)
1977
1978
1979

        if should_sync_mrope_positions:
            self._calc_mrope_positions(scheduler_output)
1980
            self.mrope_positions.copy_to_gpu(total_num_scheduled_tokens)
1981

1982
        return mm_embeds, is_mm_embed
1983

1984
1985
1986
1987
1988
1989
1990
1991
1992
1993
1994
1995
1996
1997
1998
1999
    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
2000
        model = cast(SupportsMultiModal, self.model)
2001
2002
        encoder_features = {}
        for _, _, mm_kwargs_group in group_mm_kwargs_by_modality(
2003
2004
2005
2006
            mm_kwargs,
            device=self.device,
            pin_memory=self.pin_memory,
            merge_by_field_config=model.merge_by_field_config,
2007
            multimodal_cpu_fields=model.multimodal_cpu_fields,
2008
2009
2010
2011
2012
2013
2014
2015
        ):
            # 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

2016
    def get_model(self) -> nn.Module:
2017
        # get raw model out of the cudagraph wrapper.
2018
        if isinstance(self.model, (CUDAGraphWrapper, UBatchWrapper)):
2019
            return self.model.unwrap()
2020
2021
        return self.model

2022
2023
2024
2025
2026
2027
2028
2029
2030
2031
2032
2033
2034
2035
2036
    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

2037
2038
2039
2040
2041
    def get_supported_pooling_tasks(self) -> list[PoolingTask]:
        model = self.get_model()
        if not is_pooling_model(model):
            return []

2042
2043
        supported_tasks = list(model.pooler.get_supported_tasks())

2044
2045
2046
2047
2048
        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")
2049

2050
2051
            logger.debug_once(
                "Chunked prefill is not supported with "
2052
2053
                "token_embed and token_classify tasks "
                "which using ALL pooling. "
2054
2055
2056
                "Please turn off chunked prefill by "
                "`--no-enable-chunked-prefill` before using it."
            )
2057
2058
2059
2060
2061

        if "score" in supported_tasks:
            num_labels = getattr(self.model_config.hf_config, "num_labels", 0)
            if num_labels != 1:
                supported_tasks.remove("score")
2062
                logger.debug_once("Score API is only enabled for num_labels == 1.")
2063
2064

        return supported_tasks
2065

2066
2067
2068
2069
2070
2071
2072
2073
2074
2075
    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)

2076
    def sync_and_slice_intermediate_tensors(
2077
2078
2079
2080
2081
        self,
        num_tokens: int,
        intermediate_tensors: IntermediateTensors,
        sync_self: bool,
    ) -> IntermediateTensors:
2082
2083
2084
        assert self.intermediate_tensors is not None

        tp = self.vllm_config.parallel_config.tensor_parallel_size
2085
        is_rs = is_residual_scattered_for_sp(self.vllm_config, num_tokens)
2086
2087
2088
2089
2090
2091

        # 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():
2092
                is_scattered = k == "residual" and is_rs
2093
                copy_len = num_tokens // tp if is_scattered else num_tokens
2094
                self.intermediate_tensors[k][:copy_len].copy_(
2095
2096
2097
2098
2099
2100
2101
2102
2103
2104
2105
2106
2107
                    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:
2108
2109
2110
2111
2112
2113
2114
        """
        Step for the EPLB (Expert Parallelism Load Balancing) state.
        """
        if not self.parallel_config.enable_eplb:
            return

        assert self.eplb_state is not None
2115
2116
        model = self.get_model()
        assert is_mixture_of_experts(model)
2117
2118
2119
        self.eplb_state.step(
            is_dummy,
            is_profile,
2120
            log_stats=self.parallel_config.eplb_config.log_balancedness,
2121
2122
        )

2123
2124
2125
2126
    # 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)
2127
2128
    @staticmethod
    def pad_out_ubatch_slice(ubatch_slices: UBatchSlices, num_total_tokens: int):
2129
2130
2131
2132
2133
2134
        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
        )
2135

2136
2137
2138
2139
2140
2141
    def _pool(
        self,
        hidden_states: torch.Tensor,
        num_scheduled_tokens: int,
        num_scheduled_tokens_np: np.ndarray,
    ) -> ModelRunnerOutput:
2142
2143
2144
        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"
        )
2145

2146
        hidden_states = hidden_states[:num_scheduled_tokens]
2147
        pooling_metadata = self.input_batch.get_pooling_metadata()
2148
2149
2150
2151
        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]
2152

2153
2154
2155
2156
2157
2158
2159
2160
2161
2162
        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()
2163

2164
        pooler_output: list[torch.Tensor | None] = []
2165
        for raw_output, seq_len, prompt_len in zip(
2166
2167
            raw_pooler_output, seq_lens_cpu, pooling_metadata.prompt_lens
        ):
2168
            output = raw_output if seq_len == prompt_len else None
2169
            pooler_output.append(output)
2170
2171
2172
2173
2174
2175
2176
2177
2178
2179

        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,
        )

2180
    def _get_num_input_tokens(self, num_scheduled_tokens: int) -> int:
2181
2182
2183
2184
2185
2186
        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]
        ):
2187
2188
2189
2190
2191
2192
2193
2194
            # 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
2195
2196
2197
2198
        if (
            self.compilation_config.pass_config.enable_sequence_parallelism
            and tp_size > 1
        ):
2199
2200
2201
            return round_up(num_scheduled_tokens, tp_size)
        return num_scheduled_tokens

2202
    def _preprocess(
2203
2204
        self,
        scheduler_output: "SchedulerOutput",
2205
        num_input_tokens: int,  # Padded
2206
        intermediate_tensors: IntermediateTensors | None = None,
2207
    ) -> tuple[
2208
2209
        torch.Tensor | None,
        torch.Tensor | None,
2210
        torch.Tensor,
2211
        IntermediateTensors | None,
2212
2213
        dict[str, Any],
    ]:
2214
        num_scheduled_tokens = scheduler_output.total_num_scheduled_tokens
2215
        is_first_rank = get_pp_group().is_first_rank
2216

2217
2218
        # _prepare_inputs may reorder the batch, so we must gather multi
        # modal outputs after that to ensure the correct order
2219
2220
        if (
            self.supports_mm_inputs
2221
            and is_first_rank
2222
2223
            and not self.model_config.is_encoder_decoder
        ):
2224
2225
            # Run the multimodal encoder if any.
            self._execute_mm_encoder(scheduler_output)
2226
            mm_embeds, is_mm_embed = self._gather_mm_embeddings(scheduler_output)
2227

2228
2229
2230
            # 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.
2231
            inputs_embeds_scheduled = self.model.get_input_embeddings(
2232
2233
2234
                self.input_ids.gpu[:num_scheduled_tokens],
                multimodal_embeddings=mm_embeds,
                is_multimodal=is_mm_embed,
2235
            )
2236

2237
            # TODO(woosuk): Avoid the copy. Optimize.
2238
            self.inputs_embeds.gpu[:num_scheduled_tokens].copy_(inputs_embeds_scheduled)
2239

2240
            input_ids = None
2241
            inputs_embeds = self.inputs_embeds.gpu[:num_input_tokens]
2242
2243
2244
2245
            model_kwargs = {
                **self._init_model_kwargs(num_scheduled_tokens),
                **self._extract_mm_kwargs(scheduler_output),
            }
2246
        elif self.enable_prompt_embeds and is_first_rank:
2247
2248
2249
2250
2251
2252
2253
2254
2255
2256
2257
2258
            # 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).
2259
2260
2261
            token_ids_idx = (
                self.is_token_ids.gpu[:num_scheduled_tokens]
                .nonzero(as_tuple=False)
2262
                .squeeze(1)
2263
            )
2264
2265
2266
            # Some tokens ids may need to become embeds
            if token_ids_idx.numel() > 0:
                token_ids = self.input_ids.gpu[token_ids_idx]
2267
                tokens_to_embeds = self.model.get_input_embeddings(input_ids=token_ids)
2268
2269
2270
2271
2272
                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
2273
        else:
2274
2275
2276
2277
            # 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.
2278
            input_ids = self.input_ids.gpu[:num_input_tokens]
2279
            inputs_embeds = None
2280
            model_kwargs = self._init_model_kwargs(num_input_tokens)
2281
        if self.uses_mrope:
2282
            positions = self.mrope_positions.gpu[:, :num_input_tokens]
2283
        else:
2284
            positions = self.positions.gpu[:num_input_tokens]
2285

2286
        if is_first_rank:
2287
2288
            intermediate_tensors = None
        else:
2289
            intermediate_tensors = self.sync_and_slice_intermediate_tensors(
2290
2291
                num_input_tokens, intermediate_tensors, True
            )
2292

2293
2294
2295
2296
        if (
            self.model_config.is_encoder_decoder
            and scheduler_output.scheduled_encoder_inputs
        ):
2297
2298
2299
            encoder_inputs = self._extract_encoder_inputs(scheduler_output)
            model_kwargs.update(encoder_inputs)

2300
2301
2302
2303
2304
2305
2306
        return (
            input_ids,
            inputs_embeds,
            positions,
            intermediate_tensors,
            model_kwargs,
        )
2307

2308
    def _sample(
2309
        self,
2310
2311
        logits: torch.Tensor | None,
        spec_decode_metadata: SpecDecodeMetadata | None,
2312
    ) -> SamplerOutput:
2313
        # Sample the next token and get logprobs if needed.
2314
        sampling_metadata = self.input_batch.sampling_metadata
2315
        if spec_decode_metadata is None:
2316
2317
2318
            # 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()
2319
            return self.sampler(
2320
2321
2322
                logits=logits,
                sampling_metadata=sampling_metadata,
            )
2323

2324
        sampler_output = self.rejection_sampler(
2325
2326
            spec_decode_metadata,
            None,  # draft_probs
2327
            logits,
2328
2329
            sampling_metadata,
        )
2330
        self._update_states_after_model_execute(sampler_output.sampled_token_ids)
2331
2332
2333
        return sampler_output

    def _bookkeeping_sync(
2334
2335
2336
        self,
        scheduler_output: "SchedulerOutput",
        sampler_output: SamplerOutput,
2337
        logits: torch.Tensor | None,
2338
2339
        hidden_states: torch.Tensor,
        num_scheduled_tokens: int,
2340
        spec_decode_metadata: SpecDecodeMetadata | None,
2341
    ) -> tuple[
2342
        dict[str, int],
2343
        LogprobsLists | None,
2344
        list[list[int]],
2345
        dict[str, LogprobsTensors | None],
2346
2347
2348
        list[str],
        dict[str, int],
        list[int],
2349
    ]:
2350
2351
2352
2353
        num_nans_in_logits = {}
        if envs.VLLM_COMPUTE_NANS_IN_LOGITS:
            num_nans_in_logits = self._get_nans_in_logits(logits)

2354
2355
2356
        discard_sampled_tokens_req_indices = self.discard_request_indices.np[
            : self.num_discarded_requests
        ]
2357
2358
2359
2360
        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)
2361

2362
2363
2364
        # 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()
2365
        req_id_to_index_output_copy = self.input_batch.req_id_to_index.copy()
2366
2367

        num_sampled_tokens = sampler_output.sampled_token_ids.shape[0]
2368
        sampled_token_ids = sampler_output.sampled_token_ids
2369
        invalid_req_indices = []
2370
2371
2372
2373
2374
2375
2376
2377
2378
2379
2380
2381
2382
2383
        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:
2384
                valid_sampled_token_ids[int(i)].clear()
2385
        else:
2386
            valid_sampled_token_ids = []
2387
            invalid_req_indices = discard_sampled_tokens_req_indices.tolist()
2388
2389
2390
2391
2392
2393
            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.
2394
            self.input_batch.prev_sampled_token_ids = sampled_token_ids
2395
2396
2397
2398
2399
            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
            }
2400

2401
2402
2403
2404
2405
        # 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.
2406
        req_ids = self.input_batch.req_ids
2407
2408
2409
2410
        logprobs_tensors = sampler_output.logprobs_tensors
        cu_num_accepted_tokens = (
            [0] if spec_decode_metadata and logprobs_tensors else None
        )
2411
2412
        for req_idx in range(num_sampled_tokens):
            if self.use_async_scheduling:
2413
                sampled_ids = [-1] if req_idx not in invalid_req_indices_set else None
2414
2415
            else:
                sampled_ids = valid_sampled_token_ids[req_idx]
2416
2417
2418
2419
2420
2421
2422
2423

            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
                )

2424
2425
2426
2427
            if not sampled_ids:
                continue

            start_idx = self.input_batch.num_tokens_no_spec[req_idx]
2428
            end_idx = start_idx + num_sampled_ids
2429
2430
2431
2432
            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}"
2433
            )
2434

2435
2436
            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
2437
2438
            self.input_batch.num_tokens_no_spec[req_idx] = end_idx
            self.input_batch.num_tokens[req_idx] = end_idx
2439

2440
            req_id = req_ids[req_idx]
2441
2442
2443
            req_state = self.requests[req_id]
            req_state.output_token_ids.extend(sampled_ids)

2444
2445
        logprobs_lists = (
            logprobs_tensors.tolists(cu_num_accepted_tokens)
2446
            if not self.use_async_scheduling and logprobs_tensors is not None
2447
2448
2449
2450
2451
2452
2453
2454
2455
            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,
        )

2456
2457
2458
2459
2460
2461
2462
2463
2464
2465
        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,
        )

2466
2467
2468
2469
2470
2471
2472
2473
2474
2475
2476
2477
2478
2479
2480
    @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()

2481
2482
    def _model_forward(
        self,
2483
2484
2485
2486
        input_ids: torch.Tensor | None = None,
        positions: torch.Tensor | None = None,
        intermediate_tensors: IntermediateTensors | None = None,
        inputs_embeds: torch.Tensor | None = None,
2487
2488
2489
2490
2491
        **model_kwargs: dict[str, Any],
    ) -> Any:
        """Helper method to call the model forward pass.

        This method can be overridden by subclasses for model execution.
2492
        Motivation: We can inspect only this method versus
2493
2494
2495
2496
2497
2498
2499
2500
2501
2502
2503
2504
2505
2506
2507
2508
2509
2510
2511
2512
        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,
        )

2513
2514
2515
2516
    @torch.inference_mode()
    def execute_model(
        self,
        scheduler_output: "SchedulerOutput",
2517
        intermediate_tensors: IntermediateTensors | None = None,
2518
2519
2520
2521
2522
2523
2524
    ) -> 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
2525
        with record_function_or_nullcontext("Preprocess"):
2526
2527
2528
2529
            with self.synchronize_input_prep():
                # Update persistent batch states.
                self._update_states(scheduler_output)

2530
                if not num_scheduled_tokens:
2531
2532
2533
2534
                    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(
2535
2536
                        scheduler_output, self.vllm_config
                    )
2537
2538
2539
2540
                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 "
2541
2542
                        "it when the requests need prompt logprobs"
                    )
2543

2544
2545
2546
2547
2548
2549
                num_reqs = self.input_batch.num_reqs
                req_ids = self.input_batch.req_ids
                tokens = [scheduler_output.num_scheduled_tokens[i] for i in req_ids]
                num_scheduled_tokens_np = np.array(tokens, dtype=np.int32)
                max_num_scheduled_tokens = int(num_scheduled_tokens_np.max())

2550
2551
2552
2553
                (
                    logits_indices,
                    spec_decode_metadata,
                    ubatch_slices,
2554
                    num_tokens_across_dp,
2555
2556
2557
2558
2559
2560
2561
2562
2563
2564
2565
2566
2567
2568
2569
2570
2571
2572
2573
2574
2575
2576
2577
2578
2579
2580
2581
2582
2583
2584
2585
                ) = self._prepare_inputs(
                    scheduler_output, num_scheduled_tokens_np, max_num_scheduled_tokens
                )

                cascade_attn_prefix_lens = None
                # Disable cascade attention when using microbatching (DBO)
                if self.cascade_attn_enabled and ubatch_slices is None:
                    # Pre-compute cascade attention prefix lengths
                    # NOTE: Must be AFTER _prepare_inputs uses self.input_batch state
                    cascade_attn_prefix_lens = self._compute_cascade_attn_prefix_lens(
                        num_scheduled_tokens_np,
                        scheduler_output.num_common_prefix_blocks,
                    )

                # TODO(lucas): move cudagraph dispatching here:
                #   https://github.com/vllm-project/vllm/issues/23789

                total_num_scheduled_tokens = scheduler_output.total_num_scheduled_tokens
                use_spec_decode = len(scheduler_output.scheduled_spec_decode_tokens) > 0
                attn_metadata, spec_decode_common_attn_metadata = (
                    self._build_attention_metadata(
                        total_num_scheduled_tokens=total_num_scheduled_tokens,
                        max_num_scheduled_tokens=max_num_scheduled_tokens,
                        num_reqs=num_reqs,
                        ubatch_slices=ubatch_slices,
                        logits_indices=logits_indices,
                        use_spec_decode=use_spec_decode,
                        scheduled_encoder_inputs=scheduler_output.scheduled_encoder_inputs,
                        cascade_attn_prefix_lens=cascade_attn_prefix_lens,
                    )
                )
2586

2587
            dp_rank = self.parallel_config.data_parallel_rank
2588
2589
            if ubatch_slices:
                assert num_tokens_across_dp is not None
2590
                num_input_tokens = int(num_tokens_across_dp[dp_rank].item())
2591
2592
                self.pad_out_ubatch_slice(ubatch_slices, num_input_tokens)
            elif num_tokens_across_dp is not None:
2593
                num_input_tokens = int(num_tokens_across_dp[dp_rank].item())
2594
2595
2596
2597
2598
            else:
                num_input_tokens = self._get_num_input_tokens(
                    scheduler_output.total_num_scheduled_tokens
                )

2599
2600
2601
2602
2603
2604
            (
                input_ids,
                inputs_embeds,
                positions,
                intermediate_tensors,
                model_kwargs,
2605
            ) = self._preprocess(
2606
                scheduler_output, num_input_tokens, intermediate_tensors
2607
2608
            )

2609
2610
2611
            uniform_decode = (
                max_num_scheduled_tokens == self.uniform_decode_query_len
            ) and (num_scheduled_tokens == num_reqs * max_num_scheduled_tokens)
2612
            batch_descriptor = BatchDescriptor(
2613
2614
2615
                num_tokens=num_input_tokens,
                uniform_decode=uniform_decode,
                has_lora=len(self.input_batch.lora_id_to_lora_request) > 0,
2616
2617
            )
            cudagraph_runtime_mode, batch_descriptor = (
2618
2619
2620
2621
                self.cudagraph_dispatcher.dispatch(
                    batch_descriptor,
                    use_cascade_attn=cascade_attn_prefix_lens is not None,
                )
2622
            )
2623

2624
2625
        # Set cudagraph mode to none if calc_kv_scales is true.
        if attn_metadata is not None:
2626
2627
2628
2629
2630
            metadata_list = (
                attn_metadata.values()
                if isinstance(attn_metadata, dict)
                else [attn_metadata]
            )
2631
            if any(
2632
2633
                getattr(m, "enable_kv_scales_calculation", False) for m in metadata_list
            ):
2634
2635
                cudagraph_runtime_mode = CUDAGraphMode.NONE

2636
2637
        # Run the model.
        # Use persistent buffers for CUDA graphs.
2638
2639
        with (
            set_forward_context(
2640
2641
2642
2643
2644
2645
                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,
2646
                ubatch_slices=ubatch_slices,
2647
2648
2649
2650
            ),
            record_function_or_nullcontext("Forward"),
            self.maybe_get_kv_connector_output(scheduler_output) as kv_connector_output,
        ):
2651
            model_output = self._model_forward(
2652
2653
2654
2655
2656
2657
2658
2659
2660
                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:
2661
                # True when EAGLE 3 is used.
2662
2663
                hidden_states, aux_hidden_states = model_output
            else:
2664
                # Common case.
2665
2666
2667
                hidden_states = model_output
                aux_hidden_states = None

2668
2669
2670
2671
2672
            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)
2673
2674
                    hidden_states.kv_connector_output = kv_connector_output
                    return hidden_states
2675

2676
                if self.is_pooling_model:
2677
                    # Return the pooling output.
2678
2679
2680
                    output = self._pool(
                        hidden_states, num_scheduled_tokens, num_scheduled_tokens_np
                    )
2681
2682
                    output.kv_connector_output = kv_connector_output
                    return output
2683
2684

                sample_hidden_states = hidden_states[logits_indices]
2685
                logits = self.model.compute_logits(sample_hidden_states)
2686
2687
2688
2689
            else:
                # Rare case.
                assert not self.is_pooling_model

2690
                sample_hidden_states = hidden_states[logits_indices]
2691
                if not get_pp_group().is_last_rank:
2692
                    all_gather_tensors = {
2693
2694
2695
                        "residual": not is_residual_scattered_for_sp(
                            self.vllm_config, num_input_tokens
                        )
2696
                    }
2697
                    get_pp_group().send_tensor_dict(
2698
2699
                        hidden_states.tensors,
                        all_gather_group=get_tp_group(),
2700
2701
                        all_gather_tensors=all_gather_tensors,
                    )
2702
2703
                    logits = None
                else:
2704
                    logits = self.model.compute_logits(sample_hidden_states)
2705
2706
2707
2708
2709

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

2710
2711
2712
                model_output_broadcast_data = get_pp_group().broadcast_tensor_dict(
                    model_output_broadcast_data, src=len(get_pp_group().ranks) - 1
                )
2713
2714
2715
                assert model_output_broadcast_data is not None
                logits = model_output_broadcast_data["logits"]

2716
2717
2718
2719
2720
2721
2722
2723
2724
2725
2726
2727
2728
2729
2730
2731
2732
2733
2734
2735
2736
2737
2738
2739
2740
2741
2742
2743
2744
2745
2746
2747
2748
2749
2750
2751
2752
2753
2754
        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
            )
2755
2756
2757
2758

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

2759
2760
2761
2762
2763
2764
2765
2766
2767
2768
2769
2770
2771
2772
        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,
                )

2773
2774
2775
2776
2777
        use_padded_batch_for_eagle = (
            self.speculative_config
            and self.speculative_config.use_eagle()
            and not self.speculative_config.disable_padded_drafter_batch
        )
2778
2779
2780
        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
2781
2782
2783
2784
2785
        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
        ):
2786
            effective_drafter_max_model_len = (
2787
2788
                self.speculative_config.draft_model_config.max_model_len
            )
2789
        input_fits_in_drafter = spec_decode_common_attn_metadata and (
2790
2791
2792
2793
            spec_decode_common_attn_metadata.max_seq_len
            + self.speculative_config.num_speculative_tokens
            <= effective_drafter_max_model_len
        )
2794
        if use_padded_batch_for_eagle and input_fits_in_drafter:
2795
2796
2797
2798
            # 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)

2799
2800
2801
2802
2803
2804
2805
2806
2807
        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,
2808
2809
2810
2811
2812
            ) = self._bookkeeping_sync(
                scheduler_output,
                sampler_output,
                logits,
                hidden_states,
2813
                scheduler_output.total_num_scheduled_tokens,
2814
                spec_decode_metadata,
2815
            )
2816

2817
2818
2819
2820
2821
        if (
            self.speculative_config
            and not use_padded_batch_for_eagle
            and input_fits_in_drafter
        ):
2822
2823
2824
            # 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)
2825

2826
2827
        with record_function_or_nullcontext("EPLB"):
            self.eplb_step()
2828

2829
2830
2831
        output = ModelRunnerOutput(
            req_ids=req_ids_output_copy,
            req_id_to_index=req_id_to_index_output_copy,
2832
2833
2834
2835
            sampled_token_ids=valid_sampled_token_ids,
            logprobs=logprobs_lists,
            prompt_logprobs_dict=prompt_logprobs_dict,
            pooler_output=[],
2836
            kv_connector_output=kv_connector_output,
2837
2838
2839
            num_nans_in_logits=num_nans_in_logits,
        )

2840
2841
2842
        if not self.use_async_scheduling:
            return output

2843
        async_output = AsyncGPUModelRunnerOutput(
2844
            model_runner_output=output,
2845
            sampled_token_ids=sampler_output.sampled_token_ids,
2846
            logprobs_tensors=sampler_output.logprobs_tensors,
2847
2848
2849
2850
            invalid_req_indices=invalid_req_indices,
            async_output_copy_stream=self.async_output_copy_stream,
        )

2851
2852
2853
2854
2855
2856
2857
2858
2859
        # 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

2860
    def take_draft_token_ids(self) -> DraftTokenIds | None:
2861
2862
2863
2864
2865
2866
2867
2868
2869
2870
        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)

2871
2872
2873
    def propose_draft_token_ids(
        self,
        scheduler_output: "SchedulerOutput",
2874
        sampled_token_ids: torch.Tensor | list[list[int]],
2875
2876
2877
        sampling_metadata: SamplingMetadata,
        hidden_states: torch.Tensor,
        sample_hidden_states: torch.Tensor,
2878
2879
        aux_hidden_states: list[torch.Tensor] | None,
        spec_decode_metadata: SpecDecodeMetadata | None,
2880
        common_attn_metadata: CommonAttentionMetadata,
2881
    ) -> list[list[int]] | torch.Tensor:
2882
2883
        num_scheduled_tokens = scheduler_output.total_num_scheduled_tokens
        if self.speculative_config.method == "ngram":
2884
            assert isinstance(sampled_token_ids, list)
2885
            assert isinstance(self.drafter, NgramProposer)
2886
            draft_token_ids = self.drafter.propose(
2887
2888
                sampled_token_ids,
                self.input_batch.req_ids,
2889
2890
                self.input_batch.num_tokens_no_spec,
                self.input_batch.token_ids_cpu,
2891
2892
                self.input_batch.spec_decode_unsupported_reqs,
            )
2893
2894
2895
2896
        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)
2897
        elif self.speculative_config.method == "medusa":
2898
            assert isinstance(sampled_token_ids, list)
2899
            assert isinstance(self.drafter, MedusaProposer)
2900

2901
2902
            if sample_hidden_states.shape[0] == len(sampled_token_ids):
                # The input to the target model does not include draft tokens.
2903
2904
2905
2906
                hidden_states = sample_hidden_states
            else:
                indices = []
                offset = 0
2907
2908
2909
                assert spec_decode_metadata is not None, (
                    "No spec decode metadata for medusa"
                )
2910
                for num_draft, tokens in zip(
2911
2912
                    spec_decode_metadata.num_draft_tokens, sampled_token_ids
                ):
2913
2914
                    indices.append(offset + len(tokens) - 1)
                    offset += num_draft + 1
2915
                indices = torch.tensor(indices, device=self.device)
2916
2917
                hidden_states = sample_hidden_states[indices]

2918
            draft_token_ids = self.drafter.propose(
2919
2920
2921
                target_hidden_states=hidden_states,
                sampling_metadata=sampling_metadata,
            )
2922
        elif self.speculative_config.use_eagle():
2923
            assert isinstance(self.drafter, EagleProposer)
2924
2925
2926
2927
2928

            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.
2929
2930
                assert isinstance(sampled_token_ids, list), (
                    "sampled_token_ids should be a python list when"
2931
                    "padded-batch is disabled."
2932
                )
2933
                next_token_ids = self.drafter.prepare_next_token_ids_cpu(
2934
2935
2936
2937
2938
                    sampled_token_ids,
                    self.requests,
                    self.input_batch,
                    scheduler_output.num_scheduled_tokens,
                )
2939
2940
2941
2942
2943
            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.
2944
2945
                assert isinstance(sampled_token_ids, torch.Tensor), (
                    "sampled_token_ids should be a torch.Tensor when"
2946
                    "padded-batch is enabled."
2947
2948
                )
                next_token_ids, valid_sampled_tokens_count = (
2949
2950
2951
2952
2953
2954
                    self.drafter.prepare_next_token_ids_padded(
                        common_attn_metadata,
                        sampled_token_ids,
                        self.requests,
                        self.input_batch,
                        self.discard_request_indices.gpu,
2955
                        self.num_discarded_requests,
2956
                    )
2957
                )
Jiayi Yao's avatar
Jiayi Yao committed
2958

2959
            if spec_decode_metadata is None:
2960
                token_indices_to_sample = None
2961
                # input_ids can be None for multimodal models.
2962
                target_token_ids = self.input_ids.gpu[:num_scheduled_tokens]
2963
                target_positions = self._get_positions(num_scheduled_tokens)
2964
                if self.use_aux_hidden_state_outputs:
Wentao Ye's avatar
Wentao Ye committed
2965
                    assert aux_hidden_states is not None
2966
                    target_hidden_states = torch.cat(
2967
2968
                        [h[:num_scheduled_tokens] for h in aux_hidden_states], dim=-1
                    )
2969
2970
                else:
                    target_hidden_states = hidden_states[:num_scheduled_tokens]
2971
            else:
2972
2973
                if self.speculative_config.disable_padded_drafter_batch:
                    token_indices_to_sample = None
2974
2975
2976
2977
2978
                    common_attn_metadata, token_indices = self.drafter.prepare_inputs(
                        common_attn_metadata,
                        sampled_token_ids,
                        spec_decode_metadata.num_draft_tokens,
                    )
2979
                else:
2980
                    common_attn_metadata, token_indices, token_indices_to_sample = (
2981
2982
2983
                        self.drafter.prepare_inputs_padded(
                            common_attn_metadata,
                            spec_decode_metadata,
2984
2985
2986
                            valid_sampled_tokens_count,
                        )
                    )
2987

2988
                target_token_ids = self.input_ids.gpu[token_indices]
2989
                target_positions = self._get_positions(token_indices)
2990
                if self.use_aux_hidden_state_outputs:
Wentao Ye's avatar
Wentao Ye committed
2991
                    assert aux_hidden_states is not None
2992
                    target_hidden_states = torch.cat(
2993
2994
                        [h[token_indices] for h in aux_hidden_states], dim=-1
                    )
2995
2996
                else:
                    target_hidden_states = hidden_states[token_indices]
2997

2998
            if self.supports_mm_inputs:
2999
3000
3001
3002
3003
3004
                mm_embed_inputs = self._gather_mm_embeddings(
                    scheduler_output,
                    shift_computed_tokens=1,
                )
            else:
                mm_embed_inputs = None
3005

3006
            draft_token_ids = self.drafter.propose(
3007
3008
3009
3010
                target_token_ids=target_token_ids,
                target_positions=target_positions,
                target_hidden_states=target_hidden_states,
                next_token_ids=next_token_ids,
3011
                last_token_indices=token_indices_to_sample,
3012
                sampling_metadata=sampling_metadata,
3013
                common_attn_metadata=common_attn_metadata,
3014
                mm_embed_inputs=mm_embed_inputs,
3015
            )
3016

3017
        return draft_token_ids
3018

3019
3020
3021
    def update_config(self, overrides: dict[str, Any]) -> None:
        allowed_config_names = {"load_config", "model_config"}
        for config_name, config_overrides in overrides.items():
3022
3023
            assert config_name in allowed_config_names, (
                f"Config `{config_name}` not supported. "
3024
                f"Allowed configs: {allowed_config_names}"
3025
            )
3026
3027
3028
3029
            config = getattr(self, config_name)
            new_config = update_config(config, config_overrides)
            setattr(self, config_name, new_config)

3030
3031
3032
3033
3034
    def load_model(self, eep_scale_up: bool = False) -> None:
        """
        Args:
            eep_scale_up: the model loading is for elastic EP scale up.
        """
3035
3036
3037
3038
3039
        logger.info_once(
            "Starting to load model %s...",
            self.model_config.model,
            scope="global",
        )
3040
3041
3042
3043
3044
        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)
        )
3045

3046
3047
3048
        if self.parallel_config.enable_eplb:
            self.eplb_state = EplbState(self.parallel_config, self.device)
            eplb_models = 0
3049
        with DeviceMemoryProfiler() as m:
3050
            time_before_load = time.perf_counter()
3051
            model_loader = get_model_loader(self.load_config)
3052
            self.model = model_loader.load_model(
3053
3054
                vllm_config=self.vllm_config, model_config=self.model_config
            )
3055
            if self.lora_config:
3056
3057
3058
                self.model = self.load_lora_model(
                    self.model, self.vllm_config, self.device
                )
3059
            if hasattr(self, "drafter"):
3060
                logger.info_once("Loading drafter model...")
3061
                self.drafter.load_model(self.model)
3062
3063
3064
3065
3066
3067
3068
3069
3070
3071
3072
3073
3074
3075
3076
3077
3078
3079
3080
3081
3082
3083
3084
3085
3086
3087
3088
3089
3090
3091
3092
                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

3093
            if self.use_aux_hidden_state_outputs:
3094
                if not supports_eagle3(self.get_model()):
3095
3096
                    raise RuntimeError(
                        "Model does not support EAGLE3 interface but "
3097
3098
                        "aux_hidden_state_outputs was requested"
                    )
3099
3100
3101
3102
3103
3104
3105
3106
3107
3108
3109
3110
3111

                # 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)
3112
            time_after_load = time.perf_counter()
3113
        self.model_memory_usage = m.consumed_memory
3114
        logger.info_once(
3115
            "Model loading took %.4f GiB memory and %.6f seconds",
3116
3117
            self.model_memory_usage / GiB_bytes,
            time_after_load - time_before_load,
3118
            scope="local",
3119
        )
3120
        prepare_communication_buffer_for_model(self.model)
3121
        self.is_multimodal_pruning_enabled = (
3122
            supports_multimodal_pruning(self.get_model())
3123
3124
            and self.model_config.multimodal_config.is_multimodal_pruning_enabled()
        )
3125

3126
        if is_mixture_of_experts(self.model) and self.parallel_config.enable_eplb:
3127
3128
3129
3130
3131
3132
3133
3134
3135
3136
3137
            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(
3138
                self.model,
3139
                self.model_config,
3140
3141
3142
                global_expert_load,
                old_global_expert_indices,
                rank_mapping,
3143
3144
            )

3145
        if (
3146
3147
            self.vllm_config.compilation_config.mode
            == CompilationMode.STOCK_TORCH_COMPILE
3148
            and supports_dynamo()
3149
        ):
3150
            backend = self.vllm_config.compilation_config.init_backend(self.vllm_config)
3151
            compilation_counter.stock_torch_compile_count += 1
3152
            self.model.compile(fullgraph=True, backend=backend)
3153
            return
3154
        # for other compilation modes, cudagraph behavior is controlled by
3155
3156
3157
        # CudagraphWraper and CudagraphDispatcher of vllm.

        # wrap the model with full cudagraph wrapper if needed.
3158
3159
3160
3161
3162
3163
3164
        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
            )
3165
3166
        elif self.parallel_config.enable_dbo:
            if self.compilation_config.cudagraph_mode.has_full_cudagraphs():
3167
3168
3169
                self.model = UBatchWrapper(
                    self.model, self.vllm_config, CUDAGraphMode.FULL, self.device
                )
3170
            else:
3171
3172
3173
                self.model = UBatchWrapper(
                    self.model, self.vllm_config, CUDAGraphMode.NONE, self.device
                )
3174

3175
    def _get_eagle3_aux_layers_from_config(self) -> tuple[int, ...] | None:
3176
3177
3178
3179
3180
3181
3182
3183
3184
3185
3186
3187
3188
3189
3190
3191
3192
3193
3194
3195
3196
3197
3198
        """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

3199
    def reload_weights(self) -> None:
3200
        assert getattr(self, "model", None) is not None, (
3201
            "Cannot reload weights before model is loaded."
3202
        )
3203
3204
        model_loader = get_model_loader(self.load_config)
        logger.info("Reloading weights inplace...")
3205
        model_loader.load_weights(self.get_model(), model_config=self.model_config)
3206

3207
3208
3209
3210
3211
    def save_tensorized_model(
        self,
        tensorizer_config: "TensorizerConfig",
    ) -> None:
        TensorizerLoader.save_model(
3212
            self.get_model(),
3213
            tensorizer_config=tensorizer_config,
3214
            model_config=self.model_config,
3215
3216
        )

3217
3218
3219
    def _get_prompt_logprobs_dict(
        self,
        hidden_states: torch.Tensor,
3220
        num_scheduled_tokens: dict[str, int],
3221
    ) -> dict[str, LogprobsTensors | None]:
3222
3223
3224
3225
        num_prompt_logprobs_dict = self.input_batch.num_prompt_logprobs
        if not num_prompt_logprobs_dict:
            return {}

3226
        in_progress_dict = self.input_batch.in_progress_prompt_logprobs_cpu
3227
        prompt_logprobs_dict: dict[str, LogprobsTensors | None] = {}
3228
3229
3230
3231
3232

        # 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():
3233
            num_tokens = num_scheduled_tokens[req_id]
3234
3235
3236

            # Get metadata for this request.
            request = self.requests[req_id]
3237
3238
3239
3240
            if request.prompt_token_ids is None:
                # Prompt logprobs is incompatible with prompt embeddings
                continue

3241
3242
            num_prompt_tokens = len(request.prompt_token_ids)
            prompt_token_ids = torch.tensor(request.prompt_token_ids).to(
3243
3244
                self.device, non_blocking=True
            )
3245

3246
3247
3248
3249
3250
3251
            # 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(
3252
3253
                    num_prompt_tokens - 1, num_prompt_logprobs + 1
                )
3254
3255
                in_progress_dict[req_id] = logprobs_tensors

3256
            # Determine number of logits to retrieve.
3257
3258
            start_idx = request.num_computed_tokens
            start_tok = start_idx + 1
3259
            num_remaining_tokens = num_prompt_tokens - start_tok
3260
            if num_tokens <= num_remaining_tokens:
3261
                # This is a chunk, more tokens remain.
3262
3263
3264
                # 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.
3265
3266
3267
3268
3269
                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)
3270
3271
3272
3273
3274
3275
3276
                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
3277
3278
3279
3280
3281

            # 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]
3282
            offset = self.query_start_loc.np[req_idx].item()
3283
            prompt_hidden_states = hidden_states[offset : offset + num_logits]
3284
            logits = self.model.compute_logits(prompt_hidden_states)
3285
3286
3287
3288

            # 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.
3289
            tgt_token_ids = prompt_token_ids[start_tok : start_tok + num_logits]
3290
3291

            # Compute prompt logprobs.
3292
3293
            logprobs = self.sampler.compute_logprobs(logits)
            token_ids, logprobs, ranks = self.sampler.gather_logprobs(
3294
3295
                logprobs, num_prompt_logprobs, tgt_token_ids
            )
3296
3297

            # Transfer GPU->CPU async.
3298
3299
            chunk_slice = slice(start_idx, start_idx + num_logits)
            logprobs_tensors.logprob_token_ids[chunk_slice].copy_(
3300
3301
3302
                token_ids, non_blocking=True
            )
            logprobs_tensors.logprobs[chunk_slice].copy_(logprobs, non_blocking=True)
3303
            logprobs_tensors.selected_token_ranks[chunk_slice].copy_(
3304
3305
                ranks, non_blocking=True
            )
3306
3307
3308
3309
3310

        # 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]
3311
            del in_progress_dict[req_id]
3312
3313

        # Must synchronize the non-blocking GPU->CPU transfers.
3314
        if prompt_logprobs_dict:
3315
            self._sync_device()
3316
3317
3318

        return prompt_logprobs_dict

3319
3320
    def _get_nans_in_logits(
        self,
3321
        logits: torch.Tensor | None,
3322
3323
3324
3325
3326
3327
3328
3329
3330
3331
3332
    ) -> 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])
3333
3334
3335
                    if num_nans_for_index is not None and req_index < logits.shape[0]
                    else 0
                )
3336
3337
3338
3339
            return num_nans_in_logits
        except IndexError:
            return {}

3340
3341
3342
3343
3344
3345
    @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
3346
         - during DP rank dummy run
3347
3348
3349
3350
3351
3352
3353
3354
3355
3356
3357
        """
        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(
3358
                    self.input_ids.gpu,
3359
3360
                    low=0,
                    high=self.model_config.get_vocab_size(),
3361
3362
                    dtype=input_ids.dtype,
                )
3363

3364
            logger.debug_once("Randomizing dummy data for DP Rank")
3365
            input_ids.copy_(rand_input_ids()[: input_ids.size(0)], non_blocking=True)
3366
3367
3368
            yield
            input_ids.fill_(0)

3369
3370
3371
3372
3373
3374
    def _get_mm_dummy_batch(
        self,
        modality: str,
        max_items_per_batch: int,
    ) -> BatchedTensorInputs:
        """Dummy data for profiling and precompiling multimodal models."""
3375
3376
        assert self.mm_budget is not None

3377
3378
        dummy_decoder_data = self.mm_registry.get_decoder_dummy_data(
            model_config=self.model_config,
3379
            seq_len=self.max_model_len,
3380
            mm_counts={modality: 1},
3381
            cache=self.mm_budget.cache,
3382
3383
3384
3385
        )
        dummy_mm_data = dummy_decoder_data.multi_modal_data

        # Result in the maximum GPU consumption of the model
3386
3387
        dummy_mm_item = dummy_mm_data[modality][0]
        dummy_mm_items = [dummy_mm_item] * max_items_per_batch
3388

3389
        model = cast(SupportsMultiModal, self.model)
3390
3391
3392
3393
3394
3395
3396
        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,
3397
                multimodal_cpu_fields=model.multimodal_cpu_fields,
3398
3399
            )
        )
3400

3401
3402
3403
3404
    @torch.inference_mode()
    def _dummy_run(
        self,
        num_tokens: int,
3405
        cudagraph_runtime_mode: CUDAGraphMode | None = None,
3406
3407
        force_attention: bool = False,
        uniform_decode: bool = False,
3408
        allow_microbatching: bool = True,
3409
3410
        skip_eplb: bool = False,
        is_profile: bool = False,
3411
        create_mixed_batch: bool = False,
3412
        remove_lora: bool = True,
3413
        activate_lora: bool = False,
3414
    ) -> tuple[torch.Tensor, torch.Tensor]:
3415
3416
3417
3418
3419
3420
3421
        """
        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.
3422
                - if not set will determine the cudagraph mode based on using
3423
                    the self.cudagraph_dispatcher.
3424
3425
3426
3427
                - CUDAGraphMode.NONE: No cudagraph, for warm up and profile run
                - CUDAGraphMode.PIECEWISE: Piecewise cudagraph.
                - CUDAGraphMode.FULL: Full cudagraph, attention metadata is
                    needed.
3428
            force_attention: If True, always create attention metadata. Used to
3429
3430
3431
3432
                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.
3433
3434
            create_mixed_batch: If True, create a mixed batch with both decode
                (1 token) and prefill (multiple tokens) requests.
3435
            remove_lora: If False, dummy LoRAs are not destroyed after the run
3436
            activate_lora: If False, dummy_run is performed without LoRAs.
3437
        """
3438
3439
3440
3441
        assert (
            cudagraph_runtime_mode is None
            or cudagraph_runtime_mode.valid_runtime_modes()
        )
3442

3443
        # If cudagraph_mode.decode_mode() == FULL and
3444
        # cudagraph_mode.separate_routine(). This means that we are using
3445
3446
3447
3448
3449
3450
3451
3452
3453
3454
3455
        # 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.
3456
        max_query_len = self.uniform_decode_query_len if uniform_decode else num_tokens
3457

3458
3459
3460
3461
3462
        # 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
3463
3464
3465
3466
        if create_mixed_batch:
            assert not uniform_decode
            # Create mixed batch:
            # first half decode tokens, second half one prefill
3467
            num_decode_tokens = min(max_num_reqs - 1, num_tokens // 2)
3468
3469
3470
3471
            num_prefill_tokens = num_tokens - num_decode_tokens
            num_reqs = num_decode_tokens + 1

            # Create decode requests (1 token each) followed by prefill request
3472
            num_scheduled_tokens_list = [1] * num_decode_tokens + [num_prefill_tokens]
3473
3474
3475
            # Note: Overriding max_query_len to be the prefill tokens
            max_query_len = num_prefill_tokens
        elif uniform_decode:
3476
            assert not create_mixed_batch
3477
            num_reqs = min(max_num_reqs, cdiv(num_tokens, max_query_len))
3478
3479
            num_scheduled_tokens_list = [max_query_len] * num_reqs
            if num_tokens % max_query_len != 0:
3480
                num_scheduled_tokens_list[-1] = num_tokens % max_query_len
3481
3482
3483
3484
3485
3486
        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

3487
3488
        assert sum(num_scheduled_tokens_list) == num_tokens
        assert len(num_scheduled_tokens_list) == num_reqs
3489
        num_scheduled_tokens = np.array(num_scheduled_tokens_list, dtype=np.int32)
3490
        total_num_scheduled_tokens = int(num_scheduled_tokens.sum())
3491
        num_sampled_tokens = np.ones(num_reqs, dtype=np.int32)
3492

3493
3494
3495
        # Disable DP padding when running eager
        allow_dp_padding = self.compilation_config.cudagraph_mode != CUDAGraphMode.NONE

3496
3497
        # We currently only microbatch if the number of tokens is
        # over a certain threshold.
3498
        ubatch_slices, num_tokens_across_dp = coordinate_batch_across_dp(
3499
3500
3501
3502
3503
3504
3505
            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,
3506
3507
3508
        )
        num_tokens_after_padding = num_tokens
        if num_tokens_across_dp is not None:
3509
3510
            dp_rank = self.parallel_config.data_parallel_rank
            num_tokens_after_padding = int(num_tokens_across_dp[dp_rank])
3511

3512
        attn_metadata: PerLayerAttnMetadata | None = None
3513
3514
3515

        # If force_attention is True, we always capture attention. Otherwise,
        # it only happens for cudagraph_runtime_mode=FULL.
3516
        if force_attention or cudagraph_runtime_mode == CUDAGraphMode.FULL:
3517
3518
3519
3520
3521
3522
            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:
3523
                seq_lens = max_query_len
3524
            self.seq_lens.np[:num_reqs] = seq_lens
3525
3526
            self.seq_lens.np[num_reqs:] = 0
            self.seq_lens.copy_to_gpu()
3527

3528
3529
            cum_num_tokens, _ = self._get_cumsum_and_arange(num_scheduled_tokens)
            self.query_start_loc.np[1 : num_reqs + 1] = cum_num_tokens
3530
3531
            self.query_start_loc.copy_to_gpu()

3532
3533
3534
3535
3536
3537
3538
            attn_metadata, _ = self._build_attention_metadata(
                total_num_scheduled_tokens=num_tokens,
                max_num_scheduled_tokens=max_query_len,
                num_reqs=num_reqs,
                ubatch_slices=ubatch_slices,
                for_cudagraph_capture=True,
            )
3539

3540
        with self.maybe_dummy_run_with_lora(
3541
3542
3543
3544
3545
            self.lora_config,
            num_scheduled_tokens,
            num_sampled_tokens,
            activate_lora,
            remove_lora,
3546
        ):
3547
3548
3549
            # 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)
3550
            if self.supports_mm_inputs and not self.model_config.is_encoder_decoder:
3551
                input_ids = None
3552
                inputs_embeds = self.inputs_embeds.gpu[:num_tokens_after_padding]
3553
                model_kwargs = {
3554
                    **model_kwargs,
3555
3556
                    **self._dummy_mm_kwargs(num_reqs),
                }
3557
3558
            elif self.enable_prompt_embeds:
                input_ids = None
3559
3560
                inputs_embeds = self.inputs_embeds.gpu[:num_tokens_after_padding]
                model_kwargs = self._init_model_kwargs(num_tokens_after_padding)
3561
            else:
3562
                input_ids = self.input_ids.gpu[:num_tokens_after_padding]
3563
                inputs_embeds = None
3564

3565
            if self.uses_mrope:
3566
                positions = self.mrope_positions.gpu[:, :num_tokens_after_padding]
3567
            else:
3568
                positions = self.positions.gpu[:num_tokens_after_padding]
3569
3570
3571
3572
3573
3574
3575
3576
3577

            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,
3578
3579
3580
                            device=self.device,
                        )
                    )
3581
3582

                intermediate_tensors = self.sync_and_slice_intermediate_tensors(
3583
                    num_tokens_after_padding, None, False
3584
                )
3585
3586

            # filter out the valid batch descriptor
3587
3588
3589
3590
3591
            _cg_mode, batch_descriptor = (
                self.cudagraph_dispatcher.dispatch(
                    BatchDescriptor(
                        num_tokens=num_tokens_after_padding,
                        uniform_decode=uniform_decode,
3592
                        has_lora=activate_lora and self.lora_config is not None,
3593
3594
3595
3596
3597
                    )
                )
                if not is_profile
                else (CUDAGraphMode.NONE, None)
            )
3598
3599
3600
            if cudagraph_runtime_mode is not None:
                # we allow forcing NONE when the dispatcher disagrees to support
                # warm ups for cudagraph capture
3601
3602
3603
3604
                assert (
                    cudagraph_runtime_mode == CUDAGraphMode.NONE
                    or cudagraph_runtime_mode == _cg_mode
                ), (
3605
                    f"Cudagraph runtime mode mismatch at dummy_run. "
3606
3607
                    f"Expected {_cg_mode}, but got {cudagraph_runtime_mode}."
                )
3608
3609
            else:
                cudagraph_runtime_mode = _cg_mode
3610

3611
            if ubatch_slices is not None:
3612
3613
3614
3615
3616
3617
3618
                # 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

3619
3620
3621
            with (
                self.maybe_randomize_inputs(input_ids),
                set_forward_context(
3622
3623
                    attn_metadata,
                    self.vllm_config,
3624
                    num_tokens=num_tokens_after_padding,
3625
3626
                    num_tokens_across_dp=num_tokens_across_dp,
                    cudagraph_runtime_mode=cudagraph_runtime_mode,
3627
                    batch_descriptor=batch_descriptor,
3628
3629
3630
                    ubatch_slices=ubatch_slices,
                ),
            ):
3631
                outputs = self.model(
3632
3633
3634
3635
                    input_ids=input_ids,
                    positions=positions,
                    intermediate_tensors=intermediate_tensors,
                    inputs_embeds=inputs_embeds,
3636
                    **model_kwargs,
3637
                )
3638

3639
3640
3641
3642
            if self.use_aux_hidden_state_outputs:
                hidden_states, _ = outputs
            else:
                hidden_states = outputs
3643

3644
            if self.speculative_config and self.speculative_config.use_eagle():
3645
                assert isinstance(self.drafter, EagleProposer)
3646
3647
3648
3649
                use_cudagraphs = (
                    cudagraph_runtime_mode == CUDAGraphMode.PIECEWISE
                    and not self.speculative_config.enforce_eager
                )
3650
3651
3652
3653
3654
3655
3656
3657
3658
3659
3660
3661

                # Note(gnovack) - We need to disable cudagraphs for one of the two
                # lora cases when cudagraph_specialize_lora is enabled. This is a
                # short term mitigation for issue mentioned in
                # https://github.com/vllm-project/vllm/issues/28334
                if self.compilation_config.cudagraph_specialize_lora and activate_lora:
                    use_cudagraphs = False

                self.drafter.dummy_run(
                    num_tokens,
                    use_cudagraphs=use_cudagraphs,
                )
3662

3663
3664
3665
3666
3667
3668
3669
3670
3671
3672
        # 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)

3673
        logit_indices = np.cumsum(num_scheduled_tokens) - 1
3674
3675
3676
3677
        logit_indices_device = torch.from_numpy(logit_indices).to(
            self.device, non_blocking=True
        )
        return hidden_states, hidden_states[logit_indices_device]
3678
3679
3680
3681
3682
3683

    @torch.inference_mode()
    def _dummy_sampler_run(
        self,
        hidden_states: torch.Tensor,
    ) -> torch.Tensor:
3684
3685
3686
3687
        # 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)
3688

3689
        logits = self.model.compute_logits(hidden_states)
3690
3691
        num_reqs = logits.size(0)

3692
        dummy_tensors = lambda v: torch.full((num_reqs,), v, device=self.device)
3693
3694
3695
3696
3697
3698
3699
3700
3701
3702
3703
3704
3705
3706
3707

        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)],
3708
            spec_token_ids=[[] for _ in range(num_reqs)],
3709
3710
            allowed_token_ids_mask=None,
            bad_words_token_ids={},
3711
            logitsprocs=LogitsProcessors(),
3712
        )
3713
        try:
3714
3715
3716
            sampler_output = self.sampler(
                logits=logits, sampling_metadata=dummy_metadata
            )
3717
        except RuntimeError as e:
3718
            if "out of memory" in str(e):
3719
3720
3721
3722
                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 "
3723
3724
                    "initializing the engine."
                ) from e
3725
3726
            else:
                raise e
3727
        if self.speculative_config:
3728
3729
            draft_token_ids = [[0] for _ in range(num_reqs)]
            dummy_spec_decode_metadata = SpecDecodeMetadata.make_dummy(
3730
3731
                draft_token_ids, self.device
            )
3732
3733
3734
3735
3736
3737

            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
3738
3739
3740
3741
3742
            logits = torch.randn(
                num_tokens + num_reqs,
                logits.shape[-1],
                device=self.device,
                dtype=logits.dtype,
3743
            )
3744
3745
3746
            self.rejection_sampler(
                dummy_spec_decode_metadata,
                draft_probs,
3747
                logits,
3748
3749
                dummy_metadata,
            )
3750
        return sampler_output
3751

3752
    def _dummy_pooler_run_task(
3753
3754
        self,
        hidden_states: torch.Tensor,
3755
3756
        task: PoolingTask,
    ) -> PoolerOutput:
3757
3758
3759
3760
3761
3762
3763
3764
3765
3766
3767
        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

3768
        dummy_prompt_lens = torch.tensor(
3769
3770
            num_scheduled_tokens_list,
            device="cpu",
3771
        )
3772
3773
3774
        dummy_token_ids = torch.zeros(
            (num_reqs, req_num_tokens), dtype=torch.int32, device=self.device
        )
3775

3776
        model = cast(VllmModelForPooling, self.get_model())
3777
        dummy_pooling_params = PoolingParams(task=task)
3778
        dummy_pooling_params.verify(task=task, model_config=self.model_config)
3779
        to_update = model.pooler.get_pooling_updates(task)
3780
3781
        to_update.apply(dummy_pooling_params)

3782
        dummy_metadata = PoolingMetadata(
3783
3784
3785
3786
            prompt_lens=dummy_prompt_lens,
            prompt_token_ids=dummy_token_ids,
            pooling_params=[dummy_pooling_params] * num_reqs,
        )
3787

3788
3789
3790
        dummy_metadata.build_pooling_cursor(
            num_scheduled_tokens_list, device=hidden_states.device
        )
3791

3792
        try:
3793
3794
3795
            return model.pooler(
                hidden_states=hidden_states, pooling_metadata=dummy_metadata
            )
3796
        except RuntimeError as e:
3797
            if "out of memory" in str(e):
3798
                raise RuntimeError(
3799
3800
3801
                    "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 "
3802
3803
                    "initializing the engine."
                ) from e
3804
3805
            else:
                raise e
3806
3807
3808
3809
3810
3811
3812

    @torch.inference_mode()
    def _dummy_pooler_run(
        self,
        hidden_states: torch.Tensor,
    ) -> PoolerOutput:
        # Find the task that has the largest output for subsequent steps
3813
3814
3815
3816
3817
3818
3819
3820
3821
3822
3823
3824
3825
3826
3827
3828
3829
3830
3831
3832
        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."
                )

3833
        output_size = dict[PoolingTask, float]()
3834
        for task in supported_pooling_tasks:
3835
3836
            # Run a full batch with each task to ensure none of them OOMs
            output = self._dummy_pooler_run_task(hidden_states, task)
3837
            output_size[task] = sum(o.nbytes for o in output)
3838
3839
3840
3841
            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)
3842

3843
    def profile_run(self) -> None:
3844
        # Profile with multimodal encoder & encoder cache.
3845
        if self.supports_mm_inputs:
3846
            if self.model_config.multimodal_config.skip_mm_profiling:
3847
                logger.info(
3848
                    "Skipping memory profiling for multimodal encoder and "
3849
3850
                    "encoder cache."
                )
3851
3852
3853
3854
3855
3856
3857
3858
            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.
3859
                    dummy_modality = mm_budget.get_modality_with_max_tokens()
3860
3861
3862
                    max_mm_items_per_batch = mm_budget.max_items_per_batch_by_modality[
                        dummy_modality
                    ]
3863
3864
3865
3866
3867
3868
3869
3870
3871

                    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,
                    )
3872

3873
3874
3875
3876
3877
                    # Create dummy batch of multimodal inputs.
                    batched_dummy_mm_inputs = self._get_mm_dummy_batch(
                        dummy_modality,
                        max_mm_items_per_batch,
                    )
3878

3879
                    # Run multimodal encoder.
3880
3881
3882
                    dummy_encoder_outputs = self.model.get_multimodal_embeddings(
                        **batched_dummy_mm_inputs
                    )
3883

3884
3885
3886
3887
                    sanity_check_mm_encoder_outputs(
                        dummy_encoder_outputs,
                        expected_num_items=max_mm_items_per_batch,
                    )
3888

3889
3890
3891
3892
3893
3894
3895
3896
3897
3898
                    # 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(
3899
3900
                                (encoder_budget, encoder_output_shape[-1])
                            )
3901
3902
3903
3904
3905
3906
                            num_tokens = output.shape[0]
                            expanded[:num_tokens].copy_(output)
                            expanded_outputs.append(expanded)

                        dummy_encoder_outputs = expanded_outputs

3907
                    # Cache the dummy encoder outputs.
3908
                    self.encoder_cache["tmp"] = dict(enumerate(dummy_encoder_outputs))
3909

3910
        # Add `is_profile` here to pre-allocate communication buffers
3911
3912
3913
        hidden_states, last_hidden_states = self._dummy_run(
            self.max_num_tokens, is_profile=True
        )
3914
        if get_pp_group().is_last_rank:
3915
3916
3917
3918
            if self.is_pooling_model:
                output = self._dummy_pooler_run(hidden_states)
            else:
                output = self._dummy_sampler_run(last_hidden_states)
3919
        else:
3920
            output = None
3921
        self._sync_device()
3922
        del hidden_states, output
3923
        self.encoder_cache.clear()
3924
        gc.collect()
3925

3926
    def capture_model(self) -> int:
3927
        if self.compilation_config.cudagraph_mode == CUDAGraphMode.NONE:
3928
            logger.warning(
3929
                "Skipping CUDA graph capture. To turn on CUDA graph capture, "
3930
3931
                "ensure `cudagraph_mode` was not manually set to `NONE`"
            )
3932
            return 0
3933

3934
3935
        compilation_counter.num_gpu_runner_capture_triggers += 1

3936
3937
        start_time = time.perf_counter()

3938
3939
3940
3941
3942
3943
3944
3945
3946
3947
3948
3949
3950
3951
        @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()
3952
                    gc.collect()
3953

3954
3955
3956
        # 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.
3957
        set_cudagraph_capturing_enabled(True)
3958
        with freeze_gc(), graph_capture(device=self.device):
3959
            start_free_gpu_memory = torch.cuda.mem_get_info()[0]
3960
            cudagraph_mode = self.compilation_config.cudagraph_mode
3961
            assert cudagraph_mode is not None
3962
3963
3964
3965
3966
3967
3968
3969
3970

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

3971
3972
            if cudagraph_mode.mixed_mode() != CUDAGraphMode.NONE:
                cudagraph_runtime_mode = cudagraph_mode.mixed_mode()
3973
                # make sure we capture the largest batch size first
3974
3975
3976
                compilation_cases = list(
                    product(reversed(self.cudagraph_batch_sizes), lora_cases)
                )
3977
3978
3979
                self._capture_cudagraphs(
                    compilation_cases,
                    cudagraph_runtime_mode=cudagraph_runtime_mode,
3980
3981
                    uniform_decode=False,
                )
3982

3983
3984
            # Capture full cudagraph for uniform decode batches if we
            # don't already have full mixed prefill-decode cudagraphs.
3985
3986
3987
3988
3989
3990
3991
            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
                )
3992
                decode_cudagraph_batch_sizes = [
3993
3994
                    x
                    for x in self.cudagraph_batch_sizes
3995
                    if max_num_tokens >= x >= self.uniform_decode_query_len
3996
                ]
3997
3998
3999
                compilation_cases_decode = list(
                    product(reversed(decode_cudagraph_batch_sizes), lora_cases)
                )
4000
4001
4002
                self._capture_cudagraphs(
                    compilation_cases=compilation_cases_decode,
                    cudagraph_runtime_mode=CUDAGraphMode.FULL,
4003
4004
                    uniform_decode=True,
                )
4005

4006
4007
4008
            torch.cuda.synchronize()
            end_free_gpu_memory = torch.cuda.mem_get_info()[0]

4009
4010
4011
        # 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
4012
        # we may do lazy capturing in future that still allows capturing
4013
4014
        # after here.
        set_cudagraph_capturing_enabled(False)
4015
4016
4017
4018
4019

        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.
4020
        logger.info_once(
4021
4022
4023
            "Graph capturing finished in %.0f secs, took %.2f GiB",
            elapsed_time,
            cuda_graph_size / (1 << 30),
4024
            scope="local",
4025
        )
4026
        return cuda_graph_size
4027

4028
4029
    def _capture_cudagraphs(
        self,
4030
        compilation_cases: list[tuple[int, bool]],
4031
4032
4033
4034
4035
4036
4037
        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}"
4038
4039
4040
4041
4042
4043
4044
4045

        # 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",
4046
4047
4048
                    cudagraph_runtime_mode.name,
                ),
            )
4049

4050
        # We skip EPLB here since we don't want to record dummy metrics
4051
        for num_tokens, activate_lora in compilation_cases:
4052
            # We currently only capture ubatched graphs when its a FULL
4053
4054
4055
            # 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
4056
4057
4058
4059
            allow_microbatching = (
                self.parallel_config.enable_dbo
                and cudagraph_runtime_mode == CUDAGraphMode.FULL
                and uniform_decode
4060
4061
4062
4063
4064
                and check_ubatch_thresholds(
                    config=self.vllm_config.parallel_config,
                    num_tokens=num_tokens,
                    uniform_decode=uniform_decode,
                )
4065
            )
4066

4067
4068
4069
4070
4071
4072
            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.
4073
4074
4075
4076
4077
4078
4079
4080
4081
                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,
4082
                    activate_lora=activate_lora,
4083
4084
4085
4086
4087
4088
4089
4090
                )
            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,
4091
                activate_lora=activate_lora,
4092
            )
4093
        self.maybe_remove_all_loras(self.lora_config)
4094

4095
4096
4097
4098
    def initialize_attn_backend(self, kv_cache_config: KVCacheConfig) -> None:
        """
        Initialize the attention backends and attention metadata builders.
        """
4099
        assert len(self.attn_groups) == 0, "Attention backends are already initialized"
4100

4101
4102
4103
4104
4105
4106
        class AttentionGroupKey(NamedTuple):
            attn_backend: type[AttentionBackend]
            kv_cache_spec: KVCacheSpec

        def get_attn_backends_for_group(
            kv_cache_group_spec: KVCacheGroupSpec,
4107
        ) -> tuple[dict[AttentionGroupKey, list[str]], set[type[AttentionBackend]]]:
4108
            layers = get_layers_from_vllm_config(
4109
4110
                self.vllm_config, AttentionLayerBase, kv_cache_group_spec.layer_names
            )
4111
4112
            attn_backends = {}
            attn_backend_layers = defaultdict(list)
4113
            # Dedupe based on full class name; this is a bit safer than
4114
4115
4116
4117
            # 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.
4118
            for layer_name in kv_cache_group_spec.layer_names:
4119
                attn_backend = layers[layer_name].get_attn_backend()
4120
4121
4122
4123
4124
4125
4126

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

4127
4128
4129
                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):
4130
                    layer_kv_cache_spec = layer_kv_cache_spec.kv_cache_specs[layer_name]
4131
                key = (full_cls_name, layer_kv_cache_spec)
4132
4133
4134
                attn_backends[key] = AttentionGroupKey(
                    attn_backend, layer_kv_cache_spec
                )
4135
                attn_backend_layers[key].append(layer_name)
4136
4137
4138
4139
            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()),
            )
4140
4141

        def create_attn_groups(
4142
            attn_backends_map: dict[AttentionGroupKey, list[str]],
4143
            kv_cache_group_id: int,
4144
4145
        ) -> list[AttentionGroup]:
            attn_groups: list[AttentionGroup] = []
4146
            for (attn_backend, kv_cache_spec), layer_names in attn_backends_map.items():
4147
                attn_group = AttentionGroup(
4148
                    attn_backend,
4149
                    layer_names,
4150
                    kv_cache_spec,
4151
                    kv_cache_group_id,
4152
4153
                )

4154
4155
4156
                attn_groups.append(attn_group)
            return attn_groups

4157
4158
        attention_backend_maps = []
        attention_backend_set: set[type[AttentionBackend]] = set()
4159
        for kv_cache_group_spec in kv_cache_config.kv_cache_groups:
4160
            attn_backends = get_attn_backends_for_group(kv_cache_group_spec)
4161
4162
4163
4164
4165
4166
            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)

4167
4168
        for i, attn_backend_map in enumerate(attention_backend_maps):
            self.attn_groups.append(create_attn_groups(attn_backend_map, i))
4169

4170
4171
4172
4173
4174
4175
4176
4177
4178
4179
4180
4181
4182
4183
4184
4185
4186
4187
    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
4188
        # Calculate reorder batch threshold (if needed)
4189
4190
        # Note (tdoublep): do this *after* constructing builders,
        # because some of them change the threshold at init time.
4191
4192
        self.calculate_reorder_batch_threshold()

4193
4194
4195
    def _check_and_update_cudagraph_mode(
        self, attention_backends: set[type[AttentionBackend]]
    ) -> None:
4196
        """
4197
        Resolve the cudagraph_mode when there are multiple attention
4198
4199
4200
4201
        backends with potential conflicting CUDA graph support.
        Then initialize the cudagraph_dispatcher based on the resolved
        cudagraph_mode.
        """
4202
        min_cg_support = AttentionCGSupport.ALWAYS
4203
        min_cg_backend_name = None
4204

4205
4206
4207
4208
4209
        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__
4210
4211
4212
        # Flexible resolve the cudagraph mode
        cudagraph_mode = self.compilation_config.cudagraph_mode
        # check cudagraph for mixed batch is supported
4213
4214
4215
4216
4217
4218
        if (
            cudagraph_mode.mixed_mode() == CUDAGraphMode.FULL
            and min_cg_support != AttentionCGSupport.ALWAYS
        ):
            msg = (
                f"CUDAGraphMode.{cudagraph_mode.name} is not supported "
4219
                f"with {min_cg_backend_name} backend (support: "
4220
4221
                f"{min_cg_support})"
            )
4222
4223
            if min_cg_support == AttentionCGSupport.NEVER:
                # if not supported any full cudagraphs, just raise it.
4224
4225
                msg += (
                    "; please try cudagraph_mode=PIECEWISE, and "
4226
                    "make sure compilation mode is VLLM_COMPILE"
4227
                )
4228
4229
4230
4231
4232
                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"
4233
                cudagraph_mode = self.compilation_config.cudagraph_mode = (
4234
                    CUDAGraphMode.FULL_AND_PIECEWISE
4235
                )
4236
4237
            else:
                msg += "; setting cudagraph_mode=FULL_DECODE_ONLY"
4238
                cudagraph_mode = self.compilation_config.cudagraph_mode = (
4239
                    CUDAGraphMode.FULL_DECODE_ONLY
4240
                )
4241
4242
            logger.warning(msg)

4243
        # check that if we are doing decode full-cudagraphs it is supported
4244
4245
4246
4247
4248
4249
        if (
            cudagraph_mode.decode_mode() == CUDAGraphMode.FULL
            and min_cg_support == AttentionCGSupport.NEVER
        ):
            msg = (
                f"CUDAGraphMode.{cudagraph_mode.name} is not supported "
4250
                f"with {min_cg_backend_name} backend (support: "
4251
4252
                f"{min_cg_support})"
            )
4253
            if self.compilation_config.mode == CompilationMode.VLLM_COMPILE and (
4254
4255
4256
4257
4258
                self.compilation_config.splitting_ops_contain_attention()
                or self.compilation_config.use_inductor_graph_partition
            ):
                msg += (
                    "; setting cudagraph_mode=PIECEWISE because "
4259
                    "attention is compiled piecewise"
4260
4261
                )
                cudagraph_mode = self.compilation_config.cudagraph_mode = (
4262
                    CUDAGraphMode.PIECEWISE
4263
                )
4264
            else:
4265
4266
                msg += (
                    "; setting cudagraph_mode=NONE because "
4267
                    "attention is not compiled piecewise"
4268
4269
                )
                cudagraph_mode = self.compilation_config.cudagraph_mode = (
4270
                    CUDAGraphMode.NONE
4271
                )
4272
4273
            logger.warning(msg)

4274
4275
        # check that if we are doing spec-decode + decode full-cudagraphs it is
        # supported
4276
4277
4278
4279
4280
4281
4282
4283
        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 "
4284
                f"{min_cg_backend_name} (support: {min_cg_support})"
4285
            )
4286
4287
            if self.compilation_config.splitting_ops_contain_attention():
                msg += "; setting cudagraph_mode=PIECEWISE"
4288
                cudagraph_mode = self.compilation_config.cudagraph_mode = (
4289
                    CUDAGraphMode.PIECEWISE
4290
                )
4291
4292
            else:
                msg += "; setting cudagraph_mode=NONE"
4293
                cudagraph_mode = self.compilation_config.cudagraph_mode = (
4294
                    CUDAGraphMode.NONE
4295
                )
4296
4297
4298
4299
            logger.warning(msg)

        # double check that we can support full cudagraph if they are requested
        # even after automatic downgrades
4300
4301
4302
4303
4304
4305
        if (
            cudagraph_mode.has_full_cudagraphs()
            and min_cg_support == AttentionCGSupport.NEVER
        ):
            raise ValueError(
                f"CUDAGraphMode.{cudagraph_mode.name} is not "
4306
                f"supported with {min_cg_backend_name} backend ("
4307
4308
                f"support:{min_cg_support}) "
                "; please try cudagraph_mode=PIECEWISE, "
4309
                "and make sure compilation mode is VLLM_COMPILE"
4310
            )
4311

4312
4313
        # Trigger cudagraph dispatching keys initialization after
        # resolved cudagraph mode.
4314
        self.cudagraph_dispatcher.initialize_cudagraph_keys(
4315
4316
            self.compilation_config.cudagraph_mode, self.uniform_decode_query_len
        )
4317

4318
4319
    def calculate_reorder_batch_threshold(self) -> None:
        """
4320
4321
4322
4323
        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.
4324
        """
4325
4326
4327
4328
4329
4330
        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()
        ]
4331
4332
4333
4334
4335
        # 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
4336
        self.reorder_batch_threshold = reduce(min_none_high, reorder_batch_thresholds)
4337

4338
4339
4340
    @staticmethod
    def select_common_block_size(
        kv_manager_block_size: int, attn_groups: list[AttentionGroup]
4341
4342
    ) -> int:
        """
4343
4344
4345
4346
4347
        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.
4348
4349
4350
4351
4352
4353

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

        Returns:
4354
            The selected block size
4355
4356

        Raises:
4357
            ValueError: If no valid block size found
4358
4359
        """

4360
4361
4362
4363
4364
4365
4366
4367
4368
4369
4370
4371
4372
4373
4374
4375
4376
4377
4378
4379
4380
4381
4382
4383
4384
4385
4386
4387
4388
4389
4390
4391
4392
4393
4394
4395
4396
4397
4398
4399
4400
4401
        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)
        )
4402

4403
4404
4405
4406
4407
4408
        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}. ")
4409

4410
4411
4412
    def may_reinitialize_input_batch(
        self, kv_cache_config: KVCacheConfig, kernel_block_sizes: list[int]
    ) -> None:
4413
4414
4415
4416
4417
4418
4419
        """
        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.
4420
            kernel_block_sizes: The kernel block sizes for each KV cache group.
4421
4422
4423
4424
        """
        block_sizes = [
            kv_cache_group.kv_cache_spec.block_size
            for kv_cache_group in kv_cache_config.kv_cache_groups
4425
            if not isinstance(kv_cache_group.kv_cache_spec, EncoderOnlyAttentionSpec)
4426
        ]
4427
4428
4429
4430

        if block_sizes != [self.cache_config.block_size] or kernel_block_sizes != [
            self.cache_config.block_size
        ]:
4431
4432
4433
            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
4434
4435
                "for more details."
            )
4436
4437
            self.input_batch = InputBatch(
                max_num_reqs=self.max_num_reqs,
4438
                max_model_len=max(self.max_model_len, self.max_encoder_len),
4439
4440
4441
4442
4443
                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,
4444
                kernel_block_sizes=kernel_block_sizes,
4445
                is_spec_decode=bool(self.vllm_config.speculative_config),
4446
                logitsprocs=self.input_batch.logitsprocs,
4447
                logitsprocs_need_output_token_ids=self.input_batch.logitsprocs_need_output_token_ids,
4448
                is_pooling_model=self.is_pooling_model,
4449
4450
                num_speculative_tokens=(
                    self.vllm_config.speculative_config.num_speculative_tokens
4451
4452
4453
                    if self.vllm_config.speculative_config
                    else 0
                ),
4454
4455
            )

4456
    def _allocate_kv_cache_tensors(
4457
4458
        self, kv_cache_config: KVCacheConfig
    ) -> dict[str, torch.Tensor]:
4459
        """
4460
4461
4462
        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.

4463
        Args:
4464
            kv_cache_config: The KV cache config
4465
        Returns:
4466
            dict[str, torch.Tensor]: A map between layer names to their
4467
            corresponding memory buffer for KV cache.
4468
        """
4469
4470
        kv_cache_raw_tensors: dict[str, torch.Tensor] = {}
        for kv_cache_tensor in kv_cache_config.kv_cache_tensors:
4471
4472
4473
            tensor = torch.zeros(
                kv_cache_tensor.size, dtype=torch.int8, device=self.device
            )
4474
4475
4476
4477
4478
            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:
4479
4480
4481
4482
            for layer_name in group.layer_names:
                if layer_name in self.runner_only_attn_layers:
                    continue
                layer_names.add(layer_name)
4483
4484
4485
        assert layer_names == set(kv_cache_raw_tensors.keys()), (
            "Some layers are not correctly initialized"
        )
4486
4487
        return kv_cache_raw_tensors

4488
4489
4490
    def _attn_group_iterator(self) -> Iterator[AttentionGroup]:
        return itertools.chain.from_iterable(self.attn_groups)

4491
    def _kv_cache_spec_attn_group_iterator(self) -> Iterator[AttentionGroup]:
4492
4493
        if not self.kv_cache_config.kv_cache_groups:
            return
4494
4495
        for attn_groups in self.attn_groups:
            yield from attn_groups
4496

4497
4498
4499
4500
4501
4502
4503
4504
4505
4506
4507
4508
4509
4510
4511
    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 = []
4512
        for kv_cache_gid, kv_cache_group in enumerate(kv_cache_config.kv_cache_groups):
4513
4514
4515
4516
4517
4518
            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):
4519
                continue
4520
            elif isinstance(kv_cache_spec, AttentionSpec):
4521
4522
4523
                # This is an attention backend that supports virtual
                # block splitting. Get the supported block sizes from
                # all backends in the group.
4524
                attn_groups = self.attn_groups[kv_cache_gid]
4525
                kv_manager_block_size = kv_cache_group.kv_cache_spec.block_size
4526
                selected_kernel_size = self.select_common_block_size(
4527
4528
4529
                    kv_manager_block_size, attn_groups
                )
                kernel_block_sizes.append(selected_kernel_size)
4530
            elif isinstance(kv_cache_spec, MambaSpec):
4531
4532
                # This is likely Mamba or other non-attention cache,
                # no splitting.
4533
                kernel_block_sizes.append(kv_cache_spec.block_size)
4534
4535
4536
4537
4538
4539
            else:
                raise NotImplementedError(
                    f"unknown kv cache spec {kv_cache_group.kv_cache_spec}"
                )
        return kernel_block_sizes

4540
4541
4542
4543
    def _reshape_kv_cache_tensors(
        self,
        kv_cache_config: KVCacheConfig,
        kv_cache_raw_tensors: dict[str, torch.Tensor],
4544
        kernel_block_sizes: list[int],
4545
    ) -> dict[str, torch.Tensor]:
4546
        """
4547
        Reshape the KV cache tensors to the desired shape and dtype.
4548

4549
        Args:
4550
4551
            kv_cache_config: The KV cache config
            kv_cache_raw_tensors: The KV cache buffer of each layer, with
4552
                correct size but uninitialized shape.
4553
            kernel_block_sizes: The kernel block sizes for each KV cache group.
4554
        Returns:
4555
            Dict[str, torch.Tensor]: A map between layer names to their
4556
4557
            corresponding memory buffer for KV cache.
        """
4558
        kv_caches: dict[str, torch.Tensor] = {}
4559
        has_attn, has_mamba = False, False
4560
4561
        for group in self._kv_cache_spec_attn_group_iterator():
            kv_cache_spec = group.kv_cache_spec
4562
            attn_backend = group.backend
4563
4564
4565
4566
            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]
4567
            for layer_name in group.layer_names:
4568
4569
                if layer_name in self.runner_only_attn_layers:
                    continue
4570
4571
                raw_tensor = kv_cache_raw_tensors[layer_name]
                assert raw_tensor.numel() % kv_cache_spec.page_size_bytes == 0
4572
                num_blocks = raw_tensor.numel() // kv_cache_spec.page_size_bytes
4573
                if isinstance(kv_cache_spec, AttentionSpec):
4574
                    has_attn = True
4575
4576
                    num_blocks_per_kv_block = (
                        kv_cache_spec.block_size // kernel_block_size
4577
4578
4579
                    )
                    kernel_num_blocks = num_blocks * num_blocks_per_kv_block

4580
                    kv_cache_shape = attn_backend.get_kv_cache_shape(
4581
                        kernel_num_blocks,
4582
                        kernel_block_size,
4583
4584
                        kv_cache_spec.num_kv_heads,
                        kv_cache_spec.head_size,
4585
4586
                        cache_dtype_str=self.cache_config.cache_dtype,
                    )
4587
                    dtype = kv_cache_spec.dtype
4588
                    try:
4589
                        kv_cache_stride_order = attn_backend.get_kv_cache_stride_order()
4590
                        assert len(kv_cache_stride_order) == len(kv_cache_shape)
4591
                    except (AttributeError, NotImplementedError):
4592
                        kv_cache_stride_order = tuple(range(len(kv_cache_shape)))
4593
4594
4595
4596
4597
                    # 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.
4598
4599
4600
                    kv_cache_shape = tuple(
                        kv_cache_shape[i] for i in kv_cache_stride_order
                    )
4601
4602
4603
4604
4605
                    # Maintain original KV shape view.
                    inv_order = [
                        kv_cache_stride_order.index(i)
                        for i in range(len(kv_cache_stride_order))
                    ]
4606
4607
4608
4609
4610
4611
                    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
4612
                elif isinstance(kv_cache_spec, MambaSpec):
4613
                    has_mamba = True
Chen Zhang's avatar
Chen Zhang committed
4614
4615
                    raw_tensor = kv_cache_raw_tensors[layer_name]
                    state_tensors = []
4616
                    storage_offset_bytes = 0
4617
                    for shape, dtype in zip(kv_cache_spec.shapes, kv_cache_spec.dtypes):
4618
4619
                        dtype_size = get_dtype_size(dtype)
                        num_element_per_page = (
4620
4621
                            kv_cache_spec.page_size_bytes // dtype_size
                        )
Chen Zhang's avatar
Chen Zhang committed
4622
                        target_shape = (num_blocks, *shape)
4623
4624
                        stride = torch.empty(target_shape).stride()
                        target_stride = (num_element_per_page, *stride[1:])
4625
                        assert storage_offset_bytes % dtype_size == 0
4626
4627
4628
4629
                        tensor = torch.as_strided(
                            raw_tensor.view(dtype),
                            size=target_shape,
                            stride=target_stride,
4630
                            storage_offset=storage_offset_bytes // dtype_size,
4631
                        )
Chen Zhang's avatar
Chen Zhang committed
4632
                        state_tensors.append(tensor)
4633
                        storage_offset_bytes += stride[0] * dtype_size
4634
4635

                    kv_caches[layer_name] = state_tensors
4636
                else:
4637
                    raise NotImplementedError
4638
4639

        if has_attn and has_mamba:
4640
            self._update_hybrid_attention_mamba_layout(kv_caches)
4641

4642
4643
        return kv_caches

4644
    def _update_hybrid_attention_mamba_layout(
4645
4646
        self, kv_caches: dict[str, torch.Tensor]
    ) -> None:
4647
        """
4648
4649
        Update the layout of attention layers from (2, num_blocks, ...) to
        (num_blocks, 2, ...).
4650
4651

        Args:
4652
            kv_caches: The KV cache buffer of each layer.
4653
4654
        """

4655
4656
        for group in self._kv_cache_spec_attn_group_iterator():
            kv_cache_spec = group.kv_cache_spec
4657
            for layer_name in group.layer_names:
4658
                kv_cache = kv_caches[layer_name]
4659
4660
4661
4662
                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 "
4663
                        f"a tensor of shape {kv_cache.shape}"
4664
                    )
4665
                    hidden_size = kv_cache.shape[2:].numel()
4666
4667
4668
4669
                    kv_cache.as_strided_(
                        size=kv_cache.shape,
                        stride=(hidden_size, 2 * hidden_size, *kv_cache.stride()[2:]),
                    )
4670

4671
    def initialize_kv_cache_tensors(
4672
        self, kv_cache_config: KVCacheConfig, kernel_block_sizes: list[int]
4673
    ) -> dict[str, torch.Tensor]:
4674
4675
4676
4677
4678
        """
        Initialize the memory buffer for KV cache.

        Args:
            kv_cache_config: The KV cache config
4679
4680
            kernel_block_sizes: The kernel block sizes for each KV cache group.

4681
        Returns:
4682
            Dict[str, torch.Tensor]: A map between layer names to their
4683
4684
4685
4686
4687
            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
4688
        kv_caches = self._reshape_kv_cache_tensors(
4689
            kv_cache_config, kv_cache_raw_tensors, kernel_block_sizes
4690
        )
4691

4692
        # Set up cross-layer KV cache sharing
4693
4694
        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)
4695
4696
            kv_caches[layer_name] = kv_caches[target_layer_name]

4697
4698
4699
4700
4701
4702
4703
4704
4705
        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,
        )
4706
4707
4708
        return kv_caches

    def maybe_add_kv_sharing_layers_to_kv_cache_groups(
4709
4710
        self, kv_cache_config: KVCacheConfig
    ) -> None:
4711
4712
4713
4714
4715
4716
4717
4718
4719
4720
4721
4722
4723
4724
4725
4726
4727
4728
        """
        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.
4729
            attn_layers = get_layers_from_vllm_config(self.vllm_config, Attention)
4730
4731
            for layer_name in reversed(attn_layers):
                if layer_name in self.shared_kv_cache_layers:
4732
                    self.kv_sharing_fast_prefill_eligible_layers.add(layer_name)
4733
4734
                else:
                    break
4735

4736
4737
4738
4739
4740
4741
4742
    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
        """
4743
        kv_cache_config = deepcopy(kv_cache_config)
4744
        self.kv_cache_config = kv_cache_config
4745
        self.may_add_encoder_only_layers_to_kv_cache_config()
4746
        self.maybe_add_kv_sharing_layers_to_kv_cache_groups(kv_cache_config)
4747
        self.initialize_attn_backend(kv_cache_config)
4748
4749
4750
4751
4752
4753
        # 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)
4754
4755
4756
4757

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

4758
        # Reinitialize need to after initialize_attn_backend
4759
4760
4761
4762
        self.may_reinitialize_input_batch(kv_cache_config, kernel_block_sizes)
        kv_caches = self.initialize_kv_cache_tensors(
            kv_cache_config, kernel_block_sizes
        )
4763

4764
4765
4766
4767
4768
4769
        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
4770
        if has_kv_transfer_group():
4771
4772
4773
            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
4774

4775
        if self.dcp_world_size > 1:
4776
            layers = get_layers_from_vllm_config(self.vllm_config, AttentionLayerBase)
4777
4778
4779
4780
4781
            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__} "
4782
4783
                    "does not return the softmax lse for decode."
                )
4784

4785
4786
4787
4788
4789
    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
4790
        encoder_only_attn_specs: dict[AttentionSpec, list[str]] = defaultdict(list)
4791
4792
4793
        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:
4794
                attn_spec: AttentionSpec = EncoderOnlyAttentionSpec(
4795
4796
4797
                    block_size=block_size,
                    num_kv_heads=attn_module.num_kv_heads,
                    head_size=attn_module.head_size,
4798
4799
                    dtype=self.kv_cache_dtype,
                )
4800
4801
4802
                encoder_only_attn_specs[attn_spec].append(layer_name)
                self.runner_only_attn_layers.add(layer_name)
        if len(encoder_only_attn_specs) > 0:
4803
4804
4805
            assert len(encoder_only_attn_specs) == 1, (
                "Only support one encoder-only attention spec now"
            )
4806
4807
            spec, layer_names = encoder_only_attn_specs.popitem()
            self.kv_cache_config.kv_cache_groups.append(
4808
4809
                KVCacheGroupSpec(layer_names=layer_names, kv_cache_spec=spec)
            )
4810

4811
    def get_kv_cache_spec(self) -> dict[str, KVCacheSpec]:
4812
        """
4813
        Generates the KVCacheSpec by parsing the kv cache format from each
4814
4815
        Attention module in the static forward context.
        Returns:
4816
            KVCacheSpec: A dictionary mapping layer names to their KV cache
4817
4818
4819
            format. Layers that do not need KV cache are not included.
        """

4820
        kv_cache_spec: dict[str, KVCacheSpec] = {}
4821
        attn_layers = get_layers_from_vllm_config(self.vllm_config, AttentionLayerBase)
Chen Zhang's avatar
Chen Zhang committed
4822
        for layer_name, attn_module in attn_layers.items():
4823
4824
4825
4826
4827
4828
4829
4830
4831
4832
4833
4834
4835
4836
4837
            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
4838

4839
        return kv_cache_spec
4840
4841
4842
4843
4844
4845
4846
4847
4848
4849

    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.
4850
        pinned = self.sampled_token_ids_pinned_cpu[: sampled_token_ids.shape[0]]
4851
4852
4853
4854
        pinned.copy_(sampled_token_ids, non_blocking=True)
        self.transfer_event.record()
        self.transfer_event.synchronize()
        return pinned.tolist()