gpu_model_runner.py 221 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
from collections import defaultdict
8
from collections.abc import Iterator, Sequence
9
from contextlib import contextmanager
10
from copy import copy, 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
24
25
26
27
from vllm.attention.backends.abstract import (
    AttentionBackend,
    AttentionMetadata,
    MultipleOf,
)
28
from vllm.compilation.counter import compilation_counter
29
30
from vllm.compilation.cuda_graph import CUDAGraphWrapper
from vllm.compilation.monitor import set_cudagraph_capturing_enabled
31
from vllm.config import (
32
    CompilationMode,
33
34
35
36
37
    CUDAGraphMode,
    VllmConfig,
    get_layers_from_vllm_config,
    update_config,
)
38
from vllm.distributed.ec_transfer import get_ec_transfer, has_ec_transfer
39
from vllm.distributed.eplb.eplb_state import EplbState
40
from vllm.distributed.kv_transfer import get_kv_transfer_group, has_kv_transfer_group
41
from vllm.distributed.kv_transfer.kv_connector.utils import copy_kv_blocks
42
from vllm.distributed.parallel_state import (
43
    get_dcp_group,
44
45
46
47
48
49
    get_pp_group,
    get_tp_group,
    graph_capture,
    is_global_first_rank,
    prepare_communication_buffer_for_model,
)
50
from vllm.forward_context import BatchDescriptor, set_forward_context
51
from vllm.logger import init_logger
52
from vllm.model_executor.layers.attention_layer_base import AttentionLayerBase
53
from vllm.model_executor.layers.rotary_embedding import MRotaryEmbedding
54
from vllm.model_executor.model_loader import TensorizerLoader, get_model_loader
55
from vllm.model_executor.models.interfaces import (
56
    SupportsMRoPE,
57
58
59
60
61
62
63
    SupportsMultiModal,
    is_mixture_of_experts,
    supports_eagle3,
    supports_mrope,
    supports_multimodal_pruning,
    supports_transcription,
)
64
from vllm.model_executor.models.interfaces_base import (
65
66
67
68
    VllmModelForPooling,
    is_pooling_model,
    is_text_generation_model,
)
69
from vllm.multimodal import MULTIMODAL_REGISTRY
70
71
72
73
74
from vllm.multimodal.inputs import (
    BatchedTensorInputs,
    MultiModalKwargsItem,
    PlaceholderRange,
)
75
from vllm.multimodal.utils import group_mm_kwargs_by_modality
76
from vllm.pooling_params import PoolingParams
77
from vllm.sampling_params import SamplingType
78
from vllm.sequence import IntermediateTensors
79
from vllm.tasks import GenerationTask, PoolingTask, SupportedTask
80
from vllm.utils import length_from_prompt_token_ids_or_embeds
81
from vllm.utils.jsontree import json_map_leaves
82
from vllm.utils.math_utils import cdiv, round_up
83
84
from vllm.utils.mem_constants import GiB_bytes
from vllm.utils.mem_utils import DeviceMemoryProfiler
85
from vllm.utils.platform_utils import is_pin_memory_available
86
87
88
89
90
from vllm.utils.torch_utils import (
    get_dtype_size,
    kv_cache_dtype_str_to_dtype,
    supports_dynamo,
)
91
from vllm.v1.attention.backends.gdn_attn import GDNAttentionMetadataBuilder
92
from vllm.v1.attention.backends.utils import (
93
94
95
    AttentionCGSupport,
    AttentionMetadataBuilder,
    CommonAttentionMetadata,
96
    create_fast_prefill_custom_backend,
97
    get_dcp_local_seq_lens,
98
99
100
    reorder_batch_to_split_decodes_and_prefills,
    split_attn_metadata,
)
101
from vllm.v1.cudagraph_dispatcher import CudagraphDispatcher
102
103
104
105
106
107
108
109
110
111
112
113
114
115
116
117
118
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,
119
    ECConnectorOutput,
120
    KVConnectorOutput,
121
122
123
124
125
    LogprobsLists,
    LogprobsTensors,
    ModelRunnerOutput,
    PoolerOutput,
    SamplerOutput,
126
    make_empty_encoder_model_runner_output,
127
)
128
from vllm.v1.pool.metadata import PoolingMetadata
129
from vllm.v1.sample.logits_processor import LogitsProcessors, build_logitsprocs
130
from vllm.v1.sample.logits_processor.interface import LogitsProcessor
131
from vllm.v1.sample.metadata import SamplingMetadata
132
from vllm.v1.sample.rejection_sampler import RejectionSampler
133
from vllm.v1.sample.sampler import Sampler
134
from vllm.v1.spec_decode.eagle import EagleProposer
135
from vllm.v1.spec_decode.medusa import MedusaProposer
136
from vllm.v1.spec_decode.metadata import SpecDecodeMetadata
137
from vllm.v1.spec_decode.ngram_proposer import NgramProposer
138
from vllm.v1.spec_decode.suffix_decoding import SuffixDecodingProposer
139
from vllm.v1.structured_output.utils import apply_grammar_bitmask
140
from vllm.v1.utils import CpuGpuBuffer, record_function_or_nullcontext
141
from vllm.v1.worker.dp_utils import coordinate_batch_across_dp
142
from vllm.v1.worker.ec_connector_model_runner_mixin import ECConnectorModelRunnerMixin
143
from vllm.v1.worker.gpu_input_batch import CachedRequestState, InputBatch
144
from vllm.v1.worker.gpu_ubatch_wrapper import UBatchWrapper
145
from vllm.v1.worker.kv_connector_model_runner_mixin import KVConnectorModelRunnerMixin
146
from vllm.v1.worker.lora_model_runner_mixin import LoRAModelRunnerMixin
147
148
149
150
151
from vllm.v1.worker.ubatch_utils import (
    UBatchSlice,
    UBatchSlices,
    check_ubatch_thresholds,
)
152
from vllm.v1.worker.utils import is_residual_scattered_for_sp
153

154
155
156
157
158
159
160
161
162
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,
)
163

164
if TYPE_CHECKING:
165
    from vllm.model_executor.model_loader.tensorizer import TensorizerConfig
166
    from vllm.v1.core.sched.output import GrammarOutput, SchedulerOutput
167
168
169

logger = init_logger(__name__)

170
171
AttnMetadataDict: TypeAlias = dict[str, AttentionMetadata]
# list when ubatching is enabled
172
PerLayerAttnMetadata: TypeAlias = list[AttnMetadataDict] | AttnMetadataDict
173

174

175
176
177
178
179
180
# Wrapper for ModelRunnerOutput to support overlapped execution.
class AsyncGPUModelRunnerOutput(AsyncModelRunnerOutput):
    def __init__(
        self,
        model_runner_output: ModelRunnerOutput,
        sampled_token_ids: torch.Tensor,
181
        logprobs_tensors: torch.Tensor | None,
182
183
        invalid_req_indices: list[int],
        async_output_copy_stream: torch.cuda.Stream,
184
        vocab_size: int,
185
186
187
188
189
    ):
        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.
190
        self.async_copy_ready_event = torch.Event()
191
192
193
194

        # 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
195
        self.vocab_size = vocab_size
196
        self._logprobs_tensors = logprobs_tensors
197
198
199
200
201

        # 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)
202
            self.sampled_token_ids_cpu = self._sampled_token_ids.to(
203
204
                "cpu", non_blocking=True
            )
205
206
207
208
209
            self._logprobs_tensors_cpu = (
                self._logprobs_tensors.to_cpu_nonblocking()
                if self._logprobs_tensors
                else None
            )
210
            self.async_copy_ready_event.record()
211
212
213

    def get_output(self) -> ModelRunnerOutput:
        """Copy the device tensors to the host and return a ModelRunnerOutput.
214

215
216
        This function blocks until the copy is finished.
        """
217
        self.async_copy_ready_event.synchronize()
218

219
220
        # Release the device tensors once the copy has completed.
        del self._logprobs_tensors
221
        del self._sampled_token_ids
222
223
        max_gen_len = self.sampled_token_ids_cpu.shape[-1]
        if max_gen_len == 1:
224
            valid_sampled_token_ids = self.sampled_token_ids_cpu.tolist()
225
226
227
228
229
        else:
            valid_sampled_token_ids = RejectionSampler.parse_output(
                self.sampled_token_ids_cpu,
                self.vocab_size,
            )
230
        for i in self._invalid_req_indices:
231
            valid_sampled_token_ids[i].clear()
232
233
234

        output = self._model_runner_output
        output.sampled_token_ids = valid_sampled_token_ids
235
236
237
238
        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()
239
240
241
        return output


242
243
244
245
246
247
248
249
250
251
252
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
253
    ec_connector_output: ECConnectorOutput | None
254
255


256
257
258
class GPUModelRunner(
    LoRAModelRunnerMixin, KVConnectorModelRunnerMixin, ECConnectorModelRunnerMixin
):
259
260
    def __init__(
        self,
261
        vllm_config: VllmConfig,
262
        device: torch.device,
263
    ):
264
265
266
        self.vllm_config = vllm_config
        self.model_config = vllm_config.model_config
        self.cache_config = vllm_config.cache_config
267
        self.compilation_config = vllm_config.compilation_config
268
269
270
271
272
273
        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
274

275
        from vllm.model_executor.models.utils import set_cpu_offload_max_bytes
276
277

        set_cpu_offload_max_bytes(int(self.cache_config.cpu_offload_gb * 1024**3))
278

279
280
281
282
        model_config = self.model_config
        cache_config = self.cache_config
        scheduler_config = self.scheduler_config
        parallel_config = self.parallel_config
283
        self.device = device
284
285
        self.pin_memory = is_pin_memory_available()
        self.dtype = self.model_config.dtype
286
287
288
        self.kv_cache_dtype = kv_cache_dtype_str_to_dtype(
            cache_config.cache_dtype, self.model_config
        )
289

290
        self.is_pooling_model = model_config.runner_type == "pooling"
291
        self.enable_prompt_embeds = model_config.enable_prompt_embeds
292
        self.is_multimodal_raw_input_only_model = (
293
294
            model_config.is_multimodal_raw_input_only_model
        )
295
296
        # This will be overridden in load_model()
        self.is_multimodal_pruning_enabled = False
297
        self.max_model_len = model_config.max_model_len
298
299
300

        # Always set to false after the first forward pass
        self.calculate_kv_scales = self.cache_config.calculate_kv_scales
301
        self.dcp_world_size = self.parallel_config.decode_context_parallel_size
302
        self.dcp_rank = 0 if self.dcp_world_size <= 1 else get_dcp_group().rank_in_group
303
        self.max_num_tokens = scheduler_config.max_num_batched_tokens
304
        self.max_num_reqs = scheduler_config.max_num_seqs
305

306
307
308
309
310
        # 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 = (
311
312
313
            self.parallel_config.distributed_executor_backend == "external_launcher"
            and len(get_pp_group().ranks) > 0
        )
314

315
        # Model-related.
316
        self.num_query_heads = model_config.get_num_attention_heads(parallel_config)
317
        self.hidden_size = model_config.get_hidden_size()
318
        self.attention_chunk_size = model_config.attention_chunk_size
319
        # Only relevant for models using ALiBi (e.g, MPT)
320
        self.use_alibi = model_config.uses_alibi
321

322
        self.cascade_attn_enabled = not self.model_config.disable_cascade_attn
323

324
        # Multi-modal data support
325
        self.mm_registry = MULTIMODAL_REGISTRY
326
        self.uses_mrope = model_config.uses_mrope
327
        self.supports_mm_inputs = self.mm_registry.supports_multimodal_inputs(
328
329
            model_config
        )
330

331
332
333
        if self.model_config.is_encoder_decoder:
            # Maximum length of the encoder input, only for encoder-decoder
            # models.
334
            self.max_encoder_len = scheduler_config.max_num_encoder_input_tokens
335
336
337
        else:
            self.max_encoder_len = 0

338
        # Sampler
339
        self.sampler = Sampler(logprobs_mode=self.model_config.logprobs_mode)
340

341
        self.eplb_state: EplbState | None = None
342
343
344
345
346
347
        """
        State of the expert parallelism load balancer.

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

348
        # Lazy initializations
349
        # self.model: nn.Module  # Set after load_model
350
        # Initialize in initialize_kv_cache
351
        self.kv_caches: list[torch.Tensor] = []
352
353
354
        # Initialize in initialize_kv_cache_tensors
        self.cross_layers_kv_cache: torch.Tensor | None = None
        self.cross_layers_attn_backend: type[AttentionBackend] | None = None
355
356
        # indexes: [kv_cache_group_id][attn_group]
        self.attn_groups: list[list[AttentionGroup]] = []
357
358
        # self.kv_cache_config: KVCacheConfig

359
360
        # mm_hash ->  encoder_output
        self.encoder_cache: dict[str, torch.Tensor] = {}
361

362
        self.use_aux_hidden_state_outputs = False
363
364
365
366
367
        # 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:
368
369
370
            self.drafter: (
                NgramProposer | SuffixDecodingProposer | EagleProposer | MedusaProposer
            )
371
372
            if self.speculative_config.method == "ngram":
                self.drafter = NgramProposer(self.vllm_config)
373
374
            elif self.speculative_config.method == "suffix":
                self.drafter = SuffixDecodingProposer(self.vllm_config)
375
            elif self.speculative_config.use_eagle():
376
                self.drafter = EagleProposer(self.vllm_config, self.device, self)
377
378
379
380
                if self.speculative_config.method == "eagle3":
                    self.use_aux_hidden_state_outputs = True
            elif self.speculative_config.method == "medusa":
                self.drafter = MedusaProposer(
381
                    vllm_config=self.vllm_config, device=self.device
382
                )
383
            else:
384
385
386
387
                raise ValueError(
                    "Unknown speculative decoding method: "
                    f"{self.speculative_config.method}"
                )
388
            self.rejection_sampler = RejectionSampler(self.sampler)
389

390
391
392
393
        self.num_spec_tokens = 0
        if self.speculative_config:
            self.num_spec_tokens = self.speculative_config.num_speculative_tokens

394
        # Request states.
395
        self.requests: dict[str, CachedRequestState] = {}
396
397
398
        # NOTE(rob): num_prompt_logprobs only includes reqs
        # that are currently in the prefill phase.
        self.num_prompt_logprobs: dict[str, int] = {}
399
        self.comm_stream = torch.cuda.Stream()
400

401
402
403
404
405
406
407
408
409
        # 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.
410
411
412
413
        logits_processors = model_config.logits_processors
        custom_logitsprocs: Sequence[str | type[LogitsProcessor]] = (
            tuple(logits_processors) if logits_processors is not None else ()
        )
414
415
        self.input_batch = InputBatch(
            max_num_reqs=self.max_num_reqs,
416
417
418
            # 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),
419
420
421
            max_num_batched_tokens=self.max_num_tokens,
            device=self.device,
            pin_memory=self.pin_memory,
422
            vocab_size=self.model_config.get_vocab_size(),
423
            block_sizes=[self.cache_config.block_size],
424
            kernel_block_sizes=[self.cache_config.block_size],
425
            is_spec_decode=bool(self.vllm_config.speculative_config),
426
            logitsprocs=build_logitsprocs(
427
428
429
                self.vllm_config,
                self.device,
                self.pin_memory,
430
                self.is_pooling_model,
431
                custom_logitsprocs,
432
            ),
433
434
435
            # 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),
436
            is_pooling_model=self.is_pooling_model,
437
            cp_kv_cache_interleave_size=self.parallel_config.cp_kv_cache_interleave_size,
438
        )
439

440
        self.use_async_scheduling = self.scheduler_config.async_scheduling
441
442
443
444
445
        # 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.
446
        self.prepare_inputs_event: torch.Event | None = None
447
448
        if self.use_async_scheduling:
            self.async_output_copy_stream = torch.cuda.Stream()
449
            self.prepare_inputs_event = torch.Event()
450

451
        # self.cudagraph_batch_sizes sorts in ascending order.
452
453
454
455
        if (
            self.compilation_config.cudagraph_capture_sizes
            and self.compilation_config.cudagraph_mode != CUDAGraphMode.NONE
        ):
456
457
            self.cudagraph_batch_sizes = sorted(
                self.compilation_config.cudagraph_capture_sizes
458
            )
459

460
        # Cache the device properties.
461
        self._init_device_properties()
462

463
        # Persistent buffers for CUDA graphs.
464
465
466
467
468
        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
        )
469
        self.seq_lens = self._make_buffer(self.max_num_reqs, dtype=torch.int32)
470
471
472
473
        if self.dcp_world_size > 1:
            self.dcp_local_seq_lens = self._make_buffer(
                self.max_num_reqs, dtype=torch.int32
            )
474
475
476
        # 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.
477
478
479
480
481
482
483
        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
        )
484
485
        self.num_discarded_requests = 0

486
487
488
489
490
491
        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
        )
492

493
494
        # Only relevant for multimodal models
        if self.supports_mm_inputs:
495
            self.is_mm_embed = self._make_buffer(self.max_num_tokens, dtype=torch.bool)
496

497
        # Only relevant for models using M-RoPE (e.g, Qwen2-VL)
498
        if self.uses_mrope:
Roger Wang's avatar
Roger Wang committed
499
500
501
502
            # 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
503
504
505
506
507
508

            # 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
509
            self.mrope_positions = self._make_buffer(
510
511
                (3, self.max_num_tokens + 1), dtype=torch.int64
            )
512

513
        # None in the first PP rank. The rest are set after load_model.
514
        self.intermediate_tensors: IntermediateTensors | None = None
515

516
        # OPTIMIZATION: Cache the tensors rather than creating them every step.
517
        # Keep in int64 to avoid overflow with long context
518
519
520
521
        self.arange_np = np.arange(
            max(self.max_num_reqs + 1, self.max_model_len, self.max_num_tokens),
            dtype=np.int64,
        )
522

523
524
525
526
527
        # 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] = {}
528
529
530
531
532
        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(
533
534
                self.max_num_tokens, dtype=torch.int32, device=self.device
            )
535

536
        self.uniform_decode_query_len = 1 + self.num_spec_tokens
537
538
539
540

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

541
542
543
544
545
546
547
548
549
        self.mm_budget = (
            MultiModalBudget(
                self.model_config,
                self.scheduler_config,
                self.mm_registry,
            )
            if self.supports_mm_inputs
            else None
        )
550

551
        self.reorder_batch_threshold: int | None = None
552

553
554
555
556
557
        # 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()

558
        # Cached outputs.
559
        self._draft_token_ids: list[list[int]] | torch.Tensor | None = None
560
        self.transfer_event = torch.Event()
561
        self.sampled_token_ids_pinned_cpu = torch.empty(
562
            (self.max_num_reqs, 1),
563
564
            dtype=torch.int64,
            device="cpu",
565
566
            pin_memory=self.pin_memory,
        )
567

568
569
        # Pre-allocated tensor for copying valid sampled token counts to CPU,
        # with dedicated stream for overlapping and event for coordination.
570
        self.valid_sampled_token_count_event: torch.Event | None = None
571
572
        self.valid_sampled_token_count_copy_stream: torch.cuda.Stream | None = None
        if self.use_async_scheduling and self.num_spec_tokens:
573
            self.valid_sampled_token_count_event = torch.Event()
574
575
576
577
578
579
580
581
            self.valid_sampled_token_count_copy_stream = torch.cuda.Stream()
        self.valid_sampled_token_count_cpu = torch.empty(
            self.max_num_reqs,
            dtype=torch.int64,
            device="cpu",
            pin_memory=self.pin_memory,
        )

582
583
        # Ephemeral state transferred between execute_model() and sample_tokens().
        self.execute_model_state: ExecuteModelState | None = None
584
        self.kv_connector_output: KVConnectorOutput | None = None
585

586
587
588
589
    def reset_mm_cache(self) -> None:
        if self.mm_budget:
            self.mm_budget.reset_cache()

590
591
592
593
594
595
596
597
598
599
    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]

600
    def _make_buffer(
601
        self, *size: int | torch.SymInt, dtype: torch.dtype, numpy: bool = True
602
603
604
605
606
607
608
609
    ) -> CpuGpuBuffer:
        return CpuGpuBuffer(
            *size,
            dtype=dtype,
            device=self.device,
            pin_memory=self.pin_memory,
            with_numpy=numpy,
        )
610

611
612
613
    def _init_model_kwargs(self, num_tokens: int):
        model_kwargs = dict[str, Any]()

614
        if not self.is_pooling_model:
615
616
            return model_kwargs

617
618
        num_reqs = self.input_batch.num_reqs
        pooling_params = self.input_batch.get_pooling_params()
619
620
621

        token_type_id_requests = dict[int, Any]()
        for i, param in enumerate(pooling_params):
622
623
624
625
626
            if (
                param.extra_kwargs is not None
                and (token_types := param.extra_kwargs.get("compressed_token_type_ids"))
                is not None
            ):
627
628
629
630
631
                token_type_id_requests[i] = token_types

        if len(token_type_id_requests) == 0:
            return model_kwargs

632
        seq_lens = self.seq_lens.gpu[:num_reqs]
633
634
635
636
637
638
639
640
        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(
641
642
            device=self.device
        )
643
644
        return model_kwargs

645
    def _may_reorder_batch(self, scheduler_output: "SchedulerOutput") -> None:
646
647
        """
        Update the order of requests in the batch based on the attention
648
        backend's needs. For example, some attention backends (namely MLA) may
649
650
651
652
653
654
        want to separate requests based on if the attention computation will be
        compute-bound or memory-bound.

        Args:
            scheduler_output: The scheduler output.
        """
655
656
657
658
659
660
661
662
        # 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

663
664
665
666
        if self.reorder_batch_threshold is not None:
            reorder_batch_to_split_decodes_and_prefills(
                self.input_batch,
                scheduler_output,
667
668
                decode_threshold=self.reorder_batch_threshold,
            )
669

670
671
    # Note: used for model runner override.
    def _init_device_properties(self) -> None:
672
        """Initialize attributes from torch.cuda.get_device_properties"""
673
674
675
676
677
678
679
        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()

680
    def _update_states(self, scheduler_output: "SchedulerOutput") -> None:
681
682
683
684
685
686
        """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.

687
688
        The SamplingMetadata is updated and copied to the GPU if there is a
        new/resumed/paused/finished request in the batch.
689
690
        """
        # Remove finished requests from the cached states.
691
692
        for req_id in scheduler_output.finished_req_ids:
            self.requests.pop(req_id, None)
693
            self.num_prompt_logprobs.pop(req_id, None)
694
695
696
697
698
699
700
        # 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:
701
            self.input_batch.remove_request(req_id)
702
703

        # Free the cached encoder outputs.
704
705
        for mm_hash in scheduler_output.free_encoder_mm_hashes:
            self.encoder_cache.pop(mm_hash, None)
706

707
708
709
710
711
712
713
714
715
716
717
718
719
        # 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:
720
            self.input_batch.remove_request(req_id)
721

722
        reqs_to_add: list[CachedRequestState] = []
723
        # Add new requests to the cached states.
724
725
726
        for new_req_data in scheduler_output.scheduled_new_reqs:
            req_id = new_req_data.req_id
            sampling_params = new_req_data.sampling_params
727
            pooling_params = new_req_data.pooling_params
728

729
730
731
732
            if (
                sampling_params
                and sampling_params.sampling_type == SamplingType.RANDOM_SEED
            ):
733
734
735
736
737
                generator = torch.Generator(device=self.device)
                generator.manual_seed(sampling_params.seed)
            else:
                generator = None

738
739
            if self.is_pooling_model:
                assert pooling_params is not None
740
741
                task = pooling_params.task
                assert task is not None, "You did not set `task` in the API"
742

743
                model = cast(VllmModelForPooling, self.get_model())
744
                to_update = model.pooler.get_pooling_updates(task)
745
746
                to_update.apply(pooling_params)

747
            req_state = CachedRequestState(
748
                req_id=req_id,
749
                prompt_token_ids=new_req_data.prompt_token_ids,
750
                prompt_embeds=new_req_data.prompt_embeds,
751
                mm_features=new_req_data.mm_features,
752
                sampling_params=sampling_params,
753
                pooling_params=pooling_params,
754
                generator=generator,
755
756
                block_ids=new_req_data.block_ids,
                num_computed_tokens=new_req_data.num_computed_tokens,
757
                output_token_ids=[],
758
                lora_request=new_req_data.lora_request,
759
            )
760
761
            self.requests[req_id] = req_state

762
763
764
765
766
767
768
            if sampling_params and sampling_params.prompt_logprobs is not None:
                self.num_prompt_logprobs[req_id] = (
                    self.input_batch.vocab_size
                    if sampling_params.prompt_logprobs == -1
                    else sampling_params.prompt_logprobs
                )

769
            # Only relevant for models using M-RoPE (e.g, Qwen2-VL)
770
            if self.uses_mrope:
771
                self._init_mrope_positions(req_state)
772

773
            reqs_to_add.append(req_state)
774

775
        # Update the states of the running/resumed requests.
776
        is_last_rank = get_pp_group().is_last_rank
777
        req_data = scheduler_output.scheduled_cached_reqs
778
779
780
781
782

        # Wait until valid_sampled_tokens_count is copied to cpu,
        # then use it to update actual num_computed_tokens of each request.
        valid_sampled_token_count = self._get_valid_sampled_token_count()

783
        for i, req_id in enumerate(req_data.req_ids):
784
            req_state = self.requests[req_id]
785
786
            num_computed_tokens = req_data.num_computed_tokens[i]
            new_block_ids = req_data.new_block_ids[i]
787
            resumed_from_preemption = req_id in req_data.resumed_req_ids
788
            num_output_tokens = req_data.num_output_tokens[i]
789
            req_index = self.input_batch.req_id_to_index.get(req_id)
790

791
792
793
794
795
796
797
798
799
800
801
802
803
804
805
806
807
808
809
810
811
812
813
            # prev_num_draft_len is used in async scheduling mode with
            # spec decode. it indicates if need to update num_computed_tokens
            # of the request. for example:
            # fist step: num_computed_tokens = 0, spec_tokens = [],
            # prev_num_draft_len = 0.
            # second step: num_computed_tokens = 100(prompt lenth),
            # spec_tokens = [a,b], prev_num_draft_len = 0.
            # third step: num_computed_tokens = 100 + 2, spec_tokens = [c,d],
            # prev_num_draft_len = 2.
            # num_computed_tokens in first step and second step does't contain
            # the spec tokens length, but in third step it contains the
            # spec tokens length. we only need to update num_computed_tokens
            # when prev_num_draft_len > 0.
            if req_state.prev_num_draft_len:
                if req_index is None:
                    req_state.prev_num_draft_len = 0
                else:
                    assert self.input_batch.prev_req_id_to_index is not None
                    prev_req_index = self.input_batch.prev_req_id_to_index[req_id]
                    num_accepted = valid_sampled_token_count[prev_req_index] - 1
                    num_rejected = req_state.prev_num_draft_len - num_accepted
                    num_computed_tokens -= num_rejected
                    req_state.output_token_ids.extend([-1] * num_accepted)
814

815
            # Update the cached states.
816
            req_state.num_computed_tokens = num_computed_tokens
817
818
819
820
821
822
823
824

            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.
825
826
827
                num_new_tokens = (
                    num_computed_tokens + len(new_token_ids) - req_state.num_tokens
                )
828
829
830
831
                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:
832
                    req_state.output_token_ids.extend(new_token_ids[-num_new_tokens:])
833
834
835
836
837
            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:
838
839
840
841
                    end_idx = (
                        self.input_batch.num_prompt_tokens[req_index]
                        + num_output_tokens
                    )
842
843
                    self.input_batch.num_tokens[req_index] = end_idx
                    self.input_batch.num_tokens_no_spec[req_index] = end_idx
844

845
            # Update the block IDs.
846
            if not resumed_from_preemption:
847
848
                if new_block_ids is not None:
                    # Append the new blocks to the existing block IDs.
849
                    for block_ids, new_ids in zip(req_state.block_ids, new_block_ids):
850
                        block_ids.extend(new_ids)
851
            else:
852
                assert req_index is None
853
                assert new_block_ids is not None
854
855
                # The request is resumed from preemption.
                # Replace the existing block IDs with the new ones.
856
                req_state.block_ids = new_block_ids
857
858
859
860
861

            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.
862
863
864
865
866
867
868

                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:]

869
                reqs_to_add.append(req_state)
870
871
872
                continue

            # Update the persistent batch.
873
            self.input_batch.num_computed_tokens_cpu[req_index] = num_computed_tokens
874
            if new_block_ids is not None:
875
                self.input_batch.block_table.append_row(new_block_ids, req_index)
876
877
878
879
880
881
882

            # 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)
883
                self.input_batch.token_ids_cpu[
884
885
886
                    req_index, start_token_index:end_token_index
                ] = new_token_ids
                self.input_batch.num_tokens_no_spec[req_index] = end_token_index
887
                self.input_batch.num_tokens[req_index] = end_token_index
888

889
            # Add spec_token_ids to token_ids_cpu.
890
            spec_token_ids = scheduler_output.scheduled_spec_decode_tokens.get(
891
                req_id, []
892
            )
893
894
895
896
897
            num_spec_tokens = len(spec_token_ids)
            # For async scheduling, token_ids_cpu assigned from
            # spec_token_ids are placeholders and will be overwritten in
            # _prepare_input_ids.
            if num_spec_tokens:
898
899
900
                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[
901
902
                    req_index, start_index:end_token_index
                ] = spec_token_ids
903
904
                # NOTE(woosuk): `num_tokens` here may include spec tokens.
                self.input_batch.num_tokens[req_index] += num_spec_tokens
905
906
907
908
909
910

            # 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.
911
912
            self.input_batch.spec_token_ids[req_index].clear()
            self.input_batch.spec_token_ids[req_index].extend(spec_token_ids)
913

914
915
916
917
918
919
920
921
922
            # there are no draft tokens with async scheduling,
            # we clear the spec_decoding info in scheduler_output and
            # use normal sampling but rejection_sampling.
            if self.use_async_scheduling:
                req_state.prev_num_draft_len = num_spec_tokens
                if num_spec_tokens and self._draft_token_ids is None:
                    scheduler_output.total_num_scheduled_tokens -= num_spec_tokens
                    scheduler_output.num_scheduled_tokens[req_id] -= num_spec_tokens
                    scheduler_output.scheduled_spec_decode_tokens.pop(req_id, None)
923
924
        # Add the new or resumed requests to the persistent batch.
        # The smaller empty indices are filled first.
925
926
        for request in reqs_to_add:
            self.input_batch.add_request(request)
927

928
929
930
931
932
933
        # 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()
934

935
    def _update_states_after_model_execute(
936
937
        self, output_token_ids: torch.Tensor
    ) -> None:
938
939
940
941
942
943
944
945
946
947
948
949
        """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.
950
951
952
953
954
955
956
957
958
959
960
961
962
963
964
965
966
967
968
969
        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()
        )
970
971
972
        for i, num_tokens in enumerate(num_accepted_tokens):
            self.input_batch.num_accepted_tokens_cpu[i] = num_tokens

973
    def _init_mrope_positions(self, req_state: CachedRequestState):
974
975
        model = self.get_model()
        assert supports_mrope(model), "M-RoPE support is not implemented."
976
977
978
979
        assert req_state.prompt_token_ids is not None, (
            "M-RoPE requires prompt_token_ids to be available."
        )
        mrope_model = cast(SupportsMRoPE, model)
980
981

        req_state.mrope_positions, req_state.mrope_position_delta = (
982
            mrope_model.get_mrope_input_positions(
983
                req_state.prompt_token_ids,
984
                req_state.mm_features,
985
            )
986
        )
987

988
    def _extract_mm_kwargs(
989
        self,
990
991
        scheduler_output: "SchedulerOutput",
    ) -> BatchedTensorInputs:
992
        if not scheduler_output or not self.is_multimodal_raw_input_only_model:
993
            return {}
994

995
996
        mm_kwargs = list[MultiModalKwargsItem]()
        for req in scheduler_output.scheduled_new_reqs:
997
998
999
            for feature in req.mm_features:
                if feature.data is not None:
                    mm_kwargs.append(feature.data)
1000

1001
        # Input all modalities at once
1002
        model = cast(SupportsMultiModal, self.model)
1003
1004
        mm_kwargs_combined: BatchedTensorInputs = {}
        for _, _, mm_kwargs_group in group_mm_kwargs_by_modality(
1005
1006
1007
1008
            mm_kwargs,
            device=self.device,
            pin_memory=self.pin_memory,
            merge_by_field_config=model.merge_by_field_config,
1009
            multimodal_cpu_fields=model.multimodal_cpu_fields,
1010
1011
        ):
            mm_kwargs_combined.update(mm_kwargs_group)
1012

1013
        return mm_kwargs_combined
1014

1015
    def _dummy_mm_kwargs(self, num_seqs: int) -> BatchedTensorInputs:
1016
        if not self.is_multimodal_raw_input_only_model:
1017
            return {}
1018

1019
1020
1021
1022
1023
        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)
1024

1025
1026
1027
    def _get_cumsum_and_arange(
        self,
        num_tokens: np.ndarray,
1028
        cumsum_dtype: np.dtype | None = None,
1029
1030
1031
1032
1033
1034
1035
1036
1037
1038
1039
1040
1041
1042
1043
1044
    ) -> 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

1045
    def _prepare_input_ids(
1046
1047
1048
1049
        self,
        scheduler_output: "SchedulerOutput",
        total_num_scheduled_tokens: int,
        cu_num_tokens: np.ndarray,
1050
    ) -> None:
1051
        """Prepare the input IDs for the current batch.
1052

1053
1054
1055
1056
1057
1058
1059
        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)
1060
1061
1062
            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)
1063
1064
1065
1066
1067
1068
1069
            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
1070
1071
1072
1073
        sample_flattened_indices: list[int] = []
        spec_flattened_indices: list[int] = []
        prev_common_req_indices: list[int] = []
        prev_draft_token_indices: list[int] = []
1074
1075
        indices_match = True
        max_flattened_index = -1
1076
1077
1078
        total_num_spec_tokens = 0
        scheduled_spec_tokens = scheduler_output.scheduled_spec_decode_tokens

1079
1080
1081
1082
1083
        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.
1084
1085
                draft_len = len(scheduled_spec_tokens.get(req_id, ()))
                total_num_spec_tokens += draft_len
1086
                flattened_index = cu_num_tokens[cur_index].item() - 1
1087
1088
1089
1090
1091
1092
1093
1094
1095
1096
1097
1098
1099
1100
1101
                # example: cu_num_tokens = [2, 5, 8], draft_tokens = [1, 2, 2]
                # sample_flattened_indices = [0, 2, 5]
                # spec_flattened_indices = [1,   3, 4,    6, 7]
                sample_flattened_indices.append(flattened_index - draft_len)
                spec_flattened_indices.extend(
                    range(flattened_index - draft_len + 1, flattened_index + 1)
                )
                start = prev_index * self.num_spec_tokens
                # prev_draft_token_indices is used to find which draft_tokens_id
                # should be copied to input_ids
                # example: prev draft_tokens_id [[1,2], [3,4], [5, 6]]
                # flatten draft_tokens_id [1,2,3,4,5,6]
                # draft_len of each request [1, 2, 1]
                # then prev_draft_token_indices is [0,   2, 3,   4]
                prev_draft_token_indices.extend(range(start, start + draft_len))
1102
                indices_match &= prev_index == flattened_index
1103
                max_flattened_index = max(max_flattened_index, flattened_index)
1104
1105
1106
        num_commmon_tokens = len(sample_flattened_indices)
        total_without_spec = total_num_scheduled_tokens - total_num_spec_tokens
        if num_commmon_tokens < total_without_spec:
1107
1108
1109
            # 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)
1110
1111
1112
            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)
1113
1114
        if num_commmon_tokens == 0:
            # No requests in common with the previous iteration
1115
            # So input_ids.cpu will have all the input ids.
1116
1117
1118
1119
1120
1121
1122
            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_(
1123
1124
1125
                self.input_batch.prev_sampled_token_ids[:num_commmon_tokens, 0],
                non_blocking=True,
            )
1126
1127
            if self.enable_prompt_embeds:
                self.is_token_ids.gpu[:num_commmon_tokens] = True
1128
            return
1129
        # Upload the index tensors asynchronously so the scatter can be non-blocking.
1130
1131
        sampled_tokens_index_tensor = torch.tensor(
            sample_flattened_indices, dtype=torch.int64, pin_memory=self.pin_memory
1132
        ).to(self.device, non_blocking=True)
1133
        prev_common_req_indices_tensor = torch.tensor(
1134
1135
            prev_common_req_indices, dtype=torch.int64, pin_memory=self.pin_memory
        ).to(self.device, non_blocking=True)
1136
1137
        self.input_ids.gpu.scatter_(
            dim=0,
1138
            index=sampled_tokens_index_tensor,
1139
            src=self.input_batch.prev_sampled_token_ids[
1140
1141
1142
                prev_common_req_indices_tensor, 0
            ],
        )
1143

1144
1145
1146
1147
1148
1149
1150
1151
1152
1153
1154
1155
1156
1157
1158
1159
1160
1161
1162
1163
1164
1165
1166
        # Scatter the draft tokens after the sampled tokens are scattered.
        if self._draft_token_ids is None or not spec_flattened_indices:
            return

        assert isinstance(self._draft_token_ids, torch.Tensor)
        draft_tokens_index_tensor = torch.tensor(
            spec_flattened_indices, dtype=torch.int64, pin_memory=self.pin_memory
        ).to(self.device, non_blocking=True)
        prev_draft_token_indices_tensor = torch.tensor(
            prev_draft_token_indices, dtype=torch.int64, pin_memory=self.pin_memory
        ).to(self.device, non_blocking=True)

        # because input_ids dtype is torch.int32,
        # so convert draft_token_ids to torch.int32 here.
        draft_token_ids = self._draft_token_ids.to(dtype=torch.int32)
        self._draft_token_ids = None

        self.input_ids.gpu.scatter_(
            dim=0,
            index=draft_tokens_index_tensor,
            src=draft_token_ids.flatten()[prev_draft_token_indices_tensor],
        )

1167
1168
    def _get_encoder_seq_lens(
        self,
1169
        scheduled_encoder_inputs: dict[str, list[int]],
1170
1171
        kv_cache_spec: KVCacheSpec,
        num_reqs: int,
1172
    ) -> np.ndarray | None:
1173
1174
1175
1176
1177
1178
        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)
1179
        for req_id in scheduled_encoder_inputs:
1180
1181
1182
1183
1184
            req_index = self.input_batch.req_id_to_index[req_id]
            encoder_seq_lens[req_index] = self.max_encoder_len

        return encoder_seq_lens

1185
    def _prepare_inputs(
1186
1187
1188
1189
        self,
        scheduler_output: "SchedulerOutput",
        num_scheduled_tokens: np.ndarray,
        max_num_scheduled_tokens: int,
1190
1191
    ) -> tuple[
        torch.Tensor,
1192
1193
1194
        SpecDecodeMetadata | None,
        UBatchSlices | None,
        torch.Tensor | None,
1195
    ]:
1196
1197
        """
        :return: tuple[
1198
            logits_indices, spec_decode_metadata,
1199
            ubatch_slices, num_tokens_across_dp,
1200
1201
        ]
        """
1202
1203
1204
1205
1206
1207
1208
        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.
1209
        self.input_batch.block_table.commit_block_table(num_reqs)
1210
1211
1212

        # Get request indices.
        # E.g., [2, 5, 3] -> [0, 0, 1, 1, 1, 1, 1, 2, 2, 2]
1213
        req_indices = np.repeat(self.arange_np[:num_reqs], num_scheduled_tokens)
1214

1215
1216
        # cu_num_tokens: [2, 5, 3] -> [2, 7, 10]
        # arange: [0, 1, 0, 1, 2, 3, 4, 0, 1, 2]
1217
        cu_num_tokens, arange = self._get_cumsum_and_arange(num_scheduled_tokens)
1218
1219

        # Get positions.
1220
        positions_np = self.positions.np[:total_num_scheduled_tokens]
1221
1222
1223
1224
1225
        np.add(
            self.input_batch.num_computed_tokens_cpu[req_indices],
            arange,
            out=positions_np,
        )
1226

1227
1228
        # Calculate M-RoPE positions.
        # Only relevant for models using M-RoPE (e.g, Qwen2-VL)
1229
        if self.uses_mrope:
1230
1231
            self._calc_mrope_positions(scheduler_output)

1232
1233
1234
1235
        # 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.
1236
1237
1238
        token_indices = (
            positions_np + req_indices * self.input_batch.token_ids_cpu.shape[1]
        )
1239
        token_indices_tensor = torch.from_numpy(token_indices)
1240

1241
1242
1243
        # 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.
1244
1245
1246
1247
1248
1249
        torch.index_select(
            self.input_batch.token_ids_cpu_tensor.flatten(),
            0,
            token_indices_tensor,
            out=self.input_ids.cpu[:total_num_scheduled_tokens],
        )
1250
        if self.enable_prompt_embeds:
1251
            is_token_ids = self.input_batch.is_token_ids_tensor.flatten()
1252
1253
1254
1255
            torch.index_select(
                is_token_ids,
                0,
                token_indices_tensor,
1256
1257
                out=self.is_token_ids.cpu[:total_num_scheduled_tokens],
            )
1258
1259
1260
1261
1262
1263
1264
1265
1266
1267
1268
1269
1270
1271
1272
1273
1274
1275
1276
1277
1278
1279
1280
1281
1282
1283
1284
1285
1286
1287
1288
1289
1290

        # 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:
1291
1292
1293
                    self.inputs_embeds.cpu[
                        output_idx : output_idx + actual_num_sched
                    ].copy_(req_embeds[start_pos:actual_end])
1294
1295

                output_idx += num_sched
1296

1297
1298
        self.input_batch.block_table.compute_slot_mapping(req_indices, positions_np)
        self.input_batch.block_table.commit_slot_mapping(total_num_scheduled_tokens)
1299
1300

        # Prepare the attention metadata.
1301
        self.query_start_loc.np[0] = 0
1302
        self.query_start_loc.np[1 : num_reqs + 1] = cu_num_tokens
1303
1304
        # Note: pad query_start_loc to be non-decreasing, as kernels
        # like FlashAttention requires that
1305
        self.query_start_loc.np[num_reqs + 1 :].fill(cu_num_tokens[-1])
1306
        self.query_start_loc.copy_to_gpu()
1307
        query_start_loc = self.query_start_loc.gpu[: num_reqs + 1]
1308

1309
        num_tokens_unpadded = scheduler_output.total_num_scheduled_tokens
1310
        num_tokens_padded = self._get_num_input_tokens(num_tokens_unpadded)
1311
1312
1313
        uniform_decode = (
            max_num_scheduled_tokens == self.uniform_decode_query_len
        ) and (total_num_scheduled_tokens == num_reqs * max_num_scheduled_tokens)
1314
1315
1316
1317
1318
1319
1320

        # 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

1321
        ubatch_slices, num_tokens_across_dp = coordinate_batch_across_dp(
1322
1323
1324
1325
1326
1327
1328
            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,
1329
        )
1330

1331
        self.seq_lens.np[:num_reqs] = (
1332
1333
            self.input_batch.num_computed_tokens_cpu[:num_reqs] + num_scheduled_tokens
        )
1334
        # Fill unused with 0 for full cuda graph mode.
1335
1336
        self.seq_lens.np[num_reqs:].fill(0)
        self.seq_lens.copy_to_gpu()
1337

1338
        num_tokens = [self.requests[r].num_tokens for r in self.input_batch.req_ids]
1339
1340
1341
1342
1343
1344
1345
        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)
1346
1347
1348
        self.discard_request_indices.np[: self.num_discarded_requests] = (
            discard_request_indices
        )
1349
1350
1351

        self.discard_request_indices.copy_to_gpu(self.num_discarded_requests)

1352
        # Copy the tensors to the GPU.
1353
1354
1355
1356
1357
        self._prepare_input_ids(
            scheduler_output,
            total_num_scheduled_tokens,
            cu_num_tokens,
        )
1358

1359
        if self.uses_mrope:
1360
            # Only relevant for models using M-RoPE (e.g, Qwen2-VL)
1361
1362
            self.mrope_positions.gpu[:, :total_num_scheduled_tokens].copy_(
                self.mrope_positions.cpu[:, :total_num_scheduled_tokens],
1363
1364
                non_blocking=True,
            )
1365
1366
        else:
            # Common case (1D positions)
1367
            self.positions.copy_to_gpu(total_num_scheduled_tokens)
1368

1369
        use_spec_decode = len(scheduler_output.scheduled_spec_decode_tokens) > 0
1370
1371
1372
1373
1374
1375
1376
        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
1377
            num_draft_tokens = None
1378
            spec_decode_metadata = None
1379
            num_sampled_tokens = np.ones(num_reqs, dtype=np.int32)
1380
1381
1382
1383
1384
        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)
1385
1386
1387
            # 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)
1388
1389
1390
1391
            for (
                req_id,
                draft_token_ids,
            ) in scheduler_output.scheduled_spec_decode_tokens.items():
1392
1393
                req_idx = self.input_batch.req_id_to_index[req_id]
                num_draft_tokens[req_idx] = len(draft_token_ids)
1394
1395
1396
1397
1398
1399
1400
1401
                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
                )
1402
            spec_decode_metadata = self._calc_spec_decode_metadata(
1403
1404
                num_draft_tokens, cu_num_tokens
            )
1405
            logits_indices = spec_decode_metadata.logits_indices
1406
            num_sampled_tokens = num_draft_tokens + 1
1407
            # For DECODE only cuda graph of some attention backends (e.g., GDN).
1408
            self.num_decode_draft_tokens.np[:num_reqs] = num_decode_draft_tokens
1409
1410
            self.num_decode_draft_tokens.np[num_reqs:].fill(-1)
            self.num_decode_draft_tokens.copy_to_gpu()
1411

1412
1413
1414
1415
1416
        # Hot-Swap lora model
        if self.lora_config:
            assert (
                np.sum(num_sampled_tokens)
                <= self.vllm_config.scheduler_config.max_num_batched_tokens
1417
            )
1418
1419
1420
1421
1422
1423
1424
1425
1426
1427
1428
1429
1430
1431
1432
1433
1434
1435
1436
1437
1438
1439
1440
1441
1442
1443
1444
            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
1445
        num_logits_indices = None
1446
1447
1448
1449
1450
1451
        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
                )
1452

1453
1454
1455
1456
1457
1458
        # 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,
1459
                self.parallel_config.cp_kv_cache_interleave_size,
1460
1461
1462
            )
            self.dcp_local_seq_lens.copy_to_gpu(num_reqs)

1463
1464
1465
        attn_metadata: PerLayerAttnMetadata = {}
        if ubatch_slices is not None:
            attn_metadata = [dict() for _ in range(len(ubatch_slices))]
1466

1467
1468
        # Used in the below loop
        query_start_loc = self.query_start_loc.gpu[: num_reqs + 1]
1469
        query_start_loc_cpu = self.query_start_loc.cpu[: num_reqs + 1]
1470
        seq_lens = self.seq_lens.gpu[:num_reqs]
1471
        seq_lens_cpu = self.seq_lens.cpu[:num_reqs]
1472
1473
1474
        num_computed_tokens_cpu = self.input_batch.num_computed_tokens_cpu_tensor[
            :num_reqs
        ]
1475
1476
1477
1478
1479
1480

        dcp_local_seq_lens, dcp_local_seq_lens_cpu = None, None
        if self.dcp_world_size > 1:
            dcp_local_seq_lens = self.dcp_local_seq_lens.gpu[:num_reqs]
            dcp_local_seq_lens_cpu = self.dcp_local_seq_lens.cpu[:num_reqs]

1481
        spec_decode_common_attn_metadata = None
1482
1483
1484
1485
1486
1487
1488
1489
1490

        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()

1491
1492
        if use_spec_decode:
            self.num_accepted_tokens.np[:num_reqs] = (
1493
1494
                self.input_batch.num_accepted_tokens_cpu[:num_reqs]
            )
1495
1496
            self.num_accepted_tokens.np[num_reqs:].fill(1)
            self.num_accepted_tokens.copy_to_gpu()
1497

1498
1499
        # Prepare the attention metadata for each KV cache group and make layers
        # in the same group share the same metadata.
1500
        for kv_cache_gid, kv_cache_group in enumerate(
1501
1502
            self.kv_cache_config.kv_cache_groups
        ):
1503
            encoder_seq_lens = self._get_encoder_seq_lens(
1504
1505
1506
                scheduled_encoder_inputs or {},
                kv_cache_group.kv_cache_spec,
                num_reqs,
1507
            )
1508

1509
            if isinstance(kv_cache_group.kv_cache_spec, EncoderOnlyAttentionSpec):
1510
1511
1512
1513
1514
                # 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,
1515
1516
1517
                    device=self.device,
                )
                slot_mapping = torch.zeros(
1518
                    (total_num_scheduled_tokens,),
1519
1520
1521
                    dtype=torch.int64,
                    device=self.device,
                )
1522
            else:
1523
                blk_table = self.input_batch.block_table[kv_cache_gid]
1524
                blk_table_tensor = blk_table.get_device_tensor(num_reqs)
1525
                slot_mapping = blk_table.slot_mapping.gpu[:total_num_scheduled_tokens]
1526
1527
1528

                # Fill unused with -1. Needed for reshape_and_cache in full cuda
                # graph mode.
1529
                blk_table.slot_mapping.gpu[total_num_scheduled_tokens:].fill_(-1)
1530

1531
            common_attn_metadata = CommonAttentionMetadata(
1532
1533
1534
1535
1536
                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,
1537
1538
1539
                num_reqs=num_reqs,
                num_actual_tokens=total_num_scheduled_tokens,
                max_query_len=max_num_scheduled_tokens,
1540
                max_seq_len=max_seq_len,
1541
1542
                block_table_tensor=blk_table_tensor,
                slot_mapping=slot_mapping,
1543
                logits_indices_padded=logits_indices_padded,
1544
                num_logits_indices=num_logits_indices,
1545
                causal=True,
1546
                encoder_seq_lens=encoder_seq_lens,
1547
                dcp_local_seq_lens=dcp_local_seq_lens,
1548
                dcp_local_seq_lens_cpu=dcp_local_seq_lens_cpu,
1549
1550
            )

1551
            if self.speculative_config and spec_decode_common_attn_metadata is None:
1552
                if isinstance(self.drafter, EagleProposer):
1553
                    if self.drafter.attn_layer_names[0] in kv_cache_group.layer_names:
1554
1555
1556
                        spec_decode_common_attn_metadata = common_attn_metadata
                else:
                    spec_decode_common_attn_metadata = common_attn_metadata
1557

1558
1559
1560
1561
1562
1563
            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
                )
1564
                builder = attn_group.get_metadata_builder()
1565

1566
                extra_attn_metadata_args = {}
1567
                if use_spec_decode and isinstance(builder, GDNAttentionMetadataBuilder):
1568
                    extra_attn_metadata_args = dict(
1569
1570
1571
1572
                        num_accepted_tokens=self.num_accepted_tokens.gpu[:num_reqs],
                        num_decode_draft_tokens_cpu=self.num_decode_draft_tokens.cpu[
                            :num_reqs
                        ],
1573
1574
                    )

1575
1576
                if ubatch_slices is not None:
                    common_attn_metadata_list = split_attn_metadata(
1577
1578
                        ubatch_slices, common_attn_metadata
                    )
1579
                    for ubid, common_attn_metadata in enumerate(
1580
1581
                        common_attn_metadata_list
                    ):
1582
1583
1584
1585
1586
1587
1588
1589
1590
1591
1592
                        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:
1593
1594
1595
1596
                            assert type(attn_metadata) is list
                            attn_metadata[ubid][layer_name] = attn_metadata_i
                else:
                    assert isinstance(attn_metadata, dict)
1597
1598
1599
1600
1601
1602
1603
1604
1605
1606
                    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,
                        )
1607
1608
                    for layer_name in attn_group.layer_names:
                        attn_metadata[layer_name] = attn_metadata_i
1609

1610
        return attn_metadata, spec_decode_common_attn_metadata
1611

1612
1613
1614
1615
1616
1617
1618
1619
1620
1621
    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
        """
1622

1623
1624
1625
1626
1627
1628
1629
1630
1631
1632
1633
1634
1635
1636
1637
1638
1639
1640
1641
1642
1643
1644
        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
1645

1646
1647
1648
1649
    def _compute_cascade_attn_prefix_len(
        self,
        num_scheduled_tokens: np.ndarray,
        num_common_prefix_blocks: int,
1650
1651
        kv_cache_spec: KVCacheSpec,
        attn_metadata_builder: AttentionMetadataBuilder,
1652
1653
1654
1655
1656
1657
1658
1659
1660
1661
1662
1663
1664
1665
1666
1667
1668
1669
    ) -> 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.
        """
1670

1671
        common_prefix_len = num_common_prefix_blocks * kv_cache_spec.block_size
1672
1673
1674
1675
1676
1677
1678
1679
1680
1681
1682
1683
1684
1685
1686
1687
1688
1689
1690
1691
1692
1693
1694
1695
1696
1697
1698
1699
1700
1701
1702
1703
1704
1705
1706
1707
1708
        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]
1709
        # Request 3's num_computed_tokens: 3 (i.e., [A, B, C])
1710
1711
1712
1713
1714
1715
1716
        # 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(
1717
1718
            common_prefix_len, self.input_batch.num_computed_tokens_cpu[:num_reqs].min()
        )
1719
        # common_prefix_len should be a multiple of the block size.
1720
1721
1722
1723
1724
1725
1726
1727
1728
1729
1730
        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
        )
1731
1732
        assert isinstance(kv_cache_spec, AttentionSpec)
        use_cascade = attn_metadata_builder.use_cascade_attention(
1733
1734
1735
            common_prefix_len=common_prefix_len,
            query_lens=num_scheduled_tokens,
            num_query_heads=self.num_query_heads,
1736
            num_kv_heads=kv_cache_spec.num_kv_heads,
1737
            use_alibi=self.use_alibi,
1738
            use_sliding_window=use_sliding_window,
1739
            use_local_attention=use_local_attention,
1740
            num_sms=self.num_sms,
1741
            dcp_world_size=self.dcp_world_size,
1742
1743
1744
        )
        return common_prefix_len if use_cascade else 0

1745
1746
    def _calc_mrope_positions(self, scheduler_output: "SchedulerOutput"):
        mrope_pos_ptr = 0
1747
        for index, req_id in enumerate(self.input_batch.req_ids):
1748
1749
1750
            req = self.requests[req_id]
            assert req.mrope_positions is not None

1751
1752
            num_computed_tokens = self.input_batch.num_computed_tokens_cpu[index]
            num_scheduled_tokens = scheduler_output.num_scheduled_tokens[req_id]
1753
            num_prompt_tokens = length_from_prompt_token_ids_or_embeds(
1754
1755
                req.prompt_token_ids, req.prompt_embeds
            )
1756
1757

            if num_computed_tokens + num_scheduled_tokens > num_prompt_tokens:
1758
1759
                prompt_part_len = max(0, num_prompt_tokens - num_computed_tokens)
                completion_part_len = max(0, num_scheduled_tokens - prompt_part_len)
1760
1761
1762
1763
1764
1765
1766
1767
1768
1769
1770
1771
1772
            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

1773
1774
1775
                self.mrope_positions.cpu[:, dst_start:dst_end] = req.mrope_positions[
                    :, src_start:src_end
                ]
1776
1777
1778
1779
1780
1781
1782
                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

1783
                assert req.mrope_position_delta is not None
1784
                MRotaryEmbedding.get_next_input_positions_tensor(
1785
                    out=self.mrope_positions.np,
1786
1787
1788
1789
1790
                    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,
                )
1791
1792
1793

                mrope_pos_ptr += completion_part_len

1794
1795
    def _calc_spec_decode_metadata(
        self,
1796
1797
1798
1799
1800
1801
1802
1803
1804
1805
1806
1807
1808
1809
1810
1811
        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
1812
1813
1814
1815

        # 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(
1816
1817
            num_sampled_tokens, cumsum_dtype=np.int32
        )
1818
        # Step 2. [0, 0, 0, 0, 103, 104, 104, 104, 206, 207, 207]
1819
        logits_indices = np.repeat(
1820
1821
            cu_num_scheduled_tokens - num_sampled_tokens, num_sampled_tokens
        )
1822
        # Step 3. [0, 1, 2, 3, 103, 104, 105, 106, 206, 207, 208]
1823
1824
1825
1826
1827
1828
        logits_indices += arange

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

        # Compute the draft logits indices.
1829
1830
1831
        # 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(
1832
1833
            num_draft_tokens, cumsum_dtype=np.int32
        )
1834
1835
        # [0, 0, 0, 5, 5, 9]
        target_logits_indices = np.repeat(
1836
1837
            cu_num_sampled_tokens - num_sampled_tokens, num_draft_tokens
        )
1838
1839
1840
1841
1842
        # [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(
1843
1844
            self.device, non_blocking=True
        )
1845
1846
1847
        cu_num_sampled_tokens = torch.from_numpy(cu_num_sampled_tokens).to(
            self.device, non_blocking=True
        )
1848
1849
1850
        logits_indices = torch.from_numpy(logits_indices).to(
            self.device, non_blocking=True
        )
1851
        target_logits_indices = torch.from_numpy(target_logits_indices).to(
1852
1853
            self.device, non_blocking=True
        )
1854
        bonus_logits_indices = torch.from_numpy(bonus_logits_indices).to(
1855
1856
            self.device, non_blocking=True
        )
1857

1858
1859
        # Compute the draft token ids.
        # draft_token_indices:      [  1,   2,   3, 105, 106, 208]
1860
        draft_token_ids = self.input_ids.gpu[logits_indices]
1861
1862
        draft_token_ids = draft_token_ids[target_logits_indices + 1]

1863
        return SpecDecodeMetadata(
1864
1865
1866
            draft_token_ids=draft_token_ids,
            num_draft_tokens=num_draft_tokens.tolist(),
            cu_num_draft_tokens=cu_num_draft_tokens,
1867
            cu_num_sampled_tokens=cu_num_sampled_tokens,
1868
1869
1870
1871
1872
            target_logits_indices=target_logits_indices,
            bonus_logits_indices=bonus_logits_indices,
            logits_indices=logits_indices,
        )

1873
1874
1875
1876
1877
1878
1879
    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
1880
        self.kv_sharing_fast_prefill_logits_indices[:num_logits].copy_(logits_indices)
1881
1882
1883
1884
1885
        # 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_(
1886
1887
1888
1889
1890
1891
            logits_indices[-1].item()
        )
        if (
            self.compilation_config.cudagraph_mode != CUDAGraphMode.NONE
            and num_logits <= self.cudagraph_batch_sizes[-1]
        ):
1892
1893
1894
1895
1896
            # 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
1897
1898
1899
        logits_indices_padded = self.kv_sharing_fast_prefill_logits_indices[
            :num_logits_padded
        ]
1900
1901
        return logits_indices_padded

1902
1903
1904
1905
1906
1907
1908
1909
    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
1910
                inputs.
1911
1912
1913
1914
1915
1916

        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
        """
1917
1918
        scheduled_encoder_inputs = scheduler_output.scheduled_encoder_inputs
        if not scheduled_encoder_inputs:
1919
            return [], []
1920
        # Batch the multi-modal inputs.
1921
        mm_kwargs = list[MultiModalKwargsItem]()
1922
1923
        # list of tuple (mm_hash, position_info)
        mm_hashes_pos = list[tuple[str, PlaceholderRange]]()
1924
1925
        for req_id, encoder_input_ids in scheduled_encoder_inputs.items():
            req_state = self.requests[req_id]
1926
1927

            for mm_input_id in encoder_input_ids:
1928
                mm_feature = req_state.mm_features[mm_input_id]
1929
1930
                if mm_feature.data is None:
                    continue
1931
1932
1933
                mm_hash = mm_feature.identifier
                mm_kwargs.append(mm_feature.data)
                mm_hashes_pos.append((mm_hash, mm_feature.mm_position))
1934

1935
1936
        return mm_kwargs, mm_hashes_pos

1937
1938
1939
    def _execute_mm_encoder(
        self, scheduler_output: "SchedulerOutput"
    ) -> list[torch.Tensor]:
1940
1941
        # Batch the multi-modal inputs using the helper method.
        mm_kwargs, mm_hashes_pos = self._batch_mm_kwargs_from_scheduler(
1942
1943
            scheduler_output
        )
1944
1945

        if not mm_kwargs:
1946
            return []
1947

1948
1949
1950
1951
1952
1953
1954
        # 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.
1955
        model = cast(SupportsMultiModal, self.model)
1956
        encoder_outputs: list[torch.Tensor] = []
1957
        for modality, num_items, mm_kwargs_group in group_mm_kwargs_by_modality(
1958
1959
1960
1961
            mm_kwargs,
            device=self.device,
            pin_memory=self.pin_memory,
            merge_by_field_config=model.merge_by_field_config,
1962
            multimodal_cpu_fields=model.multimodal_cpu_fields,
1963
        ):
1964
            curr_group_outputs: list[torch.Tensor] = []
1965
1966

            # EVS-related change.
1967
            # (ekhvedchenia): Temporary hack to limit peak memory usage when
1968
            # processing multimodal data. This solves the issue with scheduler
1969
1970
1971
1972
            # 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)
1973
1974
1975
1976
1977
1978
1979
1980
1981
1982
1983
1984
1985
1986
1987
1988
            # 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,
1989
                            multimodal_cpu_fields=model.multimodal_cpu_fields,
1990
                        )
1991
                    )
1992

1993
                    micro_batch_outputs = model.embed_multimodal(
1994
1995
                        **micro_batch_mm_inputs
                    )
1996
1997
1998
1999
2000
2001
2002
2003
2004
2005

                    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.
2006
                curr_group_outputs = model.embed_multimodal(**mm_kwargs_group)  # type: ignore[assignment]
2007

2008
2009
            sanity_check_mm_encoder_outputs(
                curr_group_outputs,
2010
                expected_num_items=num_items,
2011
            )
2012
            encoder_outputs.extend(curr_group_outputs)
2013

2014
2015
2016
        # 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(
2017
2018
2019
                output,
                is_embed=pos_info.is_embed,
            )
2020
2021
            logger.debug("Finish execute for mm hash %s", mm_hash)
            self.maybe_save_ec_to_connector(self.encoder_cache, mm_hash)
2022

2023
2024
        return encoder_outputs

2025
    def _gather_mm_embeddings(
2026
2027
        self,
        scheduler_output: "SchedulerOutput",
2028
        shift_computed_tokens: int = 0,
2029
2030
2031
2032
2033
2034
2035
2036
    ) -> 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
2037
        should_sync_mrope_positions = False
2038

2039
        for req_id in self.input_batch.req_ids:
2040
2041
            mm_embeds_req: list[torch.Tensor] = []

2042
            num_scheduled_tokens = scheduler_output.num_scheduled_tokens[req_id]
2043
            req_state = self.requests[req_id]
2044
            num_computed_tokens = req_state.num_computed_tokens + shift_computed_tokens
2045

2046
2047
            for mm_feature in req_state.mm_features:
                pos_info = mm_feature.mm_position
2048
2049
                start_pos = pos_info.offset
                num_encoder_tokens = pos_info.length
2050
2051
2052
2053
2054
2055
2056
2057
2058
2059
2060
2061
2062
2063
2064
2065

                # 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,
2066
2067
                    num_encoder_tokens,
                )
2068
                assert start_idx < end_idx
2069

2070
                mm_hash = mm_feature.identifier
2071
                encoder_output = self.encoder_cache.get(mm_hash, None)
2072
                assert encoder_output is not None, f"Encoder cache miss for {mm_hash}."
2073
2074
2075
2076

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

2077
                req_start_pos = req_start_idx + start_pos - num_computed_tokens
2078
2079
2080
                is_mm_embed[req_start_pos + start_idx : req_start_pos + end_idx] = (
                    True if is_embed is None else is_embed
                )
2081

2082
2083
2084
2085
                mm_embeds_item = gather_mm_placeholders(
                    encoder_output[start_idx:end_idx],
                    is_embed=is_embed,
                )
2086
2087
2088
                mm_embeds_req.append(mm_embeds_item)

            if self.is_multimodal_pruning_enabled and self.uses_mrope:
2089
                assert req_state.mrope_positions is not None
2090
2091
2092
2093
2094
2095
2096
                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,
2097
2098
                    )
                )
2099
2100
2101
2102
                req_state.mrope_positions.copy_(new_mrope_positions)
                req_state.mrope_position_delta = new_delta

            mm_embeds.extend(mm_embeds_req)
2103
2104
2105
            req_start_idx += num_scheduled_tokens

        is_mm_embed = self.is_mm_embed.copy_to_gpu(total_num_scheduled_tokens)
2106
2107
2108

        if should_sync_mrope_positions:
            self._calc_mrope_positions(scheduler_output)
2109
            self.mrope_positions.copy_to_gpu(total_num_scheduled_tokens)
2110

2111
        return mm_embeds, is_mm_embed
2112

2113
    def get_model(self) -> nn.Module:
2114
        # get raw model out of the cudagraph wrapper.
2115
        if isinstance(self.model, (CUDAGraphWrapper, UBatchWrapper)):
2116
            return self.model.unwrap()
2117
2118
        return self.model

2119
2120
2121
2122
2123
2124
2125
2126
2127
2128
2129
2130
2131
2132
2133
    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

2134
2135
2136
2137
2138
    def get_supported_pooling_tasks(self) -> list[PoolingTask]:
        model = self.get_model()
        if not is_pooling_model(model):
            return []

2139
2140
        supported_tasks = list(model.pooler.get_supported_tasks())

2141
        if self.scheduler_config.enable_chunked_prefill:
2142
2143
2144
2145
            if "token_embed" in supported_tasks:
                supported_tasks.remove("token_embed")
            if "token_classify" in supported_tasks:
                supported_tasks.remove("token_classify")
2146

2147
2148
            logger.debug_once(
                "Chunked prefill is not supported with "
2149
2150
                "token_embed and token_classify tasks "
                "which using ALL pooling. "
2151
2152
2153
                "Please turn off chunked prefill by "
                "`--no-enable-chunked-prefill` before using it."
            )
2154
2155
2156
2157
2158

        if "score" in supported_tasks:
            num_labels = getattr(self.model_config.hf_config, "num_labels", 0)
            if num_labels != 1:
                supported_tasks.remove("score")
2159
                logger.debug_once("Score API is only enabled for num_labels == 1.")
2160
2161

        return supported_tasks
2162

2163
2164
2165
2166
2167
2168
2169
2170
2171
2172
    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)

2173
    def sync_and_slice_intermediate_tensors(
2174
2175
        self,
        num_tokens: int,
2176
        intermediate_tensors: IntermediateTensors | None,
2177
2178
        sync_self: bool,
    ) -> IntermediateTensors:
2179
2180
2181
        assert self.intermediate_tensors is not None

        tp = self.vllm_config.parallel_config.tensor_parallel_size
2182
        is_rs = is_residual_scattered_for_sp(self.vllm_config, num_tokens)
2183
2184
2185
2186
2187
2188

        # 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():
2189
                is_scattered = k == "residual" and is_rs
2190
                copy_len = num_tokens // tp if is_scattered else num_tokens
2191
                self.intermediate_tensors[k][:copy_len].copy_(
2192
2193
2194
2195
2196
2197
2198
2199
2200
2201
2202
2203
2204
                    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:
2205
2206
2207
2208
2209
2210
2211
        """
        Step for the EPLB (Expert Parallelism Load Balancing) state.
        """
        if not self.parallel_config.enable_eplb:
            return

        assert self.eplb_state is not None
2212
2213
        model = self.get_model()
        assert is_mixture_of_experts(model)
2214
2215
2216
        self.eplb_state.step(
            is_dummy,
            is_profile,
2217
            log_stats=self.parallel_config.eplb_config.log_balancedness,
2218
2219
        )

2220
2221
2222
2223
    # 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)
2224
2225
    @staticmethod
    def pad_out_ubatch_slice(ubatch_slices: UBatchSlices, num_total_tokens: int):
2226
2227
2228
2229
2230
2231
        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
        )
2232

2233
2234
2235
2236
2237
2238
    def _pool(
        self,
        hidden_states: torch.Tensor,
        num_scheduled_tokens: int,
        num_scheduled_tokens_np: np.ndarray,
    ) -> ModelRunnerOutput:
2239
2240
2241
        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"
        )
2242

2243
        hidden_states = hidden_states[:num_scheduled_tokens]
2244
        pooling_metadata = self.input_batch.get_pooling_metadata()
2245
2246
2247
2248
        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]
2249

2250
2251
2252
2253
2254
2255
2256
2257
2258
2259
        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()
2260

2261
        pooler_output: list[torch.Tensor | None] = []
2262
        for raw_output, seq_len, prompt_len in zip(
2263
2264
            raw_pooler_output, seq_lens_cpu, pooling_metadata.prompt_lens
        ):
2265
            output = raw_output if seq_len == prompt_len else None
2266
            pooler_output.append(output)
2267
2268
2269
2270
2271
2272
2273
2274
2275
2276

        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,
        )

2277
    def _get_num_input_tokens(self, num_scheduled_tokens: int) -> int:
2278
2279
2280
2281
2282
2283
        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]
        ):
2284
2285
2286
2287
2288
2289
2290
2291
            # 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
2292
2293
2294
2295
        if (
            self.compilation_config.pass_config.enable_sequence_parallelism
            and tp_size > 1
        ):
2296
2297
2298
            return round_up(num_scheduled_tokens, tp_size)
        return num_scheduled_tokens

2299
    def _preprocess(
2300
2301
        self,
        scheduler_output: "SchedulerOutput",
2302
        num_input_tokens: int,  # Padded
2303
        intermediate_tensors: IntermediateTensors | None = None,
2304
    ) -> tuple[
2305
2306
        torch.Tensor | None,
        torch.Tensor | None,
2307
        torch.Tensor,
2308
        IntermediateTensors | None,
2309
        dict[str, Any],
2310
        ECConnectorOutput | None,
2311
    ]:
2312
        num_scheduled_tokens = scheduler_output.total_num_scheduled_tokens
2313
        is_first_rank = get_pp_group().is_first_rank
2314

2315
2316
        # _prepare_inputs may reorder the batch, so we must gather multi
        # modal outputs after that to ensure the correct order
2317
2318
        ec_connector_output = None

2319
2320
        if (
            self.supports_mm_inputs
2321
            and is_first_rank
2322
2323
            and not self.model_config.is_encoder_decoder
        ):
2324
            # Run the multimodal encoder if any.
2325
2326
2327
2328
2329
2330
            with self.maybe_get_ec_connector_output(
                scheduler_output,
                encoder_cache=self.encoder_cache,
            ) as ec_connector_output:
                self._execute_mm_encoder(scheduler_output)
                mm_embeds, is_mm_embed = self._gather_mm_embeddings(scheduler_output)
2331

2332
2333
2334
            # 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.
2335
            inputs_embeds_scheduled = self.model.embed_input_ids(
2336
2337
2338
                self.input_ids.gpu[:num_scheduled_tokens],
                multimodal_embeddings=mm_embeds,
                is_multimodal=is_mm_embed,
2339
            )
2340

2341
            # TODO(woosuk): Avoid the copy. Optimize.
2342
            self.inputs_embeds.gpu[:num_scheduled_tokens].copy_(inputs_embeds_scheduled)
2343

2344
            input_ids = None
2345
            inputs_embeds = self.inputs_embeds.gpu[:num_input_tokens]
2346
2347
2348
2349
            model_kwargs = {
                **self._init_model_kwargs(num_scheduled_tokens),
                **self._extract_mm_kwargs(scheduler_output),
            }
2350
        elif self.enable_prompt_embeds and is_first_rank:
2351
2352
2353
2354
2355
2356
2357
2358
2359
2360
2361
2362
            # 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).
2363
2364
2365
            token_ids_idx = (
                self.is_token_ids.gpu[:num_scheduled_tokens]
                .nonzero(as_tuple=False)
2366
                .squeeze(1)
2367
            )
2368
2369
2370
            # Some tokens ids may need to become embeds
            if token_ids_idx.numel() > 0:
                token_ids = self.input_ids.gpu[token_ids_idx]
2371
                tokens_to_embeds = self.model.embed_input_ids(input_ids=token_ids)
2372
2373
2374
2375
2376
                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
2377
        else:
2378
2379
2380
2381
            # 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.
2382
            input_ids = self.input_ids.gpu[:num_input_tokens]
2383
            inputs_embeds = None
2384
            model_kwargs = self._init_model_kwargs(num_input_tokens)
2385
        if self.uses_mrope:
2386
            positions = self.mrope_positions.gpu[:, :num_input_tokens]
2387
        else:
2388
            positions = self.positions.gpu[:num_input_tokens]
2389

2390
        if is_first_rank:
2391
2392
            intermediate_tensors = None
        else:
2393
            assert intermediate_tensors is not None
2394
            intermediate_tensors = self.sync_and_slice_intermediate_tensors(
2395
2396
                num_input_tokens, intermediate_tensors, True
            )
2397

2398
2399
2400
2401
        if (
            self.model_config.is_encoder_decoder
            and scheduler_output.scheduled_encoder_inputs
        ):
2402
2403
2404
2405
2406
2407
2408
            # Run the encoder, just like we do with other multimodal inputs.
            # For an encoder-decoder model, our processing here is a bit
            # simpler, because the outputs are just passed to the decoder.
            # We are not doing any prompt replacement. We also will only
            # ever have a single encoder input.
            encoder_outputs = self._execute_mm_encoder(scheduler_output)
            model_kwargs.update({"encoder_outputs": encoder_outputs})
2409

2410
2411
2412
2413
2414
2415
        return (
            input_ids,
            inputs_embeds,
            positions,
            intermediate_tensors,
            model_kwargs,
2416
            ec_connector_output,
2417
        )
2418

2419
    def _sample(
2420
        self,
2421
2422
        logits: torch.Tensor | None,
        spec_decode_metadata: SpecDecodeMetadata | None,
2423
    ) -> SamplerOutput:
2424
        # Sample the next token and get logprobs if needed.
2425
        sampling_metadata = self.input_batch.sampling_metadata
2426
        if spec_decode_metadata is None:
2427
2428
2429
            # 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()
2430
            return self.sampler(
2431
2432
2433
                logits=logits,
                sampling_metadata=sampling_metadata,
            )
2434

2435
        sampler_output = self.rejection_sampler(
2436
2437
            spec_decode_metadata,
            None,  # draft_probs
2438
            logits,
2439
2440
            sampling_metadata,
        )
2441
        self._update_states_after_model_execute(sampler_output.sampled_token_ids)
2442
2443
2444
        return sampler_output

    def _bookkeeping_sync(
2445
2446
2447
        self,
        scheduler_output: "SchedulerOutput",
        sampler_output: SamplerOutput,
2448
        logits: torch.Tensor | None,
2449
2450
        hidden_states: torch.Tensor,
        num_scheduled_tokens: int,
2451
        spec_decode_metadata: SpecDecodeMetadata | None,
2452
    ) -> tuple[
2453
        dict[str, int],
2454
        LogprobsLists | None,
2455
        list[list[int]],
2456
        dict[str, LogprobsTensors | None],
2457
2458
2459
        list[str],
        dict[str, int],
        list[int],
2460
    ]:
2461
2462
2463
2464
        num_nans_in_logits = {}
        if envs.VLLM_COMPUTE_NANS_IN_LOGITS:
            num_nans_in_logits = self._get_nans_in_logits(logits)

2465
2466
2467
        discard_sampled_tokens_req_indices = self.discard_request_indices.np[
            : self.num_discarded_requests
        ]
2468
2469
2470
2471
        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)
2472

2473
2474
2475
        # 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()
2476
        req_id_to_index_output_copy = self.input_batch.req_id_to_index.copy()
2477
2478

        num_sampled_tokens = sampler_output.sampled_token_ids.shape[0]
2479
        sampled_token_ids = sampler_output.sampled_token_ids
2480
        logprobs_tensors = sampler_output.logprobs_tensors
2481
        invalid_req_indices = []
2482
        cu_num_new_tokens: list[int] | None = None
2483
2484
2485
2486
2487
2488
2489
2490
2491
2492
2493
2494
        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,
                )
2495
2496
2497
2498
2499
2500
                if logprobs_tensors:
                    # Needed for extracting logprobs when spec decoding.
                    # This must be done prior to discarding sampled tokens.
                    cu_num_new_tokens = [0]
                    for toks in valid_sampled_token_ids:
                        cu_num_new_tokens.append(cu_num_new_tokens[-1] + len(toks))
2501
2502
            # Mask out the sampled tokens that should not be sampled.
            for i in discard_sampled_tokens_req_indices:
2503
                valid_sampled_token_ids[int(i)].clear()
2504
        else:
2505
            valid_sampled_token_ids = []
2506
            invalid_req_indices = discard_sampled_tokens_req_indices.tolist()
2507
2508
2509
2510
2511
            invalid_req_indices_set = set(invalid_req_indices)

            # 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.
2512
2513
2514
2515
            # With spec decoding, this is done in propose_draft_token_ids().
            if self.input_batch.prev_sampled_token_ids is None:
                assert sampled_token_ids.shape[-1] == 1
                self.input_batch.prev_sampled_token_ids = sampled_token_ids
2516
2517
2518
2519
2520
            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
            }
2521

2522
2523
2524
2525
2526
        # 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.
2527
        req_ids = self.input_batch.req_ids
2528
2529
        for req_idx in range(num_sampled_tokens):
            if self.use_async_scheduling:
2530
                sampled_ids = [-1] if req_idx not in invalid_req_indices_set else None
2531
2532
            else:
                sampled_ids = valid_sampled_token_ids[req_idx]
2533

2534
            num_sampled_ids: int = len(sampled_ids) if sampled_ids else 0
2535

2536
            if not sampled_ids:
2537
2538
2539
                continue

            start_idx = self.input_batch.num_tokens_no_spec[req_idx]
2540
            end_idx = start_idx + num_sampled_ids
2541
2542
2543
2544
            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}"
2545
            )
2546

2547
2548
            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
2549
2550
            self.input_batch.num_tokens_no_spec[req_idx] = end_idx
            self.input_batch.num_tokens[req_idx] = end_idx
2551

2552
            req_id = req_ids[req_idx]
2553
2554
2555
            req_state = self.requests[req_id]
            req_state.output_token_ids.extend(sampled_ids)

2556
        logprobs_lists = (
2557
            logprobs_tensors.tolists(cu_num_new_tokens)
2558
            if not self.use_async_scheduling and logprobs_tensors is not None
2559
2560
2561
2562
2563
2564
2565
2566
2567
            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,
        )

2568
2569
2570
2571
2572
2573
2574
2575
2576
2577
        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,
        )

2578
2579
2580
2581
2582
2583
2584
2585
2586
2587
2588
2589
2590
2591
2592
    @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()

2593
2594
    def _model_forward(
        self,
2595
2596
2597
2598
        input_ids: torch.Tensor | None = None,
        positions: torch.Tensor | None = None,
        intermediate_tensors: IntermediateTensors | None = None,
        inputs_embeds: torch.Tensor | None = None,
2599
2600
2601
2602
2603
        **model_kwargs: dict[str, Any],
    ) -> Any:
        """Helper method to call the model forward pass.

        This method can be overridden by subclasses for model execution.
2604
        Motivation: We can inspect only this method versus
2605
2606
2607
2608
2609
2610
2611
2612
2613
2614
2615
2616
2617
2618
2619
2620
2621
2622
2623
2624
        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,
        )

2625
2626
2627
2628
    @torch.inference_mode()
    def execute_model(
        self,
        scheduler_output: "SchedulerOutput",
2629
        intermediate_tensors: IntermediateTensors | None = None,
2630
2631
2632
2633
2634
2635
    ) -> 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."
            )
2636
2637
2638
2639
2640
2641
2642
2643
2644
2645
2646
2647
2648
2649
2650

        # self._draft_token_ids is None when `input_fits_in_drafter=False`
        # and there is no draft tokens scheduled. so it need to update the
        # spec_decoding info in scheduler_output with async_scheduling.
        # use deepcopy to avoid the modification has influence on the
        # scheduler_output in engine core process.
        # TODO(Ronald1995): deepcopy is expensive when there is a large
        # number of requests, optimize it later.
        if (
            self.use_async_scheduling
            and self.num_spec_tokens
            and self._draft_token_ids is None
        ):
            scheduler_output = deepcopy(scheduler_output)

2651
        num_scheduled_tokens = scheduler_output.total_num_scheduled_tokens
2652
        with record_function_or_nullcontext("gpu_model_runner: preprocess"):
2653
2654
2655
2656
            with self.synchronize_input_prep():
                # Update persistent batch states.
                self._update_states(scheduler_output)

2657
2658
2659
2660
2661
2662
2663
2664
                if has_ec_transfer() and get_ec_transfer().is_producer:
                    with self.maybe_get_ec_connector_output(
                        scheduler_output,
                        encoder_cache=self.encoder_cache,
                    ) as ec_connector_output:
                        self._execute_mm_encoder(scheduler_output)
                        return make_empty_encoder_model_runner_output(scheduler_output)

2665
                if not num_scheduled_tokens:
2666
2667
2668
2669
2670
2671
2672
2673
2674
2675
2676
2677
                    if (
                        self.parallel_config.distributed_executor_backend
                        == "external_launcher"
                        and self.parallel_config.data_parallel_size > 1
                    ):
                        # this is a corner case when both external launcher
                        # and DP are enabled, num_scheduled_tokens could be
                        # 0, and has_unfinished_requests in the outer loop
                        # returns True. before returning early here we call
                        # dummy run to ensure coordinate_batch_across_dp
                        # is called into to avoid out of sync issues.
                        self._dummy_run(1)
2678
2679
2680
2681
                    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(
2682
2683
                        scheduler_output, self.vllm_config
                    )
2684
                if self.cache_config.kv_sharing_fast_prefill:
2685
                    assert not self.num_prompt_logprobs, (
2686
2687
                        "--kv-sharing-fast-prefill produces incorrect "
                        "logprobs for prompt tokens, tokens, please disable "
2688
2689
                        "it when the requests need prompt logprobs"
                    )
2690

2691
2692
2693
2694
2695
2696
                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())

2697
2698
2699
2700
                (
                    logits_indices,
                    spec_decode_metadata,
                    ubatch_slices,
2701
                    num_tokens_across_dp,
2702
2703
2704
2705
2706
2707
2708
2709
2710
2711
2712
2713
2714
2715
2716
2717
2718
2719
2720
2721
2722
2723
2724
2725
2726
2727
2728
2729
2730
2731
2732
                ) = 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,
                    )
                )
2733

2734
2735
2736
2737
2738
2739
2740
2741
2742
2743
2744
                dp_rank = self.parallel_config.data_parallel_rank
                if ubatch_slices:
                    assert num_tokens_across_dp is not None
                    num_input_tokens = int(num_tokens_across_dp[dp_rank].item())
                    self.pad_out_ubatch_slice(ubatch_slices, num_input_tokens)
                elif num_tokens_across_dp is not None:
                    num_input_tokens = int(num_tokens_across_dp[dp_rank].item())
                else:
                    num_input_tokens = self._get_num_input_tokens(
                        scheduler_output.total_num_scheduled_tokens
                    )
2745

2746
2747
2748
2749
2750
2751
2752
2753
2754
2755
                (
                    input_ids,
                    inputs_embeds,
                    positions,
                    intermediate_tensors,
                    model_kwargs,
                    ec_connector_output,
                ) = self._preprocess(
                    scheduler_output, num_input_tokens, intermediate_tensors
                )
2756

2757
2758
2759
            uniform_decode = (
                max_num_scheduled_tokens == self.uniform_decode_query_len
            ) and (num_scheduled_tokens == num_reqs * max_num_scheduled_tokens)
2760
            batch_desc = BatchDescriptor(
2761
2762
2763
                num_tokens=num_input_tokens,
                uniform_decode=uniform_decode,
                has_lora=len(self.input_batch.lora_id_to_lora_request) > 0,
2764
2765
            )
            cudagraph_runtime_mode, batch_descriptor = (
2766
                self.cudagraph_dispatcher.dispatch(
2767
                    batch_desc,
2768
2769
                    use_cascade_attn=cascade_attn_prefix_lens is not None,
                )
2770
            )
2771

2772
        # Set cudagraph mode to none if calc_kv_scales is true.
2773
2774
2775
2776
2777
2778
        # KV scales calculation involves dynamic operations that are incompatible
        # with CUDA graph capture.
        if self.calculate_kv_scales:
            cudagraph_runtime_mode = CUDAGraphMode.NONE
            # Mark KV scales as calculated after the first forward pass
            self.calculate_kv_scales = False
2779

2780
2781
        # Run the model.
        # Use persistent buffers for CUDA graphs.
2782
2783
        with (
            set_forward_context(
2784
2785
2786
2787
2788
2789
                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,
2790
                ubatch_slices=ubatch_slices,
2791
            ),
2792
            record_function_or_nullcontext("gpu_model_runner: forward"),
2793
2794
            self.maybe_get_kv_connector_output(scheduler_output) as kv_connector_output,
        ):
2795
            model_output = self._model_forward(
2796
2797
2798
2799
2800
2801
2802
                input_ids=input_ids,
                positions=positions,
                intermediate_tensors=intermediate_tensors,
                inputs_embeds=inputs_embeds,
                **model_kwargs,
            )

2803
        with record_function_or_nullcontext("gpu_model_runner: postprocess"):
2804
            if self.use_aux_hidden_state_outputs:
2805
                # True when EAGLE 3 is used.
2806
2807
                hidden_states, aux_hidden_states = model_output
            else:
2808
                # Common case.
2809
2810
2811
                hidden_states = model_output
                aux_hidden_states = None

2812
2813
2814
2815
2816
            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)
2817
                    hidden_states.kv_connector_output = kv_connector_output
2818
                    self.kv_connector_output = kv_connector_output
2819
                    return hidden_states
2820

2821
                if self.is_pooling_model:
2822
                    # Return the pooling output.
2823
2824
2825
                    output = self._pool(
                        hidden_states, num_scheduled_tokens, num_scheduled_tokens_np
                    )
2826
2827
                    output.kv_connector_output = kv_connector_output
                    return output
2828
2829

                sample_hidden_states = hidden_states[logits_indices]
2830
                logits = self.model.compute_logits(sample_hidden_states)
2831
2832
2833
2834
            else:
                # Rare case.
                assert not self.is_pooling_model

2835
                sample_hidden_states = hidden_states[logits_indices]
2836
                if not get_pp_group().is_last_rank:
2837
                    all_gather_tensors = {
2838
2839
2840
                        "residual": not is_residual_scattered_for_sp(
                            self.vllm_config, num_input_tokens
                        )
2841
                    }
2842
                    get_pp_group().send_tensor_dict(
2843
2844
                        hidden_states.tensors,
                        all_gather_group=get_tp_group(),
2845
2846
                        all_gather_tensors=all_gather_tensors,
                    )
2847
2848
                    logits = None
                else:
2849
                    logits = self.model.compute_logits(sample_hidden_states)
2850

2851
                model_output_broadcast_data: dict[str, Any] = {}
2852
2853
2854
                if logits is not None:
                    model_output_broadcast_data["logits"] = logits.contiguous()

2855
                broadcasted = get_pp_group().broadcast_tensor_dict(
2856
2857
                    model_output_broadcast_data, src=len(get_pp_group().ranks) - 1
                )
2858
2859
                assert broadcasted is not None
                logits = broadcasted["logits"]
2860

2861
2862
2863
2864
2865
2866
2867
2868
        self.execute_model_state = ExecuteModelState(
            scheduler_output,
            logits,
            spec_decode_metadata,
            spec_decode_common_attn_metadata,
            hidden_states,
            sample_hidden_states,
            aux_hidden_states,
2869
            ec_connector_output,
2870
        )
2871
        self.kv_connector_output = kv_connector_output
2872
2873
2874
2875
2876
2877
        return None

    @torch.inference_mode
    def sample_tokens(
        self, grammar_output: "GrammarOutput | None"
    ) -> ModelRunnerOutput | AsyncModelRunnerOutput | IntermediateTensors:
2878
2879
2880
        kv_connector_output = self.kv_connector_output
        self.kv_connector_output = None

2881
2882
        if self.execute_model_state is None:
            # Nothing to do (PP non-final rank case), output isn't used.
2883
            if not kv_connector_output:
2884
                return None  # type: ignore[return-value]
2885
2886
2887
2888
2889
2890
2891
2892
2893

            # In case of PP with kv transfer, we need to pass through the
            # kv_connector_output
            if kv_connector_output.is_empty():
                return EMPTY_MODEL_RUNNER_OUTPUT

            output = copy(EMPTY_MODEL_RUNNER_OUTPUT)
            output.kv_connector_output = kv_connector_output
            return output
2894
2895
2896
2897
2898
2899
2900
2901
2902
2903

        # Unpack ephemeral state.
        (
            scheduler_output,
            logits,
            spec_decode_metadata,
            spec_decode_common_attn_metadata,
            hidden_states,
            sample_hidden_states,
            aux_hidden_states,
2904
            ec_connector_output,
2905
2906
2907
2908
2909
2910
2911
2912
2913
        ) = 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
            )
2914

2915
        with record_function_or_nullcontext("gpu_model_runner: sample"):
2916
2917
            sampler_output = self._sample(logits, spec_decode_metadata)

2918
2919
        self.input_batch.prev_sampled_token_ids = None

2920
        def propose_draft_token_ids(sampled_token_ids):
2921
            assert spec_decode_common_attn_metadata is not None
2922
            with record_function_or_nullcontext("gpu_model_runner: draft"):
2923
2924
2925
2926
2927
2928
2929
2930
2931
2932
2933
                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,
                )

2934
        spec_config = self.speculative_config
2935
        use_padded_batch_for_eagle = (
2936
2937
2938
            spec_config is not None
            and spec_config.use_eagle()
            and not spec_config.disable_padded_drafter_batch
2939
        )
2940
2941
2942
        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
2943
        if (
2944
2945
2946
            spec_config is not None
            and spec_config.draft_model_config is not None
            and spec_config.draft_model_config.max_model_len is not None
2947
        ):
2948
            effective_drafter_max_model_len = (
2949
                spec_config.draft_model_config.max_model_len
2950
            )
2951
        input_fits_in_drafter = spec_decode_common_attn_metadata and (
2952
            spec_decode_common_attn_metadata.max_seq_len + self.num_spec_tokens
2953
2954
            <= effective_drafter_max_model_len
        )
2955
        if use_padded_batch_for_eagle:
2956
2957
            assert self.speculative_config is not None
            assert isinstance(self.drafter, EagleProposer)
2958
2959
2960
2961
2962
2963
            sampled_token_ids = sampler_output.sampled_token_ids
            if input_fits_in_drafter:
                # 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(sampled_token_ids)
            elif self.valid_sampled_token_count_event is not None:
2964
                assert spec_decode_common_attn_metadata is not None
2965
2966
2967
2968
2969
2970
2971
2972
2973
2974
2975
2976
2977
                next_token_ids, valid_sampled_tokens_count = (
                    self.drafter.prepare_next_token_ids_padded(
                        spec_decode_common_attn_metadata,
                        sampled_token_ids,
                        self.requests,
                        self.input_batch,
                        self.discard_request_indices.gpu,
                        self.num_discarded_requests,
                    )
                )
                self._copy_valid_sampled_token_count(
                    next_token_ids, valid_sampled_tokens_count
                )
2978

2979
        with record_function_or_nullcontext("gpu_model_runner: bookkeep"):
2980
2981
2982
2983
2984
2985
2986
2987
            (
                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,
2988
2989
2990
2991
2992
            ) = self._bookkeeping_sync(
                scheduler_output,
                sampler_output,
                logits,
                hidden_states,
2993
                scheduler_output.total_num_scheduled_tokens,
2994
                spec_decode_metadata,
2995
            )
2996

2997
2998
2999
3000
3001
        if (
            self.speculative_config
            and not use_padded_batch_for_eagle
            and input_fits_in_drafter
        ):
3002
3003
3004
            # 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)
3005

3006
        with record_function_or_nullcontext("gpu_model_runner: eplb"):
3007
            self.eplb_step()
3008
3009
3010
3011
3012
3013
3014
3015
3016
        with record_function_or_nullcontext("gpu_model_runner: ModelRunnerOutput"):
            output = ModelRunnerOutput(
                req_ids=req_ids_output_copy,
                req_id_to_index=req_id_to_index_output_copy,
                sampled_token_ids=valid_sampled_token_ids,
                logprobs=logprobs_lists,
                prompt_logprobs_dict=prompt_logprobs_dict,
                pooler_output=[],
                kv_connector_output=kv_connector_output,
3017
3018
3019
                ec_connector_output=ec_connector_output
                if self.supports_mm_inputs
                else None,
3020
3021
                num_nans_in_logits=num_nans_in_logits,
            )
3022

3023
3024
        if not self.use_async_scheduling:
            return output
3025
3026
3027
3028
3029
3030
3031
3032
3033
        with record_function_or_nullcontext(
            "gpu_model_runner: AsyncGPUModelRunnerOutput"
        ):
            async_output = AsyncGPUModelRunnerOutput(
                model_runner_output=output,
                sampled_token_ids=sampler_output.sampled_token_ids,
                logprobs_tensors=sampler_output.logprobs_tensors,
                invalid_req_indices=invalid_req_indices,
                async_output_copy_stream=self.async_output_copy_stream,
3034
                vocab_size=self.input_batch.vocab_size,
3035
3036
3037
3038
3039
            )
        with record_function_or_nullcontext(
            "gpu_model_runner: set_async_sampled_token_ids"
        ):
            # Save ref of sampled_token_ids CPU tensor if the batch contains
3040
            # any requests with sampling params that require output ids.
3041
3042
3043
3044
            self.input_batch.set_async_sampled_token_ids(
                async_output.sampled_token_ids_cpu,
                async_output.async_copy_ready_event,
            )
3045
3046
3047

        return async_output

3048
    def take_draft_token_ids(self) -> DraftTokenIds | None:
3049
3050
3051
3052
3053
3054
3055
3056
3057
3058
        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)

3059
3060
3061
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
    def _copy_valid_sampled_token_count(
        self, next_token_ids: torch.Tensor, valid_sampled_tokens_count: torch.Tensor
    ) -> None:
        if self.valid_sampled_token_count_event is None:
            return

        default_stream = torch.cuda.current_stream()
        # Initialize a new stream to overlap the copy operation with
        # prepare_input of draft model.
        with torch.cuda.stream(self.valid_sampled_token_count_copy_stream):
            self.valid_sampled_token_count_copy_stream.wait_stream(default_stream)  # type: ignore
            counts = valid_sampled_tokens_count
            counts_cpu = self.valid_sampled_token_count_cpu
            counts_cpu[: counts.shape[0]].copy_(counts, non_blocking=True)
            self.valid_sampled_token_count_event.record()

        self.input_batch.prev_sampled_token_ids = next_token_ids.unsqueeze(1)

    def _get_valid_sampled_token_count(self) -> list[int]:
        # Wait until valid_sampled_tokens_count is copied to cpu,
        prev_sampled_token_ids = self.input_batch.prev_sampled_token_ids
        if (
            self.valid_sampled_token_count_event is None
            or prev_sampled_token_ids is None
        ):
            return []

        counts_cpu = self.valid_sampled_token_count_cpu
        self.valid_sampled_token_count_event.synchronize()
        return counts_cpu[: prev_sampled_token_ids.shape[0]].tolist()

3090
3091
3092
    def propose_draft_token_ids(
        self,
        scheduler_output: "SchedulerOutput",
3093
        sampled_token_ids: torch.Tensor | list[list[int]],
3094
3095
3096
        sampling_metadata: SamplingMetadata,
        hidden_states: torch.Tensor,
        sample_hidden_states: torch.Tensor,
3097
3098
        aux_hidden_states: list[torch.Tensor] | None,
        spec_decode_metadata: SpecDecodeMetadata | None,
3099
        common_attn_metadata: CommonAttentionMetadata,
3100
    ) -> list[list[int]] | torch.Tensor:
3101
        num_scheduled_tokens = scheduler_output.total_num_scheduled_tokens
3102
3103
3104
        spec_config = self.speculative_config
        assert spec_config is not None
        if spec_config.method == "ngram":
3105
            assert isinstance(sampled_token_ids, list)
3106
            assert isinstance(self.drafter, NgramProposer)
3107
            draft_token_ids = self.drafter.propose(
3108
3109
                sampled_token_ids,
                self.input_batch.req_ids,
3110
3111
                self.input_batch.num_tokens_no_spec,
                self.input_batch.token_ids_cpu,
3112
3113
                self.input_batch.spec_decode_unsupported_reqs,
            )
3114
        elif spec_config.method == "suffix":
3115
3116
3117
            assert isinstance(sampled_token_ids, list)
            assert isinstance(self.drafter, SuffixDecodingProposer)
            draft_token_ids = self.drafter.propose(self.input_batch, sampled_token_ids)
3118
        elif spec_config.method == "medusa":
3119
            assert isinstance(sampled_token_ids, list)
3120
            assert isinstance(self.drafter, MedusaProposer)
3121

3122
3123
            if sample_hidden_states.shape[0] == len(sampled_token_ids):
                # The input to the target model does not include draft tokens.
3124
3125
3126
3127
                hidden_states = sample_hidden_states
            else:
                indices = []
                offset = 0
3128
3129
3130
                assert spec_decode_metadata is not None, (
                    "No spec decode metadata for medusa"
                )
3131
                for num_draft, tokens in zip(
3132
3133
                    spec_decode_metadata.num_draft_tokens, sampled_token_ids
                ):
3134
                    indices.append(offset + len(tokens) - 1)
3135
                    offset += num_draft + 1
3136
                indices = torch.tensor(indices, device=self.device)
3137
3138
                hidden_states = sample_hidden_states[indices]

3139
            draft_token_ids = self.drafter.propose(
3140
3141
3142
                target_hidden_states=hidden_states,
                sampling_metadata=sampling_metadata,
            )
3143
        elif spec_config.use_eagle():
3144
            assert isinstance(self.drafter, EagleProposer)
3145

3146
            if spec_config.disable_padded_drafter_batch:
3147
3148
3149
                # 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.
3150
3151
                assert isinstance(sampled_token_ids, list), (
                    "sampled_token_ids should be a python list when"
3152
                    "padded-batch is disabled."
3153
                )
3154
                next_token_ids = self.drafter.prepare_next_token_ids_cpu(
3155
3156
3157
3158
3159
                    sampled_token_ids,
                    self.requests,
                    self.input_batch,
                    scheduler_output.num_scheduled_tokens,
                )
3160
3161
3162
3163
3164
            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.
3165
3166
                assert isinstance(sampled_token_ids, torch.Tensor), (
                    "sampled_token_ids should be a torch.Tensor when"
3167
                    "padded-batch is enabled."
3168
3169
                )
                next_token_ids, valid_sampled_tokens_count = (
3170
3171
3172
3173
3174
3175
                    self.drafter.prepare_next_token_ids_padded(
                        common_attn_metadata,
                        sampled_token_ids,
                        self.requests,
                        self.input_batch,
                        self.discard_request_indices.gpu,
3176
                        self.num_discarded_requests,
3177
                    )
3178
                )
3179
3180
3181
                self._copy_valid_sampled_token_count(
                    next_token_ids, valid_sampled_tokens_count
                )
Jiayi Yao's avatar
Jiayi Yao committed
3182

3183
            if spec_decode_metadata is None:
3184
                token_indices_to_sample = None
3185
                # input_ids can be None for multimodal models.
3186
                target_token_ids = self.input_ids.gpu[:num_scheduled_tokens]
3187
                target_positions = self._get_positions(num_scheduled_tokens)
3188
                if self.use_aux_hidden_state_outputs:
Wentao Ye's avatar
Wentao Ye committed
3189
                    assert aux_hidden_states is not None
3190
                    target_hidden_states = torch.cat(
3191
3192
                        [h[:num_scheduled_tokens] for h in aux_hidden_states], dim=-1
                    )
3193
3194
                else:
                    target_hidden_states = hidden_states[:num_scheduled_tokens]
3195
            else:
3196
                if spec_config.disable_padded_drafter_batch:
3197
                    token_indices_to_sample = None
3198
3199
3200
3201
3202
                    common_attn_metadata, token_indices = self.drafter.prepare_inputs(
                        common_attn_metadata,
                        sampled_token_ids,
                        spec_decode_metadata.num_draft_tokens,
                    )
3203
                else:
3204
                    common_attn_metadata, token_indices, token_indices_to_sample = (
3205
3206
3207
                        self.drafter.prepare_inputs_padded(
                            common_attn_metadata,
                            spec_decode_metadata,
3208
3209
3210
                            valid_sampled_tokens_count,
                        )
                    )
3211

3212
                target_token_ids = self.input_ids.gpu[token_indices]
3213
                target_positions = self._get_positions(token_indices)
3214
                if self.use_aux_hidden_state_outputs:
Wentao Ye's avatar
Wentao Ye committed
3215
                    assert aux_hidden_states is not None
3216
                    target_hidden_states = torch.cat(
3217
3218
                        [h[token_indices] for h in aux_hidden_states], dim=-1
                    )
3219
3220
                else:
                    target_hidden_states = hidden_states[token_indices]
3221

3222
            if self.supports_mm_inputs:
3223
3224
3225
3226
3227
3228
                mm_embed_inputs = self._gather_mm_embeddings(
                    scheduler_output,
                    shift_computed_tokens=1,
                )
            else:
                mm_embed_inputs = None
3229

3230
            draft_token_ids = self.drafter.propose(
3231
3232
3233
3234
                target_token_ids=target_token_ids,
                target_positions=target_positions,
                target_hidden_states=target_hidden_states,
                next_token_ids=next_token_ids,
3235
                last_token_indices=token_indices_to_sample,
3236
                sampling_metadata=sampling_metadata,
3237
                common_attn_metadata=common_attn_metadata,
3238
                mm_embed_inputs=mm_embed_inputs,
3239
            )
3240

3241
        return draft_token_ids
3242

3243
3244
3245
    def update_config(self, overrides: dict[str, Any]) -> None:
        allowed_config_names = {"load_config", "model_config"}
        for config_name, config_overrides in overrides.items():
3246
3247
            assert config_name in allowed_config_names, (
                f"Config `{config_name}` not supported. "
3248
                f"Allowed configs: {allowed_config_names}"
3249
            )
3250
3251
3252
3253
            config = getattr(self, config_name)
            new_config = update_config(config, config_overrides)
            setattr(self, config_name, new_config)

3254
3255
3256
3257
3258
    def load_model(self, eep_scale_up: bool = False) -> None:
        """
        Args:
            eep_scale_up: the model loading is for elastic EP scale up.
        """
3259
3260
3261
3262
3263
        logger.info_once(
            "Starting to load model %s...",
            self.model_config.model,
            scope="global",
        )
3264
3265
3266
3267
3268
        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)
        )
3269

3270
3271
3272
        if self.parallel_config.enable_eplb:
            self.eplb_state = EplbState(self.parallel_config, self.device)
            eplb_models = 0
3273
        with DeviceMemoryProfiler() as m:
3274
            time_before_load = time.perf_counter()
3275
            model_loader = get_model_loader(self.load_config)
3276
            self.model = model_loader.load_model(
3277
3278
                vllm_config=self.vllm_config, model_config=self.model_config
            )
3279
            if self.lora_config:
3280
3281
3282
                self.model = self.load_lora_model(
                    self.model, self.vllm_config, self.device
                )
3283
            if hasattr(self, "drafter"):
3284
                logger.info_once("Loading drafter model...")
3285
                self.drafter.load_model(self.model)
3286
3287
3288
3289
3290
                if (
                    hasattr(self.drafter, "model")
                    and is_mixture_of_experts(self.drafter.model)
                    and self.parallel_config.enable_eplb
                ):
3291
3292
3293
                    spec_config = self.vllm_config.speculative_config
                    assert spec_config is not None
                    assert spec_config.draft_model_config is not None
3294
3295
                    logger.info_once(
                        "EPLB is enabled for drafter model %s.",
3296
                        spec_config.draft_model_config.model,
3297
3298
3299
3300
3301
3302
3303
3304
3305
3306
3307
3308
3309
3310
3311
3312
                    )

                    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,
3313
                        spec_config.draft_model_config,
3314
3315
3316
3317
3318
3319
                        global_expert_load,
                        old_global_expert_indices,
                        rank_mapping,
                    )
                    eplb_models += 1

3320
            if self.use_aux_hidden_state_outputs:
3321
                if not supports_eagle3(self.get_model()):
3322
3323
                    raise RuntimeError(
                        "Model does not support EAGLE3 interface but "
3324
3325
                        "aux_hidden_state_outputs was requested"
                    )
3326
3327
3328
3329
3330
3331
3332
3333
3334
3335
3336
3337
3338

                # 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)
3339
            time_after_load = time.perf_counter()
3340
        self.model_memory_usage = m.consumed_memory
3341
        logger.info_once(
3342
            "Model loading took %.4f GiB memory and %.6f seconds",
3343
3344
            self.model_memory_usage / GiB_bytes,
            time_after_load - time_before_load,
3345
            scope="local",
3346
        )
3347
        prepare_communication_buffer_for_model(self.model)
3348
        mm_config = self.model_config.multimodal_config
3349
        self.is_multimodal_pruning_enabled = (
3350
            supports_multimodal_pruning(self.get_model())
3351
3352
            and mm_config is not None
            and mm_config.is_multimodal_pruning_enabled()
3353
        )
3354

3355
        if is_mixture_of_experts(self.model) and self.parallel_config.enable_eplb:
3356
3357
3358
3359
3360
3361
3362
3363
3364
3365
3366
            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(
3367
                self.model,
3368
                self.model_config,
3369
3370
3371
                global_expert_load,
                old_global_expert_indices,
                rank_mapping,
3372
3373
            )

3374
        if (
3375
3376
            self.vllm_config.compilation_config.mode
            == CompilationMode.STOCK_TORCH_COMPILE
3377
            and supports_dynamo()
3378
        ):
3379
            backend = self.vllm_config.compilation_config.init_backend(self.vllm_config)
3380
            compilation_counter.stock_torch_compile_count += 1
3381
            self.model.compile(fullgraph=True, backend=backend)
3382
            return
3383
        # for other compilation modes, cudagraph behavior is controlled by
3384
3385
3386
        # CudagraphWraper and CudagraphDispatcher of vllm.

        # wrap the model with full cudagraph wrapper if needed.
3387
3388
3389
        cudagraph_mode = self.compilation_config.cudagraph_mode
        assert cudagraph_mode is not None
        if cudagraph_mode.has_full_cudagraphs() and not self.parallel_config.enable_dbo:
3390
3391
3392
            self.model = CUDAGraphWrapper(
                self.model, self.vllm_config, runtime_mode=CUDAGraphMode.FULL
            )
3393
        elif self.parallel_config.enable_dbo:
3394
            if cudagraph_mode.has_full_cudagraphs():
3395
3396
3397
                self.model = UBatchWrapper(
                    self.model, self.vllm_config, CUDAGraphMode.FULL, self.device
                )
3398
            else:
3399
3400
3401
                self.model = UBatchWrapper(
                    self.model, self.vllm_config, CUDAGraphMode.NONE, self.device
                )
3402

3403
    def _get_eagle3_aux_layers_from_config(self) -> tuple[int, ...] | None:
3404
3405
3406
3407
3408
3409
3410
3411
3412
3413
3414
3415
3416
3417
3418
3419
3420
3421
3422
3423
3424
3425
3426
        """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

3427
    def reload_weights(self) -> None:
3428
        assert getattr(self, "model", None) is not None, (
3429
            "Cannot reload weights before model is loaded."
3430
        )
3431
3432
        model_loader = get_model_loader(self.load_config)
        logger.info("Reloading weights inplace...")
3433
        model_loader.load_weights(self.get_model(), model_config=self.model_config)
3434

3435
3436
3437
3438
3439
    def save_tensorized_model(
        self,
        tensorizer_config: "TensorizerConfig",
    ) -> None:
        TensorizerLoader.save_model(
3440
            self.get_model(),
3441
            tensorizer_config=tensorizer_config,
3442
            model_config=self.model_config,
3443
3444
        )

3445
3446
3447
    def _get_prompt_logprobs_dict(
        self,
        hidden_states: torch.Tensor,
3448
        num_scheduled_tokens: dict[str, int],
3449
    ) -> dict[str, LogprobsTensors | None]:
3450
        num_prompt_logprobs_dict = self.num_prompt_logprobs
3451
3452
3453
        if not num_prompt_logprobs_dict:
            return {}

3454
        in_progress_dict = self.input_batch.in_progress_prompt_logprobs_cpu
3455
        prompt_logprobs_dict: dict[str, LogprobsTensors | None] = {}
3456
3457
3458
3459
3460

        # 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():
3461
3462
3463
3464
            num_tokens = num_scheduled_tokens.get(req_id)
            if num_tokens is None:
                # This can happen if the request was preempted in prefill stage.
                continue
3465
3466
3467

            # Get metadata for this request.
            request = self.requests[req_id]
3468
3469
3470
3471
            if request.prompt_token_ids is None:
                # Prompt logprobs is incompatible with prompt embeddings
                continue

3472
3473
            num_prompt_tokens = len(request.prompt_token_ids)
            prompt_token_ids = torch.tensor(request.prompt_token_ids).to(
3474
3475
                self.device, non_blocking=True
            )
3476

3477
3478
3479
3480
3481
3482
            # 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(
3483
3484
                    num_prompt_tokens - 1, num_prompt_logprobs + 1
                )
3485
3486
                in_progress_dict[req_id] = logprobs_tensors

3487
            # Determine number of logits to retrieve.
3488
3489
            start_idx = request.num_computed_tokens
            start_tok = start_idx + 1
3490
            num_remaining_tokens = num_prompt_tokens - start_tok
3491
            if num_tokens <= num_remaining_tokens:
3492
                # This is a chunk, more tokens remain.
3493
3494
3495
                # 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.
3496
3497
3498
3499
3500
                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)
3501
3502
3503
3504
3505
3506
3507
                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
3508
3509
3510
3511
3512

            # 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]
3513
            offset = self.query_start_loc.np[req_idx].item()
3514
            prompt_hidden_states = hidden_states[offset : offset + num_logits]
3515
            logits = self.model.compute_logits(prompt_hidden_states)
3516
3517
3518
3519

            # 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.
3520
            tgt_token_ids = prompt_token_ids[start_tok : start_tok + num_logits]
3521
3522

            # Compute prompt logprobs.
3523
3524
            logprobs = self.sampler.compute_logprobs(logits)
            token_ids, logprobs, ranks = self.sampler.gather_logprobs(
3525
3526
                logprobs, num_prompt_logprobs, tgt_token_ids
            )
3527
3528

            # Transfer GPU->CPU async.
3529
3530
            chunk_slice = slice(start_idx, start_idx + num_logits)
            logprobs_tensors.logprob_token_ids[chunk_slice].copy_(
3531
3532
3533
                token_ids, non_blocking=True
            )
            logprobs_tensors.logprobs[chunk_slice].copy_(logprobs, non_blocking=True)
3534
            logprobs_tensors.selected_token_ranks[chunk_slice].copy_(
3535
3536
                ranks, non_blocking=True
            )
3537
3538
3539
3540
3541

        # 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]
3542
            del in_progress_dict[req_id]
3543
3544

        # Must synchronize the non-blocking GPU->CPU transfers.
3545
        if prompt_logprobs_dict:
3546
            self._sync_device()
3547
3548
3549

        return prompt_logprobs_dict

3550
3551
    def _get_nans_in_logits(
        self,
3552
        logits: torch.Tensor | None,
3553
3554
3555
3556
3557
3558
3559
3560
3561
3562
3563
    ) -> 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])
3564
3565
3566
                    if num_nans_for_index is not None and req_index < logits.shape[0]
                    else 0
                )
3567
3568
3569
3570
            return num_nans_in_logits
        except IndexError:
            return {}

3571
3572
3573
3574
3575
3576
    @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
3577
         - during DP rank dummy run
3578
3579
3580
3581
3582
3583
3584
3585
3586
3587
3588
        """
        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(
3589
                    self.input_ids.gpu,
3590
3591
                    low=0,
                    high=self.model_config.get_vocab_size(),
3592
3593
                    dtype=input_ids.dtype,
                )
3594

3595
            logger.debug_once("Randomizing dummy data for DP Rank")
3596
            input_ids.copy_(rand_input_ids()[: input_ids.size(0)], non_blocking=True)
3597
3598
3599
            yield
            input_ids.fill_(0)

3600
3601
3602
3603
3604
3605
    def _get_mm_dummy_batch(
        self,
        modality: str,
        max_items_per_batch: int,
    ) -> BatchedTensorInputs:
        """Dummy data for profiling and precompiling multimodal models."""
3606
3607
        assert self.mm_budget is not None

3608
3609
        dummy_decoder_data = self.mm_registry.get_decoder_dummy_data(
            model_config=self.model_config,
3610
            seq_len=self.max_model_len,
3611
            mm_counts={modality: 1},
3612
            cache=self.mm_budget.cache,
3613
3614
3615
3616
        )
        dummy_mm_data = dummy_decoder_data.multi_modal_data

        # Result in the maximum GPU consumption of the model
3617
3618
        dummy_mm_item = dummy_mm_data[modality][0]
        dummy_mm_items = [dummy_mm_item] * max_items_per_batch
3619

3620
        model = cast(SupportsMultiModal, self.model)
3621
3622
3623
3624
3625
3626
3627
        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,
3628
                multimodal_cpu_fields=model.multimodal_cpu_fields,
3629
3630
            )
        )
3631

3632
3633
3634
3635
    @torch.inference_mode()
    def _dummy_run(
        self,
        num_tokens: int,
3636
        cudagraph_runtime_mode: CUDAGraphMode | None = None,
3637
3638
        force_attention: bool = False,
        uniform_decode: bool = False,
3639
        allow_microbatching: bool = True,
3640
3641
        skip_eplb: bool = False,
        is_profile: bool = False,
3642
        create_mixed_batch: bool = False,
3643
        remove_lora: bool = True,
3644
        activate_lora: bool = False,
3645
    ) -> tuple[torch.Tensor, torch.Tensor]:
3646
3647
3648
3649
3650
3651
3652
        """
        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.
3653
                - if not set will determine the cudagraph mode based on using
3654
                    the self.cudagraph_dispatcher.
3655
3656
3657
3658
                - CUDAGraphMode.NONE: No cudagraph, for warm up and profile run
                - CUDAGraphMode.PIECEWISE: Piecewise cudagraph.
                - CUDAGraphMode.FULL: Full cudagraph, attention metadata is
                    needed.
3659
            force_attention: If True, always create attention metadata. Used to
3660
3661
3662
3663
                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.
3664
3665
            create_mixed_batch: If True, create a mixed batch with both decode
                (1 token) and prefill (multiple tokens) requests.
3666
            remove_lora: If False, dummy LoRAs are not destroyed after the run
3667
            activate_lora: If False, dummy_run is performed without LoRAs.
3668
        """
3669
3670
3671
3672
        assert (
            cudagraph_runtime_mode is None
            or cudagraph_runtime_mode.valid_runtime_modes()
        )
3673

3674
        # If cudagraph_mode.decode_mode() == FULL and
3675
        # cudagraph_mode.separate_routine(). This means that we are using
3676
3677
3678
3679
3680
3681
3682
3683
3684
3685
3686
        # 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.
3687
        max_query_len = self.uniform_decode_query_len if uniform_decode else num_tokens
3688

3689
3690
3691
3692
3693
        # 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
3694
3695
3696
3697
        if create_mixed_batch:
            assert not uniform_decode
            # Create mixed batch:
            # first half decode tokens, second half one prefill
3698
            num_decode_tokens = min(max_num_reqs - 1, num_tokens // 2)
3699
3700
3701
3702
            num_prefill_tokens = num_tokens - num_decode_tokens
            num_reqs = num_decode_tokens + 1

            # Create decode requests (1 token each) followed by prefill request
3703
            num_scheduled_tokens_list = [1] * num_decode_tokens + [num_prefill_tokens]
3704
3705
3706
            # Note: Overriding max_query_len to be the prefill tokens
            max_query_len = num_prefill_tokens
        elif uniform_decode:
3707
            assert not create_mixed_batch
3708
            num_reqs = min(max_num_reqs, cdiv(num_tokens, max_query_len))
3709
3710
            num_scheduled_tokens_list = [max_query_len] * num_reqs
            if num_tokens % max_query_len != 0:
3711
                num_scheduled_tokens_list[-1] = num_tokens % max_query_len
3712
3713
3714
3715
3716
3717
        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

3718
3719
        assert sum(num_scheduled_tokens_list) == num_tokens
        assert len(num_scheduled_tokens_list) == num_reqs
3720
        num_scheduled_tokens = np.array(num_scheduled_tokens_list, dtype=np.int32)
3721
        total_num_scheduled_tokens = int(num_scheduled_tokens.sum())
3722
        num_sampled_tokens = np.ones(num_reqs, dtype=np.int32)
3723

3724
3725
3726
        # Disable DP padding when running eager
        allow_dp_padding = self.compilation_config.cudagraph_mode != CUDAGraphMode.NONE

3727
3728
        # We currently only microbatch if the number of tokens is
        # over a certain threshold.
3729
        ubatch_slices, num_tokens_across_dp = coordinate_batch_across_dp(
3730
3731
3732
3733
3734
3735
3736
            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,
3737
3738
3739
        )
        num_tokens_after_padding = num_tokens
        if num_tokens_across_dp is not None:
3740
3741
            dp_rank = self.parallel_config.data_parallel_rank
            num_tokens_after_padding = int(num_tokens_across_dp[dp_rank])
3742

3743
3744
3745
3746
3747
3748
3749
3750
3751
3752
3753
3754
3755
3756
3757
3758
3759
3760
3761
3762
3763
3764
3765
3766
3767
        # filter out the valid batch descriptor
        _cg_mode, batch_descriptor = (
            self.cudagraph_dispatcher.dispatch(
                BatchDescriptor(
                    num_tokens=num_tokens_after_padding,
                    uniform_decode=uniform_decode,
                    has_lora=activate_lora and self.lora_config is not None,
                )
            )
            if not is_profile
            else (CUDAGraphMode.NONE, None)
        )
        if cudagraph_runtime_mode is not None:
            # we allow forcing NONE when the dispatcher disagrees to support
            # warm ups for cudagraph capture
            assert (
                cudagraph_runtime_mode == CUDAGraphMode.NONE
                or cudagraph_runtime_mode == _cg_mode
            ), (
                f"Cudagraph runtime mode mismatch at dummy_run. "
                f"Expected {_cg_mode}, but got {cudagraph_runtime_mode}."
            )
        else:
            cudagraph_runtime_mode = _cg_mode

3768
        attn_metadata: PerLayerAttnMetadata | None = None
3769
3770
3771

        # If force_attention is True, we always capture attention. Otherwise,
        # it only happens for cudagraph_runtime_mode=FULL.
3772
        if force_attention or cudagraph_runtime_mode == CUDAGraphMode.FULL:
3773
3774
3775
3776
3777
3778
            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:
3779
                seq_lens = max_query_len  # type: ignore[assignment]
3780
            self.seq_lens.np[:num_reqs] = seq_lens
3781
3782
            self.seq_lens.np[num_reqs:] = 0
            self.seq_lens.copy_to_gpu()
3783

3784
3785
            cum_num_tokens, _ = self._get_cumsum_and_arange(num_scheduled_tokens)
            self.query_start_loc.np[1 : num_reqs + 1] = cum_num_tokens
3786
3787
            self.query_start_loc.copy_to_gpu()

3788
3789
3790
3791
3792
3793
3794
            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,
            )
3795

3796
        with self.maybe_dummy_run_with_lora(
3797
3798
3799
3800
3801
            self.lora_config,
            num_scheduled_tokens,
            num_sampled_tokens,
            activate_lora,
            remove_lora,
3802
        ):
3803
3804
3805
            # 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)
3806
            if self.supports_mm_inputs and not self.model_config.is_encoder_decoder:
3807
                input_ids = None
3808
                inputs_embeds = self.inputs_embeds.gpu[:num_tokens_after_padding]
3809
                model_kwargs = {
3810
                    **model_kwargs,
3811
3812
                    **self._dummy_mm_kwargs(num_reqs),
                }
3813
3814
            elif self.enable_prompt_embeds:
                input_ids = None
3815
3816
                inputs_embeds = self.inputs_embeds.gpu[:num_tokens_after_padding]
                model_kwargs = self._init_model_kwargs(num_tokens_after_padding)
3817
            else:
3818
                input_ids = self.input_ids.gpu[:num_tokens_after_padding]
3819
                inputs_embeds = None
3820

3821
            if self.uses_mrope:
3822
                positions = self.mrope_positions.gpu[:, :num_tokens_after_padding]
3823
            else:
3824
                positions = self.positions.gpu[:num_tokens_after_padding]
3825
3826
3827
3828
3829
3830
3831
3832
3833

            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,
3834
3835
3836
                            device=self.device,
                        )
                    )
3837
3838

                intermediate_tensors = self.sync_and_slice_intermediate_tensors(
3839
                    num_tokens_after_padding, None, False
3840
                )
3841

3842
            if ubatch_slices is not None:
3843
3844
3845
3846
3847
3848
3849
                # 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

3850
3851
3852
            with (
                self.maybe_randomize_inputs(input_ids),
                set_forward_context(
3853
3854
                    attn_metadata,
                    self.vllm_config,
3855
                    num_tokens=num_tokens_after_padding,
3856
3857
                    num_tokens_across_dp=num_tokens_across_dp,
                    cudagraph_runtime_mode=cudagraph_runtime_mode,
3858
                    batch_descriptor=batch_descriptor,
3859
3860
3861
                    ubatch_slices=ubatch_slices,
                ),
            ):
3862
                outputs = self.model(
3863
3864
3865
3866
                    input_ids=input_ids,
                    positions=positions,
                    intermediate_tensors=intermediate_tensors,
                    inputs_embeds=inputs_embeds,
3867
                    **model_kwargs,
3868
                )
3869

3870
3871
3872
3873
            if self.use_aux_hidden_state_outputs:
                hidden_states, _ = outputs
            else:
                hidden_states = outputs
3874

3875
            if self.speculative_config and self.speculative_config.use_eagle():
3876
                assert isinstance(self.drafter, EagleProposer)
3877
3878
3879
3880
                use_cudagraphs = (
                    cudagraph_runtime_mode == CUDAGraphMode.PIECEWISE
                    and not self.speculative_config.enforce_eager
                )
3881
3882
3883
3884
3885
3886
3887
3888
3889
3890
3891
3892

                # 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,
                )
3893

3894
3895
3896
3897
3898
3899
3900
3901
3902
3903
        # 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)

3904
        logit_indices = np.cumsum(num_scheduled_tokens) - 1
3905
3906
3907
3908
        logit_indices_device = torch.from_numpy(logit_indices).to(
            self.device, non_blocking=True
        )
        return hidden_states, hidden_states[logit_indices_device]
3909
3910
3911
3912
3913
3914

    @torch.inference_mode()
    def _dummy_sampler_run(
        self,
        hidden_states: torch.Tensor,
    ) -> torch.Tensor:
3915
3916
3917
3918
        # 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)
3919

3920
        logits = self.model.compute_logits(hidden_states)
3921
3922
        num_reqs = logits.size(0)

3923
        dummy_tensors = lambda v: torch.full((num_reqs,), v, device=self.device)
3924
3925
3926
3927
3928
3929
3930
3931
3932
3933
3934
3935
3936
3937
3938

        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)],
3939
            spec_token_ids=[[] for _ in range(num_reqs)],
3940
3941
            allowed_token_ids_mask=None,
            bad_words_token_ids={},
3942
            logitsprocs=LogitsProcessors(),
3943
        )
3944
        try:
3945
3946
3947
            sampler_output = self.sampler(
                logits=logits, sampling_metadata=dummy_metadata
            )
3948
        except RuntimeError as e:
3949
            if "out of memory" in str(e):
3950
3951
3952
3953
                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 "
3954
3955
                    "initializing the engine."
                ) from e
3956
3957
            else:
                raise e
3958
        if self.speculative_config:
3959
3960
            draft_token_ids = [[0] for _ in range(num_reqs)]
            dummy_spec_decode_metadata = SpecDecodeMetadata.make_dummy(
3961
3962
                draft_token_ids, self.device
            )
3963
3964
3965
3966
3967
3968

            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
3969
3970
3971
3972
3973
            logits = torch.randn(
                num_tokens + num_reqs,
                logits.shape[-1],
                device=self.device,
                dtype=logits.dtype,
3974
            )
3975
3976
3977
            self.rejection_sampler(
                dummy_spec_decode_metadata,
                draft_probs,
3978
                logits,
3979
3980
                dummy_metadata,
            )
3981
        return sampler_output
3982

3983
    def _dummy_pooler_run_task(
3984
3985
        self,
        hidden_states: torch.Tensor,
3986
3987
        task: PoolingTask,
    ) -> PoolerOutput:
3988
3989
3990
3991
3992
3993
3994
3995
3996
3997
3998
        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

3999
        dummy_prompt_lens = torch.tensor(
4000
4001
            num_scheduled_tokens_list,
            device="cpu",
4002
        )
4003
4004
4005
        dummy_token_ids = torch.zeros(
            (num_reqs, req_num_tokens), dtype=torch.int32, device=self.device
        )
4006

4007
        model = cast(VllmModelForPooling, self.get_model())
4008
        dummy_pooling_params = PoolingParams(task=task)
4009
        dummy_pooling_params.verify(task=task, model_config=self.model_config)
4010
        to_update = model.pooler.get_pooling_updates(task)
4011
4012
        to_update.apply(dummy_pooling_params)

4013
        dummy_metadata = PoolingMetadata(
4014
4015
4016
4017
            prompt_lens=dummy_prompt_lens,
            prompt_token_ids=dummy_token_ids,
            pooling_params=[dummy_pooling_params] * num_reqs,
        )
4018

4019
4020
4021
        dummy_metadata.build_pooling_cursor(
            num_scheduled_tokens_list, device=hidden_states.device
        )
4022

4023
        try:
4024
4025
4026
            return model.pooler(
                hidden_states=hidden_states, pooling_metadata=dummy_metadata
            )
4027
        except RuntimeError as e:
4028
            if "out of memory" in str(e):
4029
                raise RuntimeError(
4030
4031
4032
                    "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 "
4033
4034
                    "initializing the engine."
                ) from e
4035
4036
            else:
                raise e
4037
4038
4039
4040
4041
4042
4043

    @torch.inference_mode()
    def _dummy_pooler_run(
        self,
        hidden_states: torch.Tensor,
    ) -> PoolerOutput:
        # Find the task that has the largest output for subsequent steps
4044
4045
4046
        supported_pooling_tasks = self.get_supported_pooling_tasks()

        if not supported_pooling_tasks:
4047
            if self.scheduler_config.enable_chunked_prefill:
4048
4049
4050
4051
4052
4053
4054
4055
4056
4057
4058
4059
4060
4061
4062
4063
                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."
                )

4064
        output_size = dict[PoolingTask, float]()
4065
        for task in supported_pooling_tasks:
4066
4067
            # Run a full batch with each task to ensure none of them OOMs
            output = self._dummy_pooler_run_task(hidden_states, task)
4068
            output_size[task] = sum(o.nbytes for o in output)
4069
4070
4071
4072
            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)
4073

4074
    def profile_run(self) -> None:
4075
        # Profile with multimodal encoder & encoder cache.
4076
        if self.supports_mm_inputs:
4077
4078
            mm_config = self.model_config.multimodal_config
            if mm_config is not None and mm_config.skip_mm_profiling:
4079
                logger.info(
4080
                    "Skipping memory profiling for multimodal encoder and "
4081
4082
                    "encoder cache."
                )
4083
4084
4085
4086
4087
4088
4089
4090
            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.
4091
                    dummy_modality = mm_budget.get_modality_with_max_tokens()
4092
4093
4094
                    max_mm_items_per_batch = mm_budget.max_items_per_batch_by_modality[
                        dummy_modality
                    ]
4095
4096
4097
4098
4099
4100
4101
4102
4103

                    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,
                    )
4104

4105
4106
4107
4108
4109
                    # Create dummy batch of multimodal inputs.
                    batched_dummy_mm_inputs = self._get_mm_dummy_batch(
                        dummy_modality,
                        max_mm_items_per_batch,
                    )
4110

4111
                    # Run multimodal encoder.
4112
                    dummy_encoder_outputs = self.model.embed_multimodal(
4113
4114
                        **batched_dummy_mm_inputs
                    )
4115

4116
4117
4118
4119
                    sanity_check_mm_encoder_outputs(
                        dummy_encoder_outputs,
                        expected_num_items=max_mm_items_per_batch,
                    )
4120

4121
4122
4123
4124
4125
4126
4127
4128
4129
4130
                    # 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(
4131
4132
                                (encoder_budget, encoder_output_shape[-1])
                            )
4133
4134
4135
4136
4137
4138
                            num_tokens = output.shape[0]
                            expanded[:num_tokens].copy_(output)
                            expanded_outputs.append(expanded)

                        dummy_encoder_outputs = expanded_outputs

4139
                    # Cache the dummy encoder outputs.
4140
                    self.encoder_cache["tmp"] = dict(enumerate(dummy_encoder_outputs))
4141

4142
        # Add `is_profile` here to pre-allocate communication buffers
4143
4144
4145
        hidden_states, last_hidden_states = self._dummy_run(
            self.max_num_tokens, is_profile=True
        )
4146
        if get_pp_group().is_last_rank:
4147
4148
4149
4150
            if self.is_pooling_model:
                output = self._dummy_pooler_run(hidden_states)
            else:
                output = self._dummy_sampler_run(last_hidden_states)
4151
        else:
4152
            output = None
4153
        self._sync_device()
4154
        del hidden_states, output
4155
        self.encoder_cache.clear()
4156
        gc.collect()
4157

4158
    def capture_model(self) -> int:
4159
        if self.compilation_config.cudagraph_mode == CUDAGraphMode.NONE:
4160
            logger.warning(
4161
                "Skipping CUDA graph capture. To turn on CUDA graph capture, "
4162
4163
                "ensure `cudagraph_mode` was not manually set to `NONE`"
            )
4164
            return 0
4165

4166
4167
        compilation_counter.num_gpu_runner_capture_triggers += 1

4168
4169
        start_time = time.perf_counter()

4170
4171
4172
4173
4174
4175
4176
4177
4178
4179
4180
4181
4182
4183
        @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()
4184
                    gc.collect()
4185

4186
4187
4188
        # 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.
4189
        set_cudagraph_capturing_enabled(True)
4190
        with freeze_gc(), graph_capture(device=self.device):
4191
            start_free_gpu_memory = torch.cuda.mem_get_info()[0]
4192
            cudagraph_mode = self.compilation_config.cudagraph_mode
4193
            assert cudagraph_mode is not None
4194
4195
4196
4197
4198
4199
4200
4201
4202

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

4203
4204
            if cudagraph_mode.mixed_mode() != CUDAGraphMode.NONE:
                cudagraph_runtime_mode = cudagraph_mode.mixed_mode()
4205
                # make sure we capture the largest batch size first
4206
4207
4208
                compilation_cases = list(
                    product(reversed(self.cudagraph_batch_sizes), lora_cases)
                )
4209
4210
4211
                self._capture_cudagraphs(
                    compilation_cases,
                    cudagraph_runtime_mode=cudagraph_runtime_mode,
4212
4213
                    uniform_decode=False,
                )
4214

4215
4216
            # Capture full cudagraph for uniform decode batches if we
            # don't already have full mixed prefill-decode cudagraphs.
4217
4218
4219
4220
4221
4222
4223
            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
                )
4224
                decode_cudagraph_batch_sizes = [
4225
4226
                    x
                    for x in self.cudagraph_batch_sizes
4227
                    if max_num_tokens >= x >= self.uniform_decode_query_len
4228
                ]
4229
4230
4231
                compilation_cases_decode = list(
                    product(reversed(decode_cudagraph_batch_sizes), lora_cases)
                )
4232
4233
4234
                self._capture_cudagraphs(
                    compilation_cases=compilation_cases_decode,
                    cudagraph_runtime_mode=CUDAGraphMode.FULL,
4235
4236
                    uniform_decode=True,
                )
4237

4238
4239
4240
            torch.cuda.synchronize()
            end_free_gpu_memory = torch.cuda.mem_get_info()[0]

4241
4242
4243
        # 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
4244
        # we may do lazy capturing in future that still allows capturing
4245
4246
        # after here.
        set_cudagraph_capturing_enabled(False)
4247
4248
4249
4250
4251

        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.
4252
        logger.info_once(
4253
4254
4255
            "Graph capturing finished in %.0f secs, took %.2f GiB",
            elapsed_time,
            cuda_graph_size / (1 << 30),
4256
            scope="local",
4257
        )
4258
        return cuda_graph_size
4259

4260
4261
    def _capture_cudagraphs(
        self,
4262
        compilation_cases: list[tuple[int, bool]],
4263
4264
4265
4266
4267
4268
4269
        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}"
4270
4271
4272
4273
4274
4275
4276
4277

        # 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",
4278
4279
4280
                    cudagraph_runtime_mode.name,
                ),
            )
4281

4282
        # We skip EPLB here since we don't want to record dummy metrics
4283
        for num_tokens, activate_lora in compilation_cases:
4284
            # We currently only capture ubatched graphs when its a FULL
4285
4286
4287
            # 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
4288
4289
4290
4291
            allow_microbatching = (
                self.parallel_config.enable_dbo
                and cudagraph_runtime_mode == CUDAGraphMode.FULL
                and uniform_decode
4292
4293
4294
4295
4296
                and check_ubatch_thresholds(
                    config=self.vllm_config.parallel_config,
                    num_tokens=num_tokens,
                    uniform_decode=uniform_decode,
                )
4297
            )
4298

4299
4300
4301
4302
4303
4304
            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.
4305
4306
4307
4308
4309
4310
4311
4312
4313
                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,
4314
                    activate_lora=activate_lora,
4315
4316
4317
4318
4319
4320
4321
4322
                )
            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,
4323
                activate_lora=activate_lora,
4324
            )
4325
        self.maybe_remove_all_loras(self.lora_config)
4326

4327
4328
4329
4330
    def initialize_attn_backend(self, kv_cache_config: KVCacheConfig) -> None:
        """
        Initialize the attention backends and attention metadata builders.
        """
4331
        assert len(self.attn_groups) == 0, "Attention backends are already initialized"
4332

4333
4334
4335
4336
4337
4338
        class AttentionGroupKey(NamedTuple):
            attn_backend: type[AttentionBackend]
            kv_cache_spec: KVCacheSpec

        def get_attn_backends_for_group(
            kv_cache_group_spec: KVCacheGroupSpec,
4339
        ) -> tuple[dict[AttentionGroupKey, list[str]], set[type[AttentionBackend]]]:
4340
            layer_type = cast(type[Any], AttentionLayerBase)
4341
            layers = get_layers_from_vllm_config(
4342
                self.vllm_config, layer_type, kv_cache_group_spec.layer_names
4343
            )
4344
4345
            attn_backends = {}
            attn_backend_layers = defaultdict(list)
4346
            # Dedupe based on full class name; this is a bit safer than
4347
4348
4349
4350
            # 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.
4351
            for layer_name in kv_cache_group_spec.layer_names:
4352
                attn_backend = layers[layer_name].get_attn_backend()
4353
4354
4355
4356

                if layer_name in self.kv_sharing_fast_prefill_eligible_layers:
                    attn_backend = create_fast_prefill_custom_backend(
                        "FastPrefill",
4357
                        attn_backend,  # type: ignore[arg-type]
4358
4359
                    )

4360
4361
4362
                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):
4363
                    layer_kv_cache_spec = layer_kv_cache_spec.kv_cache_specs[layer_name]
4364
                key = (full_cls_name, layer_kv_cache_spec)
4365
4366
4367
                attn_backends[key] = AttentionGroupKey(
                    attn_backend, layer_kv_cache_spec
                )
4368
                attn_backend_layers[key].append(layer_name)
4369
4370
4371
4372
            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()),
            )
4373
4374

        def create_attn_groups(
4375
            attn_backends_map: dict[AttentionGroupKey, list[str]],
4376
            kv_cache_group_id: int,
4377
4378
        ) -> list[AttentionGroup]:
            attn_groups: list[AttentionGroup] = []
4379
            for (attn_backend, kv_cache_spec), layer_names in attn_backends_map.items():
4380
                attn_group = AttentionGroup(
4381
                    attn_backend,
4382
                    layer_names,
4383
                    kv_cache_spec,
4384
                    kv_cache_group_id,
4385
4386
                )

4387
4388
4389
                attn_groups.append(attn_group)
            return attn_groups

4390
        attention_backend_maps = []
4391
        attention_backend_list = []
4392
        for kv_cache_group_spec in kv_cache_config.kv_cache_groups:
4393
            attn_backends = get_attn_backends_for_group(kv_cache_group_spec)
4394
            attention_backend_maps.append(attn_backends[0])
4395
            attention_backend_list.append(attn_backends[1])
4396
4397

        # Resolve cudagraph_mode before actually initialize metadata_builders
4398
4399
4400
        self._check_and_update_cudagraph_mode(
            attention_backend_list, kv_cache_config.kv_cache_groups
        )
4401

4402
4403
        for i, attn_backend_map in enumerate(attention_backend_maps):
            self.attn_groups.append(create_attn_groups(attn_backend_map, i))
4404

4405
4406
4407
4408
4409
4410
4411
4412
4413
4414
4415
4416
4417
4418
4419
4420
4421
4422
    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
4423
        # Calculate reorder batch threshold (if needed)
4424
4425
        # Note (tdoublep): do this *after* constructing builders,
        # because some of them change the threshold at init time.
4426
4427
        self.calculate_reorder_batch_threshold()

4428
    def _check_and_update_cudagraph_mode(
4429
4430
4431
        self,
        attention_backends: list[set[type[AttentionBackend]]],
        kv_cache_groups: list[KVCacheGroupSpec],
4432
    ) -> None:
4433
        """
4434
        Resolve the cudagraph_mode when there are multiple attention
4435
        groups with potential conflicting CUDA graph support.
4436
4437
4438
        Then initialize the cudagraph_dispatcher based on the resolved
        cudagraph_mode.
        """
4439
        min_cg_support = AttentionCGSupport.ALWAYS
4440
        min_cg_backend_name = None
4441

4442
4443
4444
4445
4446
4447
4448
4449
4450
4451
4452
4453
        for attn_backend_set, kv_cache_group in zip(
            attention_backends, kv_cache_groups
        ):
            for attn_backend in attn_backend_set:
                builder_cls = attn_backend.get_builder_cls()

                cg_support = builder_cls.get_cudagraph_support(
                    self.vllm_config, kv_cache_group.kv_cache_spec
                )
                if cg_support.value < min_cg_support.value:
                    min_cg_support = cg_support
                    min_cg_backend_name = attn_backend.__name__
4454
4455
        # Flexible resolve the cudagraph mode
        cudagraph_mode = self.compilation_config.cudagraph_mode
4456
        assert cudagraph_mode is not None
4457
        # check cudagraph for mixed batch is supported
4458
4459
4460
4461
4462
4463
        if (
            cudagraph_mode.mixed_mode() == CUDAGraphMode.FULL
            and min_cg_support != AttentionCGSupport.ALWAYS
        ):
            msg = (
                f"CUDAGraphMode.{cudagraph_mode.name} is not supported "
4464
                f"with {min_cg_backend_name} backend (support: "
4465
4466
                f"{min_cg_support})"
            )
4467
4468
            if min_cg_support == AttentionCGSupport.NEVER:
                # if not supported any full cudagraphs, just raise it.
4469
4470
                msg += (
                    "; please try cudagraph_mode=PIECEWISE, and "
4471
                    "make sure compilation mode is VLLM_COMPILE"
4472
                )
4473
4474
4475
4476
4477
                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"
4478
                cudagraph_mode = self.compilation_config.cudagraph_mode = (
4479
                    CUDAGraphMode.FULL_AND_PIECEWISE
4480
                )
4481
4482
            else:
                msg += "; setting cudagraph_mode=FULL_DECODE_ONLY"
4483
                cudagraph_mode = self.compilation_config.cudagraph_mode = (
4484
                    CUDAGraphMode.FULL_DECODE_ONLY
4485
                )
4486
4487
            logger.warning(msg)

4488
        # check that if we are doing decode full-cudagraphs it is supported
4489
4490
4491
4492
4493
4494
        if (
            cudagraph_mode.decode_mode() == CUDAGraphMode.FULL
            and min_cg_support == AttentionCGSupport.NEVER
        ):
            msg = (
                f"CUDAGraphMode.{cudagraph_mode.name} is not supported "
4495
                f"with {min_cg_backend_name} backend (support: "
4496
4497
                f"{min_cg_support})"
            )
4498
            if self.compilation_config.mode == CompilationMode.VLLM_COMPILE and (
4499
4500
4501
4502
4503
                self.compilation_config.splitting_ops_contain_attention()
                or self.compilation_config.use_inductor_graph_partition
            ):
                msg += (
                    "; setting cudagraph_mode=PIECEWISE because "
4504
                    "attention is compiled piecewise"
4505
4506
                )
                cudagraph_mode = self.compilation_config.cudagraph_mode = (
4507
                    CUDAGraphMode.PIECEWISE
4508
                )
4509
            else:
4510
4511
                msg += (
                    "; setting cudagraph_mode=NONE because "
4512
                    "attention is not compiled piecewise"
4513
4514
                )
                cudagraph_mode = self.compilation_config.cudagraph_mode = (
4515
                    CUDAGraphMode.NONE
4516
                )
4517
4518
            logger.warning(msg)

4519
4520
        # check that if we are doing spec-decode + decode full-cudagraphs it is
        # supported
4521
4522
4523
4524
4525
4526
4527
4528
        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 "
4529
                f"{min_cg_backend_name} (support: {min_cg_support})"
4530
            )
4531
4532
            if self.compilation_config.splitting_ops_contain_attention():
                msg += "; setting cudagraph_mode=PIECEWISE"
4533
                cudagraph_mode = self.compilation_config.cudagraph_mode = (
4534
                    CUDAGraphMode.PIECEWISE
4535
                )
4536
4537
            else:
                msg += "; setting cudagraph_mode=NONE"
4538
                cudagraph_mode = self.compilation_config.cudagraph_mode = (
4539
                    CUDAGraphMode.NONE
4540
                )
4541
4542
4543
4544
            logger.warning(msg)

        # double check that we can support full cudagraph if they are requested
        # even after automatic downgrades
4545
4546
4547
4548
4549
4550
        if (
            cudagraph_mode.has_full_cudagraphs()
            and min_cg_support == AttentionCGSupport.NEVER
        ):
            raise ValueError(
                f"CUDAGraphMode.{cudagraph_mode.name} is not "
4551
                f"supported with {min_cg_backend_name} backend ("
4552
4553
                f"support:{min_cg_support}) "
                "; please try cudagraph_mode=PIECEWISE, "
4554
                "and make sure compilation mode is VLLM_COMPILE"
4555
            )
4556

4557
4558
4559
4560
4561
4562
4563
4564
4565
4566
4567
4568
4569
4570
        # if we have dedicated decode cudagraphs, and spec-decode is enabled,
        # we need to adjust the cudagraph sizes to be a multiple of the uniform
        # decode query length to avoid: https://github.com/vllm-project/vllm/issues/28207
        # temp-fix: https://github.com/vllm-project/vllm/issues/28207#issuecomment-3504004536
        # Will be removed in the near future when we have seperate cudagraph capture
        # sizes for decode and mixed prefill-decode.
        if (
            cudagraph_mode.decode_mode() == CUDAGraphMode.FULL
            and cudagraph_mode.separate_routine()
            and self.uniform_decode_query_len > 1
        ):
            self.compilation_config.adjust_cudagraph_sizes_for_spec_decode(
                self.uniform_decode_query_len, self.parallel_config.tensor_parallel_size
            )
4571
4572
4573
4574
            capture_sizes = self.compilation_config.cudagraph_capture_sizes
            self.cudagraph_batch_sizes = (
                capture_sizes if capture_sizes is not None else []
            )
4575

4576
4577
        # Trigger cudagraph dispatching keys initialization after
        # resolved cudagraph mode.
4578
4579
        cudagraph_mode = self.compilation_config.cudagraph_mode
        assert cudagraph_mode is not None
4580
        self.cudagraph_dispatcher.initialize_cudagraph_keys(
4581
            cudagraph_mode, self.uniform_decode_query_len
4582
        )
4583

4584
4585
    def calculate_reorder_batch_threshold(self) -> None:
        """
4586
4587
4588
4589
        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.
4590
        """
4591
4592
        min_none_high = lambda a, b: a if b is None else b if a is None else min(a, b)

4593
        reorder_batch_thresholds: list[int | None] = [
4594
4595
4596
            group.get_metadata_builder().reorder_batch_threshold
            for group in self._attn_group_iterator()
        ]
4597
4598
4599
4600
4601
        # 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
4602
        self.reorder_batch_threshold = reduce(min_none_high, reorder_batch_thresholds)  # type: ignore[assignment]
4603

4604
4605
4606
    @staticmethod
    def select_common_block_size(
        kv_manager_block_size: int, attn_groups: list[AttentionGroup]
4607
4608
    ) -> int:
        """
4609
4610
4611
4612
4613
        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.
4614
4615
4616
4617
4618
4619

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

        Returns:
4620
            The selected block size
4621
4622

        Raises:
4623
            ValueError: If no valid block size found
4624
4625
        """

4626
4627
4628
4629
4630
4631
4632
4633
        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
4634
                for supported_size in backend.get_supported_kernel_block_sizes():
4635
4636
4637
4638
4639
4640
4641
4642
4643
4644
4645
4646
4647
4648
4649
4650
4651
4652
4653
4654
4655
4656
4657
4658
4659
4660
4661
4662
4663
4664
                    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
4665
            for supported_size in backend.get_supported_kernel_block_sizes()
4666
4667
            if isinstance(supported_size, int)
        )
4668

4669
4670
4671
4672
4673
4674
        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}. ")
4675

4676
4677
4678
    def may_reinitialize_input_batch(
        self, kv_cache_config: KVCacheConfig, kernel_block_sizes: list[int]
    ) -> None:
4679
4680
4681
4682
4683
4684
4685
        """
        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.
4686
            kernel_block_sizes: The kernel block sizes for each KV cache group.
4687
4688
4689
4690
        """
        block_sizes = [
            kv_cache_group.kv_cache_spec.block_size
            for kv_cache_group in kv_cache_config.kv_cache_groups
4691
            if not isinstance(kv_cache_group.kv_cache_spec, EncoderOnlyAttentionSpec)
4692
        ]
4693
4694
4695
4696

        if block_sizes != [self.cache_config.block_size] or kernel_block_sizes != [
            self.cache_config.block_size
        ]:
4697
4698
4699
            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
4700
4701
                "for more details."
            )
4702
4703
            self.input_batch = InputBatch(
                max_num_reqs=self.max_num_reqs,
4704
                max_model_len=max(self.max_model_len, self.max_encoder_len),
4705
4706
4707
4708
4709
                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,
4710
                kernel_block_sizes=kernel_block_sizes,
4711
                is_spec_decode=bool(self.vllm_config.speculative_config),
4712
                logitsprocs=self.input_batch.logitsprocs,
4713
                logitsprocs_need_output_token_ids=self.input_batch.logitsprocs_need_output_token_ids,
4714
                is_pooling_model=self.is_pooling_model,
4715
                num_speculative_tokens=self.num_spec_tokens,
4716
4717
            )

4718
    def _allocate_kv_cache_tensors(
4719
4720
        self, kv_cache_config: KVCacheConfig
    ) -> dict[str, torch.Tensor]:
4721
        """
4722
4723
4724
        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.

4725
        Args:
4726
            kv_cache_config: The KV cache config
4727
        Returns:
4728
            dict[str, torch.Tensor]: A map between layer names to their
4729
            corresponding memory buffer for KV cache.
4730
        """
4731
4732
        kv_cache_raw_tensors: dict[str, torch.Tensor] = {}
        for kv_cache_tensor in kv_cache_config.kv_cache_tensors:
4733
4734
4735
            tensor = torch.zeros(
                kv_cache_tensor.size, dtype=torch.int8, device=self.device
            )
4736
4737
4738
4739
4740
            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:
4741
4742
4743
4744
            for layer_name in group.layer_names:
                if layer_name in self.runner_only_attn_layers:
                    continue
                layer_names.add(layer_name)
4745
4746
4747
        assert layer_names == set(kv_cache_raw_tensors.keys()), (
            "Some layers are not correctly initialized"
        )
4748
4749
        return kv_cache_raw_tensors

4750
4751
4752
    def _attn_group_iterator(self) -> Iterator[AttentionGroup]:
        return itertools.chain.from_iterable(self.attn_groups)

4753
    def _kv_cache_spec_attn_group_iterator(self) -> Iterator[AttentionGroup]:
4754
4755
        if not self.kv_cache_config.kv_cache_groups:
            return
4756
4757
        for attn_groups in self.attn_groups:
            yield from attn_groups
4758

4759
4760
4761
4762
4763
4764
4765
4766
4767
4768
4769
4770
4771
4772
4773
    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 = []
4774
        for kv_cache_gid, kv_cache_group in enumerate(kv_cache_config.kv_cache_groups):
4775
4776
4777
4778
4779
4780
            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):
4781
                continue
4782
            elif isinstance(kv_cache_spec, AttentionSpec):
4783
4784
4785
                # This is an attention backend that supports virtual
                # block splitting. Get the supported block sizes from
                # all backends in the group.
4786
                attn_groups = self.attn_groups[kv_cache_gid]
4787
                kv_manager_block_size = kv_cache_group.kv_cache_spec.block_size
4788
                selected_kernel_size = self.select_common_block_size(
4789
4790
4791
                    kv_manager_block_size, attn_groups
                )
                kernel_block_sizes.append(selected_kernel_size)
4792
            elif isinstance(kv_cache_spec, MambaSpec):
4793
4794
                # This is likely Mamba or other non-attention cache,
                # no splitting.
4795
                kernel_block_sizes.append(kv_cache_spec.block_size)
4796
4797
4798
4799
4800
4801
            else:
                raise NotImplementedError(
                    f"unknown kv cache spec {kv_cache_group.kv_cache_spec}"
                )
        return kernel_block_sizes

4802
4803
4804
4805
    def _reshape_kv_cache_tensors(
        self,
        kv_cache_config: KVCacheConfig,
        kv_cache_raw_tensors: dict[str, torch.Tensor],
4806
        kernel_block_sizes: list[int],
4807
    ) -> dict[str, torch.Tensor]:
4808
        """
4809
        Reshape the KV cache tensors to the desired shape and dtype.
4810

4811
        Args:
4812
4813
            kv_cache_config: The KV cache config
            kv_cache_raw_tensors: The KV cache buffer of each layer, with
4814
                correct size but uninitialized shape.
4815
            kernel_block_sizes: The kernel block sizes for each KV cache group.
4816
        Returns:
4817
            Dict[str, torch.Tensor]: A map between layer names to their
4818
4819
            corresponding memory buffer for KV cache.
        """
4820
        kv_caches: dict[str, torch.Tensor] = {}
4821
        has_attn, has_mamba = False, False
4822
4823
        for group in self._kv_cache_spec_attn_group_iterator():
            kv_cache_spec = group.kv_cache_spec
4824
            attn_backend = group.backend
4825
4826
4827
4828
            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]
4829
            for layer_name in group.layer_names:
4830
4831
                if layer_name in self.runner_only_attn_layers:
                    continue
4832
4833
                raw_tensor = kv_cache_raw_tensors[layer_name]
                assert raw_tensor.numel() % kv_cache_spec.page_size_bytes == 0
4834
                num_blocks = raw_tensor.numel() // kv_cache_spec.page_size_bytes
4835
                if isinstance(kv_cache_spec, AttentionSpec):
4836
                    has_attn = True
4837
4838
                    num_blocks_per_kv_block = (
                        kv_cache_spec.block_size // kernel_block_size
4839
4840
4841
                    )
                    kernel_num_blocks = num_blocks * num_blocks_per_kv_block

4842
                    kv_cache_shape = attn_backend.get_kv_cache_shape(
4843
                        kernel_num_blocks,
4844
                        kernel_block_size,
4845
4846
                        kv_cache_spec.num_kv_heads,
                        kv_cache_spec.head_size,
4847
4848
                        cache_dtype_str=self.cache_config.cache_dtype,
                    )
4849
                    dtype = kv_cache_spec.dtype
4850
                    try:
4851
                        kv_cache_stride_order = attn_backend.get_kv_cache_stride_order()
4852
                        assert len(kv_cache_stride_order) == len(kv_cache_shape)
4853
                    except (AttributeError, NotImplementedError):
4854
                        kv_cache_stride_order = tuple(range(len(kv_cache_shape)))
4855
4856
4857
4858
4859
                    # 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.
4860
4861
4862
                    kv_cache_shape = tuple(
                        kv_cache_shape[i] for i in kv_cache_stride_order
                    )
4863
4864
4865
4866
4867
                    # Maintain original KV shape view.
                    inv_order = [
                        kv_cache_stride_order.index(i)
                        for i in range(len(kv_cache_stride_order))
                    ]
4868
4869
4870
4871
4872
4873
                    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
4874
                elif isinstance(kv_cache_spec, MambaSpec):
4875
                    has_mamba = True
Chen Zhang's avatar
Chen Zhang committed
4876
4877
                    raw_tensor = kv_cache_raw_tensors[layer_name]
                    state_tensors = []
4878
                    storage_offset_bytes = 0
4879
                    for shape, dtype in zip(kv_cache_spec.shapes, kv_cache_spec.dtypes):
4880
4881
                        dtype_size = get_dtype_size(dtype)
                        num_element_per_page = (
4882
4883
                            kv_cache_spec.page_size_bytes // dtype_size
                        )
Chen Zhang's avatar
Chen Zhang committed
4884
                        target_shape = (num_blocks, *shape)
4885
4886
                        stride = torch.empty(target_shape).stride()
                        target_stride = (num_element_per_page, *stride[1:])
4887
                        assert storage_offset_bytes % dtype_size == 0
4888
4889
4890
4891
                        tensor = torch.as_strided(
                            raw_tensor.view(dtype),
                            size=target_shape,
                            stride=target_stride,
4892
                            storage_offset=storage_offset_bytes // dtype_size,
4893
                        )
Chen Zhang's avatar
Chen Zhang committed
4894
                        state_tensors.append(tensor)
4895
                        storage_offset_bytes += stride[0] * dtype_size
4896
4897

                    kv_caches[layer_name] = state_tensors
4898
                else:
4899
                    raise NotImplementedError
4900
4901

        if has_attn and has_mamba:
4902
            self._update_hybrid_attention_mamba_layout(kv_caches)
4903

4904
4905
        return kv_caches

4906
    def _update_hybrid_attention_mamba_layout(
4907
4908
        self, kv_caches: dict[str, torch.Tensor]
    ) -> None:
4909
        """
4910
4911
        Update the layout of attention layers from (2, num_blocks, ...) to
        (num_blocks, 2, ...).
4912
4913

        Args:
4914
            kv_caches: The KV cache buffer of each layer.
4915
4916
        """

4917
4918
        for group in self._kv_cache_spec_attn_group_iterator():
            kv_cache_spec = group.kv_cache_spec
4919
            for layer_name in group.layer_names:
4920
                kv_cache = kv_caches[layer_name]
4921
4922
4923
4924
                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 "
4925
                        f"a tensor of shape {kv_cache.shape}"
4926
                    )
4927
                    hidden_size = kv_cache.shape[2:].numel()
4928
4929
4930
4931
                    kv_cache.as_strided_(
                        size=kv_cache.shape,
                        stride=(hidden_size, 2 * hidden_size, *kv_cache.stride()[2:]),
                    )
4932

4933
    def initialize_kv_cache_tensors(
4934
        self, kv_cache_config: KVCacheConfig, kernel_block_sizes: list[int]
4935
    ) -> dict[str, torch.Tensor]:
4936
4937
4938
4939
4940
        """
        Initialize the memory buffer for KV cache.

        Args:
            kv_cache_config: The KV cache config
4941
4942
            kernel_block_sizes: The kernel block sizes for each KV cache group.

4943
        Returns:
4944
            Dict[str, torch.Tensor]: A map between layer names to their
4945
4946
            corresponding memory buffer for KV cache.
        """
4947
4948
4949
4950
4951
4952
4953
4954
4955
4956
4957
4958
4959
4960
4961
4962
4963
4964
4965
4966
4967
4968
4969
4970

        # Try creating KV caches optimized for kv-connector transfers
        cache_dtype = self.cache_config.cache_dtype
        if self.use_uniform_kv_cache(self.attn_groups, cache_dtype):
            kv_caches, cross_layers_kv_cache, attn_backend = (
                self.allocate_uniform_kv_caches(
                    kv_cache_config,
                    self.attn_groups,
                    cache_dtype,
                    self.device,
                    kernel_block_sizes,
                )
            )
            self.cross_layers_kv_cache = cross_layers_kv_cache
            self.cross_layers_attn_backend = attn_backend
        else:
            # Fallback to the general case
            # 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
            kv_caches = self._reshape_kv_cache_tensors(
                kv_cache_config, kv_cache_raw_tensors, kernel_block_sizes
            )
4971

4972
        # Set up cross-layer KV cache sharing
4973
4974
        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)
4975
4976
            kv_caches[layer_name] = kv_caches[target_layer_name]

4977
4978
4979
4980
4981
4982
4983
4984
4985
        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,
        )
4986
4987
4988
        return kv_caches

    def maybe_add_kv_sharing_layers_to_kv_cache_groups(
4989
4990
        self, kv_cache_config: KVCacheConfig
    ) -> None:
4991
4992
4993
4994
4995
4996
4997
4998
4999
5000
5001
5002
5003
5004
5005
5006
5007
5008
        """
        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.
5009
            attn_layers = get_layers_from_vllm_config(self.vllm_config, Attention)
5010
5011
            for layer_name in reversed(attn_layers):
                if layer_name in self.shared_kv_cache_layers:
5012
                    self.kv_sharing_fast_prefill_eligible_layers.add(layer_name)
5013
5014
                else:
                    break
5015

5016
5017
5018
5019
5020
5021
5022
    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
        """
5023
        kv_cache_config = deepcopy(kv_cache_config)
5024
        self.kv_cache_config = kv_cache_config
5025
        self.may_add_encoder_only_layers_to_kv_cache_config()
5026
        self.maybe_add_kv_sharing_layers_to_kv_cache_groups(kv_cache_config)
5027
        self.initialize_attn_backend(kv_cache_config)
5028
5029
5030
5031
5032
5033
        # 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)
5034
5035
5036
5037

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

5038
        # Reinitialize need to after initialize_attn_backend
5039
5040
5041
5042
        self.may_reinitialize_input_batch(kv_cache_config, kernel_block_sizes)
        kv_caches = self.initialize_kv_cache_tensors(
            kv_cache_config, kernel_block_sizes
        )
5043

5044
5045
5046
5047
5048
5049
        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
5050
        if has_kv_transfer_group():
5051
            kv_transfer_group = get_kv_transfer_group()
5052
5053
5054
5055
5056
5057
5058
            if self.cross_layers_kv_cache is not None:
                assert self.cross_layers_attn_backend is not None
                kv_transfer_group.register_cross_layers_kv_cache(
                    self.cross_layers_kv_cache, self.cross_layers_attn_backend
                )
            else:
                kv_transfer_group.register_kv_caches(kv_caches)
5059
            kv_transfer_group.set_host_xfer_buffer_ops(copy_kv_blocks)
Robert Shaw's avatar
Robert Shaw committed
5060

5061
        if self.dcp_world_size > 1:
5062
5063
            layer_type = cast(type[Any], AttentionLayerBase)
            layers = get_layers_from_vllm_config(self.vllm_config, layer_type)
5064
            for layer in layers.values():
5065
5066
5067
5068
                layer_impl = getattr(layer, "impl", None)
                if layer_impl is None:
                    continue
                assert layer_impl.need_to_return_lse_for_decode, (
5069
5070
                    "DCP requires attention impls to return"
                    " the softmax lse for decode, but the impl "
5071
                    f"{layer_impl.__class__.__name__} "
5072
5073
                    "does not return the softmax lse for decode."
                )
5074

5075
5076
5077
5078
5079
    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
5080
        encoder_only_attn_specs: dict[AttentionSpec, list[str]] = defaultdict(list)
5081
5082
5083
        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:
5084
                attn_spec: AttentionSpec = EncoderOnlyAttentionSpec(
5085
5086
5087
                    block_size=block_size,
                    num_kv_heads=attn_module.num_kv_heads,
                    head_size=attn_module.head_size,
5088
5089
                    dtype=self.kv_cache_dtype,
                )
5090
5091
5092
                encoder_only_attn_specs[attn_spec].append(layer_name)
                self.runner_only_attn_layers.add(layer_name)
        if len(encoder_only_attn_specs) > 0:
5093
5094
5095
            assert len(encoder_only_attn_specs) == 1, (
                "Only support one encoder-only attention spec now"
            )
5096
5097
            spec, layer_names = encoder_only_attn_specs.popitem()
            self.kv_cache_config.kv_cache_groups.append(
5098
5099
                KVCacheGroupSpec(layer_names=layer_names, kv_cache_spec=spec)
            )
5100

5101
    def get_kv_cache_spec(self) -> dict[str, KVCacheSpec]:
5102
        """
5103
        Generates the KVCacheSpec by parsing the kv cache format from each
5104
5105
        Attention module in the static forward context.
        Returns:
5106
            KVCacheSpec: A dictionary mapping layer names to their KV cache
5107
5108
            format. Layers that do not need KV cache are not included.
        """
5109
5110
        if has_ec_transfer() and get_ec_transfer().is_producer:
            return {}
5111
        kv_cache_spec: dict[str, KVCacheSpec] = {}
5112
5113
        layer_type = cast(type[Any], AttentionLayerBase)
        attn_layers = get_layers_from_vllm_config(self.vllm_config, layer_type)
Chen Zhang's avatar
Chen Zhang committed
5114
        for layer_name, attn_module in attn_layers.items():
5115
5116
5117
5118
5119
5120
5121
5122
5123
5124
5125
5126
5127
5128
5129
            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
5130

5131
        return kv_cache_spec
5132

5133
    def _to_list(self, sampled_token_ids: torch.Tensor) -> list[list[int]]:
5134
5135
5136
5137
5138
5139
5140
5141
        # 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.
5142
        pinned = self.sampled_token_ids_pinned_cpu[: sampled_token_ids.shape[0]]
5143
5144
5145
        pinned.copy_(sampled_token_ids, non_blocking=True)
        self.transfer_event.record()
        self.transfer_event.synchronize()
5146
        return pinned.tolist()