gpu_model_runner.py 190 KB
Newer Older
1
# SPDX-License-Identifier: Apache-2.0
2
# SPDX-FileCopyrightText: Copyright contributors to the vLLM project
3

4
import gc
5
import itertools
6
import time
7
8
from collections import defaultdict
from collections.abc import Iterator
9
from contextlib import contextmanager
10
from copy import deepcopy
11
from typing import TYPE_CHECKING, Any, NamedTuple, Optional, Union, cast
12
13
14
15
16

import numpy as np
import torch
import torch.distributed
import torch.nn as nn
17
from tqdm import tqdm
18
from typing_extensions import TypeAlias
19

20
import vllm.envs as envs
21
from vllm.attention import Attention, AttentionType
22
from vllm.attention.backends.abstract import AttentionBackend
23
from vllm.attention.layers.chunked_local_attention import ChunkedLocalAttention
24
from vllm.compilation.counter import compilation_counter
25
26
27
from vllm.compilation.cuda_graph import CUDAGraphWrapper
from vllm.compilation.monitor import set_cudagraph_capturing_enabled
from vllm.config import (CompilationLevel, CUDAGraphMode, VllmConfig,
28
                         get_layers_from_vllm_config, update_config)
29
from vllm.distributed.eplb.eplb_state import EplbState
30
31
from vllm.distributed.kv_transfer import (get_kv_transfer_group,
                                          has_kv_transfer_group)
32
from vllm.distributed.kv_transfer.kv_connector.utils import copy_kv_blocks
33
from vllm.distributed.parallel_state import (
34
    get_pp_group, get_tp_group, graph_capture, is_global_first_rank,
35
    prepare_communication_buffer_for_model)
36
37
from vllm.forward_context import (BatchDescriptor, DPMetadata,
                                  set_forward_context)
38
from vllm.logger import init_logger
39
40
from vllm.model_executor.layers.attention_layer_base import AttentionLayerBase
from vllm.model_executor.layers.mamba.abstract import MambaBase
41
from vllm.model_executor.layers.rotary_embedding import MRotaryEmbedding
42
from vllm.model_executor.model_loader import TensorizerLoader, get_model_loader
43
44
# yapf conflicts with isort for this block
# yapf: disable
45
46
from vllm.model_executor.models.interfaces import (SupportsMultiModal,
                                                   is_mixture_of_experts,
47
                                                   supports_eagle3,
48
                                                   supports_mrope,
49
                                                   supports_multimodal_pruning,
50
                                                   supports_transcription)
51
# yapf: enable
52
53
from vllm.model_executor.models.interfaces_base import (
    VllmModelForPooling, is_pooling_model, is_text_generation_model)
54
from vllm.multimodal import MULTIMODAL_REGISTRY
55
from vllm.multimodal.inputs import (BatchedTensorInputs, MultiModalKwargsItem,
56
                                    PlaceholderRange)
57
from vllm.multimodal.utils import group_mm_kwargs_by_modality
58
from vllm.pooling_params import PoolingParams
59
from vllm.sampling_params import SamplingType
60
from vllm.sequence import IntermediateTensors
61
from vllm.tasks import GenerationTask, PoolingTask, SupportedTask
62
from vllm.utils import (STR_DTYPE_TO_TORCH_DTYPE, DeviceMemoryProfiler,
63
                        GiB_bytes, cdiv, check_use_alibi, get_dtype_size,
64
65
66
                        is_pin_memory_available,
                        length_from_prompt_token_ids_or_embeds, round_up,
                        supports_dynamo)
67
from vllm.utils.jsontree import json_map_leaves
68
from vllm.v1.attention.backends.flash_attn import AttentionMetadata
69
from vllm.v1.attention.backends.gdn_attn import GDNAttentionMetadataBuilder
70
from vllm.v1.attention.backends.utils import (
71
    AttentionCGSupport, AttentionMetadataBuilder, CommonAttentionMetadata,
72
    create_fast_prefill_custom_backend,
73
    reorder_batch_to_split_decodes_and_prefills, split_attn_metadata)
74
from vllm.v1.cudagraph_dispatcher import CudagraphDispatcher
75
76
# yapf conflicts with isort for this block
# yapf: disable
77
78
from vllm.v1.kv_cache_interface import (AttentionSpec,
                                        ChunkedLocalAttentionSpec,
79
                                        CrossAttentionSpec,
80
                                        EncoderOnlyAttentionSpec,
81
                                        FullAttentionSpec, KVCacheConfig,
82
                                        KVCacheGroupSpec, KVCacheSpec,
83
84
                                        MambaSpec, SlidingWindowSpec,
                                        UniformTypeKVCacheSpecs)
85
# yapf: enable
86
from vllm.v1.outputs import (EMPTY_MODEL_RUNNER_OUTPUT, AsyncModelRunnerOutput,
87
                             DraftTokenIds, LogprobsLists, LogprobsTensors,
88
                             ModelRunnerOutput, PoolerOutput, SamplerOutput)
89
from vllm.v1.pool.metadata import PoolingMetadata
90
from vllm.v1.sample.logits_processor import LogitsProcessors, build_logitsprocs
91
from vllm.v1.sample.metadata import SamplingMetadata
92
from vllm.v1.sample.rejection_sampler import RejectionSampler
93
from vllm.v1.sample.sampler import Sampler
94
from vllm.v1.spec_decode.eagle import EagleProposer
95
from vllm.v1.spec_decode.medusa import MedusaProposer
96
from vllm.v1.spec_decode.metadata import SpecDecodeMetadata
97
from vllm.v1.spec_decode.ngram_proposer import NgramProposer
98
from vllm.v1.structured_output.utils import apply_grammar_bitmask
99
from vllm.v1.utils import CpuGpuBuffer, record_function_or_nullcontext
100
from vllm.v1.worker.gpu_input_batch import CachedRequestState, InputBatch
101
from vllm.v1.worker.gpu_ubatch_wrapper import UBatchWrapper
102
from vllm.v1.worker.kv_connector_model_runner_mixin import (
103
    KVConnectorModelRunnerMixin)
104
from vllm.v1.worker.lora_model_runner_mixin import LoRAModelRunnerMixin
105
106
from vllm.v1.worker.ubatch_splitting import (check_ubatch_thresholds,
                                             ubatch_split)
107
from vllm.v1.worker.ubatch_utils import UBatchSlice, UBatchSlices
108
from vllm.v1.worker.utils import is_residual_scattered_for_sp
109

110
111
112
113
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)
114

115
if TYPE_CHECKING:
116
    from vllm.model_executor.model_loader.tensorizer import TensorizerConfig
117
    from vllm.v1.core.sched.output import SchedulerOutput
118
119
120

logger = init_logger(__name__)

121
122
123
124
125
AttnMetadataDict: TypeAlias = dict[str, AttentionMetadata]
# list when ubatching is enabled
PerLayerAttnMetadata: TypeAlias = Union[list[AttnMetadataDict],
                                        AttnMetadataDict]

126

127
128
129
130
131
132
133
134
135
136
137
138
139
140
141
142
143
144
145
146
147
148
149
150
151
152
153
154
155
156
157
158
159
160
161
162
163
164
165
166
167
168
169
170
171
172
173
# Wrapper for ModelRunnerOutput to support overlapped execution.
class AsyncGPUModelRunnerOutput(AsyncModelRunnerOutput):

    def __init__(
        self,
        model_runner_output: ModelRunnerOutput,
        sampled_token_ids: torch.Tensor,
        invalid_req_indices: list[int],
        async_output_copy_stream: torch.cuda.Stream,
    ):
        self._model_runner_output = model_runner_output
        self._invalid_req_indices = invalid_req_indices

        # Event on the copy stream so we can synchronize the non-blocking copy.
        self._async_copy_ready_event = torch.cuda.Event()

        # Keep a reference to the device tensor to avoid it being
        # deallocated until we finish copying it to the host.
        self._sampled_token_ids = sampled_token_ids

        # Initiate the copy on a separate stream, but do not synchronize it.
        default_stream = torch.cuda.current_stream()
        with torch.cuda.stream(async_output_copy_stream):
            async_output_copy_stream.wait_stream(default_stream)
            self._sampled_token_ids_cpu = self._sampled_token_ids.to(
                'cpu', non_blocking=True)
            self._async_copy_ready_event.record()

    def get_output(self) -> ModelRunnerOutput:
        """Copy the device tensors to the host and return a ModelRunnerOutput.
        
        This function blocks until the copy is finished.
        """
        self._async_copy_ready_event.synchronize()

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

        valid_sampled_token_ids = self._sampled_token_ids_cpu.tolist()
        for i in self._invalid_req_indices:
            valid_sampled_token_ids[i].clear()

        output = self._model_runner_output
        output.sampled_token_ids = valid_sampled_token_ids
        return output


174
class GPUModelRunner(LoRAModelRunnerMixin, KVConnectorModelRunnerMixin):
175
176
177

    def __init__(
        self,
178
        vllm_config: VllmConfig,
179
        device: torch.device,
180
    ):
181
182
183
        self.vllm_config = vllm_config
        self.model_config = vllm_config.model_config
        self.cache_config = vllm_config.cache_config
184
        self.compilation_config = vllm_config.compilation_config
185
186
187
188
189
190
        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
191

192
193
194
        from vllm.model_executor.models.utils import set_cpu_offload_max_bytes
        set_cpu_offload_max_bytes(
            int(self.cache_config.cpu_offload_gb * 1024**3))
195
196
197
        from vllm.model_executor.layers.batch_invariant import (
            init_batch_invariance)
        init_batch_invariance()
198

199
200
201
202
        model_config = self.model_config
        cache_config = self.cache_config
        scheduler_config = self.scheduler_config
        parallel_config = self.parallel_config
203
        self.device = device
204
205
206
207
208
209
210
211
        self.pin_memory = is_pin_memory_available()
        self.dtype = self.model_config.dtype
        if cache_config.cache_dtype == "auto":
            self.kv_cache_dtype = self.dtype
        else:
            self.kv_cache_dtype = STR_DTYPE_TO_TORCH_DTYPE[
                cache_config.cache_dtype]

212
        self.is_pooling_model = (model_config.runner_type == 'pooling')
213
        self.enable_prompt_embeds = model_config.enable_prompt_embeds
214
215
        self.is_multimodal_raw_input_only_model = (
            model_config.is_multimodal_raw_input_only_model)
216
217
        # This will be overridden in load_model()
        self.is_multimodal_pruning_enabled = False
218
        self.max_model_len = model_config.max_model_len
219
        self.dcp_world_size = self.parallel_config.decode_context_parallel_size
220
        self.max_num_tokens = scheduler_config.max_num_batched_tokens
221
        self.max_num_reqs = scheduler_config.max_num_seqs
222

223
224
225
226
227
228
229
230
        # 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 = (
            self.parallel_config.distributed_executor_backend
            == "external_launcher" and len(get_pp_group().ranks) > 0)

231
        # Model-related.
232
233
        self.num_query_heads = model_config.get_num_attention_heads(
            parallel_config)
234
        self.hidden_size = model_config.get_hidden_size()
235
        self.attention_chunk_size = model_config.attention_chunk_size
236
237
        # Only relevant for models using ALiBi (e.g, MPT)
        self.use_alibi = check_use_alibi(model_config)
238

239
        self.cascade_attn_enabled = not self.model_config.disable_cascade_attn
240

241
        # Multi-modal data support
242
        self.mm_registry = MULTIMODAL_REGISTRY
243
        self.uses_mrope = model_config.uses_mrope
244
245
        self.supports_mm_inputs = self.mm_registry.supports_multimodal_inputs(
            model_config)
246

247
248
249
        if self.model_config.is_encoder_decoder:
            # Maximum length of the encoder input, only for encoder-decoder
            # models.
250
251
            self.max_encoder_len = scheduler_config.\
                            max_num_encoder_input_tokens
252
253
254
        else:
            self.max_encoder_len = 0

255
        # Sampler
256
        self.sampler = Sampler(logprobs_mode=self.model_config.logprobs_mode)
257

258
259
260
261
262
263
264
        self.eplb_state: Optional[EplbState] = None
        """
        State of the expert parallelism load balancer.

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

265
        # Lazy initializations
266
        # self.model: nn.Module  # Set after load_model
267
        # Initialize in initialize_kv_cache
268
        self.kv_caches: list[torch.Tensor] = []
269
270
        # indexes: [kv_cache_group_id][attn_group]
        self.attn_groups: list[list[AttentionGroup]] = []
271
272
        # self.kv_cache_config: KVCacheConfig

273
274
        # mm_hash ->  encoder_output
        self.encoder_cache: dict[str, torch.Tensor] = {}
275

276
        self.use_aux_hidden_state_outputs = False
277
278
279
280
281
282
283
284
285
286
287
288
289
290
291
292
293
294
295
296
        # Set up speculative decoding.
        # NOTE(Jiayi): currently we put the entire draft model on
        # the last PP rank. This is not ideal if there are many
        # layers in the draft model.
        if self.speculative_config and get_pp_group().is_last_rank:
            if self.speculative_config.method == "ngram":
                self.drafter = NgramProposer(self.vllm_config)
            elif self.speculative_config.use_eagle():
                self.drafter = EagleProposer(self.vllm_config, self.device,
                                             self)  # type: ignore
                if self.speculative_config.method == "eagle3":
                    self.use_aux_hidden_state_outputs = True
            elif self.speculative_config.method == "medusa":
                self.drafter = MedusaProposer(
                    vllm_config=self.vllm_config,
                    device=self.device)  # type: ignore
            else:
                raise ValueError("Unknown speculative decoding method: "
                                 f"{self.speculative_config.method}")
            self.rejection_sampler = RejectionSampler()
297

298
        # Request states.
299
        self.requests: dict[str, CachedRequestState] = {}
300
        self.comm_stream = torch.cuda.Stream()
301

302
303
304
305
306
307
308
309
310
        # 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.
311
312
        self.input_batch = InputBatch(
            max_num_reqs=self.max_num_reqs,
313
314
315
            # 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),
316
317
318
            max_num_batched_tokens=self.max_num_tokens,
            device=self.device,
            pin_memory=self.pin_memory,
319
            vocab_size=self.model_config.get_vocab_size(),
320
            block_sizes=[self.cache_config.block_size],
321
            is_spec_decode=bool(self.vllm_config.speculative_config),
322
323
324
325
326
            logitsprocs=build_logitsprocs(
                self.vllm_config, self.device, self.pin_memory,
                self.is_pooling_model,
                self.vllm_config.model_config.logits_processors),
            is_pooling_model=self.is_pooling_model,
327
        )
328

329
330
331
332
        self.use_async_scheduling = self.scheduler_config.async_scheduling
        self.async_output_copy_stream = torch.cuda.Stream() if \
            self.use_async_scheduling else None

333
        # TODO(woosuk): Provide an option to tune the max cudagraph batch size.
334
335
336
        # The convention is different.
        # self.cudagraph_batch_sizes sorts in ascending order.
        # The batch sizes in the config are in descending order.
337
338
339
340
        if self.compilation_config.cudagraph_capture_sizes and \
                self.compilation_config.cudagraph_mode != CUDAGraphMode.NONE:
            self.cudagraph_batch_sizes = list(
                reversed(self.compilation_config.cudagraph_capture_sizes))
341

342
        # Cache the device properties.
343
        self._init_device_properties()
344

345
        # Persistent buffers for CUDA graphs.
346
347
348
349
350
351
352
        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)
        self.seq_lens = self._make_buffer(self.max_num_reqs, dtype=torch.int32)
353
354
355
356
357
358
359
        # 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.
        self.inputs_embeds = self._make_buffer(self.max_num_tokens,
                                               self.hidden_size,
                                               dtype=self.dtype,
                                               numpy=False)
360
361
        self.is_token_ids = self._make_buffer(self.max_num_tokens,
                                              dtype=torch.bool)
362
363
364
365
        self.discard_request_indices = self._make_buffer(self.max_num_reqs,
                                                         dtype=torch.int64)
        self.num_discarded_requests = 0

366
367
        self.num_decode_draft_tokens = self._make_buffer(self.max_num_reqs,
                                                         dtype=torch.int32)
368
369
        self.num_accepted_tokens = self._make_buffer(self.max_num_reqs,
                                                     dtype=torch.int64)
370
371

        # Only relevant for models using M-RoPE (e.g, Qwen2-VL)
372
        if self.uses_mrope:
Roger Wang's avatar
Roger Wang committed
373
374
375
376
            # 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
377
378
379
380
381
382

            # 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
383
384
            self.mrope_positions = self._make_buffer(
                (3, self.max_num_tokens + 1), dtype=torch.int64)
385

386
387
388
389
390
391
392
393
        # CUDA event to synchronize use of reused CPU tensors between steps
        # when async scheduling is enabled.
        self.prepare_inputs_event: Optional[torch.cuda.Event] = None
        if self.use_async_scheduling:
            self.prepare_inputs_event = torch.cuda.Event()
            # Start in a completed state.
            self.prepare_inputs_event.record(torch.cuda.default_stream())

394
395
        # None in the first PP rank. The rest are set after load_model.
        self.intermediate_tensors: Optional[IntermediateTensors] = None
396

397
        # OPTIMIZATION: Cache the tensors rather than creating them every step.
398
        # Keep in int64 to avoid overflow with long context
399
        self.arange_np = np.arange(max(self.max_num_reqs + 1,
400
401
                                       self.max_model_len,
                                       self.max_num_tokens),
402
                                   dtype=np.int64)
403

404
405
406
407
408
        # 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] = {}
409
410
411
412
413
414
        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(
                self.max_num_tokens, dtype=torch.int32, device=self.device)
415

416
417
418
419
420
421
        self.uniform_decode_query_len = 1 if not self.speculative_config else \
            1 + self.speculative_config.num_speculative_tokens

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

422
        self.mm_budget = MultiModalBudget(
423
424
425
            self.model_config,
            self.scheduler_config,
            self.mm_registry,
426
        ) if self.supports_mm_inputs else None
427

428
429
        self.reorder_batch_threshold: Optional[int] = None

430
431
432
433
434
        # 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()

435
436
437
        # Cached outputs.
        self._draft_token_ids: Optional[Union[list[list[int]],
                                              torch.Tensor]] = None
438
439
440
441
442
        self.transfer_event = torch.cuda.Event()
        self.sampled_token_ids_pinned_cpu = torch.empty(
            (self.max_model_len, 1),
            dtype=torch.int64,
            device="cpu",
443
            pin_memory=self.pin_memory)
444

445
446
447
448
449
450
451
452
453
454
    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]

455
456
457
458
459
    def _make_buffer(self,
                     *size: Union[int, torch.SymInt],
                     dtype: torch.dtype,
                     numpy: bool = True) -> CpuGpuBuffer:
        return CpuGpuBuffer(*size,
460
461
                            dtype=dtype,
                            device=self.device,
462
463
                            pin_memory=self.pin_memory,
                            with_numpy=numpy)
464

465
466
467
    def _init_model_kwargs(self, num_tokens: int):
        model_kwargs = dict[str, Any]()

468
        if not self.is_pooling_model:
469
470
            return model_kwargs

471
472
        num_reqs = self.input_batch.num_reqs
        pooling_params = self.input_batch.get_pooling_params()
473
474
475
476
477
478
479
480
481
482
483

        token_type_id_requests = dict[int, Any]()
        for i, param in enumerate(pooling_params):
            if param.extra_kwargs is not None and \
            (token_types := param.extra_kwargs.get(
                "compressed_token_type_ids")) is not None:
                token_type_id_requests[i] = token_types

        if len(token_type_id_requests) == 0:
            return model_kwargs

484
        seq_lens = self.seq_lens.gpu[:num_reqs]
485
486
487
488
489
490
491
492
493
494
495
        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(
            device=self.device)
        return model_kwargs

496
    def _may_reorder_batch(self, scheduler_output: "SchedulerOutput") -> None:
497
498
        """
        Update the order of requests in the batch based on the attention
499
        backend's needs. For example, some attention backends (namely MLA) may
500
501
502
503
504
505
        want to separate requests based on if the attention computation will be
        compute-bound or memory-bound.

        Args:
            scheduler_output: The scheduler output.
        """
506
507
508
509
510
511
512
513
        # 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

514
        if self.reorder_batch_threshold is not None:
515
516
517
            # NOTE(lucas): currently no backend supports the custom masking
            #  required for DCP with q_len > 1, so we assert here. Remove this
            #  assert once the custom mask is support is added to FA3.
518
519
520
            if self.dcp_world_size > 1:
                assert self.reorder_batch_threshold == 1, \
                    "DCP not support reorder_batch_threshold > 1 now."
521
522
523
524
            reorder_batch_to_split_decodes_and_prefills(
                self.input_batch,
                scheduler_output,
                decode_threshold=self.reorder_batch_threshold)
525

526
527
528
529
530
531
532
533
534
535
536
    # Note: used for model runner override.
    def _init_device_properties(self) -> None:
        """Initialize attributes from torch.cuda.get_device_properties
        """
        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()

537
    def _update_states(self, scheduler_output: "SchedulerOutput") -> None:
538
539
540
541
542
543
        """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.

544
545
        The SamplingMetadata is updated and copied to the GPU if there is a
        new/resumed/paused/finished request in the batch.
546
547
        """
        # Remove finished requests from the cached states.
548
549
        for req_id in scheduler_output.finished_req_ids:
            self.requests.pop(req_id, None)
550
551
552
553
554
555
556
        # 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:
557
            self.input_batch.remove_request(req_id)
558
559

        # Free the cached encoder outputs.
560
561
        for mm_hash in scheduler_output.free_encoder_mm_hashes:
            self.encoder_cache.pop(mm_hash, None)
562

563
564
565
566
567
568
569
570
571
572
573
574
575
        # 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:
576
            self.input_batch.remove_request(req_id)
577

578
        reqs_to_add: list[CachedRequestState] = []
579
        # Add new requests to the cached states.
580
581
582
        for new_req_data in scheduler_output.scheduled_new_reqs:
            req_id = new_req_data.req_id
            sampling_params = new_req_data.sampling_params
583
            pooling_params = new_req_data.pooling_params
584

585
586
            if sampling_params and \
                sampling_params.sampling_type == SamplingType.RANDOM_SEED:
587
588
589
590
591
                generator = torch.Generator(device=self.device)
                generator.manual_seed(sampling_params.seed)
            else:
                generator = None

592
593
            if self.is_pooling_model:
                assert pooling_params is not None
594
595
                task = pooling_params.task
                assert task is not None, "You did not set `task` in the API"
596

597
                model = cast(VllmModelForPooling, self.get_model())
598
                to_update = model.pooler.get_pooling_updates(task)
599
600
                to_update.apply(pooling_params)

601
            req_state = CachedRequestState(
602
                req_id=req_id,
603
                prompt_token_ids=new_req_data.prompt_token_ids,
604
                prompt_embeds=new_req_data.prompt_embeds,
605
                mm_features=new_req_data.mm_features,
606
                sampling_params=sampling_params,
607
                pooling_params=pooling_params,
608
                generator=generator,
609
610
                block_ids=new_req_data.block_ids,
                num_computed_tokens=new_req_data.num_computed_tokens,
611
                output_token_ids=[],
612
                lora_request=new_req_data.lora_request,
613
            )
614
615
            self.requests[req_id] = req_state

616
            # Only relevant for models using M-RoPE (e.g, Qwen2-VL)
617
            if self.uses_mrope:
618
                self._init_mrope_positions(req_state)
619

620
            reqs_to_add.append(req_state)
621

622
        # Update the states of the running/resumed requests.
623
        is_last_rank = get_pp_group().is_last_rank
624
625
        req_data = scheduler_output.scheduled_cached_reqs
        for i, req_id in enumerate(req_data.req_ids):
626
            req_state = self.requests[req_id]
627
628
629
            num_computed_tokens = req_data.num_computed_tokens[i]
            new_block_ids = req_data.new_block_ids[i]
            resumed_from_preemption = req_data.resumed_from_preemption[i]
630

631
            # Update the cached states.
632
            req_state.num_computed_tokens = num_computed_tokens
633
634
635
636
637
638
639
640
641
642
643
644
645
646
647
648
649

            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.
                num_new_tokens = (num_computed_tokens + len(new_token_ids) -
                                  req_state.num_tokens)
                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:
                    req_state.output_token_ids.extend(
                        new_token_ids[-num_new_tokens:])

650
            # Update the block IDs.
651
            if not resumed_from_preemption:
652
653
654
655
656
                if new_block_ids is not None:
                    # Append the new blocks to the existing block IDs.
                    for block_ids, new_ids in zip(req_state.block_ids,
                                                  new_block_ids):
                        block_ids.extend(new_ids)
657
            else:
658
                assert new_block_ids is not None
659
660
                # The request is resumed from preemption.
                # Replace the existing block IDs with the new ones.
661
                req_state.block_ids = new_block_ids
662
663
664
665
666
667

            req_index = self.input_batch.req_id_to_index.get(req_id)
            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.
668
                reqs_to_add.append(req_state)
669
670
671
672
                continue

            # Update the persistent batch.
            self.input_batch.num_computed_tokens_cpu[req_index] = (
673
                num_computed_tokens)
674
675
676
            if new_block_ids is not None:
                self.input_batch.block_table.append_row(
                    new_block_ids, req_index)
677
678
679
680
681
682
683

            # 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)
684
                self.input_batch.token_ids_cpu[
685
686
687
688
689
                    req_index,
                    start_token_index:end_token_index] = new_token_ids
                self.input_batch.num_tokens_no_spec[
                    req_index] = end_token_index
                self.input_batch.num_tokens[req_index] = end_token_index
690

691
692
693
694
695
696
697
698
699
700
701
702
            # Add spec_token_ids to token_ids_cpu.
            spec_token_ids = (
                scheduler_output.scheduled_spec_decode_tokens.get(req_id, ()))
            if spec_token_ids:
                num_spec_tokens = len(spec_token_ids)
                start_index = self.input_batch.num_tokens_no_spec[req_index]
                end_token_index = start_index + num_spec_tokens
                self.input_batch.token_ids_cpu[
                    req_index, start_index:end_token_index] = spec_token_ids
                # NOTE(woosuk): `num_tokens` here may include spec tokens.
                self.input_batch.num_tokens[req_index] += num_spec_tokens

703
704
        # Add the new or resumed requests to the persistent batch.
        # The smaller empty indices are filled first.
705
706
        for request in reqs_to_add:
            self.input_batch.add_request(request)
707

708
709
710
711
712
713
        # 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()
714

715
716
717
718
719
720
721
722
723
724
725
726
727
728
729
730
731
732
733
734
735
736
737
738
739
    def _update_states_after_model_execute(
            self, output_token_ids: torch.Tensor) -> None:
        """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.
        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()
        for i, num_tokens in enumerate(num_accepted_tokens):
            self.input_batch.num_accepted_tokens_cpu[i] = num_tokens

740
741
742
743
744
745
    def _init_mrope_positions(self, req_state: CachedRequestState):
        image_grid_thw = []
        video_grid_thw = []
        second_per_grid_ts = []
        audio_feature_lengths = []
        use_audio_in_video = False
746
747
748
749
        for mm_feature in req_state.mm_features:
            mm_item = mm_feature.data
            if mm_item is None:
                continue
750
751
752
753
754
755
756
757
758
759
760
761
            mm_input = mm_item.get_data()
            if (t := mm_input.get("image_grid_thw")) is not None:
                image_grid_thw.append(t.tolist())
            if (t := mm_input.get("video_grid_thw")) is not None:
                video_grid_thw.append(t.tolist())
            if (t := mm_input.get("second_per_grid_ts")) is not None:
                second_per_grid_ts.append(t)
            if (t := mm_input.get("audio_feature_lengths")) is not None:
                audio_feature_lengths.append(t)
            if mm_input.get("use_audio_in_video") is True:
                use_audio_in_video = True

762
763
764
765
766
767
768
769
770
771
772
773
774
775
776
777
778
779
780
781
782
783
        if supports_mrope(self.model):
            req_state.mrope_positions, req_state.mrope_position_delta = \
                self.model.get_mrope_input_positions(
                    req_state.prompt_token_ids,
                    hf_config=self.model_config.hf_config,
                    image_grid_thw=image_grid_thw,
                    video_grid_thw=video_grid_thw,
                    second_per_grid_ts=second_per_grid_ts,
                    audio_feature_lengths=audio_feature_lengths,
                    use_audio_in_video=use_audio_in_video,
                )
        else:
            req_state.mrope_positions, req_state.mrope_position_delta = \
                MRotaryEmbedding.get_input_positions_tensor(
                    req_state.prompt_token_ids,
                    hf_config=self.model_config.hf_config,
                    image_grid_thw=image_grid_thw,
                    video_grid_thw=video_grid_thw,
                    second_per_grid_ts=second_per_grid_ts,
                    audio_feature_lengths=audio_feature_lengths,
                    use_audio_in_video=use_audio_in_video,
                )
784

785
    def _extract_mm_kwargs(
786
        self,
787
788
        scheduler_output: "SchedulerOutput",
    ) -> BatchedTensorInputs:
789
        if not scheduler_output or not self.is_multimodal_raw_input_only_model:
790
            return {}
791

792
793
        mm_kwargs = list[MultiModalKwargsItem]()
        for req in scheduler_output.scheduled_new_reqs:
794
795
796
            for feature in req.mm_features:
                if feature.data is not None:
                    mm_kwargs.append(feature.data)
797

798
        # Input all modalities at once
799
        model = cast(SupportsMultiModal, self.model)
800
801
802
803
804
        mm_kwargs_combined: BatchedTensorInputs = {}
        for _, _, mm_kwargs_group in group_mm_kwargs_by_modality(
                mm_kwargs,
                device=self.device,
                pin_memory=self.pin_memory,
805
                merge_by_field_config=model.merge_by_field_config,
806
807
        ):
            mm_kwargs_combined.update(mm_kwargs_group)
808

809
        return mm_kwargs_combined
810

811
    def _dummy_mm_kwargs(self, num_seqs: int) -> BatchedTensorInputs:
812
        if not self.is_multimodal_raw_input_only_model:
813
            return {}
814

815
816
817
818
819
        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)
820

821
822
823
824
825
826
827
828
829
830
831
832
833
834
835
836
837
838
839
840
    def _get_cumsum_and_arange(
        self,
        num_tokens: np.ndarray,
        cumsum_dtype: Optional[np.dtype] = None,
    ) -> tuple[np.ndarray, np.ndarray]:
        """Get the cumulative sum and batched arange of the given array.
        # E.g., [2, 5, 3] -> ([2, 7, 10], [0, 1, 0, 1, 2, 3, 4, 0, 1, 2])
        # Equivalent to but faster than:
        # np.concatenate([np.arange(n) for n in num_tokens])
        """
        # Step 1. [2, 5, 3] -> [2, 7, 10]
        cu_num_tokens = np.cumsum(num_tokens, dtype=cumsum_dtype)
        total_num_tokens = cu_num_tokens[-1]
        # Step 2. [2, 7, 10] -> [0, 0, 2, 2, 2, 2, 2, 7, 7, 7]
        cumsums_offsets = np.repeat(cu_num_tokens - num_tokens, num_tokens)
        # Step 3. [0, 1, 0, 1, 2, 3, 4, 0, 1, 2]
        arange = self.arange_np[:total_num_tokens] - cumsums_offsets

        return cu_num_tokens, arange

841
842
843
844
845
846
847
848
849
850
851
    def _prepare_input_ids(self, total_num_scheduled_tokens: int,
                           cu_num_tokens: np.ndarray) -> None:
        """Prepare the input IDs for the current batch.
        
        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)
852
853
854
            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)
855
856
857
858
859
860
861
862
863
864
865
866
867
868
869
870
871
872
873
874
875
876
877
878
879
            return

        # Async scheduling case, where some decode requests from the previous
        # iteration won't have entries in input_ids_cpu and need to be copied
        # on the GPU from prev_sampled_token_ids.
        prev_req_id_to_index = self.input_batch.prev_req_id_to_index
        assert prev_req_id_to_index is not None
        flattened_indices = []
        prev_common_req_indices = []
        indices_match = True
        max_flattened_index = -1
        for req_id, cur_index in self.input_batch.req_id_to_index.items():
            if (prev_index := prev_req_id_to_index.get(req_id)) is not None:
                prev_common_req_indices.append(prev_index)
                # We need to compute the flattened input_ids index of the
                # last token in each common request.
                flattened_index = cu_num_tokens[cur_index].item() - 1
                flattened_indices.append(flattened_index)
                indices_match &= (prev_index == flattened_index)
                max_flattened_index = max(max_flattened_index, flattened_index)
        num_commmon_tokens = len(flattened_indices)
        if num_commmon_tokens < total_num_scheduled_tokens:
            # If not all requests are decodes from the last iteration,
            # We need to copy the input_ids_cpu to the GPU first.
            self.input_ids.copy_to_gpu(total_num_scheduled_tokens)
880
881
882
            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)
883
884
885
886
887
888
889
890
891
892
893
894
895
        if num_commmon_tokens == 0:
            # No requests in common with the previous iteration
            # So input_ids_cpu will have all the input ids.
            return
        if indices_match and max_flattened_index == (num_commmon_tokens - 1):
            # Common-case optimization: the batch is unchanged
            # and no reordering happened.
            # The indices are both the same permutation of 0..N-1 so
            # we can copy directly using a single slice.
            self.input_ids.gpu[:num_commmon_tokens].copy_(
                self.input_batch.prev_sampled_token_ids[:num_commmon_tokens,
                                                        0],
                non_blocking=True)
896
897
            if self.enable_prompt_embeds:
                self.is_token_ids.gpu[:num_commmon_tokens] = True
898
899
900
901
902
903
904
905
906
907
908
909
910
911
912
913
914
915
            return
        # Upload the index tensors asynchronously
        # so the scatter can be non-blocking.
        input_ids_index_tensor = torch.tensor(flattened_indices,
                                              dtype=torch.int64,
                                              pin_memory=self.pin_memory).to(
                                                  self.device,
                                                  non_blocking=True)
        prev_common_req_indices_tensor = torch.tensor(
            prev_common_req_indices,
            dtype=torch.int64,
            pin_memory=self.pin_memory).to(self.device, non_blocking=True)
        self.input_ids.gpu.scatter_(
            dim=0,
            index=input_ids_index_tensor,
            src=self.input_batch.prev_sampled_token_ids[
                prev_common_req_indices_tensor, 0])

916
917
918
919
920
921
922
923
924
925
926
927
928
929
930
931
932
933
    def _get_encoder_seq_lens(
        self,
        scheduler_output: "SchedulerOutput",
        kv_cache_spec: KVCacheSpec,
        num_reqs: int,
    ) -> Optional[np.ndarray]:
        if not isinstance(kv_cache_spec, CrossAttentionSpec):
            return None

        # Build encoder_seq_lens array mapping request indices to
        # encoder lengths for inputs scheduled in this batch
        encoder_seq_lens = np.zeros(num_reqs, dtype=np.int32)
        for req_id in scheduler_output.scheduled_encoder_inputs:
            req_index = self.input_batch.req_id_to_index[req_id]
            encoder_seq_lens[req_index] = self.max_encoder_len

        return encoder_seq_lens

934
    def _prepare_inputs(
935
936
937
938
        self, scheduler_output: "SchedulerOutput"
    ) -> tuple[PerLayerAttnMetadata, torch.Tensor,
               Optional[SpecDecodeMetadata], np.ndarray,
               Optional[CommonAttentionMetadata], int, Optional[UBatchSlices],
939
               Optional[torch.Tensor], bool]:
940
941
942
        """
        :return: tuple[
            attn_metadata: layer-to-attention_metadata mapping,
943
944
945
            logits_indices, spec_decode_metadata,
            num_scheduled_tokens, spec_decode_common_attn_metadata,
            max_num_scheduled_tokens, use_cascade_attn
946
947
        ]
        """
948
949
950
951
952
953
954
        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.
955
        self.input_batch.block_table.commit_block_table(num_reqs)
956
957

        # Get the number of scheduled tokens for each request.
958
959
960
961
        req_ids = self.input_batch.req_ids
        tokens = [scheduler_output.num_scheduled_tokens[i] for i in req_ids]
        num_scheduled_tokens = np.array(tokens, dtype=np.int32)
        max_num_scheduled_tokens = max(tokens)
962
963
964

        # Get request indices.
        # E.g., [2, 5, 3] -> [0, 0, 1, 1, 1, 1, 1, 2, 2, 2]
965
966
        req_indices = np.repeat(self.arange_np[:num_reqs],
                                num_scheduled_tokens)
967

968
969
970
971
        # cu_num_tokens: [2, 5, 3] -> [2, 7, 10]
        # arange: [0, 1, 0, 1, 2, 3, 4, 0, 1, 2]
        cu_num_tokens, arange = self._get_cumsum_and_arange(
            num_scheduled_tokens)
972
973

        # Get positions.
974
        positions_np = self.positions.np[:total_num_scheduled_tokens]
975
976
977
978
        np.add(self.input_batch.num_computed_tokens_cpu[req_indices],
               arange,
               out=positions_np)

979
980
        # Calculate M-RoPE positions.
        # Only relevant for models using M-RoPE (e.g, Qwen2-VL)
981
        if self.uses_mrope:
982
983
            self._calc_mrope_positions(scheduler_output)

984
985
986
987
        # 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.
988
989
        token_indices = (positions_np +
                         req_indices * self.input_batch.token_ids_cpu.shape[1])
990
        token_indices_tensor = torch.from_numpy(token_indices)
991

992
993
994
995
        # 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.
        torch.index_select(self.input_batch.token_ids_cpu_tensor.flatten(),
996
                           0,
997
                           token_indices_tensor,
998
                           out=self.input_ids.cpu[:total_num_scheduled_tokens])
999
1000
1001
1002
1003
1004
1005
        if self.enable_prompt_embeds:
            is_token_ids = self.input_batch.is_token_ids.flatten()
            torch.index_select(
                is_token_ids,
                0,
                token_indices_tensor,
                out=self.is_token_ids.cpu[:total_num_scheduled_tokens])
1006
1007
1008
1009
1010
1011
1012
1013
1014
1015
1016
1017
1018
1019
1020
1021
1022
1023
1024
1025
1026
1027
1028
1029
1030
1031
1032
1033
1034
1035
1036
1037
1038
1039
1040
1041
1042
1043
1044

        # 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:
                    self.inputs_embeds.cpu[output_idx:output_idx +
                                           actual_num_sched].copy_(
                                               req_embeds[start_pos:actual_end]
                                           )

                output_idx += num_sched
1045

1046
1047
1048
1049
        self.input_batch.block_table.compute_slot_mapping(
            req_indices, positions_np)
        self.input_batch.block_table.commit_slot_mapping(
            total_num_scheduled_tokens)
1050
1051

        # Prepare the attention metadata.
1052
1053
        self.query_start_loc.np[0] = 0
        self.query_start_loc.np[1:num_reqs + 1] = cu_num_tokens
1054
1055
        # Note: pad query_start_loc to be non-decreasing, as kernels
        # like FlashAttention requires that
1056
1057
1058
        self.query_start_loc.np[num_reqs + 1:].fill(cu_num_tokens[-1])
        self.query_start_loc.copy_to_gpu()
        query_start_loc = self.query_start_loc.gpu[:num_reqs + 1]
1059

1060
1061
1062
        num_tokens_unpadded = scheduler_output.total_num_scheduled_tokens
        num_tokens_padded = num_tokens_unpadded + self.get_local_padding(
            num_tokens_unpadded)
1063
1064
1065
        uniform_decode = \
            (max_num_scheduled_tokens == self.uniform_decode_query_len) and \
            (total_num_scheduled_tokens == num_reqs * max_num_scheduled_tokens)
1066
        ubatch_slices, num_tokens_after_padding = \
1067
            ubatch_split(num_scheduled_tokens,
1068
1069
                         num_tokens_unpadded,
                         num_tokens_padded,
1070
1071
                         uniform_decode=uniform_decode,
                         vllm_config=self.vllm_config)
1072

1073
        self.seq_lens.np[:num_reqs] = (
1074
1075
            self.input_batch.num_computed_tokens_cpu[:num_reqs] +
            num_scheduled_tokens)
1076
        # Fill unused with 0 for full cuda graph mode.
1077
1078
1079
1080
        self.seq_lens.np[num_reqs:].fill(0)
        self.seq_lens.copy_to_gpu()
        seq_lens = self.seq_lens.gpu[:num_reqs]
        max_seq_len = self.seq_lens.np[:num_reqs].max().item()
1081

1082
1083
1084
1085
1086
1087
1088
1089
1090
1091
1092
1093
1094
1095
1096
        num_tokens = [
            self.requests[r].num_tokens for r in self.input_batch.req_ids
        ]
        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)
        self.discard_request_indices.np[:self.num_discarded_requests] = (
            discard_request_indices)

        self.discard_request_indices.copy_to_gpu(self.num_discarded_requests)

1097
        # Copy the tensors to the GPU.
1098
1099
        self._prepare_input_ids(total_num_scheduled_tokens, cu_num_tokens)

1100
        if self.uses_mrope:
1101
            # Only relevant for models using M-RoPE (e.g, Qwen2-VL)
1102
1103
            self.mrope_positions.gpu[:, :total_num_scheduled_tokens].copy_(
                self.mrope_positions.cpu[:, :total_num_scheduled_tokens],
1104
1105
1106
                non_blocking=True)
        else:
            # Common case (1D positions)
1107
            self.positions.copy_to_gpu(total_num_scheduled_tokens)
1108

1109
1110
1111
1112
1113
1114
1115
1116
1117
        use_spec_decode = len(
            scheduler_output.scheduled_spec_decode_tokens) > 0
        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
1118
            num_draft_tokens = None
1119
1120
1121
1122
1123
1124
            spec_decode_metadata = None
        else:
            # Get the number of draft tokens for each request.
            # Iterate over the dictionary rather than all requests since not all
            # requests have draft tokens.
            num_draft_tokens = np.zeros(num_reqs, dtype=np.int32)
1125
1126
1127
            # 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)
1128
1129
1130
1131
            for req_id, draft_token_ids in (
                    scheduler_output.scheduled_spec_decode_tokens.items()):
                req_idx = self.input_batch.req_id_to_index[req_id]
                num_draft_tokens[req_idx] = len(draft_token_ids)
1132
1133
1134
                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)
1135
1136
1137
            spec_decode_metadata = self._calc_spec_decode_metadata(
                num_draft_tokens, cu_num_tokens)
            logits_indices = spec_decode_metadata.logits_indices
1138
1139
1140
1141
1142
1143

            # For DECODE only cuda graph of some attention backends (e.g., GDN).
            self.num_decode_draft_tokens.np[:
                                            num_reqs] = num_decode_draft_tokens
            self.num_decode_draft_tokens.np[num_reqs:].fill(-1)
            self.num_decode_draft_tokens.copy_to_gpu()
1144
1145
1146

        logits_indices_padded = None
        if self.cache_config.kv_sharing_fast_prefill:
1147
            logits_indices_padded = self._prepare_kv_sharing_fast_prefill(
1148
1149
                logits_indices)

1150
1151
1152
        attn_metadata: PerLayerAttnMetadata = {}
        if ubatch_slices is not None:
            attn_metadata = [dict() for _ in range(len(ubatch_slices))]
1153
        use_cascade_attn = False
1154

1155
        # Used in the below loop.
1156
1157
        query_start_loc_cpu = self.query_start_loc.cpu[:num_reqs + 1]
        seq_lens_cpu = self.seq_lens.cpu[:num_reqs]
1158
1159
1160
        num_computed_tokens_cpu = (
            self.input_batch.num_computed_tokens_cpu_tensor[:num_reqs])
        spec_decode_common_attn_metadata = None
1161
1162
1163
1164
1165
        if use_spec_decode:
            self.num_accepted_tokens.np[:num_reqs] = (
                self.input_batch.num_accepted_tokens_cpu[:num_reqs])
            self.num_accepted_tokens.np[num_reqs:].fill(1)
            self.num_accepted_tokens.copy_to_gpu()
1166

1167
1168
1169
1170
        # Prepare the attention metadata for each KV cache group and make layers
        # in the same group share the same metadata.
        for kv_cache_group_id, kv_cache_group_spec in enumerate(
                self.kv_cache_config.kv_cache_groups):
1171
1172
            encoder_seq_lens = self._get_encoder_seq_lens(
                scheduler_output, kv_cache_group_spec.kv_cache_spec, num_reqs)
1173

1174
1175
1176
1177
1178
1179
1180
            if isinstance(kv_cache_group_spec.kv_cache_spec,
                          EncoderOnlyAttentionSpec):
                # 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,
1181
1182
1183
1184
1185
1186
1187
                    device=self.device,
                )
                slot_mapping = torch.zeros(
                    (total_num_scheduled_tokens, ),
                    dtype=torch.int64,
                    device=self.device,
                )
1188
1189
1190
                num_common_prefix_blocks = 0
            else:
                blk_table = self.input_batch.block_table[kv_cache_group_id]
1191
1192
1193
                blk_table_tensor = blk_table.get_device_tensor(num_reqs)
                slot_mapping = blk_table.slot_mapping.gpu[:
                                                          total_num_scheduled_tokens]
1194
1195
1196

                # Fill unused with -1. Needed for reshape_and_cache in full cuda
                # graph mode.
1197
1198
                blk_table.slot_mapping.gpu[total_num_scheduled_tokens:].fill_(
                    -1)
1199
1200
1201
                num_common_prefix_blocks = (
                    scheduler_output.
                    num_common_prefix_blocks[kv_cache_group_id])
1202

1203
            common_attn_metadata = CommonAttentionMetadata(
1204
1205
1206
1207
1208
                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,
1209
1210
1211
                num_reqs=num_reqs,
                num_actual_tokens=total_num_scheduled_tokens,
                max_query_len=max_num_scheduled_tokens,
1212
                max_seq_len=max_seq_len,
1213
1214
                block_table_tensor=blk_table_tensor,
                slot_mapping=slot_mapping,
1215
1216
                logits_indices_padded=logits_indices_padded,
                num_logits_indices=logits_indices.size(0),
1217
                causal=True,
1218
                encoder_seq_lens=encoder_seq_lens,
1219
1220
            )

1221
1222
1223
1224
1225
1226
1227
1228
            if (self.speculative_config
                    and spec_decode_common_attn_metadata is None):
                if isinstance(self.drafter, EagleProposer):
                    if (self.drafter.attn_layer_names[0]
                            in kv_cache_group_spec.layer_names):
                        spec_decode_common_attn_metadata = common_attn_metadata
                else:
                    spec_decode_common_attn_metadata = common_attn_metadata
1229

1230
1231
1232
            for attn_group in self.attn_groups[kv_cache_group_id]:
                # Prepare for cascade attention if enabled & beneficial.
                common_prefix_len = 0
1233
                builder = attn_group.get_metadata_builder()
1234
1235
1236
                if self.cascade_attn_enabled:
                    common_prefix_len = self._compute_cascade_attn_prefix_len(
                        num_scheduled_tokens,
1237
                        num_common_prefix_blocks,
1238
                        attn_group.kv_cache_spec,
1239
1240
                        builder,
                    )
1241

1242
1243
1244
1245
1246
1247
                extra_attn_metadata_args = {}
                if use_spec_decode and isinstance(builder,
                                                  GDNAttentionMetadataBuilder):
                    extra_attn_metadata_args = dict(
                        num_accepted_tokens=self.num_accepted_tokens.
                        gpu[:num_reqs],
1248
1249
                        num_decode_draft_tokens_cpu=self.
                        num_decode_draft_tokens.cpu[:num_reqs],
1250
1251
                    )

1252
1253
1254
1255
1256
1257
1258
1259
1260
1261
1262
1263
1264
1265
1266
1267
1268
1269
                if ubatch_slices is not None:
                    common_attn_metadata_list = split_attn_metadata(
                        ubatch_slices, common_attn_metadata)
                    for ubid, common_attn_metadata in enumerate(
                            common_attn_metadata_list):
                        attn_metadata_i = (attn_group.get_metadata_builder(
                            ubatch_id=ubid).build(
                                common_prefix_len=common_prefix_len,
                                common_attn_metadata=common_attn_metadata))
                        for layer_name in kv_cache_group_spec.layer_names:
                            assert type(attn_metadata) is list
                            attn_metadata[ubid][layer_name] = attn_metadata_i
                else:
                    assert isinstance(attn_metadata, dict)
                    attn_metadata_i = builder.build(
                        common_prefix_len=common_prefix_len,
                        common_attn_metadata=common_attn_metadata,
                        **extra_attn_metadata_args)
1270
1271
                    use_cascade_attn |= getattr(attn_metadata_i, "use_cascade",
                                                False)
1272
1273
                    for layer_name in attn_group.layer_names:
                        attn_metadata[layer_name] = attn_metadata_i
1274

1275
1276
1277
1278
        # disable cascade attention when DBO
        if ubatch_slices is not None:
            use_cascade_attn = False

1279
1280
1281
1282
        # Hot-Swap lora model
        if self.lora_config:
            self.set_active_loras(self.input_batch, num_scheduled_tokens)

1283
1284
        return (attn_metadata, logits_indices, spec_decode_metadata,
                num_scheduled_tokens, spec_decode_common_attn_metadata,
1285
                max_num_scheduled_tokens, ubatch_slices,
1286
                num_tokens_after_padding, use_cascade_attn)
1287

1288
1289
1290
1291
    def _compute_cascade_attn_prefix_len(
        self,
        num_scheduled_tokens: np.ndarray,
        num_common_prefix_blocks: int,
1292
1293
        kv_cache_spec: KVCacheSpec,
        attn_metadata_builder: AttentionMetadataBuilder,
1294
1295
1296
1297
1298
1299
1300
1301
1302
1303
1304
1305
1306
1307
1308
1309
1310
1311
    ) -> 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.
        """
1312
        common_prefix_len = num_common_prefix_blocks * kv_cache_spec.block_size
1313
1314
1315
1316
1317
1318
1319
1320
1321
1322
1323
1324
1325
1326
1327
1328
1329
1330
1331
1332
1333
1334
1335
1336
1337
1338
1339
1340
1341
1342
1343
1344
1345
1346
1347
1348
1349
        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]
1350
        # Request 3's num_computed_tokens: 3 (i.e., [A, B, C])
1351
1352
1353
1354
1355
1356
1357
1358
1359
1360
        # 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(
            common_prefix_len,
            self.input_batch.num_computed_tokens_cpu[:num_reqs].min())
        # common_prefix_len should be a multiple of the block size.
1361
1362
1363
1364
1365
        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))
1366
1367
1368
1369
        use_local_attention = (
            isinstance(kv_cache_spec, ChunkedLocalAttentionSpec)
            or (isinstance(kv_cache_spec, FullAttentionSpec)
                and kv_cache_spec.attention_chunk_size is not None))
1370
1371
        assert isinstance(kv_cache_spec, AttentionSpec)
        use_cascade = attn_metadata_builder.use_cascade_attention(
1372
1373
1374
            common_prefix_len=common_prefix_len,
            query_lens=num_scheduled_tokens,
            num_query_heads=self.num_query_heads,
1375
            num_kv_heads=kv_cache_spec.num_kv_heads,
1376
            use_alibi=self.use_alibi,
1377
            use_sliding_window=use_sliding_window,
1378
            use_local_attention=use_local_attention,
1379
1380
1381
1382
            num_sms=self.num_sms,
        )
        return common_prefix_len if use_cascade else 0

1383
1384
    def _calc_mrope_positions(self, scheduler_output: "SchedulerOutput"):
        mrope_pos_ptr = 0
1385
        for index, req_id in enumerate(self.input_batch.req_ids):
1386
1387
1388
1389
1390
1391
1392
            req = self.requests[req_id]
            assert req.mrope_positions is not None

            num_computed_tokens = \
                self.input_batch.num_computed_tokens_cpu[index]
            num_scheduled_tokens = \
                scheduler_output.num_scheduled_tokens[req_id]
1393
1394
            num_prompt_tokens = length_from_prompt_token_ids_or_embeds(
                req.prompt_token_ids, req.prompt_embeds)
1395
1396
1397
1398
1399
1400
1401
1402
1403
1404
1405
1406
1407
1408
1409
1410
1411
1412
1413

            if num_computed_tokens + num_scheduled_tokens > num_prompt_tokens:
                prompt_part_len = max(0,
                                      num_prompt_tokens - num_computed_tokens)
                completion_part_len = max(
                    0, num_scheduled_tokens - prompt_part_len)
            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

1414
1415
                self.mrope_positions.cpu[:, dst_start:dst_end] = (
                    req.mrope_positions[:, src_start:src_end])
1416
1417
1418
1419
1420
1421
1422
                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

1423
                MRotaryEmbedding.get_next_input_positions_tensor(
1424
                    out=self.mrope_positions.np,
1425
1426
1427
1428
1429
                    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,
                )
1430
1431
1432

                mrope_pos_ptr += completion_part_len

1433
1434
    def _calc_spec_decode_metadata(
        self,
1435
1436
1437
1438
1439
1440
1441
1442
1443
1444
1445
1446
1447
1448
1449
1450
        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
1451
1452
1453
1454
1455
1456

        # 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(
            num_sampled_tokens, cumsum_dtype=np.int32)
        # Step 2. [0, 0, 0, 0, 103, 104, 104, 104, 206, 207, 207]
1457
1458
        logits_indices = np.repeat(
            cu_num_scheduled_tokens - num_sampled_tokens, num_sampled_tokens)
1459
        # Step 3. [0, 1, 2, 3, 103, 104, 105, 106, 206, 207, 208]
1460
1461
1462
1463
1464
1465
        logits_indices += arange

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

        # Compute the draft logits indices.
1466
1467
1468
1469
        # 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(
            num_draft_tokens, cumsum_dtype=np.int32)
1470
1471
1472
1473
1474
1475
1476
1477
1478
1479
1480
1481
1482
1483
        # [0, 0, 0, 5, 5, 9]
        target_logits_indices = np.repeat(
            cu_num_sampled_tokens - num_sampled_tokens, num_draft_tokens)
        # [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(
            self.device, non_blocking=True)
        logits_indices = torch.from_numpy(logits_indices).to(self.device,
                                                             non_blocking=True)
        target_logits_indices = torch.from_numpy(target_logits_indices).to(
            self.device, non_blocking=True)
        bonus_logits_indices = torch.from_numpy(bonus_logits_indices).to(
1484
1485
            self.device, non_blocking=True)

1486
1487
        # Compute the draft token ids.
        # draft_token_indices:      [  1,   2,   3, 105, 106, 208]
1488
        draft_token_ids = self.input_ids.gpu[logits_indices]
1489
1490
1491
1492
1493
1494
1495
1496
1497
1498
1499
1500
        draft_token_ids = draft_token_ids[target_logits_indices + 1]

        metadata = SpecDecodeMetadata(
            draft_token_ids=draft_token_ids,
            num_draft_tokens=num_draft_tokens.tolist(),
            cu_num_draft_tokens=cu_num_draft_tokens,
            target_logits_indices=target_logits_indices,
            bonus_logits_indices=bonus_logits_indices,
            logits_indices=logits_indices,
        )
        return metadata

1501
1502
1503
1504
1505
1506
1507
1508
1509
1510
1511
1512
1513
1514
1515
1516
1517
1518
1519
1520
1521
1522
1523
1524
1525
1526
    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
        self.kv_sharing_fast_prefill_logits_indices[:num_logits].copy_(
            logits_indices)
        # 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_(
            logits_indices[-1].item())
        if (self.compilation_config.cudagraph_mode != CUDAGraphMode.NONE
                and num_logits <= self.cudagraph_batch_sizes[-1]):
            # 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
        logits_indices_padded = (
            self.kv_sharing_fast_prefill_logits_indices[:num_logits_padded])
        return logits_indices_padded

1527
1528
1529
1530
1531
1532
1533
1534
    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
1535
                inputs.
1536
1537
1538
1539
1540
1541

        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
        """
1542
1543
        scheduled_encoder_inputs = scheduler_output.scheduled_encoder_inputs
        if not scheduled_encoder_inputs:
1544
            return [], []
1545
        # Batch the multi-modal inputs.
1546
        mm_kwargs = list[MultiModalKwargsItem]()
1547
1548
        # list of tuple (mm_hash, position_info)
        mm_hashes_pos = list[tuple[str, PlaceholderRange]]()
1549
1550
        for req_id, encoder_input_ids in scheduled_encoder_inputs.items():
            req_state = self.requests[req_id]
1551
1552

            for mm_input_id in encoder_input_ids:
1553
1554
1555
1556
                mm_feature = req_state.mm_features[mm_input_id]
                mm_hash = mm_feature.identifier
                mm_kwargs.append(mm_feature.data)
                mm_hashes_pos.append((mm_hash, mm_feature.mm_position))
1557

1558
1559
1560
1561
1562
1563
1564
1565
1566
1567
        return mm_kwargs, mm_hashes_pos

    def _execute_mm_encoder(self, scheduler_output: "SchedulerOutput"):
        # Batch the multi-modal inputs using the helper method.
        mm_kwargs, mm_hashes_pos = self._batch_mm_kwargs_from_scheduler(
            scheduler_output)

        if not mm_kwargs:
            return

1568
1569
1570
1571
1572
1573
1574
        # 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.
1575
        model = cast(SupportsMultiModal, self.model)
1576
        encoder_outputs = []
1577
        for modality, num_items, mm_kwargs_group in group_mm_kwargs_by_modality(
1578
                mm_kwargs,
1579
                device=self.device,
1580
                pin_memory=self.pin_memory,
1581
                merge_by_field_config=model.merge_by_field_config,
1582
        ):
1583
1584
1585
1586
1587
1588
1589
1590
1591
1592
1593
1594
1595
1596
1597
1598
1599
1600
1601
1602
1603
1604
1605
1606
1607
1608
1609
1610
1611
            # (ekhvedchenia): Temporary hack to limit peak memory usage when
            # processing multimodal data.This solves the issue with scheduler
            # putting too many video samples into a single batch. Scheduler
            # uses pruned vision tokens count to compare it versus compute
            # budget which is incorrect (Either input media size or non-pruned
            # output vision tokens count should be considered)
            curr_group_outputs = []

            if self.is_multimodal_pruning_enabled and modality == "video":
                micro_batch_size = 1
                for i in range(0, num_items, micro_batch_size):
                    micro_batch_mm_inputs = dict(
                        (k, v[i:i + micro_batch_size])
                        for k, v in mm_kwargs_group.items())

                    micro_batch_outputs = model.get_multimodal_embeddings(
                        **micro_batch_mm_inputs)

                    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.
                curr_group_outputs = model.get_multimodal_embeddings(
                    **mm_kwargs_group)
1612

1613
1614
            sanity_check_mm_encoder_outputs(
                curr_group_outputs,
1615
                expected_num_items=num_items,
1616
            )
1617
            encoder_outputs.extend(curr_group_outputs)
1618

1619
1620
1621
        # 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(
1622
1623
1624
1625
1626
                output,
                is_embed=pos_info.is_embed,
            )

    def _gather_mm_embeddings(
1627
1628
        self,
        scheduler_output: "SchedulerOutput",
1629
        shift_computed_tokens: int = 0,
1630
    ) -> list[torch.Tensor]:
1631
        should_sync_mrope_positions = False
1632
        mm_embeds: list[torch.Tensor] = []
1633
        for req_id in self.input_batch.req_ids:
1634
1635
            mm_embeds_req: list[torch.Tensor] = []

1636
1637
1638
            num_scheduled_tokens = scheduler_output.num_scheduled_tokens[
                req_id]
            req_state = self.requests[req_id]
1639
1640
            num_computed_tokens = \
                req_state.num_computed_tokens + shift_computed_tokens
1641
1642
            for mm_feature in req_state.mm_features:
                pos_info = mm_feature.mm_position
1643
1644
                start_pos = pos_info.offset
                num_encoder_tokens = pos_info.length
1645
1646
1647
1648
1649
1650
1651
1652
1653
1654
1655
1656
1657
1658
1659
1660

                # 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,
1661
1662
                    num_encoder_tokens,
                )
1663
                assert start_idx < end_idx
1664

1665
                mm_hash = mm_feature.identifier
1666
1667
1668
                encoder_output = self.encoder_cache.get(mm_hash, None)
                assert encoder_output is not None,\
                    f"Encoder cache miss for {mm_hash}."
1669
1670
1671
1672
1673
1674
1675
1676

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

                mm_embeds_item = gather_mm_placeholders(
                    encoder_output[start_idx:end_idx],
                    is_embed=is_embed,
                )
1677
1678
1679
1680
1681
1682
1683
1684
1685
1686
1687
1688
1689
1690
1691
1692
1693
1694
1695
1696
1697
1698
                mm_embeds_req.append(mm_embeds_item)

            if self.is_multimodal_pruning_enabled and self.uses_mrope:
                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,
                    ))
                assert req_state.mrope_positions is not None
                req_state.mrope_positions.copy_(new_mrope_positions)
                req_state.mrope_position_delta = new_delta

            mm_embeds.extend(mm_embeds_req)

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

1699
        return mm_embeds
1700

1701
1702
1703
1704
1705
1706
1707
1708
1709
1710
1711
1712
1713
1714
1715
1716
    def _extract_encoder_inputs(
        self,
        scheduler_output: "SchedulerOutput",
    ) -> dict[str, torch.Tensor]:
        """Extract encoder inputs for encoder-decoder models.

        This method extracts multimodal input features from scheduled encoder
        inputs and formats them for the encoder-decoder model forward pass.
        """
        # Batch the multi-modal inputs using the helper method.
        mm_kwargs, _ = self._batch_mm_kwargs_from_scheduler(scheduler_output)

        if not mm_kwargs:
            return {}

        # Group MM kwargs by modality and extract features
1717
        model = cast(SupportsMultiModal, self.model)
1718
1719
1720
1721
1722
        encoder_features = {}
        for _, _, mm_kwargs_group in group_mm_kwargs_by_modality(
                mm_kwargs,
                device=self.device,
                pin_memory=self.pin_memory,
1723
                merge_by_field_config=model.merge_by_field_config,
1724
1725
1726
1727
1728
1729
1730
1731
        ):
            # Add the grouped features to encoder_features dict
            # This allows the model to receive them as kwargs (e.g.,
            # input_features=...)
            encoder_features.update(mm_kwargs_group)

        return encoder_features

1732
    def get_model(self) -> nn.Module:
1733
        # get raw model out of the cudagraph wrapper.
1734
        if isinstance(self.model, (CUDAGraphWrapper, UBatchWrapper)):
1735
            return self.model.unwrap()
1736
1737
        return self.model

1738
1739
1740
1741
1742
1743
1744
1745
1746
1747
1748
1749
1750
1751
1752
    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

1753
1754
1755
1756
1757
    def get_supported_pooling_tasks(self) -> list[PoolingTask]:
        model = self.get_model()
        if not is_pooling_model(model):
            return []

1758
1759
1760
1761
1762
1763
        supported_tasks = list(model.pooler.get_supported_tasks())

        if (self.scheduler_config.chunked_prefill_enabled
                and "encode" in supported_tasks):
            supported_tasks.remove("encode")

1764
1765
1766
1767
1768
1769
1770
1771
1772
1773
1774
            logger.debug_once("Chunked prefill is not supported with "
                              "encode task which using ALL pooling. "
                              "Please turn off chunked prefill by "
                              "`--no-enable-chunked-prefill` before using it.")

        if "score" in supported_tasks:
            num_labels = getattr(self.model_config.hf_config, "num_labels", 0)
            if num_labels != 1:
                supported_tasks.remove("score")
                logger.debug_once(
                    "Score API is only enabled for num_labels == 1.")
1775
1776

        return supported_tasks
1777

1778
1779
1780
1781
1782
1783
1784
1785
1786
1787
    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)

1788
1789
1790
1791
1792
1793
1794
    def sync_and_slice_intermediate_tensors(
            self, num_tokens: int, intermediate_tensors: IntermediateTensors,
            sync_self: bool) -> IntermediateTensors:

        assert self.intermediate_tensors is not None

        tp = self.vllm_config.parallel_config.tensor_parallel_size
1795
        is_rs = is_residual_scattered_for_sp(self.vllm_config, num_tokens)
1796
1797
1798
1799
1800
1801

        # 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():
1802
                is_scattered = k == "residual" and is_rs
1803
1804
1805
1806
1807
1808
1809
                copy_len = num_tokens // tp if is_scattered else \
                    num_tokens
                self.intermediate_tensors[k][:copy_len].copy_(
                    v[:copy_len], non_blocking=True)

        return IntermediateTensors({
            k:
1810
1811
            v[:num_tokens //
              tp] if k == "residual" and is_rs else v[:num_tokens]
1812
1813
1814
            for k, v in self.intermediate_tensors.items()
        })

1815
1816
1817
1818
1819
1820
1821
1822
1823
1824
    def eplb_step(self,
                  is_dummy: bool = False,
                  is_profile: bool = False) -> None:
        """
        Step for the EPLB (Expert Parallelism Load Balancing) state.
        """
        if not self.parallel_config.enable_eplb:
            return

        assert self.eplb_state is not None
1825
1826
        model = self.get_model()
        assert is_mixture_of_experts(model)
1827
        self.eplb_state.step(
1828
            model,
1829
1830
            is_dummy,
            is_profile,
1831
            log_stats=self.parallel_config.eplb_config.log_balancedness,
1832
1833
        )

1834
1835
    def get_dp_padding(self,
                       num_tokens: int) -> tuple[int, Optional[torch.Tensor]]:
1836
1837
1838
1839
1840
1841
1842
1843
1844
1845
1846
        """
        Determines the total number of tokens that each rank will run.
        All ranks will be padded out so that they run with the same number
        of tokens

        Returns: tuple[
            num_pad_tokens: The number of tokens that will be added to the batch
            num_tokens_after_padding: A tensor containing the total number of
            tokens for each DP rank including padding.
        ]
        """
1847
1848
        dp_size = self.vllm_config.parallel_config.data_parallel_size
        dp_rank = self.vllm_config.parallel_config.data_parallel_rank
1849
1850
1851
1852
1853
1854
1855
1856
1857

        # For DP: Don't pad when setting enforce_eager.
        # 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.
        #
        # TODO(tms) : There are many cases where padding is enabled for
        # prefills, causing unnecessary and excessive padding of activations.

        if dp_size == 1 or self.vllm_config.model_config.enforce_eager:
1858
            # Early exit.
1859
            return 0, None
1860
1861
1862
1863

        num_tokens_across_dp = DPMetadata.num_tokens_across_dp(
            num_tokens, dp_size, dp_rank)
        max_tokens_across_dp_cpu = torch.max(num_tokens_across_dp).item()
1864
1865
1866
1867
1868
        num_tokens_after_padding = torch.tensor([max_tokens_across_dp_cpu] *
                                                dp_size,
                                                device="cpu",
                                                dtype=torch.int32)
        return max_tokens_across_dp_cpu - num_tokens, num_tokens_after_padding
1869

1870
1871
1872
1873
1874
1875
1876
1877
1878
1879
1880
1881
1882
1883
1884
1885
1886
1887
1888
1889
1890
1891
1892
1893
1894
1895
1896
1897
1898
1899
1900
1901
1902
    def get_local_padding(self, num_tokens_unpadded: int) -> int:

        num_tokens_padded = num_tokens_unpadded

        if (self.compilation_config.cudagraph_mode != CUDAGraphMode.NONE
                and num_tokens_unpadded <= self.cudagraph_batch_sizes[-1]):
            # Use piecewise CUDA graphs.
            # Add padding to the batch size.
            num_tokens_padded = self.vllm_config.pad_for_cudagraph(
                num_tokens_unpadded)
        else:
            # 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
            if self.vllm_config.compilation_config.pass_config. \
                enable_sequence_parallelism and tp_size > 1:
                num_tokens_padded = round_up(num_tokens_unpadded, tp_size)

        num_pad_tokens = num_tokens_padded - num_tokens_unpadded
        return num_pad_tokens

    # 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)
    def pad_out_ubatch_slice(self, ubatch_slices: UBatchSlices,
                             num_total_tokens: int):
        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)

1903
1904
1905
1906
1907
1908
1909
1910
1911
1912
1913
    def _pool(
        self,
        hidden_states: torch.Tensor,
        num_scheduled_tokens: int,
        num_scheduled_tokens_np: np.ndarray,
    ) -> ModelRunnerOutput:
        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"

1914
        hidden_states = hidden_states[:num_scheduled_tokens]
1915
        pooling_metadata = self.input_batch.get_pooling_metadata()
1916
1917
        pooling_metadata.build_pooling_cursor(num_scheduled_tokens_np.tolist(),
                                              device=hidden_states.device)
1918
        seq_lens_cpu = self.seq_lens.cpu[:self.input_batch.num_reqs]
1919

1920
1921
1922
1923
1924
1925
1926
1927
1928
1929
        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()
1930
1931
1932

        pooler_output: list[Optional[torch.Tensor]] = []
        for raw_output, seq_len, prompt_len in zip(
1933
                raw_pooler_output, seq_lens_cpu, pooling_metadata.prompt_lens):
1934

1935
            output = raw_output if seq_len == prompt_len else None
1936
            pooler_output.append(output)
1937
1938
1939
1940
1941
1942
1943
1944
1945
1946

        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,
        )

1947
1948
1949
1950
1951
1952
1953
1954
1955
1956
1957
1958
1959
1960
1961
1962
1963
1964
1965
    def _get_num_input_tokens(self, num_scheduled_tokens: int) -> int:
        if (self.compilation_config.cudagraph_mode != CUDAGraphMode.NONE
                and not envs.VLLM_DISABLE_PAD_FOR_CUDAGRAPH
                and hasattr(self, "cudagraph_batch_sizes")
                and self.cudagraph_batch_sizes
                and num_scheduled_tokens <= self.cudagraph_batch_sizes[-1]):
            # 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
        if (self.compilation_config.pass_config.enable_sequence_parallelism
                and tp_size > 1):
            return round_up(num_scheduled_tokens, tp_size)
        return num_scheduled_tokens

1966
    def _preprocess(
1967
1968
        self,
        scheduler_output: "SchedulerOutput",
1969
        intermediate_tensors: Optional[IntermediateTensors] = None,
1970
1971
        ubatch_slices: Optional[UBatchSlices] = None,
        num_tokens_after_padding: Optional[torch.Tensor] = None,
1972
1973
1974
    ) -> tuple[int, int, Optional[torch.Tensor], Optional[torch.Tensor],
               Optional[torch.Tensor], torch.Tensor,
               Optional[IntermediateTensors], dict[str, Any]]:
1975

1976
        num_scheduled_tokens = scheduler_output.total_num_scheduled_tokens
1977
1978
1979
1980
1981
1982
1983
1984
1985
        if ubatch_slices:
            assert num_tokens_after_padding is not None
            num_input_tokens = int(num_tokens_after_padding[0].item() * 2)
            self.pad_out_ubatch_slice(ubatch_slices, num_input_tokens)
        elif ubatch_slices is None:
            num_input_tokens = self._get_num_input_tokens(num_scheduled_tokens)
            num_pad, num_tokens_after_padding = self.get_dp_padding(
                num_input_tokens)
            num_input_tokens += num_pad
1986

1987
1988
        # _prepare_inputs may reorder the batch, so we must gather multi
        # modal outputs after that to ensure the correct order
1989
1990
        if (self.supports_mm_inputs and get_pp_group().is_first_rank
                and not self.model_config.is_encoder_decoder):
1991
1992
1993
1994
            # Run the multimodal encoder if any.
            self._execute_mm_encoder(scheduler_output)
            mm_embeds = self._gather_mm_embeddings(scheduler_output)

1995
1996
1997
            # 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.
1998
            inputs_embeds_scheduled = self.model.get_input_embeddings(
1999
                input_ids=self.input_ids.gpu[:num_scheduled_tokens],
2000
2001
                multimodal_embeddings=mm_embeds or None,
            )
2002

2003
            # TODO(woosuk): Avoid the copy. Optimize.
2004
            self.inputs_embeds.gpu[:num_scheduled_tokens].copy_(
2005
2006
                inputs_embeds_scheduled)

2007
            input_ids = None
2008
            inputs_embeds = self.inputs_embeds.gpu[:num_input_tokens]
2009
2010
2011
2012
            model_kwargs = {
                **self._init_model_kwargs(num_scheduled_tokens),
                **self._extract_mm_kwargs(scheduler_output),
            }
2013
        elif self.enable_prompt_embeds and get_pp_group().is_first_rank:
2014
2015
2016
2017
2018
2019
2020
2021
2022
2023
2024
2025
2026
2027
2028
2029
2030
2031
2032
2033
2034
2035
2036
2037
2038
            # 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).
            token_ids_idx = self.is_token_ids.gpu[:num_scheduled_tokens] \
                .nonzero(as_tuple=False) \
                .squeeze(1)
            # Some tokens ids may need to become embeds
            if token_ids_idx.numel() > 0:
                token_ids = self.input_ids.gpu[token_ids_idx]
                tokens_to_embeds = self.model.get_input_embeddings(
                    input_ids=token_ids)
                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
2039
        else:
2040
2041
2042
2043
            # 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.
2044
            input_ids = self.input_ids.gpu[:num_input_tokens]
2045
            inputs_embeds = None
2046
            model_kwargs = self._init_model_kwargs(num_input_tokens)
2047
        if self.uses_mrope:
2048
            positions = self.mrope_positions.gpu[:, :num_input_tokens]
2049
        else:
2050
            positions = self.positions.gpu[:num_input_tokens]
2051

2052
2053
2054
        if get_pp_group().is_first_rank:
            intermediate_tensors = None
        else:
2055
2056
            intermediate_tensors = self.sync_and_slice_intermediate_tensors(
                num_input_tokens, intermediate_tensors, True)
2057

2058
2059
2060
2061
2062
        if (self.model_config.is_encoder_decoder
                and scheduler_output.scheduled_encoder_inputs):
            encoder_inputs = self._extract_encoder_inputs(scheduler_output)
            model_kwargs.update(encoder_inputs)

2063
2064
2065
        return (
            num_scheduled_tokens,
            num_input_tokens,
2066
            num_tokens_after_padding,
2067
2068
2069
2070
2071
2072
            input_ids,
            inputs_embeds,
            positions,
            intermediate_tensors,
            model_kwargs,
        )
2073

2074
2075
2076
2077
    def _sample(
            self, logits: Optional[torch.Tensor],
            spec_decode_metadata: Optional[SpecDecodeMetadata]
    ) -> SamplerOutput:
2078
        # Sample the next token and get logprobs if needed.
2079
        sampling_metadata = self.input_batch.sampling_metadata
2080
        if spec_decode_metadata is None:
2081
            sampler_output = self.sampler(
2082
2083
2084
2085
                logits=logits,
                sampling_metadata=sampling_metadata,
            )
        else:
2086
2087
2088
2089
            # When indexing with a tensor (bonus_logits_indices), PyTorch
            # creates a new tensor with separate storage from the original
            # logits tensor. This means any in-place operations on bonus_logits
            # won't affect the original logits tensor.
2090
            assert logits is not None
2091
            bonus_logits = logits[spec_decode_metadata.bonus_logits_indices]
2092
            sampler_output = self.sampler(
2093
                logits=bonus_logits,
2094
2095
2096
                sampling_metadata=sampling_metadata,
            )
            bonus_token_ids = sampler_output.sampled_token_ids
2097

2098
2099
2100
            # Just like `bonus_logits`, `target_logits` is a new tensor with
            # separate storage from the original `logits` tensor. Therefore,
            # it is safe to update `target_logits` in place.
2101
            target_logits = logits[spec_decode_metadata.target_logits_indices]
2102
            output_token_ids = self.rejection_sampler(
2103
                spec_decode_metadata,
2104
                None,  # draft_probs
2105
                target_logits,
2106
                bonus_token_ids,
2107
2108
                sampling_metadata,
            )
2109
            sampler_output.sampled_token_ids = output_token_ids
2110
            self._update_states_after_model_execute(output_token_ids)
2111

2112
2113
2114
2115
2116
2117
2118
2119
2120
2121
2122
2123
2124
2125
2126
        return sampler_output

    def _bookkeeping_sync(
        self, scheduler_output: "SchedulerOutput",
        sampler_output: SamplerOutput, logits: Optional[torch.Tensor],
        hidden_states: torch.Tensor, num_scheduled_tokens: int
    ) -> tuple[
            dict[str, int],
            Optional[LogprobsLists],
            list[list[int]],
            dict[str, Optional[LogprobsTensors]],
            list[str],
            dict[str, int],
            list[int],
    ]:
2127
2128
2129
2130
        num_nans_in_logits = {}
        if envs.VLLM_COMPUTE_NANS_IN_LOGITS:
            num_nans_in_logits = self._get_nans_in_logits(logits)

2131
2132
2133
2134
2135
2136
        discard_sampled_tokens_req_indices = \
            self.discard_request_indices.np[:self.num_discarded_requests]
        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)
2137

2138
2139
2140
2141
2142
2143
        # 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()
        req_id_to_index_output_copy = \
            self.input_batch.req_id_to_index.copy()

2144
2145
        # NOTE: GPU -> CPU Sync happens here.
        # Move as many CPU operations as possible before this sync point.
2146
2147
2148
2149
2150
2151
        logprobs_tensors = sampler_output.logprobs_tensors
        logprobs_lists = logprobs_tensors.tolists() \
            if logprobs_tensors is not None else None

        # Compute prompt logprobs if needed.
        prompt_logprobs_dict = self._get_prompt_logprobs_dict(
2152
            hidden_states[:num_scheduled_tokens],
2153
            scheduler_output.num_scheduled_tokens,
2154
2155
        )

2156
        num_sampled_tokens = sampler_output.sampled_token_ids.shape[0]
2157
        sampled_token_ids = sampler_output.sampled_token_ids
2158
        invalid_req_indices = []
2159
2160
2161
2162
2163
2164
2165
2166
2167
2168
2169
2170
2171
2172
        if not self.use_async_scheduling:
            # Get the valid generated tokens.
            max_gen_len = sampled_token_ids.shape[-1]
            if max_gen_len == 1:
                # No spec decode tokens.
                valid_sampled_token_ids = self._to_list(sampled_token_ids)
            else:
                # Includes spec decode tokens.
                valid_sampled_token_ids = self.rejection_sampler.parse_output(
                    sampled_token_ids,
                    self.input_batch.vocab_size,
                )
            # Mask out the sampled tokens that should not be sampled.
            for i in discard_sampled_tokens_req_indices:
2173
                valid_sampled_token_ids[int(i)].clear()
2174
        else:
2175
            valid_sampled_token_ids = []
2176
            invalid_req_indices = discard_sampled_tokens_req_indices.tolist()
2177
2178
2179
2180
2181
2182
2183
2184
2185
2186
2187
2188
2189
2190
2191
            invalid_req_indices_set = set(invalid_req_indices)
            assert sampled_token_ids.shape[-1] == 1

            # Cache the sampled tokens on the GPU and avoid CPU sync.
            # These will be copied into input_ids in the next step
            # when preparing inputs.
            self.input_batch.prev_sampled_token_ids = \
                sampled_token_ids
            self.input_batch.prev_sampled_token_ids_invalid_indices = \
                invalid_req_indices_set
            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
            }
2192

2193
2194
2195
2196
2197
        # 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.
2198
        req_ids = self.input_batch.req_ids
2199
2200
2201
2202
2203
2204
        for req_idx in range(num_sampled_tokens):
            if self.use_async_scheduling:
                sampled_ids = [-1] if \
                    req_idx not in invalid_req_indices_set else None
            else:
                sampled_ids = valid_sampled_token_ids[req_idx]
2205
2206
2207
2208
2209
2210
2211
2212
2213
2214
2215
2216
            if not sampled_ids:
                continue

            start_idx = self.input_batch.num_tokens_no_spec[req_idx]
            end_idx = start_idx + len(sampled_ids)
            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}")

            self.input_batch.token_ids_cpu[req_idx,
                                           start_idx:end_idx] = sampled_ids
2217
            self.input_batch.is_token_ids[req_idx, start_idx:end_idx] = True
2218
2219
            self.input_batch.num_tokens_no_spec[req_idx] = end_idx
            self.input_batch.num_tokens[req_idx] = end_idx
2220

2221
            req_id = req_ids[req_idx]
2222
2223
2224
            req_state = self.requests[req_id]
            req_state.output_token_ids.extend(sampled_ids)

2225
2226
2227
2228
2229
2230
2231
2232
2233
2234
        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,
        )

2235
2236
2237
2238
2239
2240
2241
2242
2243
2244
2245
2246
2247
2248
2249
    @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()

2250
2251
2252
2253
2254
2255
2256
    @torch.inference_mode()
    def execute_model(
        self,
        scheduler_output: "SchedulerOutput",
        intermediate_tensors: Optional[IntermediateTensors] = None,
    ) -> Union[ModelRunnerOutput, AsyncModelRunnerOutput, IntermediateTensors]:
        with record_function_or_nullcontext("Preprocess"):
2257
2258
2259
2260
2261
2262
2263
2264
2265
2266
2267
2268
2269
2270
2271
2272
            with self.synchronize_input_prep():
                # Update persistent batch states.
                self._update_states(scheduler_output)

                if not scheduler_output.total_num_scheduled_tokens:
                    if not has_kv_transfer_group():
                        # Return empty ModelRunnerOutput if no work to do.
                        return EMPTY_MODEL_RUNNER_OUTPUT
                    return self.kv_connector_no_forward(
                        scheduler_output, self.vllm_config)
                if self.cache_config.kv_sharing_fast_prefill:
                    assert not self.input_batch.num_prompt_logprobs, (
                        "--kv-sharing-fast-prefill produces incorrect "
                        "logprobs for prompt tokens, tokens, please disable "
                        "it when the requests need prompt logprobs")

2273
2274
2275
                # Prepare the decoder inputs.
                (attn_metadata, logits_indices, spec_decode_metadata,
                 num_scheduled_tokens_np, spec_decode_common_attn_metadata,
2276
2277
                 max_query_len, ubatch_slices, num_tokens_after_padding,
                 use_cascade_attn) = self._prepare_inputs(scheduler_output)
2278

2279
2280
2281
2282
2283
2284
2285
2286
2287
            (
                num_scheduled_tokens,
                num_input_tokens,
                num_tokens_across_dp,
                input_ids,
                inputs_embeds,
                positions,
                intermediate_tensors,
                model_kwargs,
2288
2289
2290
            ) = self._preprocess(scheduler_output, intermediate_tensors,
                                 ubatch_slices, num_tokens_after_padding)

2291
2292
2293
2294
2295
2296
2297
            uniform_decode = (max_query_len
                              == self.uniform_decode_query_len) and (
                                  num_scheduled_tokens
                                  == self.input_batch.num_reqs * max_query_len)
            batch_descriptor = BatchDescriptor(num_tokens=num_input_tokens,
                                               uniform_decode=uniform_decode)
            cudagraph_runtime_mode, batch_descriptor = \
2298
2299
                self.cudagraph_dispatcher.dispatch(batch_descriptor,
                                                   use_cascade_attn)
2300

2301
2302
2303
2304
2305
        # This is currently to get around the assert in the DPMetadata
        # where it wants `num_tokens_across_dp` to align with `num_tokens`
        if ubatch_slices is not None:
            num_input_tokens = ubatch_slices[0].num_tokens

2306
2307
2308
2309
2310
2311
2312
2313
2314
        # Run the model.
        # Use persistent buffers for CUDA graphs.
        with (set_forward_context(
                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,
2315
                ubatch_slices=ubatch_slices,
2316
2317
2318
2319
2320
2321
2322
2323
2324
2325
2326
2327
2328
        ), record_function_or_nullcontext("Forward"),
              self.maybe_get_kv_connector_output(scheduler_output) as
              kv_connector_output):
            model_output = self.model(
                input_ids=input_ids,
                positions=positions,
                intermediate_tensors=intermediate_tensors,
                inputs_embeds=inputs_embeds,
                **model_kwargs,
            )

        with record_function_or_nullcontext("Postprocess"):
            if self.use_aux_hidden_state_outputs:
2329
                # True when EAGLE 3 is used.
2330
2331
                hidden_states, aux_hidden_states = model_output
            else:
2332
                # Common case.
2333
2334
2335
                hidden_states = model_output
                aux_hidden_states = None

2336
2337
2338
2339
2340
            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)
2341
2342
                    hidden_states.kv_connector_output = kv_connector_output
                    return hidden_states
2343

2344
                if self.is_pooling_model:
2345
2346
2347
2348
2349
                    # Return the pooling output.
                    output = self._pool(hidden_states, num_scheduled_tokens,
                                        num_scheduled_tokens_np)
                    output.kv_connector_output = kv_connector_output
                    return output
2350
2351

                sample_hidden_states = hidden_states[logits_indices]
2352
                logits = self.model.compute_logits(sample_hidden_states)
2353
2354
2355
2356
2357
            else:
                # Rare case.
                assert not self.is_pooling_model

                if not get_pp_group().is_last_rank:
2358
2359
2360
2361
2362
                    all_gather_tensors = {
                        "residual":
                        not is_residual_scattered_for_sp(
                            self.vllm_config, num_input_tokens)
                    }
2363
                    get_pp_group().send_tensor_dict(
2364
2365
2366
                        hidden_states.tensors,
                        all_gather_group=get_tp_group(),
                        all_gather_tensors=all_gather_tensors)
2367
2368
2369
                    logits = None
                else:
                    sample_hidden_states = hidden_states[logits_indices]
2370
                    logits = self.model.compute_logits(sample_hidden_states)
2371
2372
2373
2374
2375

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

2376
2377
2378
2379
2380
2381
2382
2383
                model_output_broadcast_data = get_pp_group(
                ).broadcast_tensor_dict(model_output_broadcast_data,
                                        src=len(get_pp_group().ranks) - 1)
                assert model_output_broadcast_data is not None
                logits = model_output_broadcast_data["logits"]

            # Apply structured output bitmasks if present
            if scheduler_output.grammar_bitmask is not None:
2384
2385
                apply_grammar_bitmask(scheduler_output, self.input_batch,
                                      logits, self.device)
2386
2387
2388
2389

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

2390
2391
2392
2393
2394
2395
2396
2397
2398
2399
2400
2401
2402
2403
2404
2405
2406
        def propose_draft_token_ids(sampled_token_ids):
            assert spec_decode_common_attn_metadata is not None
            with record_function_or_nullcontext("Draft"):
                self._draft_token_ids = self.propose_draft_token_ids(
                    scheduler_output,
                    sampled_token_ids,
                    self.input_batch.sampling_metadata,
                    hidden_states,
                    sample_hidden_states,
                    aux_hidden_states,
                    spec_decode_metadata,
                    spec_decode_common_attn_metadata,
                )

        use_padded_batch_for_eagle = self.speculative_config and \
            self.speculative_config.use_eagle() and \
            not self.speculative_config.disable_padded_drafter_batch
2407
2408
2409
2410
2411
2412
2413
2414
2415
2416
2417
2418
2419
2420
        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
        if (self.speculative_config
                and self.speculative_config.draft_model_config is not None
                and self.speculative_config.draft_model_config.max_model_len
                is not None):
            effective_drafter_max_model_len = (
                self.speculative_config.draft_model_config.max_model_len)
        input_fits_in_drafter = spec_decode_common_attn_metadata and (
            spec_decode_common_attn_metadata.seq_lens.max() +
            self.speculative_config.num_speculative_tokens
            <= effective_drafter_max_model_len)
        if use_padded_batch_for_eagle and input_fits_in_drafter:
2421
2422
2423
2424
            # EAGLE speculative decoding can use the GPU sampled tokens
            # as inputs, and does not need to wait for bookkeeping to finish.
            propose_draft_token_ids(sampler_output.sampled_token_ids)

2425
2426
2427
2428
2429
2430
2431
2432
2433
2434
2435
2436
2437
        with record_function_or_nullcontext("Bookkeep"):
            (
                num_nans_in_logits,
                logprobs_lists,
                valid_sampled_token_ids,
                prompt_logprobs_dict,
                req_ids_output_copy,
                req_id_to_index_output_copy,
                invalid_req_indices,
            ) = self._bookkeeping_sync(scheduler_output, sampler_output,
                                       logits, hidden_states,
                                       num_scheduled_tokens)

2438
2439
        if (self.speculative_config and not use_padded_batch_for_eagle
                and input_fits_in_drafter):
2440
2441
2442
            # 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)
2443

2444
2445
        with record_function_or_nullcontext("EPLB"):
            self.eplb_step()
2446

2447
2448
2449
        output = ModelRunnerOutput(
            req_ids=req_ids_output_copy,
            req_id_to_index=req_id_to_index_output_copy,
2450
2451
2452
2453
            sampled_token_ids=valid_sampled_token_ids,
            logprobs=logprobs_lists,
            prompt_logprobs_dict=prompt_logprobs_dict,
            pooler_output=[],
2454
            kv_connector_output=kv_connector_output,
2455
2456
2457
            num_nans_in_logits=num_nans_in_logits,
        )

2458
2459
2460
2461
2462
        if not self.use_async_scheduling:
            return output

        return AsyncGPUModelRunnerOutput(
            model_runner_output=output,
2463
            sampled_token_ids=sampler_output.sampled_token_ids,
2464
2465
2466
2467
            invalid_req_indices=invalid_req_indices,
            async_output_copy_stream=self.async_output_copy_stream,
        )

2468
2469
2470
2471
2472
2473
2474
2475
2476
2477
2478
    def take_draft_token_ids(self) -> Optional[DraftTokenIds]:
        if self._draft_token_ids is None:
            return None
        req_ids = self.input_batch.req_ids
        if isinstance(self._draft_token_ids, torch.Tensor):
            draft_token_ids = self._draft_token_ids.tolist()
        else:
            draft_token_ids = self._draft_token_ids
        self._draft_token_ids = None
        return DraftTokenIds(req_ids, draft_token_ids)

2479
2480
2481
    def propose_draft_token_ids(
        self,
        scheduler_output: "SchedulerOutput",
2482
        sampled_token_ids: Union[torch.Tensor, list[list[int]]],
2483
2484
2485
        sampling_metadata: SamplingMetadata,
        hidden_states: torch.Tensor,
        sample_hidden_states: torch.Tensor,
Wentao Ye's avatar
Wentao Ye committed
2486
        aux_hidden_states: Optional[list[torch.Tensor]],
2487
        spec_decode_metadata: Optional[SpecDecodeMetadata],
2488
        common_attn_metadata: CommonAttentionMetadata,
2489
    ) -> Union[list[list[int]], torch.Tensor]:
2490
2491
        num_scheduled_tokens = scheduler_output.total_num_scheduled_tokens
        if self.speculative_config.method == "ngram":
2492
            assert isinstance(sampled_token_ids, list)
2493
            assert isinstance(self.drafter, NgramProposer)
2494
2495
2496
2497
2498
            draft_token_ids = self.drafter.propose(
                sampled_token_ids, self.input_batch.req_ids,
                self.input_batch.num_tokens_no_spec,
                self.input_batch.token_ids_cpu,
                self.input_batch.spec_decode_unsupported_reqs)
2499
        elif self.speculative_config.method == "medusa":
2500
            assert isinstance(sampled_token_ids, list)
2501
            assert isinstance(self.drafter, MedusaProposer)
2502

2503
2504
            if sample_hidden_states.shape[0] == len(sampled_token_ids):
                # The input to the target model does not include draft tokens.
2505
2506
2507
2508
                hidden_states = sample_hidden_states
            else:
                indices = []
                offset = 0
Wentao Ye's avatar
Wentao Ye committed
2509
                assert spec_decode_metadata is not None
2510
2511
                for num_draft, tokens in zip(
                        spec_decode_metadata.num_draft_tokens,
2512
                        sampled_token_ids):
2513
2514
                    indices.append(offset + len(tokens) - 1)
                    offset += num_draft + 1
2515
                indices = torch.tensor(indices, device=self.device)
2516
2517
                hidden_states = sample_hidden_states[indices]

2518
            draft_token_ids = self.drafter.propose(
2519
2520
2521
                target_hidden_states=hidden_states,
                sampling_metadata=sampling_metadata,
            )
2522
        elif self.speculative_config.use_eagle():
2523
            assert isinstance(self.drafter, EagleProposer)
2524
2525
2526
2527
2528
2529
2530
2531
2532
2533
2534
2535
2536
2537
2538
2539
2540
2541
2542
2543
2544
2545
2546
2547
2548
2549
2550
2551

            if self.speculative_config.disable_padded_drafter_batch:
                # When padded-batch is disabled, the sampled_token_ids should be
                # the cpu-side list[list[int]] of valid sampled tokens for each
                # request, with invalid requests having empty lists.
                assert isinstance(sampled_token_ids, list), \
                    "sampled_token_ids should be a python list when" \
                    "padded-batch is disabled."
                next_token_ids = self.drafter.prepare_next_token_ids_cpu(
                    sampled_token_ids, self.requests, self.input_batch,
                    scheduler_output.num_scheduled_tokens)
            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.
                assert isinstance(sampled_token_ids, torch.Tensor), \
                    "sampled_token_ids should be a torch.Tensor when" \
                    "padded-batch is enabled."
                next_token_ids, valid_sampled_tokens_count = \
                    self.drafter.prepare_next_token_ids_padded(
                        common_attn_metadata,
                        sampled_token_ids,
                        self.requests,
                        self.input_batch,
                        self.discard_request_indices.gpu,
                        self.num_discarded_requests
                    )
Jiayi Yao's avatar
Jiayi Yao committed
2552

2553
            if spec_decode_metadata is None:
2554
                token_indices_to_sample = None
2555
                # input_ids can be None for multimodal models.
2556
                target_token_ids = self.input_ids.gpu[:num_scheduled_tokens]
2557
                target_positions = self._get_positions(num_scheduled_tokens)
2558
                if self.use_aux_hidden_state_outputs:
Wentao Ye's avatar
Wentao Ye committed
2559
                    assert aux_hidden_states is not None
2560
2561
2562
                    target_hidden_states = torch.cat(
                        [h[:num_scheduled_tokens] for h in aux_hidden_states],
                        dim=-1)
2563
2564
                else:
                    target_hidden_states = hidden_states[:num_scheduled_tokens]
2565
            else:
2566
2567
2568
2569
2570
2571
2572
2573
2574
2575
2576
2577
2578
2579
                if self.speculative_config.disable_padded_drafter_batch:
                    token_indices_to_sample = None
                    common_attn_metadata, token_indices =\
                        self.drafter.prepare_inputs(
                            common_attn_metadata,
                            sampled_token_ids,
                            spec_decode_metadata.num_draft_tokens)
                else:
                    common_attn_metadata, token_indices, \
                        token_indices_to_sample =\
                        self.drafter.prepare_inputs_padded(
                            common_attn_metadata,
                            spec_decode_metadata,
                            valid_sampled_tokens_count)
2580

2581
                target_token_ids = self.input_ids.gpu[token_indices]
2582
                target_positions = self._get_positions(token_indices)
2583
                if self.use_aux_hidden_state_outputs:
Wentao Ye's avatar
Wentao Ye committed
2584
                    assert aux_hidden_states is not None
2585
2586
                    target_hidden_states = torch.cat(
                        [h[token_indices] for h in aux_hidden_states], dim=-1)
2587
2588
                else:
                    target_hidden_states = hidden_states[token_indices]
2589
            mm_embeds = None
2590
            if self.supports_mm_inputs:
2591
2592
2593
                mm_embeds = self._gather_mm_embeddings(scheduler_output,
                                                       shift_computed_tokens=1)

2594
            draft_token_ids = self.drafter.propose(
2595
2596
2597
2598
                target_token_ids=target_token_ids,
                target_positions=target_positions,
                target_hidden_states=target_hidden_states,
                next_token_ids=next_token_ids,
2599
                last_token_indices=token_indices_to_sample,
2600
                sampling_metadata=sampling_metadata,
2601
                common_attn_metadata=common_attn_metadata,
2602
                mm_embeds=mm_embeds,
2603
            )
2604
        return draft_token_ids
2605

2606
2607
2608
2609
2610
2611
2612
2613
2614
2615
    def update_config(self, overrides: dict[str, Any]) -> None:
        allowed_config_names = {"load_config", "model_config"}
        for config_name, config_overrides in overrides.items():
            assert config_name in allowed_config_names, \
                f"Config `{config_name}` not supported. " \
                f"Allowed configs: {allowed_config_names}"
            config = getattr(self, config_name)
            new_config = update_config(config, config_overrides)
            setattr(self, config_name, new_config)

2616
2617
2618
2619
2620
    def load_model(self, eep_scale_up: bool = False) -> None:
        """
        Args:
            eep_scale_up: the model loading is for elastic EP scale up.
        """
2621
        logger.info("Starting to load model %s...", self.model_config.model)
2622
2623
2624
2625
2626
2627
2628
2629
2630
2631
2632
2633
2634
        if eep_scale_up:
            from vllm.distributed.parallel_state import get_ep_group
            num_local_physical_experts = torch.empty(1,
                                                     dtype=torch.int32,
                                                     device="cpu")
            torch.distributed.broadcast(num_local_physical_experts,
                                        group=get_ep_group().cpu_group,
                                        group_src=0)
            num_local_physical_experts = int(num_local_physical_experts.item())
            new_ep_size = get_ep_group().world_size
            global_expert_load, old_global_expert_indices = (
                EplbState.recv_state())
            num_logical_experts = global_expert_load.shape[1]
2635
            self.parallel_config.eplb_config.num_redundant_experts = (
2636
2637
2638
2639
2640
2641
2642
2643
2644
2645
2646
2647
2648
2649
                num_local_physical_experts * new_ep_size - num_logical_experts)
            assert old_global_expert_indices.shape[
                1] % num_local_physical_experts == 0
            old_ep_size = old_global_expert_indices.shape[
                1] // num_local_physical_experts
            rank_mapping = {
                old_ep_rank: old_ep_rank
                for old_ep_rank in range(old_ep_size)
            }
        else:
            global_expert_load = None
            old_global_expert_indices = None
            rank_mapping = None

2650
        with DeviceMemoryProfiler() as m:
2651
            time_before_load = time.perf_counter()
2652
            model_loader = get_model_loader(self.load_config)
2653
2654
2655
            logger.info("Loading model from scratch...")
            self.model = model_loader.load_model(
                vllm_config=self.vllm_config, model_config=self.model_config)
2656
            if self.lora_config:
2657
                self.model = self.load_lora_model(self.model, self.vllm_config,
2658
                                                  self.device)
2659
2660
2661
            if hasattr(self, "drafter"):
                logger.info("Loading drafter model...")
                self.drafter.load_model(self.model)
2662
            if self.use_aux_hidden_state_outputs:
2663
2664
2665
2666
2667
2668
2669
                if supports_eagle3(self.model):
                    self.model.set_aux_hidden_state_layers(
                        self.model.get_eagle3_aux_hidden_state_layers())
                else:
                    raise RuntimeError(
                        "Model does not support EAGLE3 interface but "
                        "aux_hidden_state_outputs was requested")
2670
            time_after_load = time.perf_counter()
2671
        self.model_memory_usage = m.consumed_memory
2672
2673
        logger.info("Model loading took %.4f GiB and %.6f seconds",
                    self.model_memory_usage / GiB_bytes,
2674
                    time_after_load - time_before_load)
2675
        prepare_communication_buffer_for_model(self.model)
2676

2677
2678
2679
2680
        self.is_multimodal_pruning_enabled = (supports_multimodal_pruning(
            self.model) and self.model_config.multimodal_config.
                                              is_multimodal_pruning_enabled())

2681
2682
2683
2684
2685
2686
2687
2688
        if is_mixture_of_experts(
                self.model) and self.parallel_config.enable_eplb:
            logger.info("EPLB is enabled for model %s.",
                        self.model_config.model)
            self.eplb_state = EplbState.build(
                self.model,
                self.device,
                self.parallel_config,
2689
2690
2691
                global_expert_load,
                old_global_expert_indices,
                rank_mapping,
2692
2693
            )

2694
2695
2696
2697
2698
2699
2700
        if (
            self.vllm_config.compilation_config.level == \
                CompilationLevel.DYNAMO_AS_IS and supports_dynamo()
        ):
            backend = self.vllm_config.compilation_config.init_backend(
                self.vllm_config)
            compilation_counter.dynamo_as_is_count += 1
2701
            self.model.compile(fullgraph=True, backend=backend)
2702
2703
2704
2705
2706
            return
        # for other compilation levels, cudagraph behavior is controlled by
        # CudagraphWraper and CudagraphDispatcher of vllm.

        # wrap the model with full cudagraph wrapper if needed.
2707
2708
        if self.compilation_config.cudagraph_mode.has_full_cudagraphs() \
            and not self.parallel_config.enable_dbo:
2709
2710
2711
            self.model = CUDAGraphWrapper(self.model,
                                          self.vllm_config,
                                          runtime_mode=CUDAGraphMode.FULL)
2712
2713
2714
2715
2716
2717
2718
        elif self.parallel_config.enable_dbo:
            if self.compilation_config.cudagraph_mode.has_full_cudagraphs():
                self.model = UBatchWrapper(self.model, self.vllm_config,
                                           CUDAGraphMode.FULL, self.device)
            else:
                self.model = UBatchWrapper(self.model, self.vllm_config,
                                           CUDAGraphMode.NONE, self.device)
2719

2720
2721
2722
2723
2724
    def reload_weights(self) -> None:
        assert getattr(self, "model", None) is not None, \
            "Cannot reload weights before model is loaded."
        model_loader = get_model_loader(self.load_config)
        logger.info("Reloading weights inplace...")
2725
2726
        model_loader.load_weights(self.get_model(),
                                  model_config=self.model_config)
2727

2728
2729
2730
2731
2732
    def save_tensorized_model(
        self,
        tensorizer_config: "TensorizerConfig",
    ) -> None:
        TensorizerLoader.save_model(
2733
            self.get_model(),
2734
            tensorizer_config=tensorizer_config,
2735
            model_config=self.model_config,
2736
2737
        )

2738
2739
2740
    def _get_prompt_logprobs_dict(
        self,
        hidden_states: torch.Tensor,
2741
        num_scheduled_tokens: dict[str, int],
2742
    ) -> dict[str, Optional[LogprobsTensors]]:
2743
2744
2745
2746
        num_prompt_logprobs_dict = self.input_batch.num_prompt_logprobs
        if not num_prompt_logprobs_dict:
            return {}

2747
        in_progress_dict = self.input_batch.in_progress_prompt_logprobs_cpu
2748
        prompt_logprobs_dict: dict[str, Optional[LogprobsTensors]] = {}
2749
2750
2751
2752
2753

        # 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():
2754
            num_tokens = num_scheduled_tokens[req_id]
2755
2756
2757

            # Get metadata for this request.
            request = self.requests[req_id]
2758
2759
2760
2761
            if request.prompt_token_ids is None:
                # Prompt logprobs is incompatible with prompt embeddings
                continue

2762
2763
2764
2765
            num_prompt_tokens = len(request.prompt_token_ids)
            prompt_token_ids = torch.tensor(request.prompt_token_ids).to(
                self.device, non_blocking=True)

2766
2767
2768
2769
2770
2771
2772
2773
2774
            # 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(
                    num_prompt_tokens - 1, num_prompt_logprobs + 1)
                in_progress_dict[req_id] = logprobs_tensors

2775
            # Determine number of logits to retrieve.
2776
2777
            start_idx = request.num_computed_tokens
            start_tok = start_idx + 1
2778
            num_remaining_tokens = num_prompt_tokens - start_tok
2779
            if num_tokens <= num_remaining_tokens:
2780
                # This is a chunk, more tokens remain.
2781
2782
2783
                # 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.
2784
2785
2786
2787
2788
                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)
2789
2790
2791
2792
2793
2794
2795
                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
2796
2797
2798
2799
2800

            # 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]
2801
            offset = self.query_start_loc.np[req_idx].item()
2802
            prompt_hidden_states = hidden_states[offset:offset + num_logits]
2803
            logits = self.model.compute_logits(prompt_hidden_states)
2804
2805
2806
2807
2808
2809
2810

            # 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.
            tgt_token_ids = prompt_token_ids[start_tok:start_tok + num_logits]

            # Compute prompt logprobs.
2811
2812
            logprobs = self.sampler.compute_logprobs(logits)
            token_ids, logprobs, ranks = self.sampler.gather_logprobs(
2813
2814
2815
                logprobs, num_prompt_logprobs, tgt_token_ids)

            # Transfer GPU->CPU async.
2816
2817
2818
2819
2820
2821
2822
            chunk_slice = slice(start_idx, start_idx + num_logits)
            logprobs_tensors.logprob_token_ids[chunk_slice].copy_(
                token_ids, non_blocking=True)
            logprobs_tensors.logprobs[chunk_slice].copy_(logprobs,
                                                         non_blocking=True)
            logprobs_tensors.selected_token_ranks[chunk_slice].copy_(
                ranks, non_blocking=True)
2823
2824
2825
2826
2827

        # 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]
2828
            del in_progress_dict[req_id]
2829
2830

        # Must synchronize the non-blocking GPU->CPU transfers.
2831
        if prompt_logprobs_dict:
2832
            self._sync_device()
2833
2834
2835

        return prompt_logprobs_dict

2836
2837
2838
2839
2840
2841
2842
2843
2844
2845
2846
2847
2848
2849
2850
2851
2852
2853
2854
2855
    def _get_nans_in_logits(
        self,
        logits: Optional[torch.Tensor],
    ) -> dict[str, int]:
        try:
            if logits is None:
                return {req_id: 0 for req_id in self.input_batch.req_ids}

            num_nans_in_logits = {}
            num_nans_for_index = logits.isnan().sum(dim=-1).cpu().numpy()
            for req_id in self.input_batch.req_ids:
                req_index = self.input_batch.req_id_to_index[req_id]
                num_nans_in_logits[req_id] = (
                    int(num_nans_for_index[req_index])
                    if num_nans_for_index is not None
                    and req_index < logits.shape[0] else 0)
            return num_nans_in_logits
        except IndexError:
            return {}

2856
2857
2858
2859
2860
2861
    @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
2862
         - during DP rank dummy run
2863
2864
2865
2866
2867
2868
2869
2870
2871
2872
2873
        """
        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(
2874
                    self.input_ids.gpu,
2875
2876
2877
2878
                    low=0,
                    high=self.model_config.get_vocab_size(),
                    dtype=input_ids.dtype)

2879
            logger.debug_once("Randomizing dummy data for DP Rank")
2880
2881
2882
2883
2884
            input_ids.copy_(rand_input_ids()[:input_ids.size(0)],
                            non_blocking=True)
            yield
            input_ids.fill_(0)

2885
2886
2887
2888
2889
2890
    def _get_mm_dummy_batch(
        self,
        modality: str,
        max_items_per_batch: int,
    ) -> BatchedTensorInputs:
        """Dummy data for profiling and precompiling multimodal models."""
2891
2892
        assert self.mm_budget is not None

2893
2894
        dummy_decoder_data = self.mm_registry.get_decoder_dummy_data(
            model_config=self.model_config,
2895
            seq_len=self.max_model_len,
2896
            mm_counts={modality: 1},
2897
            cache=self.mm_budget.cache,
2898
2899
2900
2901
        )
        dummy_mm_data = dummy_decoder_data.multi_modal_data

        # Result in the maximum GPU consumption of the model
2902
2903
        dummy_mm_item = dummy_mm_data[modality][0]
        dummy_mm_items = [dummy_mm_item] * max_items_per_batch
2904

2905
        model = cast(SupportsMultiModal, self.model)
2906
2907
        return next(mm_kwargs_group
                    for _, _, mm_kwargs_group in group_mm_kwargs_by_modality(
2908
                        dummy_mm_items,
2909
2910
                        device=self.device,
                        pin_memory=self.pin_memory,
2911
                        merge_by_field_config=model.merge_by_field_config,
2912
                    ))
2913

2914
2915
2916
2917
    @torch.inference_mode()
    def _dummy_run(
        self,
        num_tokens: int,
2918
        cudagraph_runtime_mode: Optional[CUDAGraphMode] = None,
2919
2920
        force_attention: bool = False,
        uniform_decode: bool = False,
2921
        allow_microbatching: bool = True,
2922
2923
        skip_eplb: bool = False,
        is_profile: bool = False,
2924
        create_mixed_batch: bool = False,
2925
        remove_lora: bool = True,
2926
    ) -> tuple[torch.Tensor, torch.Tensor]:
2927
2928
2929
2930
2931
2932
2933
        """
        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.
2934
                - if not set will determine the cudagraph mode based on using
2935
                    the self.cudagraph_dispatcher.
2936
2937
2938
2939
                - CUDAGraphMode.NONE: No cudagraph, for warm up and profile run
                - CUDAGraphMode.PIECEWISE: Piecewise cudagraph.
                - CUDAGraphMode.FULL: Full cudagraph, attention metadata is
                    needed.
2940
            force_attention: If True, always create attention metadata. Used to
2941
2942
2943
2944
                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.
2945
2946
            create_mixed_batch: If True, create a mixed batch with both decode
                (1 token) and prefill (multiple tokens) requests.
2947
            remove_lora: If False, dummy LoRAs are not destroyed after the run
2948
        """
2949
2950
        assert cudagraph_runtime_mode is None or \
            cudagraph_runtime_mode.valid_runtime_modes()
2951

2952
        # If cudagraph_mode.decode_mode() == FULL and
2953
        # cudagraph_mode.separate_routine(). This means that we are using
2954
2955
2956
2957
2958
2959
2960
2961
2962
2963
2964
2965
2966
2967
        # 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.
        max_query_len = self.uniform_decode_query_len if uniform_decode else \
                                                                num_tokens

2968
2969
2970
2971
2972
        # 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
2973
2974
2975
2976
2977
2978
2979
2980
2981
2982
2983
2984
2985
2986
2987
        if create_mixed_batch:
            assert not uniform_decode
            # Create mixed batch:
            # first half decode tokens, second half one prefill
            num_decode_tokens = num_tokens // 2
            num_prefill_tokens = num_tokens - num_decode_tokens
            num_reqs = num_decode_tokens + 1

            # Create decode requests (1 token each) followed by prefill request
            num_scheduled_tokens_list = [1] * num_decode_tokens + [
                num_prefill_tokens
            ]
            # Note: Overriding max_query_len to be the prefill tokens
            max_query_len = num_prefill_tokens
        elif uniform_decode:
2988
2989
            assert not create_mixed_batch
            num_reqs = cdiv(num_tokens, max_query_len)
2990
2991
            num_scheduled_tokens_list = [max_query_len] * num_reqs
            if num_tokens % max_query_len != 0:
2992
                num_scheduled_tokens_list[-1] = num_tokens % max_query_len
2993
2994
2995
2996
2997
2998
        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

2999
3000
3001
3002
        assert sum(num_scheduled_tokens_list) == num_tokens
        assert len(num_scheduled_tokens_list) == num_reqs
        num_scheduled_tokens = np.array(num_scheduled_tokens_list,
                                        dtype=np.int32)
3003
        total_num_scheduled_tokens = int(num_scheduled_tokens.sum())
3004

3005
        ubatch_slices = None
3006
3007
        num_tokens_after_padding = None

3008
3009
        # We currently only microbatch if the number of tokens is
        # over a certain threshold.
3010
3011
3012
3013
3014
        if self.parallel_config.enable_dbo and allow_microbatching:
            ubatch_slices, num_tokens_after_padding = ubatch_split(
                num_scheduled_tokens,
                total_num_scheduled_tokens,
                total_num_scheduled_tokens,
3015
3016
                uniform_decode=uniform_decode,
                vllm_config=self.vllm_config,
3017
3018
3019
3020
3021
3022
3023
3024
3025
3026
3027
3028
            )

        # If we failed to microbatch, currently need to resynchronize
        # TODO(lucas,sage): we should be able to avoid this second sync by
        #  refactoring `get_dp_padding_ubatch` and `get_dp_padding` into
        #  a single `coordinate_batch_across_dp` function.
        if num_tokens_after_padding is None:
            num_pad, num_tokens_across_dp = self.get_dp_padding(num_tokens)
            num_tokens_after_padding = num_tokens + num_pad
        else:
            num_tokens_across_dp = num_tokens_after_padding
            num_tokens_after_padding = int(num_tokens_after_padding[0].item())
3029
3030

        attn_metadata: Optional[PerLayerAttnMetadata] = None
3031
3032
3033

        # If force_attention is True, we always capture attention. Otherwise,
        # it only happens for cudagraph_runtime_mode=FULL.
3034
        if force_attention or cudagraph_runtime_mode == CUDAGraphMode.FULL:
3035
            attn_metadata = {}
3036
3037
            if ubatch_slices is not None:
                attn_metadata = [dict() for _ in range(len(ubatch_slices))]
3038

3039
3040
3041
3042
3043
3044
            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:
3045
                seq_lens = max_query_len
3046
            self.seq_lens.np[:num_reqs] = seq_lens
3047
3048
            self.seq_lens.np[num_reqs:] = 0
            self.seq_lens.copy_to_gpu()
3049

3050
3051
3052
3053
3054
            cum_num_tokens, _ = self._get_cumsum_and_arange(
                num_scheduled_tokens)
            self.query_start_loc.np[1:num_reqs + 1] = cum_num_tokens
            self.query_start_loc.copy_to_gpu()

3055
3056
            for kv_cache_group_id, kv_cache_group_spec in enumerate(
                    self.kv_cache_config.kv_cache_groups):
3057
                common_attn_metadata = CommonAttentionMetadata(
3058
3059
                    query_start_loc=self.query_start_loc.gpu[:num_reqs + 1],
                    query_start_loc_cpu=self.query_start_loc.cpu[:num_reqs +
3060
                                                                 1],
3061
3062
                    seq_lens=self.seq_lens.gpu[:num_reqs],
                    seq_lens_cpu=self.seq_lens.cpu[:num_reqs],
3063
3064
3065
3066
                    num_computed_tokens_cpu=self.input_batch.
                    num_computed_tokens_cpu_tensor[:num_reqs],
                    num_reqs=num_reqs,
                    num_actual_tokens=num_tokens,
3067
                    max_query_len=max_query_len,
3068
                    max_seq_len=self.max_model_len,
3069
3070
3071
3072
                    block_table_tensor=self.input_batch.
                    block_table[kv_cache_group_id].get_device_tensor(num_reqs),
                    slot_mapping=self.input_batch.block_table[
                        kv_cache_group_id].slot_mapping.gpu[:num_tokens],
3073
                    causal=True)
3074
                for attn_group in self.attn_groups[kv_cache_group_id]:
3075
3076
3077
3078
3079
3080
3081
3082
3083
3084
3085
3086
3087
3088
3089
3090
3091
3092
3093
                    if ubatch_slices is not None:
                        common_attn_metadata_list = split_attn_metadata(
                            ubatch_slices, common_attn_metadata)
                        for ubid, common_attn_metadata in enumerate(
                                common_attn_metadata_list):
                            assert common_attn_metadata.max_query_len == 1
                            attn_metadata_i = (attn_group\
                                               .get_metadata_builder(ubatch_id=ubid)\
                                               .build_for_cudagraph_capture(common_attn_metadata))
                            for layer_name in kv_cache_group_spec.layer_names:
                                assert type(attn_metadata) is list
                                attn_metadata[ubid][
                                    layer_name] = attn_metadata_i
                    else:
                        assert type(attn_metadata) is dict
                        attn_metadata_i = attn_group.get_metadata_builder()\
                            .build_for_cudagraph_capture(common_attn_metadata)
                        for layer_name in kv_cache_group_spec.layer_names:
                            attn_metadata[layer_name] = attn_metadata_i
3094

3095
        with self.maybe_dummy_run_with_lora(self.lora_config,
3096
                                            num_scheduled_tokens, remove_lora):
3097
3098
3099
            model_kwargs = self._init_model_kwargs(num_tokens)
            if (self.supports_mm_inputs
                    and not self.model_config.is_encoder_decoder):
3100
                input_ids = None
3101
                inputs_embeds = self.inputs_embeds.gpu[:num_tokens]
3102
                model_kwargs = {
3103
                    **model_kwargs,
3104
3105
                    **self._dummy_mm_kwargs(num_reqs),
                }
3106
3107
3108
3109
            elif self.enable_prompt_embeds:
                input_ids = None
                inputs_embeds = self.inputs_embeds.gpu[:num_tokens]
                model_kwargs = self._init_model_kwargs(num_tokens)
3110
            else:
3111
                input_ids = self.input_ids.gpu[:num_tokens]
3112
                inputs_embeds = None
3113

3114
            if self.uses_mrope:
3115
                positions = self.mrope_positions.gpu[:, :num_tokens]
3116
            else:
3117
                positions = self.positions.gpu[:num_tokens]
3118
3119
3120
3121
3122
3123
3124
3125
3126
3127

            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,
                            device=self.device))
3128
3129
3130

                intermediate_tensors = self.sync_and_slice_intermediate_tensors(
                    num_tokens, None, False)
3131
3132
3133
3134

            # filter out the valid batch descriptor
            _cg_mode, batch_descriptor = self.cudagraph_dispatcher.dispatch(
                BatchDescriptor(num_tokens=num_tokens,
3135
3136
                                uniform_decode=uniform_decode)) \
                if not is_profile else (CUDAGraphMode.NONE, None)
3137
3138
3139
3140
3141
            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, (
3142
3143
                    f"Cudagraph runtime mode mismatch at dummy_run. "
                    f"Expected {_cg_mode}, but got {cudagraph_runtime_mode}.")
3144
3145
            else:
                cudagraph_runtime_mode = _cg_mode
3146

3147
3148
            if ubatch_slices is not None:
                num_tokens = num_tokens // 2
3149
            with self.maybe_randomize_inputs(input_ids), set_forward_context(
3150
3151
                    attn_metadata,
                    self.vllm_config,
3152
                    num_tokens=num_tokens_after_padding,
3153
3154
                    num_tokens_across_dp=num_tokens_across_dp,
                    cudagraph_runtime_mode=cudagraph_runtime_mode,
3155
3156
                    batch_descriptor=batch_descriptor,
                    ubatch_slices=ubatch_slices):
3157
                outputs = self.model(
3158
3159
3160
3161
                    input_ids=input_ids,
                    positions=positions,
                    intermediate_tensors=intermediate_tensors,
                    inputs_embeds=inputs_embeds,
3162
                    **model_kwargs,
3163
                )
3164

3165
3166
3167
3168
            if self.use_aux_hidden_state_outputs:
                hidden_states, _ = outputs
            else:
                hidden_states = outputs
3169

3170
            if self.speculative_config and self.speculative_config.use_eagle():
3171
3172
3173
                assert isinstance(self.drafter, EagleProposer)
                self.drafter.dummy_run(num_tokens)

3174
3175
3176
3177
3178
3179
3180
3181
3182
3183
        # 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)

3184
        logit_indices = np.cumsum(num_scheduled_tokens) - 1
3185
        return hidden_states, hidden_states[logit_indices]
3186
3187
3188
3189
3190
3191

    @torch.inference_mode()
    def _dummy_sampler_run(
        self,
        hidden_states: torch.Tensor,
    ) -> torch.Tensor:
3192
3193
3194
3195
        # 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)
3196

3197
        logits = self.model.compute_logits(hidden_states)
3198
3199
3200
3201
3202
3203
3204
3205
3206
3207
3208
3209
3210
3211
3212
3213
3214
3215
3216
3217
3218
        num_reqs = logits.size(0)

        dummy_tensors = lambda v: torch.full(
            (num_reqs, ), v, device=self.device)

        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)],
            allowed_token_ids_mask=None,
            bad_words_token_ids={},
3219
            logitsprocs=LogitsProcessors(),
3220
        )
3221
        try:
3222
3223
            sampler_output = self.sampler(logits=logits,
                                          sampling_metadata=dummy_metadata)
3224
3225
3226
3227
3228
3229
3230
3231
3232
        except RuntimeError as e:
            if 'out of memory' in str(e):
                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 "
                    "initializing the engine.") from e
            else:
                raise e
3233
        if self.speculative_config:
3234
3235
3236
3237
3238
3239
3240
3241
3242
3243
3244
3245
3246
3247
3248
3249
3250
3251
3252
3253
3254
3255
3256
3257
3258
3259
            draft_token_ids = [[0] for _ in range(num_reqs)]
            dummy_spec_decode_metadata = SpecDecodeMetadata.make_dummy(
                draft_token_ids, self.device)

            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
            target_logits = torch.randn(num_tokens,
                                        logits.shape[-1],
                                        device=self.device,
                                        dtype=logits.dtype)
            # NOTE(woosuk): Here, we should use int32 because the sampler uses
            # int32 for bonus_token_ids. If the dtype mismatches, re-compilation
            # will occur at runtime.
            bonus_token_ids = torch.zeros(num_reqs,
                                          device=self.device,
                                          dtype=torch.int32)
            self.rejection_sampler(
                dummy_spec_decode_metadata,
                draft_probs,
                target_logits,
                bonus_token_ids,
                dummy_metadata,
            )
3260
        return sampler_output
3261

3262
    def _dummy_pooler_run_task(
3263
3264
        self,
        hidden_states: torch.Tensor,
3265
3266
        task: PoolingTask,
    ) -> PoolerOutput:
3267
3268
3269
3270
3271
3272
3273
3274
3275
3276
3277
        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

3278
        dummy_prompt_lens = torch.tensor(
3279
3280
            num_scheduled_tokens_list,
            device="cpu",
3281
3282
3283
3284
        )
        dummy_token_ids = torch.zeros((num_reqs, req_num_tokens),
                                      dtype=torch.int32,
                                      device=self.device)
3285

3286
        model = cast(VllmModelForPooling, self.get_model())
3287
        dummy_pooling_params = PoolingParams(task=task)
3288
        dummy_pooling_params.verify(task=task, model_config=self.model_config)
3289
        to_update = model.pooler.get_pooling_updates(task)
3290
3291
        to_update.apply(dummy_pooling_params)

3292
        dummy_metadata = PoolingMetadata(
3293
3294
3295
3296
            prompt_lens=dummy_prompt_lens,
            prompt_token_ids=dummy_token_ids,
            pooling_params=[dummy_pooling_params] * num_reqs,
        )
3297

3298
3299
3300
        dummy_metadata.build_pooling_cursor(num_scheduled_tokens_list,
                                            device=hidden_states.device)

3301
        try:
3302
            return model.pooler(hidden_states=hidden_states,
3303
                                pooling_metadata=dummy_metadata)
3304
3305
3306
        except RuntimeError as e:
            if 'out of memory' in str(e):
                raise RuntimeError(
3307
3308
3309
                    "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 "
3310
3311
3312
                    "initializing the engine.") from e
            else:
                raise e
3313
3314
3315
3316
3317
3318
3319
3320
3321
3322
3323

    @torch.inference_mode()
    def _dummy_pooler_run(
        self,
        hidden_states: torch.Tensor,
    ) -> PoolerOutput:
        # Find the task that has the largest output for subsequent steps
        output_size = dict[PoolingTask, float]()
        for task in self.get_supported_pooling_tasks():
            # Run a full batch with each task to ensure none of them OOMs
            output = self._dummy_pooler_run_task(hidden_states, task)
3324
            output_size[task] = sum(o.nbytes for o in output)
3325
3326
3327
3328
            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)
3329

3330
    def profile_run(self) -> None:
3331
        # Profile with multimodal encoder & encoder cache.
3332
        if self.supports_mm_inputs:
3333
            if self.model_config.multimodal_config.skip_mm_profiling:
3334
                logger.info(
3335
3336
3337
3338
3339
3340
3341
3342
3343
3344
                    "Skipping memory profiling for multimodal encoder and "
                    "encoder cache.")
            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.
3345
3346
3347
                    dummy_modality = mm_budget.get_modality_with_max_tokens()
                    max_mm_items_per_batch = mm_budget \
                        .max_items_per_batch_by_modality[dummy_modality]
3348
3349
3350
3351
3352
3353
3354
3355
3356

                    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,
                    )
3357

3358
3359
3360
3361
3362
                    # Create dummy batch of multimodal inputs.
                    batched_dummy_mm_inputs = self._get_mm_dummy_batch(
                        dummy_modality,
                        max_mm_items_per_batch,
                    )
3363

3364
3365
3366
3367
                    # Run multimodal encoder.
                    dummy_encoder_outputs = \
                        self.model.get_multimodal_embeddings(
                        **batched_dummy_mm_inputs)
3368

3369
3370
3371
3372
                    sanity_check_mm_encoder_outputs(
                        dummy_encoder_outputs,
                        expected_num_items=max_mm_items_per_batch,
                    )
3373

3374
3375
3376
                    # Cache the dummy encoder outputs.
                    self.encoder_cache["tmp"] = dict(
                        enumerate(dummy_encoder_outputs))
3377

3378
        # Add `is_profile` here to pre-allocate communication buffers
3379
        hidden_states, last_hidden_states \
3380
            = self._dummy_run(self.max_num_tokens, is_profile=True)
3381
        if get_pp_group().is_last_rank:
3382
3383
3384
3385
            if self.is_pooling_model:
                output = self._dummy_pooler_run(hidden_states)
            else:
                output = self._dummy_sampler_run(last_hidden_states)
3386
        else:
3387
            output = None
3388
        self._sync_device()
3389
        del hidden_states, output
3390
        self.encoder_cache.clear()
3391
        gc.collect()
3392

3393
    def capture_model(self) -> int:
3394
        if self.compilation_config.cudagraph_mode == CUDAGraphMode.NONE:
3395
            logger.warning(
3396
                "Skipping CUDA graph capture. To turn on CUDA graph capture, "
3397
                "ensure `cudagraph_mode` was not manually set to `NONE`")
3398
            return 0
3399
3400
        else:
            self.initialize_cudagraph_capture()
3401

3402
3403
        compilation_counter.num_gpu_runner_capture_triggers += 1

3404
3405
3406
        start_time = time.perf_counter()
        start_free_gpu_memory = torch.cuda.mem_get_info()[0]

3407
3408
3409
3410
3411
3412
3413
3414
3415
3416
3417
3418
3419
3420
        @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()
3421
                    gc.collect()
3422

3423
3424
3425
        # 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.
3426
        set_cudagraph_capturing_enabled(True)
3427
        with freeze_gc(), graph_capture(device=self.device):
3428
            cudagraph_mode = self.compilation_config.cudagraph_mode
3429
            assert cudagraph_mode is not None
3430
3431
3432
3433
3434
3435
3436
3437
3438
            if cudagraph_mode.mixed_mode() != CUDAGraphMode.NONE:
                cudagraph_runtime_mode = cudagraph_mode.mixed_mode()

                compilation_cases = list(reversed(self.cudagraph_batch_sizes))
                self._capture_cudagraphs(
                    compilation_cases,
                    cudagraph_runtime_mode=cudagraph_runtime_mode,
                    uniform_decode=False)

3439
3440
            # Capture full cudagraph for uniform decode batches if we
            # don't already have full mixed prefill-decode cudagraphs.
3441
3442
3443
3444
3445
3446
3447
3448
3449
3450
3451
3452
3453
3454
3455
3456
3457
3458
            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
                decode_cudagraph_batch_sizes = [
                    x for x in self.cudagraph_batch_sizes if
                    x <= max_num_tokens and x >= self.uniform_decode_query_len
                ]
                compilation_cases_decode = list(
                    reversed(decode_cudagraph_batch_sizes))
                self._capture_cudagraphs(
                    compilation_cases=compilation_cases_decode,
                    cudagraph_runtime_mode=CUDAGraphMode.FULL,
                    uniform_decode=True)

        # 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
3459
        # we may do lazy capturing in future that still allows capturing
3460
3461
        # after here.
        set_cudagraph_capturing_enabled(False)
3462
3463
3464
3465
3466
3467
3468
3469

        end_time = time.perf_counter()
        end_free_gpu_memory = torch.cuda.mem_get_info()[0]
        elapsed_time = end_time - start_time
        cuda_graph_size = start_free_gpu_memory - end_free_gpu_memory
        # This usually takes 5~20 seconds.
        logger.info("Graph capturing finished in %.0f secs, took %.2f GiB",
                    elapsed_time, cuda_graph_size / (1 << 30))
3470
        return cuda_graph_size
3471

3472
3473
3474
3475
    def _capture_cudagraphs(self, compilation_cases: list[int],
                            cudagraph_runtime_mode: CUDAGraphMode,
                            uniform_decode: bool):
        assert cudagraph_runtime_mode != CUDAGraphMode.NONE and \
3476
3477
            cudagraph_runtime_mode.valid_runtime_modes(), \
            f"Invalid cudagraph runtime mode: {cudagraph_runtime_mode}"
3478
3479
3480
3481
3482
3483
3484
3485
3486

        # 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",
                    cudagraph_runtime_mode.name))
3487

3488
3489
3490
        # We skip EPLB here since we don't want to record dummy metrics
        for num_tokens in compilation_cases:
            # We currently only capture ubatched graphs when its a FULL
3491
3492
3493
3494
            # 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
            allow_microbatching = self.parallel_config.enable_dbo \
3495
3496
3497
3498
3499
3500
3501
3502
                and cudagraph_runtime_mode == CUDAGraphMode.FULL \
                and uniform_decode \
                and check_ubatch_thresholds(
                    config=self.vllm_config.parallel_config,
                    num_tokens=num_tokens,
                    uniform_decode=uniform_decode,
                )

3503
3504
3505
3506
3507
3508
3509
3510
            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.
                force_attention = (
                    cudagraph_runtime_mode == CUDAGraphMode.FULL)
3511
                self._dummy_run(num_tokens,
3512
3513
                                cudagraph_runtime_mode=CUDAGraphMode.NONE,
                                force_attention=force_attention,
3514
                                uniform_decode=uniform_decode,
3515
                                allow_microbatching=allow_microbatching,
3516
3517
                                skip_eplb=True,
                                remove_lora=False)
3518
3519
3520
3521
3522
3523
            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)
3524
        self.maybe_remove_all_loras(self.lora_config)
3525

3526
3527
3528
3529
    def initialize_attn_backend(self, kv_cache_config: KVCacheConfig) -> None:
        """
        Initialize the attention backends and attention metadata builders.
        """
3530
3531
3532
        assert len(self.attn_groups) == 0, \
            "Attention backends are already initialized"

3533
3534
3535
3536
3537
3538
3539
3540
3541
3542
        class AttentionGroupKey(NamedTuple):
            attn_backend: type[AttentionBackend]
            kv_cache_spec: KVCacheSpec

        def get_attn_backends_for_group(
            kv_cache_group_spec: KVCacheGroupSpec,
        ) -> dict[AttentionGroupKey, list[str]]:
            layers = get_layers_from_vllm_config(
                self.vllm_config, AttentionLayerBase,
                kv_cache_group_spec.layer_names)
3543
3544
            attn_backends = {}
            attn_backend_layers = defaultdict(list)
3545
            # Dedupe based on full class name; this is a bit safer than
3546
3547
3548
3549
            # 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.
3550
            for layer_name in kv_cache_group_spec.layer_names:
3551
                attn_backend = layers[layer_name].get_attn_backend()
3552
3553
3554
3555
3556
3557
3558

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

3559
3560
3561
3562
3563
3564
3565
3566
                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):
                    layer_kv_cache_spec = layer_kv_cache_spec.kv_cache_specs[
                        layer_name]
                key = (full_cls_name, layer_kv_cache_spec)
                attn_backends[key] = AttentionGroupKey(attn_backend,
                                                       layer_kv_cache_spec)
3567
3568
3569
3570
3571
3572
3573
                attn_backend_layers[key].append(layer_name)
            return {
                attn_backends[k]: v
                for k, v in attn_backend_layers.items()
            }

        def create_attn_groups(
3574
            attn_backends_map: dict[AttentionGroupKey, list[str]],
3575
3576
        ) -> list[AttentionGroup]:
            attn_groups: list[AttentionGroup] = []
3577
3578
            for (attn_backend,
                 kv_cache_spec), layer_names in attn_backends_map.items():
3579
3580
                attn_group = AttentionGroup.create_with_metadata_builders(
                    attn_backend,
3581
                    layer_names,
3582
                    kv_cache_spec,
3583
3584
                    self.vllm_config,
                    self.device,
3585
3586
3587
3588
                    num_metadata_builders=1
                    if not self.parallel_config.enable_dbo else 2,
                )

3589
3590
3591
3592
                attn_groups.append(attn_group)
            return attn_groups

        for kv_cache_group_spec in kv_cache_config.kv_cache_groups:
3593
3594
            attn_backends = get_attn_backends_for_group(kv_cache_group_spec)
            self.attn_groups.append(create_attn_groups(attn_backends))
3595

co63oc's avatar
co63oc committed
3596
        # Calculate reorder batch threshold (if needed)
3597
3598
        self.calculate_reorder_batch_threshold()

3599
    def initialize_cudagraph_capture(self) -> None:
3600
3601
3602
3603
3604
3605
        """
        Resolve the cudagraph_mode when there are multiple attention 
        backends with potential conflicting CUDA graph support.
        Then initialize the cudagraph_dispatcher based on the resolved
        cudagraph_mode.
        """
3606
3607
3608
3609
        min_cg_support = AttentionCGSupport.ALWAYS
        min_cg_builder_name = None

        for attn_group in self._attn_group_iterator():
3610
            builder = attn_group.get_metadata_builder()
3611
3612
3613
3614
3615
3616
3617
3618
3619
3620
3621
3622
3623
3624
3625
3626
3627
3628
3629
3630
3631
3632
3633
3634
3635
3636
3637
3638
            if builder.cudagraph_support.value < min_cg_support.value:
                min_cg_support = builder.cudagraph_support
                min_cg_builder_name = builder.__class__.__name__
        # Flexible resolve the cudagraph mode
        cudagraph_mode = self.compilation_config.cudagraph_mode
        # check cudagraph for mixed batch is supported
        if cudagraph_mode.mixed_mode() == CUDAGraphMode.FULL \
            and min_cg_support != AttentionCGSupport.ALWAYS:
            msg = (f"CUDAGraphMode.{cudagraph_mode.name} is not supported "
                   f"with {min_cg_builder_name} backend (support: "
                   f"{min_cg_support})")
            if min_cg_support == AttentionCGSupport.NEVER:
                # if not supported any full cudagraphs, just raise it.
                msg += "; please try cudagraph_mode=PIECEWISE, and "\
                    "make sure compilation level is piecewise"
                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"
                cudagraph_mode = self.compilation_config.cudagraph_mode = \
                    CUDAGraphMode.FULL_AND_PIECEWISE
            else:
                msg += "; setting cudagraph_mode=FULL_DECODE_ONLY"
                cudagraph_mode = self.compilation_config.cudagraph_mode = \
                    CUDAGraphMode.FULL_DECODE_ONLY
            logger.warning(msg)

3639
3640
3641
3642
3643
3644
3645
3646
3647
3648
3649
3650
3651
3652
3653
3654
3655
3656
3657
3658
        # check that if we are doing decode full-cudagraphs it is supported
        if (cudagraph_mode.decode_mode() == CUDAGraphMode.FULL
                and min_cg_support == AttentionCGSupport.NEVER):
            msg = (f"CUDAGraphMode.{cudagraph_mode.name} is not supported "
                   f"with {min_cg_builder_name} backend (support: "
                   f"{min_cg_support})")
            if (self.compilation_config.level == CompilationLevel.PIECEWISE and
                (self.compilation_config.splitting_ops_contain_attention()
                 or self.compilation_config.use_inductor_graph_partition)):
                msg += "; setting cudagraph_mode=PIECEWISE because "\
                    "attention is compiled piecewise"
                cudagraph_mode = self.compilation_config.cudagraph_mode = \
                    CUDAGraphMode.PIECEWISE
            else:
                msg += "; setting cudagraph_mode=NONE because "\
                    "attention is not compiled piecewise"
                cudagraph_mode = self.compilation_config.cudagraph_mode = \
                    CUDAGraphMode.NONE
            logger.warning(msg)

3659
3660
3661
3662
3663
3664
3665
3666
3667
3668
3669
3670
3671
3672
3673
3674
3675
3676
3677
3678
3679
3680
3681
3682
3683
3684
3685
3686
3687
3688
3689
3690
3691
3692
        # check that if we are doing spec-decode + decode full-cudagraphs it is
        # supported
        if (cudagraph_mode.decode_mode() == CUDAGraphMode.FULL
                and self.uniform_decode_query_len > 1 and min_cg_support.value
                < AttentionCGSupport.UNIFORM_BATCH.value):
            msg = (f"CUDAGraphMode.{cudagraph_mode.name} is not supported"
                   f" with spec-decode for attention backend "
                   f"{min_cg_builder_name} (support: {min_cg_support})")
            if self.compilation_config.splitting_ops_contain_attention():
                msg += "; setting cudagraph_mode=PIECEWISE"
                cudagraph_mode = self.compilation_config.cudagraph_mode = \
                    CUDAGraphMode.PIECEWISE
            else:
                msg += "; setting cudagraph_mode=NONE"
                cudagraph_mode = self.compilation_config.cudagraph_mode = \
                    CUDAGraphMode.NONE
            logger.warning(msg)

        # double check that we can support full cudagraph if they are requested
        # even after automatic downgrades
        if cudagraph_mode.has_full_cudagraphs() \
            and min_cg_support == AttentionCGSupport.NEVER:
            raise ValueError(f"CUDAGraphMode.{cudagraph_mode.name} is not "
                             f"supported with {min_cg_builder_name} backend ("
                             f"support:{min_cg_support}) "
                             "; please try cudagraph_mode=PIECEWISE, "
                             "and make sure compilation level is piecewise")

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

3693
3694
3695
3696
3697
    def calculate_reorder_batch_threshold(self) -> None:
        """
        Check that if any backends reorder batches; that the reordering
        is compatible (e.g., decode threshold is the same)
        """
3698
        for group in self._attn_group_iterator():
3699
            attn_metadata_builder_i = group.get_metadata_builder()
3700

3701
3702
3703
3704
3705
3706
3707
3708
3709
3710
3711
3712
3713
3714
3715
3716
            # check that if any backends reorder batches; that the reordering
            # is compatible (e.g., decode threshold is the same)
            reorder_batch_threshold_i = (
                attn_metadata_builder_i.reorder_batch_threshold)
            if reorder_batch_threshold_i is not None:
                if self.reorder_batch_threshold is not None:
                    if reorder_batch_threshold_i != \
                        self.reorder_batch_threshold:
                        raise ValueError(
                            f"Attention backend reorders decodes with "
                            f"threshold {reorder_batch_threshold_i} but other "
                            f"backend uses threshold "
                            f"{self.reorder_batch_threshold}")
                else:
                    self.reorder_batch_threshold = reorder_batch_threshold_i

3717
3718
3719
3720
3721
3722
3723
3724
3725
3726
3727
3728
3729
3730
3731
3732
3733
3734
3735
3736
3737
    def may_reinitialize_input_batch(self,
                                     kv_cache_config: KVCacheConfig) -> None:
        """
        Re-initialize the input batch if the block sizes are different from
        `[self.cache_config.block_size]`. This usually happens when there
        are multiple KV cache groups.

        Args:
            kv_cache_config: The KV cache configuration.
        """
        block_sizes = [
            kv_cache_group.kv_cache_spec.block_size
            for kv_cache_group in kv_cache_config.kv_cache_groups
        ]
        if block_sizes != [self.cache_config.block_size]:
            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
                "for more details.")
            self.input_batch = InputBatch(
                max_num_reqs=self.max_num_reqs,
3738
                max_model_len=max(self.max_model_len, self.max_encoder_len),
3739
3740
3741
3742
3743
                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,
3744
                is_spec_decode=bool(self.vllm_config.speculative_config),
3745
3746
                logitsprocs=self.input_batch.logitsprocs,
                is_pooling_model=self.is_pooling_model,
3747
3748
3749
                num_speculative_tokens=(
                    self.vllm_config.speculative_config.num_speculative_tokens
                    if self.vllm_config.speculative_config else 0),
3750
3751
            )

3752
3753
    def _allocate_kv_cache_tensors(
            self, kv_cache_config: KVCacheConfig) -> dict[str, torch.Tensor]:
3754
        """
3755
3756
3757
        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.

3758
        Args:
3759
            kv_cache_config: The KV cache config
3760
        Returns:
3761
            dict[str, torch.Tensor]: A map between layer names to their
3762
3763
3764
3765
3766
3767
3768
3769
3770
3771
3772
3773
            corresponding memory buffer for KV cache.
         """
        kv_cache_raw_tensors: dict[str, torch.Tensor] = {}
        for kv_cache_tensor in kv_cache_config.kv_cache_tensors:
            tensor = torch.zeros(kv_cache_tensor.size,
                                 dtype=torch.int8,
                                 device=self.device)
            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:
3774
3775
3776
3777
            for layer_name in group.layer_names:
                if layer_name in self.runner_only_attn_layers:
                    continue
                layer_names.add(layer_name)
3778
3779
3780
3781
        assert layer_names == set(kv_cache_raw_tensors.keys(
        )), "Some layers are not correctly initialized"
        return kv_cache_raw_tensors

3782
3783
3784
    def _attn_group_iterator(self) -> Iterator[AttentionGroup]:
        return itertools.chain.from_iterable(self.attn_groups)

3785
    def _kv_cache_spec_attn_group_iterator(self) -> Iterator[AttentionGroup]:
3786
3787
        if not self.kv_cache_config.kv_cache_groups:
            return
3788
3789
        for attn_groups in self.attn_groups:
            yield from attn_groups
3790

3791
3792
3793
3794
3795
    def _reshape_kv_cache_tensors(
        self,
        kv_cache_config: KVCacheConfig,
        kv_cache_raw_tensors: dict[str, torch.Tensor],
    ) -> dict[str, torch.Tensor]:
3796
        """
3797
        Reshape the KV cache tensors to the desired shape and dtype.
3798

3799
        Args:
3800
3801
            kv_cache_config: The KV cache config
            kv_cache_raw_tensors: The KV cache buffer of each layer, with
3802
                correct size but uninitialized shape.
3803
        Returns:
3804
            Dict[str, torch.Tensor]: A map between layer names to their
3805
3806
            corresponding memory buffer for KV cache.
        """
3807
        kv_caches: dict[str, torch.Tensor] = {}
3808
        has_attn, has_mamba = False, False
3809
3810
        for group in self._kv_cache_spec_attn_group_iterator():
            kv_cache_spec = group.kv_cache_spec
3811
3812
            attn_backend = group.backend
            for layer_name in group.layer_names:
3813
3814
                if layer_name in self.runner_only_attn_layers:
                    continue
3815
3816
3817
3818
                raw_tensor = kv_cache_raw_tensors[layer_name]
                assert raw_tensor.numel() % kv_cache_spec.page_size_bytes == 0
                num_blocks = (raw_tensor.numel() //
                              kv_cache_spec.page_size_bytes)
3819
                if isinstance(kv_cache_spec, AttentionSpec):
3820
                    has_attn = True
3821
                    kv_cache_shape = attn_backend.get_kv_cache_shape(
3822
3823
3824
                        num_blocks, kv_cache_spec.block_size,
                        kv_cache_spec.num_kv_heads, kv_cache_spec.head_size)
                    dtype = kv_cache_spec.dtype
3825
                    try:
3826
3827
                        kv_cache_stride_order = \
                            attn_backend.get_kv_cache_stride_order()
3828
3829
3830
3831
3832
3833
3834
3835
3836
3837
3838
3839
3840
3841
3842
3843
3844
                        assert len(kv_cache_stride_order) == len(
                            kv_cache_shape)
                    except (AttributeError, NotImplementedError):
                        kv_cache_stride_order = tuple(
                            range(len(kv_cache_shape)))
                    # 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.
                    kv_cache_shape = tuple(kv_cache_shape[i]
                                           for i in kv_cache_stride_order)
                    # Maintain original KV shape view.
                    inv_order = [
                        kv_cache_stride_order.index(i)
                        for i in range(len(kv_cache_stride_order))
                    ]
3845
3846
3847
                    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
3848
                elif isinstance(kv_cache_spec, MambaSpec):
3849
                    has_mamba = True
Chen Zhang's avatar
Chen Zhang committed
3850
3851
                    raw_tensor = kv_cache_raw_tensors[layer_name]
                    state_tensors = []
3852
3853
3854
3855
3856
3857
                    storage_offset_bytes = 0
                    for (shape, dtype) in zip(kv_cache_spec.shapes,
                                              kv_cache_spec.dtypes):
                        dtype_size = get_dtype_size(dtype)
                        num_element_per_page = (
                            kv_cache_spec.page_size_bytes // dtype_size)
Chen Zhang's avatar
Chen Zhang committed
3858
                        target_shape = (num_blocks, *shape)
3859
3860
                        stride = torch.empty(target_shape).stride()
                        target_stride = (num_element_per_page, *stride[1:])
3861
                        assert storage_offset_bytes % dtype_size == 0
3862
3863
3864
3865
                        tensor = torch.as_strided(
                            raw_tensor.view(dtype),
                            size=target_shape,
                            stride=target_stride,
3866
                            storage_offset=storage_offset_bytes // dtype_size,
3867
                        )
Chen Zhang's avatar
Chen Zhang committed
3868
                        state_tensors.append(tensor)
3869
                        storage_offset_bytes += stride[0] * dtype_size
3870
3871

                    kv_caches[layer_name] = state_tensors
3872
                else:
3873
                    raise NotImplementedError
3874
3875

        if has_attn and has_mamba:
3876
            self._update_hybrid_attention_mamba_layout(kv_caches)
3877

3878
3879
        return kv_caches

3880
3881
    def _update_hybrid_attention_mamba_layout(
            self, kv_caches: dict[str, torch.Tensor]) -> None:
3882
        """
3883
3884
        Update the layout of attention layers from (2, num_blocks, ...) to
        (num_blocks, 2, ...).
3885
3886

        Args:
3887
            kv_caches: The KV cache buffer of each layer.
3888
3889
        """

3890
3891
        for group in self._kv_cache_spec_attn_group_iterator():
            kv_cache_spec = group.kv_cache_spec
3892
            for layer_name in group.layer_names:
3893
3894
3895
3896
3897
3898
3899
3900
3901
3902
3903
                kv_cache = kv_caches[layer_name]
                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 " \
                        f"a tensor of shape {kv_cache.shape}"
                    hidden_size = kv_cache.shape[2:].numel()
                    kv_cache.as_strided_(size=kv_cache.shape,
                                         stride=(hidden_size, 2 * hidden_size,
                                                 *kv_cache.stride()[2:]))
3904

3905
3906
3907
3908
3909
3910
3911
3912
    def initialize_kv_cache_tensors(
            self, kv_cache_config: KVCacheConfig) -> dict[str, torch.Tensor]:
        """
        Initialize the memory buffer for KV cache.

        Args:
            kv_cache_config: The KV cache config
        Returns:
3913
            Dict[str, torch.Tensor]: A map between layer names to their
3914
3915
3916
3917
3918
3919
3920
            corresponding memory buffer for KV cache.
        """
        # Initialize the memory buffer for KV cache
        kv_cache_raw_tensors = self._allocate_kv_cache_tensors(kv_cache_config)
        # Change the memory buffer to the desired shape
        kv_caches = self._reshape_kv_cache_tensors(kv_cache_config,
                                                   kv_cache_raw_tensors)
3921

3922
3923
3924
3925
3926
3927
3928
        # Set up cross-layer KV cache sharing
        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)
            kv_caches[layer_name] = kv_caches[target_layer_name]

XuruiYang's avatar
XuruiYang committed
3929
3930
        num_attn_module = 2 \
            if self.model_config.hf_config.model_type == "longcat_flash" else 1
3931
3932
        bind_kv_cache(kv_caches,
                      self.compilation_config.static_forward_context,
XuruiYang's avatar
XuruiYang committed
3933
                      self.kv_caches, num_attn_module)
3934
3935
3936
3937
3938
3939
3940
3941
3942
3943
3944
3945
3946
3947
3948
3949
3950
3951
3952
3953
3954
3955
        return kv_caches

    def maybe_add_kv_sharing_layers_to_kv_cache_groups(
            self, kv_cache_config: KVCacheConfig) -> None:
        """
        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.
3956
3957
3958
3959
3960
3961
3962
3963
            attn_layers = get_layers_from_vllm_config(self.vllm_config,
                                                      Attention)
            for layer_name in reversed(attn_layers):
                if layer_name in self.shared_kv_cache_layers:
                    self.kv_sharing_fast_prefill_eligible_layers.add(
                        layer_name)
                else:
                    break
3964

3965
3966
3967
3968
3969
3970
3971
    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
        """
3972
        kv_cache_config = deepcopy(kv_cache_config)
3973
3974
        self.kv_cache_config = kv_cache_config
        self.may_reinitialize_input_batch(kv_cache_config)
3975
        self.may_add_encoder_only_layers_to_kv_cache_config()
3976
        self.maybe_add_kv_sharing_layers_to_kv_cache_groups(kv_cache_config)
3977
3978
3979
        self.initialize_attn_backend(kv_cache_config)
        kv_caches = self.initialize_kv_cache_tensors(kv_cache_config)

3980
3981
3982
3983
3984
3985
        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
3986
3987
        if has_kv_transfer_group():
            get_kv_transfer_group().register_kv_caches(kv_caches)
3988
3989
3990
            if self.device.type == 'xpu':
                get_kv_transfer_group().set_host_xfer_buffer_ops(
                    copy_kv_blocks)
Robert Shaw's avatar
Robert Shaw committed
3991

3992
        if self.dcp_world_size > 1:
3993
3994
3995
3996
3997
3998
3999
4000
4001
4002
            layer_names = self.attn_groups[0][0].layer_names
            layers = get_layers_from_vllm_config(self.vllm_config,
                                                 AttentionLayerBase,
                                                 layer_names)
            for layer in layers.values():
                assert layer.impl.need_to_return_lse_for_decode, (
                    "DCP requires attention impls to return"
                    " the softmax lse for decode, but the impl "
                    f"{layer.impl.__class__.__name__} "
                    "does not return the softmax lse for decode.")
4003

4004
4005
4006
4007
4008
4009
4010
4011
4012
4013
4014
    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
        use_mla = self.vllm_config.model_config.use_mla
        encoder_only_attn_specs: dict[AttentionSpec,
                                      list[str]] = defaultdict(list)
        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:
4015
                attn_spec: AttentionSpec = EncoderOnlyAttentionSpec(
4016
4017
4018
4019
4020
4021
4022
4023
4024
4025
4026
4027
4028
4029
4030
                    block_size=block_size,
                    num_kv_heads=attn_module.num_kv_heads,
                    head_size=attn_module.head_size,
                    dtype=self.kv_cache_dtype,
                    use_mla=use_mla)
                encoder_only_attn_specs[attn_spec].append(layer_name)
                self.runner_only_attn_layers.add(layer_name)
        if len(encoder_only_attn_specs) > 0:
            assert len(
                encoder_only_attn_specs
            ) == 1, "Only support one encoder-only attention spec now"
            spec, layer_names = encoder_only_attn_specs.popitem()
            self.kv_cache_config.kv_cache_groups.append(
                KVCacheGroupSpec(layer_names=layer_names, kv_cache_spec=spec))

4031
    def get_kv_cache_spec(self) -> dict[str, KVCacheSpec]:
4032
        """
4033
        Generates the KVCacheSpec by parsing the kv cache format from each
4034
4035
        Attention module in the static forward context.
        Returns:
4036
            KVCacheSpec: A dictionary mapping layer names to their KV cache
4037
4038
4039
4040
            format. Layers that do not need KV cache are not included.
        """

        block_size = self.vllm_config.cache_config.block_size
4041
        use_mla = self.vllm_config.model_config.use_mla
4042
        kv_cache_spec: dict[str, KVCacheSpec] = {}
Chen Zhang's avatar
Chen Zhang committed
4043
4044
        attn_layers = get_layers_from_vllm_config(self.vllm_config, Attention)
        for layer_name, attn_module in attn_layers.items():
4045
4046
4047
4048
4049
4050
4051
4052
4053
4054
4055
4056
            if (kv_tgt_layer :=
                    attn_module.kv_sharing_target_layer_name) is not None:
                # The layer doesn't need its own KV cache and will use that of
                # the target layer. We skip creating a KVCacheSpec for it, so
                # that KV cache management logic will act as this layer does
                # not exist, and doesn't allocate KV cache for the layer. This
                # enables the memory saving of cross-layer kv sharing, allowing
                # a given amount of memory to accommodate longer context lengths
                # or enable more requests to be processed simultaneously.
                self.shared_kv_cache_layers[layer_name] = kv_tgt_layer
                continue

4057
4058
            # TODO(lucas): move the attention specs into the model layers like
            # the attention backends
4059
            if attn_module.attn_type == AttentionType.DECODER:
4060
4061
4062
4063
4064
4065
4066
4067
                if attn_module.sliding_window is not None:
                    kv_cache_spec[layer_name] = SlidingWindowSpec(
                        block_size=block_size,
                        num_kv_heads=attn_module.num_kv_heads,
                        head_size=attn_module.head_size,
                        dtype=self.kv_cache_dtype,
                        sliding_window=attn_module.sliding_window,
                        use_mla=use_mla)
4068
4069
                elif self.attention_chunk_size is not None \
                        and isinstance(attn_module, ChunkedLocalAttention):
4070
                    kv_cache_spec[layer_name] = ChunkedLocalAttentionSpec(
4071
4072
4073
4074
4075
                        block_size=block_size,
                        num_kv_heads=attn_module.num_kv_heads,
                        head_size=attn_module.head_size,
                        dtype=self.kv_cache_dtype,
                        attention_chunk_size=self.attention_chunk_size,
4076
                        use_mla=use_mla)
4077
4078
4079
4080
4081
4082
4083
                else:
                    kv_cache_spec[layer_name] = FullAttentionSpec(
                        block_size=block_size,
                        num_kv_heads=attn_module.num_kv_heads,
                        head_size=attn_module.head_size,
                        dtype=self.kv_cache_dtype,
                        use_mla=use_mla)
4084
4085
4086
4087
4088
4089
4090
            elif attn_module.attn_type == AttentionType.ENCODER_DECODER:
                kv_cache_spec[layer_name] = CrossAttentionSpec(
                    block_size=block_size,
                    num_kv_heads=attn_module.num_kv_heads,
                    head_size=attn_module.head_size,
                    dtype=self.kv_cache_dtype,
                    use_mla=use_mla)
4091
4092
4093
4094
4095
4096
4097
4098
            elif attn_module.attn_type in (AttentionType.ENCODER,
                                           AttentionType.ENCODER_ONLY):
                # encoder-only attention does not need KV cache.
                continue
            else:
                raise ValueError(
                    f"Unknown attention type: {attn_module.attn_type}")

4099
        mamba_layers = get_layers_from_vllm_config(self.vllm_config, MambaBase)
Chen Zhang's avatar
Chen Zhang committed
4100
        if len(mamba_layers) > 0:
4101
4102
4103
            if (self.vllm_config.speculative_config is not None
                    and self.vllm_config.model_config.hf_config.model_type
                    not in ["qwen3_next"]):
Chen Zhang's avatar
Chen Zhang committed
4104
4105
4106
4107
4108
4109
                raise NotImplementedError(
                    "Mamba with speculative decoding is not supported yet.")
            if self.vllm_config.cache_config.enable_prefix_caching:
                raise NotImplementedError(
                    "Prefix caching is not supported for Mamba yet.")
            max_model_len = self.vllm_config.model_config.max_model_len
4110

4111
4112
            page_size_padded = (
                self.vllm_config.cache_config.mamba_page_size_padded)
4113

Chen Zhang's avatar
Chen Zhang committed
4114
4115
4116
4117
4118
            # Set block_size to max_model_len, so that mamba model will always
            # have only one block in the KV cache.
            for layer_name, mamba_module in mamba_layers.items():
                kv_cache_spec[layer_name] = MambaSpec(
                    shapes=mamba_module.get_state_shape(),
4119
                    dtypes=mamba_module.get_state_dtype(),
4120
                    block_size=max_model_len,
4121
                    page_size_padded=page_size_padded,
4122
4123
4124
4125
4126
                    mamba_type=mamba_module.mamba_type,
                    num_speculative_blocks=(
                        self.speculative_config.num_speculative_tokens
                        if self.speculative_config else 0),
                )
4127

4128
        return kv_cache_spec
4129
4130
4131
4132
4133
4134
4135
4136
4137
4138
4139
4140
4141
4142
4143

    def _to_list(self, sampled_token_ids: torch.Tensor) -> list[list[int]]:
        # This is a short term mitigation for issue mentioned in
        # https://github.com/vllm-project/vllm/issues/22754.
        # `tolist` would trigger a cuda wise stream sync, which
        # would block other copy ops from other cuda streams.
        # A cuda event sync would avoid such a situation. Since
        # this is in the critical path of every single model
        # forward loop, this has caused perf issue for a disagg
        # setup.
        pinned = self.sampled_token_ids_pinned_cpu[:sampled_token_ids.shape[0]]
        pinned.copy_(sampled_token_ids, non_blocking=True)
        self.transfer_event.record()
        self.transfer_event.synchronize()
        return pinned.tolist()