gpu_model_runner.py 156 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, 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

19
import vllm.envs as envs
20
from vllm.attention import Attention, AttentionType
21
from vllm.attention.backends.abstract import AttentionBackend
22
from vllm.attention.layers.chunked_local_attention import ChunkedLocalAttention
23
from vllm.compilation.counter import compilation_counter
24
25
26
from vllm.compilation.cuda_graph import CUDAGraphWrapper
from vllm.compilation.monitor import set_cudagraph_capturing_enabled
from vllm.config import (CompilationLevel, CUDAGraphMode, VllmConfig,
27
                         get_layers_from_vllm_config, update_config)
28
from vllm.distributed.eplb.eplb_state import EplbState
29
30
from vllm.distributed.kv_transfer import (get_kv_transfer_group,
                                          has_kv_transfer_group)
31
from vllm.distributed.kv_transfer.kv_connector.utils import copy_kv_blocks
32
from vllm.distributed.parallel_state import (
33
    get_pp_group, get_tp_group, graph_capture, is_global_first_rank,
34
    prepare_communication_buffer_for_model)
35
36
from vllm.forward_context import (BatchDescriptor, DPMetadata,
                                  set_forward_context)
37
from vllm.logger import init_logger
38
39
from vllm.model_executor.layers.attention_layer_base import AttentionLayerBase
from vllm.model_executor.layers.mamba.abstract import MambaBase
40
from vllm.model_executor.layers.rotary_embedding import MRotaryEmbedding
41
from vllm.model_executor.model_loader import TensorizerLoader, get_model_loader
42
from vllm.model_executor.models.interfaces import (is_mixture_of_experts,
43
                                                   supports_eagle3,
44
45
46
                                                   supports_transcription)
from vllm.model_executor.models.interfaces_base import (
    VllmModelForPooling, is_pooling_model, is_text_generation_model)
47
from vllm.multimodal import MULTIMODAL_REGISTRY
48
from vllm.multimodal.inputs import (BatchedTensorInputs, MultiModalKwargsItem,
49
                                    PlaceholderRange)
50
from vllm.multimodal.utils import group_mm_kwargs_by_modality
51
from vllm.pooling_params import PoolingParams
52
from vllm.sampling_params import SamplingType
53
from vllm.sequence import IntermediateTensors, PoolerOutput
54
from vllm.tasks import GenerationTask, PoolingTask, SupportedTask
55
from vllm.utils import (STR_DTYPE_TO_TORCH_DTYPE, DeviceMemoryProfiler,
56
57
58
                        GiB_bytes, LazyLoader, cdiv, check_use_alibi,
                        get_dtype_size, is_pin_memory_available, round_up,
                        supports_dynamo)
59
from vllm.v1.attention.backends.mla.flashmla import FlashMLABackend
60
from vllm.v1.attention.backends.utils import (
61
    AttentionCGSupport, AttentionMetadataBuilder, CommonAttentionMetadata,
62
    create_fast_prefill_custom_backend,
63
    reorder_batch_to_split_decodes_and_prefills)
64
from vllm.v1.cudagraph_dispatcher import CudagraphDispatcher
65
66
from vllm.v1.kv_cache_interface import (AttentionSpec,
                                        ChunkedLocalAttentionSpec,
67
                                        EncoderOnlyAttentionSpec,
68
                                        FullAttentionSpec, KVCacheConfig,
69
70
                                        KVCacheGroupSpec, KVCacheSpec,
                                        MambaSpec, SlidingWindowSpec)
71
72
from vllm.v1.outputs import (EMPTY_MODEL_RUNNER_OUTPUT, AsyncModelRunnerOutput,
                             DraftTokenIds, LogprobsTensors, ModelRunnerOutput)
73
from vllm.v1.pool.metadata import PoolingMetadata
74
from vllm.v1.sample.logits_processor import LogitsProcessors, build_logitsprocs
75
from vllm.v1.sample.metadata import SamplingMetadata
76
from vllm.v1.sample.rejection_sampler import RejectionSampler
77
from vllm.v1.sample.sampler import Sampler
78
from vllm.v1.spec_decode.eagle import EagleProposer
79
from vllm.v1.spec_decode.medusa import MedusaProposer
80
from vllm.v1.spec_decode.metadata import SpecDecodeMetadata
81
from vllm.v1.spec_decode.ngram_proposer import NgramProposer
82
from vllm.v1.utils import CpuGpuBuffer
83
from vllm.v1.worker.gpu_input_batch import CachedRequestState, InputBatch
84
from vllm.v1.worker.kv_connector_model_runner_mixin import (
85
    KVConnectorModelRunnerMixin, KVConnectorOutput)
86
from vllm.v1.worker.lora_model_runner_mixin import LoRAModelRunnerMixin
87

88
89
90
91
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)
92

93
if TYPE_CHECKING:
94
95
    import xgrammar as xgr

96
    from vllm.model_executor.model_loader.tensorizer import TensorizerConfig
97
    from vllm.v1.core.sched.output import SchedulerOutput
98
99
else:
    xgr = LazyLoader("xgr", globals(), "xgrammar")
100
101
102
103

logger = init_logger(__name__)


104
105
106
107
108
109
110
111
112
113
114
115
116
117
118
119
120
121
122
123
124
125
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
# 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


151
class GPUModelRunner(LoRAModelRunnerMixin, KVConnectorModelRunnerMixin):
152
153
154

    def __init__(
        self,
155
        vllm_config: VllmConfig,
156
        device: torch.device,
157
    ):
158
159
160
        self.vllm_config = vllm_config
        self.model_config = vllm_config.model_config
        self.cache_config = vllm_config.cache_config
161
        self.compilation_config = vllm_config.compilation_config
162
163
164
165
166
167
        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
168

169
170
171
172
        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))

173
174
175
176
        model_config = self.model_config
        cache_config = self.cache_config
        scheduler_config = self.scheduler_config
        parallel_config = self.parallel_config
177
        self.device = device
178
179
180
181
182
183
184
185
        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]

186
        self.is_pooling_model = (model_config.runner_type == 'pooling')
187
188
189
        self.is_multimodal_raw_input_only_model = (
            model_config.is_multimodal_raw_input_only_model)

190
        self.max_model_len = model_config.max_model_len
191
        self.dcp_world_size = self.parallel_config.decode_context_parallel_size
192
        self.max_num_tokens = scheduler_config.max_num_batched_tokens
193
        self.max_num_reqs = scheduler_config.max_num_seqs
194
195

        # Model-related.
196
197
        self.num_query_heads = model_config.get_num_attention_heads(
            parallel_config)
198
        self.hidden_size = model_config.get_hidden_size()
199
        self.attention_chunk_size = model_config.attention_chunk_size
200
201
        # Only relevant for models using ALiBi (e.g, MPT)
        self.use_alibi = check_use_alibi(model_config)
202

203
        self.cascade_attn_enabled = not self.model_config.disable_cascade_attn
204

205
        # Multi-modal data support
206
        self.mm_registry = MULTIMODAL_REGISTRY
207
        self.uses_mrope = model_config.uses_mrope
208
209
        self.supports_mm_inputs = self.mm_registry.supports_multimodal_inputs(
            model_config)
210

211
        # Sampler
212
        self.sampler = Sampler(logprobs_mode=self.model_config.logprobs_mode)
213

214
215
216
217
218
219
220
        self.eplb_state: Optional[EplbState] = None
        """
        State of the expert parallelism load balancer.

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

221
        # Lazy initializations
222
        # self.model: nn.Module  # Set after load_model
223
        # Initialize in initialize_kv_cache
224
        self.kv_caches: list[torch.Tensor] = []
225
226
        # indexes: [kv_cache_group_id][attn_group]
        self.attn_groups: list[list[AttentionGroup]] = []
227
228
        # self.kv_cache_config: KVCacheConfig

229
230
        # mm_hash ->  encoder_output
        self.encoder_cache: dict[str, torch.Tensor] = {}
231

232
        self.use_aux_hidden_state_outputs = False
233
234
235
236
237
238
239
240
241
242
243
244
245
246
247
248
249
250
251
252
        # 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()
253

254
        # Request states.
255
        self.requests: dict[str, CachedRequestState] = {}
256

257
258
259
260
261
262
263
264
265
        # 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.
266
267
268
269
270
271
        self.input_batch = InputBatch(
            max_num_reqs=self.max_num_reqs,
            max_model_len=self.max_model_len,
            max_num_batched_tokens=self.max_num_tokens,
            device=self.device,
            pin_memory=self.pin_memory,
272
            vocab_size=self.model_config.get_vocab_size(),
273
            block_sizes=[self.cache_config.block_size],
274
            is_spec_decode=bool(self.vllm_config.speculative_config),
275
276
277
278
279
            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,
280
        )
281

282
283
284
285
        self.use_async_scheduling = self.scheduler_config.async_scheduling
        self.async_output_copy_stream = torch.cuda.Stream() if \
            self.use_async_scheduling else None

286
        # TODO(woosuk): Provide an option to tune the max cudagraph batch size.
287
288
289
        # The convention is different.
        # self.cudagraph_batch_sizes sorts in ascending order.
        # The batch sizes in the config are in descending order.
290
291
292
293
        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))
294

295
        # Cache the device properties.
296
        self._init_device_properties()
297

298
        # Persistent buffers for CUDA graphs.
299
300
301
302
303
304
305
306
307
308
309
        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)
        self.inputs_embeds = torch.zeros(
            (self.max_num_tokens, self.hidden_size),
            dtype=self.dtype,
            device=self.device)
310
311

        # Only relevant for models using M-RoPE (e.g, Qwen2-VL)
312
        if self.uses_mrope:
Roger Wang's avatar
Roger Wang committed
313
314
315
316
            # 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
317
318
319
320
321
322

            # 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
323
324
            self.mrope_positions = self._make_buffer(
                (3, self.max_num_tokens + 1), dtype=torch.int64)
325

326
327
        # None in the first PP rank. The rest are set after load_model.
        self.intermediate_tensors: Optional[IntermediateTensors] = None
328

329
        # OPTIMIZATION: Cache the tensors rather than creating them every step.
330
        # Keep in int64 to avoid overflow with long context
331
        self.arange_np = np.arange(max(self.max_num_reqs + 1,
332
333
                                       self.max_model_len,
                                       self.max_num_tokens),
334
                                   dtype=np.int64)
335

336
337
338
339
340
        # 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] = {}
341
342
343
344
345
346
        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)
347

348
349
350
351
352
353
        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)

354
355
356
357
        self.mm_budget = (MultiModalBudget(
            self.model_config,
            self.scheduler_config,
            self.mm_registry,
358
        ) if self.supports_mm_inputs else None)
359

360
361
        self.reorder_batch_threshold: Optional[int] = None

362
363
364
365
366
        # 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()

367
368
369
        # Cached outputs.
        self._draft_token_ids: Optional[Union[list[list[int]],
                                              torch.Tensor]] = None
370
371
372
373
374
        self.transfer_event = torch.cuda.Event()
        self.sampled_token_ids_pinned_cpu = torch.empty(
            (self.max_model_len, 1),
            dtype=torch.int64,
            device="cpu",
375
            pin_memory=self.pin_memory)
376

377
378
379
380
381
382
    def _make_buffer(self, *args, dtype: torch.dtype) -> CpuGpuBuffer:
        return CpuGpuBuffer(*args,
                            dtype=dtype,
                            device=self.device,
                            pin_memory=self.pin_memory)

383
384
385
    def _init_model_kwargs(self, num_tokens: int):
        model_kwargs = dict[str, Any]()

386
        if not self.is_pooling_model:
387
388
            return model_kwargs

389
390
        num_reqs = self.input_batch.num_reqs
        pooling_params = self.input_batch.get_pooling_params()
391
392
393
394
395
396
397
398
399
400
401

        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

402
        seq_lens = self.seq_lens.gpu[:num_reqs]
403
404
405
406
407
408
409
410
411
412
413
        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

414
    def _may_reorder_batch(self, scheduler_output: "SchedulerOutput") -> None:
415
416
        """
        Update the order of requests in the batch based on the attention
417
        backend's needs. For example, some attention backends (namely MLA) may
418
419
420
421
422
423
        want to separate requests based on if the attention computation will be
        compute-bound or memory-bound.

        Args:
            scheduler_output: The scheduler output.
        """
424
425
426
427
428
429
430
431
        # 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

432
        if self.reorder_batch_threshold is not None:
433
434
435
            if self.dcp_world_size > 1:
                assert self.reorder_batch_threshold == 1, \
                    "DCP not support reorder_batch_threshold > 1 now."
436
437
438
439
            reorder_batch_to_split_decodes_and_prefills(
                self.input_batch,
                scheduler_output,
                decode_threshold=self.reorder_batch_threshold)
440

441
442
443
444
445
446
447
448
449
450
451
    # 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()

452
    def _update_states(self, scheduler_output: "SchedulerOutput") -> None:
453
454
455
456
457
458
        """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.

459
460
        The SamplingMetadata is updated and copied to the GPU if there is a
        new/resumed/paused/finished request in the batch.
461
462
        """
        # Remove finished requests from the cached states.
463
464
        for req_id in scheduler_output.finished_req_ids:
            self.requests.pop(req_id, None)
465
466
467
468
469
470
471
        # 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:
472
            self.input_batch.remove_request(req_id)
473
474

        # Free the cached encoder outputs.
475
476
        for mm_hash in scheduler_output.free_encoder_mm_hashes:
            self.encoder_cache.pop(mm_hash, None)
477

478
479
480
481
482
483
484
485
486
487
488
489
490
        # 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:
491
            self.input_batch.remove_request(req_id)
492

493
        reqs_to_add: list[CachedRequestState] = []
494
        # Add new requests to the cached states.
495
496
497
        for new_req_data in scheduler_output.scheduled_new_reqs:
            req_id = new_req_data.req_id
            sampling_params = new_req_data.sampling_params
498
            pooling_params = new_req_data.pooling_params
499

500
501
            if sampling_params and \
                sampling_params.sampling_type == SamplingType.RANDOM_SEED:
502
503
504
505
506
                generator = torch.Generator(device=self.device)
                generator.manual_seed(sampling_params.seed)
            else:
                generator = None

507
508
            if self.is_pooling_model:
                assert pooling_params is not None
509
510
                task = pooling_params.task
                assert task is not None, "You did not set `task` in the API"
511

512
                model = cast(VllmModelForPooling, self.get_model())
513
                to_update = model.pooler.get_pooling_updates(task)
514
515
                to_update.apply(pooling_params)

516
            req_state = CachedRequestState(
517
                req_id=req_id,
518
                prompt_token_ids=new_req_data.prompt_token_ids,
519
                mm_kwargs=new_req_data.mm_kwargs,
520
                mm_positions=new_req_data.mm_positions,
521
                mm_hashes=new_req_data.mm_hashes,
522
                sampling_params=sampling_params,
523
                pooling_params=pooling_params,
524
                generator=generator,
525
526
                block_ids=new_req_data.block_ids,
                num_computed_tokens=new_req_data.num_computed_tokens,
527
                output_token_ids=[],
528
                lora_request=new_req_data.lora_request,
529
            )
530
531
            self.requests[req_id] = req_state

532
            # Only relevant for models using M-RoPE (e.g, Qwen2-VL)
533
            if self.uses_mrope:
534
                self._init_mrope_positions(req_state)
535

536
            reqs_to_add.append(req_state)
537

538
        # Update the states of the running/resumed requests.
539
        is_last_rank = get_pp_group().is_last_rank
540
541
        req_data = scheduler_output.scheduled_cached_reqs
        for i, req_id in enumerate(req_data.req_ids):
542
            req_state = self.requests[req_id]
543
544
545
            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]
546

547
            # Update the cached states.
548
            req_state.num_computed_tokens = num_computed_tokens
549
550
551
552
553
554
555
556
557
558
559
560
561
562
563
564
565

            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:])

566
            # Update the block IDs.
567
            if not resumed_from_preemption:
568
569
570
571
572
                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)
573
            else:
574
                assert new_block_ids is not None
575
576
                # The request is resumed from preemption.
                # Replace the existing block IDs with the new ones.
577
                req_state.block_ids = new_block_ids
578
579
580
581
582
583

            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.
584
                reqs_to_add.append(req_state)
585
586
587
588
                continue

            # Update the persistent batch.
            self.input_batch.num_computed_tokens_cpu[req_index] = (
589
                num_computed_tokens)
590
591
592
            if new_block_ids is not None:
                self.input_batch.block_table.append_row(
                    new_block_ids, req_index)
593
594
595
596
597
598
599

            # 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)
600
                self.input_batch.token_ids_cpu[
601
602
603
604
605
                    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
606

607
608
609
610
611
612
613
614
615
616
617
618
            # 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

619
620
        # Add the new or resumed requests to the persistent batch.
        # The smaller empty indices are filled first.
621
622
        for request in reqs_to_add:
            self.input_batch.add_request(request)
623

624
625
626
627
628
629
        # 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()
630

631
632
633
634
635
636
637
638
639
640
641
642
643
644
645
646
647
648
649
650
651
652
653
654
655
656
657
658
659
660
    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
        for mm_item in req_state.mm_kwargs:
            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

        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,
            )

661
    def _extract_mm_kwargs(
662
        self,
663
664
        scheduler_output: "SchedulerOutput",
    ) -> BatchedTensorInputs:
665
        if not scheduler_output or not self.is_multimodal_raw_input_only_model:
666
            return {}
667

668
669
670
        mm_kwargs = list[MultiModalKwargsItem]()
        for req in scheduler_output.scheduled_new_reqs:
            mm_kwargs.extend(req.mm_kwargs)
671

672
673
674
675
676
677
678
679
        # Input all modalities at once
        mm_kwargs_combined: BatchedTensorInputs = {}
        for _, _, mm_kwargs_group in group_mm_kwargs_by_modality(
                mm_kwargs,
                device=self.device,
                pin_memory=self.pin_memory,
        ):
            mm_kwargs_combined.update(mm_kwargs_group)
680

681
        return mm_kwargs_combined
682

683
    def _dummy_mm_kwargs(self, num_seqs: int) -> BatchedTensorInputs:
684
        if not self.is_multimodal_raw_input_only_model:
685
            return {}
686

687
688
689
690
691
        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)
692

693
694
695
696
697
698
699
700
701
702
703
704
705
706
707
708
709
710
711
712
    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

713
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
740
741
742
743
744
745
746
747
748
749
750
751
752
753
754
755
756
757
758
759
760
761
762
763
764
765
766
767
768
769
770
771
772
773
774
775
776
777
778
779
    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)
            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)
        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)
            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])

780
    def _prepare_inputs(
781
782
        self,
        scheduler_output: "SchedulerOutput",
783
784
    ) -> tuple[dict[str, Any], torch.Tensor, Optional[SpecDecodeMetadata],
               np.ndarray, Optional[CommonAttentionMetadata], int]:
785
786
787
788
789
790
        """
        :return: tuple[
            attn_metadata: layer-to-attention_metadata mapping,
            logits_indices, spec_decode_metadata
        ]
        """
791
792
793
794
795
796
797
        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.
798
        self.input_batch.block_table.commit_block_table(num_reqs)
799
800

        # Get the number of scheduled tokens for each request.
801
802
803
804
        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)
805
806
807

        # Get request indices.
        # E.g., [2, 5, 3] -> [0, 0, 1, 1, 1, 1, 1, 2, 2, 2]
808
809
        req_indices = np.repeat(self.arange_np[:num_reqs],
                                num_scheduled_tokens)
810

811
812
813
814
        # 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)
815
816

        # Get positions.
817
        positions_np = self.positions.np[:total_num_scheduled_tokens]
818
819
820
821
        np.add(self.input_batch.num_computed_tokens_cpu[req_indices],
               arange,
               out=positions_np)

822
823
        # Calculate M-RoPE positions.
        # Only relevant for models using M-RoPE (e.g, Qwen2-VL)
824
        if self.uses_mrope:
825
826
            self._calc_mrope_positions(scheduler_output)

827
828
829
830
        # 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.
831
832
        token_indices = (positions_np +
                         req_indices * self.input_batch.token_ids_cpu.shape[1])
833

834
835
836
837
        # 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(),
838
                           0,
839
                           torch.from_numpy(token_indices),
840
                           out=self.input_ids.cpu[:total_num_scheduled_tokens])
841

842
843
844
845
        self.input_batch.block_table.compute_slot_mapping(
            req_indices, positions_np)
        self.input_batch.block_table.commit_slot_mapping(
            total_num_scheduled_tokens)
846
847

        # Prepare the attention metadata.
848
849
        self.query_start_loc.np[0] = 0
        self.query_start_loc.np[1:num_reqs + 1] = cu_num_tokens
850
851
        # Note: pad query_start_loc to be non-decreasing, as kernels
        # like FlashAttention requires that
852
853
854
        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]
855

856
        self.seq_lens.np[:num_reqs] = (
857
858
            self.input_batch.num_computed_tokens_cpu[:num_reqs] +
            num_scheduled_tokens)
859
        # Fill unused with 0 for full cuda graph mode.
860
861
862
863
        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()
864
865

        # Copy the tensors to the GPU.
866
867
        self._prepare_input_ids(total_num_scheduled_tokens, cu_num_tokens)

868
        if self.uses_mrope:
869
            # Only relevant for models using M-RoPE (e.g, Qwen2-VL)
870
871
            self.mrope_positions.gpu[:, :total_num_scheduled_tokens].copy_(
                self.mrope_positions.cpu[:, :total_num_scheduled_tokens],
872
873
874
                non_blocking=True)
        else:
            # Common case (1D positions)
875
            self.positions.copy_to_gpu(total_num_scheduled_tokens)
876

877
878
879
880
881
882
883
884
885
886
887
888
889
890
891
892
893
894
895
896
897
898
899
900
901
902
        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
            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)
            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)

            spec_decode_metadata = self._calc_spec_decode_metadata(
                num_draft_tokens, cu_num_tokens)
            logits_indices = spec_decode_metadata.logits_indices

        logits_indices_padded = None
        if self.cache_config.kv_sharing_fast_prefill:
903
            logits_indices_padded = self._prepare_kv_sharing_fast_prefill(
904
905
                logits_indices)

906
        attn_metadata: dict[str, Any] = {}
907

908
        # Used in the below loop.
909
910
        query_start_loc_cpu = self.query_start_loc.cpu[:num_reqs + 1]
        seq_lens_cpu = self.seq_lens.cpu[:num_reqs]
911
912
913
914
        num_computed_tokens_cpu = (
            self.input_batch.num_computed_tokens_cpu_tensor[:num_reqs])
        spec_decode_common_attn_metadata = None

915
916
917
918
919
        # 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):

920
921
922
923
924
925
926
            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,
927
928
929
930
931
932
933
                    device=self.device,
                )
                slot_mapping = torch.zeros(
                    (total_num_scheduled_tokens, ),
                    dtype=torch.int64,
                    device=self.device,
                )
934
935
936
937
938
939
940
941
942
943
944
945
946
                num_common_prefix_blocks = 0
            else:
                blk_table = self.input_batch.block_table[kv_cache_group_id]
                blk_table_tensor = blk_table.get_device_tensor()[:num_reqs]
                slot_mapping = blk_table.slot_mapping[:
                                                      total_num_scheduled_tokens]

                # Fill unused with -1. Needed for reshape_and_cache in full cuda
                # graph mode.
                blk_table.slot_mapping[total_num_scheduled_tokens:].fill_(-1)
                num_common_prefix_blocks = (
                    scheduler_output.
                    num_common_prefix_blocks[kv_cache_group_id])
947

948
            common_attn_metadata = CommonAttentionMetadata(
949
950
951
952
953
                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,
954
955
956
                num_reqs=num_reqs,
                num_actual_tokens=total_num_scheduled_tokens,
                max_query_len=max_num_scheduled_tokens,
957
                max_seq_len=max_seq_len,
958
959
                block_table_tensor=blk_table_tensor,
                slot_mapping=slot_mapping,
960
961
                logits_indices_padded=logits_indices_padded,
                num_logits_indices=logits_indices.size(0),
962
                causal=True,
963
964
965
966
967
968
            )

            if self.speculative_config and \
                spec_decode_common_attn_metadata is None:
                spec_decode_common_attn_metadata = common_attn_metadata

969
970
971
972
973
974
975
            for attn_group in self.attn_groups[kv_cache_group_id]:
                # Prepare for cascade attention if enabled & beneficial.
                common_prefix_len = 0
                builder = attn_group.metadata_builder
                if self.cascade_attn_enabled:
                    common_prefix_len = self._compute_cascade_attn_prefix_len(
                        num_scheduled_tokens,
976
                        num_common_prefix_blocks,
977
978
979
                        kv_cache_group_spec.kv_cache_spec,
                        builder,
                    )
980

981
982
983
                attn_metadata_i = (builder.build(
                    common_prefix_len=common_prefix_len,
                    common_attn_metadata=common_attn_metadata,
984
985
                ))

986
987
                for layer_name in attn_group.layer_names:
                    attn_metadata[layer_name] = attn_metadata_i
988

989
990
991
992
        # Hot-Swap lora model
        if self.lora_config:
            self.set_active_loras(self.input_batch, num_scheduled_tokens)

993
994
995
        return (attn_metadata, logits_indices, spec_decode_metadata,
                num_scheduled_tokens, spec_decode_common_attn_metadata,
                max_num_scheduled_tokens)
996

997
998
999
1000
    def _compute_cascade_attn_prefix_len(
        self,
        num_scheduled_tokens: np.ndarray,
        num_common_prefix_blocks: int,
1001
1002
        kv_cache_spec: KVCacheSpec,
        attn_metadata_builder: AttentionMetadataBuilder,
1003
1004
1005
1006
1007
1008
1009
1010
1011
1012
1013
1014
1015
1016
1017
1018
1019
1020
    ) -> 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.
        """
1021
        common_prefix_len = num_common_prefix_blocks * kv_cache_spec.block_size
1022
1023
1024
1025
1026
1027
1028
1029
1030
1031
1032
1033
1034
1035
1036
1037
1038
1039
1040
1041
1042
1043
1044
1045
1046
1047
1048
1049
1050
1051
1052
1053
1054
1055
1056
1057
1058
        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]
1059
        # Request 3's num_computed_tokens: 3 (i.e., [A, B, C])
1060
1061
1062
1063
1064
1065
1066
1067
1068
1069
        # 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.
1070
1071
1072
1073
1074
        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))
1075
1076
1077
1078
        use_local_attention = (
            isinstance(kv_cache_spec, ChunkedLocalAttentionSpec)
            or (isinstance(kv_cache_spec, FullAttentionSpec)
                and kv_cache_spec.attention_chunk_size is not None))
1079
1080
        assert isinstance(kv_cache_spec, AttentionSpec)
        use_cascade = attn_metadata_builder.use_cascade_attention(
1081
1082
1083
            common_prefix_len=common_prefix_len,
            query_lens=num_scheduled_tokens,
            num_query_heads=self.num_query_heads,
1084
            num_kv_heads=kv_cache_spec.num_kv_heads,
1085
            use_alibi=self.use_alibi,
1086
            use_sliding_window=use_sliding_window,
1087
            use_local_attention=use_local_attention,
1088
1089
1090
1091
            num_sms=self.num_sms,
        )
        return common_prefix_len if use_cascade else 0

1092
1093
    def _calc_mrope_positions(self, scheduler_output: "SchedulerOutput"):
        mrope_pos_ptr = 0
1094
        for index, req_id in enumerate(self.input_batch.req_ids):
1095
1096
1097
1098
1099
1100
1101
1102
1103
1104
1105
1106
1107
1108
1109
1110
1111
1112
1113
1114
1115
1116
1117
1118
1119
1120
1121
            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]
            num_prompt_tokens = len(req.prompt_token_ids)

            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

1122
1123
                self.mrope_positions.cpu[:, dst_start:dst_end] = (
                    req.mrope_positions[:, src_start:src_end])
1124
1125
1126
1127
1128
1129
1130
                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

1131
                MRotaryEmbedding.get_next_input_positions_tensor(
1132
                    out=self.mrope_positions.np,
1133
1134
1135
1136
1137
                    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,
                )
1138
1139
1140

                mrope_pos_ptr += completion_part_len

1141
1142
    def _calc_spec_decode_metadata(
        self,
1143
1144
1145
1146
1147
1148
1149
1150
1151
1152
1153
1154
1155
1156
1157
1158
        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
1159
1160
1161
1162
1163
1164

        # 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]
1165
1166
        logits_indices = np.repeat(
            cu_num_scheduled_tokens - num_sampled_tokens, num_sampled_tokens)
1167
        # Step 3. [0, 1, 2, 3, 103, 104, 105, 106, 206, 207, 208]
1168
1169
1170
1171
1172
1173
        logits_indices += arange

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

        # Compute the draft logits indices.
1174
1175
1176
1177
        # 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)
1178
1179
1180
1181
1182
1183
1184
1185
1186
1187
1188
1189
1190
1191
        # [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(
1192
1193
            self.device, non_blocking=True)

1194
1195
        # Compute the draft token ids.
        # draft_token_indices:      [  1,   2,   3, 105, 106, 208]
1196
        draft_token_ids = self.input_ids.gpu[logits_indices]
1197
1198
1199
1200
1201
1202
1203
1204
1205
1206
1207
1208
        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

1209
1210
1211
1212
1213
1214
1215
1216
1217
1218
1219
1220
1221
1222
1223
1224
1225
1226
1227
1228
1229
1230
1231
1232
1233
1234
    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

1235
    def _execute_mm_encoder(self, scheduler_output: "SchedulerOutput"):
1236
1237
1238
1239
        scheduled_encoder_inputs = scheduler_output.scheduled_encoder_inputs
        if not scheduled_encoder_inputs:
            return
        # Batch the multi-modal inputs.
1240
        mm_kwargs = list[MultiModalKwargsItem]()
1241
1242
        # list of tuple (mm_hash, position_info)
        mm_hashes_pos = list[tuple[str, PlaceholderRange]]()
1243
1244
        for req_id, encoder_input_ids in scheduled_encoder_inputs.items():
            req_state = self.requests[req_id]
1245
1246

            for mm_input_id in encoder_input_ids:
1247
                mm_hash = req_state.mm_hashes[mm_input_id]
1248
                mm_kwargs.append(req_state.mm_kwargs[mm_input_id])
1249
1250
                mm_hashes_pos.append(
                    (mm_hash, req_state.mm_positions[mm_input_id]))
1251
1252
1253
1254
1255
1256
1257
1258
1259

        # 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.
        encoder_outputs = []
1260
1261
        for _, num_items, mm_kwargs_group in group_mm_kwargs_by_modality(
                mm_kwargs,
1262
                device=self.device,
1263
1264
                pin_memory=self.pin_memory,
        ):
1265
1266
1267
1268
1269
1270
1271
1272
            # 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 = self.model.get_multimodal_embeddings(
1273
                **mm_kwargs_group)
1274

1275
1276
            sanity_check_mm_encoder_outputs(
                curr_group_outputs,
1277
                expected_num_items=num_items,
1278
1279
            )

1280
1281
            for output in curr_group_outputs:
                encoder_outputs.append(output)
1282

1283
1284
1285
        # 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(
1286
1287
1288
1289
1290
                output,
                is_embed=pos_info.is_embed,
            )

    def _gather_mm_embeddings(
1291
1292
        self,
        scheduler_output: "SchedulerOutput",
1293
        shift_computed_tokens: int = 0,
1294
    ) -> list[torch.Tensor]:
1295
        mm_embeds: list[torch.Tensor] = []
1296
        for req_id in self.input_batch.req_ids:
1297
1298
1299
            num_scheduled_tokens = scheduler_output.num_scheduled_tokens[
                req_id]
            req_state = self.requests[req_id]
1300
1301
            num_computed_tokens = \
                req_state.num_computed_tokens + shift_computed_tokens
1302
            mm_positions = req_state.mm_positions
1303
            mm_hashes = req_state.mm_hashes
1304
            for i, pos_info in enumerate(mm_positions):
1305
1306
                start_pos = pos_info.offset
                num_encoder_tokens = pos_info.length
1307
1308
1309
1310
1311
1312
1313
1314
1315
1316
1317
1318
1319
1320
1321
1322

                # 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,
1323
1324
                    num_encoder_tokens,
                )
1325
                assert start_idx < end_idx
1326
1327
1328
1329
1330

                mm_hash = mm_hashes[i]
                encoder_output = self.encoder_cache.get(mm_hash, None)
                assert encoder_output is not None,\
                    f"Encoder cache miss for {mm_hash}."
1331
1332
1333
1334
1335
1336
1337
1338
1339
1340

                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,
                )
                mm_embeds.append(mm_embeds_item)
        return mm_embeds
1341

1342
    def get_model(self) -> nn.Module:
1343
1344
1345
        # get raw model out of the cudagraph wrapper.
        if isinstance(self.model, CUDAGraphWrapper):
            return self.model.unwrap()
1346
1347
        return self.model

1348
1349
1350
1351
1352
1353
1354
1355
1356
1357
1358
1359
1360
1361
1362
    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

1363
1364
1365
1366
1367
    def get_supported_pooling_tasks(self) -> list[PoolingTask]:
        model = self.get_model()
        if not is_pooling_model(model):
            return []

1368
1369
1370
1371
1372
1373
        supported_tasks = list(model.pooler.get_supported_tasks())

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

1374
1375
1376
1377
1378
1379
1380
1381
1382
1383
1384
            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.")
1385
1386

        return supported_tasks
1387

1388
1389
1390
1391
1392
1393
1394
1395
1396
1397
    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)

1398
1399
1400
1401
1402
1403
1404
1405
1406
    def apply_grammar_bitmask(
        self,
        scheduler_output: "SchedulerOutput",
        logits: torch.Tensor,
    ):
        grammar_bitmask = scheduler_output.grammar_bitmask
        if grammar_bitmask is None:
            return

1407
1408
1409
1410
1411
1412
1413
1414
1415
        # We receive the structured output bitmask from the scheduler,
        # compacted to contain bitmasks only for structured output requests.
        # The order of the requests in the bitmask is not guaranteed to be the
        # same as the order of the requests in the gpu runner's batch. We need
        # to sort the bitmask to match the order of the requests used here.

        # Get the batch indices of the structured output requests.
        # Keep track of the number of speculative tokens scheduled for every
        # request in the batch, as the logit indices are offset by this amount.
1416
        struct_out_req_batch_indices: dict[str, int] = {}
1417
1418
1419
1420
1421
1422
1423
1424
1425
1426
1427
1428
1429
        cumulative_offset = 0
        seq = sorted(self.input_batch.req_id_to_index.items(),
                     key=lambda x: x[1])
        for req_id, batch_index in seq:
            logit_index = batch_index + cumulative_offset
            cumulative_offset += len(
                scheduler_output.scheduled_spec_decode_tokens.get(req_id, []))
            if req_id in scheduler_output.structured_output_request_ids:
                struct_out_req_batch_indices[req_id] = logit_index

        out_indices = []

        # Reorder the bitmask to match the order of the requests in the batch.
1430
1431
1432
1433
        sorted_bitmask = np.full(shape=(logits.shape[0],
                                        grammar_bitmask.shape[1]),
                                 fill_value=-1,
                                 dtype=grammar_bitmask.dtype)
1434
1435
1436
1437
1438
1439
1440
1441
1442
1443
1444
1445
1446
        cumulative_index = 0
        seq = sorted(scheduler_output.structured_output_request_ids.items(),
                     key=lambda x: x[1])
        for req_id, _ in seq:
            logit_index = struct_out_req_batch_indices[req_id]
            num_spec_tokens = len(
                scheduler_output.scheduled_spec_decode_tokens.get(req_id, []))
            for i in range(1 + num_spec_tokens):
                sorted_bitmask[logit_index + i] = \
                    grammar_bitmask[cumulative_index + i]
                out_indices.append(logit_index + i)
            cumulative_index += 1 + num_spec_tokens
        grammar_bitmask = sorted_bitmask
1447

1448
        # If the length of out indices and the logits have the same shape
1449
1450
        # we don't need to pass indices to the kernel,
        # since the bitmask is already aligned with the logits.
1451
        skip_out_indices = len(out_indices) == logits.shape[0]
1452

1453
1454
        # Serialization of np.ndarray is much more efficient than a tensor,
        # so we receive it in that format.
1455
        grammar_bitmask = torch.from_numpy(grammar_bitmask).contiguous()
1456

1457
        xgr.apply_token_bitmask_inplace(
1458
1459
            logits,
            grammar_bitmask.to(self.device, non_blocking=True),
1460
            indices=out_indices if not skip_out_indices else None,
1461
1462
        )

1463
1464
1465
1466
1467
1468
1469
    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
1470
        enabled_sp = self.compilation_config.pass_config. \
1471
1472
1473
1474
1475
1476
1477
1478
1479
1480
1481
1482
1483
            enable_sequence_parallelism
        if enabled_sp:
            # When sequence parallelism is enabled, we always pad num_tokens
            # to be a multiple of tensor_parallel_size (tp) earlier
            assert num_tokens % tp == 0
        is_residual_scattered = tp > 1 and enabled_sp \
            and num_tokens % tp == 0

        # 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():
1484
                is_scattered = k == "residual" and is_residual_scattered
1485
1486
1487
1488
1489
1490
1491
1492
1493
1494
1495
1496
                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:
            v[:num_tokens // tp]
            if k == "residual" and is_residual_scattered else v[:num_tokens]
            for k, v in self.intermediate_tensors.items()
        })

1497
1498
1499
1500
1501
1502
1503
1504
1505
1506
    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
1507
1508
        model = self.get_model()
        assert is_mixture_of_experts(model)
1509
        self.eplb_state.step(
1510
            model,
1511
1512
            is_dummy,
            is_profile,
1513
            log_stats=self.parallel_config.eplb_config.log_balancedness,
1514
1515
        )

1516
1517
    def get_dp_padding(self,
                       num_tokens: int) -> tuple[int, Optional[torch.Tensor]]:
1518
1519
        dp_size = self.vllm_config.parallel_config.data_parallel_size
        dp_rank = self.vllm_config.parallel_config.data_parallel_rank
1520
1521
1522
1523
1524
1525
1526
1527
1528

        # 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:
1529
            # Early exit.
1530
            return 0, None
1531
1532
1533
1534

        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()
1535
1536
1537
1538
1539
        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
1540

1541
1542
1543
1544
1545
    def _pool(
        self,
        hidden_states: torch.Tensor,
        num_scheduled_tokens: int,
        num_scheduled_tokens_np: np.ndarray,
1546
        kv_connector_output: Optional[KVConnectorOutput],
1547
1548
1549
1550
1551
1552
    ) -> 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"

1553
        hidden_states = hidden_states[:num_scheduled_tokens]
1554
        pooling_metadata = self.input_batch.get_pooling_metadata()
1555
1556
        pooling_metadata.build_pooling_cursor(num_scheduled_tokens_np.tolist(),
                                              device=hidden_states.device)
1557
        seq_lens_cpu = self.seq_lens.cpu[:self.input_batch.num_reqs]
1558

1559
        # Pooling models D2H & synchronize occurs in pooler.py:build_output
1560
        raw_pooler_output = self.model.pooler(
1561
            hidden_states=hidden_states, pooling_metadata=pooling_metadata)
1562
1563
1564

        pooler_output: list[Optional[torch.Tensor]] = []
        for raw_output, seq_len, prompt_len in zip(
1565
                raw_pooler_output, seq_lens_cpu, pooling_metadata.prompt_lens):
1566

1567
1568
            output = raw_output.data if seq_len == prompt_len else None
            pooler_output.append(output)
1569
1570
1571
1572
1573
1574
1575
1576

        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,
1577
            kv_connector_output=kv_connector_output,
1578
1579
        )

1580
1581
1582
1583
    @torch.inference_mode()
    def execute_model(
        self,
        scheduler_output: "SchedulerOutput",
1584
        intermediate_tensors: Optional[IntermediateTensors] = None,
1585
    ) -> Union[ModelRunnerOutput, AsyncModelRunnerOutput, IntermediateTensors]:
1586
        self._update_states(scheduler_output)
1587
        if not scheduler_output.total_num_scheduled_tokens:
1588
1589
1590
            if not has_kv_transfer_group():
                # Return empty ModelRunnerOutput if there's no work to do.
                return EMPTY_MODEL_RUNNER_OUTPUT
Robert Shaw's avatar
Robert Shaw committed
1591

1592
1593
            return self.kv_connector_no_forward(scheduler_output,
                                                self.vllm_config)
1594

1595
1596
1597
1598
1599
1600
        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")

1601
        # Prepare the decoder inputs.
1602
1603
        (attn_metadata, logits_indices, spec_decode_metadata,
         num_scheduled_tokens_np, spec_decode_common_attn_metadata,
1604
         max_query_len) = self._prepare_inputs(scheduler_output)
1605

1606
        num_scheduled_tokens = scheduler_output.total_num_scheduled_tokens
1607
        if (self.compilation_config.cudagraph_mode != CUDAGraphMode.NONE
1608
                and not envs.VLLM_DISABLE_PAD_FOR_CUDAGRAPH
1609
                and num_scheduled_tokens <= self.cudagraph_batch_sizes[-1]):
1610
            # Use CUDA graphs.
1611
            # Add padding to the batch size.
1612
            num_input_tokens = self.vllm_config.pad_for_cudagraph(
1613
1614
1615
                num_scheduled_tokens)
        else:
            # Eager mode.
1616
1617
1618
            # Pad tokens to multiple of tensor_parallel_size when
            # enabled collective fusion for SP
            tp_size = self.vllm_config.parallel_config.tensor_parallel_size
1619
            if self.compilation_config.pass_config. \
1620
1621
1622
1623
                enable_sequence_parallelism and tp_size > 1:
                num_input_tokens = round_up(num_scheduled_tokens, tp_size)
            else:
                num_input_tokens = num_scheduled_tokens
1624

1625
        # Padding for DP
1626
1627
        num_pad, num_tokens_across_dp = self.get_dp_padding(num_input_tokens)
        num_input_tokens += num_pad
1628

1629
1630
        # _prepare_inputs may reorder the batch, so we must gather multi
        # modal outputs after that to ensure the correct order
1631
        if self.supports_mm_inputs:
1632
1633
1634
1635
1636
1637
            # Run the multimodal encoder if any.
            self._execute_mm_encoder(scheduler_output)
            mm_embeds = self._gather_mm_embeddings(scheduler_output)
        else:
            mm_embeds = []

1638
        if self.supports_mm_inputs and get_pp_group().is_first_rank:
1639
1640
1641
            # 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.
1642
            inputs_embeds_scheduled = self.model.get_input_embeddings(
1643
                input_ids=self.input_ids.gpu[:num_scheduled_tokens],
1644
1645
                multimodal_embeddings=mm_embeds or None,
            )
1646

1647
            # TODO(woosuk): Avoid the copy. Optimize.
1648
1649
1650
            self.inputs_embeds[:num_scheduled_tokens].copy_(
                inputs_embeds_scheduled)

1651
            input_ids = None
1652
            inputs_embeds = self.inputs_embeds[:num_input_tokens]
1653
1654
1655
1656
            model_kwargs = {
                **self._init_model_kwargs(num_scheduled_tokens),
                **self._extract_mm_kwargs(scheduler_output),
            }
1657
        else:
1658
1659
1660
1661
            # 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.
1662
            input_ids = self.input_ids.gpu[:num_input_tokens]
1663
            inputs_embeds = None
1664
            model_kwargs = self._init_model_kwargs(num_input_tokens)
1665
        if self.uses_mrope:
1666
            positions = self.mrope_positions.gpu[:, :num_input_tokens]
1667
        else:
1668
            positions = self.positions.gpu[:num_input_tokens]
1669

1670
1671
1672
        if get_pp_group().is_first_rank:
            intermediate_tensors = None
        else:
1673
1674
            intermediate_tensors = self.sync_and_slice_intermediate_tensors(
                num_input_tokens, intermediate_tensors, True)
1675

1676
1677
1678
1679
1680
1681
        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 = \
            self.cudagraph_dispatcher.dispatch(batch_descriptor)
1682

1683
        # Run the model.
1684
        # Use persistent buffers for CUDA graphs.
1685
1686
1687
1688
1689
        with set_forward_context(
                attn_metadata,
                self.vllm_config,
                num_tokens=num_input_tokens,
                num_tokens_across_dp=num_tokens_across_dp,
1690
1691
                cudagraph_runtime_mode=cudagraph_runtime_mode,
                batch_descriptor=batch_descriptor,
1692
1693
        ), self.maybe_get_kv_connector_output(
                scheduler_output) as kv_connector_output:
1694

Robert Shaw's avatar
Robert Shaw committed
1695
            model_output = self.model(
1696
                input_ids=input_ids,
1697
                positions=positions,
1698
                intermediate_tensors=intermediate_tensors,
1699
                inputs_embeds=inputs_embeds,
1700
                **model_kwargs,
1701
            )
1702
1703

        if self.use_aux_hidden_state_outputs:
Robert Shaw's avatar
Robert Shaw committed
1704
            hidden_states, aux_hidden_states = model_output
1705
        else:
Robert Shaw's avatar
Robert Shaw committed
1706
            hidden_states = model_output
1707
1708
            aux_hidden_states = None

1709
1710
1711
1712
1713
1714
1715
        # 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
        broadcast_pp_output = \
            self.parallel_config.distributed_executor_backend \
            == "external_launcher" and len(get_pp_group().ranks) > 0
1716
        if not get_pp_group().is_last_rank:
1717
            # For mid-pipeline stages, return the hidden states.
1718
            assert isinstance(hidden_states, IntermediateTensors)
1719
            if not broadcast_pp_output:
1720
                hidden_states.kv_connector_output = kv_connector_output
1721
1722
1723
1724
1725
                return hidden_states
            get_pp_group().send_tensor_dict(hidden_states.tensors,
                                            all_gather_group=get_tp_group())
            logits = None
        else:
1726
            if self.is_pooling_model:
1727
                return self._pool(hidden_states, num_scheduled_tokens,
1728
                                  num_scheduled_tokens_np, kv_connector_output)
1729

1730
1731
1732
1733
1734
1735
1736
1737
1738
1739
            sample_hidden_states = hidden_states[logits_indices]
            logits = self.model.compute_logits(sample_hidden_states, None)
        if broadcast_pp_output:
            model_output_broadcast_data = {
                "logits": logits.contiguous(),
            } if logits is not None else {}
            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"]
1740

1741
1742
1743
1744
        # Apply structured output bitmasks if present
        if scheduler_output.grammar_bitmask is not None:
            self.apply_grammar_bitmask(scheduler_output, logits)

1745
        # Sample the next token and get logprobs if needed.
1746
        sampling_metadata = self.input_batch.sampling_metadata
1747
        if spec_decode_metadata is None:
1748
            sampler_output = self.sampler(
1749
1750
1751
1752
                logits=logits,
                sampling_metadata=sampling_metadata,
            )
        else:
1753
1754
1755
1756
            # 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.
1757
            assert logits is not None
1758
            bonus_logits = logits[spec_decode_metadata.bonus_logits_indices]
1759
            sampler_output = self.sampler(
1760
                logits=bonus_logits,
1761
1762
1763
                sampling_metadata=sampling_metadata,
            )
            bonus_token_ids = sampler_output.sampled_token_ids
1764

1765
1766
1767
            # 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.
1768
            target_logits = logits[spec_decode_metadata.target_logits_indices]
1769
            output_token_ids = self.rejection_sampler(
1770
                spec_decode_metadata,
1771
                None,  # draft_probs
1772
                target_logits,
1773
                bonus_token_ids,
1774
1775
                sampling_metadata,
            )
1776
            sampler_output.sampled_token_ids = output_token_ids
1777

1778
1779
1780
1781
        num_nans_in_logits = {}
        if envs.VLLM_COMPUTE_NANS_IN_LOGITS:
            num_nans_in_logits = self._get_nans_in_logits(logits)

1782
1783
        # TODO(woosuk): The following loop can be slow since it iterates over
        # the requests one by one. Optimize.
1784
1785
        discard_sampled_tokens_req_indices = []
        for i, req_id in enumerate(self.input_batch.req_ids):
1786
1787
1788
            req_state = self.requests[req_id]
            seq_len = (req_state.num_computed_tokens +
                       scheduler_output.num_scheduled_tokens[req_id])
1789
            if seq_len < req_state.num_tokens:
1790
                # Ignore the sampled token for partial prefills.
1791
                # Rewind the generator state as if the token was not sampled.
1792
                # This relies on cuda-specific torch-internal impl details
1793
1794
1795
1796
1797
1798
                generator = self.input_batch.generators.get(i)
                if generator is not None:
                    generator.set_offset(generator.get_offset() - 4)
                # Record the index of the request that should not be sampled,
                # so that we could clear the sampled tokens before returning.
                discard_sampled_tokens_req_indices.append(i)
1799

1800
1801
1802
1803
1804
1805
        # 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()

1806
1807
        # NOTE: GPU -> CPU Sync happens here.
        # Move as many CPU operations as possible before this sync point.
1808
1809
1810
1811
1812
1813
        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(
1814
            hidden_states[:num_scheduled_tokens],
1815
            scheduler_output.num_scheduled_tokens,
1816
1817
        )

1818
        num_sampled_tokens = sampler_output.sampled_token_ids.shape[0]
1819
        sampled_token_ids = sampler_output.sampled_token_ids
1820
1821
1822
1823
1824
1825
1826
1827
1828
1829
1830
1831
1832
1833
1834
        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:
                valid_sampled_token_ids[i].clear()
1835
        else:
1836
1837
1838
1839
1840
1841
1842
1843
1844
1845
1846
1847
1848
1849
1850
1851
1852
            valid_sampled_token_ids = []
            invalid_req_indices = list(discard_sampled_tokens_req_indices)
            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
            }
1853

1854
1855
1856
1857
1858
        # 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.
1859
        req_ids = self.input_batch.req_ids
1860
1861
1862
1863
1864
1865
        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]
1866
1867
1868
1869
1870
1871
1872
1873
1874
1875
1876
1877
1878
1879
            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
            self.input_batch.num_tokens_no_spec[req_idx] = end_idx
            self.input_batch.num_tokens[req_idx] = end_idx
1880

1881
            req_id = req_ids[req_idx]
1882
1883
1884
            req_state = self.requests[req_id]
            req_state.output_token_ids.extend(sampled_ids)

1885
        if self.speculative_config:
1886
            assert spec_decode_common_attn_metadata is not None
1887
            self._draft_token_ids = self.propose_draft_token_ids(
1888
1889
1890
1891
1892
1893
1894
                scheduler_output,
                valid_sampled_token_ids,
                sampling_metadata,
                hidden_states,
                sample_hidden_states,
                aux_hidden_states,
                spec_decode_metadata,
1895
                spec_decode_common_attn_metadata,
1896
1897
1898
1899
            )

        self.eplb_step()

1900
1901
1902
        output = ModelRunnerOutput(
            req_ids=req_ids_output_copy,
            req_id_to_index=req_id_to_index_output_copy,
1903
1904
1905
1906
            sampled_token_ids=valid_sampled_token_ids,
            logprobs=logprobs_lists,
            prompt_logprobs_dict=prompt_logprobs_dict,
            pooler_output=[],
1907
            kv_connector_output=kv_connector_output,
1908
1909
1910
            num_nans_in_logits=num_nans_in_logits,
        )

1911
1912
1913
1914
1915
1916
1917
1918
1919
1920
        if not self.use_async_scheduling:
            return output

        return AsyncGPUModelRunnerOutput(
            model_runner_output=output,
            sampled_token_ids=sampled_token_ids,
            invalid_req_indices=invalid_req_indices,
            async_output_copy_stream=self.async_output_copy_stream,
        )

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

1932
1933
1934
1935
1936
1937
1938
1939
1940
    def propose_draft_token_ids(
        self,
        scheduler_output: "SchedulerOutput",
        sampled_token_ids: list[list[int]],
        sampling_metadata: SamplingMetadata,
        hidden_states: torch.Tensor,
        sample_hidden_states: torch.Tensor,
        aux_hidden_states: Optional[torch.Tensor],
        spec_decode_metadata: Optional[SpecDecodeMetadata],
1941
        common_attn_metadata: CommonAttentionMetadata,
1942
    ) -> Union[list[list[int]], torch.Tensor]:
1943
1944
        num_scheduled_tokens = scheduler_output.total_num_scheduled_tokens
        if self.speculative_config.method == "ngram":
1945
            assert isinstance(self.drafter, NgramProposer)
1946
            draft_token_ids = self.propose_ngram_draft_token_ids(
1947
                sampled_token_ids)
1948
1949
        elif self.speculative_config.method == "medusa":
            assert isinstance(self.drafter, MedusaProposer)
1950
1951
            if sample_hidden_states.shape[0] == len(sampled_token_ids):
                # The input to the target model does not include draft tokens.
1952
1953
1954
1955
1956
1957
                hidden_states = sample_hidden_states
            else:
                indices = []
                offset = 0
                for num_draft, tokens in zip(
                        spec_decode_metadata.num_draft_tokens,
1958
                        sampled_token_ids):
1959
1960
                    indices.append(offset + len(tokens) - 1)
                    offset += num_draft + 1
1961
                indices = torch.tensor(indices, device=self.device)
1962
1963
                hidden_states = sample_hidden_states[indices]

1964
            draft_token_ids = self.drafter.propose(
1965
1966
1967
                target_hidden_states=hidden_states,
                sampling_metadata=sampling_metadata,
            )
1968
        elif self.speculative_config.use_eagle():
1969
1970
            assert isinstance(self.drafter, EagleProposer)
            # TODO(woosuk): Refactor the loop.
1971
            req_ids = self.input_batch.req_ids
1972
            next_token_ids: list[int] = []
1973
            for i, token_ids in enumerate(sampled_token_ids):
1974
1975
1976
1977
1978
1979
                if token_ids:
                    # Common case.
                    next_token_id = token_ids[-1]
                else:
                    # Partial prefill (rare case).
                    # Get the next token id from the request state.
1980
                    req_id = req_ids[i]
1981
1982
1983
1984
1985
                    req_state = self.requests[req_id]
                    seq_len = (req_state.num_computed_tokens +
                               scheduler_output.num_scheduled_tokens[req_id])
                    next_token_id = req_state.get_token_id(seq_len)
                next_token_ids.append(next_token_id)
1986
1987
1988
            next_token_ids = torch.tensor(next_token_ids,
                                          dtype=torch.int32,
                                          device=self.device)
Jiayi Yao's avatar
Jiayi Yao committed
1989

1990
1991
            if spec_decode_metadata is None:
                # input_ids can be None for multimodal models.
1992
                target_token_ids = self.input_ids.gpu[:num_scheduled_tokens]
1993
                # TODO(woosuk): Support M-RoPE.
1994
                target_positions = self.positions.gpu[:num_scheduled_tokens]
1995
                if self.use_aux_hidden_state_outputs:
1996
1997
1998
                    target_hidden_states = torch.cat(
                        [h[:num_scheduled_tokens] for h in aux_hidden_states],
                        dim=-1)
1999
2000
                else:
                    target_hidden_states = hidden_states[:num_scheduled_tokens]
2001
2002
2003
2004
            else:
                # TODO(woosuk): Refactor this.
                num_draft_tokens = spec_decode_metadata.num_draft_tokens
                num_rejected_tokens = [
2005
                    n + 1 - len(sampled_token_ids[i]) if n > 0 else 0
2006
2007
                    for i, n in enumerate(num_draft_tokens)
                ]
2008
2009
2010
2011
2012
2013
                num_rejected_tokens_cpu = torch.tensor(num_rejected_tokens,
                                                       dtype=torch.int32)
                common_attn_metadata, token_indices =\
                    self.drafter.prepare_inputs(
                    common_attn_metadata, num_rejected_tokens_cpu)

2014
                target_token_ids = self.input_ids.gpu[token_indices]
2015
                # TODO(woosuk): Support M-RoPE.
2016
                target_positions = self.positions.gpu[token_indices]
2017
                if self.use_aux_hidden_state_outputs:
2018
2019
                    target_hidden_states = torch.cat(
                        [h[token_indices] for h in aux_hidden_states], dim=-1)
2020
2021
                else:
                    target_hidden_states = hidden_states[token_indices]
2022
            mm_embeds = None
2023
            if self.supports_mm_inputs:
2024
2025
2026
                mm_embeds = self._gather_mm_embeddings(scheduler_output,
                                                       shift_computed_tokens=1)

2027
            draft_token_ids = self.drafter.propose(
2028
2029
2030
2031
2032
                target_token_ids=target_token_ids,
                target_positions=target_positions,
                target_hidden_states=target_hidden_states,
                next_token_ids=next_token_ids,
                sampling_metadata=sampling_metadata,
2033
                common_attn_metadata=common_attn_metadata,
2034
                mm_embeds=mm_embeds,
2035
            )
2036
        return draft_token_ids
2037

2038
    def propose_ngram_draft_token_ids(
2039
        self,
2040
2041
        sampled_token_ids: list[list[int]],
    ) -> list[list[int]]:
2042
        # TODO(woosuk): Optimize.
2043
        req_ids = self.input_batch.req_ids
2044
        draft_token_ids: list[list[int]] = []
2045
2046
2047
        for i, sampled_ids in enumerate(sampled_token_ids):
            num_sampled_ids = len(sampled_ids)
            if not num_sampled_ids:
2048
2049
2050
2051
                # Skip speculative decoding.
                draft_token_ids.append([])
                continue

2052
2053
            # Skip requests that require sampling parameters that are not
            # supported with speculative decoding.
2054
            req_id = req_ids[i]
2055
            if req_id in self.input_batch.spec_decode_unsupported_reqs:
2056
2057
2058
                draft_token_ids.append([])
                continue

2059
2060
            num_tokens = self.input_batch.num_tokens_no_spec[i]
            if num_tokens >= self.max_model_len:
2061
2062
2063
2064
                # Skip requests that have already reached the max model length.
                draft_token_ids.append([])
                continue

2065
            drafter_output = self.drafter.propose(
2066
                self.input_batch.token_ids_cpu[i, :num_tokens])
2067
2068
2069
2070
2071
2072
            if drafter_output is None or len(drafter_output) == 0:
                draft_token_ids.append([])
            else:
                draft_token_ids.append(drafter_output.tolist())
        return draft_token_ids

2073
2074
2075
2076
2077
2078
2079
2080
2081
2082
    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)

2083
2084
2085
2086
2087
    def load_model(self, eep_scale_up: bool = False) -> None:
        """
        Args:
            eep_scale_up: the model loading is for elastic EP scale up.
        """
2088
        logger.info("Starting to load model %s...", self.model_config.model)
2089
2090
2091
2092
2093
2094
2095
2096
2097
2098
2099
2100
2101
        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]
2102
            self.parallel_config.eplb_config.num_redundant_experts = (
2103
2104
2105
2106
2107
2108
2109
2110
2111
2112
2113
2114
2115
2116
                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

2117
        with DeviceMemoryProfiler() as m:
2118
            time_before_load = time.perf_counter()
2119
            model_loader = get_model_loader(self.load_config)
2120
2121
2122
            logger.info("Loading model from scratch...")
            self.model = model_loader.load_model(
                vllm_config=self.vllm_config, model_config=self.model_config)
2123
2124
2125
2126
2127
2128
            if self.lora_config:
                self.model = self.load_lora_model(self.model,
                                                  self.model_config,
                                                  self.scheduler_config,
                                                  self.lora_config,
                                                  self.device)
2129
2130
2131
            if hasattr(self, "drafter"):
                logger.info("Loading drafter model...")
                self.drafter.load_model(self.model)
2132
            if self.use_aux_hidden_state_outputs:
2133
2134
2135
2136
2137
2138
2139
                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")
2140
            time_after_load = time.perf_counter()
2141
        self.model_memory_usage = m.consumed_memory
2142
2143
        logger.info("Model loading took %.4f GiB and %.6f seconds",
                    self.model_memory_usage / GiB_bytes,
2144
                    time_after_load - time_before_load)
2145
        prepare_communication_buffer_for_model(self.model)
2146

2147
2148
2149
2150
2151
2152
2153
2154
        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,
2155
2156
2157
                global_expert_load,
                old_global_expert_indices,
                rank_mapping,
2158
2159
            )

2160
2161
2162
2163
2164
2165
2166
2167
2168
2169
        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
            self.model.compile(
                fullgraph=envs.VLLM_TEST_DYNAMO_FULLGRAPH_CAPTURE,
                backend=backend)
2170
2171
2172
2173
2174
2175
2176
2177
2178
            return
        # for other compilation levels, cudagraph behavior is controlled by
        # CudagraphWraper and CudagraphDispatcher of vllm.

        # wrap the model with full cudagraph wrapper if needed.
        if self.compilation_config.cudagraph_mode.has_full_cudagraphs():
            self.model = CUDAGraphWrapper(self.model,
                                          self.vllm_config,
                                          runtime_mode=CUDAGraphMode.FULL)
2179

2180
2181
2182
2183
2184
    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...")
2185
2186
        model = self.get_model()
        model_loader.load_weights(model, model_config=self.model_config)
2187

2188
2189
2190
2191
    def save_tensorized_model(
        self,
        tensorizer_config: "TensorizerConfig",
    ) -> None:
2192
        model = self.get_model()
2193
        TensorizerLoader.save_model(
2194
            model,
2195
            tensorizer_config=tensorizer_config,
2196
            model_config=self.model_config,
2197
2198
        )

2199
2200
2201
    def _get_prompt_logprobs_dict(
        self,
        hidden_states: torch.Tensor,
2202
        num_scheduled_tokens: dict[str, int],
2203
    ) -> dict[str, Optional[LogprobsTensors]]:
2204
2205
2206
2207
        num_prompt_logprobs_dict = self.input_batch.num_prompt_logprobs
        if not num_prompt_logprobs_dict:
            return {}

2208
        in_progress_dict = self.input_batch.in_progress_prompt_logprobs_cpu
2209
        prompt_logprobs_dict: dict[str, Optional[LogprobsTensors]] = {}
2210
2211
2212
2213
2214

        # 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():
2215
            num_tokens = num_scheduled_tokens[req_id]
2216
2217
2218
2219
2220
2221
2222

            # Get metadata for this request.
            request = self.requests[req_id]
            num_prompt_tokens = len(request.prompt_token_ids)
            prompt_token_ids = torch.tensor(request.prompt_token_ids).to(
                self.device, non_blocking=True)

2223
2224
2225
2226
2227
2228
2229
2230
2231
            # 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

2232
            # Determine number of logits to retrieve.
2233
2234
            start_idx = request.num_computed_tokens
            start_tok = start_idx + 1
2235
            num_remaining_tokens = num_prompt_tokens - start_tok
2236
            if num_tokens <= num_remaining_tokens:
2237
                # This is a chunk, more tokens remain.
2238
2239
2240
                # 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.
2241
2242
2243
2244
2245
                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)
2246
2247
2248
2249
2250
2251
2252
                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
2253
2254
2255
2256
2257

            # 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]
2258
            offset = self.query_start_loc.np[req_idx].item()
2259
2260
2261
2262
2263
2264
2265
2266
2267
            prompt_hidden_states = hidden_states[offset:offset + num_logits]
            logits = self.model.compute_logits(prompt_hidden_states, None)

            # 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.
2268
2269
            logprobs = self.sampler.compute_logprobs(logits)
            token_ids, logprobs, ranks = self.sampler.gather_logprobs(
2270
2271
2272
                logprobs, num_prompt_logprobs, tgt_token_ids)

            # Transfer GPU->CPU async.
2273
2274
2275
2276
2277
2278
2279
            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)
2280
2281
2282
2283
2284

        # 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]
2285
            del in_progress_dict[req_id]
2286
2287

        # Must synchronize the non-blocking GPU->CPU transfers.
2288
        if prompt_logprobs_dict:
2289
            self._sync_device()
2290
2291
2292

        return prompt_logprobs_dict

2293
2294
2295
2296
2297
2298
2299
2300
2301
2302
2303
2304
2305
2306
2307
2308
2309
2310
2311
2312
    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 {}

2313
2314
2315
2316
2317
2318
    @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
2319
         - during DP rank dummy run
2320
2321
2322
2323
2324
2325
2326
2327
2328
2329
2330
        """
        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(
2331
                    self.input_ids.gpu,
2332
2333
2334
2335
                    low=0,
                    high=self.model_config.get_vocab_size(),
                    dtype=input_ids.dtype)

2336
            logger.debug_once("Randomizing dummy data for DP Rank")
2337
2338
2339
2340
2341
            input_ids.copy_(rand_input_ids()[:input_ids.size(0)],
                            non_blocking=True)
            yield
            input_ids.fill_(0)

2342
2343
2344
2345
2346
2347
    def _get_mm_dummy_batch(
        self,
        modality: str,
        max_items_per_batch: int,
    ) -> BatchedTensorInputs:
        """Dummy data for profiling and precompiling multimodal models."""
2348
2349
        assert self.mm_budget is not None

2350
2351
2352
2353
        dummy_decoder_data = self.mm_registry.get_decoder_dummy_data(
            model_config=self.model_config,
            seq_len=self.max_num_tokens,
            mm_counts={modality: 1},
2354
            cache=self.mm_budget.cache,
2355
2356
2357
2358
        )
        dummy_mm_data = dummy_decoder_data.multi_modal_data

        # Result in the maximum GPU consumption of the model
2359
2360
        dummy_mm_item = dummy_mm_data[modality][0]
        dummy_mm_items = [dummy_mm_item] * max_items_per_batch
2361

2362
2363
        return next(mm_kwargs_group
                    for _, _, mm_kwargs_group in group_mm_kwargs_by_modality(
2364
                        dummy_mm_items,
2365
2366
2367
                        device=self.device,
                        pin_memory=self.pin_memory,
                    ))
2368

2369
2370
2371
2372
    @torch.inference_mode()
    def _dummy_run(
        self,
        num_tokens: int,
2373
2374
2375
        cudagraph_runtime_mode: CUDAGraphMode = CUDAGraphMode.NONE,
        force_attention: bool = False,
        uniform_decode: bool = False,
2376
2377
        skip_eplb: bool = False,
        is_profile: bool = False,
2378
        remove_lora: bool = True,
2379
    ) -> tuple[torch.Tensor, torch.Tensor]:
2380
2381
2382
2383
2384
2385
2386
2387
2388
2389
2390
        """
        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.
                - CUDAGraphMode.NONE: No cudagraph, for warm up and profile run
                - CUDAGraphMode.PIECEWISE: Piecewise cudagraph.
                - CUDAGraphMode.FULL: Full cudagraph, attention metadata is
                    needed.
2391
            force_attention: If True, always create attention metadata. Used to
2392
2393
2394
2395
                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.
2396
            remove_lora: If False, dummy LoRAs are not destroyed after the run
2397
2398
2399
2400
        """
        assert cudagraph_runtime_mode in {
            CUDAGraphMode.NONE, CUDAGraphMode.PIECEWISE, CUDAGraphMode.FULL
        }
2401

2402
        # Padding for DP
2403
2404
        num_pad, num_tokens_across_dp = self.get_dp_padding(num_tokens)
        num_tokens += num_pad
2405

2406
2407
2408
2409
2410
2411
2412
2413
2414
2415
2416
2417
2418
2419
2420
2421
        # If cudagraph_mode.decode_mode() == FULL and
        # cudagraph_mode.seperate_routine(). This means that we are using
        # 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

2422
2423
2424
2425
2426
        # 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
2427
2428
2429
2430
2431
2432
2433
2434
2435
2436
2437
2438
2439
        if uniform_decode:
            num_reqs = cdiv(num_tokens, max_query_len)
            assert num_reqs <= max_num_reqs, \
                "Do not capture num_reqs > max_num_reqs for uniform batch"
            num_scheduled_tokens_list = [max_query_len] * num_reqs
            if num_tokens % max_query_len != 0:
                num_scheduled_tokens_list[-1] = num_tokens % max_query_len
        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

2440
2441
2442
2443
        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)
2444

2445
        attn_metadata: Optional[dict[str, Any]] = None
2446
2447
2448

        # If force_attention is True, we always capture attention. Otherwise,
        # it only happens for cudagraph_runtime_mode=FULL.
2449
        if force_attention or cudagraph_runtime_mode == CUDAGraphMode.FULL:
2450
2451
            attn_metadata = {}

2452
            # Make sure max_model_len is used at the graph capture time.
2453
2454
2455
            self.seq_lens.np[:num_reqs] = self.max_model_len
            self.seq_lens.np[num_reqs:] = 0
            self.seq_lens.copy_to_gpu()
2456

2457
2458
            for kv_cache_group_id, kv_cache_group_spec in enumerate(
                    self.kv_cache_config.kv_cache_groups):
2459
                common_attn_metadata = CommonAttentionMetadata(
2460
2461
                    query_start_loc=self.query_start_loc.gpu[:num_reqs + 1],
                    query_start_loc_cpu=self.query_start_loc.cpu[:num_reqs +
2462
                                                                 1],
2463
2464
                    seq_lens=self.seq_lens.gpu[:num_reqs],
                    seq_lens_cpu=self.seq_lens.cpu[:num_reqs],
2465
2466
2467
2468
                    num_computed_tokens_cpu=self.input_batch.
                    num_computed_tokens_cpu_tensor[:num_reqs],
                    num_reqs=num_reqs,
                    num_actual_tokens=num_tokens,
2469
                    max_query_len=max_query_len,
2470
                    max_seq_len=self.max_model_len,
2471
2472
2473
                    block_table_tensor=self.input_batch.block_table[
                        kv_cache_group_id].get_device_tensor()[:num_reqs],
                    slot_mapping=self.input_batch.
2474
2475
                    block_table[kv_cache_group_id].slot_mapping[:num_tokens],
                    causal=True)
2476

2477
2478
2479
2480
2481
                for attn_group in self.attn_groups[kv_cache_group_id]:
                    attn_metadata_i = attn_group.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
2482

2483
        with self.maybe_dummy_run_with_lora(self.lora_config,
2484
                                            num_scheduled_tokens, remove_lora):
2485
            if self.supports_mm_inputs:
2486
2487
                input_ids = None
                inputs_embeds = self.inputs_embeds[:num_tokens]
2488
2489
2490
2491
                model_kwargs = {
                    **self._init_model_kwargs(num_tokens),
                    **self._dummy_mm_kwargs(num_reqs),
                }
2492
            else:
2493
                input_ids = self.input_ids.gpu[:num_tokens]
2494
                inputs_embeds = None
2495
                model_kwargs = self._init_model_kwargs(num_tokens)
2496

2497
            if self.uses_mrope:
2498
                positions = self.mrope_positions.gpu[:, :num_tokens]
2499
            else:
2500
                positions = self.positions.gpu[:num_tokens]
2501
2502
2503
2504
2505
2506
2507
2508
2509
2510

            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))
2511
2512
2513

                intermediate_tensors = self.sync_and_slice_intermediate_tensors(
                    num_tokens, None, False)
2514
2515
2516
2517
2518
2519
2520
2521
2522
2523
2524
2525
            if cudagraph_runtime_mode == CUDAGraphMode.NONE:
                batch_descriptor = None
            else:
                # filter out the valid batch descriptor
                _cg_mode, batch_descriptor = \
                    self.cudagraph_dispatcher.dispatch(
                        BatchDescriptor(num_tokens=num_tokens,
                                        uniform_decode=uniform_decode))
                # sanity check
                assert cudagraph_runtime_mode == _cg_mode, (
                    f"Cudagraph runtime mode mismatch at dummy_run. "
                    f"Expected {_cg_mode}, but got {cudagraph_runtime_mode}.")
2526

2527
            with self.maybe_randomize_inputs(input_ids), set_forward_context(
2528
2529
2530
                    attn_metadata,
                    self.vllm_config,
                    num_tokens=num_tokens,
2531
2532
2533
                    num_tokens_across_dp=num_tokens_across_dp,
                    cudagraph_runtime_mode=cudagraph_runtime_mode,
                    batch_descriptor=batch_descriptor):
2534
                outputs = self.model(
2535
2536
2537
2538
                    input_ids=input_ids,
                    positions=positions,
                    intermediate_tensors=intermediate_tensors,
                    inputs_embeds=inputs_embeds,
2539
                    **model_kwargs,
2540
                )
2541

2542
2543
2544
2545
            if self.use_aux_hidden_state_outputs:
                hidden_states, _ = outputs
            else:
                hidden_states = outputs
2546

2547
            if self.speculative_config and self.speculative_config.use_eagle():
2548
2549
2550
                assert isinstance(self.drafter, EagleProposer)
                self.drafter.dummy_run(num_tokens)

2551
2552
2553
2554
2555
2556
2557
2558
2559
2560
        # 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)

2561
        logit_indices = np.cumsum(num_scheduled_tokens) - 1
2562
        return hidden_states, hidden_states[logit_indices]
2563
2564
2565
2566
2567
2568

    @torch.inference_mode()
    def _dummy_sampler_run(
        self,
        hidden_states: torch.Tensor,
    ) -> torch.Tensor:
2569
2570
2571
2572
        # 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)
2573
2574
2575
2576
2577
2578
2579
2580
2581
2582
2583
2584
2585
2586
2587
2588
2589
2590
2591
2592
2593
2594
2595

        logits = self.model.compute_logits(hidden_states, None)
        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={},
2596
            logitsprocs=LogitsProcessors(),
2597
        )
2598
        try:
2599
2600
            sampler_output = self.sampler(logits=logits,
                                          sampling_metadata=dummy_metadata)
2601
2602
2603
2604
2605
2606
2607
2608
2609
        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
2610
        if self.speculative_config:
2611
2612
2613
2614
2615
2616
2617
2618
2619
2620
2621
2622
2623
2624
2625
2626
2627
2628
2629
2630
2631
2632
2633
2634
2635
2636
            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,
            )
2637
        return sampler_output
2638

2639
    def _dummy_pooler_run_task(
2640
2641
        self,
        hidden_states: torch.Tensor,
2642
2643
        task: PoolingTask,
    ) -> PoolerOutput:
2644
2645
2646
2647
2648
2649
2650
2651
2652
2653
2654
        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

2655
        dummy_prompt_lens = torch.tensor(
2656
2657
            num_scheduled_tokens_list,
            device="cpu",
2658
2659
2660
2661
        )
        dummy_token_ids = torch.zeros((num_reqs, req_num_tokens),
                                      dtype=torch.int32,
                                      device=self.device)
2662

2663
        model = cast(VllmModelForPooling, self.get_model())
2664
2665
        dummy_pooling_params = PoolingParams(task=task)
        to_update = model.pooler.get_pooling_updates(task)
2666
2667
        to_update.apply(dummy_pooling_params)

2668
        dummy_metadata = PoolingMetadata(
2669
2670
2671
2672
            prompt_lens=dummy_prompt_lens,
            prompt_token_ids=dummy_token_ids,
            pooling_params=[dummy_pooling_params] * num_reqs,
        )
2673

2674
2675
2676
        dummy_metadata.build_pooling_cursor(num_scheduled_tokens_list,
                                            device=hidden_states.device)

2677
        try:
2678
            return model.pooler(hidden_states=hidden_states,
2679
                                pooling_metadata=dummy_metadata)
2680
2681
2682
        except RuntimeError as e:
            if 'out of memory' in str(e):
                raise RuntimeError(
2683
2684
2685
                    "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 "
2686
2687
2688
                    "initializing the engine.") from e
            else:
                raise e
2689
2690
2691
2692
2693
2694
2695
2696
2697
2698
2699
2700
2701
2702
2703
2704

    @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)
            output_size[task] = output.get_data_nbytes()
            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)
2705

2706
    def profile_run(self) -> None:
2707
        # Profile with multimodal encoder & encoder cache.
2708
        if self.supports_mm_inputs:
2709
            if self.model_config.multimodal_config.skip_mm_profiling:
2710
                logger.info(
2711
2712
2713
2714
2715
2716
2717
2718
2719
2720
2721
                    "Skipping memory profiling for multimodal encoder and "
                    "encoder cache.")
            else:
                mm_budget = self.mm_budget
                assert mm_budget is not None

                # TODO: handle encoder-decoder models once we support them.
                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.
2722
2723
2724
                    dummy_modality = mm_budget.get_modality_with_max_tokens()
                    max_mm_items_per_batch = mm_budget \
                        .max_items_per_batch_by_modality[dummy_modality]
2725
2726
2727
2728
2729
2730
2731
2732
2733

                    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,
                    )
2734

2735
2736
2737
2738
2739
                    # Create dummy batch of multimodal inputs.
                    batched_dummy_mm_inputs = self._get_mm_dummy_batch(
                        dummy_modality,
                        max_mm_items_per_batch,
                    )
2740

2741
2742
2743
2744
                    # Run multimodal encoder.
                    dummy_encoder_outputs = \
                        self.model.get_multimodal_embeddings(
                        **batched_dummy_mm_inputs)
2745

2746
2747
2748
2749
                    sanity_check_mm_encoder_outputs(
                        dummy_encoder_outputs,
                        expected_num_items=max_mm_items_per_batch,
                    )
2750

2751
2752
2753
                    # Cache the dummy encoder outputs.
                    self.encoder_cache["tmp"] = dict(
                        enumerate(dummy_encoder_outputs))
2754

2755
        # Add `is_profile` here to pre-allocate communication buffers
2756
        hidden_states, last_hidden_states \
2757
            = self._dummy_run(self.max_num_tokens, is_profile=True)
2758
        if get_pp_group().is_last_rank:
2759
2760
2761
2762
            if self.is_pooling_model:
                output = self._dummy_pooler_run(hidden_states)
            else:
                output = self._dummy_sampler_run(last_hidden_states)
2763
        else:
2764
            output = None
2765
        self._sync_device()
2766
        del hidden_states, output
2767
        self.encoder_cache.clear()
2768
        gc.collect()
2769
2770

    def capture_model(self) -> None:
2771
        if self.compilation_config.cudagraph_mode == CUDAGraphMode.NONE:
2772
            logger.warning(
2773
                "Skipping CUDA graph capture. To turn on CUDA graph capture, "
2774
                "ensure `cudagraph_mode` was not manually set to `NONE`")
2775
            return
2776
2777
        else:
            self.initialize_cudagraph_capture()
2778

2779
2780
        compilation_counter.num_gpu_runner_capture_triggers += 1

2781
2782
2783
        start_time = time.perf_counter()
        start_free_gpu_memory = torch.cuda.mem_get_info()[0]

2784
2785
2786
2787
2788
2789
2790
2791
2792
2793
2794
2795
2796
2797
2798
        @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()

2799
2800
2801
        # 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.
2802
        set_cudagraph_capturing_enabled(True)
2803
        with freeze_gc(), graph_capture(device=self.device):
2804
2805
2806
2807
2808
2809
2810
2811
2812
2813
2814
2815
2816
2817
2818
2819
2820
2821
2822
2823
2824
2825
2826
2827
2828
2829
2830
2831
2832
2833
            cudagraph_mode = self.compilation_config.cudagraph_mode
            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)

            # Capture full cudagraph for uniform decode batches if we have
            # dont already have full mixed prefill-decode cudagraphs
            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
2834
        # we may do lazy capturing in future that still allows capturing
2835
2836
        # after here.
        set_cudagraph_capturing_enabled(False)
2837
2838
2839
2840
2841
2842
2843
2844

        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))
2845

2846
2847
2848
2849
2850
2851
2852
2853
2854
2855
2856
2857
2858
2859
2860
2861
2862
2863
2864
2865
2866
2867
2868
2869
2870
2871
2872
2873
2874
    def _capture_cudagraphs(self, compilation_cases: list[int],
                            cudagraph_runtime_mode: CUDAGraphMode,
                            uniform_decode: bool):
        assert cudagraph_runtime_mode != CUDAGraphMode.NONE and \
            cudagraph_runtime_mode in [CUDAGraphMode.FULL,
                                        CUDAGraphMode.PIECEWISE]

        # 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))
        # We skip EPLB here since we don't want to record dummy metrics
        for num_tokens in compilation_cases:
            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)
                self._dummy_run(num_tokens,
                                cudagraph_runtime_mode=CUDAGraphMode.NONE,
                                force_attention=force_attention,
                                uniform_decode=uniform_decode,
2875
2876
                                skip_eplb=True,
                                remove_lora=False)
2877
2878
2879
            self._dummy_run(num_tokens,
                            cudagraph_runtime_mode=cudagraph_runtime_mode,
                            uniform_decode=uniform_decode,
2880
2881
2882
                            skip_eplb=True,
                            remove_lora=False)
        self.maybe_remove_all_loras(self.lora_config)
2883

2884
2885
2886
2887
    def initialize_attn_backend(self, kv_cache_config: KVCacheConfig) -> None:
        """
        Initialize the attention backends and attention metadata builders.
        """
2888
2889
2890
2891
2892
2893
        assert len(self.attn_groups) == 0, \
            "Attention backends are already initialized"

        def get_attn_backends_for_layers(
                layer_names: list[str]
        ) -> dict[type[AttentionBackend], list[str]]:
2894
2895
2896
            layers = get_layers_from_vllm_config(self.vllm_config,
                                                 AttentionLayerBase,
                                                 layer_names)
2897
2898
            attn_backends = {}
            attn_backend_layers = defaultdict(list)
2899
            # Dedupe based on full class name; this is a bit safer than
2900
2901
2902
2903
2904
            # 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.
            for layer_name in layer_names:
2905
                attn_backend = layers[layer_name].get_attn_backend()
2906
2907
2908
2909
2910
2911
2912

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

2913
2914
2915
2916
2917
2918
2919
2920
2921
2922
2923
2924
2925
2926
2927
2928
2929
2930
2931
2932
2933
2934
2935
2936
2937
2938
2939
                key = attn_backend.full_cls_name()
                attn_backends[key] = attn_backend
                attn_backend_layers[key].append(layer_name)
            return {
                attn_backends[k]: v
                for k, v in attn_backend_layers.items()
            }

        def create_attn_groups(
            attn_backends_map: dict[AttentionBackend, list[str]],
            kv_cache_spec: KVCacheSpec,
        ) -> list[AttentionGroup]:
            attn_groups: list[AttentionGroup] = []
            for attn_backend, layer_names in attn_backends_map.items():
                attn_metadata_builder_i = attn_backend.get_builder_cls()(
                    kv_cache_spec,
                    layer_names,
                    self.vllm_config,
                    self.device,
                )
                attn_group = AttentionGroup(attn_backend,
                                            attn_metadata_builder_i,
                                            layer_names)
                attn_groups.append(attn_group)
            return attn_groups

        for kv_cache_group_spec in kv_cache_config.kv_cache_groups:
2940
            kv_cache_spec = kv_cache_group_spec.kv_cache_spec
2941
2942
            attn_backends = get_attn_backends_for_layers(
                kv_cache_group_spec.layer_names)
2943
2944
            self.attn_groups.append(
                create_attn_groups(attn_backends, kv_cache_spec))
2945

co63oc's avatar
co63oc committed
2946
        # Calculate reorder batch threshold (if needed)
2947
2948
        self.calculate_reorder_batch_threshold()

2949
2950
2951
2952
2953
2954
2955
2956
2957
2958
2959
2960
2961
2962
2963
2964
2965
2966
2967
2968
2969
2970
2971
2972
2973
2974
2975
2976
2977
2978
2979
2980
2981
2982
2983
2984
2985
2986
2987
2988
2989
2990
2991
2992
2993
2994
2995
2996
2997
2998
2999
3000
3001
3002
3003
3004
3005
3006
3007
3008
3009
3010
3011
3012
3013
3014
3015
3016
3017
    def initialize_cudagraph_capture(self) -> None:
        min_cg_support = AttentionCGSupport.ALWAYS
        min_cg_builder_name = None

        for attn_group in self._attn_group_iterator():
            builder = attn_group.metadata_builder
            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)

        # 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)

3018
3019
3020
3021
3022
    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)
        """
3023
3024
3025
        for group in self._attn_group_iterator():
            attn_metadata_builder_i = group.metadata_builder

3026
3027
3028
3029
3030
3031
3032
3033
3034
3035
3036
3037
3038
3039
3040
3041
            # 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

3042
3043
3044
3045
3046
3047
3048
3049
3050
3051
3052
3053
3054
3055
3056
3057
3058
3059
3060
3061
3062
3063
3064
3065
3066
3067
3068
    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,
                max_model_len=self.max_model_len,
                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,
3069
                is_spec_decode=bool(self.vllm_config.speculative_config),
3070
3071
                logitsprocs=self.input_batch.logitsprocs,
                is_pooling_model=self.is_pooling_model,
3072
3073
            )

3074
3075
    def _allocate_kv_cache_tensors(
            self, kv_cache_config: KVCacheConfig) -> dict[str, torch.Tensor]:
3076
        """
3077
3078
3079
        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.

3080
        Args:
3081
            kv_cache_config: The KV cache config
3082
        Returns:
3083
            dict[str, torch.Tensor]: A map between layer names to their
3084
3085
3086
3087
3088
3089
3090
3091
3092
3093
3094
3095
            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:
3096
3097
3098
3099
            for layer_name in group.layer_names:
                if layer_name in self.runner_only_attn_layers:
                    continue
                layer_names.add(layer_name)
3100
3101
3102
3103
        assert layer_names == set(kv_cache_raw_tensors.keys(
        )), "Some layers are not correctly initialized"
        return kv_cache_raw_tensors

3104
3105
3106
3107
3108
3109
3110
3111
3112
3113
3114
3115
    def _attn_group_iterator(self) -> Iterator[AttentionGroup]:
        return itertools.chain.from_iterable(self.attn_groups)

    def _kv_cache_spec_attn_group_iterator(
            self) -> Iterator[tuple[KVCacheSpec, AttentionGroup]]:
        if not self.kv_cache_config.kv_cache_groups:
            return
        for kv_cache_spec_id, attn_groups in enumerate(self.attn_groups):
            for attn_group in attn_groups:
                yield self.kv_cache_config.kv_cache_groups[
                    kv_cache_spec_id].kv_cache_spec, attn_group

3116
3117
3118
3119
3120
    def _reshape_kv_cache_tensors(
        self,
        kv_cache_config: KVCacheConfig,
        kv_cache_raw_tensors: dict[str, torch.Tensor],
    ) -> dict[str, torch.Tensor]:
3121
        """
3122
        Reshape the KV cache tensors to the desired shape and dtype.
3123

3124
        Args:
3125
3126
            kv_cache_config: The KV cache config
            kv_cache_raw_tensors: The KV cache buffer of each layer, with
3127
                correct size but uninitialized shape.
3128
        Returns:
3129
            Dict[str, torch.Tensor]: A map between layer names to their
3130
3131
            corresponding memory buffer for KV cache.
        """
3132
        kv_caches: dict[str, torch.Tensor] = {}
3133
        has_attn, has_mamba = False, False
3134
3135
3136
        for kv_cache_spec, group in self._kv_cache_spec_attn_group_iterator():
            attn_backend = group.backend
            for layer_name in group.layer_names:
3137
3138
                if layer_name in self.runner_only_attn_layers:
                    continue
3139
3140
3141
3142
                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)
3143
                if isinstance(kv_cache_spec, AttentionSpec):
3144
                    has_attn = True
3145
                    kv_cache_shape = attn_backend.get_kv_cache_shape(
3146
3147
3148
                        num_blocks, kv_cache_spec.block_size,
                        kv_cache_spec.num_kv_heads, kv_cache_spec.head_size)
                    dtype = kv_cache_spec.dtype
3149
                    try:
3150
3151
                        kv_cache_stride_order = \
                            attn_backend.get_kv_cache_stride_order()
3152
3153
3154
3155
3156
3157
3158
3159
3160
3161
3162
3163
3164
3165
3166
3167
3168
                        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))
                    ]
3169
3170
3171
                    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
3172
                elif isinstance(kv_cache_spec, MambaSpec):
3173
                    has_mamba = True
Chen Zhang's avatar
Chen Zhang committed
3174
3175
                    raw_tensor = kv_cache_raw_tensors[layer_name]
                    state_tensors = []
3176
3177
3178
3179
3180
3181
                    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
3182
                        target_shape = (num_blocks, *shape)
3183
3184
                        stride = torch.empty(target_shape).stride()
                        target_stride = (num_element_per_page, *stride[1:])
3185
                        assert storage_offset_bytes % dtype_size == 0
3186
3187
3188
3189
                        tensor = torch.as_strided(
                            raw_tensor.view(dtype),
                            size=target_shape,
                            stride=target_stride,
3190
                            storage_offset=storage_offset_bytes // dtype_size,
3191
                        )
Chen Zhang's avatar
Chen Zhang committed
3192
                        state_tensors.append(tensor)
3193
                        storage_offset_bytes += stride[0] * dtype_size
3194
3195

                    kv_caches[layer_name] = state_tensors
3196
                else:
3197
                    raise NotImplementedError
3198
3199

        if has_attn and has_mamba:
3200
            self._update_hybrid_attention_mamba_layout(kv_caches)
3201

3202
3203
        return kv_caches

3204
3205
    def _update_hybrid_attention_mamba_layout(
            self, kv_caches: dict[str, torch.Tensor]) -> None:
3206
        """
3207
3208
        Update the layout of attention layers from (2, num_blocks, ...) to
        (num_blocks, 2, ...).
3209
3210

        Args:
3211
            kv_caches: The KV cache buffer of each layer.
3212
3213
        """

3214
3215
        for kv_cache_spec, group in self._kv_cache_spec_attn_group_iterator():
            for layer_name in group.layer_names:
3216
3217
3218
3219
3220
3221
3222
3223
3224
3225
3226
                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:]))
3227

3228
3229
3230
3231
3232
3233
3234
3235
    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:
3236
            Dict[str, torch.Tensor]: A map between layer names to their
3237
3238
3239
3240
3241
3242
3243
            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)
3244

3245
3246
3247
3248
3249
3250
3251
3252
3253
3254
3255
3256
3257
3258
3259
3260
3261
3262
3263
3264
3265
3266
3267
3268
3269
3270
3271
3272
3273
3274
3275
3276
        # 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]

        bind_kv_cache(kv_caches,
                      self.compilation_config.static_forward_context,
                      self.kv_caches)
        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.
3277
3278
3279
3280
3281
3282
3283
3284
            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
3285

3286
3287
3288
3289
3290
3291
3292
    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
        """
3293
        kv_cache_config = deepcopy(kv_cache_config)
3294
3295
        self.kv_cache_config = kv_cache_config
        self.may_reinitialize_input_batch(kv_cache_config)
3296
        self.may_add_encoder_only_layers_to_kv_cache_config()
3297
        self.maybe_add_kv_sharing_layers_to_kv_cache_groups(kv_cache_config)
3298
3299
3300
        self.initialize_attn_backend(kv_cache_config)
        kv_caches = self.initialize_kv_cache_tensors(kv_cache_config)

3301
3302
3303
3304
3305
3306
        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
3307
3308
        if has_kv_transfer_group():
            get_kv_transfer_group().register_kv_caches(kv_caches)
3309
3310
3311
            if self.device.type == 'xpu':
                get_kv_transfer_group().set_host_xfer_buffer_ops(
                    copy_kv_blocks)
Robert Shaw's avatar
Robert Shaw committed
3312

3313
3314
3315
3316
3317
3318
        if self.dcp_world_size > 1:
            assert self.attn_groups[0][0].backend is FlashMLABackend, (
                "DCP only support flashmla now."
                "For a mla backend want to enable DCP, it is mandatory that the"
                "corresponding decode attn kernel return the softmax lse.")

3319
3320
3321
3322
3323
3324
3325
3326
3327
3328
3329
3330
3331
3332
3333
3334
3335
3336
3337
3338
3339
3340
3341
3342
3343
3344
3345
    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:
                attn_spec = EncoderOnlyAttentionSpec(
                    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))

3346
    def get_kv_cache_spec(self) -> dict[str, KVCacheSpec]:
3347
        """
3348
        Generates the KVCacheSpec by parsing the kv cache format from each
3349
3350
        Attention module in the static forward context.
        Returns:
3351
            KVCacheSpec: A dictionary mapping layer names to their KV cache
3352
3353
3354
3355
            format. Layers that do not need KV cache are not included.
        """

        block_size = self.vllm_config.cache_config.block_size
3356
        use_mla = self.vllm_config.model_config.use_mla
3357
        kv_cache_spec: dict[str, KVCacheSpec] = {}
Chen Zhang's avatar
Chen Zhang committed
3358
3359
        attn_layers = get_layers_from_vllm_config(self.vllm_config, Attention)
        for layer_name, attn_module in attn_layers.items():
3360
3361
3362
3363
3364
3365
3366
3367
3368
3369
3370
3371
            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

3372
            # TODO: Support other attention modules, e.g., cross-attention
3373
3374
            # TODO(lucas): move the attention specs into the model layers like
            # the attention backends
3375
            if attn_module.attn_type == AttentionType.DECODER:
3376
3377
3378
3379
3380
3381
3382
3383
                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)
3384
3385
                elif self.attention_chunk_size is not None \
                        and isinstance(attn_module, ChunkedLocalAttention):
3386
                    kv_cache_spec[layer_name] = ChunkedLocalAttentionSpec(
3387
3388
3389
3390
3391
                        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,
3392
                        use_mla=use_mla)
3393
3394
3395
3396
3397
3398
3399
                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)
3400
3401
3402
3403
3404
3405
3406
3407
3408
3409
            elif attn_module.attn_type in (AttentionType.ENCODER,
                                           AttentionType.ENCODER_ONLY):
                # encoder-only attention does not need KV cache.
                continue
            elif attn_module.attn_type == AttentionType.ENCODER_DECODER:
                raise NotImplementedError
            else:
                raise ValueError(
                    f"Unknown attention type: {attn_module.attn_type}")

3410
        mamba_layers = get_layers_from_vllm_config(self.vllm_config, MambaBase)
Chen Zhang's avatar
Chen Zhang committed
3411
3412
3413
3414
3415
3416
3417
3418
        if len(mamba_layers) > 0:
            if self.vllm_config.speculative_config is not None:
                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
3419

3420
3421
            page_size_padded = (
                self.vllm_config.cache_config.mamba_page_size_padded)
3422

Chen Zhang's avatar
Chen Zhang committed
3423
3424
3425
3426
3427
            # 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(),
3428
                    dtypes=mamba_module.get_state_dtype(),
3429
                    block_size=max_model_len,
3430
3431
                    page_size_padded=page_size_padded,
                    mamba_type=mamba_module.mamba_type)
3432

3433
        return kv_cache_spec
3434
3435
3436
3437
3438
3439
3440
3441
3442
3443
3444
3445
3446
3447
3448

    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()