gpu_model_runner.py 138 KB
Newer Older
1
# SPDX-License-Identifier: Apache-2.0
2
# SPDX-FileCopyrightText: Copyright contributors to the vLLM project
3

4
import dataclasses
5
import gc
6
import time
7
from contextlib import contextmanager
8
from typing import TYPE_CHECKING, Any, Optional, Union, cast
9
10
11
12
13

import numpy as np
import torch
import torch.distributed
import torch.nn as nn
14
from tqdm import tqdm
15

16
import vllm.envs as envs
17
from vllm.attention import AttentionType, get_attn_backend
18
from vllm.attention.backends.abstract import AttentionBackend
19
from vllm.attention.layer import Attention
20
from vllm.compilation.counter import compilation_counter
21
from vllm.config import (CompilationLevel, VllmConfig,
22
                         get_layers_from_vllm_config, update_config)
23
from vllm.distributed.eplb.eplb_state import EplbState
24
25
from vllm.distributed.kv_transfer import (get_kv_transfer_group,
                                          has_kv_transfer_group)
26
from vllm.distributed.parallel_state import (
27
    get_pp_group, get_tp_group, graph_capture, is_global_first_rank,
28
    prepare_communication_buffer_for_model)
29
from vllm.forward_context import DPMetadata, set_forward_context
30
from vllm.logger import init_logger
31
from vllm.model_executor.layers.mamba.mamba_mixer2 import MambaBase
32
from vllm.model_executor.layers.rotary_embedding import MRotaryEmbedding
33
from vllm.model_executor.model_loader import TensorizerLoader, get_model_loader
34
35
36
37
from vllm.model_executor.models.interfaces import (is_mixture_of_experts,
                                                   supports_transcription)
from vllm.model_executor.models.interfaces_base import (
    VllmModelForPooling, is_pooling_model, is_text_generation_model)
38
from vllm.multimodal import MULTIMODAL_REGISTRY
39
40
from vllm.multimodal.inputs import (BatchedTensorInputs, MultiModalKwargs,
                                    PlaceholderRange)
41
from vllm.multimodal.utils import group_mm_inputs_by_modality
42
from vllm.pooling_params import PoolingParams
43
from vllm.sampling_params import SamplingType
44
from vllm.sequence import IntermediateTensors, PoolerOutput
45
from vllm.tasks import GenerationTask, PoolingTask, SupportedTask
46
from vllm.utils import (STR_DTYPE_TO_TORCH_DTYPE, DeviceMemoryProfiler,
47
                        GiB_bytes, LazyLoader, check_use_alibi, get_dtype_size,
48
                        is_pin_memory_available, round_up, supports_dynamo)
49
from vllm.v1.attention.backends.mamba_selectors import get_mamba_attn_backend
50
from vllm.v1.attention.backends.utils import (
51
    AttentionCGSupport, AttentionMetadataBuilder, CommonAttentionMetadata,
52
    make_kv_sharing_fast_prefill_attention_metadata,
53
54
    make_local_attention_virtual_batches,
    reorder_batch_to_split_decodes_and_prefills)
55
56
57
58
from vllm.v1.kv_cache_interface import (AttentionSpec,
                                        ChunkedLocalAttentionSpec,
                                        FullAttentionSpec, KVCacheConfig,
                                        KVCacheSpec, MambaSpec,
59
                                        SlidingWindowSpec)
60
61
from vllm.v1.outputs import (EMPTY_MODEL_RUNNER_OUTPUT, LogprobsTensors,
                             ModelRunnerOutput)
62
from vllm.v1.pool.metadata import PoolingMetadata
63
from vllm.v1.sample.metadata import SamplingMetadata
64
from vllm.v1.sample.rejection_sampler import RejectionSampler
65
from vllm.v1.sample.sampler import Sampler
66
from vllm.v1.spec_decode.eagle import EagleProposer
67
from vllm.v1.spec_decode.medusa import MedusaProposer
68
from vllm.v1.spec_decode.metadata import SpecDecodeMetadata
69
from vllm.v1.spec_decode.ngram_proposer import NgramProposer
70
from vllm.v1.worker.gpu_input_batch import CachedRequestState, InputBatch
71
from vllm.v1.worker.kv_connector_model_runner_mixin import (
72
    KVConnectorModelRunnerMixin, KVConnectorOutput)
73
from vllm.v1.worker.lora_model_runner_mixin import LoRAModelRunnerMixin
74

75
from ..sample.logits_processor import LogitsProcessorManager
76
from .utils import (MultiModalBudget, bind_kv_cache, gather_mm_placeholders,
77
                    initialize_kv_cache_for_kv_sharing,
78
                    sanity_check_mm_encoder_outputs, scatter_mm_placeholders)
79

80
if TYPE_CHECKING:
81
    import xgrammar as xgr
82
    import xgrammar.kernels.apply_token_bitmask_inplace_torch_compile as xgr_torch_compile  # noqa: E501
83

84
    from vllm.model_executor.model_loader.tensorizer import TensorizerConfig
85
    from vllm.v1.core.sched.output import SchedulerOutput
86
87
else:
    xgr = LazyLoader("xgr", globals(), "xgrammar")
88
89
90
    xgr_torch_compile = LazyLoader(
        "xgr_torch_compile", globals(),
        "xgrammar.kernels.apply_token_bitmask_inplace_torch_compile")
91
92
93
94

logger = init_logger(__name__)


95
class GPUModelRunner(LoRAModelRunnerMixin, KVConnectorModelRunnerMixin):
96
97
98

    def __init__(
        self,
99
        vllm_config: VllmConfig,
100
        device: torch.device,
101
    ):
102
103
104
        self.vllm_config = vllm_config
        self.model_config = vllm_config.model_config
        self.cache_config = vllm_config.cache_config
105
        self.compilation_config = vllm_config.compilation_config
106
107
108
109
110
111
        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
112

113
114
115
116
        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))

117
118
119
120
        model_config = self.model_config
        cache_config = self.cache_config
        scheduler_config = self.scheduler_config
        parallel_config = self.parallel_config
121
        self.device = device
122
123
124
125
126
127
128
129
        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]

130
        self.is_multimodal_model = model_config.is_multimodal_model
131
        self.is_pooling_model = model_config.pooler_config is not None
132
        self.is_encoder_only_model = False
133
134
        self.is_multimodal_raw_input_supported = (
            model_config.is_multimodal_raw_input_supported)
135
136
        self.max_model_len = model_config.max_model_len
        self.max_num_tokens = scheduler_config.max_num_batched_tokens
137
        self.max_num_reqs = scheduler_config.max_num_seqs
138
139

        # Model-related.
140
141
        self.num_query_heads = model_config.get_num_attention_heads(
            parallel_config)
142
        self.hidden_size = model_config.get_hidden_size()
143
        self.attention_chunk_size = model_config.attention_chunk_size
144

145
        self.cascade_attn_enabled = not self.model_config.disable_cascade_attn
146

147
        # Multi-modal data support
148
        self.mm_registry = MULTIMODAL_REGISTRY
149
        self.uses_mrope = model_config.uses_mrope
150

151
        # Sampler
152
        self.sampler = Sampler(logprobs_mode=self.model_config.logprobs_mode)
153

154
155
156
157
158
159
160
        self.eplb_state: Optional[EplbState] = None
        """
        State of the expert parallelism load balancer.

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

161
        # Lazy initializations
162
        # self.model: nn.Module  # Set after load_model
163
        # Initialize in initialize_kv_cache
164
        self.kv_caches: list[torch.Tensor] = []
165
166
        self.attn_metadata_builders: list[AttentionMetadataBuilder] = []
        self.attn_backends: list[type[AttentionBackend]] = []
167
168
        # self.kv_cache_config: KVCacheConfig

169
        # req_id -> (input_id -> encoder_output)
170
        self.encoder_cache: dict[str, dict[int, torch.Tensor]] = {}
171

172
        self.use_aux_hidden_state_outputs = False
173
174
175
176
177
178
179
180
181
182
183
184
185
186
187
188
189
190
191
192
        # 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()
193

194
        # Request states.
195
        self.requests: dict[str, CachedRequestState] = {}
196

197
198
199
200
201
202
203
204
205
        # 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.
206
207
208
209
210
211
        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,
212
            vocab_size=self.model_config.get_vocab_size(),
213
            block_sizes=[self.cache_config.block_size],
214
            is_spec_decode=bool(self.vllm_config.speculative_config),
215
        )
216

217
218
219
220
221
        self.use_cuda_graph = (
            self.vllm_config.compilation_config.level
            == CompilationLevel.PIECEWISE
            and self.vllm_config.compilation_config.use_cudagraph
            and not self.model_config.enforce_eager)
222
        # TODO(woosuk): Provide an option to tune the max cudagraph batch size.
223
224
225
226
        # The convention is different.
        # self.cudagraph_batch_sizes sorts in ascending order.
        # The batch sizes in the config are in descending order.
        self.cudagraph_batch_sizes = list(
227
228
229
            reversed(self.compilation_config.cudagraph_capture_sizes))

        self.full_cuda_graph = self.compilation_config.full_cuda_graph
230

231
        # Cache the device properties.
232
        self._init_device_properties()
233

234
235
236
237
        # Persistent buffers for CUDA graphs.
        self.input_ids = torch.zeros(self.max_num_tokens,
                                     dtype=torch.int32,
                                     device=self.device)
238
239
240
        self.positions = torch.zeros(self.max_num_tokens,
                                     dtype=torch.int64,
                                     device=self.device)
241
242
243
244
245
246
247
248
249
250
        self.query_start_loc = torch.zeros(self.max_num_reqs + 1,
                                           dtype=torch.int32,
                                           device=self.device)
        self.seq_lens = torch.zeros(self.max_num_reqs,
                                    dtype=torch.int32,
                                    device=self.device)
        self.slot_mapping = torch.zeros(self.max_num_tokens,
                                        dtype=torch.int64,
                                        device=self.device)

251
252
        # None in the first PP rank. The rest are set after load_model.
        self.intermediate_tensors: Optional[IntermediateTensors] = None
253
254

        # Only relevant for models using M-RoPE (e.g, Qwen2-VL)
255
        if self.uses_mrope:
Roger Wang's avatar
Roger Wang committed
256
257
258
259
            # 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
260
261
262
263
264
265

            # 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
Roger Wang's avatar
Roger Wang committed
266
            self.mrope_positions = torch.zeros((3, self.max_num_tokens + 1),
267
268
                                               dtype=torch.int64,
                                               device=self.device)
Roger Wang's avatar
Roger Wang committed
269
270
271
272
273
            self.mrope_positions_cpu = torch.zeros(
                (3, self.max_num_tokens + 1),
                dtype=torch.int64,
                device="cpu",
                pin_memory=self.pin_memory)
274
            self.mrope_positions_np = self.mrope_positions_cpu.numpy()
275

276
277
278
        # Only relevant for models using ALiBi (e.g, MPT)
        self.use_alibi = check_use_alibi(model_config)

279
280
281
282
        self.inputs_embeds = torch.zeros(
            (self.max_num_tokens, self.hidden_size),
            dtype=self.dtype,
            device=self.device)
283

284
        # OPTIMIZATION: Cache the tensors rather than creating them every step.
285
        # Keep in int64 to avoid overflow with long context
286
        self.arange_np = np.arange(max(self.max_num_reqs + 1,
287
288
                                       self.max_model_len,
                                       self.max_num_tokens),
289
                                   dtype=np.int64)
290
291
292
        # NOTE(woosuk): These tensors are "stateless", i.e., they are literally
        # a faster version of creating a new tensor every time. Thus, we should
        # not make any assumptions about the values in these tensors.
293
294
295
296
297
298
299
300
301
302
303
304
305
306
        self.input_ids_cpu = torch.zeros(self.max_num_tokens,
                                         dtype=torch.int32,
                                         device="cpu",
                                         pin_memory=self.pin_memory)
        self.positions_cpu = torch.zeros(self.max_num_tokens,
                                         dtype=torch.int64,
                                         device="cpu",
                                         pin_memory=self.pin_memory)
        self.positions_np = self.positions_cpu.numpy()
        self.query_start_loc_cpu = torch.zeros(self.max_num_reqs + 1,
                                               dtype=torch.int32,
                                               device="cpu",
                                               pin_memory=self.pin_memory)
        self.query_start_loc_np = self.query_start_loc_cpu.numpy()
307
308
309
310
311
        self.seq_lens_cpu = torch.zeros(self.max_num_reqs,
                                        dtype=torch.int32,
                                        device="cpu",
                                        pin_memory=self.pin_memory)
        self.seq_lens_np = self.seq_lens_cpu.numpy()
312

313
314
315
316
317
        # 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] = {}
318
319
320
321
322
323
        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)
324

325
326
327
328
329
330
331
332
        self.mm_budget = (MultiModalBudget(
            self.model_config,
            self.scheduler_config,
            self.mm_registry,
            max_model_len=self.max_model_len,
            max_num_reqs=self.max_num_reqs,
        ) if self.is_multimodal_model else None)

333
334
        self.reorder_batch_threshold: Optional[int] = None

335
    def _may_reorder_batch(self, scheduler_output: "SchedulerOutput") -> None:
336
337
        """
        Update the order of requests in the batch based on the attention
338
        backend's needs. For example, some attention backends (namely MLA) may
339
340
341
342
343
344
        want to separate requests based on if the attention computation will be
        compute-bound or memory-bound.

        Args:
            scheduler_output: The scheduler output.
        """
345
346
347
348
349
350
351
352
        # 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

353
354
355
356
357
        if self.reorder_batch_threshold is not None:
            reorder_batch_to_split_decodes_and_prefills(
                self.input_batch,
                scheduler_output,
                decode_threshold=self.reorder_batch_threshold)
358

359
360
361
362
363
364
365
366
367
368
369
    # 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()

370
    def _update_states(self, scheduler_output: "SchedulerOutput") -> None:
371
372
373
374
375
376
        """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.

377
378
        The SamplingMetadata is updated and copied to the GPU if there is a
        new/resumed/paused/finished request in the batch.
379
380
        """
        # Remove finished requests from the cached states.
381
382
        for req_id in scheduler_output.finished_req_ids:
            self.requests.pop(req_id, None)
383
            self.encoder_cache.pop(req_id, None)
384
385
386
387
388
389
390
        # 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:
391
            self.input_batch.remove_request(req_id)
392
393
394
395
396
397
398
399

        # Free the cached encoder outputs.
        for req_id, input_id in scheduler_output.free_encoder_input_ids:
            encoder_outputs = self.encoder_cache.get(req_id)
            if encoder_outputs is not None:
                encoder_outputs.pop(input_id, None)
                if not encoder_outputs:
                    self.encoder_cache.pop(req_id, None)
400

401
402
403
404
405
406
407
408
409
410
411
412
413
        # 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:
414
            self.input_batch.remove_request(req_id)
415

416
        req_ids_to_add: list[str] = []
417
        # Add new requests to the cached states.
418
419
420
        for new_req_data in scheduler_output.scheduled_new_reqs:
            req_id = new_req_data.req_id
            sampling_params = new_req_data.sampling_params
421
            pooling_params = new_req_data.pooling_params
422

423
424
            if sampling_params and \
                sampling_params.sampling_type == SamplingType.RANDOM_SEED:
425
426
427
428
429
                generator = torch.Generator(device=self.device)
                generator.manual_seed(sampling_params.seed)
            else:
                generator = None

430
            if pooling_params:
431
                assert (task := pooling_params.task) is not None, (
432
433
434
                    "You did not set `task` in the API")

                model = cast(VllmModelForPooling, self.model)
435
                to_update = model.pooler.get_pooling_updates(task)
436
437
                to_update.apply(pooling_params)

438
439
            self.requests[req_id] = CachedRequestState(
                req_id=req_id,
440
441
442
                prompt_token_ids=new_req_data.prompt_token_ids,
                mm_inputs=new_req_data.mm_inputs,
                mm_positions=new_req_data.mm_positions,
443
                sampling_params=sampling_params,
444
                pooling_params=pooling_params,
445
                generator=generator,
446
447
                block_ids=new_req_data.block_ids,
                num_computed_tokens=new_req_data.num_computed_tokens,
448
                output_token_ids=[],
449
                lora_request=new_req_data.lora_request,
450
            )
451
452

            # Only relevant for models using M-RoPE (e.g, Qwen2-VL)
453
            if self.uses_mrope:
454
455
                image_grid_thw = []
                video_grid_thw = []
Roger Wang's avatar
Roger Wang committed
456
                second_per_grid_ts = []
457
458
                audio_feature_lengths = []
                use_audio_in_video = False
459
460
461
462
463
464
465
                for mm_input in self.requests[req_id].mm_inputs:
                    if mm_input.get("image_grid_thw") is not None:
                        image_grid_thw.extend(
                            mm_input["image_grid_thw"].tolist())
                    if mm_input.get("video_grid_thw") is not None:
                        video_grid_thw.extend(
                            mm_input["video_grid_thw"].tolist())
Roger Wang's avatar
Roger Wang committed
466
467
468
                    if mm_input.get("second_per_grid_ts") is not None:
                        second_per_grid_ts.extend(
                            mm_input["second_per_grid_ts"])
469
470
471
472
473
                    if mm_input.get("audio_feature_lengths") is not None:
                        audio_feature_lengths.extend(
                            mm_input["audio_feature_lengths"])
                    if mm_input.get("use_audio_in_video") is True:
                        use_audio_in_video = True
474
475
476
477
478
479
480

                hf_config = self.model_config.hf_config

                self.requests[req_id].mrope_positions, \
                    self.requests[req_id].mrope_position_delta = \
                    MRotaryEmbedding.get_input_positions_tensor(
                        self.requests[req_id].prompt_token_ids,
Roger Wang's avatar
Roger Wang committed
481
                        hf_config=hf_config,
482
483
                        image_grid_thw=image_grid_thw,
                        video_grid_thw=video_grid_thw,
Roger Wang's avatar
Roger Wang committed
484
                        second_per_grid_ts=second_per_grid_ts,
485
486
                        audio_feature_lengths=audio_feature_lengths,
                        use_audio_in_video=use_audio_in_video,
487
488
                    )

489
490
            req_ids_to_add.append(req_id)

491
        # Update the states of the running/resumed requests.
492
        is_last_rank = get_pp_group().is_last_rank
493
494
        req_data = scheduler_output.scheduled_cached_reqs
        for i, req_id in enumerate(req_data.req_ids):
495
            req_state = self.requests[req_id]
496
497
498
            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]
499

500
            # Update the cached states.
501
            req_state.num_computed_tokens = num_computed_tokens
502
503
504
505
506
507
508
509
510
511
512
513
514
515
516
517
518

            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:])

519
            # Update the block IDs.
520
            if not resumed_from_preemption:
521
                # Append the new blocks to the existing block IDs.
522
523
524
                for block_ids, new_ids in zip(req_state.block_ids,
                                              new_block_ids):
                    block_ids.extend(new_ids)
525
526
527
            else:
                # The request is resumed from preemption.
                # Replace the existing block IDs with the new ones.
528
                req_state.block_ids = new_block_ids
529
530
531
532
533
534
535
536
537
538
539

            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.
                req_ids_to_add.append(req_id)
                continue

            # Update the persistent batch.
            self.input_batch.num_computed_tokens_cpu[req_index] = (
540
                num_computed_tokens)
541
            self.input_batch.block_table.append_row(new_block_ids, req_index)
542
543
544
545
546
547
548

            # 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)
549
                self.input_batch.token_ids_cpu[
550
551
552
553
554
                    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
555

556
557
558
559
560
561
562
563
564
565
566
567
            # 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

568
569
570
571
        # Add the new or resumed requests to the persistent batch.
        # The smaller empty indices are filled first.
        for req_id in req_ids_to_add:
            req_state = self.requests[req_id]
572
            self.input_batch.add_request(req_state)
573

574
575
576
577
578
579
        # 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()
580

581
    def _extract_mm_kwargs(
582
        self,
583
584
585
        scheduler_output: "SchedulerOutput",
    ) -> BatchedTensorInputs:
        if self.is_multimodal_raw_input_supported:  # noqa: SIM102
586
            if scheduler_output:
587
                multi_modal_kwargs_list = list[MultiModalKwargs]()
588
589
590
591
592
593
                for req in scheduler_output.scheduled_new_reqs:
                    req_mm_inputs = req.mm_inputs
                    if not isinstance(req_mm_inputs, list):
                        req_mm_inputs = list(req_mm_inputs)
                    multi_modal_kwargs_list.extend(req_mm_inputs)

594
                return MultiModalKwargs.batch(multi_modal_kwargs_list)
595

596
597
598
599
600
601
602
603
604
605
606
607
        return {}

    def _dummy_mm_kwargs(self, num_seqs: int) -> BatchedTensorInputs:
        if self.is_multimodal_raw_input_supported:
            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)

        return {}
608

609
610
611
612
613
614
615
616
617
618
619
620
621
622
623
624
625
626
627
628
    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

629
    def _prepare_inputs(
630
631
        self,
        scheduler_output: "SchedulerOutput",
632
633
634
    ) -> tuple[dict[str,
                    Any], bool, torch.Tensor, Optional[SpecDecodeMetadata],
               np.ndarray, Optional[CommonAttentionMetadata]]:
635
636
637
638
639
640
641
        """
        :return: tuple[
            attn_metadata: layer-to-attention_metadata mapping,
            attention_cuda_graphs: whether attention can run in cudagraph
            logits_indices, spec_decode_metadata
        ]
        """
642
643
644
645
646
647
648
        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.
649
        self.input_batch.block_table.commit_block_table(num_reqs)
650
651

        # Get the number of scheduled tokens for each request.
652
653
654
655
        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)
656
657
658

        # Get request indices.
        # E.g., [2, 5, 3] -> [0, 0, 1, 1, 1, 1, 1, 2, 2, 2]
659
660
        req_indices = np.repeat(self.arange_np[:num_reqs],
                                num_scheduled_tokens)
661

662
663
664
665
        # 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)
666
667

        # Get positions.
668
        positions_np = self.positions_np[:total_num_scheduled_tokens]
669
670
671
672
        np.add(self.input_batch.num_computed_tokens_cpu[req_indices],
               arange,
               out=positions_np)

673
674
        # Calculate M-RoPE positions.
        # Only relevant for models using M-RoPE (e.g, Qwen2-VL)
675
        if self.uses_mrope:
676
677
            self._calc_mrope_positions(scheduler_output)

678
679
680
681
        # 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.
682
683
        token_indices = (positions_np +
                         req_indices * self.input_batch.token_ids_cpu.shape[1])
684

685
686
687
688
        # 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(),
689
                           0,
690
691
                           torch.from_numpy(token_indices),
                           out=self.input_ids_cpu[:total_num_scheduled_tokens])
692

693
694
695
696
        self.input_batch.block_table.compute_slot_mapping(
            req_indices, positions_np)
        self.input_batch.block_table.commit_slot_mapping(
            total_num_scheduled_tokens)
697
698

        # Prepare the attention metadata.
699
        self.query_start_loc_np[0] = 0
700
        self.query_start_loc_np[1:num_reqs + 1] = cu_num_tokens
701

702
703
704
        self.seq_lens_np[:num_reqs] = (
            self.input_batch.num_computed_tokens_cpu[:num_reqs] +
            num_scheduled_tokens)
705
706
707
708

        # Copy the tensors to the GPU.
        self.input_ids[:total_num_scheduled_tokens].copy_(
            self.input_ids_cpu[:total_num_scheduled_tokens], non_blocking=True)
709
        if self.uses_mrope:
710
711
712
713
714
715
716
717
718
            # Only relevant for models using M-RoPE (e.g, Qwen2-VL)
            self.mrope_positions[:, :total_num_scheduled_tokens].copy_(
                self.mrope_positions_cpu[:, :total_num_scheduled_tokens],
                non_blocking=True)
        else:
            # Common case (1D positions)
            self.positions[:total_num_scheduled_tokens].copy_(
                self.positions_cpu[:total_num_scheduled_tokens],
                non_blocking=True)
719

720
721
722
723
724
        self.query_start_loc[:num_reqs + 1].copy_(
            self.query_start_loc_cpu[:num_reqs + 1], non_blocking=True)
        self.seq_lens[:num_reqs].copy_(self.seq_lens_cpu[:num_reqs],
                                       non_blocking=True)

725
        # Fill unused with 0 for full cuda graph mode.
726
        self.seq_lens[num_reqs:].fill_(0)
727
728
729
730
        # Note: pad query_start_loc to be non-decreasing, as kernels
        # like FlashAttention requires that
        self.query_start_loc[num_reqs + 1:].fill_(
            self.query_start_loc_cpu[num_reqs].item())
731
732

        query_start_loc = self.query_start_loc[:num_reqs + 1]
733
734

        spec_decode_common_attn_metadata = None
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
780
781
782
783
784
        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:
            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.use_cuda_graph
                    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]
            )

785
        attn_metadata: dict[str, Any] = {}
786
787
788
789
790
791
792
793
794
795
796
797
798
799
800

        # Prepare encoder attention metadata separately
        # (encoder layers are not in KV cache groups)
        if self.is_encoder_only_model:
            common_attn_metadata, encoder_attn_metadata = \
                self._build_encoder_only_attn_metadata(
                scheduler_output)

            # Add encoder attention metadata for all encoder layers
            attention_layers = get_layers_from_vllm_config(
                self.vllm_config, Attention)
            for layer_name, attn_module in attention_layers.items():
                if attn_module.attn_type == AttentionType.ENCODER_ONLY:
                    attn_metadata[layer_name] = encoder_attn_metadata

801
802
803
804
805
        # 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):

806
807
808
            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]
809
810
811
812
813

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

814
815
816
817
818
819
820
821
822
823
824
825
            common_attn_metadata = CommonAttentionMetadata(
                query_start_loc=self.query_start_loc[:num_reqs + 1],
                query_start_loc_cpu=self.query_start_loc_cpu[:num_reqs + 1],
                seq_lens=self.seq_lens[:num_reqs],
                seq_lens_cpu=self.seq_lens_cpu[:num_reqs],
                num_computed_tokens_cpu=self.input_batch.
                num_computed_tokens_cpu_tensor[:num_reqs],
                num_reqs=num_reqs,
                num_actual_tokens=total_num_scheduled_tokens,
                max_query_len=max_num_scheduled_tokens,
                block_table_tensor=blk_table_tensor,
                slot_mapping=slot_mapping,
826
                causal=True,
827
828
829
830
831
832
            )

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

833
834
835
836
837
838
            if isinstance(kv_cache_group_spec.kv_cache_spec,
                          ChunkedLocalAttentionSpec):
                common_attn_metadata = make_local_attention_virtual_batches(
                    kv_cache_group_spec.kv_cache_spec.attention_chunk_size,
                    common_attn_metadata, self.cache_config.block_size)

839
840
            # Prepare for cascade attention if enabled & beneficial.
            common_prefix_len = 0
841
            builder = self.attn_metadata_builders[kv_cache_group_id]
842
843
844
            if self.cascade_attn_enabled:
                common_prefix_len = self._compute_cascade_attn_prefix_len(
                    num_scheduled_tokens,
845
846
847
                    scheduler_output.
                    num_common_prefix_blocks[kv_cache_group_id],
                    kv_cache_group_spec.kv_cache_spec,
848
                    builder,
849
                )
850

851
852
853
854
855
            attn_metadata_i = (builder.build(
                common_prefix_len=common_prefix_len,
                common_attn_metadata=common_attn_metadata,
            ))

856
857
858
859
860
861
862
863
864
865
866
867
868
869
870
871
            fast_prefill_metadata = attn_metadata_i
            if (self.cache_config.kv_sharing_fast_prefill
                    and self.kv_sharing_fast_prefill_eligible_layers):
                # Dynamically create a a dataclass type that inherits
                # from attention metadata type but includes additional
                # fields logits_indices_padded and num_logits_indices
                # which are required for prefill truncation
                fast_prefill_metadata_type = (
                    make_kv_sharing_fast_prefill_attention_metadata(
                        metadata_cls=type(attn_metadata_i), ))
                fast_prefill_metadata = fast_prefill_metadata_type(
                    **dataclasses.asdict(attn_metadata_i),
                    logits_indices_padded=logits_indices_padded,
                    num_logits_indices=logits_indices.size(0),
                )

872
            for layer_name in kv_cache_group_spec.layer_names:
873
874
875
876
877
                if (self.cache_config.kv_sharing_fast_prefill and layer_name
                        in self.kv_sharing_fast_prefill_eligible_layers):
                    attn_metadata[layer_name] = fast_prefill_metadata
                    continue

878
                attn_metadata[layer_name] = attn_metadata_i
879

880
881
882
883
884
885
886
887
888
889
890
891
892
893
894
895
896
897
898
899
900
901
902
903
904
            # Hack for now to fix chunked local attention + no hybrid kv cache
            # manager we can remove this once
            # https://github.com/vllm-project/vllm/pull/21588
            # is merged (i.e. properly handle different attention backends for
            # the same kv_cache_spec)
            if self.attention_chunk_size is not None \
                    and self.scheduler_config.disable_hybrid_kv_cache_manager:
                if not hasattr(self, "local_attention_layers"):
                    self.local_attention_layers = []
                    attn_layers = get_layers_from_vllm_config(
                        self.vllm_config, Attention)
                    for layer_name, attn_module in attn_layers.items():
                        if attn_module.use_irope:
                            self.local_attention_layers.append(layer_name)

                local_attn_metadata_i = (builder.build(
                    common_prefix_len=0,
                    common_attn_metadata=make_local_attention_virtual_batches(
                        self.attention_chunk_size, common_attn_metadata,
                        self.cache_config.block_size),
                ))

                for layer_name in self.local_attention_layers:
                    attn_metadata[layer_name] = local_attn_metadata_i

905
906
907
908
        attention_cuda_graphs = all(
            b.can_run_in_cudagraph(common_attn_metadata)
            for b in self.attn_metadata_builders)

909
910
911
912
        # Hot-Swap lora model
        if self.lora_config:
            self.set_active_loras(self.input_batch, num_scheduled_tokens)

913
        return (attn_metadata, attention_cuda_graphs, logits_indices,
914
915
                spec_decode_metadata, num_scheduled_tokens,
                spec_decode_common_attn_metadata)
916

917
918
919
920
    def _compute_cascade_attn_prefix_len(
        self,
        num_scheduled_tokens: np.ndarray,
        num_common_prefix_blocks: int,
921
922
        kv_cache_spec: KVCacheSpec,
        attn_metadata_builder: AttentionMetadataBuilder,
923
924
925
926
927
928
929
930
931
932
933
934
935
936
937
938
939
940
    ) -> 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.
        """
941
        common_prefix_len = num_common_prefix_blocks * kv_cache_spec.block_size
942
943
944
945
946
947
948
949
950
951
952
953
954
955
956
957
958
959
960
961
962
963
964
965
966
967
968
969
970
971
972
973
974
975
976
977
978
        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]
979
        # Request 3's num_computed_tokens: 3 (i.e., [A, B, C])
980
981
982
983
984
985
986
987
988
989
        # 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.
990
991
992
993
994
        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))
995
996
997
998
        use_local_attention = (
            isinstance(kv_cache_spec, ChunkedLocalAttentionSpec)
            or (isinstance(kv_cache_spec, FullAttentionSpec)
                and kv_cache_spec.attention_chunk_size is not None))
999
1000
        assert isinstance(kv_cache_spec, AttentionSpec)
        use_cascade = attn_metadata_builder.use_cascade_attention(
1001
1002
1003
            common_prefix_len=common_prefix_len,
            query_lens=num_scheduled_tokens,
            num_query_heads=self.num_query_heads,
1004
            num_kv_heads=kv_cache_spec.num_kv_heads,
1005
            use_alibi=self.use_alibi,
1006
            use_sliding_window=use_sliding_window,
1007
            use_local_attention=use_local_attention,
1008
1009
1010
1011
            num_sms=self.num_sms,
        )
        return common_prefix_len if use_cascade else 0

1012
1013
    def _calc_mrope_positions(self, scheduler_output: "SchedulerOutput"):
        mrope_pos_ptr = 0
1014
        for index, req_id in enumerate(self.input_batch.req_ids):
1015
1016
1017
1018
1019
1020
1021
1022
1023
1024
1025
1026
1027
1028
1029
1030
1031
1032
1033
1034
1035
1036
1037
1038
1039
1040
1041
1042
1043
1044
1045
1046
1047
1048
1049
1050
1051
            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

                self.mrope_positions_cpu[:, dst_start:dst_end] = \
                    req.mrope_positions[:,src_start:src_end]

                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

1052
1053
1054
1055
1056
1057
1058
                MRotaryEmbedding.get_next_input_positions_tensor(
                    out=self.mrope_positions_np,
                    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,
                )
1059
1060
1061

                mrope_pos_ptr += completion_part_len

1062
1063
    def _calc_spec_decode_metadata(
        self,
1064
1065
1066
1067
1068
1069
1070
1071
1072
1073
1074
1075
1076
1077
1078
1079
        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
1080
1081
1082
1083
1084
1085

        # 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]
1086
1087
        logits_indices = np.repeat(
            cu_num_scheduled_tokens - num_sampled_tokens, num_sampled_tokens)
1088
        # Step 3. [0, 1, 2, 3, 103, 104, 105, 106, 206, 207, 208]
1089
1090
1091
1092
1093
1094
        logits_indices += arange

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

        # Compute the draft logits indices.
1095
1096
1097
1098
        # 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)
1099
1100
1101
1102
1103
1104
1105
1106
1107
1108
1109
1110
1111
1112
        # [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(
1113
1114
            self.device, non_blocking=True)

1115
1116
1117
1118
1119
1120
1121
1122
1123
1124
1125
1126
1127
1128
1129
        # Compute the draft token ids.
        # draft_token_indices:      [  1,   2,   3, 105, 106, 208]
        draft_token_ids = self.input_ids[logits_indices]
        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

1130
    def _execute_mm_encoder(self, scheduler_output: "SchedulerOutput"):
1131
1132
1133
1134
1135
        scheduled_encoder_inputs = scheduler_output.scheduled_encoder_inputs
        if not scheduled_encoder_inputs:
            return

        # Batch the multi-modal inputs.
1136
1137
        mm_inputs = list[MultiModalKwargs]()
        req_ids_pos = list[tuple[str, int, PlaceholderRange]]()
1138
1139
        for req_id, encoder_input_ids in scheduled_encoder_inputs.items():
            req_state = self.requests[req_id]
1140
1141
1142
1143
1144

            for mm_input_id in encoder_input_ids:
                mm_inputs.append(req_state.mm_inputs[mm_input_id])
                req_ids_pos.append(
                    (req_id, mm_input_id, req_state.mm_positions[mm_input_id]))
1145
1146
1147
1148
1149
1150
1151
1152
1153
1154
1155
1156

        # 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.
        grouped_mm_inputs_list = group_mm_inputs_by_modality(mm_inputs)

        encoder_outputs = []
        for grouped_mm_inputs in grouped_mm_inputs_list:
1157
1158
            batched_mm_inputs = MultiModalKwargs.batch(
                grouped_mm_inputs, pin_memory=self.pin_memory)
1159
1160
1161
1162
            batched_mm_inputs = MultiModalKwargs.as_kwargs(
                batched_mm_inputs,
                device=self.device,
            )
1163
1164
1165
1166
1167
1168
1169
1170
1171
1172
1173

            # 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(
                **batched_mm_inputs)

1174
1175
1176
1177
1178
            sanity_check_mm_encoder_outputs(
                curr_group_outputs,
                expected_num_items=len(grouped_mm_inputs),
            )

1179
1180
            for output in curr_group_outputs:
                encoder_outputs.append(output)
1181
1182

        # Cache the encoder outputs.
1183
1184
1185
1186
        for (req_id, input_id, pos_info), output in zip(
                req_ids_pos,
                encoder_outputs,
        ):
1187
1188
1189
            if req_id not in self.encoder_cache:
                self.encoder_cache[req_id] = {}

1190
1191
1192
1193
1194
1195
            self.encoder_cache[req_id][input_id] = scatter_mm_placeholders(
                output,
                is_embed=pos_info.is_embed,
            )

    def _gather_mm_embeddings(
1196
1197
        self,
        scheduler_output: "SchedulerOutput",
1198
        shift_computed_tokens: int = 0,
1199
    ) -> list[torch.Tensor]:
1200
        mm_embeds: list[torch.Tensor] = []
1201
        for req_id in self.input_batch.req_ids:
1202
1203
1204
            num_scheduled_tokens = scheduler_output.num_scheduled_tokens[
                req_id]
            req_state = self.requests[req_id]
1205
1206
            num_computed_tokens = \
                req_state.num_computed_tokens + shift_computed_tokens
1207
1208
            mm_positions = req_state.mm_positions
            for i, pos_info in enumerate(mm_positions):
1209
1210
                start_pos = pos_info.offset
                num_encoder_tokens = pos_info.length
1211
1212
1213
1214
1215
1216
1217
1218
1219
1220
1221
1222
1223
1224
1225
1226
1227
1228
1229
1230
1231

                # 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,
                    num_encoder_tokens)
                assert start_idx < end_idx
                assert req_id in self.encoder_cache
                assert i in self.encoder_cache[req_id]
                encoder_output = self.encoder_cache[req_id][i]
1232
1233
1234
1235
1236
1237
1238
1239
1240
1241

                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
1242

1243
1244
1245
    def get_model(self) -> nn.Module:
        return self.model

1246
1247
1248
1249
1250
1251
1252
1253
1254
1255
1256
1257
1258
1259
1260
    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

1261
1262
1263
1264
1265
    def get_supported_pooling_tasks(self) -> list[PoolingTask]:
        model = self.get_model()
        if not is_pooling_model(model):
            return []

1266
        return list(model.pooler.get_supported_tasks())
1267

1268
1269
1270
1271
1272
1273
1274
1275
1276
1277
    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)

1278
1279
1280
1281
1282
1283
1284
1285
1286
    def apply_grammar_bitmask(
        self,
        scheduler_output: "SchedulerOutput",
        logits: torch.Tensor,
    ):
        grammar_bitmask = scheduler_output.grammar_bitmask
        if grammar_bitmask is None:
            return

1287
1288
1289
1290
1291
1292
1293
1294
1295
        # 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.
1296
        struct_out_req_batch_indices: dict[str, int] = {}
1297
1298
1299
1300
1301
1302
1303
1304
1305
1306
1307
1308
1309
1310
1311
1312
1313
1314
1315
1316
1317
1318
1319
1320
1321
1322
1323
1324
1325
        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.
        sorted_bitmask = np.zeros_like(grammar_bitmask,
                                       shape=(logits.shape[0],
                                              grammar_bitmask.shape[1]))
        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
1326

1327
1328
        # Serialization of np.ndarray is much more efficient than a tensor,
        # so we receive it in that format.
1329
1330
        grammar_bitmask = torch.from_numpy(grammar_bitmask)

1331
1332
1333
1334
        # Force use of the torch.compile implementation from xgrammar to work
        # around issues with the Triton kernel in concurrent structured output
        # scenarios. See PR #19565 and issues #19493, #18376 for details.
        xgr_torch_compile.apply_token_bitmask_inplace_torch_compile(
1335
1336
            logits,
            grammar_bitmask.to(self.device, non_blocking=True),
1337
            indices=out_indices,
1338
1339
        )

1340
1341
1342
1343
1344
1345
1346
    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
1347
        enabled_sp = self.compilation_config.pass_config. \
1348
1349
1350
1351
1352
1353
1354
1355
1356
1357
1358
1359
1360
            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():
1361
                is_scattered = k == "residual" and is_residual_scattered
1362
1363
1364
1365
1366
1367
1368
1369
1370
1371
1372
1373
                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()
        })

1374
1375
1376
1377
1378
1379
1380
1381
1382
1383
1384
1385
1386
1387
1388
1389
1390
1391
    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
        assert is_mixture_of_experts(self.model)
        self.eplb_state.step(
            self.model,
            is_dummy,
            is_profile,
            log_stats=self.parallel_config.eplb_log_balancedness,
        )

1392
1393
    def get_dp_padding(self,
                       num_tokens: int) -> tuple[int, Optional[torch.Tensor]]:
1394
1395
        dp_size = self.vllm_config.parallel_config.data_parallel_size
        dp_rank = self.vllm_config.parallel_config.data_parallel_rank
1396
1397
1398
1399
1400
1401
1402
1403
1404

        # 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:
1405
            # Early exit.
1406
            return 0, None
1407
1408
1409
1410

        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()
1411
1412
1413
1414
1415
        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
1416

1417
1418
1419
1420
1421
    def _pool(
        self,
        hidden_states: torch.Tensor,
        num_scheduled_tokens: int,
        num_scheduled_tokens_np: np.ndarray,
1422
        kv_connector_output: Optional[KVConnectorOutput],
1423
1424
1425
1426
1427
1428
1429
1430
1431
1432
1433
1434
1435
1436
1437
1438
1439
1440
1441
1442
1443
1444
1445
1446
1447
1448
1449
1450
1451
1452
1453
1454
1455
1456
    ) -> 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"

        extracted_hidden_states = list(
            torch.split(hidden_states[:num_scheduled_tokens],
                        num_scheduled_tokens_np.tolist()))

        pooling_metadata = self.input_batch.pooling_metadata

        raw_pooler_output = self.model.pooler(
            hidden_states=extracted_hidden_states,
            pooling_metadata=pooling_metadata)

        pooler_output: list[Optional[torch.Tensor]] = []
        seq_lens = self.seq_lens[:self.input_batch.num_reqs]
        for raw_output, seq_len, prompt_len in zip(
                raw_pooler_output, seq_lens, pooling_metadata.prompt_lens):

            if seq_len == prompt_len:
                pooler_output.append(raw_output.data.cpu())
            else:
                pooler_output.append(None)

        return ModelRunnerOutput(
            req_ids=self.input_batch.req_ids,
            req_id_to_index=self.input_batch.req_id_to_index,
            sampled_token_ids=[],
            spec_token_ids=None,
            logprobs=None,
            prompt_logprobs_dict={},
            pooler_output=pooler_output,
1457
            kv_connector_output=kv_connector_output,
1458
1459
        )

1460
1461
1462
1463
    @torch.inference_mode()
    def execute_model(
        self,
        scheduler_output: "SchedulerOutput",
1464
        intermediate_tensors: Optional[IntermediateTensors] = None,
1465
    ) -> Union[ModelRunnerOutput, IntermediateTensors]:
1466
        self._update_states(scheduler_output)
1467
        if not scheduler_output.total_num_scheduled_tokens:
1468
1469
1470
            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
1471

1472
1473
            return self.kv_connector_no_forward(scheduler_output,
                                                self.vllm_config)
1474
1475

        # Prepare the decoder inputs.
1476
        (attn_metadata, attention_cuda_graphs, logits_indices,
1477
1478
1479
         spec_decode_metadata, num_scheduled_tokens_np,
         spec_decode_common_attn_metadata) = (
             self._prepare_inputs(scheduler_output))
1480

1481
1482
1483
1484
1485
        num_scheduled_tokens = scheduler_output.total_num_scheduled_tokens
        if (self.use_cuda_graph
                and num_scheduled_tokens <= self.cudagraph_batch_sizes[-1]):
            # Use piecewise CUDA graphs.
            # Add padding to the batch size.
1486
            num_input_tokens = self.vllm_config.pad_for_cudagraph(
1487
1488
1489
                num_scheduled_tokens)
        else:
            # Eager mode.
1490
1491
1492
            # Pad tokens to multiple of tensor_parallel_size when
            # enabled collective fusion for SP
            tp_size = self.vllm_config.parallel_config.tensor_parallel_size
1493
            if self.compilation_config.pass_config. \
1494
1495
1496
1497
                enable_sequence_parallelism and tp_size > 1:
                num_input_tokens = round_up(num_scheduled_tokens, tp_size)
            else:
                num_input_tokens = num_scheduled_tokens
1498

1499
        # Padding for DP
1500
1501
        num_pad, num_tokens_across_dp = self.get_dp_padding(num_input_tokens)
        num_input_tokens += num_pad
1502

1503
1504
1505
1506
1507
1508
1509
1510
1511
        # _prepare_inputs may reorder the batch, so we must gather multi
        # modal outputs after that to ensure the correct order
        if self.is_multimodal_model:
            # Run the multimodal encoder if any.
            self._execute_mm_encoder(scheduler_output)
            mm_embeds = self._gather_mm_embeddings(scheduler_output)
        else:
            mm_embeds = []

1512
        if self.is_multimodal_model and get_pp_group().is_first_rank:
1513
1514
1515
            # 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.
1516
1517
            inputs_embeds_scheduled = self.model.get_input_embeddings(
                input_ids=self.input_ids[:num_scheduled_tokens],
1518
1519
                multimodal_embeddings=mm_embeds or None,
            )
1520

1521
            # TODO(woosuk): Avoid the copy. Optimize.
1522
1523
1524
            self.inputs_embeds[:num_scheduled_tokens].copy_(
                inputs_embeds_scheduled)

1525
            input_ids = None
1526
1527
            inputs_embeds = self.inputs_embeds[:num_input_tokens]
            model_mm_kwargs = self._extract_mm_kwargs(scheduler_output)
1528
        else:
1529
1530
1531
1532
1533
1534
            # 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.
            input_ids = self.input_ids[:num_input_tokens]
            inputs_embeds = None
1535
            model_mm_kwargs = {}
1536
1537
1538
1539
        if self.uses_mrope:
            positions = self.mrope_positions[:, :num_input_tokens]
        else:
            positions = self.positions[:num_input_tokens]
1540

1541
1542
1543
        if get_pp_group().is_first_rank:
            intermediate_tensors = None
        else:
1544
1545
            intermediate_tensors = self.sync_and_slice_intermediate_tensors(
                num_input_tokens, intermediate_tensors, True)
1546

1547
1548
1549
1550
1551
        # Some attention backends only support CUDA Graphs in pure decode.
        # If attention doesn't support CUDA Graphs for this batch, but we
        # compiled with full CUDA graphs, we have to skip them entirely.
        skip_cuda_graphs = self.full_cuda_graph and not attention_cuda_graphs

1552
        # Run the model.
1553
        # Use persistent buffers for CUDA graphs.
1554
1555
1556
1557
1558
1559
        with set_forward_context(
                attn_metadata,
                self.vllm_config,
                num_tokens=num_input_tokens,
                num_tokens_across_dp=num_tokens_across_dp,
                skip_cuda_graphs=skip_cuda_graphs,
1560
1561
        ), self.maybe_get_kv_connector_output(
                scheduler_output) as kv_connector_output:
Robert Shaw's avatar
Robert Shaw committed
1562
1563

            model_output = self.model(
1564
                input_ids=input_ids,
1565
                positions=positions,
1566
                intermediate_tensors=intermediate_tensors,
1567
                inputs_embeds=inputs_embeds,
1568
                **MultiModalKwargs.as_kwargs(
1569
                    model_mm_kwargs,
1570
1571
                    device=self.device,
                ),
1572
            )
1573
1574

        if self.use_aux_hidden_state_outputs:
Robert Shaw's avatar
Robert Shaw committed
1575
            hidden_states, aux_hidden_states = model_output
1576
        else:
Robert Shaw's avatar
Robert Shaw committed
1577
            hidden_states = model_output
1578
1579
            aux_hidden_states = None

1580
1581
1582
1583
1584
1585
1586
        # 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
1587
        if not get_pp_group().is_last_rank:
1588
            # For mid-pipeline stages, return the hidden states.
1589
            assert isinstance(hidden_states, IntermediateTensors)
1590
            if not broadcast_pp_output:
1591
                hidden_states.kv_connector_output = kv_connector_output
1592
1593
1594
1595
1596
                return hidden_states
            get_pp_group().send_tensor_dict(hidden_states.tensors,
                                            all_gather_group=get_tp_group())
            logits = None
        else:
1597
1598
            if self.input_batch.pooling_params:
                return self._pool(hidden_states, num_scheduled_tokens,
1599
                                  num_scheduled_tokens_np, kv_connector_output)
1600

1601
1602
1603
1604
1605
1606
1607
1608
1609
1610
            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"]
1611

1612
1613
1614
1615
        # Apply structured output bitmasks if present
        if scheduler_output.grammar_bitmask is not None:
            self.apply_grammar_bitmask(scheduler_output, logits)

1616
        # Sample the next token and get logprobs if needed.
1617
        sampling_metadata = self.input_batch.sampling_metadata
1618
        if spec_decode_metadata is None:
1619
            sampler_output = self.sampler(
1620
1621
1622
1623
                logits=logits,
                sampling_metadata=sampling_metadata,
            )
        else:
1624
1625
1626
1627
            # 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.
1628
            assert logits is not None
1629
            bonus_logits = logits[spec_decode_metadata.bonus_logits_indices]
1630
            sampler_output = self.sampler(
1631
                logits=bonus_logits,
1632
1633
1634
                sampling_metadata=sampling_metadata,
            )
            bonus_token_ids = sampler_output.sampled_token_ids
1635

1636
1637
1638
            # 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.
1639
            target_logits = logits[spec_decode_metadata.target_logits_indices]
1640
            output_token_ids = self.rejection_sampler(
1641
                spec_decode_metadata,
1642
                None,  # draft_probs
1643
                target_logits,
1644
                bonus_token_ids,
1645
1646
                sampling_metadata,
            )
1647
            sampler_output.sampled_token_ids = output_token_ids
1648

1649
1650
1651
1652
        num_nans_in_logits = {}
        if envs.VLLM_COMPUTE_NANS_IN_LOGITS:
            num_nans_in_logits = self._get_nans_in_logits(logits)

1653
1654
        # TODO(woosuk): The following loop can be slow since it iterates over
        # the requests one by one. Optimize.
1655
1656
        discard_sampled_tokens_req_indices = []
        for i, req_id in enumerate(self.input_batch.req_ids):
1657
1658
1659
            req_state = self.requests[req_id]
            seq_len = (req_state.num_computed_tokens +
                       scheduler_output.num_scheduled_tokens[req_id])
1660
            if seq_len < req_state.num_tokens:
1661
                # Ignore the sampled token for partial prefills.
1662
                # Rewind the generator state as if the token was not sampled.
1663
                # This relies on cuda-specific torch-internal impl details
1664
1665
1666
1667
1668
1669
                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)
1670

1671
1672
        # NOTE: GPU -> CPU Sync happens here.
        # Move as many CPU operations as possible before this sync point.
1673
1674
1675
1676
1677
1678
        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(
1679
            hidden_states[:num_scheduled_tokens],
1680
1681
1682
            scheduler_output,
        )

1683
        # Get the valid generated tokens.
1684
1685
1686
        sampled_token_ids = sampler_output.sampled_token_ids
        max_gen_len = sampled_token_ids.shape[-1]
        if max_gen_len == 1:
1687
            # No spec decode tokens.
1688
1689
            valid_sampled_token_ids = sampled_token_ids.tolist()
        else:
1690
            # Includes spec decode tokens.
1691
            valid_sampled_token_ids = self.rejection_sampler.parse_output(
1692
1693
1694
                sampled_token_ids,
                self.input_batch.vocab_size,
            )
1695
1696
1697
        # Mask out the sampled tokens that should not be sampled.
        for i in discard_sampled_tokens_req_indices:
            valid_sampled_token_ids[i].clear()
1698

1699
1700
1701
1702
1703
1704
1705
1706
1707
1708
1709
1710
1711
1712
1713
1714
1715
1716
1717
1718
1719
1720
1721
1722
        # 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.
        for req_idx, sampled_ids in enumerate(valid_sampled_token_ids):
            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
            req_id = self.input_batch.req_ids[req_idx]
            req_state = self.requests[req_id]
            req_state.output_token_ids.extend(sampled_ids)

1723
        if not self.speculative_config:
1724
            # Speculative decoding is not enabled.
1725
            spec_token_ids = None
1726
        else:
1727
            assert spec_decode_common_attn_metadata is not None
1728
1729
1730
1731
1732
1733
1734
1735
            spec_token_ids = self.propose_draft_token_ids(
                scheduler_output,
                valid_sampled_token_ids,
                sampling_metadata,
                hidden_states,
                sample_hidden_states,
                aux_hidden_states,
                spec_decode_metadata,
1736
                spec_decode_common_attn_metadata,
1737
1738
1739
1740
1741
1742
1743
1744
1745
1746
1747
1748
            )

        self.eplb_step()

        return ModelRunnerOutput(
            req_ids=self.input_batch.req_ids,
            req_id_to_index=self.input_batch.req_id_to_index,
            sampled_token_ids=valid_sampled_token_ids,
            spec_token_ids=spec_token_ids,
            logprobs=logprobs_lists,
            prompt_logprobs_dict=prompt_logprobs_dict,
            pooler_output=[],
1749
            kv_connector_output=kv_connector_output,
1750
1751
1752
1753
1754
1755
1756
1757
1758
1759
1760
1761
            num_nans_in_logits=num_nans_in_logits,
        )

    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],
1762
        common_attn_metadata: CommonAttentionMetadata,
1763
1764
1765
    ) -> list[list[int]]:
        num_scheduled_tokens = scheduler_output.total_num_scheduled_tokens
        if self.speculative_config.method == "ngram":
1766
            assert isinstance(self.drafter, NgramProposer)
1767
1768
            spec_token_ids = self.propose_ngram_draft_token_ids(
                sampled_token_ids)
1769
1770
        elif self.speculative_config.method == "medusa":
            assert isinstance(self.drafter, MedusaProposer)
1771
1772
            if sample_hidden_states.shape[0] == len(sampled_token_ids):
                # The input to the target model does not include draft tokens.
1773
1774
1775
1776
1777
1778
                hidden_states = sample_hidden_states
            else:
                indices = []
                offset = 0
                for num_draft, tokens in zip(
                        spec_decode_metadata.num_draft_tokens,
1779
                        sampled_token_ids):
1780
1781
                    indices.append(offset + len(tokens) - 1)
                    offset += num_draft + 1
1782
                indices = torch.tensor(indices, device=self.device)
1783
1784
1785
1786
1787
1788
                hidden_states = sample_hidden_states[indices]

            spec_token_ids = self.drafter.propose(
                target_hidden_states=hidden_states,
                sampling_metadata=sampling_metadata,
            )
1789
        elif self.speculative_config.use_eagle():
1790
1791
1792
            assert isinstance(self.drafter, EagleProposer)
            # TODO(woosuk): Refactor the loop.
            next_token_ids: list[int] = []
1793
            for i, token_ids in enumerate(sampled_token_ids):
1794
1795
1796
1797
1798
1799
1800
1801
1802
1803
1804
1805
                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.
                    req_id = self.input_batch.req_ids[i]
                    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)
1806
1807
1808
            next_token_ids = torch.tensor(next_token_ids,
                                          dtype=torch.int32,
                                          device=self.device)
Jiayi Yao's avatar
Jiayi Yao committed
1809

1810
1811
1812
            if spec_decode_metadata is None:
                # input_ids can be None for multimodal models.
                target_token_ids = self.input_ids[:num_scheduled_tokens]
1813
1814
                # TODO(woosuk): Support M-RoPE.
                target_positions = self.positions[:num_scheduled_tokens]
1815
                if self.use_aux_hidden_state_outputs:
1816
1817
1818
                    target_hidden_states = torch.cat(
                        [h[:num_scheduled_tokens] for h in aux_hidden_states],
                        dim=-1)
1819
1820
                else:
                    target_hidden_states = hidden_states[:num_scheduled_tokens]
1821
1822
1823
1824
            else:
                # TODO(woosuk): Refactor this.
                num_draft_tokens = spec_decode_metadata.num_draft_tokens
                num_rejected_tokens = [
1825
                    n + 1 - len(sampled_token_ids[i]) if n > 0 else 0
1826
1827
                    for i, n in enumerate(num_draft_tokens)
                ]
1828
1829
1830
1831
1832
1833
                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)

1834
                target_token_ids = self.input_ids[token_indices]
1835
1836
                # TODO(woosuk): Support M-RoPE.
                target_positions = self.positions[token_indices]
1837
                if self.use_aux_hidden_state_outputs:
1838
1839
                    target_hidden_states = torch.cat(
                        [h[token_indices] for h in aux_hidden_states], dim=-1)
1840
1841
                else:
                    target_hidden_states = hidden_states[token_indices]
1842
1843
1844
1845
1846
            mm_embeds = None
            if self.is_multimodal_model:
                mm_embeds = self._gather_mm_embeddings(scheduler_output,
                                                       shift_computed_tokens=1)

1847
            draft_token_ids = self.drafter.propose(
1848
1849
1850
1851
1852
                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,
1853
                common_attn_metadata=common_attn_metadata,
1854
                mm_embeds=mm_embeds,
1855
1856
            )
            spec_token_ids = draft_token_ids.tolist()
1857
        return spec_token_ids
1858

1859
    def propose_ngram_draft_token_ids(
1860
        self,
1861
1862
        sampled_token_ids: list[list[int]],
    ) -> list[list[int]]:
1863
        # TODO(woosuk): Optimize.
1864
        draft_token_ids: list[list[int]] = []
1865
1866
1867
        for i, sampled_ids in enumerate(sampled_token_ids):
            num_sampled_ids = len(sampled_ids)
            if not num_sampled_ids:
1868
1869
1870
1871
                # Skip speculative decoding.
                draft_token_ids.append([])
                continue

1872
1873
            # Skip requests that require sampling parameters that are not
            # supported with speculative decoding.
1874
            req_id = self.input_batch.req_ids[i]
1875
            if req_id in self.input_batch.spec_decode_unsupported_reqs:
1876
1877
1878
                draft_token_ids.append([])
                continue

1879
1880
            num_tokens = self.input_batch.num_tokens_no_spec[i]
            if num_tokens >= self.max_model_len:
1881
1882
1883
1884
                # Skip requests that have already reached the max model length.
                draft_token_ids.append([])
                continue

1885
            drafter_output = self.drafter.propose(
1886
                self.input_batch.token_ids_cpu[i, :num_tokens])
1887
1888
1889
1890
1891
1892
            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

1893
1894
1895
1896
1897
1898
1899
1900
1901
1902
    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)

1903
1904
1905
1906
1907
    def load_model(self, eep_scale_up: bool = False) -> None:
        """
        Args:
            eep_scale_up: the model loading is for elastic EP scale up.
        """
1908
        logger.info("Starting to load model %s...", self.model_config.model)
1909
1910
1911
1912
1913
1914
1915
1916
1917
1918
1919
1920
1921
1922
1923
1924
1925
1926
1927
1928
1929
1930
1931
1932
1933
1934
1935
1936
        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]
            self.parallel_config.num_redundant_experts = (
                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

1937
        with DeviceMemoryProfiler() as m:
1938
            time_before_load = time.perf_counter()
1939
            model_loader = get_model_loader(self.load_config)
1940
1941
1942
            logger.info("Loading model from scratch...")
            self.model = model_loader.load_model(
                vllm_config=self.vllm_config, model_config=self.model_config)
1943
1944
1945
1946
1947
1948
            if self.lora_config:
                self.model = self.load_lora_model(self.model,
                                                  self.model_config,
                                                  self.scheduler_config,
                                                  self.lora_config,
                                                  self.device)
1949
1950
1951
            if hasattr(self, "drafter"):
                logger.info("Loading drafter model...")
                self.drafter.load_model(self.model)
1952
1953
1954
            if self.use_aux_hidden_state_outputs:
                self.model.set_aux_hidden_state_layers(
                    self.model.get_eagle3_aux_hidden_state_layers())
1955
            time_after_load = time.perf_counter()
1956
        self.model_memory_usage = m.consumed_memory
1957
1958
        logger.info("Model loading took %.4f GiB and %.6f seconds",
                    self.model_memory_usage / GiB_bytes,
1959
                    time_after_load - time_before_load)
1960
        prepare_communication_buffer_for_model(self.model)
1961

1962
1963
1964
1965
1966
1967
1968
1969
        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,
1970
1971
1972
                global_expert_load,
                old_global_expert_indices,
                rank_mapping,
1973
1974
            )

1975
1976
1977
1978
1979
1980
1981
1982
1983
1984
1985
        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)

1986
1987
1988
1989
1990
1991
1992
    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...")
        model_loader.load_weights(self.model, model_config=self.model_config)

1993
1994
1995
1996
1997
1998
1999
    def save_tensorized_model(
        self,
        tensorizer_config: "TensorizerConfig",
    ) -> None:
        TensorizerLoader.save_model(
            self.model,
            tensorizer_config=tensorizer_config,
2000
            model_config=self.model_config,
2001
2002
        )

2003
2004
2005
2006
    def _get_prompt_logprobs_dict(
        self,
        hidden_states: torch.Tensor,
        scheduler_output: "SchedulerOutput",
2007
    ) -> dict[str, Optional[LogprobsTensors]]:
2008
2009
2010
2011
        num_prompt_logprobs_dict = self.input_batch.num_prompt_logprobs
        if not num_prompt_logprobs_dict:
            return {}

2012
        in_progress_dict = self.input_batch.in_progress_prompt_logprobs_cpu
2013
        prompt_logprobs_dict: dict[str, Optional[LogprobsTensors]] = {}
2014
2015
2016
2017
2018
2019
2020
2021
2022
2023
2024
2025
2026
2027

        # 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():

            num_tokens = scheduler_output.num_scheduled_tokens[req_id]

            # 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)

2028
2029
2030
2031
2032
2033
2034
2035
2036
            # 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

2037
            # Determine number of logits to retrieve.
2038
2039
            start_idx = request.num_computed_tokens
            start_tok = start_idx + 1
2040
            num_remaining_tokens = num_prompt_tokens - start_tok
2041
            if num_tokens <= num_remaining_tokens:
2042
                # This is a chunk, more tokens remain.
2043
2044
2045
                # 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.
2046
2047
2048
2049
2050
                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)
2051
2052
2053
2054
2055
2056
2057
                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
2058
2059
2060
2061
2062
2063
2064
2065
2066
2067
2068
2069
2070
2071
2072

            # 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]
            offset = self.query_start_loc_np[req_idx].item()
            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.
2073
2074
            logprobs = self.sampler.compute_logprobs(logits)
            token_ids, logprobs, ranks = self.sampler.gather_logprobs(
2075
2076
2077
                logprobs, num_prompt_logprobs, tgt_token_ids)

            # Transfer GPU->CPU async.
2078
2079
2080
2081
2082
2083
2084
            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)
2085
2086
2087
2088
2089

        # 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]
2090
            del in_progress_dict[req_id]
2091
2092

        # Must synchronize the non-blocking GPU->CPU transfers.
2093
        if prompt_logprobs_dict:
2094
            self._sync_device()
2095
2096
2097

        return prompt_logprobs_dict

2098
2099
2100
2101
2102
2103
2104
2105
2106
2107
2108
2109
2110
2111
2112
2113
2114
2115
2116
2117
    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 {}

2118
2119
2120
2121
2122
2123
    @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
2124
         - during DP rank dummy run
2125
2126
2127
2128
2129
2130
2131
2132
2133
2134
2135
2136
2137
2138
2139
2140
2141
2142
2143
2144
2145
2146
        """
        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(
                    self.input_ids,
                    low=0,
                    high=self.model_config.get_vocab_size(),
                    dtype=input_ids.dtype)

            logger.debug("Randomizing dummy data for DP Rank")
            input_ids.copy_(rand_input_ids()[:input_ids.size(0)],
                            non_blocking=True)
            yield
            input_ids.fill_(0)

2147
2148
2149
2150
2151
2152
2153
2154
2155
2156
2157
2158
2159
2160
2161
2162
2163
2164
2165
2166
2167
2168
2169
2170
    def _get_mm_dummy_batch(
        self,
        modality: str,
        max_items_per_batch: int,
    ) -> BatchedTensorInputs:
        """Dummy data for profiling and precompiling multimodal models."""
        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},
        )
        dummy_mm_data = dummy_decoder_data.multi_modal_data

        # Result in the maximum GPU consumption of the model
        dummy_mm_item = dummy_mm_data.get_item(modality=modality, item_index=0)
        dummy_mm_kwargs = MultiModalKwargs.from_items([dummy_mm_item])

        batched_dummy_mm_inputs = MultiModalKwargs.batch([dummy_mm_kwargs] *
                                                         max_items_per_batch)
        return MultiModalKwargs.as_kwargs(
            batched_dummy_mm_inputs,
            device=self.device,
        )

2171
2172
2173
2174
    @torch.inference_mode()
    def _dummy_run(
        self,
        num_tokens: int,
2175
        capture_attn_cudagraph: bool = False,
2176
2177
        skip_eplb: bool = False,
        is_profile: bool = False,
2178
    ) -> tuple[torch.Tensor, torch.Tensor]:
2179

2180
        # Padding for DP
2181
2182
        num_pad, num_tokens_across_dp = self.get_dp_padding(num_tokens)
        num_tokens += num_pad
2183

2184
2185
2186
2187
2188
        # 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
2189
        num_reqs = min(num_tokens, max_num_reqs)
2190
2191
2192
2193
2194
2195
2196
        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
        num_scheduled_tokens = np.array(num_scheduled_tokens_list,
                                        dtype=np.int32)
2197

2198
2199
2200
2201
        attn_metadata: Optional[dict[str, Any]] = None
        if capture_attn_cudagraph:
            attn_metadata = {}

2202
2203
2204
2205
2206
            # Make sure max_model_len is used at the graph capture time.
            self.seq_lens_np[:num_reqs] = self.max_model_len
            self.seq_lens_np[num_reqs:] = 0
            self.seq_lens[:num_reqs].copy_(self.seq_lens_cpu[:num_reqs],
                                           non_blocking=True)
2207

2208
2209
            for kv_cache_group_id, kv_cache_group_spec in enumerate(
                    self.kv_cache_config.kv_cache_groups):
2210
2211
2212
2213
2214
2215
2216
2217
2218
2219
2220
2221
2222
2223
                common_attn_metadata = CommonAttentionMetadata(
                    query_start_loc=self.query_start_loc[:num_reqs + 1],
                    query_start_loc_cpu=self.query_start_loc_cpu[:num_reqs +
                                                                 1],
                    seq_lens=self.seq_lens[:num_reqs],
                    seq_lens_cpu=self.seq_lens_cpu[:num_reqs],
                    num_computed_tokens_cpu=self.input_batch.
                    num_computed_tokens_cpu_tensor[:num_reqs],
                    num_reqs=num_reqs,
                    num_actual_tokens=num_tokens,
                    max_query_len=num_tokens,
                    block_table_tensor=self.input_batch.block_table[
                        kv_cache_group_id].get_device_tensor()[:num_reqs],
                    slot_mapping=self.input_batch.
2224
2225
                    block_table[kv_cache_group_id].slot_mapping[:num_tokens],
                    causal=True)
2226
2227
2228
2229

                attn_metadata_i = self.attn_metadata_builders[
                    kv_cache_group_id].build_for_cudagraph_capture(
                        common_attn_metadata)
2230
2231
                for layer_name in kv_cache_group_spec.layer_names:
                    attn_metadata[layer_name] = attn_metadata_i
2232

2233
2234
2235
2236
2237
        with self.maybe_dummy_run_with_lora(self.lora_config,
                                            num_scheduled_tokens):
            if self.is_multimodal_model:
                input_ids = None
                inputs_embeds = self.inputs_embeds[:num_tokens]
2238
                model_mm_kwargs = self._dummy_mm_kwargs(num_reqs)
2239
2240
2241
            else:
                input_ids = self.input_ids[:num_tokens]
                inputs_embeds = None
2242
                model_mm_kwargs = {}
2243

2244
2245
2246
2247
2248
2249
2250
2251
2252
2253
2254
2255
2256
2257
            if self.uses_mrope:
                positions = self.mrope_positions[:, :num_tokens]
            else:
                positions = self.positions[:num_tokens]

            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))
2258
2259
2260

                intermediate_tensors = self.sync_and_slice_intermediate_tensors(
                    num_tokens, None, False)
2261

2262
            with self.maybe_randomize_inputs(input_ids), set_forward_context(
2263
2264
2265
2266
                    attn_metadata,
                    self.vllm_config,
                    num_tokens=num_tokens,
                    num_tokens_across_dp=num_tokens_across_dp):
2267
                outputs = self.model(
2268
2269
2270
2271
                    input_ids=input_ids,
                    positions=positions,
                    intermediate_tensors=intermediate_tensors,
                    inputs_embeds=inputs_embeds,
2272
                    **MultiModalKwargs.as_kwargs(
2273
                        model_mm_kwargs,
2274
2275
                        device=self.device,
                    ),
2276
                )
2277

2278
2279
2280
2281
            if self.use_aux_hidden_state_outputs:
                hidden_states, _ = outputs
            else:
                hidden_states = outputs
2282

2283
            if self.speculative_config and self.speculative_config.use_eagle():
2284
2285
2286
                assert isinstance(self.drafter, EagleProposer)
                self.drafter.dummy_run(num_tokens)

2287
2288
2289
2290
2291
2292
2293
2294
2295
2296
        # 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)

2297
        logit_indices = np.cumsum(num_scheduled_tokens) - 1
2298
        return hidden_states, hidden_states[logit_indices]
2299
2300
2301
2302
2303
2304

    @torch.inference_mode()
    def _dummy_sampler_run(
        self,
        hidden_states: torch.Tensor,
    ) -> torch.Tensor:
2305
2306
2307
2308
        # 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)
2309
2310
2311
2312
2313
2314
2315
2316
2317
2318
2319
2320
2321
2322
2323
2324
2325
2326
2327
2328
2329
2330
2331

        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={},
2332
            logitsprocs=LogitsProcessorManager(),
2333
        )
2334
        try:
2335
2336
            sampler_output = self.sampler(logits=logits,
                                          sampling_metadata=dummy_metadata)
2337
2338
2339
2340
2341
2342
2343
2344
2345
        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
2346
        if self.speculative_config:
2347
2348
2349
2350
2351
2352
2353
2354
2355
2356
2357
2358
2359
2360
2361
2362
2363
2364
2365
2366
2367
2368
2369
2370
2371
2372
            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,
            )
2373
        return sampler_output
2374

2375
    def _dummy_pooler_run_task(
2376
2377
        self,
        hidden_states: torch.Tensor,
2378
2379
        task: PoolingTask,
    ) -> PoolerOutput:
2380
2381
2382
2383
2384
2385
2386
2387
2388
2389
2390
2391
2392
        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

        hidden_states_list = list(
            torch.split(hidden_states, num_scheduled_tokens_list))
        req_num_tokens = num_tokens // num_reqs

2393
2394
2395
2396
2397
2398
2399
        dummy_prompt_lens = torch.tensor(
            [h.shape[0] for h in hidden_states_list],
            device=self.device,
        )
        dummy_token_ids = torch.zeros((num_reqs, req_num_tokens),
                                      dtype=torch.int32,
                                      device=self.device)
2400

2401
2402
2403
        model = cast(VllmModelForPooling, self.model)
        dummy_pooling_params = PoolingParams(task=task)
        to_update = model.pooler.get_pooling_updates(task)
2404
2405
        to_update.apply(dummy_pooling_params)

2406
        dummy_metadata = PoolingMetadata(
2407
2408
2409
2410
            prompt_lens=dummy_prompt_lens,
            prompt_token_ids=dummy_token_ids,
            pooling_params=[dummy_pooling_params] * num_reqs,
        )
2411
2412

        try:
2413
2414
            return model.pooler(hidden_states=hidden_states_list,
                                pooling_metadata=dummy_metadata)
2415
2416
2417
        except RuntimeError as e:
            if 'out of memory' in str(e):
                raise RuntimeError(
2418
2419
2420
                    "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 "
2421
2422
2423
                    "initializing the engine.") from e
            else:
                raise e
2424
2425
2426
2427
2428
2429
2430
2431
2432
2433
2434
2435
2436
2437
2438
2439

    @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)
2440

2441
    def profile_run(self) -> None:
2442
        # Profile with multimodal encoder & encoder cache.
2443
2444
2445
2446
2447
2448
2449
2450
2451
2452
2453
2454
2455
2456
2457
2458
2459
2460
2461
2462
2463
2464
2465
2466
2467
2468
        if self.is_multimodal_model:
            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.
                (
                    dummy_modality,
                    max_tokens,
                ) = mm_budget.get_modality_with_max_tokens()
                (
                    max_mm_items_per_prompt,
                    max_mm_items_per_batch,
                ) = mm_budget.get_max_items(dummy_modality, max_tokens)

                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,
                )
2469

2470
2471
2472
2473
2474
                # Create dummy batch of multimodal inputs.
                batched_dummy_mm_inputs = self._get_mm_dummy_batch(
                    dummy_modality,
                    max_mm_items_per_batch,
                )
2475

2476
2477
2478
2479
2480
2481
2482
2483
                # Run multimodal encoder.
                dummy_encoder_outputs = self.model.get_multimodal_embeddings(
                    **batched_dummy_mm_inputs)

                sanity_check_mm_encoder_outputs(
                    dummy_encoder_outputs,
                    expected_num_items=max_mm_items_per_batch,
                )
2484

2485
2486
2487
                # Cache the dummy encoder outputs.
                self.encoder_cache["tmp"] = dict(
                    enumerate(dummy_encoder_outputs))
2488

2489
        # Add `is_profile` here to pre-allocate communication buffers
2490
        hidden_states, last_hidden_states \
2491
            = self._dummy_run(self.max_num_tokens, is_profile=True)
2492
        if get_pp_group().is_last_rank:
2493
2494
2495
2496
            if self.is_pooling_model:
                output = self._dummy_pooler_run(hidden_states)
            else:
                output = self._dummy_sampler_run(last_hidden_states)
2497
        else:
2498
            output = None
2499
        self._sync_device()
2500
        del hidden_states, output
2501
        self.encoder_cache.clear()
2502
        gc.collect()
2503
2504

    def capture_model(self) -> None:
2505
2506
        if not self.use_cuda_graph:
            logger.warning(
2507
2508
2509
                "Skipping CUDA graph capture. To turn on CUDA graph capture, "
                "set -O %s and ensure `use_cudagraph` was not manually set to "
                "False", CompilationLevel.PIECEWISE)
2510
2511
            return

2512
2513
        compilation_counter.num_gpu_runner_capture_triggers += 1

2514
2515
2516
        start_time = time.perf_counter()
        start_free_gpu_memory = torch.cuda.mem_get_info()[0]

2517
2518
2519
2520
2521
2522
2523
2524
2525
2526
2527
2528
2529
2530
2531
        @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()

2532
2533
2534
        # 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.
2535
        with freeze_gc(), graph_capture(device=self.device):
2536
            full_cg = self.full_cuda_graph
2537
2538
2539
            # Only rank 0 should print progress bar during capture
            compilation_cases = reversed(self.cudagraph_batch_sizes)
            if is_global_first_rank():
2540
2541
2542
2543
                compilation_cases = tqdm(
                    list(compilation_cases),
                    disable=not self.load_config.use_tqdm_on_load,
                    desc="Capturing CUDA graph shapes")
2544
            for num_tokens in compilation_cases:
2545
                # We skip EPLB here since we don't want to record dummy metrics
2546
2547
                for _ in range(
                        self.compilation_config.cudagraph_num_of_warmups):
2548
2549
2550
2551
2552
2553
                    self._dummy_run(num_tokens,
                                    capture_attn_cudagraph=full_cg,
                                    skip_eplb=True)
                self._dummy_run(num_tokens,
                                capture_attn_cudagraph=full_cg,
                                skip_eplb=True)
2554
2555
2556
2557
2558
2559
2560
2561

        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))
2562

2563
    def _initialize_single_attn_backend(
2564
        self, kv_cache_spec: KVCacheSpec, layer_names: list[str]
2565
2566
2567
2568
2569
2570
2571
2572
2573
2574
2575
2576
2577
2578
2579
2580
2581
2582
2583
2584
2585
2586
    ) -> tuple[AttentionBackend, AttentionMetadataBuilder]:
        if isinstance(kv_cache_spec, AttentionSpec):
            attn_backend_i = get_attn_backend(
                kv_cache_spec.head_size,
                self.dtype,
                kv_cache_spec.dtype,
                kv_cache_spec.block_size,
                self.model_config.is_attention_free,
                use_mla=kv_cache_spec.use_mla,
            )
            if attn_backend_i is None:
                error_msg = (f"Error with get_attn_backend: "
                             f"{kv_cache_spec.head_size=}, "
                             f"{self.dtype=}, {kv_cache_spec.dtype=}, "
                             f"{kv_cache_spec.block_size=}, "
                             f"{self.model_config.is_attention_free=}, "
                             f"{kv_cache_spec.use_mla=}")
                logger.error(error_msg)
                raise NotImplementedError(
                    "Non-Attention backend is not supported by V1 "
                    "GPUModelRunner.")
        elif isinstance(kv_cache_spec, MambaSpec):
2587
            attn_backend_i = get_mamba_attn_backend(kv_cache_spec.mamba_type)
2588
2589
2590
2591
2592
2593
        else:
            raise ValueError(
                f"Unknown KV cache spec type: {type(kv_cache_spec)}")

        attn_metadata_builder_i = attn_backend_i.get_builder_cls()(
            kv_cache_spec,
2594
            layer_names,
2595
2596
2597
2598
            self.vllm_config,
            self.device,
        )

2599
2600
2601
2602
2603
2604
2605
2606
2607
2608
2609
2610
2611
2612
2613
2614
        if self.full_cuda_graph:
            if attn_metadata_builder_i.attn_cudagraph_support == \
                AttentionCGSupport.NEVER:
                raise ValueError(f"Full CUDAGraph not supported for "
                                 f"{attn_backend_i.__name__}. Turn off "
                                 f"CompilationConfig.full_cuda_graph or use a "
                                 f" different attention backend.")
            if attn_metadata_builder_i.attn_cudagraph_support == \
                AttentionCGSupport.PURE_DECODE_ONLY:
                # Limit the max cudagraph size to the max number of
                # sequences for pure decode only cudagraph backend,
                # whose max_query_len is 1.
                self.cudagraph_batch_sizes = [
                    size for size in self.cudagraph_batch_sizes
                    if size <= self.scheduler_config.max_num_seqs
                ]
2615
2616
        return attn_backend_i, attn_metadata_builder_i

2617
2618
2619
2620
2621
2622
2623
2624
2625
2626
    def initialize_attn_backend(self, kv_cache_config: KVCacheConfig) -> None:
        """
        Initialize the attention backends and attention metadata builders.
        """
        assert len(self.attn_backends) == 0 and len(
            self.attn_metadata_builders
        ) == 0, "Attention backends are already initialized"
        for i, kv_cache_group_spec in enumerate(
                kv_cache_config.kv_cache_groups):
            kv_cache_spec = kv_cache_group_spec.kv_cache_spec
2627

2628
2629
2630
            attn_backend_i, attn_metadata_builder_i = (
                self._initialize_single_attn_backend(
                    kv_cache_spec, kv_cache_group_spec.layer_names))
2631
2632
2633
            self.attn_backends.append(attn_backend_i)
            self.attn_metadata_builders.append(attn_metadata_builder_i)

2634
2635
2636
        # Calculate reorder batch threshold (if neeeded)
        self.calculate_reorder_batch_threshold()

2637
2638
2639
2640
2641
2642
2643
2644
2645
2646
2647
2648
2649
2650
2651
2652
2653
2654
2655
2656
2657
2658
2659
2660
2661
2662
2663
        if len(self.attn_backends) > 0:
            return

        # Check if model is encoder-only
        block_size = self.vllm_config.cache_config.block_size
        use_mla = self.vllm_config.model_config.use_mla
        attn_specs = list[AttentionSpec]()
        attn_layers = get_layers_from_vllm_config(self.vllm_config, Attention)
        for attn_module in attn_layers.values():

            if attn_module.attn_type == AttentionType.ENCODER_ONLY:
                assert attn_module.sliding_window is None, "Sliding "
                "window attention is not supported for encoder-only models"

                attn_specs.append(
                    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))
            else:
                raise ValueError("Expected only encoder-only layers")

        if len(attn_specs) > 0:
            assert len(attn_specs) == len(attn_layers), \
                "All or none of the layers are expected to be encoder-only"

2664
2665
2666
            attn_backend, attn_metadata_builder = (
                self._initialize_single_attn_backend(attn_specs[0],
                                                     attn_layers.keys()))
2667
2668
2669
2670
            self.attn_backends.append(attn_backend)
            self.attn_metadata_builders.append(attn_metadata_builder)
            self.is_encoder_only_model = True

2671
2672
2673
2674
2675
2676
2677
2678
2679
2680
2681
2682
2683
2684
2685
2686
2687
2688
2689
2690
2691
2692
    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)
        """
        for attn_metadata_builder_i in self.attn_metadata_builders:
            # 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

2693
2694
2695
2696
2697
2698
2699
2700
2701
2702
2703
2704
2705
2706
2707
2708
2709
2710
2711
2712
2713
2714
2715
2716
2717
2718
2719
    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,
2720
                is_spec_decode=bool(self.vllm_config.speculative_config),
2721
2722
            )

2723
2724
    def _allocate_kv_cache_tensors(
            self, kv_cache_config: KVCacheConfig) -> dict[str, torch.Tensor]:
2725
        """
2726
2727
2728
        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.

2729
        Args:
2730
            kv_cache_config: The KV cache config
2731
        Returns:
2732
            dict[str, torch.Tensor]: A map between layer names to their
2733
2734
2735
2736
2737
2738
2739
2740
2741
2742
2743
2744
2745
2746
2747
2748
2749
2750
2751
2752
2753
2754
            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:
            layer_names.update(group.layer_names)
        assert layer_names == set(kv_cache_raw_tensors.keys(
        )), "Some layers are not correctly initialized"
        return kv_cache_raw_tensors

    def _reshape_kv_cache_tensors(
        self,
        kv_cache_config: KVCacheConfig,
        kv_cache_raw_tensors: dict[str, torch.Tensor],
    ) -> dict[str, torch.Tensor]:
2755
        """
2756
        Reshape the KV cache tensors to the desired shape and dtype.
2757

2758
        Args:
2759
2760
            kv_cache_config: The KV cache config
            kv_cache_raw_tensors: The KV cache buffer of each layer, with
2761
2762
            correct size but uninitialized shape.
        Returns:
2763
            Dict[str, torch.Tensor]: A map between layer names to their
2764
2765
            corresponding memory buffer for KV cache.
        """
2766
        kv_caches: dict[str, torch.Tensor] = {}
2767
        has_attn, has_mamba = False, False
2768
2769
2770
2771
2772
2773
2774
2775
        for i, kv_cache_group_spec in enumerate(
                kv_cache_config.kv_cache_groups):
            kv_cache_spec = kv_cache_group_spec.kv_cache_spec
            for layer_name in kv_cache_group_spec.layer_names:
                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)
2776
                if isinstance(kv_cache_spec, AttentionSpec):
2777
                    has_attn = True
2778
                    kv_cache_shape = self.attn_backends[i].get_kv_cache_shape(
2779
2780
2781
                        num_blocks, kv_cache_spec.block_size,
                        kv_cache_spec.num_kv_heads, kv_cache_spec.head_size)
                    dtype = kv_cache_spec.dtype
2782
2783
2784
2785
2786
2787
2788
2789
2790
2791
2792
2793
2794
2795
2796
2797
2798
2799
2800
2801
                    try:
                        kv_cache_stride_order = self.attn_backends[
                            i].get_kv_cache_stride_order()
                        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))
                    ]
2802
2803
2804
                    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
2805
                elif isinstance(kv_cache_spec, MambaSpec):
2806
                    has_mamba = True
Chen Zhang's avatar
Chen Zhang committed
2807
2808
                    raw_tensor = kv_cache_raw_tensors[layer_name]
                    dtype = kv_cache_spec.dtype
2809
2810
                    num_element_per_page = (kv_cache_spec.page_size_bytes //
                                            get_dtype_size(dtype))
Chen Zhang's avatar
Chen Zhang committed
2811
                    state_tensors = []
2812
                    storage_offset = 0
Chen Zhang's avatar
Chen Zhang committed
2813
2814
                    for shape in kv_cache_spec.shapes:
                        target_shape = (num_blocks, *shape)
2815
2816
2817
2818
2819
2820
2821
2822
                        stride = torch.empty(target_shape).stride()
                        target_stride = (num_element_per_page, *stride[1:])
                        tensor = torch.as_strided(
                            raw_tensor.view(dtype),
                            size=target_shape,
                            stride=target_stride,
                            storage_offset=storage_offset,
                        )
Chen Zhang's avatar
Chen Zhang committed
2823
                        state_tensors.append(tensor)
2824
2825
2826
                        storage_offset += stride[0]

                    kv_caches[layer_name] = state_tensors
2827
                else:
2828
                    raise NotImplementedError
2829
2830
2831
2832
2833

        if has_attn and has_mamba:
            self._verify_hybrid_attention_mamba_layout(kv_cache_config,
                                                       kv_cache_raw_tensors)

2834
2835
        return kv_caches

2836
2837
2838
2839
2840
2841
2842
2843
2844
2845
2846
2847
2848
2849
2850
2851
2852
2853
2854
2855
2856
2857
2858
2859
2860
2861
2862
2863
2864
2865
2866
    def _verify_hybrid_attention_mamba_layout(
            self, kv_cache_config: KVCacheConfig,
            kv_cache_raw_tensors: dict[str, torch.Tensor]) -> None:
        """
        Verify that the KV cache memory layout is compatible for
        models with both attention and mamba KV cache groups.

        Args:
            kv_cache_config: The KV cache config
            kv_cache_raw_tensors: The KV cache buffer of each layer.
        """

        for i, kv_cache_group_spec in enumerate(
                kv_cache_config.kv_cache_groups):
            kv_cache_spec = kv_cache_group_spec.kv_cache_spec
            for layer_name in kv_cache_group_spec.layer_names:
                raw_tensor = kv_cache_raw_tensors[layer_name]
                num_blocks = (raw_tensor.numel() //
                              kv_cache_spec.page_size_bytes)
                if isinstance(kv_cache_spec, AttentionSpec):
                    kv_cache_shape = self.attn_backends[i].get_kv_cache_shape(
                        num_blocks, kv_cache_spec.block_size,
                        kv_cache_spec.num_kv_heads, kv_cache_spec.head_size)
                    if kv_cache_shape[0] != num_blocks or kv_cache_shape[
                            1] != 2:
                        raise ValueError(
                            "Hybrid models in V1 require an attention "
                            "backend with kv_cache_shape="
                            "(num_blocks, 2, ...). Please try setting "
                            "VLLM_ATTENTION_BACKEND=FLASHINFER")

2867
2868
2869
2870
2871
2872
2873
2874
    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:
2875
            Dict[str, torch.Tensor]: A map between layer names to their
2876
2877
2878
2879
2880
2881
2882
            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)
2883

2884
2885
2886
2887
2888
2889
2890
2891
        # Setup `kv_cache_config` and `kv_caches` for models
        # with cross-layer KV sharing
        if self.shared_kv_cache_layers:
            initialize_kv_cache_for_kv_sharing(
                self.shared_kv_cache_layers,
                kv_cache_config.kv_cache_groups,
                kv_caches,
            )
2892
2893
2894
2895
2896
2897
2898
2899
2900
2901
            attn_layers = get_layers_from_vllm_config(self.vllm_config,
                                                      Attention)
            # Iterate in reversed order and add layers that re-use KV cache
            # e.g. in YOCO-like KV sharing setups (e.g. Gemma3n)
            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
2902

2903
2904
2905
        bind_kv_cache(kv_caches,
                      self.compilation_config.static_forward_context,
                      self.kv_caches)
2906
2907
2908
2909
2910
2911
2912
2913
2914
2915
2916
2917
2918
2919
        return kv_caches

    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
        """
        self.kv_cache_config = kv_cache_config
        self.may_reinitialize_input_batch(kv_cache_config)
        self.initialize_attn_backend(kv_cache_config)
        kv_caches = self.initialize_kv_cache_tensors(kv_cache_config)

2920
2921
2922
2923
2924
2925
        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
2926
2927
2928
        if has_kv_transfer_group():
            get_kv_transfer_group().register_kv_caches(kv_caches)

2929
    def get_kv_cache_spec(self) -> dict[str, KVCacheSpec]:
2930
        """
2931
        Generates the KVCacheSpec by parsing the kv cache format from each
2932
2933
        Attention module in the static forward context.
        Returns:
2934
            KVCacheSpec: A dictionary mapping layer names to their KV cache
2935
2936
2937
2938
            format. Layers that do not need KV cache are not included.
        """

        block_size = self.vllm_config.cache_config.block_size
2939
        use_mla = self.vllm_config.model_config.use_mla
2940
        kv_cache_spec: dict[str, KVCacheSpec] = {}
Chen Zhang's avatar
Chen Zhang committed
2941
2942
        attn_layers = get_layers_from_vllm_config(self.vllm_config, Attention)
        for layer_name, attn_module in attn_layers.items():
2943
2944
2945
2946
2947
2948
2949
2950
2951
2952
2953
2954
            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

2955
            # TODO: Support other attention modules, e.g., cross-attention
2956
            if attn_module.attn_type == AttentionType.DECODER:
2957
                use_local_attention = (self.attention_chunk_size is not None
2958
                                       and attn_module.use_irope)
2959
2960
2961
2962
2963
2964
2965
2966
                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)
2967
2968
2969
                    assert not use_local_attention, (
                        "attention module can not be with ",
                        "both local attention and sliding window")
2970
                elif use_local_attention:
2971
                    kv_cache_spec[layer_name] = ChunkedLocalAttentionSpec(
2972
2973
2974
2975
2976
                        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,
2977
                        use_mla=use_mla)
2978
2979
2980
2981
2982
2983
2984
                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)
2985
2986
2987
2988
2989
2990
2991
2992
2993
2994
            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}")

2995
        mamba_layers = get_layers_from_vllm_config(self.vllm_config, MambaBase)
Chen Zhang's avatar
Chen Zhang committed
2996
2997
2998
2999
3000
3001
3002
3003
        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
3004

3005
3006
            page_size_padded = (
                self.vllm_config.cache_config.mamba_page_size_padded)
3007

Chen Zhang's avatar
Chen Zhang committed
3008
3009
3010
3011
3012
3013
            # 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(),
                    dtype=self.kv_cache_dtype,
3014
                    block_size=max_model_len,
3015
3016
                    page_size_padded=page_size_padded,
                    mamba_type=mamba_module.mamba_type)
3017

3018
        return kv_cache_spec
3019
3020
3021
3022
3023
3024
3025
3026
3027
3028
3029
3030
3031
3032
3033
3034
3035
3036
3037
3038
3039
3040
3041
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 _build_encoder_only_attn_metadata(
            self, scheduler_output: "SchedulerOutput") -> \
                tuple[CommonAttentionMetadata, Any]:
        """Prepare encoder attention metadata for encoder-only models.

        Args:
            scheduler_output: Scheduler output

        Returns:
            dict[str, Any]: Encoder attention metadata
        """
        num_reqs = self.input_batch.num_reqs
        total_num_scheduled_tokens = scheduler_output.total_num_scheduled_tokens

        # Get the number of scheduled tokens for each request.
        req_ids = self.input_batch.req_ids
        tokens = [scheduler_output.num_scheduled_tokens[i] for i in req_ids]
        max_num_scheduled_tokens = max(tokens)

        # Use the first attention metadata builder
        # to create encoder attention metadata
        builder = self.attn_metadata_builders[0]

        dummy_block_table = torch.zeros((num_reqs, 1),
                                        dtype=torch.int32,
                                        device=self.device)
        dummy_slot_mapping = torch.zeros((total_num_scheduled_tokens, ),
                                         dtype=torch.int32,
                                         device=self.device)

        common_metadata = CommonAttentionMetadata(
            query_start_loc=self.query_start_loc[:num_reqs + 1],
            query_start_loc_cpu=self.query_start_loc_cpu[:num_reqs + 1],
            seq_lens=self.seq_lens[:num_reqs],
            seq_lens_cpu=self.seq_lens_cpu[:num_reqs],
            num_computed_tokens_cpu=self.input_batch.
            num_computed_tokens_cpu_tensor[:num_reqs],
            num_reqs=num_reqs,
            num_actual_tokens=total_num_scheduled_tokens,
            max_query_len=max_num_scheduled_tokens,
            block_table_tensor=dummy_block_table,
            slot_mapping=dummy_slot_mapping,
            causal=False,
        )

        return common_metadata, builder.build(
            common_prefix_len=0,  # No cascade for encoder
            common_attn_metadata=common_metadata,
        )