gpu_model_runner.py 123 KB
Newer Older
1
# SPDX-License-Identifier: Apache-2.0
2
# SPDX-FileCopyrightText: Copyright contributors to the vLLM project
3

Robert Shaw's avatar
Robert Shaw committed
4
import copy
5
import gc
6
import time
7
import weakref
8
from contextlib import contextmanager
9
from typing import TYPE_CHECKING, Any, Optional, Union
10
11
12
13
14

import numpy as np
import torch
import torch.distributed
import torch.nn as nn
15
from tqdm import tqdm
16

17
import vllm.envs as envs
18
from vllm.attention import AttentionType, get_attn_backend
19
from vllm.attention.backends.abstract import AttentionBackend
20
from vllm.attention.layer import Attention
21
from vllm.compilation.counter import compilation_counter
22
23
from vllm.config import (CompilationLevel, VllmConfig,
                         get_layers_from_vllm_config)
24
from vllm.distributed.eplb.eplb_state import EplbState
25
26
from vllm.distributed.kv_transfer import (get_kv_transfer_group,
                                          has_kv_transfer_group)
Robert Shaw's avatar
Robert Shaw committed
27
from vllm.distributed.kv_transfer.kv_connector.v1 import KVConnectorBase_V1
28
from vllm.distributed.parallel_state import (
29
    get_pp_group, get_tp_group, graph_capture, is_global_first_rank,
30
    prepare_communication_buffer_for_model)
31
32
from vllm.forward_context import (DPMetadata, get_forward_context,
                                  set_forward_context)
33
from vllm.logger import init_logger
Chen Zhang's avatar
Chen Zhang committed
34
from vllm.model_executor.layers.mamba.mamba_mixer2 import MambaMixer2
35
from vllm.model_executor.layers.rotary_embedding import MRotaryEmbedding
36
from vllm.model_executor.model_loader import TensorizerLoader, get_model_loader
37
38
from vllm.model_executor.models.interfaces import (has_step_pooler,
                                                   is_mixture_of_experts)
39
40
from vllm.multimodal import MULTIMODAL_REGISTRY
from vllm.multimodal.inputs import 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
45
from vllm.utils import (STR_DTYPE_TO_TORCH_DTYPE, DeviceMemoryProfiler,
46
                        GiB_bytes, LazyLoader, async_tensor_h2d, cdiv,
Chen Zhang's avatar
Chen Zhang committed
47
                        check_use_alibi, get_dtype_size,
Woosuk Kwon's avatar
Woosuk Kwon committed
48
                        is_pin_memory_available, round_up)
Chen Zhang's avatar
Chen Zhang committed
49
from vllm.v1.attention.backends.mamba_attn import Mamba2AttentionBackend
50
51
from vllm.v1.attention.backends.utils import (AttentionMetadataBuilder,
                                              CommonAttentionMetadata)
52
from vllm.v1.core.encoder_cache_manager import compute_encoder_budget
53
from vllm.v1.kv_cache_interface import (AttentionSpec, FullAttentionSpec,
Chen Zhang's avatar
Chen Zhang committed
54
                                        KVCacheConfig, KVCacheSpec, MambaSpec,
55
                                        SlidingWindowSpec)
56
57
from vllm.v1.outputs import (EMPTY_MODEL_RUNNER_OUTPUT, LogprobsTensors,
                             ModelRunnerOutput)
58
from vllm.v1.pool.metadata import PoolingMetadata
59
from vllm.v1.sample.metadata import SamplingMetadata
60
from vllm.v1.sample.rejection_sampler import RejectionSampler
61
from vllm.v1.sample.sampler import Sampler
62
from vllm.v1.spec_decode.eagle import EagleProposer
63
from vllm.v1.spec_decode.medusa import MedusaProposer
64
from vllm.v1.spec_decode.metadata import SpecDecodeMetadata
65
from vllm.v1.spec_decode.ngram_proposer import NgramProposer
66
from vllm.v1.utils import bind_kv_cache
67
from vllm.v1.worker.block_table import BlockTable
68
from vllm.v1.worker.gpu_input_batch import CachedRequestState, InputBatch
69
from vllm.v1.worker.lora_model_runner_mixin import LoRAModelRunnerMixin
70

71
from ..sample.logits_processor import LogitsProcessorManager
72
73
from .utils import (gather_mm_placeholders, initialize_kv_cache_for_kv_sharing,
                    sanity_check_mm_encoder_outputs, scatter_mm_placeholders)
74

75
if TYPE_CHECKING:
76
    import xgrammar as xgr
77
    import xgrammar.kernels.apply_token_bitmask_inplace_torch_compile as xgr_torch_compile  # noqa: E501
78

79
    from vllm.model_executor.model_loader.tensorizer import TensorizerConfig
80
    from vllm.v1.core.sched.output import SchedulerOutput
81
82
else:
    xgr = LazyLoader("xgr", globals(), "xgrammar")
83
84
85
    xgr_torch_compile = LazyLoader(
        "xgr_torch_compile", globals(),
        "xgrammar.kernels.apply_token_bitmask_inplace_torch_compile")
86
87
88
89

logger = init_logger(__name__)


90
class GPUModelRunner(LoRAModelRunnerMixin):
91
92
93

    def __init__(
        self,
94
        vllm_config: VllmConfig,
95
        device: torch.device,
96
    ):
97
98
99
        self.vllm_config = vllm_config
        self.model_config = vllm_config.model_config
        self.cache_config = vllm_config.cache_config
100
        self.compilation_config = vllm_config.compilation_config
101
102
103
104
105
106
107
        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.prompt_adapter_config = vllm_config.prompt_adapter_config
        self.observability_config = vllm_config.observability_config
108

109
110
111
112
        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))

113
114
115
116
        model_config = self.model_config
        cache_config = self.cache_config
        scheduler_config = self.scheduler_config
        parallel_config = self.parallel_config
117
        self.device = device
118
119
120
121
122
123
124
125
        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]

126
        self.is_multimodal_model = model_config.is_multimodal_model
127
        self.is_pooling_model = model_config.pooler_config is not None
128
129
        self.max_model_len = model_config.max_model_len
        self.max_num_tokens = scheduler_config.max_num_batched_tokens
130
        self.max_num_reqs = scheduler_config.max_num_seqs
131
132

        # Model-related.
133
134
        self.num_query_heads = model_config.get_num_attention_heads(
            parallel_config)
135
        self.hidden_size = model_config.get_hidden_size()
136
        self.attention_chunk_size = model_config.attention_chunk_size
137

138
        self.cascade_attn_enabled = not self.model_config.disable_cascade_attn
139

140
        # Multi-modal data support
141
        self.mm_registry = MULTIMODAL_REGISTRY
142
        self.uses_mrope = model_config.uses_mrope
143

144
145
146
        encoder_compute_budget, encoder_cache_size = compute_encoder_budget(
            model_config=model_config,
            scheduler_config=scheduler_config,
147
            mm_registry=self.mm_registry,
148
149
150
        )
        self.max_num_encoder_input_tokens = encoder_compute_budget
        self.encoder_cache_size = encoder_cache_size
151

152
153
154
        # Sampler
        self.sampler = Sampler()

155
156
157
158
159
160
161
        self.eplb_state: Optional[EplbState] = None
        """
        State of the expert parallelism load balancer.

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

162
        # Lazy initializations
163
        # self.model: nn.Module  # Set after load_model
164
        # Initialize in initialize_kv_cache
165
        self.kv_caches: list[torch.Tensor] = []
166
167
        self.attn_metadata_builders: list[AttentionMetadataBuilder] = []
        self.attn_backends: list[type[AttentionBackend]] = []
168
169
        # self.kv_cache_config: KVCacheConfig

170
        # req_id -> (input_id -> encoder_output)
171
        self.encoder_cache: dict[str, dict[int, torch.Tensor]] = {}
172

173
        self.use_aux_hidden_state_outputs = False
174
175
176
177
178
179
180
181
182
183
184
185
186
187
188
189
190
191
192
193
        # 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()
194

195
        # Request states.
196
        self.requests: dict[str, CachedRequestState] = {}
197

198
199
200
201
202
203
204
205
206
        # 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.
207
208
209
210
211
212
        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,
213
            vocab_size=self.model_config.get_vocab_size(),
214
            block_sizes=[self.cache_config.block_size],
215
            is_spec_decode=bool(self.vllm_config.speculative_config),
216
        )
217

218
219
220
221
222
        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)
223
        # TODO(woosuk): Provide an option to tune the max cudagraph batch size.
224
225
226
227
        # 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(
228
229
230
            reversed(self.compilation_config.cudagraph_capture_sizes))

        self.full_cuda_graph = self.compilation_config.full_cuda_graph
231

232
        # Cache the device properties.
233
        self._init_device_properties()
234

235
236
237
238
        # Persistent buffers for CUDA graphs.
        self.input_ids = torch.zeros(self.max_num_tokens,
                                     dtype=torch.int32,
                                     device=self.device)
239
240
241
        self.positions = torch.zeros(self.max_num_tokens,
                                     dtype=torch.int64,
                                     device=self.device)
242
243
244
245
246
247
248
249
250
251
        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)

252
253
        # None in the first PP rank. The rest are set after load_model.
        self.intermediate_tensors: Optional[IntermediateTensors] = None
254
255

        # Only relevant for models using M-RoPE (e.g, Qwen2-VL)
256
        if self.uses_mrope:
Roger Wang's avatar
Roger Wang committed
257
258
259
260
            # 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
261
262
263
264
265
266

            # 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
267
            self.mrope_positions = torch.zeros((3, self.max_num_tokens + 1),
268
269
                                               dtype=torch.int64,
                                               device=self.device)
Roger Wang's avatar
Roger Wang committed
270
271
272
273
274
            self.mrope_positions_cpu = torch.zeros(
                (3, self.max_num_tokens + 1),
                dtype=torch.int64,
                device="cpu",
                pin_memory=self.pin_memory)
275
            self.mrope_positions_np = self.mrope_positions_cpu.numpy()
276

277
278
279
        # Only relevant for models using ALiBi (e.g, MPT)
        self.use_alibi = check_use_alibi(model_config)

280
281
282
283
        self.inputs_embeds = torch.zeros(
            (self.max_num_tokens, self.hidden_size),
            dtype=self.dtype,
            device=self.device)
284

285
        # OPTIMIZATION: Cache the tensors rather than creating them every step.
286
        # Keep in int64 to avoid overflow with long context
287
        self.arange_np = np.arange(max(self.max_num_reqs + 1,
288
289
                                       self.max_model_len,
                                       self.max_num_tokens),
290
                                   dtype=np.int64)
291
292
293
        # 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.
294
295
296
297
298
299
300
301
302
303
304
305
306
307
        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()
308
309
310
311
312
        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()
313

314
315
316
317
318
319
        # 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] = {}

320
    def _may_reorder_batch(self, scheduler_output: "SchedulerOutput") -> None:
321
322
        """
        Update the order of requests in the batch based on the attention
323
        backend's needs. For example, some attention backends (namely MLA) may
324
325
326
327
328
329
        want to separate requests based on if the attention computation will be
        compute-bound or memory-bound.

        Args:
            scheduler_output: The scheduler output.
        """
330
331
        self.attn_metadata_builders[0].reorder_batch(self.input_batch,
                                                     scheduler_output)
332
333
334
335
336

        # For models with multiple KV cache groups, the groups should agree on
        # the same order of requests. We ensure this by only allowing the first
        # group to reorder the batch and asserting that all other groups do not
        # reorder the batch.
337
338
339
        # TODO(tdoublep): make this more flexible so that any group can
        # re-order the batch (not only the first).
        # TODO(tdoublep): verify this during engine init instead of at runtime
340
        for i in range(1, len(self.kv_cache_config.kv_cache_groups)):
341
            batch_reordered = self.attn_metadata_builders[i].reorder_batch(
342
                self.input_batch, scheduler_output)
343
            assert not batch_reordered
344

345
346
347
348
349
350
351
352
353
354
355
    # 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()

356
    def _update_states(self, scheduler_output: "SchedulerOutput") -> None:
357
358
359
360
361
362
        """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.

363
364
        The SamplingMetadata is updated and copied to the GPU if there is a
        new/resumed/paused/finished request in the batch.
365
366
        """
        # Remove finished requests from the cached states.
367
368
        for req_id in scheduler_output.finished_req_ids:
            self.requests.pop(req_id, None)
369
            self.encoder_cache.pop(req_id, None)
370
371
372
373
374
375
376
        # 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:
377
            self.input_batch.remove_request(req_id)
378
379
380
381
382
383
384
385

        # 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)
386

387
388
389
390
391
392
393
394
395
396
397
398
399
        # 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:
400
            self.input_batch.remove_request(req_id)
401

402
        req_ids_to_add: list[str] = []
403
        # Add new requests to the cached states.
404
405
406
        for new_req_data in scheduler_output.scheduled_new_reqs:
            req_id = new_req_data.req_id
            sampling_params = new_req_data.sampling_params
407
408
409
            pooling_params = new_req_data.pooling_params
            if sampling_params and \
                sampling_params.sampling_type == SamplingType.RANDOM_SEED:
410
411
412
413
414
                generator = torch.Generator(device=self.device)
                generator.manual_seed(sampling_params.seed)
            else:
                generator = None

415
416
            self.requests[req_id] = CachedRequestState(
                req_id=req_id,
417
418
419
                prompt_token_ids=new_req_data.prompt_token_ids,
                mm_inputs=new_req_data.mm_inputs,
                mm_positions=new_req_data.mm_positions,
420
                sampling_params=sampling_params,
421
                pooling_params=pooling_params,
422
                generator=generator,
423
424
                block_ids=new_req_data.block_ids,
                num_computed_tokens=new_req_data.num_computed_tokens,
425
                output_token_ids=[],
426
                lora_request=new_req_data.lora_request,
427
            )
428
429

            # Only relevant for models using M-RoPE (e.g, Qwen2-VL)
430
            if self.uses_mrope:
431
432
                image_grid_thw = []
                video_grid_thw = []
Roger Wang's avatar
Roger Wang committed
433
                second_per_grid_ts = []
434
435
                audio_feature_lengths = []
                use_audio_in_video = False
436
437
438
439
440
441
442
                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
443
444
445
                    if mm_input.get("second_per_grid_ts") is not None:
                        second_per_grid_ts.extend(
                            mm_input["second_per_grid_ts"])
446
447
448
449
450
                    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
451
452
453
454
455
456
457

                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
458
                        hf_config=hf_config,
459
460
                        image_grid_thw=image_grid_thw,
                        video_grid_thw=video_grid_thw,
Roger Wang's avatar
Roger Wang committed
461
                        second_per_grid_ts=second_per_grid_ts,
462
463
                        audio_feature_lengths=audio_feature_lengths,
                        use_audio_in_video=use_audio_in_video,
464
465
                    )

466
467
            req_ids_to_add.append(req_id)

468
        # Update the states of the running/resumed requests.
469
        is_last_rank = get_pp_group().is_last_rank
470
471
        req_data = scheduler_output.scheduled_cached_reqs
        for i, req_id in enumerate(req_data.req_ids):
472
            req_state = self.requests[req_id]
473
474
475
            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]
476

477
            # Update the cached states.
478
            req_state.num_computed_tokens = num_computed_tokens
479
480
481
482
483
484
485
486
487
488
489
490
491
492
493
494
495

            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:])

496
            # Update the block IDs.
497
            if not resumed_from_preemption:
498
                # Append the new blocks to the existing block IDs.
499
500
501
                for block_ids, new_ids in zip(req_state.block_ids,
                                              new_block_ids):
                    block_ids.extend(new_ids)
502
503
504
            else:
                # The request is resumed from preemption.
                # Replace the existing block IDs with the new ones.
505
                req_state.block_ids = new_block_ids
506
507
508
509
510
511
512
513
514
515
516

            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] = (
517
                num_computed_tokens)
518
            self.input_batch.block_table.append_row(new_block_ids, req_index)
519
520
521
522
523
524
525

            # 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)
526
                self.input_batch.token_ids_cpu[
527
528
529
530
531
532
533
534
535
536
537
538
539
540
541
542
                    req_index,
                    start_token_index:end_token_index] = new_token_ids
                self.input_batch.num_tokens_no_spec[
                    req_index] = end_token_index
                # Add spec_token_ids to token_ids_cpu.
                spec_token_ids = (
                    scheduler_output.scheduled_spec_decode_tokens.get(
                        req_id, ()))
                if spec_token_ids:
                    start_index = end_token_index
                    end_token_index += len(spec_token_ids)
                    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] = end_token_index
543
544
545
546
547

        # 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]
548
            self.input_batch.add_request(req_state)
549

550
551
552
553
554
555
        # 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()
556

557
558
559
560
561
562
563
564
565
566
567
568
569
570
571
572
573
574
575
576
    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

577
    def _prepare_inputs(
578
579
        self,
        scheduler_output: "SchedulerOutput",
580
    ) -> tuple[dict[str, Any], bool, torch.Tensor,
581
               Optional[SpecDecodeMetadata], np.ndarray]:
582
583
584
585
586
587
588
        """
        :return: tuple[
            attn_metadata: layer-to-attention_metadata mapping,
            attention_cuda_graphs: whether attention can run in cudagraph
            logits_indices, spec_decode_metadata
        ]
        """
589
590
591
592
593
594
595
        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.
596
        self.input_batch.block_table.commit(num_reqs)
597
598

        # Get the number of scheduled tokens for each request.
599
600
601
602
        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)
603
604
605

        # Get request indices.
        # E.g., [2, 5, 3] -> [0, 0, 1, 1, 1, 1, 1, 2, 2, 2]
606
607
        req_indices = np.repeat(self.arange_np[:num_reqs],
                                num_scheduled_tokens)
608

609
610
611
612
        # 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)
613
614

        # Get positions.
615
        positions_np = self.positions_np[:total_num_scheduled_tokens]
616
617
618
619
        np.add(self.input_batch.num_computed_tokens_cpu[req_indices],
               arange,
               out=positions_np)

620
621
        # Calculate M-RoPE positions.
        # Only relevant for models using M-RoPE (e.g, Qwen2-VL)
622
        if self.uses_mrope:
623
624
            self._calc_mrope_positions(scheduler_output)

625
626
627
628
        # 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.
629
630
        token_indices = (positions_np +
                         req_indices * self.input_batch.token_ids_cpu.shape[1])
631

632
633
634
635
        # 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(),
636
                           0,
637
638
                           torch.from_numpy(token_indices),
                           out=self.input_ids_cpu[:total_num_scheduled_tokens])
639

640
641
642
643
644
645
646
647
648
649
650
651
652
653
654
655
656
657
658
659
660
661
662
        # Calculate the slot mapping for each KV cache group.
        for kv_cache_group_id, kv_cache_group_spec in enumerate(
                self.kv_cache_config.kv_cache_groups):
            block_size = kv_cache_group_spec.kv_cache_spec.block_size
            block_table: BlockTable = self.input_batch.block_table[
                kv_cache_group_id]
            # E.g., [0, 1, 0, 1, 2, 3, 4, 0, 1, 2]
            # -> [0, 0, K, K, K + 1, K + 1, K + 2, 2 * K, 2 * K, 2 * K + 1]
            # where K is the max_num_blocks_per_req and the block size is 2.
            # NOTE(woosuk): We can't simply use `token_indices // block_size`
            # here because M (max_model_len) is not necessarily divisible by
            # block_size.
            block_table_indices = (
                req_indices * block_table.max_num_blocks_per_req +
                positions_np // block_size)
            block_table_cpu = block_table.get_cpu_tensor()
            block_numbers = block_table_cpu.flatten(
            )[block_table_indices].numpy()
            block_offsets = positions_np % block_size
            np.add(
                block_numbers * block_size,
                block_offsets,
                out=block_table.slot_mapping_np[:total_num_scheduled_tokens])
663
664

        # Prepare the attention metadata.
665
        self.query_start_loc_np[0] = 0
666
        self.query_start_loc_np[1:num_reqs + 1] = cu_num_tokens
667

668
669
670
        self.seq_lens_np[:num_reqs] = (
            self.input_batch.num_computed_tokens_cpu[:num_reqs] +
            num_scheduled_tokens)
671
672
673
674

        # 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)
675
        if self.uses_mrope:
676
677
678
679
680
681
682
683
684
            # 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)
685

686
687
688
689
690
691
692
        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)

        # Fill unused with -1. Needed for reshape_and_cache
        self.seq_lens[num_reqs:].fill_(0)
693
694
695
696
        # 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())
697
698
699
700

        query_start_loc = self.query_start_loc[:num_reqs + 1]
        seq_lens = self.seq_lens[:num_reqs]

701
        common_attn_metadata = CommonAttentionMetadata(
702
703
704
705
706
707
            query_start_loc=query_start_loc,
            seq_lens=seq_lens,
            num_reqs=num_reqs,
            num_actual_tokens=total_num_scheduled_tokens,
            max_query_len=max_num_scheduled_tokens,
        )
708

709
        attn_metadata: dict[str, Any] = {}
710
711
712
713
714
715
716
        # 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):

            # Prepare for cascade attention if enabled & beneficial.
            common_prefix_len = 0
717
            builder = self.attn_metadata_builders[kv_cache_group_id]
718
719
720
            if self.cascade_attn_enabled:
                common_prefix_len = self._compute_cascade_attn_prefix_len(
                    num_scheduled_tokens,
721
722
723
                    scheduler_output.
                    num_common_prefix_blocks[kv_cache_group_id],
                    kv_cache_group_spec.kv_cache_spec,
724
                    builder,
725
                )
726

727
728
729
730
731
            attn_metadata_i = (builder.build(
                common_prefix_len=common_prefix_len,
                common_attn_metadata=common_attn_metadata,
            ))

732
733
            for layer_name in kv_cache_group_spec.layer_names:
                attn_metadata[layer_name] = attn_metadata_i
734

735
736
737
738
        attention_cuda_graphs = all(
            b.can_run_in_cudagraph(common_attn_metadata)
            for b in self.attn_metadata_builders)

739
740
        use_spec_decode = len(
            scheduler_output.scheduled_spec_decode_tokens) > 0
741
        if not use_spec_decode:
742
743
744
745
746
            # 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.
747
            logits_indices = query_start_loc[1:] - 1
748
749
750
751
752
753
754
755
756
757
758
759
760
761
            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
762

763
764
765
766
        # Hot-Swap lora model
        if self.lora_config:
            self.set_active_loras(self.input_batch, num_scheduled_tokens)

767
        return (attn_metadata, attention_cuda_graphs, logits_indices,
768
                spec_decode_metadata, num_scheduled_tokens)
769

770
771
772
773
    def _compute_cascade_attn_prefix_len(
        self,
        num_scheduled_tokens: np.ndarray,
        num_common_prefix_blocks: int,
774
775
        kv_cache_spec: KVCacheSpec,
        attn_metadata_builder: AttentionMetadataBuilder,
776
777
778
779
780
781
782
783
784
785
786
787
788
789
790
791
792
793
    ) -> 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.
        """
794
        common_prefix_len = num_common_prefix_blocks * kv_cache_spec.block_size
795
796
797
798
799
800
801
802
803
804
805
806
807
808
809
810
811
812
813
814
815
816
817
818
819
820
821
822
823
824
825
826
827
828
829
830
831
        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]
832
        # Request 3's num_computed_tokens: 3 (i.e., [A, B, C])
833
834
835
836
837
838
839
840
841
842
        # 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.
843
844
845
846
847
848
849
        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))
        assert isinstance(kv_cache_spec, AttentionSpec)
        use_cascade = attn_metadata_builder.use_cascade_attention(
850
851
852
            common_prefix_len=common_prefix_len,
            query_lens=num_scheduled_tokens,
            num_query_heads=self.num_query_heads,
853
            num_kv_heads=kv_cache_spec.num_kv_heads,
854
            use_alibi=self.use_alibi,
855
            use_sliding_window=use_sliding_window,
856
857
858
859
            num_sms=self.num_sms,
        )
        return common_prefix_len if use_cascade else 0

860
861
    def _calc_mrope_positions(self, scheduler_output: "SchedulerOutput"):
        mrope_pos_ptr = 0
862
        for index, req_id in enumerate(self.input_batch.req_ids):
863
864
865
866
867
868
869
870
871
872
873
874
875
876
877
878
879
880
881
882
883
884
885
886
887
888
889
890
891
892
893
894
895
896
897
898
899
            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

900
901
902
903
904
905
906
                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,
                )
907
908
909

                mrope_pos_ptr += completion_part_len

910
911
    def _calc_spec_decode_metadata(
        self,
912
913
914
915
916
917
918
919
920
921
922
923
924
925
926
927
        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
928
929
930
931
932
933

        # 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]
934
935
        logits_indices = np.repeat(
            cu_num_scheduled_tokens - num_sampled_tokens, num_sampled_tokens)
936
        # Step 3. [0, 1, 2, 3, 103, 104, 105, 106, 206, 207, 208]
937
938
939
940
941
942
        logits_indices += arange

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

        # Compute the draft logits indices.
943
944
945
946
        # 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)
947
948
949
950
951
952
953
954
955
956
957
958
959
960
        # [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(
961
962
            self.device, non_blocking=True)

963
964
965
966
967
968
969
970
971
972
973
974
975
976
977
        # 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

978
    def _execute_mm_encoder(self, scheduler_output: "SchedulerOutput"):
979
980
981
982
983
        scheduled_encoder_inputs = scheduler_output.scheduled_encoder_inputs
        if not scheduled_encoder_inputs:
            return

        # Batch the multi-modal inputs.
984
985
        mm_inputs = list[MultiModalKwargs]()
        req_ids_pos = list[tuple[str, int, PlaceholderRange]]()
986
987
        for req_id, encoder_input_ids in scheduled_encoder_inputs.items():
            req_state = self.requests[req_id]
988
989
990
991
992

            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]))
993
994
995
996
997
998
999
1000
1001
1002
1003
1004

        # 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:
1005
1006
            batched_mm_inputs = MultiModalKwargs.batch(
                grouped_mm_inputs, pin_memory=self.pin_memory)
1007
1008
1009
1010
            batched_mm_inputs = MultiModalKwargs.as_kwargs(
                batched_mm_inputs,
                device=self.device,
            )
1011
1012
1013
1014
1015
1016
1017
1018
1019
1020
1021

            # 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)

1022
1023
1024
1025
1026
            sanity_check_mm_encoder_outputs(
                curr_group_outputs,
                expected_num_items=len(grouped_mm_inputs),
            )

1027
1028
            for output in curr_group_outputs:
                encoder_outputs.append(output)
1029
1030

        # Cache the encoder outputs.
1031
1032
1033
1034
        for (req_id, input_id, pos_info), output in zip(
                req_ids_pos,
                encoder_outputs,
        ):
1035
1036
1037
            if req_id not in self.encoder_cache:
                self.encoder_cache[req_id] = {}

1038
1039
1040
1041
1042
1043
            self.encoder_cache[req_id][input_id] = scatter_mm_placeholders(
                output,
                is_embed=pos_info.is_embed,
            )

    def _gather_mm_embeddings(
1044
1045
        self,
        scheduler_output: "SchedulerOutput",
1046
    ) -> list[torch.Tensor]:
1047
        mm_embeds: list[torch.Tensor] = []
1048
        for req_id in self.input_batch.req_ids:
1049
1050
1051
1052
1053
1054
            num_scheduled_tokens = scheduler_output.num_scheduled_tokens[
                req_id]
            req_state = self.requests[req_id]
            num_computed_tokens = req_state.num_computed_tokens
            mm_positions = req_state.mm_positions
            for i, pos_info in enumerate(mm_positions):
1055
1056
                start_pos = pos_info.offset
                num_encoder_tokens = pos_info.length
1057
1058
1059
1060
1061
1062
1063
1064
1065
1066
1067
1068
1069
1070
1071
1072
1073
1074
1075
1076
1077

                # 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]
1078
1079
1080
1081
1082
1083
1084
1085
1086
1087

                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
1088

1089
1090
1091
    def get_model(self) -> nn.Module:
        return self.model

1092
1093
1094
1095
1096
1097
1098
1099
1100
    def apply_grammar_bitmask(
        self,
        scheduler_output: "SchedulerOutput",
        logits: torch.Tensor,
    ):
        grammar_bitmask = scheduler_output.grammar_bitmask
        if grammar_bitmask is None:
            return

1101
1102
1103
1104
1105
1106
1107
1108
1109
        # 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.
1110
        struct_out_req_batch_indices: dict[str, int] = {}
1111
1112
1113
1114
1115
1116
1117
1118
1119
1120
1121
1122
1123
1124
1125
1126
1127
1128
1129
1130
1131
1132
1133
1134
1135
1136
1137
1138
1139
        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
1140

1141
1142
        # Serialization of np.ndarray is much more efficient than a tensor,
        # so we receive it in that format.
1143
1144
        grammar_bitmask = torch.from_numpy(grammar_bitmask)

1145
1146
1147
1148
        # 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(
1149
1150
            logits,
            grammar_bitmask.to(self.device, non_blocking=True),
1151
            indices=out_indices,
1152
1153
        )

1154
1155
1156
1157
1158
1159
1160
    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
1161
        enabled_sp = self.compilation_config.pass_config. \
1162
1163
1164
1165
1166
1167
1168
1169
1170
1171
1172
1173
1174
1175
1176
1177
1178
1179
1180
1181
1182
1183
1184
1185
1186
1187
            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():
                is_scattered = "residual" and is_residual_scattered
                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()
        })

1188
1189
1190
1191
1192
1193
1194
1195
1196
1197
1198
1199
1200
1201
1202
1203
1204
1205
    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,
        )

1206
1207
    def get_dp_padding(self,
                       num_tokens: int) -> tuple[int, Optional[torch.Tensor]]:
1208
1209
        dp_size = self.vllm_config.parallel_config.data_parallel_size
        dp_rank = self.vllm_config.parallel_config.data_parallel_rank
1210
1211
1212
1213
1214
1215
1216
1217
1218

        # 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:
1219
            # Early exit.
1220
            return 0, None
1221
1222
1223
1224

        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()
1225
1226
1227
1228
1229
        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
1230

1231
1232
1233
1234
1235
1236
1237
1238
1239
1240
1241
1242
1243
1244
1245
1246
1247
1248
1249
1250
1251
1252
1253
1254
1255
1256
1257
1258
1259
1260
1261
1262
1263
1264
1265
1266
1267
1268
1269
1270
1271
1272
1273
1274
1275
    def _pool(
        self,
        hidden_states: torch.Tensor,
        num_scheduled_tokens: int,
        num_scheduled_tokens_np: np.ndarray,
        finished_sending: Optional[set[str]],
        finished_recving: Optional[set[str]],
    ) -> 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,
            finished_sending=finished_sending,
            finished_recving=finished_recving,
        )

1276
1277
1278
1279
    @torch.inference_mode()
    def execute_model(
        self,
        scheduler_output: "SchedulerOutput",
1280
        intermediate_tensors: Optional[IntermediateTensors] = None,
1281
    ) -> Union[ModelRunnerOutput, IntermediateTensors]:
1282
        self._update_states(scheduler_output)
1283
        if not scheduler_output.total_num_scheduled_tokens:
Robert Shaw's avatar
Robert Shaw committed
1284
1285
1286
1287
1288
            if not has_kv_transfer_group():
                # Return empty ModelRunnerOutput if there's no work to do.
                return EMPTY_MODEL_RUNNER_OUTPUT

            return self.kv_connector_no_forward(scheduler_output)
1289
1290

        # Prepare the decoder inputs.
1291
        (attn_metadata, attention_cuda_graphs, logits_indices,
1292
1293
         spec_decode_metadata,
         num_scheduled_tokens_np) = (self._prepare_inputs(scheduler_output))
1294
1295
1296
1297
1298
        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.
1299
            num_input_tokens = self.vllm_config.pad_for_cudagraph(
1300
1301
1302
                num_scheduled_tokens)
        else:
            # Eager mode.
1303
1304
1305
            # Pad tokens to multiple of tensor_parallel_size when
            # enabled collective fusion for SP
            tp_size = self.vllm_config.parallel_config.tensor_parallel_size
1306
            if self.compilation_config.pass_config. \
1307
1308
1309
1310
                enable_sequence_parallelism and tp_size > 1:
                num_input_tokens = round_up(num_scheduled_tokens, tp_size)
            else:
                num_input_tokens = num_scheduled_tokens
1311

1312
        # Padding for DP
1313
1314
        num_pad, num_tokens_across_dp = self.get_dp_padding(num_input_tokens)
        num_input_tokens += num_pad
1315

1316
1317
1318
1319
1320
1321
1322
1323
1324
        # _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 = []

1325
        if self.is_multimodal_model and get_pp_group().is_first_rank:
1326
1327
1328
1329
            # 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.
            input_ids = self.input_ids[:num_scheduled_tokens]
1330
            if mm_embeds:
1331
                inputs_embeds = self.model.get_input_embeddings(
1332
                    input_ids, mm_embeds)
1333
1334
1335
1336
1337
1338
            else:
                inputs_embeds = self.model.get_input_embeddings(input_ids)
            # TODO(woosuk): Avoid the copy. Optimize.
            self.inputs_embeds[:num_scheduled_tokens].copy_(inputs_embeds)
            inputs_embeds = self.inputs_embeds[:num_input_tokens]
            input_ids = None
1339
        else:
1340
1341
1342
1343
1344
1345
            # 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
1346
1347
1348
1349
        if self.uses_mrope:
            positions = self.mrope_positions[:, :num_input_tokens]
        else:
            positions = self.positions[:num_input_tokens]
1350

1351
1352
1353
        if get_pp_group().is_first_rank:
            intermediate_tensors = None
        else:
1354
1355
            intermediate_tensors = self.sync_and_slice_intermediate_tensors(
                num_input_tokens, intermediate_tensors, True)
1356

1357
1358
1359
1360
1361
        # 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

1362
        # Run the model.
1363
        # Use persistent buffers for CUDA graphs.
1364
1365
1366
1367
1368
1369
1370
        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,
        ):
Robert Shaw's avatar
Robert Shaw committed
1371
1372
1373
            self.maybe_setup_kv_connector(scheduler_output)

            model_output = self.model(
1374
                input_ids=input_ids,
1375
                positions=positions,
1376
                intermediate_tensors=intermediate_tensors,
1377
                inputs_embeds=inputs_embeds,
1378
            )
1379

Robert Shaw's avatar
Robert Shaw committed
1380
1381
1382
1383
            self.maybe_wait_for_kv_save()
            finished_sending, finished_recving = (
                self.get_finished_kv_transfers(scheduler_output))

1384
        if self.use_aux_hidden_state_outputs:
Robert Shaw's avatar
Robert Shaw committed
1385
            hidden_states, aux_hidden_states = model_output
1386
        else:
Robert Shaw's avatar
Robert Shaw committed
1387
            hidden_states = model_output
1388
1389
            aux_hidden_states = None

1390
1391
1392
1393
1394
1395
1396
        # 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
1397
        if not get_pp_group().is_last_rank:
1398
            # For mid-pipeline stages, return the hidden states.
1399
1400
1401
1402
1403
1404
1405
            if not broadcast_pp_output:
                return hidden_states
            assert isinstance(hidden_states, IntermediateTensors)
            get_pp_group().send_tensor_dict(hidden_states.tensors,
                                            all_gather_group=get_tp_group())
            logits = None
        else:
1406
1407
1408
1409
1410
            if self.input_batch.pooling_params:
                return self._pool(hidden_states, num_scheduled_tokens,
                                  num_scheduled_tokens_np, finished_sending,
                                  finished_recving)

1411
1412
1413
1414
1415
1416
1417
1418
1419
1420
            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"]
1421

1422
1423
1424
1425
        # Apply structured output bitmasks if present
        if scheduler_output.grammar_bitmask is not None:
            self.apply_grammar_bitmask(scheduler_output, logits)

1426
        # Sample the next token and get logprobs if needed.
1427
        sampling_metadata = self.input_batch.sampling_metadata
1428
        if spec_decode_metadata is None:
1429
            sampler_output = self.sampler(
1430
1431
1432
1433
                logits=logits,
                sampling_metadata=sampling_metadata,
            )
        else:
1434
1435
1436
1437
            # 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.
1438
            assert logits is not None
1439
            bonus_logits = logits[spec_decode_metadata.bonus_logits_indices]
1440
            sampler_output = self.sampler(
1441
                logits=bonus_logits,
1442
1443
1444
                sampling_metadata=sampling_metadata,
            )
            bonus_token_ids = sampler_output.sampled_token_ids
1445

1446
1447
1448
            # 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.
1449
            target_logits = logits[spec_decode_metadata.target_logits_indices]
1450
            output_token_ids = self.rejection_sampler(
1451
                spec_decode_metadata,
1452
                None,  # draft_probs
1453
                target_logits,
1454
                bonus_token_ids,
1455
1456
                sampling_metadata,
            )
1457
            sampler_output.sampled_token_ids = output_token_ids
1458

1459
1460
1461
1462
        num_nans_in_logits = {}
        if envs.VLLM_COMPUTE_NANS_IN_LOGITS:
            num_nans_in_logits = self._get_nans_in_logits(logits)

1463
1464
        # TODO(woosuk): The following loop can be slow since it iterates over
        # the requests one by one. Optimize.
1465
1466
        discard_sampled_tokens_req_indices = []
        for i, req_id in enumerate(self.input_batch.req_ids):
1467
1468
1469
            req_state = self.requests[req_id]
            seq_len = (req_state.num_computed_tokens +
                       scheduler_output.num_scheduled_tokens[req_id])
1470
            if seq_len < req_state.num_tokens:
1471
                # Ignore the sampled token for partial prefills.
1472
                # Rewind the generator state as if the token was not sampled.
1473
                # This relies on cuda-specific torch-internal impl details
1474
1475
1476
1477
1478
1479
                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)
1480

1481
1482
        # NOTE: GPU -> CPU Sync happens here.
        # Move as many CPU operations as possible before this sync point.
1483
1484
1485
1486
1487
1488
        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(
1489
            hidden_states[:num_scheduled_tokens],
1490
1491
1492
            scheduler_output,
        )

1493
        # Get the valid generated tokens.
1494
1495
1496
        sampled_token_ids = sampler_output.sampled_token_ids
        max_gen_len = sampled_token_ids.shape[-1]
        if max_gen_len == 1:
1497
            # No spec decode tokens.
1498
1499
            valid_sampled_token_ids = sampled_token_ids.tolist()
        else:
1500
            # Includes spec decode tokens.
1501
            valid_sampled_token_ids = self.rejection_sampler.parse_output(
1502
1503
1504
                sampled_token_ids,
                self.input_batch.vocab_size,
            )
1505
1506
1507
        # Mask out the sampled tokens that should not be sampled.
        for i in discard_sampled_tokens_req_indices:
            valid_sampled_token_ids[i].clear()
1508

1509
1510
1511
1512
1513
1514
1515
1516
1517
1518
1519
1520
1521
1522
1523
1524
1525
1526
1527
1528
1529
1530
1531
1532
        # 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)

1533
        if not self.speculative_config:
1534
            # Speculative decoding is not enabled.
1535
            spec_token_ids = None
1536
1537
1538
1539
1540
1541
1542
1543
1544
1545
1546
1547
1548
1549
1550
1551
1552
1553
1554
1555
1556
1557
1558
1559
1560
1561
1562
1563
1564
1565
1566
1567
1568
1569
1570
1571
1572
1573
1574
1575
1576
1577
1578
1579
        else:
            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,
                attn_metadata,
            )

        # Clear KVConnector state after all KVs are generated.
        if has_kv_transfer_group():
            get_kv_transfer_group().clear_connector_metadata()

        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=[],
            finished_sending=finished_sending,
            finished_recving=finished_recving,
            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],
        attn_metadata: dict[str, Any],
    ) -> list[list[int]]:
        num_scheduled_tokens = scheduler_output.total_num_scheduled_tokens
        if self.speculative_config.method == "ngram":
1580
            assert isinstance(self.drafter, NgramProposer)
1581
1582
            spec_token_ids = self.propose_ngram_draft_token_ids(
                sampled_token_ids)
1583
1584
        elif self.speculative_config.method == "medusa":
            assert isinstance(self.drafter, MedusaProposer)
1585
1586
            if sample_hidden_states.shape[0] == len(sampled_token_ids):
                # The input to the target model does not include draft tokens.
1587
1588
1589
1590
1591
1592
                hidden_states = sample_hidden_states
            else:
                indices = []
                offset = 0
                for num_draft, tokens in zip(
                        spec_decode_metadata.num_draft_tokens,
1593
                        sampled_token_ids):
1594
1595
                    indices.append(offset + len(tokens) - 1)
                    offset += num_draft + 1
1596
                indices = torch.tensor(indices, device=self.device)
1597
1598
1599
1600
1601
1602
                hidden_states = sample_hidden_states[indices]

            spec_token_ids = self.drafter.propose(
                target_hidden_states=hidden_states,
                sampling_metadata=sampling_metadata,
            )
1603
        elif self.speculative_config.use_eagle():
1604
1605
1606
            assert isinstance(self.drafter, EagleProposer)
            # TODO(woosuk): Refactor the loop.
            next_token_ids: list[int] = []
1607
            for i, token_ids in enumerate(sampled_token_ids):
1608
1609
1610
1611
1612
1613
1614
1615
1616
1617
1618
1619
                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)
1620
1621
1622
1623
1624
1625
1626
            next_token_ids = torch.tensor(next_token_ids,
                                          dtype=torch.int32,
                                          device=self.device)
            # At this moment, we assume all eagle layers belong to the same KV
            # cache group, thus using the same attention metadata.
            eagle_attn_metadata = attn_metadata[
                self.drafter.attn_layer_names[0]]
1627

Jiayi Yao's avatar
Jiayi Yao committed
1628
1629
1630
1631
1632
1633
            # NOTE: deepseek_mtp uses MLA which does not have `block_table`
            if hasattr(eagle_attn_metadata, "block_table"):
                block_table = eagle_attn_metadata.block_table
            else:
                block_table = None

1634
1635
1636
            if spec_decode_metadata is None:
                # input_ids can be None for multimodal models.
                target_token_ids = self.input_ids[:num_scheduled_tokens]
1637
1638
                # TODO(woosuk): Support M-RoPE.
                target_positions = self.positions[:num_scheduled_tokens]
1639
                if self.use_aux_hidden_state_outputs:
1640
1641
1642
                    target_hidden_states = torch.cat(
                        [h[:num_scheduled_tokens] for h in aux_hidden_states],
                        dim=-1)
1643
1644
                else:
                    target_hidden_states = hidden_states[:num_scheduled_tokens]
1645
1646
                target_slot_mapping = eagle_attn_metadata.slot_mapping
                cu_num_tokens = eagle_attn_metadata.query_start_loc
1647
1648
1649
1650
            else:
                # TODO(woosuk): Refactor this.
                num_draft_tokens = spec_decode_metadata.num_draft_tokens
                num_rejected_tokens = [
1651
                    n + 1 - len(sampled_token_ids[i]) if n > 0 else 0
1652
1653
                    for i, n in enumerate(num_draft_tokens)
                ]
1654
                num_rejected_tokens_tensor = async_tensor_h2d(
1655
1656
                    num_rejected_tokens,
                    dtype=torch.int32,
1657
1658
1659
                    target_device=self.device,
                    pin_memory=True)
                num_tokens = num_scheduled_tokens - sum(num_rejected_tokens)
1660
                cu_num_tokens, token_indices = self.drafter.prepare_inputs(
1661
                    eagle_attn_metadata.query_start_loc,
1662
1663
                    num_rejected_tokens_tensor,
                    num_tokens,
1664
1665
                )
                target_token_ids = self.input_ids[token_indices]
1666
1667
                # TODO(woosuk): Support M-RoPE.
                target_positions = self.positions[token_indices]
1668
                if self.use_aux_hidden_state_outputs:
1669
1670
                    target_hidden_states = torch.cat(
                        [h[token_indices] for h in aux_hidden_states], dim=-1)
1671
1672
                else:
                    target_hidden_states = hidden_states[token_indices]
1673
1674
                target_slot_mapping = eagle_attn_metadata.slot_mapping[
                    token_indices]
1675
            draft_token_ids = self.drafter.propose(
1676
1677
1678
1679
1680
1681
                target_token_ids=target_token_ids,
                target_positions=target_positions,
                target_hidden_states=target_hidden_states,
                target_slot_mapping=target_slot_mapping,
                next_token_ids=next_token_ids,
                cu_num_tokens=cu_num_tokens,
Jiayi Yao's avatar
Jiayi Yao committed
1682
                block_table=block_table,
1683
1684
1685
                sampling_metadata=sampling_metadata,
            )
            spec_token_ids = draft_token_ids.tolist()
1686
        return spec_token_ids
1687

Robert Shaw's avatar
Robert Shaw committed
1688
1689
1690
1691
1692
1693
1694
1695
1696
1697
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
1723
1724
1725
1726
1727
1728
1729
1730
1731
1732
1733
    def kv_connector_no_forward(
            self, scheduler_output: "SchedulerOutput") -> ModelRunnerOutput:
        # KV send/recv even if no work to do.
        with set_forward_context(None, self.vllm_config):
            self.maybe_setup_kv_connector(scheduler_output)
            finished_sending, finished_recving = (
                self.get_finished_kv_transfers(scheduler_output))

        if not finished_sending and not finished_recving:
            return EMPTY_MODEL_RUNNER_OUTPUT

        output = copy.copy(EMPTY_MODEL_RUNNER_OUTPUT)
        output.finished_sending = finished_sending
        output.finished_recving = finished_recving
        return output

    @staticmethod
    def maybe_setup_kv_connector(scheduler_output: "SchedulerOutput"):
        # Update KVConnector with the KVConnector metadata forward().
        if has_kv_transfer_group():
            kv_connector = get_kv_transfer_group()
            assert isinstance(kv_connector, KVConnectorBase_V1)
            assert scheduler_output.kv_connector_metadata is not None
            kv_connector.bind_connector_metadata(
                scheduler_output.kv_connector_metadata)

            # Background KV cache transfers happen here.
            # These transfers are designed to be async and the requests
            # involved may be disjoint from the running requests.
            # Do this here to save a collective_rpc.
            kv_connector.start_load_kv(get_forward_context())

    @staticmethod
    def maybe_wait_for_kv_save() -> None:
        if has_kv_transfer_group():
            get_kv_transfer_group().wait_for_save()

    @staticmethod
    def get_finished_kv_transfers(
        scheduler_output: "SchedulerOutput",
    ) -> tuple[Optional[set[str]], Optional[set[str]]]:
        if has_kv_transfer_group():
            return get_kv_transfer_group().get_finished(
                scheduler_output.finished_req_ids)
        return None, None

1734
    def propose_ngram_draft_token_ids(
1735
        self,
1736
1737
        sampled_token_ids: list[list[int]],
    ) -> list[list[int]]:
1738
        # TODO(woosuk): Optimize.
1739
        draft_token_ids: list[list[int]] = []
1740
1741
1742
        for i, sampled_ids in enumerate(sampled_token_ids):
            num_sampled_ids = len(sampled_ids)
            if not num_sampled_ids:
1743
1744
1745
1746
                # Skip speculative decoding.
                draft_token_ids.append([])
                continue

1747
1748
            # Skip requests that require sampling parameters that are not
            # supported with speculative decoding.
1749
            req_id = self.input_batch.req_ids[i]
1750
            if req_id in self.input_batch.spec_decode_unsupported_reqs:
1751
1752
1753
                draft_token_ids.append([])
                continue

1754
1755
            num_tokens = self.input_batch.num_tokens_no_spec[i]
            if num_tokens >= self.max_model_len:
1756
1757
1758
1759
                # Skip requests that have already reached the max model length.
                draft_token_ids.append([])
                continue

1760
            drafter_output = self.drafter.propose(
1761
                self.input_batch.token_ids_cpu[i, :num_tokens])
1762
1763
1764
1765
1766
1767
            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

1768
1769
1770
    def load_model(self) -> None:
        logger.info("Starting to load model %s...", self.model_config.model)
        with DeviceMemoryProfiler() as m:  # noqa: SIM117
1771
            time_before_load = time.perf_counter()
1772
1773
1774
1775
1776
1777
1778
1779
1780
1781
1782
1783
            model_loader = get_model_loader(self.load_config)
            if not hasattr(self, "model"):
                logger.info("Loading model from scratch...")
                self.model = model_loader.load_model(
                    vllm_config=self.vllm_config,
                    model_config=self.model_config)
            else:
                logger.info(
                    "Model was already initialized. Loading weights inplace..."
                )
                model_loader.load_weights(self.model,
                                          model_config=self.model_config)
1784
1785
            if has_step_pooler(self.model):
                self.input_batch.logits_processing_needs_token_ids = True
1786
1787
1788
1789
1790
1791
            if self.lora_config:
                self.model = self.load_lora_model(self.model,
                                                  self.model_config,
                                                  self.scheduler_config,
                                                  self.lora_config,
                                                  self.device)
1792
1793
1794
            if hasattr(self, "drafter"):
                logger.info("Loading drafter model...")
                self.drafter.load_model(self.model)
1795
1796
1797
            if self.use_aux_hidden_state_outputs:
                self.model.set_aux_hidden_state_layers(
                    self.model.get_eagle3_aux_hidden_state_layers())
1798
            time_after_load = time.perf_counter()
1799
        self.model_memory_usage = m.consumed_memory
1800
1801
        logger.info("Model loading took %.4f GiB and %.6f seconds",
                    self.model_memory_usage / GiB_bytes,
1802
                    time_after_load - time_before_load)
1803
        prepare_communication_buffer_for_model(self.model)
1804

1805
1806
1807
1808
1809
1810
1811
1812
1813
1814
        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,
            )

1815
1816
1817
1818
1819
1820
1821
1822
1823
    def save_tensorized_model(
        self,
        tensorizer_config: "TensorizerConfig",
    ) -> None:
        TensorizerLoader.save_model(
            self.model,
            tensorizer_config=tensorizer_config,
        )

1824
1825
1826
1827
    def _get_prompt_logprobs_dict(
        self,
        hidden_states: torch.Tensor,
        scheduler_output: "SchedulerOutput",
1828
    ) -> dict[str, Optional[LogprobsTensors]]:
1829
1830
1831
1832
        num_prompt_logprobs_dict = self.input_batch.num_prompt_logprobs
        if not num_prompt_logprobs_dict:
            return {}

1833
        in_progress_dict = self.input_batch.in_progress_prompt_logprobs_cpu
1834
        prompt_logprobs_dict: dict[str, Optional[LogprobsTensors]] = {}
1835
1836
1837
1838
1839
1840
1841
1842
1843
1844
1845
1846
1847
1848

        # 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)

1849
1850
1851
1852
1853
1854
1855
1856
1857
            # 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

1858
            # Determine number of logits to retrieve.
1859
1860
            start_idx = request.num_computed_tokens
            start_tok = start_idx + 1
1861
            num_remaining_tokens = num_prompt_tokens - start_tok
1862
            if num_tokens <= num_remaining_tokens:
1863
                # This is a chunk, more tokens remain.
1864
1865
1866
                # 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.
1867
1868
1869
1870
1871
                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)
1872
1873
1874
1875
1876
1877
1878
                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
1879
1880
1881
1882
1883
1884
1885
1886
1887
1888
1889
1890
1891
1892
1893

            # 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.
1894
1895
            logprobs = self.sampler.compute_logprobs(logits)
            token_ids, logprobs, ranks = self.sampler.gather_logprobs(
1896
1897
1898
                logprobs, num_prompt_logprobs, tgt_token_ids)

            # Transfer GPU->CPU async.
1899
1900
1901
1902
1903
1904
1905
            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)
1906
1907
1908
1909
1910

        # 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]
1911
            del in_progress_dict[req_id]
1912
1913

        # Must synchronize the non-blocking GPU->CPU transfers.
1914
        if prompt_logprobs_dict:
1915
            self._sync_device()
1916
1917
1918

        return prompt_logprobs_dict

1919
1920
1921
1922
1923
1924
1925
1926
1927
1928
1929
1930
1931
1932
1933
1934
1935
1936
1937
1938
    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 {}

1939
1940
1941
1942
1943
1944
1945
1946
1947
1948
1949
1950
1951
1952
1953
1954
1955
1956
1957
1958
1959
1960
1961
1962
1963
1964
1965
1966
1967
    @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
         - during DP rank dummy run 
        """
        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)

1968
1969
1970
1971
    @torch.inference_mode()
    def _dummy_run(
        self,
        num_tokens: int,
1972
        capture_attn_cudagraph: bool = False,
1973
1974
        skip_eplb: bool = False,
        is_profile: bool = False,
1975
    ) -> tuple[torch.Tensor, torch.Tensor]:
1976

1977
        # Padding for DP
1978
1979
        num_pad, num_tokens_across_dp = self.get_dp_padding(num_tokens)
        num_tokens += num_pad
1980

1981
1982
1983
1984
1985
        # 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
1986
        num_reqs = min(num_tokens, max_num_reqs)
1987
1988
1989
1990
1991
1992
1993
        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)
1994

1995
1996
1997
1998
        attn_metadata: Optional[dict[str, Any]] = None
        if capture_attn_cudagraph:
            attn_metadata = {}

1999
            query_start_loc = self.query_start_loc[:num_reqs + 1]
2000
2001
2002
2003
2004
            # 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)
2005
2006
2007
            seq_lens = self.seq_lens[:num_reqs]

            common_attn_metadata = CommonAttentionMetadata(
2008
2009
2010
2011
2012
2013
                query_start_loc=query_start_loc,
                seq_lens=seq_lens,
                num_reqs=num_reqs,
                num_actual_tokens=num_tokens,
                max_query_len=num_tokens,
            )
2014

2015
2016
            for kv_cache_group_id, kv_cache_group_spec in enumerate(
                    self.kv_cache_config.kv_cache_groups):
2017
2018
2019
2020

                attn_metadata_i = self.attn_metadata_builders[
                    kv_cache_group_id].build_for_cudagraph_capture(
                        common_attn_metadata)
2021
2022
                for layer_name in kv_cache_group_spec.layer_names:
                    attn_metadata[layer_name] = attn_metadata_i
2023

2024
2025
2026
2027
2028
2029
2030
2031
2032
2033
2034
2035
2036
2037
2038
2039
2040
2041
2042
2043
2044
2045
2046
        with self.maybe_dummy_run_with_lora(self.lora_config,
                                            num_scheduled_tokens):
            model = self.model
            if self.is_multimodal_model:
                input_ids = None
                inputs_embeds = self.inputs_embeds[:num_tokens]
            else:
                input_ids = self.input_ids[:num_tokens]
                inputs_embeds = None
            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))
2047
2048
2049

                intermediate_tensors = self.sync_and_slice_intermediate_tensors(
                    num_tokens, None, False)
2050

2051
            with self.maybe_randomize_inputs(input_ids), set_forward_context(
2052
2053
2054
2055
                    attn_metadata,
                    self.vllm_config,
                    num_tokens=num_tokens,
                    num_tokens_across_dp=num_tokens_across_dp):
2056
                outputs = model(
2057
2058
2059
2060
2061
                    input_ids=input_ids,
                    positions=positions,
                    intermediate_tensors=intermediate_tensors,
                    inputs_embeds=inputs_embeds,
                )
2062
2063
2064
2065
            if self.use_aux_hidden_state_outputs:
                hidden_states, _ = outputs
            else:
                hidden_states = outputs
2066

2067
            if self.speculative_config and self.speculative_config.use_eagle():
2068
2069
2070
                assert isinstance(self.drafter, EagleProposer)
                self.drafter.dummy_run(num_tokens)

2071
2072
2073
2074
2075
2076
2077
2078
2079
2080
        # 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)

2081
        logit_indices = np.cumsum(num_scheduled_tokens) - 1
2082
        return hidden_states, hidden_states[logit_indices]
2083
2084
2085
2086
2087
2088

    @torch.inference_mode()
    def _dummy_sampler_run(
        self,
        hidden_states: torch.Tensor,
    ) -> torch.Tensor:
2089
2090
2091
2092
        # 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)
2093
2094
2095
2096
2097
2098
2099
2100
2101
2102
2103
2104
2105
2106
2107
2108
2109
2110
2111
2112
2113
2114
2115

        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={},
2116
            logitsprocs=LogitsProcessorManager(),
2117
        )
2118
        try:
2119
2120
            sampler_output = self.sampler(logits=logits,
                                          sampling_metadata=dummy_metadata)
2121
2122
2123
2124
2125
2126
2127
2128
2129
        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
2130
        if self.speculative_config:
2131
2132
2133
2134
2135
2136
2137
2138
2139
2140
2141
2142
2143
2144
2145
2146
2147
2148
2149
2150
2151
2152
2153
2154
2155
2156
            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,
            )
2157
        return sampler_output
2158

2159
2160
2161
2162
2163
2164
2165
2166
2167
2168
2169
2170
2171
2172
2173
2174
2175
2176
2177
2178
2179
2180
2181
2182
2183
2184
2185
2186
2187
2188
2189
2190
2191
2192
2193
2194
2195
2196
2197
2198
2199
2200
    @torch.inference_mode()
    def _dummy_pooler_run(
        self,
        hidden_states: torch.Tensor,
    ) -> torch.Tensor:

        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

        dummy_metadata = PoolingMetadata(
            prompt_lens=torch.tensor([h.shape[0] for h in hidden_states_list],
                                     device=self.device),
            prompt_token_ids=torch.zeros((num_reqs, req_num_tokens),
                                         dtype=torch.int32,
                                         device=self.device),
            pooling_params=[PoolingParams()] * num_reqs)

        try:
            pooler_output = self.model.pooler(hidden_states=hidden_states_list,
                                              pooling_metadata=dummy_metadata)
        except RuntimeError as e:
            if 'out of memory' in str(e):
                raise RuntimeError(
                    "CUDA out of memory occurred when warming up pooler 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
        return pooler_output

2201
    def profile_run(self) -> None:
2202
        # Profile with multimodal encoder & encoder cache.
2203
2204
2205
        # TODO: handle encoder-decoder models once we support them.
        if (self.is_multimodal_model and self.max_num_encoder_input_tokens > 0
                and self.encoder_cache_size > 0):
2206

2207
            # NOTE: Currently model is profiled with a single non-text
2208
2209
            # modality with the max possible input tokens even when
            # it supports multiple.
2210
2211
            max_tokens_by_modality_dict = self.mm_registry \
                .get_max_tokens_per_item_by_nonzero_modality(self.model_config)
2212
2213
2214
2215
            dummy_data_modality, max_tokens_per_mm_item = max(
                max_tokens_by_modality_dict.items(), key=lambda item: item[1])

            # Check how many items of this modality can be supported by
2216
2217
2218
2219
2220
2221
            # the encoder budget.
            encoder_budget = min(self.max_num_encoder_input_tokens,
                                 self.encoder_cache_size)

            max_num_mm_items_encoder_budget = cdiv(encoder_budget,
                                                   max_tokens_per_mm_item)
2222
2223
2224

            # Check how many items of this modality can be supported by
            # the decoder budget.
2225
2226
            max_mm_items_per_req = self.mm_registry.get_mm_limits_per_prompt(
                self.model_config)[dummy_data_modality]
2227
2228
2229
2230
2231
2232
2233
2234
2235
2236

            # NOTE: We do not consider max_num_batched_tokens on purpose
            # because the multimodal embeddings can be generated in advance
            # and chunked prefilled.
            max_num_mm_items_decoder_budget = self.max_num_reqs * \
                max_mm_items_per_req

            max_num_mm_items = min(max_num_mm_items_encoder_budget,
                                   max_num_mm_items_decoder_budget)

2237
2238
2239
2240
2241
2242
            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_num_mm_items, dummy_data_modality)

            # Create dummy batch of multimodal inputs.
2243
            dummy_mm_kwargs = self.mm_registry.get_decoder_dummy_data(
2244
2245
                model_config=self.model_config,
                seq_len=self.max_num_tokens,
2246
2247
2248
2249
                mm_counts={
                    dummy_data_modality: 1
                },
            ).multi_modal_data
2250

2251
            batched_dummy_mm_inputs = MultiModalKwargs.batch(
2252
2253
                [dummy_mm_kwargs] * max_num_mm_items,
                pin_memory=self.pin_memory)
2254
            batched_dummy_mm_inputs = MultiModalKwargs.as_kwargs(
2255
2256
2257
                batched_dummy_mm_inputs,
                device=self.device,
            )
2258
2259
2260
2261

            # Run multimodal encoder.
            dummy_encoder_outputs = self.model.get_multimodal_embeddings(
                **batched_dummy_mm_inputs)
2262
2263
2264
2265
2266

            sanity_check_mm_encoder_outputs(
                dummy_encoder_outputs,
                expected_num_items=max_num_mm_items,
            )
2267
2268
2269
2270

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

2271
        # Add `is_profile` here to pre-allocate communication buffers
2272
        hidden_states, last_hidden_states \
2273
            = self._dummy_run(self.max_num_tokens, is_profile=True)
2274
        if get_pp_group().is_last_rank:
2275
2276
2277
2278
            if self.is_pooling_model:
                output = self._dummy_pooler_run(hidden_states)
            else:
                output = self._dummy_sampler_run(last_hidden_states)
2279
        else:
2280
            output = None
2281
        self._sync_device()
2282
        del hidden_states, output
2283
        self.encoder_cache.clear()
2284
        gc.collect()
2285
2286

    def capture_model(self) -> None:
2287
2288
        if not self.use_cuda_graph:
            logger.warning(
2289
2290
2291
                "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)
2292
2293
            return

2294
2295
        compilation_counter.num_gpu_runner_capture_triggers += 1

2296
2297
2298
        start_time = time.perf_counter()
        start_free_gpu_memory = torch.cuda.mem_get_info()[0]

2299
2300
2301
        # 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.
2302
        with graph_capture(device=self.device):
2303
            full_cg = self.full_cuda_graph
2304
2305
2306
2307
2308
2309
            # Only rank 0 should print progress bar during capture
            compilation_cases = reversed(self.cudagraph_batch_sizes)
            if is_global_first_rank():
                compilation_cases = tqdm(list(compilation_cases),
                                         desc="Capturing CUDA graph shapes")
            for num_tokens in compilation_cases:
2310
                # We skip EPLB here since we don't want to record dummy metrics
2311
2312
                for _ in range(
                        self.compilation_config.cudagraph_num_of_warmups):
2313
2314
2315
2316
2317
2318
                    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)
2319
2320
2321
2322
2323
2324
2325
2326

        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))
2327

2328
2329
2330
2331
2332
2333
2334
2335
2336
2337
    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
Chen Zhang's avatar
Chen Zhang committed
2338
2339
2340
2341
2342
2343
2344
2345
2346
2347
2348
2349
2350
2351
2352
2353
2354
2355
2356
2357
2358
2359
2360
2361
2362
            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):
                attn_backend_i = Mamba2AttentionBackend
            else:
                raise ValueError(
                    f"Unknown KV cache spec type: {type(kv_cache_spec)}")
2363
2364
2365

            block_table_i = self.input_batch.block_table[i]
            attn_metadata_builder_i = attn_backend_i.get_builder_cls()(
2366
2367
2368
2369
2370
2371
2372
2373
2374
2375
2376
2377
                weakref.proxy(self),
                kv_cache_spec,
                block_table_i,
            )

            if (self.full_cuda_graph
                    and not attn_metadata_builder_i.full_cudagraph_supported):
                raise ValueError(
                    f"Full CUDAGraph not supported for "
                    f"{attn_backend_i.__name__}. Turn off CompilationConfig."
                    f"full_cuda_graph or use a different attention backend.")

2378
2379
2380
            self.attn_backends.append(attn_backend_i)
            self.attn_metadata_builders.append(attn_metadata_builder_i)

2381
2382
2383
2384
2385
2386
2387
2388
2389
2390
2391
2392
2393
2394
2395
2396
2397
2398
2399
2400
2401
2402
2403
2404
2405
2406
2407
    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,
2408
                is_spec_decode=bool(self.vllm_config.speculative_config),
2409
2410
            )

2411
2412
    def _allocate_kv_cache_tensors(
            self, kv_cache_config: KVCacheConfig) -> dict[str, torch.Tensor]:
2413
        """
2414
2415
2416
        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.

2417
        Args:
2418
            kv_cache_config: The KV cache config
2419
        Returns:
2420
            dict[str, torch.Tensor]: A map between layer names to their
2421
2422
2423
2424
2425
2426
2427
2428
2429
2430
2431
2432
2433
2434
2435
2436
2437
2438
2439
2440
2441
2442
            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]:
2443
        """
2444
        Reshape the KV cache tensors to the desired shape and dtype.
2445

2446
        Args:
2447
2448
            kv_cache_config: The KV cache config
            kv_cache_raw_tensors: The KV cache buffer of each layer, with
2449
2450
            correct size but uninitialized shape.
        Returns:
2451
            Dict[str, torch.Tensor]: A map between layer names to their
2452
2453
            corresponding memory buffer for KV cache.
        """
2454
        kv_caches: dict[str, torch.Tensor] = {}
2455
        has_attn, has_mamba = False, False
2456
2457
2458
2459
2460
2461
2462
2463
        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)
2464
                if isinstance(kv_cache_spec, AttentionSpec):
2465
                    has_attn = True
2466
                    kv_cache_shape = self.attn_backends[i].get_kv_cache_shape(
2467
2468
2469
                        num_blocks, kv_cache_spec.block_size,
                        kv_cache_spec.num_kv_heads, kv_cache_spec.head_size)
                    dtype = kv_cache_spec.dtype
2470
2471
2472
2473
2474
2475
2476
2477
2478
2479
2480
2481
2482
2483
2484
2485
2486
2487
2488
2489
                    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))
                    ]
2490
2491
2492
                    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
2493
                elif isinstance(kv_cache_spec, MambaSpec):
2494
                    has_mamba = True
Chen Zhang's avatar
Chen Zhang committed
2495
2496
                    raw_tensor = kv_cache_raw_tensors[layer_name]
                    dtype = kv_cache_spec.dtype
2497
2498
                    num_element_per_page = (kv_cache_spec.page_size_bytes //
                                            get_dtype_size(dtype))
Chen Zhang's avatar
Chen Zhang committed
2499
                    state_tensors = []
2500
                    storage_offset = 0
Chen Zhang's avatar
Chen Zhang committed
2501
2502
                    for shape in kv_cache_spec.shapes:
                        target_shape = (num_blocks, *shape)
2503
2504
2505
2506
2507
2508
2509
2510
                        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
2511
                        state_tensors.append(tensor)
2512
2513
2514
                        storage_offset += stride[0]

                    kv_caches[layer_name] = state_tensors
2515
                else:
2516
                    raise NotImplementedError
2517
2518
2519
2520
2521

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

2522
2523
        return kv_caches

2524
2525
2526
2527
2528
2529
2530
2531
2532
2533
2534
2535
2536
2537
2538
2539
2540
2541
2542
2543
2544
2545
2546
2547
2548
2549
2550
2551
2552
2553
2554
    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")

2555
2556
2557
2558
2559
2560
2561
2562
    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:
2563
            Dict[str, torch.Tensor]: A map between layer names to their
2564
2565
2566
2567
2568
2569
2570
            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)
2571

2572
2573
2574
2575
2576
2577
2578
2579
2580
        # 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,
            )

2581
2582
2583
        bind_kv_cache(kv_caches,
                      self.compilation_config.static_forward_context,
                      self.kv_caches)
2584
2585
2586
2587
2588
2589
2590
2591
2592
2593
2594
2595
2596
2597
        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)

2598
2599
2600
2601
2602
2603
        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
2604
2605
2606
        if has_kv_transfer_group():
            get_kv_transfer_group().register_kv_caches(kv_caches)

2607
    def get_kv_cache_spec(self) -> dict[str, KVCacheSpec]:
2608
        """
2609
        Generates the KVCacheSpec by parsing the kv cache format from each
2610
2611
        Attention module in the static forward context.
        Returns:
2612
            KVCacheSpec: A dictionary mapping layer names to their KV cache
2613
2614
2615
2616
            format. Layers that do not need KV cache are not included.
        """

        block_size = self.vllm_config.cache_config.block_size
2617
        use_mla = self.vllm_config.model_config.use_mla
2618
        kv_cache_spec: dict[str, KVCacheSpec] = {}
Chen Zhang's avatar
Chen Zhang committed
2619
2620
        attn_layers = get_layers_from_vllm_config(self.vllm_config, Attention)
        for layer_name, attn_module in attn_layers.items():
2621
2622
2623
2624
2625
2626
2627
2628
2629
2630
2631
2632
            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

2633
            # TODO: Support other attention modules, e.g., cross-attention
2634
            if attn_module.attn_type == AttentionType.DECODER:
2635
2636
2637
2638
2639
2640
2641
2642
2643
2644
2645
2646
2647
2648
2649
                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)
                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)
2650
2651
2652
2653
2654
2655
2656
2657
2658
2659
            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}")

Chen Zhang's avatar
Chen Zhang committed
2660
2661
2662
2663
2664
2665
2666
2667
2668
2669
2670
2671
2672
        mamba_layers = get_layers_from_vllm_config(self.vllm_config,
                                                   MambaMixer2)
        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 not self.vllm_config.model_config.enforce_eager:
                raise NotImplementedError(
                    "Mamba with cuda graph 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
2673
2674
2675
2676
2677

            page_size_padded = self._maybe_pad_mamba_page_size(
                attn_layers, mamba_layers, kv_cache_spec, max_model_len,
                block_size)

Chen Zhang's avatar
Chen Zhang committed
2678
2679
2680
2681
2682
2683
            # 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,
2684
2685
2686
                    block_size=max_model_len,
                    page_size_padded=page_size_padded)

2687
        return kv_cache_spec
2688
2689
2690
2691
2692
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
2720
2721
2722
2723
2724
2725
2726
2727
2728
2729
2730
2731
2732
2733
2734
2735
2736
2737
2738

    def _maybe_pad_mamba_page_size(
        self,
        attn_layers: dict[str, Attention],
        mamba_layers: dict[str, MambaMixer2],
        kv_cache_spec: dict[str, KVCacheSpec],
        max_model_len: int,
        block_size: int,
    ) -> Optional[int]:
        """
        Ensure that page size of attention KV cache groups is greater than or
        equal to the mamba KV cache groups. If not, we suggest to the user
        how to set the attention block size to ensure that it is.

        If the attention page size is strictly greater than the mamba page size,
        we pad the mamba page size to make them equal.

        Args:
            attn_layers: Attention layers
            mamba_layers: Mamba layers
            kv_cache_spec: KV cache spec (populated with attention layers)

        Returns:
            Optional[int]: Mamba page size with padding (None if no padding).
        """

        if len(attn_layers) == 0:
            return None

        attn_layer_name = next(iter(attn_layers))
        attn_page_size = kv_cache_spec[attn_layer_name].page_size_bytes
        mamba_layer_name = next(iter(mamba_layers))
        mamba_page_size = MambaSpec(
            shapes=mamba_layers[mamba_layer_name].get_state_shape(),
            dtype=self.kv_cache_dtype,
            block_size=max_model_len).page_size_bytes
        if attn_page_size < mamba_page_size:
            # attention page size (for 16 tokens)
            attn_page_size_16 = 16 * attn_page_size // block_size
            # some attention backends (e.g. FA) only support setting
            # block size to multiple of 16, so let's suggest a value
            # that would work (note: FA is currently not compatible
            # with mamba layers, use FlashInfer instead).
            suggest_attn_block_size = 16 * cdiv(mamba_page_size,
                                                attn_page_size_16)
            raise ValueError(
                "Attention block size should be increased to at least "
                f"{suggest_attn_block_size} in order to match "
                "the mamba page size")

        return attn_page_size