gpu_model_runner.py 105 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
20
from vllm.attention.backends.abstract import (AttentionBackend,
                                              AttentionMetadataBuilder)
21
from vllm.attention.layer import Attention
22
from vllm.attention.utils.fa_utils import get_flash_attn_version
23
24
from vllm.config import (CompilationLevel, VllmConfig,
                         get_layers_from_vllm_config)
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
30
    get_pp_group, get_tp_group, graph_capture,
    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
34
from vllm.model_executor.layers.rotary_embedding import MRotaryEmbedding
35
from vllm.model_executor.model_loader import TensorizerLoader, get_model_loader
36
37
from vllm.multimodal import MULTIMODAL_REGISTRY
from vllm.multimodal.inputs import MultiModalKwargs, PlaceholderRange
38
from vllm.multimodal.utils import group_mm_inputs_by_modality
39
from vllm.sampling_params import SamplingType
40
from vllm.sequence import IntermediateTensors
41
from vllm.utils import (STR_DTYPE_TO_TORCH_DTYPE, DeviceMemoryProfiler,
42
43
                        GiB_bytes, LazyLoader, async_tensor_h2d, cdiv,
                        check_use_alibi, is_pin_memory_available)
44
from vllm.v1.attention.backends.utils import CommonAttentionMetadata
45
from vllm.v1.core.encoder_cache_manager import compute_encoder_budget
46
47
48
from vllm.v1.kv_cache_interface import (AttentionSpec, FullAttentionSpec,
                                        KVCacheConfig, KVCacheSpec,
                                        SlidingWindowSpec)
49
50
from vllm.v1.outputs import (EMPTY_MODEL_RUNNER_OUTPUT, LogprobsTensors,
                             ModelRunnerOutput)
51
from vllm.v1.sample.metadata import SamplingMetadata
52
from vllm.v1.sample.rejection_sampler import RejectionSampler
53
from vllm.v1.sample.sampler import Sampler
54
from vllm.v1.spec_decode.eagle import EagleProposer
55
from vllm.v1.spec_decode.medusa import MedusaProposer
56
from vllm.v1.spec_decode.metadata import SpecDecodeMetadata
57
from vllm.v1.spec_decode.ngram_proposer import NgramProposer
58
from vllm.v1.spec_decode.utils import is_spec_decode_supported
59
from vllm.v1.utils import bind_kv_cache
60
from vllm.v1.worker.block_table import BlockTable
61
from vllm.v1.worker.gpu_input_batch import CachedRequestState, InputBatch
62
from vllm.v1.worker.lora_model_runner_mixin import LoRAModelRunnerMixin
63

64
65
from .utils import (gather_mm_placeholders, initialize_kv_cache_for_kv_sharing,
                    sanity_check_mm_encoder_outputs, scatter_mm_placeholders)
66

67
if TYPE_CHECKING:
68
69
    import xgrammar as xgr

70
    from vllm.model_executor.model_loader.tensorizer import TensorizerConfig
71
    from vllm.v1.core.sched.output import SchedulerOutput
72
73
else:
    xgr = LazyLoader("xgr", globals(), "xgrammar")
74
75
76
77

logger = init_logger(__name__)


78
class GPUModelRunner(LoRAModelRunnerMixin):
79
80
81

    def __init__(
        self,
82
        vllm_config: VllmConfig,
83
        device: torch.device,
84
    ):
85
86
87
88
89
90
91
92
93
94
        self.vllm_config = vllm_config
        self.model_config = vllm_config.model_config
        self.cache_config = vllm_config.cache_config
        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
95

96
97
98
99
        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))

100
101
102
103
        model_config = self.model_config
        cache_config = self.cache_config
        scheduler_config = self.scheduler_config
        parallel_config = self.parallel_config
104
        self.device = device
105
106
107
108
109
110
111
112
        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]

113
        self.is_multimodal_model = model_config.is_multimodal_model
114
115
        self.max_model_len = model_config.max_model_len
        self.max_num_tokens = scheduler_config.max_num_batched_tokens
116
        self.max_num_reqs = scheduler_config.max_num_seqs
117
118

        # Model-related.
119
120
        self.num_query_heads = model_config.get_num_attention_heads(
            parallel_config)
121
        self.hidden_size = model_config.get_hidden_size()
122
        self.attention_chunk_size = model_config.attention_chunk_size
123

124
        self.cascade_attn_enabled = not self.model_config.disable_cascade_attn
125

126
        # Multi-modal data support
127
        self.mm_registry = MULTIMODAL_REGISTRY
128
        self.uses_mrope = model_config.uses_mrope
129

130
131
132
        encoder_compute_budget, encoder_cache_size = compute_encoder_budget(
            model_config=model_config,
            scheduler_config=scheduler_config,
133
            mm_registry=self.mm_registry,
134
135
136
        )
        self.max_num_encoder_input_tokens = encoder_compute_budget
        self.encoder_cache_size = encoder_cache_size
137

138
139
140
        # Sampler
        self.sampler = Sampler()

141
        # Lazy initializations
142
        # self.model: nn.Module  # Set after load_model
143
        # Initialize in initialize_kv_cache
144
        self.kv_caches: list[torch.Tensor] = []
145
146
        self.attn_metadata_builders: list[AttentionMetadataBuilder] = []
        self.attn_backends: list[type[AttentionBackend]] = []
147
148
        # self.kv_cache_config: KVCacheConfig

149
        # req_id -> (input_id -> encoder_output)
150
        self.encoder_cache: dict[str, dict[int, torch.Tensor]] = {}
151

152
        self.use_aux_hidden_state_outputs = False
153
154
155
156
157
158
159
160
161
162
163
164
165
166
167
168
169
170
171
172
        # 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()
173

174
        # Request states.
175
        self.requests: dict[str, CachedRequestState] = {}
176

177
178
179
180
181
182
183
184
185
        # 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.
186
187
188
189
190
191
        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,
192
            vocab_size=self.model_config.get_vocab_size(),
193
            block_sizes=[self.cache_config.block_size],
194
        )
195

196
        self.use_cuda_graph = (self.vllm_config.compilation_config.level
197
198
199
                               == CompilationLevel.PIECEWISE
                               and not self.model_config.enforce_eager)
        # TODO(woosuk): Provide an option to tune the max cudagraph batch size.
200
201
202
203
        # 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(
204
205
            reversed(
                self.vllm_config.compilation_config.cudagraph_capture_sizes))
206

207
        # Cache the device properties.
208
        self._init_device_properties()
209

210
211
212
213
        # Persistent buffers for CUDA graphs.
        self.input_ids = torch.zeros(self.max_num_tokens,
                                     dtype=torch.int32,
                                     device=self.device)
214
215
216
        self.positions = torch.zeros(self.max_num_tokens,
                                     dtype=torch.int64,
                                     device=self.device)
217
218
219
220
221
222
223
224
225
226
        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)

227
228
        # None in the first PP rank. The rest are set after load_model.
        self.intermediate_tensors: Optional[IntermediateTensors] = None
229
230

        # Only relevant for models using M-RoPE (e.g, Qwen2-VL)
231
        if self.uses_mrope:
Roger Wang's avatar
Roger Wang committed
232
233
234
235
            # 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
236
237
238
239
240
241

            # 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
242
            self.mrope_positions = torch.zeros((3, self.max_num_tokens + 1),
243
244
                                               dtype=torch.int64,
                                               device=self.device)
Roger Wang's avatar
Roger Wang committed
245
246
247
248
249
            self.mrope_positions_cpu = torch.zeros(
                (3, self.max_num_tokens + 1),
                dtype=torch.int64,
                device="cpu",
                pin_memory=self.pin_memory)
250

251
252
253
        # Only relevant for models using ALiBi (e.g, MPT)
        self.use_alibi = check_use_alibi(model_config)

254
255
256
257
        self.inputs_embeds = torch.zeros(
            (self.max_num_tokens, self.hidden_size),
            dtype=self.dtype,
            device=self.device)
258

259
        # OPTIMIZATION: Cache the tensors rather than creating them every step.
260
        # Keep in int64 to avoid overflow with long context
261
        self.arange_np = np.arange(max(self.max_num_reqs + 1,
262
263
                                       self.max_model_len,
                                       self.max_num_tokens),
264
                                   dtype=np.int64)
265
266
267
        # 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.
268
269
270
271
272
273
274
275
276
277
278
279
280
281
        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()
282
283
284
285
286
        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()
287

288
289
290
291
292
293
        # 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] = {}

294
295
296
    def _may_reorder_batch(self, scheduler_output: "SchedulerOutput") -> bool:
        """
        Update the order of requests in the batch based on the attention
297
        backend's needs. For example, some attention backends (namely MLA) may
298
299
300
301
302
303
304
305
306
307
308
309
310
311
312
313
314
315
316
317
318
        want to separate requests based on if the attention computation will be
        compute-bound or memory-bound.

        Args:
            scheduler_output: The scheduler output.

        Returns:
            True if the batch was reordered, False otherwise.
        """
        batch_reordered = self.attn_metadata_builders[0].reorder_batch(
            self.input_batch, scheduler_output)

        # 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.
        for i in range(1, len(self.kv_cache_config.kv_cache_groups)):
            assert not self.attn_metadata_builders[i].reorder_batch(
                self.input_batch, scheduler_output)
        return batch_reordered

319
320
321
322
323
324
325
326
327
328
329
    # 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()

330
    def _update_states(self, scheduler_output: "SchedulerOutput") -> None:
331
332
333
334
335
336
        """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.

337
338
        The SamplingMetadata is updated and copied to the GPU if there is a
        new/resumed/paused/finished request in the batch.
339
340
        """
        # Remove finished requests from the cached states.
341
342
        for req_id in scheduler_output.finished_req_ids:
            self.requests.pop(req_id, None)
343
            self.encoder_cache.pop(req_id, None)
344
345
346
347
348
349
        # 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.
350
        removed_req_indices: list[int] = []
351
352
353
354
        for req_id in scheduler_output.finished_req_ids:
            req_index = self.input_batch.remove_request(req_id)
            if req_index is not None:
                removed_req_indices.append(req_index)
355
356
357
358
359
360
361
362

        # 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)
363

364
365
366
367
368
369
370
371
372
373
374
375
376
        # 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:
377
            req_index = self.input_batch.remove_request(req_id)
378
379
            assert req_index is not None
            removed_req_indices.append(req_index)
380

381
        req_ids_to_add: list[str] = []
382
        # Add new requests to the cached states.
383
384
385
        for new_req_data in scheduler_output.scheduled_new_reqs:
            req_id = new_req_data.req_id
            sampling_params = new_req_data.sampling_params
386
            if sampling_params.sampling_type == SamplingType.RANDOM_SEED:
387
388
389
390
391
                generator = torch.Generator(device=self.device)
                generator.manual_seed(sampling_params.seed)
            else:
                generator = None

392
393
            self.requests[req_id] = CachedRequestState(
                req_id=req_id,
394
395
396
                prompt_token_ids=new_req_data.prompt_token_ids,
                mm_inputs=new_req_data.mm_inputs,
                mm_positions=new_req_data.mm_positions,
397
398
                sampling_params=sampling_params,
                generator=generator,
399
400
                block_ids=new_req_data.block_ids,
                num_computed_tokens=new_req_data.num_computed_tokens,
401
                output_token_ids=[],
402
                lora_request=new_req_data.lora_request,
403
            )
404
405

            # Only relevant for models using M-RoPE (e.g, Qwen2-VL)
406
            if self.uses_mrope:
407
408
                image_grid_thw = []
                video_grid_thw = []
Roger Wang's avatar
Roger Wang committed
409
                second_per_grid_ts = []
410
411
                audio_feature_lengths = []
                use_audio_in_video = False
412
413
414
415
416
417
418
                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
419
420
421
                    if mm_input.get("second_per_grid_ts") is not None:
                        second_per_grid_ts.extend(
                            mm_input["second_per_grid_ts"])
422
423
424
425
426
                    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
427
428
429
430
431
432
433

                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
434
                        hf_config=hf_config,
435
436
                        image_grid_thw=image_grid_thw,
                        video_grid_thw=video_grid_thw,
Roger Wang's avatar
Roger Wang committed
437
                        second_per_grid_ts=second_per_grid_ts,
438
439
                        audio_feature_lengths=audio_feature_lengths,
                        use_audio_in_video=use_audio_in_video,
440
441
                    )

442
443
            req_ids_to_add.append(req_id)

444
445
446
        # Update the states of the running/resumed requests.
        for req_data in scheduler_output.scheduled_cached_reqs:
            req_id = req_data.req_id
447
448
            req_state = self.requests[req_id]

449
            # Update the cached states.
450
451
452
453
454
455
456
            num_computed_tokens = req_data.num_computed_tokens
            req_state.num_computed_tokens = num_computed_tokens
            # Add the sampled token(s) from the previous step (if any).
            # This doesn't include "unverified" tokens like spec decode tokens.
            num_new_tokens = (num_computed_tokens +
                              len(req_data.new_token_ids) -
                              req_state.num_tokens)
457
458
459
460
461
462
            if num_new_tokens == 1:
                # Avoid slicing list in most common case.
                req_state.output_token_ids.append(req_data.new_token_ids[-1])
            elif num_new_tokens > 0:
                req_state.output_token_ids.extend(
                    req_data.new_token_ids[-num_new_tokens:])
463
            # Update the block IDs.
464
465
            if not req_data.resumed_from_preemption:
                # Append the new blocks to the existing block IDs.
466
467
468
469
470
                for block_ids, new_block_ids in zip(  # type: ignore[call-overload]
                        req_state.block_ids,
                        req_data.new_block_ids,
                        strict=True):
                    block_ids.extend(new_block_ids)
471
472
473
474
475
476
477
478
479
480
481
482
483
484
485
            else:
                # The request is resumed from preemption.
                # Replace the existing block IDs with the new ones.
                req_state.block_ids = req_data.new_block_ids

            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] = (
486
                num_computed_tokens)
487
488
            self.input_batch.block_table.append_row(req_data.new_block_ids,
                                                    req_index)
489
490
491
492
493
494
            # Add new_token_ids to token_ids_cpu.
            start_token_index = num_computed_tokens
            end_token_index = num_computed_tokens + len(req_data.new_token_ids)
            self.input_batch.token_ids_cpu[
                req_index,
                start_token_index:end_token_index] = req_data.new_token_ids
495
            self.input_batch.num_tokens_no_spec[req_index] = end_token_index
496
497
            # Add spec_token_ids to token_ids_cpu.
            spec_token_ids = scheduler_output.scheduled_spec_decode_tokens.get(
498
                req_id, ())
499
500
501
502
503
504
505
            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 decode tokens.
            self.input_batch.num_tokens[req_index] = end_token_index
506

507
508
        # Check if the batch has changed. If not, we can skip copying the
        # sampling metadata from CPU to GPU.
509
510
        batch_changed = len(removed_req_indices) > 0 or len(req_ids_to_add) > 0

511
512
        # Add the new or resumed requests to the persistent batch.
        # The smaller empty indices are filled first.
513
        removed_req_indices.sort(reverse=True)
514
515
516
517
518
519
520
521
522
523
524
525
526
        for req_id in req_ids_to_add:
            req_state = self.requests[req_id]
            if removed_req_indices:
                # Fill the empty index.
                req_index = removed_req_indices.pop()
            else:
                # Append to the end.
                req_index = None
            self.input_batch.add_request(req_state, req_index)

        # Condense the batched states if there are empty indices.
        if removed_req_indices:
            self.input_batch.condense(removed_req_indices)
527

528
        batch_reordered = self._may_reorder_batch(scheduler_output)
529
530

        if batch_changed or batch_reordered:
531
            self.input_batch.refresh_sampling_metadata()
532

533
534
535
536
537
538
539
540
541
542
543
544
545
546
547
548
549
550
551
552
    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

553
    def _prepare_inputs(
554
555
        self,
        scheduler_output: "SchedulerOutput",
556
    ) -> tuple[dict[str, Any], torch.Tensor, Optional[SpecDecodeMetadata]]:
557
558
559
560
561
562
563
        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.
564
        self.input_batch.block_table.commit(num_reqs)
565
566

        # Get the number of scheduled tokens for each request.
567
568
569
570
        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)
571
572
573

        # Get request indices.
        # E.g., [2, 5, 3] -> [0, 0, 1, 1, 1, 1, 1, 2, 2, 2]
574
575
        req_indices = np.repeat(self.arange_np[:num_reqs],
                                num_scheduled_tokens)
576

577
578
579
580
        # 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)
581
582

        # Get positions.
583
        positions_np = self.positions_np[:total_num_scheduled_tokens]
584
585
586
587
        np.add(self.input_batch.num_computed_tokens_cpu[req_indices],
               arange,
               out=positions_np)

588
589
        # Calculate M-RoPE positions.
        # Only relevant for models using M-RoPE (e.g, Qwen2-VL)
590
        if self.uses_mrope:
591
592
            self._calc_mrope_positions(scheduler_output)

593
594
595
596
        # 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.
597
598
        token_indices = (positions_np +
                         req_indices * self.input_batch.token_ids_cpu.shape[1])
599

600
601
602
603
        # 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(),
604
                           0,
605
606
                           torch.from_numpy(token_indices),
                           out=self.input_ids_cpu[:total_num_scheduled_tokens])
607

608
609
610
611
612
613
614
615
616
617
618
619
620
621
622
623
624
625
626
627
628
629
630
        # 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])
631
632

        # Prepare the attention metadata.
633
        self.query_start_loc_np[0] = 0
634
        self.query_start_loc_np[1:num_reqs + 1] = cu_num_tokens
635

636
637
638
        self.seq_lens_np[:num_reqs] = (
            self.input_batch.num_computed_tokens_cpu[:num_reqs] +
            num_scheduled_tokens)
639
640
641
642

        # 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)
643
        if self.uses_mrope:
644
645
646
647
648
649
650
651
652
            # 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)
653

654
655
656
657
658
659
660
        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)
661
662
663
664
        # 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())
665
666
667
668

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

669
670
671
        common_attn_metadata = CommonAttentionMetadata(
            query_start_loc=query_start_loc, seq_lens=seq_lens)

672
        attn_metadata: dict[str, Any] = {}
673
674
675
676
677
678
679
680
681
682
        # 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
            if self.cascade_attn_enabled:
                common_prefix_len = self._compute_cascade_attn_prefix_len(
                    num_scheduled_tokens,
683
684
685
686
                    scheduler_output.
                    num_common_prefix_blocks[kv_cache_group_id],
                    kv_cache_group_spec.kv_cache_spec,
                    self.attn_metadata_builders[kv_cache_group_id],
687
                )
688

689
690
691
692
693
694
695
            attn_metadata_i = (
                self.attn_metadata_builders[kv_cache_group_id].build(
                    num_reqs=num_reqs,
                    num_actual_tokens=total_num_scheduled_tokens,
                    max_query_len=max_num_scheduled_tokens,
                    common_prefix_len=common_prefix_len,
                    common_attn_metadata=common_attn_metadata))
696
697
            for layer_name in kv_cache_group_spec.layer_names:
                attn_metadata[layer_name] = attn_metadata_i
698

699
700
        use_spec_decode = len(
            scheduler_output.scheduled_spec_decode_tokens) > 0
701
        if not use_spec_decode:
702
703
704
705
706
            # 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.
707
            logits_indices = query_start_loc[1:] - 1
708
709
710
711
712
713
714
715
716
717
718
719
720
721
            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
722

723
724
725
726
        # Hot-Swap lora model
        if self.lora_config:
            self.set_active_loras(self.input_batch, num_scheduled_tokens)

727
        return attn_metadata, logits_indices, spec_decode_metadata
728

729
730
731
732
    def _compute_cascade_attn_prefix_len(
        self,
        num_scheduled_tokens: np.ndarray,
        num_common_prefix_blocks: int,
733
734
        kv_cache_spec: KVCacheSpec,
        attn_metadata_builder: AttentionMetadataBuilder,
735
736
737
738
739
740
741
742
743
744
745
746
747
748
749
750
751
752
    ) -> 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.
        """
753
        common_prefix_len = num_common_prefix_blocks * kv_cache_spec.block_size
754
755
756
757
758
759
760
761
762
763
764
765
766
767
768
769
770
771
772
773
774
775
776
777
778
779
780
781
782
783
784
785
786
787
788
789
790
        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]
791
        # Request 3's num_computed_tokens: 3 (i.e., [A, B, C])
792
793
794
795
796
797
798
799
800
801
        # 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.
802
803
804
805
806
807
808
        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(
809
810
811
            common_prefix_len=common_prefix_len,
            query_lens=num_scheduled_tokens,
            num_query_heads=self.num_query_heads,
812
            num_kv_heads=kv_cache_spec.num_kv_heads,
813
            use_alibi=self.use_alibi,
814
            use_sliding_window=use_sliding_window,
815
816
817
818
            num_sms=self.num_sms,
        )
        return common_prefix_len if use_cascade else 0

819
820
    def _calc_mrope_positions(self, scheduler_output: "SchedulerOutput"):
        mrope_pos_ptr = 0
821
        for index, req_id in enumerate(self.input_batch.req_ids):
822
823
824
825
826
827
828
829
830
831
832
833
834
835
836
837
838
839
840
841
842
843
844
845
846
847
848
849
850
851
852
853
854
855
856
857
858
859
860
861
862
863
864
865
866
867
868
869
870
            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

                self.mrope_positions_cpu[:, dst_start:dst_end] = \
                    MRotaryEmbedding.get_next_input_positions_tensor(
                        req.mrope_position_delta,
                        context_len=num_computed_tokens +
                        prompt_part_len,
                        seq_len=num_computed_tokens +
                        prompt_part_len +
                        completion_part_len,
                    )

                mrope_pos_ptr += completion_part_len

871
872
    def _calc_spec_decode_metadata(
        self,
873
874
875
876
877
878
879
880
881
882
883
884
885
886
887
888
        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
889
890
891
892
893
894

        # 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]
895
896
        logits_indices = np.repeat(
            cu_num_scheduled_tokens - num_sampled_tokens, num_sampled_tokens)
897
        # Step 3. [0, 1, 2, 3, 103, 104, 105, 106, 206, 207, 208]
898
899
900
901
902
903
        logits_indices += arange

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

        # Compute the draft logits indices.
904
905
906
907
        # 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)
908
909
910
911
912
913
914
915
916
917
918
919
920
921
        # [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(
922
923
            self.device, non_blocking=True)

924
925
926
927
928
929
930
931
932
933
934
935
936
937
938
        # 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

939
    def _execute_mm_encoder(self, scheduler_output: "SchedulerOutput"):
940
941
942
943
944
        scheduled_encoder_inputs = scheduler_output.scheduled_encoder_inputs
        if not scheduled_encoder_inputs:
            return

        # Batch the multi-modal inputs.
945
946
        mm_inputs = list[MultiModalKwargs]()
        req_ids_pos = list[tuple[str, int, PlaceholderRange]]()
947
948
        for req_id, encoder_input_ids in scheduled_encoder_inputs.items():
            req_state = self.requests[req_id]
949
950
951
952
953

            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]))
954
955
956
957
958
959
960
961
962
963
964
965

        # 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:
966
967
            batched_mm_inputs = MultiModalKwargs.batch(
                grouped_mm_inputs, pin_memory=self.pin_memory)
968
969
970
971
            batched_mm_inputs = MultiModalKwargs.as_kwargs(
                batched_mm_inputs,
                device=self.device,
            )
972
973
974
975
976
977
978
979
980
981
982

            # 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)

983
984
985
986
987
            sanity_check_mm_encoder_outputs(
                curr_group_outputs,
                expected_num_items=len(grouped_mm_inputs),
            )

988
989
            for output in curr_group_outputs:
                encoder_outputs.append(output)
990
991

        # Cache the encoder outputs.
992
993
994
995
        for (req_id, input_id, pos_info), output in zip(
                req_ids_pos,
                encoder_outputs,
        ):
996
997
998
            if req_id not in self.encoder_cache:
                self.encoder_cache[req_id] = {}

999
1000
1001
1002
1003
1004
            self.encoder_cache[req_id][input_id] = scatter_mm_placeholders(
                output,
                is_embed=pos_info.is_embed,
            )

    def _gather_mm_embeddings(
1005
1006
        self,
        scheduler_output: "SchedulerOutput",
1007
    ) -> list[torch.Tensor]:
1008
        mm_embeds: list[torch.Tensor] = []
1009
        for req_id in self.input_batch.req_ids:
1010
1011
1012
1013
1014
1015
            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):
1016
1017
                start_pos = pos_info.offset
                num_encoder_tokens = pos_info.length
1018
1019
1020
1021
1022
1023
1024
1025
1026
1027
1028
1029
1030
1031
1032
1033
1034
1035
1036
1037
1038

                # 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]
1039
1040
1041
1042
1043
1044
1045
1046
1047
1048

                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
1049

1050
1051
1052
    def get_model(self) -> nn.Module:
        return self.model

1053
1054
1055
1056
1057
1058
1059
1060
1061
    def apply_grammar_bitmask(
        self,
        scheduler_output: "SchedulerOutput",
        logits: torch.Tensor,
    ):
        grammar_bitmask = scheduler_output.grammar_bitmask
        if grammar_bitmask is None:
            return

1062
1063
1064
1065
1066
1067
1068
1069
1070
        # 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.
1071
        struct_out_req_batch_indices: dict[str, int] = {}
1072
1073
1074
1075
1076
1077
1078
1079
1080
1081
1082
1083
1084
1085
1086
1087
1088
1089
1090
1091
1092
1093
1094
1095
1096
1097
1098
1099
1100
        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
1101

1102
1103
        # Serialization of np.ndarray is much more efficient than a tensor,
        # so we receive it in that format.
1104
1105
1106
1107
1108
        grammar_bitmask = torch.from_numpy(grammar_bitmask)

        xgr.apply_token_bitmask_inplace(
            logits,
            grammar_bitmask.to(self.device, non_blocking=True),
1109
            indices=out_indices,
1110
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
1140
1141
1142
1143
1144
1145
    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
        enabled_sp = self.vllm_config.compilation_config.pass_config. \
            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()
        })

1146
1147
    def get_dp_padding(self,
                       num_tokens: int) -> tuple[int, Optional[torch.Tensor]]:
1148
1149
        dp_size = self.vllm_config.parallel_config.data_parallel_size
        dp_rank = self.vllm_config.parallel_config.data_parallel_rank
1150
1151
1152
1153
1154
1155
1156
1157
1158

        # 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:
1159
            # Early exit.
1160
            return 0, None
1161
1162
1163
1164

        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()
1165
1166
1167
1168
1169
        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
1170

1171
1172
1173
1174
    @torch.inference_mode()
    def execute_model(
        self,
        scheduler_output: "SchedulerOutput",
1175
        intermediate_tensors: Optional[IntermediateTensors] = None,
1176
    ) -> Union[ModelRunnerOutput, IntermediateTensors]:
1177

1178
        self._update_states(scheduler_output)
1179
        if not scheduler_output.total_num_scheduled_tokens:
Robert Shaw's avatar
Robert Shaw committed
1180
1181
1182
1183
1184
            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)
1185
1186

        # Prepare the decoder inputs.
1187
1188
        attn_metadata, logits_indices, spec_decode_metadata = (
            self._prepare_inputs(scheduler_output))
1189
1190
1191
1192
1193
        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.
1194
            num_input_tokens = self.vllm_config.pad_for_cudagraph(
1195
1196
1197
                num_scheduled_tokens)
        else:
            # Eager mode.
1198
1199
1200
1201
1202
1203
1204
1205
1206
            # Pad tokens to multiple of tensor_parallel_size when
            # enabled collective fusion for SP
            tp_size = self.vllm_config.parallel_config.tensor_parallel_size
            if self.vllm_config.compilation_config.pass_config. \
                enable_sequence_parallelism and tp_size > 1:
                from vllm.utils import round_up
                num_input_tokens = round_up(num_scheduled_tokens, tp_size)
            else:
                num_input_tokens = num_scheduled_tokens
1207

1208
        # Padding for DP
1209
1210
        num_pad, num_tokens_across_dp = self.get_dp_padding(num_input_tokens)
        num_input_tokens += num_pad
1211

1212
1213
1214
1215
1216
1217
1218
1219
1220
        # _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 = []

1221
        if self.is_multimodal_model and get_pp_group().is_first_rank:
1222
1223
1224
1225
            # 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]
1226
            if mm_embeds:
1227
                inputs_embeds = self.model.get_input_embeddings(
1228
                    input_ids, mm_embeds)
1229
1230
1231
1232
1233
1234
            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
1235
        else:
1236
1237
1238
1239
1240
1241
            # 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
1242
1243
1244
1245
        if self.uses_mrope:
            positions = self.mrope_positions[:, :num_input_tokens]
        else:
            positions = self.positions[:num_input_tokens]
1246

1247
1248
1249
        if get_pp_group().is_first_rank:
            intermediate_tensors = None
        else:
1250
1251
            intermediate_tensors = self.sync_and_slice_intermediate_tensors(
                num_input_tokens, intermediate_tensors, True)
1252

1253
1254
        # Run the decoder.
        # Use persistent buffers for CUDA graphs.
1255
1256
        with set_forward_context(attn_metadata,
                                 self.vllm_config,
1257
1258
                                 num_tokens=num_input_tokens,
                                 num_tokens_across_dp=num_tokens_across_dp):
Robert Shaw's avatar
Robert Shaw committed
1259
1260
1261
            self.maybe_setup_kv_connector(scheduler_output)

            model_output = self.model(
1262
                input_ids=input_ids,
1263
                positions=positions,
1264
                intermediate_tensors=intermediate_tensors,
1265
                inputs_embeds=inputs_embeds,
1266
            )
1267

Robert Shaw's avatar
Robert Shaw committed
1268
1269
1270
1271
            self.maybe_wait_for_kv_save()
            finished_sending, finished_recving = (
                self.get_finished_kv_transfers(scheduler_output))

1272
        if self.use_aux_hidden_state_outputs:
Robert Shaw's avatar
Robert Shaw committed
1273
            hidden_states, aux_hidden_states = model_output
1274
        else:
Robert Shaw's avatar
Robert Shaw committed
1275
            hidden_states = model_output
1276
1277
1278
1279
1280
1281
1282
        # 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
1283
        if not get_pp_group().is_last_rank:
1284
            # For mid-pipeline stages, return the hidden states.
1285
1286
1287
1288
1289
1290
1291
1292
1293
1294
1295
1296
1297
1298
1299
1300
1301
            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:
            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"]
1302

1303
1304
1305
1306
        # Apply structured output bitmasks if present
        if scheduler_output.grammar_bitmask is not None:
            self.apply_grammar_bitmask(scheduler_output, logits)

1307
        # Sample the next token and get logprobs if needed.
1308
        sampling_metadata = self.input_batch.sampling_metadata
1309
        if spec_decode_metadata is None:
1310
            sampler_output = self.sampler(
1311
1312
1313
1314
                logits=logits,
                sampling_metadata=sampling_metadata,
            )
        else:
1315
1316
1317
1318
            # 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.
1319
            assert logits is not None
1320
            bonus_logits = logits[spec_decode_metadata.bonus_logits_indices]
1321
            sampler_output = self.sampler(
1322
                logits=bonus_logits,
1323
1324
1325
                sampling_metadata=sampling_metadata,
            )
            bonus_token_ids = sampler_output.sampled_token_ids
1326

1327
1328
1329
            # 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.
1330
            target_logits = logits[spec_decode_metadata.target_logits_indices]
1331
            output_token_ids = self.rejection_sampler(
1332
                spec_decode_metadata,
1333
                None,  # draft_probs
1334
                target_logits,
1335
                bonus_token_ids,
1336
1337
                sampling_metadata,
            )
1338
            sampler_output.sampled_token_ids = output_token_ids
1339
1340
1341

        # TODO(woosuk): The following loop can be slow since it iterates over
        # the requests one by one. Optimize.
1342
1343
        discard_sampled_tokens_req_indices = []
        for i, req_id in enumerate(self.input_batch.req_ids):
1344
1345
1346
            req_state = self.requests[req_id]
            seq_len = (req_state.num_computed_tokens +
                       scheduler_output.num_scheduled_tokens[req_id])
1347
            if seq_len < req_state.num_tokens:
1348
                # Ignore the sampled token for partial prefills.
1349
                # Rewind the generator state as if the token was not sampled.
1350
                # This relies on cuda-specific torch-internal impl details
1351
1352
1353
1354
1355
1356
                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)
1357

1358
1359
        # NOTE: GPU -> CPU Sync happens here.
        # Move as many CPU operations as possible before this sync point.
1360
1361
1362
1363
1364
1365
        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(
1366
            hidden_states[:num_scheduled_tokens],
1367
1368
1369
            scheduler_output,
        )

1370
        # Get the valid generated tokens.
1371
1372
1373
        sampled_token_ids = sampler_output.sampled_token_ids
        max_gen_len = sampled_token_ids.shape[-1]
        if max_gen_len == 1:
1374
            # No spec decode tokens.
1375
1376
            valid_sampled_token_ids = sampled_token_ids.tolist()
        else:
1377
            # Includes spec decode tokens.
1378
            valid_sampled_token_ids = self.rejection_sampler.parse_output(
1379
1380
1381
                sampled_token_ids,
                self.input_batch.vocab_size,
            )
1382
1383
1384
        # Mask out the sampled tokens that should not be sampled.
        for i in discard_sampled_tokens_req_indices:
            valid_sampled_token_ids[i].clear()
1385

1386
        if not self.speculative_config:
1387
            # Speculative decoding is not enabled.
1388
            spec_token_ids = None
1389
1390
        elif self.speculative_config.method == "ngram":
            assert isinstance(self.drafter, NgramProposer)
1391
            spec_token_ids = self.generate_draft_token_ids(
1392
                valid_sampled_token_ids, sampling_metadata)
1393
1394
1395
1396
1397
1398
1399
1400
1401
1402
1403
1404
1405
1406
1407
1408
1409
1410
1411
1412
1413
        elif self.speculative_config.method == "medusa":
            assert isinstance(self.drafter, MedusaProposer)
            if max_gen_len == 1:
                hidden_states = sample_hidden_states
            else:
                indices = []
                offset = 0
                for num_draft, tokens in zip(
                        spec_decode_metadata.num_draft_tokens,
                        valid_sampled_token_ids):
                    indices.append(offset + len(tokens) - 1)
                    offset += num_draft + 1

                indices = torch.tensor(indices,
                                       device=sample_hidden_states.device)
                hidden_states = sample_hidden_states[indices]

            spec_token_ids = self.drafter.propose(
                target_hidden_states=hidden_states,
                sampling_metadata=sampling_metadata,
            )
1414
        elif self.speculative_config.use_eagle():
1415
1416
1417
1418
1419
1420
1421
1422
1423
1424
1425
1426
1427
1428
1429
1430
            assert isinstance(self.drafter, EagleProposer)
            # TODO(woosuk): Refactor the loop.
            next_token_ids: list[int] = []
            for i, token_ids in enumerate(valid_sampled_token_ids):
                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)
1431
1432
1433
1434
1435
1436
1437
            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]]
1438

Jiayi Yao's avatar
Jiayi Yao committed
1439
1440
1441
1442
1443
1444
            # 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

1445
1446
1447
            if spec_decode_metadata is None:
                # input_ids can be None for multimodal models.
                target_token_ids = self.input_ids[:num_scheduled_tokens]
1448
                target_positions = positions[:num_scheduled_tokens]
1449
                if self.use_aux_hidden_state_outputs:
1450
1451
1452
                    target_hidden_states = torch.cat(
                        [h[:num_scheduled_tokens] for h in aux_hidden_states],
                        dim=-1)
1453
1454
                else:
                    target_hidden_states = hidden_states[:num_scheduled_tokens]
1455
1456
                target_slot_mapping = eagle_attn_metadata.slot_mapping
                cu_num_tokens = eagle_attn_metadata.query_start_loc
1457
1458
1459
1460
1461
1462
1463
            else:
                # TODO(woosuk): Refactor this.
                num_draft_tokens = spec_decode_metadata.num_draft_tokens
                num_rejected_tokens = [
                    n + 1 - len(valid_sampled_token_ids[i]) if n > 0 else 0
                    for i, n in enumerate(num_draft_tokens)
                ]
1464
                num_rejected_tokens_tensor = async_tensor_h2d(
1465
1466
                    num_rejected_tokens,
                    dtype=torch.int32,
1467
1468
1469
                    target_device=self.device,
                    pin_memory=True)
                num_tokens = num_scheduled_tokens - sum(num_rejected_tokens)
1470
                cu_num_tokens, token_indices = self.drafter.prepare_inputs(
1471
                    eagle_attn_metadata.query_start_loc,
1472
1473
                    num_rejected_tokens_tensor,
                    num_tokens,
1474
1475
1476
                )
                target_token_ids = self.input_ids[token_indices]
                target_positions = positions[token_indices]
1477
                if self.use_aux_hidden_state_outputs:
1478
1479
                    target_hidden_states = torch.cat(
                        [h[token_indices] for h in aux_hidden_states], dim=-1)
1480
1481
                else:
                    target_hidden_states = hidden_states[token_indices]
1482
1483
                target_slot_mapping = eagle_attn_metadata.slot_mapping[
                    token_indices]
1484
            draft_token_ids = self.drafter.propose(
1485
1486
1487
1488
1489
1490
                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
1491
                block_table=block_table,
1492
1493
1494
                sampling_metadata=sampling_metadata,
            )
            spec_token_ids = draft_token_ids.tolist()
1495

1496
1497
1498
1499
        # Clear KVConnector state after all KVs are generated.
        if has_kv_transfer_group():
            get_kv_transfer_group().clear_connector_metadata()

1500
        return ModelRunnerOutput(
1501
            req_ids=self.input_batch.req_ids,
1502
            req_id_to_index=self.input_batch.req_id_to_index,
1503
            sampled_token_ids=valid_sampled_token_ids,
1504
            spec_token_ids=spec_token_ids,
1505
1506
            logprobs=logprobs_lists,
            prompt_logprobs_dict=prompt_logprobs_dict,
Robert Shaw's avatar
Robert Shaw committed
1507
1508
            finished_sending=finished_sending,
            finished_recving=finished_recving,
1509
1510
        )

Robert Shaw's avatar
Robert Shaw committed
1511
1512
1513
1514
1515
1516
1517
1518
1519
1520
1521
1522
1523
1524
1525
1526
1527
1528
1529
1530
1531
1532
1533
1534
1535
1536
1537
1538
1539
1540
1541
1542
1543
1544
1545
1546
1547
1548
1549
1550
1551
1552
1553
1554
1555
1556
    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

1557
1558
    def generate_draft_token_ids(
        self,
1559
        sampled_token_ids: list[list[int]],
1560
        sampling_metadata: SamplingMetadata,
1561
    ) -> list[list[int]]:
1562
        # TODO(woosuk): Optimize.
1563
        draft_token_ids: list[list[int]] = []
1564
1565
1566
        for i, sampled_ids in enumerate(sampled_token_ids):
            num_sampled_ids = len(sampled_ids)
            if not num_sampled_ids:
1567
1568
1569
1570
                # Skip speculative decoding.
                draft_token_ids.append([])
                continue

1571
1572
            # Skip requests that require sampling parameters that are not
            # supported with speculative decoding.
1573
1574
1575
1576
1577
            req_id = self.input_batch.req_ids[i]
            if not is_spec_decode_supported(req_id, self.input_batch):
                draft_token_ids.append([])
                continue

1578
1579
            # Add sampled_token_ids to token_ids_cpu.
            start_idx = self.input_batch.num_tokens_no_spec[i]
1580
            end_idx = start_idx + num_sampled_ids
1581
1582
1583
1584
1585
            if end_idx >= self.max_model_len:
                # Skip requests that have already reached the max model length.
                draft_token_ids.append([])
                continue

1586
            self.input_batch.token_ids_cpu[i, start_idx:end_idx] = sampled_ids
1587
            drafter_output = self.drafter.propose(
1588
                self.input_batch.token_ids_cpu[i, :end_idx])
1589
1590
1591
1592
1593
1594
            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

1595
1596
1597
    def load_model(self) -> None:
        logger.info("Starting to load model %s...", self.model_config.model)
        with DeviceMemoryProfiler() as m:  # noqa: SIM117
1598
            time_before_load = time.perf_counter()
1599
1600
1601
1602
1603
1604
1605
1606
1607
1608
1609
1610
            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)
1611
1612
1613
1614
1615
1616
            if self.lora_config:
                self.model = self.load_lora_model(self.model,
                                                  self.model_config,
                                                  self.scheduler_config,
                                                  self.lora_config,
                                                  self.device)
1617
1618
1619
            if hasattr(self, "drafter"):
                logger.info("Loading drafter model...")
                self.drafter.load_model(self.model)
1620
1621
1622
            if self.use_aux_hidden_state_outputs:
                self.model.set_aux_hidden_state_layers(
                    self.model.get_eagle3_aux_hidden_state_layers())
1623
            time_after_load = time.perf_counter()
1624
        self.model_memory_usage = m.consumed_memory
1625
1626
        logger.info("Model loading took %.4f GiB and %.6f seconds",
                    self.model_memory_usage / GiB_bytes,
1627
                    time_after_load - time_before_load)
1628
        prepare_communication_buffer_for_model(self.model)
1629

1630
1631
1632
1633
1634
1635
1636
1637
1638
    def save_tensorized_model(
        self,
        tensorizer_config: "TensorizerConfig",
    ) -> None:
        TensorizerLoader.save_model(
            self.model,
            tensorizer_config=tensorizer_config,
        )

1639
1640
1641
1642
    def _get_prompt_logprobs_dict(
        self,
        hidden_states: torch.Tensor,
        scheduler_output: "SchedulerOutput",
1643
    ) -> dict[str, Optional[LogprobsTensors]]:
1644
1645
1646
1647
        num_prompt_logprobs_dict = self.input_batch.num_prompt_logprobs
        if not num_prompt_logprobs_dict:
            return {}

1648
        in_progress_dict = self.input_batch.in_progress_prompt_logprobs_cpu
1649
        prompt_logprobs_dict: dict[str, Optional[LogprobsTensors]] = {}
1650
1651
1652
1653
1654
1655
1656
1657
1658
1659
1660
1661
1662
1663

        # 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)

1664
1665
1666
1667
1668
1669
1670
1671
1672
            # 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

1673
            # Determine number of logits to retrieve.
1674
1675
            start_idx = request.num_computed_tokens
            start_tok = start_idx + 1
1676
            num_remaining_tokens = num_prompt_tokens - start_tok
1677
            if num_tokens <= num_remaining_tokens:
1678
                # This is a chunk, more tokens remain.
1679
1680
1681
                # 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.
1682
1683
1684
1685
1686
                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)
1687
1688
1689
1690
1691
1692
1693
                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
1694
1695
1696
1697
1698
1699
1700
1701
1702
1703
1704
1705
1706
1707
1708

            # 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.
1709
1710
            logprobs = self.sampler.compute_logprobs(logits)
            token_ids, logprobs, ranks = self.sampler.gather_logprobs(
1711
1712
1713
                logprobs, num_prompt_logprobs, tgt_token_ids)

            # Transfer GPU->CPU async.
1714
1715
1716
1717
1718
1719
1720
            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)
1721
1722
1723
1724
1725

        # 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]
1726
            del in_progress_dict[req_id]
1727
1728

        # Must synchronize the non-blocking GPU->CPU transfers.
1729
        if prompt_logprobs_dict:
1730
            self._sync_device()
1731
1732
1733

        return prompt_logprobs_dict

1734
1735
1736
1737
1738
1739
1740
1741
1742
1743
1744
1745
1746
1747
1748
1749
1750
1751
1752
1753
1754
1755
1756
1757
1758
1759
1760
1761
1762
    @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)

1763
1764
1765
1766
    @torch.inference_mode()
    def _dummy_run(
        self,
        num_tokens: int,
1767
        skip_attn: bool = True,
1768
    ) -> torch.Tensor:
1769

1770
        # Padding for DP
1771
1772
        num_pad, num_tokens_across_dp = self.get_dp_padding(num_tokens)
        num_tokens += num_pad
1773

1774
1775
1776
1777
1778
        # 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
1779
        num_reqs = min(num_tokens, max_num_reqs)
1780
1781
1782
1783
1784
1785
1786
        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)
1787

1788
        if skip_attn:
1789
            attn_metadata: Optional[dict[str, Any]] = None
1790
1791
        else:
            query_start_loc = self.query_start_loc[:num_reqs + 1]
1792
1793
1794
1795
1796
            # 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)
1797
1798
1799
1800
1801
            seq_lens = self.seq_lens[:num_reqs]

            common_attn_metadata = CommonAttentionMetadata(
                query_start_loc=query_start_loc, seq_lens=seq_lens)

1802
1803
1804
1805
1806
            attn_metadata = {}
            for kv_cache_group_id, kv_cache_group_spec in enumerate(
                    self.kv_cache_config.kv_cache_groups):
                attn_metadata_i = (
                    self.attn_metadata_builders[kv_cache_group_id].build(
1807
                        num_reqs=num_reqs,
1808
1809
1810
1811
1812
1813
1814
                        num_actual_tokens=num_tokens,
                        max_query_len=num_tokens,
                        common_prefix_len=0,
                        common_attn_metadata=common_attn_metadata,
                    ))
                for layer_name in kv_cache_group_spec.layer_names:
                    attn_metadata[layer_name] = attn_metadata_i
1815

1816
1817
1818
1819
1820
1821
1822
1823
1824
1825
1826
1827
1828
1829
1830
1831
1832
1833
1834
1835
1836
1837
1838
        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))
1839
1840
1841

                intermediate_tensors = self.sync_and_slice_intermediate_tensors(
                    num_tokens, None, False)
1842

1843
            with self.maybe_randomize_inputs(input_ids), set_forward_context(
1844
1845
1846
1847
                    attn_metadata,
                    self.vllm_config,
                    num_tokens=num_tokens,
                    num_tokens_across_dp=num_tokens_across_dp):
1848
                outputs = model(
1849
1850
1851
1852
1853
                    input_ids=input_ids,
                    positions=positions,
                    intermediate_tensors=intermediate_tensors,
                    inputs_embeds=inputs_embeds,
                )
1854
1855
1856
1857
            if self.use_aux_hidden_state_outputs:
                hidden_states, _ = outputs
            else:
                hidden_states = outputs
1858

1859
            if self.speculative_config and self.speculative_config.use_eagle():
1860
1861
1862
                assert isinstance(self.drafter, EagleProposer)
                self.drafter.dummy_run(num_tokens)

1863
1864
1865
1866
1867
1868
1869
1870
        logit_indices = np.cumsum(num_scheduled_tokens) - 1
        return hidden_states[logit_indices]

    @torch.inference_mode()
    def _dummy_sampler_run(
        self,
        hidden_states: torch.Tensor,
    ) -> torch.Tensor:
1871
1872
1873
1874
        # 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)
1875
1876
1877
1878
1879
1880
1881
1882
1883
1884
1885
1886
1887
1888
1889
1890
1891
1892
1893
1894
1895
1896
1897
1898
1899
1900
1901

        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),
            min_p=None,
            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)],
            min_tokens={},
            logit_bias=[None for _ in range(num_reqs)],
            allowed_token_ids_mask=None,
            bad_words_token_ids={},
        )
1902
        try:
1903
1904
            sampler_output = self.sampler(logits=logits,
                                          sampling_metadata=dummy_metadata)
1905
1906
1907
1908
1909
1910
1911
1912
1913
        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
1914
        if self.speculative_config:
1915
1916
1917
1918
1919
1920
1921
1922
1923
1924
1925
1926
1927
1928
1929
1930
1931
1932
1933
1934
1935
1936
1937
1938
1939
1940
            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,
            )
1941
        return sampler_output
1942
1943

    def profile_run(self) -> None:
1944
        # Profile with multimodal encoder & encoder cache.
1945
1946
1947
        # 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):
1948

1949
            # NOTE: Currently model is profiled with a single non-text
1950
1951
            # modality with the max possible input tokens even when
            # it supports multiple.
1952
1953
            max_tokens_by_modality_dict = self.mm_registry \
                .get_max_tokens_per_item_by_nonzero_modality(self.model_config)
1954
1955
1956
1957
            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
1958
1959
1960
1961
1962
1963
            # 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)
1964
1965
1966

            # Check how many items of this modality can be supported by
            # the decoder budget.
1967
1968
            max_mm_items_per_req = self.mm_registry.get_mm_limits_per_prompt(
                self.model_config)[dummy_data_modality]
1969
1970
1971
1972
1973
1974
1975
1976
1977
1978

            # 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)

1979
1980
1981
1982
1983
1984
            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.
1985
            dummy_mm_kwargs = self.mm_registry.get_decoder_dummy_data(
1986
1987
                model_config=self.model_config,
                seq_len=self.max_num_tokens,
1988
1989
1990
1991
                mm_counts={
                    dummy_data_modality: 1
                },
            ).multi_modal_data
1992

1993
            batched_dummy_mm_inputs = MultiModalKwargs.batch(
1994
1995
                [dummy_mm_kwargs] * max_num_mm_items,
                pin_memory=self.pin_memory)
1996
            batched_dummy_mm_inputs = MultiModalKwargs.as_kwargs(
1997
1998
1999
                batched_dummy_mm_inputs,
                device=self.device,
            )
2000
2001
2002
2003

            # Run multimodal encoder.
            dummy_encoder_outputs = self.model.get_multimodal_embeddings(
                **batched_dummy_mm_inputs)
2004
2005
2006
2007
2008

            sanity_check_mm_encoder_outputs(
                dummy_encoder_outputs,
                expected_num_items=max_num_mm_items,
            )
2009
2010
2011
2012

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

2013
2014
2015
2016
2017
        hidden_states = self._dummy_run(self.max_num_tokens)
        if get_pp_group().is_last_rank:
            sampler_output = self._dummy_sampler_run(hidden_states)
        else:
            sampler_output = None
2018
        self._sync_device()
2019
2020
        del hidden_states, sampler_output
        self.encoder_cache.clear()
2021
        gc.collect()
2022
2023

    def capture_model(self) -> None:
2024
2025
        if not self.use_cuda_graph:
            logger.warning(
2026
                "Skipping CUDA graph capture. Please add "
2027
                "-O %s to use CUDA graphs.", CompilationLevel.PIECEWISE)
2028
2029
2030
2031
2032
            return

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

2033
2034
2035
        # 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.
2036
        with graph_capture(device=self.device):
2037
            skip_attn = not self.vllm_config.compilation_config.full_cuda_graph
2038
2039
2040
            for num_tokens in tqdm(reversed(self.cudagraph_batch_sizes),
                                   desc="Capturing CUDA graphs",
                                   total=len(self.cudagraph_batch_sizes)):
2041
2042
                for _ in range(self.vllm_config.compilation_config.
                               cudagraph_num_of_warmups):
2043
2044
                    self._dummy_run(num_tokens, skip_attn=skip_attn)
                self._dummy_run(num_tokens, skip_attn=skip_attn)
2045
2046
2047
2048
2049
2050
2051
2052

        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))
2053

2054
2055
2056
2057
2058
2059
2060
2061
2062
2063
2064
2065
2066
2067
2068
2069
2070
2071
2072
2073
2074
2075
2076
2077
2078
2079
2080
2081
2082
2083
2084
2085
2086
2087
2088
2089
2090
2091
2092
2093
2094
2095
2096
2097
2098
2099
2100
2101
2102
2103
    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
            if not isinstance(kv_cache_spec, AttentionSpec):
                raise NotImplementedError(
                    "Only AttentionSpec is supported for now.")
            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: {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.")

            if self.vllm_config.compilation_config.full_cuda_graph:
                attn_backend_name = attn_backend_i.__name__
                flash_attn_version = get_flash_attn_version()
                if attn_backend_name != "FlashAttentionBackend" or \
                    flash_attn_version != 3:
                    raise ValueError(
                        f"full_cuda_graph is only supported with "
                        f"FA3. Current attention backend is "
                        f"{attn_backend_name}, FlashAttention version is "
                        f"{flash_attn_version}.")

            block_table_i = self.input_batch.block_table[i]
            attn_metadata_builder_i = attn_backend_i.get_builder_cls()(
                weakref.proxy(self), kv_cache_spec, block_table_i)
            self.attn_backends.append(attn_backend_i)
            self.attn_metadata_builders.append(attn_metadata_builder_i)

2104
2105
2106
2107
2108
2109
2110
2111
2112
2113
2114
2115
2116
2117
2118
2119
2120
2121
2122
2123
2124
2125
2126
2127
2128
2129
2130
2131
2132
    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,
            )

2133
2134
    def _allocate_kv_cache_tensors(
            self, kv_cache_config: KVCacheConfig) -> dict[str, torch.Tensor]:
2135
        """
2136
2137
2138
        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.

2139
        Args:
2140
2141
2142
2143
2144
2145
2146
2147
2148
2149
2150
2151
2152
2153
2154
2155
2156
2157
2158
2159
2160
2161
2162
2163
2164
            kv_cache_config: The KV cache config 
        Returns:
            dict[str, torch.Tensor]: A map between layer names to their 
            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]:
2165
        """
2166
        Reshape the KV cache tensors to the desired shape and dtype.
2167

2168
2169
2170
2171
2172
2173
2174
2175
        Args:
            kv_cache_config: The KV cache config 
            kv_cache_raw_tensors: The KV cache buffer of each layer, with 
            correct size but uninitialized shape.
        Returns:
            Dict[str, torch.Tensor]: A map between layer names to their 
            corresponding memory buffer for KV cache.
        """
2176
        kv_caches: dict[str, torch.Tensor] = {}
2177
2178
2179
2180
2181
2182
2183
2184
        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)
2185
                if isinstance(kv_cache_spec, AttentionSpec):
2186
                    kv_cache_shape = self.attn_backends[i].get_kv_cache_shape(
2187
2188
2189
                        num_blocks, kv_cache_spec.block_size,
                        kv_cache_spec.num_kv_heads, kv_cache_spec.head_size)
                    dtype = kv_cache_spec.dtype
2190
2191
2192
2193
2194
2195
2196
2197
2198
2199
2200
2201
2202
2203
2204
2205
2206
2207
2208
2209
                    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))
                    ]
2210
2211
2212
                    kv_caches[layer_name] = kv_cache_raw_tensors[
                        layer_name].view(dtype).view(kv_cache_shape).permute(
                            *inv_order)
2213
                else:
2214
2215
2216
2217
2218
2219
2220
2221
2222
2223
2224
2225
2226
2227
2228
2229
2230
2231
2232
                    raise NotImplementedError
        return kv_caches

    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:
            Dict[str, torch.Tensor]: A map between layer names to their 
            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)
2233

2234
2235
2236
2237
2238
2239
2240
2241
2242
        # 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,
            )

2243
2244
2245
2246
2247
2248
2249
2250
2251
2252
2253
2254
2255
2256
2257
2258
2259
2260
        bind_kv_cache(
            kv_caches,
            self.vllm_config.compilation_config.static_forward_context,
            self.kv_caches)
        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)

2261
2262
2263
2264
2265
2266
        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
2267
2268
2269
        if has_kv_transfer_group():
            get_kv_transfer_group().register_kv_caches(kv_caches)

2270
    def get_kv_cache_spec(self) -> dict[str, KVCacheSpec]:
2271
        """
2272
        Generates the KVCacheSpec by parsing the kv cache format from each
2273
2274
        Attention module in the static forward context.
        Returns:
2275
            KVCacheSpec: A dictionary mapping layer names to their KV cache
2276
2277
2278
            format. Layers that do not need KV cache are not included.
        """

2279
        layers = get_layers_from_vllm_config(self.vllm_config, Attention)
2280
        block_size = self.vllm_config.cache_config.block_size
2281
        use_mla = self.vllm_config.model_config.use_mla
2282
        kv_cache_spec: dict[str, KVCacheSpec] = {}
2283
        for layer_name, attn_module in layers.items():
2284
2285
2286
2287
2288
2289
2290
2291
2292
2293
2294
2295
            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

2296
            # TODO: Support other attention modules, e.g., cross-attention
2297
            if attn_module.attn_type == AttentionType.DECODER:
2298
2299
2300
2301
2302
2303
2304
2305
2306
2307
2308
2309
2310
2311
2312
                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)
2313
2314
2315
2316
2317
2318
2319
2320
2321
2322
2323
            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}")

        return kv_cache_spec