gpu_model_runner.py 236 KB
Newer Older
1
# SPDX-License-Identifier: Apache-2.0
2
# SPDX-FileCopyrightText: Copyright contributors to the vLLM project
3

4
import functools
5
import gc
6
import itertools
7
import time
8
from collections import defaultdict
9
from collections.abc import Iterator, Sequence
10
from contextlib import contextmanager
11
from copy import copy, deepcopy
12
from functools import reduce
13
from itertools import product
14
from typing import TYPE_CHECKING, Any, NamedTuple, TypeAlias, cast
15
16
17
18
19

import numpy as np
import torch
import torch.distributed
import torch.nn as nn
20
from tqdm import tqdm
21

22
import vllm.envs as envs
23
24
25
from vllm.attention.backends.abstract import (
    AttentionBackend,
    AttentionMetadata,
26
    AttentionType,
27
28
    MultipleOf,
)
29
from vllm.attention.layer import Attention, MLAAttention
30
from vllm.compilation.counter import compilation_counter
31
from vllm.compilation.cuda_graph import CUDAGraphStat, CUDAGraphWrapper
32
from vllm.compilation.monitor import set_cudagraph_capturing_enabled
33
from vllm.config import (
34
    CompilationMode,
35
36
37
38
39
    CUDAGraphMode,
    VllmConfig,
    get_layers_from_vllm_config,
    update_config,
)
40
from vllm.distributed.ec_transfer import get_ec_transfer, has_ec_transfer
41
from vllm.distributed.eplb.eplb_state import EplbState
42
from vllm.distributed.kv_transfer import get_kv_transfer_group, has_kv_transfer_group
43
from vllm.distributed.kv_transfer.kv_connector.utils import copy_kv_blocks
44
from vllm.distributed.parallel_state import (
45
    get_dcp_group,
46
47
48
49
50
51
    get_pp_group,
    get_tp_group,
    graph_capture,
    is_global_first_rank,
    prepare_communication_buffer_for_model,
)
52
53
54
55
from vllm.forward_context import (
    BatchDescriptor,
    set_forward_context,
)
56
from vllm.logger import init_logger
57
from vllm.model_executor.layers.attention_layer_base import AttentionLayerBase
58
59
60
61
from vllm.model_executor.layers.rotary_embedding import (
    MRotaryEmbedding,
    XDRotaryEmbedding,
)
62
from vllm.model_executor.model_loader import TensorizerLoader, get_model_loader
63
from vllm.model_executor.models.interfaces import (
64
    SupportsMRoPE,
65
    SupportsMultiModal,
66
    SupportsXDRoPE,
67
68
69
70
71
    is_mixture_of_experts,
    supports_eagle3,
    supports_mrope,
    supports_multimodal_pruning,
    supports_transcription,
72
    supports_xdrope,
73
)
74
from vllm.model_executor.models.interfaces_base import (
75
76
77
78
    VllmModelForPooling,
    is_pooling_model,
    is_text_generation_model,
)
79
from vllm.multimodal import MULTIMODAL_REGISTRY
80
81
82
83
84
from vllm.multimodal.inputs import (
    BatchedTensorInputs,
    MultiModalKwargsItem,
    PlaceholderRange,
)
85
from vllm.multimodal.utils import group_mm_kwargs_by_modality
86
from vllm.pooling_params import PoolingParams
87
from vllm.sampling_params import SamplingType
88
from vllm.sequence import IntermediateTensors
89
from vllm.tasks import GenerationTask, PoolingTask, SupportedTask
90
from vllm.utils import length_from_prompt_token_ids_or_embeds
91
from vllm.utils.jsontree import json_map_leaves
92
from vllm.utils.math_utils import cdiv, round_up
93
94
from vllm.utils.mem_constants import GiB_bytes
from vllm.utils.mem_utils import DeviceMemoryProfiler
95
from vllm.utils.nvtx_pytorch_hooks import PytHooks
96
from vllm.utils.platform_utils import is_pin_memory_available
97
98
99
100
101
from vllm.utils.torch_utils import (
    get_dtype_size,
    kv_cache_dtype_str_to_dtype,
    supports_dynamo,
)
102
from vllm.v1.attention.backends.gdn_attn import GDNAttentionMetadataBuilder
103
from vllm.v1.attention.backends.utils import (
104
105
106
    AttentionCGSupport,
    AttentionMetadataBuilder,
    CommonAttentionMetadata,
107
    create_fast_prefill_custom_backend,
108
    get_dcp_local_seq_lens,
109
110
111
    reorder_batch_to_split_decodes_and_prefills,
    split_attn_metadata,
)
112
from vllm.v1.cudagraph_dispatcher import CudagraphDispatcher
113
114
115
116
117
118
119
120
121
122
123
124
125
126
127
128
129
from vllm.v1.kv_cache_interface import (
    AttentionSpec,
    ChunkedLocalAttentionSpec,
    CrossAttentionSpec,
    EncoderOnlyAttentionSpec,
    FullAttentionSpec,
    KVCacheConfig,
    KVCacheGroupSpec,
    KVCacheSpec,
    MambaSpec,
    SlidingWindowSpec,
    UniformTypeKVCacheSpecs,
)
from vllm.v1.outputs import (
    EMPTY_MODEL_RUNNER_OUTPUT,
    AsyncModelRunnerOutput,
    DraftTokenIds,
130
    ECConnectorOutput,
131
    KVConnectorOutput,
132
133
134
135
136
    LogprobsLists,
    LogprobsTensors,
    ModelRunnerOutput,
    PoolerOutput,
    SamplerOutput,
137
    make_empty_encoder_model_runner_output,
138
)
139
from vllm.v1.pool.metadata import PoolingMetadata, PoolingStates
140
from vllm.v1.sample.logits_processor import LogitsProcessors, build_logitsprocs
141
from vllm.v1.sample.logits_processor.interface import LogitsProcessor
142
from vllm.v1.sample.metadata import SamplingMetadata
143
from vllm.v1.sample.rejection_sampler import RejectionSampler
144
from vllm.v1.sample.sampler import Sampler
145
from vllm.v1.spec_decode.eagle import EagleProposer
146
from vllm.v1.spec_decode.medusa import MedusaProposer
147
from vllm.v1.spec_decode.metadata import SpecDecodeMetadata
148
from vllm.v1.spec_decode.ngram_proposer import NgramProposer
149
from vllm.v1.spec_decode.suffix_decoding import SuffixDecodingProposer
150
from vllm.v1.structured_output.utils import apply_grammar_bitmask
151
from vllm.v1.utils import CpuGpuBuffer, record_function_or_nullcontext
152
from vllm.v1.worker.cp_utils import check_attention_cp_compatibility
153
from vllm.v1.worker.dp_utils import coordinate_batch_across_dp
154
from vllm.v1.worker.ec_connector_model_runner_mixin import ECConnectorModelRunnerMixin
155
from vllm.v1.worker.gpu_input_batch import CachedRequestState, InputBatch
156
from vllm.v1.worker.gpu_ubatch_wrapper import UBatchWrapper
157
from vllm.v1.worker.kv_connector_model_runner_mixin import KVConnectorModelRunnerMixin
158
from vllm.v1.worker.lora_model_runner_mixin import LoRAModelRunnerMixin
159
160
161
from vllm.v1.worker.ubatch_utils import (
    UBatchSlices,
    check_ubatch_thresholds,
162
    maybe_create_ubatch_slices,
163
)
164
from vllm.v1.worker.utils import is_residual_scattered_for_sp
165
from vllm.v1.worker.workspace import lock_workspace
166

167
168
169
170
171
172
173
174
175
from .utils import (
    AttentionGroup,
    MultiModalBudget,
    add_kv_sharing_layers_to_kv_cache_groups,
    bind_kv_cache,
    gather_mm_placeholders,
    sanity_check_mm_encoder_outputs,
    scatter_mm_placeholders,
)
176

177
if TYPE_CHECKING:
178
    from vllm.model_executor.model_loader.tensorizer import TensorizerConfig
179
    from vllm.v1.core.sched.output import GrammarOutput, SchedulerOutput
180
181
182

logger = init_logger(__name__)

183
184
AttnMetadataDict: TypeAlias = dict[str, AttentionMetadata]
# list when ubatching is enabled
185
PerLayerAttnMetadata: TypeAlias = list[AttnMetadataDict] | AttnMetadataDict
186

187

188
189
190
191
192
193
# Wrapper for ModelRunnerOutput to support overlapped execution.
class AsyncGPUModelRunnerOutput(AsyncModelRunnerOutput):
    def __init__(
        self,
        model_runner_output: ModelRunnerOutput,
        sampled_token_ids: torch.Tensor,
194
        logprobs_tensors: LogprobsTensors | None,
195
196
        invalid_req_indices: list[int],
        async_output_copy_stream: torch.cuda.Stream,
197
        vocab_size: int,
198
199
200
201
202
    ):
        self._model_runner_output = model_runner_output
        self._invalid_req_indices = invalid_req_indices

        # Event on the copy stream so we can synchronize the non-blocking copy.
203
        self.async_copy_ready_event = torch.Event()
204
205
206
207

        # Keep a reference to the device tensor to avoid it being
        # deallocated until we finish copying it to the host.
        self._sampled_token_ids = sampled_token_ids
208
        self.vocab_size = vocab_size
209
        self._logprobs_tensors = logprobs_tensors
210
211
212
213
214

        # Initiate the copy on a separate stream, but do not synchronize it.
        default_stream = torch.cuda.current_stream()
        with torch.cuda.stream(async_output_copy_stream):
            async_output_copy_stream.wait_stream(default_stream)
215
            self.sampled_token_ids_cpu = self._sampled_token_ids.to(
216
217
                "cpu", non_blocking=True
            )
218
219
220
221
222
            self._logprobs_tensors_cpu = (
                self._logprobs_tensors.to_cpu_nonblocking()
                if self._logprobs_tensors
                else None
            )
223
            self.async_copy_ready_event.record()
224
225
226

    def get_output(self) -> ModelRunnerOutput:
        """Copy the device tensors to the host and return a ModelRunnerOutput.
227

228
229
        This function blocks until the copy is finished.
        """
230
        max_gen_len = self.sampled_token_ids_cpu.shape[-1]
231
        self.async_copy_ready_event.synchronize()
232

233
234
        # Release the device tensors once the copy has completed.
        del self._logprobs_tensors
235
        del self._sampled_token_ids
236
        if max_gen_len == 1:
237
            valid_sampled_token_ids = self.sampled_token_ids_cpu.tolist()
238
239
240
            for i in self._invalid_req_indices:
                valid_sampled_token_ids[i].clear()
            cu_num_tokens = None
241
        else:
242
            valid_sampled_token_ids, cu_num_tokens = RejectionSampler.parse_output(
243
244
                self.sampled_token_ids_cpu,
                self.vocab_size,
245
246
                self._invalid_req_indices,
                return_cu_num_tokens=self._logprobs_tensors_cpu is not None,
247
            )
248
249
250

        output = self._model_runner_output
        output.sampled_token_ids = valid_sampled_token_ids
251
        if self._logprobs_tensors_cpu:
252
            output.logprobs = self._logprobs_tensors_cpu.tolists(cu_num_tokens)
253
254
255
        return output


256
257
258
259
260
261
262
263
264
265
266
class ExecuteModelState(NamedTuple):
    """Ephemeral cached state transferred between execute_model() and
    sample_tokens(), after execute_model() returns None."""

    scheduler_output: "SchedulerOutput"
    logits: torch.Tensor
    spec_decode_metadata: SpecDecodeMetadata | None
    spec_decode_common_attn_metadata: CommonAttentionMetadata | None
    hidden_states: torch.Tensor
    sample_hidden_states: torch.Tensor
    aux_hidden_states: list[torch.Tensor] | None
267
    ec_connector_output: ECConnectorOutput | None
268
    cudagraph_stats: CUDAGraphStat | None
269
270


271
272
273
class GPUModelRunner(
    LoRAModelRunnerMixin, KVConnectorModelRunnerMixin, ECConnectorModelRunnerMixin
):
274
275
    def __init__(
        self,
276
        vllm_config: VllmConfig,
277
        device: torch.device,
278
    ):
279
280
281
        self.vllm_config = vllm_config
        self.model_config = vllm_config.model_config
        self.cache_config = vllm_config.cache_config
282
        self.compilation_config = vllm_config.compilation_config
283
284
285
286
287
288
        self.lora_config = vllm_config.lora_config
        self.load_config = vllm_config.load_config
        self.parallel_config = vllm_config.parallel_config
        self.scheduler_config = vllm_config.scheduler_config
        self.speculative_config = vllm_config.speculative_config
        self.observability_config = vllm_config.observability_config
289

290
        from vllm.model_executor.models.utils import set_cpu_offload_max_bytes
291
292

        set_cpu_offload_max_bytes(int(self.cache_config.cpu_offload_gb * 1024**3))
293

294
295
296
297
        model_config = self.model_config
        cache_config = self.cache_config
        scheduler_config = self.scheduler_config
        parallel_config = self.parallel_config
298
        self.device = device
299
300
        self.pin_memory = is_pin_memory_available()
        self.dtype = self.model_config.dtype
301

302
303
304
        self.kv_cache_dtype = kv_cache_dtype_str_to_dtype(
            cache_config.cache_dtype, self.model_config
        )
305

306
        self.is_pooling_model = model_config.runner_type == "pooling"
307
        self.enable_prompt_embeds = model_config.enable_prompt_embeds
308
        self.is_multimodal_raw_input_only_model = (
309
310
            model_config.is_multimodal_raw_input_only_model
        )
311
312
        # This will be overridden in load_model()
        self.is_multimodal_pruning_enabled = False
313
        self.max_model_len = model_config.max_model_len
314
315
316

        # Always set to false after the first forward pass
        self.calculate_kv_scales = self.cache_config.calculate_kv_scales
317
        self.dcp_world_size = self.parallel_config.decode_context_parallel_size
318
        self.dcp_rank = 0 if self.dcp_world_size <= 1 else get_dcp_group().rank_in_group
319
        self.max_num_tokens = scheduler_config.max_num_batched_tokens
320
        self.max_num_reqs = scheduler_config.max_num_seqs
321

322
323
324
325
326
        # Broadcast PP output for external_launcher (torchrun)
        # to make sure we are synced across pp ranks
        # TODO: Support overlapping mirco-batches
        # https://github.com/vllm-project/vllm/issues/18019
        self.broadcast_pp_output = (
327
328
329
            self.parallel_config.distributed_executor_backend == "external_launcher"
            and len(get_pp_group().ranks) > 0
        )
330

331
        # Model-related.
332
        self.num_query_heads = model_config.get_num_attention_heads(parallel_config)
333
        self.inputs_embeds_size = model_config.get_inputs_embeds_size()
334
        self.attention_chunk_size = model_config.attention_chunk_size
335
        # Only relevant for models using ALiBi (e.g, MPT)
336
        self.use_alibi = model_config.uses_alibi
337

338
        self.cascade_attn_enabled = not self.model_config.disable_cascade_attn
339
        self.is_mm_prefix_lm = self.model_config.is_mm_prefix_lm
340

341
        # Multi-modal data support
342
        self.mm_registry = MULTIMODAL_REGISTRY
343
        self.uses_mrope = model_config.uses_mrope
344
        self.uses_xdrope_dim = model_config.uses_xdrope_dim
345
        self.supports_mm_inputs = self.mm_registry.supports_multimodal_inputs(
346
            model_config
347
        )
348

349
350
351
        if self.model_config.is_encoder_decoder:
            # Maximum length of the encoder input, only for encoder-decoder
            # models.
352
            self.max_encoder_len = scheduler_config.max_num_encoder_input_tokens
353
354
355
        else:
            self.max_encoder_len = 0

356
        # Sampler
357
        self.sampler = Sampler(logprobs_mode=self.model_config.logprobs_mode)
358

359
        self.eplb_state: EplbState | None = None
360
361
362
363
364
365
        """
        State of the expert parallelism load balancer.

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

366
        # Lazy initializations
367
        # self.model: nn.Module  # Set after load_model
368
        # Initialize in initialize_kv_cache
369
        self.kv_caches: list[torch.Tensor] = []
370
371
372
        # Initialize in initialize_kv_cache_tensors
        self.cross_layers_kv_cache: torch.Tensor | None = None
        self.cross_layers_attn_backend: type[AttentionBackend] | None = None
373
374
        # indexes: [kv_cache_group_id][attn_group]
        self.attn_groups: list[list[AttentionGroup]] = []
375
376
        # self.kv_cache_config: KVCacheConfig

377
378
        # mm_hash ->  encoder_output
        self.encoder_cache: dict[str, torch.Tensor] = {}
379

380
        self.use_aux_hidden_state_outputs = False
381
382
383
384
385
        # 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:
386
387
388
            self.drafter: (
                NgramProposer | SuffixDecodingProposer | EagleProposer | MedusaProposer
            )
389
390
            if self.speculative_config.method == "ngram":
                self.drafter = NgramProposer(self.vllm_config)
391
392
            elif self.speculative_config.method == "suffix":
                self.drafter = SuffixDecodingProposer(self.vllm_config)
393
            elif self.speculative_config.use_eagle():
394
                self.drafter = EagleProposer(self.vllm_config, self.device, self)
395
                if self.speculative_config.method == "eagle3":
396
397
398
                    self.use_aux_hidden_state_outputs = (
                        self.drafter.eagle3_use_aux_hidden_state
                    )
399
400
            elif self.speculative_config.method == "medusa":
                self.drafter = MedusaProposer(
401
                    vllm_config=self.vllm_config, device=self.device
402
                )
403
            else:
404
405
406
407
                raise ValueError(
                    "Unknown speculative decoding method: "
                    f"{self.speculative_config.method}"
                )
408
            self.rejection_sampler = RejectionSampler(self.sampler)
409

410
411
412
413
        self.num_spec_tokens = 0
        if self.speculative_config:
            self.num_spec_tokens = self.speculative_config.num_speculative_tokens

414
        # Request states.
415
        self.requests: dict[str, CachedRequestState] = {}
416
417
418
        # NOTE(rob): num_prompt_logprobs only includes reqs
        # that are currently in the prefill phase.
        self.num_prompt_logprobs: dict[str, int] = {}
419
        self.comm_stream = torch.cuda.Stream()
420

421
422
423
424
425
426
427
428
429
        # 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.
430
431
432
433
        logits_processors = model_config.logits_processors
        custom_logitsprocs: Sequence[str | type[LogitsProcessor]] = (
            tuple(logits_processors) if logits_processors is not None else ()
        )
434
435
        self.input_batch = InputBatch(
            max_num_reqs=self.max_num_reqs,
436
437
438
            # We need to use the encoder length for encoder-decoer
            # because of KV cache for cross-attention.
            max_model_len=max(self.max_model_len, self.max_encoder_len),
439
440
441
            max_num_batched_tokens=self.max_num_tokens,
            device=self.device,
            pin_memory=self.pin_memory,
442
            vocab_size=self.model_config.get_vocab_size(),
443
            block_sizes=[self.cache_config.block_size],
444
            kernel_block_sizes=[self.cache_config.block_size],
445
            is_spec_decode=bool(self.vllm_config.speculative_config),
446
            logitsprocs=build_logitsprocs(
447
448
449
                self.vllm_config,
                self.device,
                self.pin_memory,
450
                self.is_pooling_model,
451
                custom_logitsprocs,
452
            ),
453
454
455
            # We currently don't know whether a particular custom logits processor
            # uses output token ids so we set this conservatively.
            logitsprocs_need_output_token_ids=bool(custom_logitsprocs),
456
            is_pooling_model=self.is_pooling_model,
457
            cp_kv_cache_interleave_size=self.parallel_config.cp_kv_cache_interleave_size,
458
        )
459

460
        self.use_async_scheduling = self.scheduler_config.async_scheduling
461
462
463
464
465
        # Separate cuda stream for overlapping transfer of sampled token ids from
        # GPU to CPU when async scheduling is enabled.
        self.async_output_copy_stream: torch.cuda.Stream | None = None
        # cuda event to synchronize use of reused CPU tensors between steps
        # when async scheduling is enabled.
466
        self.prepare_inputs_event: torch.Event | None = None
467
468
        if self.use_async_scheduling:
            self.async_output_copy_stream = torch.cuda.Stream()
469
            self.prepare_inputs_event = torch.Event()
470

471
        # self.cudagraph_batch_sizes sorts in ascending order.
472
473
474
475
        if (
            self.compilation_config.cudagraph_capture_sizes
            and self.compilation_config.cudagraph_mode != CUDAGraphMode.NONE
        ):
476
477
            self.cudagraph_batch_sizes = sorted(
                self.compilation_config.cudagraph_capture_sizes
478
            )
479

480
        # Cache the device properties.
481
        self._init_device_properties()
482

483
        # Persistent buffers for CUDA graphs.
484
485
486
487
488
        self.input_ids = self._make_buffer(self.max_num_tokens, dtype=torch.int32)
        self.positions = self._make_buffer(self.max_num_tokens, dtype=torch.int64)
        self.query_start_loc = self._make_buffer(
            self.max_num_reqs + 1, dtype=torch.int32
        )
489
        self.seq_lens = self._make_buffer(self.max_num_reqs, dtype=torch.int32)
490
        self.encoder_seq_lens = self._make_buffer(self.max_num_reqs, dtype=torch.int32)
491
492
493
494
        if self.dcp_world_size > 1:
            self.dcp_local_seq_lens = self._make_buffer(
                self.max_num_reqs, dtype=torch.int32
            )
495
496
497
        # Because inputs_embeds may be bfloat16 and we don't need a numpy
        # version of this tensor, avoid a RuntimeError by not creating a
        # numpy buffer.
498
        self.inputs_embeds = self._make_buffer(
499
            self.max_num_tokens, self.inputs_embeds_size, dtype=self.dtype, numpy=False
500
501
        )
        self.is_token_ids = self._make_buffer(self.max_num_tokens, dtype=torch.bool)
502
503
        self.discard_request_mask = self._make_buffer(
            self.max_num_reqs, dtype=torch.bool
504
505
506
507
508
509
510
        )
        self.num_decode_draft_tokens = self._make_buffer(
            self.max_num_reqs, dtype=torch.int32
        )
        self.num_accepted_tokens = self._make_buffer(
            self.max_num_reqs, dtype=torch.int64
        )
511

512
513
        # Only relevant for multimodal models
        if self.supports_mm_inputs:
514
            self.is_mm_embed = self._make_buffer(self.max_num_tokens, dtype=torch.bool)
515

516
        # Only relevant for models using M-RoPE (e.g, Qwen2-VL)
517
        if self.uses_mrope:
Roger Wang's avatar
Roger Wang committed
518
519
520
521
            # 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
522
523
524
525
526
527

            # 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
528
            self.mrope_positions = self._make_buffer(
529
530
                (3, self.max_num_tokens + 1), dtype=torch.int64
            )
531

532
533
534
535
536
537
538
        # Only relevant for models using XD-RoPE (e.g, HunYuan-VL)
        if self.uses_xdrope_dim > 0:
            # Similar to mrope but use assigned dimension number for RoPE, 4 as default.
            self.xdrope_positions = self._make_buffer(
                (self.uses_xdrope_dim, self.max_num_tokens + 1), dtype=torch.int64
            )

539
        # None in the first PP rank. The rest are set after load_model.
540
        self.intermediate_tensors: IntermediateTensors | None = None
541

542
        # OPTIMIZATION: Cache the tensors rather than creating them every step.
543
        # Keep in int64 to avoid overflow with long context
544
545
546
547
        self.arange_np = np.arange(
            max(self.max_num_reqs + 1, self.max_model_len, self.max_num_tokens),
            dtype=np.int64,
        )
548

549
550
551
552
553
        # 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] = {}
554
555
556
557
558
        self.kv_sharing_fast_prefill_eligible_layers: set[str] = set()

        self.kv_sharing_fast_prefill_logits_indices = None
        if self.cache_config.kv_sharing_fast_prefill:
            self.kv_sharing_fast_prefill_logits_indices = torch.zeros(
559
560
                self.max_num_tokens, dtype=torch.int32, device=self.device
            )
561

562
        self.uniform_decode_query_len = 1 + self.num_spec_tokens
563
564
565
566

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

567
568
        self.mm_budget = (
            MultiModalBudget(
569
                self.model_config,
570
571
572
573
574
575
                self.scheduler_config,
                self.mm_registry,
            )
            if self.supports_mm_inputs
            else None
        )
576

577
        self.reorder_batch_threshold: int | None = None
578

579
580
581
582
583
        # Attention layers that are only in the KVCacheConfig of the runner
        # (e.g., KV sharing, encoder-only attention), but not in the
        # KVCacheConfig of the scheduler.
        self.runner_only_attn_layers: set[str] = set()

584
        # Cached outputs.
585
        self._draft_token_ids: list[list[int]] | torch.Tensor | None = None
586
        self.transfer_event = torch.Event()
587
        self.sampled_token_ids_pinned_cpu = torch.empty(
588
            (self.max_num_reqs, 1),
589
590
            dtype=torch.int64,
            device="cpu",
591
592
            pin_memory=self.pin_memory,
        )
593

594
595
        # Pre-allocated tensor for copying valid sampled token counts to CPU,
        # with dedicated stream for overlapping and event for coordination.
596
        self.valid_sampled_token_count_event: torch.Event | None = None
597
598
        self.valid_sampled_token_count_copy_stream: torch.cuda.Stream | None = None
        if self.use_async_scheduling and self.num_spec_tokens:
599
            self.valid_sampled_token_count_event = torch.Event()
600
601
602
603
604
605
606
607
            self.valid_sampled_token_count_copy_stream = torch.cuda.Stream()
        self.valid_sampled_token_count_cpu = torch.empty(
            self.max_num_reqs,
            dtype=torch.int64,
            device="cpu",
            pin_memory=self.pin_memory,
        )

608
609
        # Ephemeral state transferred between execute_model() and sample_tokens().
        self.execute_model_state: ExecuteModelState | None = None
610
        self.kv_connector_output: KVConnectorOutput | None = None
611
        self.layerwise_nvtx_hooks_registered = False
612

613
614
615
616
    def reset_mm_cache(self) -> None:
        if self.mm_budget:
            self.mm_budget.reset_cache()

617
618
619
620
621
622
623
624
625
626
627
628
629
630
631
632
633
634
635
636
637
638
639
640
641
642
643
644
645
646
647
648
649
650
651
652
653
654
655
656
657
658
659
660
    @torch.inference_mode()
    def init_fp8_kv_scales(self) -> None:
        """
        Re-initialize the KV cache and FP8 scales after waking from sleep.
        1. Zero out the KV cache tensors to remove garbage data from re-allocation.
        2. Reset Attention layer scaling factors (_k_scale, _v_scale) to 1.0.
          If these are left at 0.0 (default after wake_up), all KV cache values
          become effectively zero, causing gibberish output.
        """
        if not self.cache_config.cache_dtype.startswith("fp8"):
            return

        kv_caches = getattr(self, "kv_caches", [])
        for cache_tensor in kv_caches:
            if cache_tensor is not None:
                cache_tensor.zero_()

        k_attr_names = ("_k_scale", "k_scale")
        v_attr_names = ("_v_scale", "v_scale")

        attn_layers = self.compilation_config.static_forward_context
        for name, module in attn_layers.items():
            if isinstance(module, (Attention, MLAAttention)):
                # TODO: Generally, scale is 1.0 if user uses on-the-fly fp8
                # kvcache quant. However, to get better accuracy, compression
                # frameworks like llm-compressors allow users to tune the
                # scale. We may need to restore the specific calibrated scales
                # here in the future.
                k_scale_val, v_scale_val = 1.0, 1.0

                # Processing K Scale
                for attr in k_attr_names:
                    if hasattr(module, attr):
                        param = getattr(module, attr)
                        if isinstance(param, torch.Tensor):
                            param.fill_(k_scale_val)

                # Processing V Scale
                for attr in v_attr_names:
                    if hasattr(module, attr):
                        param = getattr(module, attr)
                        if isinstance(param, torch.Tensor):
                            param.fill_(v_scale_val)

661
662
663
664
    def _get_positions(self, num_tokens: Any):
        if isinstance(num_tokens, int):
            if self.uses_mrope:
                return self.mrope_positions.gpu[:, :num_tokens]
665
666
            if self.uses_xdrope_dim > 0:
                return self.xdrope_positions.gpu[:, :num_tokens]
667
668
669
670
            return self.positions.gpu[:num_tokens]
        else:
            if self.uses_mrope:
                return self.mrope_positions.gpu[:, num_tokens]
671
672
            if self.uses_xdrope_dim > 0:
                return self.xdrope_positions.gpu[:, num_tokens]
673
674
            return self.positions.gpu[num_tokens]

675
    def _make_buffer(
676
        self, *size: int | torch.SymInt, dtype: torch.dtype, numpy: bool = True
677
678
679
680
681
682
683
684
    ) -> CpuGpuBuffer:
        return CpuGpuBuffer(
            *size,
            dtype=dtype,
            device=self.device,
            pin_memory=self.pin_memory,
            with_numpy=numpy,
        )
685

686
687
688
    def _init_model_kwargs(self, num_tokens: int):
        model_kwargs = dict[str, Any]()

689
        if not self.is_pooling_model:
690
691
            return model_kwargs

692
693
        num_reqs = self.input_batch.num_reqs
        pooling_params = self.input_batch.get_pooling_params()
694
695
696

        token_type_id_requests = dict[int, Any]()
        for i, param in enumerate(pooling_params):
697
698
699
700
701
            if (
                param.extra_kwargs is not None
                and (token_types := param.extra_kwargs.get("compressed_token_type_ids"))
                is not None
            ):
702
703
704
705
706
                token_type_id_requests[i] = token_types

        if len(token_type_id_requests) == 0:
            return model_kwargs

707
        seq_lens = self.seq_lens.gpu[:num_reqs]
708
709
710
711
712
713
714
715
        token_type_ids = []

        for i in range(num_reqs):
            pos = token_type_id_requests.get(i, seq_lens[i])
            ids = (torch.arange(seq_lens[i]) >= pos).int()
            token_type_ids.append(ids)

        model_kwargs["token_type_ids"] = torch.concat(token_type_ids).to(
716
717
            device=self.device
        )
718
719
        return model_kwargs

720
    def _may_reorder_batch(self, scheduler_output: "SchedulerOutput") -> None:
721
722
        """
        Update the order of requests in the batch based on the attention
723
        backend's needs. For example, some attention backends (namely MLA) may
724
725
726
727
728
729
        want to separate requests based on if the attention computation will be
        compute-bound or memory-bound.

        Args:
            scheduler_output: The scheduler output.
        """
730
731
732
733
734
735
736
737
        # Attention free models have zero kv_cache_goups, however models
        # like Mamba are also attention free but use the kv_cache for
        # keeping its internal state. This is why we check the number
        # of kv_cache groups instead of solely checking
        # for self.model_config.is_attention_free.
        if len(self.kv_cache_config.kv_cache_groups) == 0:
            return

738
739
740
741
        if self.reorder_batch_threshold is not None:
            reorder_batch_to_split_decodes_and_prefills(
                self.input_batch,
                scheduler_output,
742
743
                decode_threshold=self.reorder_batch_threshold,
            )
744

745
746
    # Note: used for model runner override.
    def _init_device_properties(self) -> None:
747
        """Initialize attributes from torch.cuda.get_device_properties"""
748
749
750
751
752
753
754
        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()

755
    def _update_states(self, scheduler_output: "SchedulerOutput") -> None:
756
757
758
759
760
761
        """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.

762
763
        The SamplingMetadata is updated and copied to the GPU if there is a
        new/resumed/paused/finished request in the batch.
764
765
        """
        # Remove finished requests from the cached states.
766
767
        for req_id in scheduler_output.finished_req_ids:
            self.requests.pop(req_id, None)
768
            self.num_prompt_logprobs.pop(req_id, None)
769
770
771
772
773
774
775
        # Remove the finished requests from the persistent batch.
        # NOTE(woosuk): There could be an edge case where finished_req_ids and
        # scheduled_req_ids overlap. This happens when a request is aborted and
        # then resubmitted with the same ID. In this case, we treat them as two
        # distinct requests - clearing the cached states for the first request
        # and handling the second as a new request.
        for req_id in scheduler_output.finished_req_ids:
776
            self.input_batch.remove_request(req_id)
777
778

        # Free the cached encoder outputs.
779
780
        for mm_hash in scheduler_output.free_encoder_mm_hashes:
            self.encoder_cache.pop(mm_hash, None)
781

782
783
784
785
786
787
788
        # 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()
789
790
791
792
793
794
795
796
        resumed_req_ids = scheduler_output.scheduled_cached_reqs.resumed_req_ids
        # NOTE(zhuohan): cached_req_ids and resumed_req_ids are usually disjoint,
        # so `(scheduled_req_ids - resumed_req_ids) == scheduled_req_ids` holds
        # apart from the forced-preemption case in reset_prefix_cache. And in
        # that case we include the resumed_req_ids in the unscheduled set so
        # that they get cleared from the persistent batch before being re-scheduled
        # in the normal resumed request path.
        unscheduled_req_ids = cached_req_ids - (scheduled_req_ids - resumed_req_ids)
797
798
799
800
801
        # 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:
802
            self.input_batch.remove_request(req_id)
803

804
        reqs_to_add: list[CachedRequestState] = []
805
        # Add new requests to the cached states.
806
807
808
        for new_req_data in scheduler_output.scheduled_new_reqs:
            req_id = new_req_data.req_id
            sampling_params = new_req_data.sampling_params
809
            pooling_params = new_req_data.pooling_params
810

811
812
813
814
            if (
                sampling_params
                and sampling_params.sampling_type == SamplingType.RANDOM_SEED
            ):
815
816
817
818
819
                generator = torch.Generator(device=self.device)
                generator.manual_seed(sampling_params.seed)
            else:
                generator = None

820
821
            if self.is_pooling_model:
                assert pooling_params is not None
822
823
                task = pooling_params.task
                assert task is not None, "You did not set `task` in the API"
824

825
                model = cast(VllmModelForPooling, self.get_model())
826
                to_update = model.pooler.get_pooling_updates(task)
827
828
                to_update.apply(pooling_params)

829
            req_state = CachedRequestState(
830
                req_id=req_id,
831
                prompt_token_ids=new_req_data.prompt_token_ids,
832
                prompt_embeds=new_req_data.prompt_embeds,
833
                mm_features=new_req_data.mm_features,
834
                sampling_params=sampling_params,
835
                pooling_params=pooling_params,
836
                generator=generator,
837
838
                block_ids=new_req_data.block_ids,
                num_computed_tokens=new_req_data.num_computed_tokens,
839
                output_token_ids=[],
840
                lora_request=new_req_data.lora_request,
841
            )
842
843
            self.requests[req_id] = req_state

844
845
846
847
848
849
850
            if sampling_params and sampling_params.prompt_logprobs is not None:
                self.num_prompt_logprobs[req_id] = (
                    self.input_batch.vocab_size
                    if sampling_params.prompt_logprobs == -1
                    else sampling_params.prompt_logprobs
                )

851
            # Only relevant for models using M-RoPE (e.g, Qwen2-VL)
852
            if self.uses_mrope:
853
                self._init_mrope_positions(req_state)
854

855
856
857
858
            # Only relevant for models using XD-RoPE (e.g, HunYuan-VL)
            if self.uses_xdrope_dim > 0:
                self._init_xdrope_positions(req_state)

859
            reqs_to_add.append(req_state)
860

861
        # Update the states of the running/resumed requests.
862
        is_last_rank = get_pp_group().is_last_rank
863
        req_data = scheduler_output.scheduled_cached_reqs
864
865
866
867
868

        # Wait until valid_sampled_tokens_count is copied to cpu,
        # then use it to update actual num_computed_tokens of each request.
        valid_sampled_token_count = self._get_valid_sampled_token_count()

869
        for i, req_id in enumerate(req_data.req_ids):
870
            req_state = self.requests[req_id]
871
872
            num_computed_tokens = req_data.num_computed_tokens[i]
            new_block_ids = req_data.new_block_ids[i]
873
            resumed_from_preemption = req_id in req_data.resumed_req_ids
874
            num_output_tokens = req_data.num_output_tokens[i]
875
            req_index = self.input_batch.req_id_to_index.get(req_id)
876

877
878
879
880
881
882
883
884
885
886
887
888
889
890
891
892
893
894
895
896
897
898
899
            # prev_num_draft_len is used in async scheduling mode with
            # spec decode. it indicates if need to update num_computed_tokens
            # of the request. for example:
            # fist step: num_computed_tokens = 0, spec_tokens = [],
            # prev_num_draft_len = 0.
            # second step: num_computed_tokens = 100(prompt lenth),
            # spec_tokens = [a,b], prev_num_draft_len = 0.
            # third step: num_computed_tokens = 100 + 2, spec_tokens = [c,d],
            # prev_num_draft_len = 2.
            # num_computed_tokens in first step and second step does't contain
            # the spec tokens length, but in third step it contains the
            # spec tokens length. we only need to update num_computed_tokens
            # when prev_num_draft_len > 0.
            if req_state.prev_num_draft_len:
                if req_index is None:
                    req_state.prev_num_draft_len = 0
                else:
                    assert self.input_batch.prev_req_id_to_index is not None
                    prev_req_index = self.input_batch.prev_req_id_to_index[req_id]
                    num_accepted = valid_sampled_token_count[prev_req_index] - 1
                    num_rejected = req_state.prev_num_draft_len - num_accepted
                    num_computed_tokens -= num_rejected
                    req_state.output_token_ids.extend([-1] * num_accepted)
900

901
            # Update the cached states.
902
            req_state.num_computed_tokens = num_computed_tokens
903
904
905
906
907
908
909
910

            if not is_last_rank:
                # When using PP, the scheduler sends the sampled tokens back,
                # because there's no direct communication between the first-
                # stage worker and the last-stage worker.
                new_token_ids = req_data.new_token_ids[i]
                # Add the sampled token(s) from the previous step (if any).
                # This doesn't include "unverified" tokens like spec tokens.
911
912
913
                num_new_tokens = (
                    num_computed_tokens + len(new_token_ids) - req_state.num_tokens
                )
914
915
916
917
                if num_new_tokens == 1:
                    # Avoid slicing list in most common case.
                    req_state.output_token_ids.append(new_token_ids[-1])
                elif num_new_tokens > 0:
918
                    req_state.output_token_ids.extend(new_token_ids[-num_new_tokens:])
919
920
921
922
923
            elif num_output_tokens < len(req_state.output_token_ids):
                # Some output tokens were discarded due to a sync-KV-load
                # failure. Align the cached state.
                del req_state.output_token_ids[num_output_tokens:]
                if req_index is not None:
924
925
926
927
                    end_idx = (
                        self.input_batch.num_prompt_tokens[req_index]
                        + num_output_tokens
                    )
928
929
                    self.input_batch.num_tokens[req_index] = end_idx
                    self.input_batch.num_tokens_no_spec[req_index] = end_idx
930

931
            # Update the block IDs.
932
            if not resumed_from_preemption:
933
934
                if new_block_ids is not None:
                    # Append the new blocks to the existing block IDs.
935
                    for block_ids, new_ids in zip(req_state.block_ids, new_block_ids):
936
                        block_ids.extend(new_ids)
937
            else:
938
                assert req_index is None
939
                assert new_block_ids is not None
940
941
                # The request is resumed from preemption.
                # Replace the existing block IDs with the new ones.
942
                req_state.block_ids = new_block_ids
943
944
945
946
947

            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.
948
949
950
951
952
953
954

                if self.use_async_scheduling and num_output_tokens > 0:
                    # We must recover the output token ids for resumed requests in the
                    # async scheduling case, so that correct input_ids are obtained.
                    resumed_token_ids = req_data.all_token_ids[req_id]
                    req_state.output_token_ids = resumed_token_ids[-num_output_tokens:]

955
                reqs_to_add.append(req_state)
956
957
958
                continue

            # Update the persistent batch.
959
            self.input_batch.num_computed_tokens_cpu[req_index] = num_computed_tokens
960
            if new_block_ids is not None:
961
                self.input_batch.block_table.append_row(new_block_ids, req_index)
962
963
964
965
966
967
968

            # For the last rank, we don't need to update the token_ids_cpu
            # because the sampled tokens are already cached.
            if not is_last_rank:
                # Add new_token_ids to token_ids_cpu.
                start_token_index = num_computed_tokens
                end_token_index = num_computed_tokens + len(new_token_ids)
969
                self.input_batch.token_ids_cpu[
970
971
972
                    req_index, start_token_index:end_token_index
                ] = new_token_ids
                self.input_batch.num_tokens_no_spec[req_index] = end_token_index
973
                self.input_batch.num_tokens[req_index] = end_token_index
974

975
            # Add spec_token_ids to token_ids_cpu.
976
            spec_token_ids = scheduler_output.scheduled_spec_decode_tokens.get(
977
                req_id, []
978
            )
979
980
981
982
983
            num_spec_tokens = len(spec_token_ids)
            # For async scheduling, token_ids_cpu assigned from
            # spec_token_ids are placeholders and will be overwritten in
            # _prepare_input_ids.
            if num_spec_tokens:
984
985
986
                start_index = self.input_batch.num_tokens_no_spec[req_index]
                end_token_index = start_index + num_spec_tokens
                self.input_batch.token_ids_cpu[
987
988
                    req_index, start_index:end_token_index
                ] = spec_token_ids
989
990
                # NOTE(woosuk): `num_tokens` here may include spec tokens.
                self.input_batch.num_tokens[req_index] += num_spec_tokens
991
992
993
994
995
996

            # When speculative decoding is used with structured output,
            # the scheduler can drop draft tokens that do not
            # conform to the schema. This can result in
            # scheduler_output.scheduled_spec_decode_tokens being empty,
            # even when speculative decoding is enabled.
997
998
            self.input_batch.spec_token_ids[req_index].clear()
            self.input_batch.spec_token_ids[req_index].extend(spec_token_ids)
999

1000
1001
1002
1003
1004
1005
1006
1007
1008
            # there are no draft tokens with async scheduling,
            # we clear the spec_decoding info in scheduler_output and
            # use normal sampling but rejection_sampling.
            if self.use_async_scheduling:
                req_state.prev_num_draft_len = num_spec_tokens
                if num_spec_tokens and self._draft_token_ids is None:
                    scheduler_output.total_num_scheduled_tokens -= num_spec_tokens
                    scheduler_output.num_scheduled_tokens[req_id] -= num_spec_tokens
                    scheduler_output.scheduled_spec_decode_tokens.pop(req_id, None)
1009
1010
        # Add the new or resumed requests to the persistent batch.
        # The smaller empty indices are filled first.
1011
1012
        for request in reqs_to_add:
            self.input_batch.add_request(request)
1013

1014
1015
1016
1017
1018
1019
        # Condense the batched states if there are gaps left by removed requests
        self.input_batch.condense()
        # Allow attention backend to reorder the batch, potentially
        self._may_reorder_batch(scheduler_output)
        # Refresh batch metadata with any pending updates.
        self.input_batch.refresh_metadata()
1020

1021
    def _update_states_after_model_execute(
1022
1023
        self, output_token_ids: torch.Tensor
    ) -> None:
1024
1025
1026
1027
1028
1029
1030
1031
1032
1033
1034
1035
        """Update the cached states after model execution.

        This is used for MTP/EAGLE for hybrid models, as in linear attention,
        only the last token's state is kept. In MTP/EAGLE, for draft tokens
        the state are kept util we decide how many tokens are accepted for
        each sequence, and a shifting is done during the next iteration
        based on the number of accepted tokens.
        """
        if not self.model_config.is_hybrid or not self.speculative_config:
            return

        # Find the number of accepted tokens for each sequence.
1036
1037
1038
1039
1040
1041
1042
1043
1044
1045
1046
1047
1048
1049
1050
1051
1052
1053
1054
1055
        num_accepted_tokens = (
            (
                torch.cat(
                    [
                        output_token_ids,
                        torch.full(
                            (output_token_ids.size(0), 1),
                            -1,
                            device=output_token_ids.device,
                        ),
                    ],
                    dim=1,
                )
                == -1
            )
            .int()
            .argmax(-1)
            .cpu()
            .numpy()
        )
1056
1057
1058
        for i, num_tokens in enumerate(num_accepted_tokens):
            self.input_batch.num_accepted_tokens_cpu[i] = num_tokens

1059
    def _init_mrope_positions(self, req_state: CachedRequestState):
1060
1061
        model = self.get_model()
        assert supports_mrope(model), "M-RoPE support is not implemented."
1062
1063
1064
1065
        assert req_state.prompt_token_ids is not None, (
            "M-RoPE requires prompt_token_ids to be available."
        )
        mrope_model = cast(SupportsMRoPE, model)
1066
1067

        req_state.mrope_positions, req_state.mrope_position_delta = (
1068
            mrope_model.get_mrope_input_positions(
1069
                req_state.prompt_token_ids,
1070
                req_state.mm_features,
1071
            )
1072
        )
1073

1074
1075
1076
1077
1078
1079
1080
1081
1082
1083
1084
1085
1086
    def _init_xdrope_positions(self, req_state: CachedRequestState):
        model = self.get_model()
        xdrope_model = cast(SupportsXDRoPE, model)
        assert req_state.prompt_token_ids is not None, (
            "XD-RoPE requires prompt_token_ids to be available."
        )
        assert supports_xdrope(model), "XD-RoPE support is not implemented."

        req_state.xdrope_positions = xdrope_model.get_xdrope_input_positions(
            req_state.prompt_token_ids,
            req_state.mm_features,
        )

1087
    def _extract_mm_kwargs(
1088
        self,
1089
1090
        scheduler_output: "SchedulerOutput",
    ) -> BatchedTensorInputs:
1091
        if not scheduler_output or not self.is_multimodal_raw_input_only_model:
1092
            return {}
1093

1094
1095
        mm_kwargs = list[MultiModalKwargsItem]()
        for req in scheduler_output.scheduled_new_reqs:
1096
1097
1098
            for feature in req.mm_features:
                if feature.data is not None:
                    mm_kwargs.append(feature.data)
1099

1100
        # Input all modalities at once
1101
        model = cast(SupportsMultiModal, self.model)
1102
1103
        mm_kwargs_combined: BatchedTensorInputs = {}
        for _, _, mm_kwargs_group in group_mm_kwargs_by_modality(
1104
1105
1106
1107
            mm_kwargs,
            device=self.device,
            pin_memory=self.pin_memory,
            merge_by_field_config=model.merge_by_field_config,
1108
1109
        ):
            mm_kwargs_combined.update(mm_kwargs_group)
1110

1111
        return mm_kwargs_combined
1112

1113
    def _dummy_mm_kwargs(self, num_seqs: int) -> BatchedTensorInputs:
1114
        if not self.is_multimodal_raw_input_only_model:
1115
            return {}
1116

1117
1118
1119
1120
1121
        mm_budget = self.mm_budget
        assert mm_budget is not None

        dummy_modality = mm_budget.get_modality_with_max_tokens()
        return self._get_mm_dummy_batch(dummy_modality, num_seqs)
1122

1123
1124
1125
    def _get_cumsum_and_arange(
        self,
        num_tokens: np.ndarray,
1126
        cumsum_dtype: np.dtype | None = None,
1127
1128
1129
1130
1131
1132
1133
1134
1135
1136
1137
1138
1139
1140
1141
1142
    ) -> 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

1143
    def _prepare_input_ids(
1144
1145
1146
1147
        self,
        scheduler_output: "SchedulerOutput",
        total_num_scheduled_tokens: int,
        cu_num_tokens: np.ndarray,
1148
    ) -> None:
1149
        """Prepare the input IDs for the current batch.
1150

1151
1152
1153
1154
1155
1156
1157
        Carefully handles the `prev_sampled_token_ids` which can be cached
        from the previous engine iteration, in which case those tokens on the
        GPU need to be copied into the corresponding slots into input_ids."""

        if self.input_batch.prev_sampled_token_ids is None:
            # Normal scheduling case
            self.input_ids.copy_to_gpu(total_num_scheduled_tokens)
1158
1159
1160
            if self.enable_prompt_embeds:
                self.inputs_embeds.copy_to_gpu(total_num_scheduled_tokens)
                self.is_token_ids.copy_to_gpu(total_num_scheduled_tokens)
1161
1162
1163
1164
1165
1166
1167
            return

        # Async scheduling case, where some decode requests from the previous
        # iteration won't have entries in input_ids_cpu and need to be copied
        # on the GPU from prev_sampled_token_ids.
        prev_req_id_to_index = self.input_batch.prev_req_id_to_index
        assert prev_req_id_to_index is not None
1168
1169
1170
1171
        sample_flattened_indices: list[int] = []
        spec_flattened_indices: list[int] = []
        prev_common_req_indices: list[int] = []
        prev_draft_token_indices: list[int] = []
1172
1173
        indices_match = True
        max_flattened_index = -1
1174
1175
1176
        total_num_spec_tokens = 0
        scheduled_spec_tokens = scheduler_output.scheduled_spec_decode_tokens

1177
1178
1179
1180
1181
        for req_id, cur_index in self.input_batch.req_id_to_index.items():
            if (prev_index := prev_req_id_to_index.get(req_id)) is not None:
                prev_common_req_indices.append(prev_index)
                # We need to compute the flattened input_ids index of the
                # last token in each common request.
1182
1183
                draft_len = len(scheduled_spec_tokens.get(req_id, ()))
                total_num_spec_tokens += draft_len
1184
                flattened_index = cu_num_tokens[cur_index].item() - 1
1185
1186
1187
1188
1189
1190
1191
1192
1193
1194
1195
1196
1197
1198
1199
                # example: cu_num_tokens = [2, 5, 8], draft_tokens = [1, 2, 2]
                # sample_flattened_indices = [0, 2, 5]
                # spec_flattened_indices = [1,   3, 4,    6, 7]
                sample_flattened_indices.append(flattened_index - draft_len)
                spec_flattened_indices.extend(
                    range(flattened_index - draft_len + 1, flattened_index + 1)
                )
                start = prev_index * self.num_spec_tokens
                # prev_draft_token_indices is used to find which draft_tokens_id
                # should be copied to input_ids
                # example: prev draft_tokens_id [[1,2], [3,4], [5, 6]]
                # flatten draft_tokens_id [1,2,3,4,5,6]
                # draft_len of each request [1, 2, 1]
                # then prev_draft_token_indices is [0,   2, 3,   4]
                prev_draft_token_indices.extend(range(start, start + draft_len))
1200
                indices_match &= prev_index == flattened_index
1201
                max_flattened_index = max(max_flattened_index, flattened_index)
1202
1203
1204
        num_commmon_tokens = len(sample_flattened_indices)
        total_without_spec = total_num_scheduled_tokens - total_num_spec_tokens
        if num_commmon_tokens < total_without_spec:
1205
1206
1207
            # If not all requests are decodes from the last iteration,
            # We need to copy the input_ids_cpu to the GPU first.
            self.input_ids.copy_to_gpu(total_num_scheduled_tokens)
1208
1209
1210
            if self.enable_prompt_embeds:
                self.inputs_embeds.copy_to_gpu(total_num_scheduled_tokens)
                self.is_token_ids.copy_to_gpu(total_num_scheduled_tokens)
1211
1212
        if num_commmon_tokens == 0:
            # No requests in common with the previous iteration
1213
            # So input_ids.cpu will have all the input ids.
1214
1215
1216
1217
1218
1219
1220
            return
        if indices_match and max_flattened_index == (num_commmon_tokens - 1):
            # Common-case optimization: the batch is unchanged
            # and no reordering happened.
            # The indices are both the same permutation of 0..N-1 so
            # we can copy directly using a single slice.
            self.input_ids.gpu[:num_commmon_tokens].copy_(
1221
1222
1223
                self.input_batch.prev_sampled_token_ids[:num_commmon_tokens, 0],
                non_blocking=True,
            )
1224
1225
            if self.enable_prompt_embeds:
                self.is_token_ids.gpu[:num_commmon_tokens] = True
1226
            return
1227
        # Upload the index tensors asynchronously so the scatter can be non-blocking.
1228
1229
        sampled_tokens_index_tensor = torch.tensor(
            sample_flattened_indices, dtype=torch.int64, pin_memory=self.pin_memory
1230
        ).to(self.device, non_blocking=True)
1231
        prev_common_req_indices_tensor = torch.tensor(
1232
1233
            prev_common_req_indices, dtype=torch.int64, pin_memory=self.pin_memory
        ).to(self.device, non_blocking=True)
1234
1235
        self.input_ids.gpu.scatter_(
            dim=0,
1236
            index=sampled_tokens_index_tensor,
1237
            src=self.input_batch.prev_sampled_token_ids[
1238
1239
1240
                prev_common_req_indices_tensor, 0
            ],
        )
1241

1242
1243
1244
1245
1246
1247
1248
1249
1250
1251
1252
1253
1254
1255
1256
1257
1258
1259
1260
1261
1262
1263
1264
        # Scatter the draft tokens after the sampled tokens are scattered.
        if self._draft_token_ids is None or not spec_flattened_indices:
            return

        assert isinstance(self._draft_token_ids, torch.Tensor)
        draft_tokens_index_tensor = torch.tensor(
            spec_flattened_indices, dtype=torch.int64, pin_memory=self.pin_memory
        ).to(self.device, non_blocking=True)
        prev_draft_token_indices_tensor = torch.tensor(
            prev_draft_token_indices, dtype=torch.int64, pin_memory=self.pin_memory
        ).to(self.device, non_blocking=True)

        # because input_ids dtype is torch.int32,
        # so convert draft_token_ids to torch.int32 here.
        draft_token_ids = self._draft_token_ids.to(dtype=torch.int32)
        self._draft_token_ids = None

        self.input_ids.gpu.scatter_(
            dim=0,
            index=draft_tokens_index_tensor,
            src=draft_token_ids.flatten()[prev_draft_token_indices_tensor],
        )

1265
1266
    def _get_encoder_seq_lens(
        self,
1267
        num_scheduled_tokens: dict[str, int],
1268
1269
        kv_cache_spec: KVCacheSpec,
        num_reqs: int,
1270
    ) -> tuple[torch.Tensor | None, np.ndarray | None]:
1271
        if not isinstance(kv_cache_spec, CrossAttentionSpec):
1272
            return None, None
1273

1274
1275
        # Zero out buffer for padding requests that are not actually scheduled (CGs)
        self.encoder_seq_lens.np[:num_reqs] = 0
1276
1277
        # Build encoder_seq_lens array mapping request indices to
        # encoder lengths for inputs scheduled in this batch
1278
        for req_id in num_scheduled_tokens:
1279
            req_index = self.input_batch.req_id_to_index[req_id]
1280
1281
1282
1283
1284
1285
1286
1287
1288
1289
1290
1291
1292
1293
1294
1295
            req_state = self.requests[req_id]
            if req_state.mm_features is None:
                self.encoder_seq_lens.np[req_index] = 0
                continue

            # Get the total number of encoder input tokens for running encoder requests
            # whether encoding is finished or not so that cross-attention knows how
            # many encoder tokens to attend to.
            encoder_input_tokens = sum(
                feature.mm_position.length for feature in req_state.mm_features
            )
            self.encoder_seq_lens.np[req_index] = encoder_input_tokens

        self.encoder_seq_lens.copy_to_gpu(num_reqs)
        encoder_seq_lens = self.encoder_seq_lens.gpu[:num_reqs]
        encoder_seq_lens_cpu = self.encoder_seq_lens.np[:num_reqs]
1296

1297
        return encoder_seq_lens, encoder_seq_lens_cpu
1298

1299
    def _prepare_inputs(
1300
1301
1302
        self,
        scheduler_output: "SchedulerOutput",
        num_scheduled_tokens: np.ndarray,
1303
1304
    ) -> tuple[
        torch.Tensor,
1305
        SpecDecodeMetadata | None,
1306
    ]:
1307
1308
        """
        :return: tuple[
1309
            logits_indices, spec_decode_metadata,
1310
1311
        ]
        """
1312
1313
1314
1315
1316
1317
1318
        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.
1319
        self.input_batch.block_table.commit_block_table(num_reqs)
1320
1321
1322

        # Get request indices.
        # E.g., [2, 5, 3] -> [0, 0, 1, 1, 1, 1, 1, 2, 2, 2]
1323
        req_indices = np.repeat(self.arange_np[:num_reqs], num_scheduled_tokens)
1324

1325
1326
        # cu_num_tokens: [2, 5, 3] -> [2, 7, 10]
        # arange: [0, 1, 0, 1, 2, 3, 4, 0, 1, 2]
1327
        cu_num_tokens, arange = self._get_cumsum_and_arange(num_scheduled_tokens)
1328
1329

        # Get positions.
1330
        positions_np = self.positions.np[:total_num_scheduled_tokens]
1331
1332
1333
1334
1335
        np.add(
            self.input_batch.num_computed_tokens_cpu[req_indices],
            arange,
            out=positions_np,
        )
1336

1337
1338
        # Calculate M-RoPE positions.
        # Only relevant for models using M-RoPE (e.g, Qwen2-VL)
1339
        if self.uses_mrope:
1340
1341
            self._calc_mrope_positions(scheduler_output)

1342
1343
1344
1345
1346
        # Calculate XD-RoPE positions.
        # Only relevant for models using XD-RoPE (e.g, HunYuan-VL)
        if self.uses_xdrope_dim > 0:
            self._calc_xdrope_positions(scheduler_output)

1347
1348
1349
1350
        # 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.
1351
1352
1353
        token_indices = (
            positions_np + req_indices * self.input_batch.token_ids_cpu.shape[1]
        )
1354
        token_indices_tensor = torch.from_numpy(token_indices)
1355

1356
1357
1358
        # 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.
1359
1360
1361
1362
1363
1364
        torch.index_select(
            self.input_batch.token_ids_cpu_tensor.flatten(),
            0,
            token_indices_tensor,
            out=self.input_ids.cpu[:total_num_scheduled_tokens],
        )
1365
        if self.enable_prompt_embeds:
1366
            is_token_ids = self.input_batch.is_token_ids_tensor.flatten()
1367
1368
1369
1370
            torch.index_select(
                is_token_ids,
                0,
                token_indices_tensor,
1371
1372
                out=self.is_token_ids.cpu[:total_num_scheduled_tokens],
            )
1373
1374
1375
1376
1377
1378
1379
1380
1381
1382
1383
1384
1385
1386
1387
1388
1389
1390
1391
1392
1393
1394
1395
1396
1397
1398
1399
1400
1401
1402
1403
1404
1405

        # Because we did not pre-allocate a massive prompt_embeds CPU tensor on
        # the InputBatch, we need to fill in the prompt embeds into the expected
        # spots in the GpuModelRunner's pre-allocated prompt_embeds tensor.
        if self.input_batch.req_prompt_embeds:
            output_idx = 0
            for req_idx in range(num_reqs):
                num_sched = num_scheduled_tokens[req_idx]

                # Skip if this request doesn't have embeddings
                if req_idx not in self.input_batch.req_prompt_embeds:
                    output_idx += num_sched
                    continue

                # Skip if no tokens scheduled
                if num_sched <= 0:
                    output_idx += num_sched
                    continue

                req_embeds = self.input_batch.req_prompt_embeds[req_idx]
                start_pos = self.input_batch.num_computed_tokens_cpu[req_idx]

                # Skip if trying to read beyond available embeddings
                if start_pos >= req_embeds.shape[0]:
                    output_idx += num_sched
                    continue

                # Copy available embeddings
                end_pos = start_pos + num_sched
                actual_end = min(end_pos, req_embeds.shape[0])
                actual_num_sched = actual_end - start_pos

                if actual_num_sched > 0:
1406
1407
1408
                    self.inputs_embeds.cpu[
                        output_idx : output_idx + actual_num_sched
                    ].copy_(req_embeds[start_pos:actual_end])
1409
1410

                output_idx += num_sched
1411

1412
1413
        self.input_batch.block_table.compute_slot_mapping(req_indices, positions_np)
        self.input_batch.block_table.commit_slot_mapping(total_num_scheduled_tokens)
1414
1415

        # Prepare the attention metadata.
1416
        self.query_start_loc.np[0] = 0
1417
        self.query_start_loc.np[1 : num_reqs + 1] = cu_num_tokens
1418
1419
        # Note: pad query_start_loc to be non-decreasing, as kernels
        # like FlashAttention requires that
1420
        self.query_start_loc.np[num_reqs + 1 :].fill(cu_num_tokens[-1])
1421
        self.query_start_loc.copy_to_gpu()
1422
        query_start_loc = self.query_start_loc.gpu[: num_reqs + 1]
1423

1424
        self.seq_lens.np[:num_reqs] = (
1425
1426
            self.input_batch.num_computed_tokens_cpu[:num_reqs] + num_scheduled_tokens
        )
1427
        # Fill unused with 0 for full cuda graph mode.
1428
1429
        self.seq_lens.np[num_reqs:].fill(0)
        self.seq_lens.copy_to_gpu()
1430

1431
        num_tokens = [self.requests[r].num_tokens for r in self.input_batch.req_ids]
1432
1433
        num_tokens_np = np.array(num_tokens, dtype=np.int32)

1434
        # Record which requests should not be sampled,
1435
        # so that we could clear the sampled tokens before returning
1436
1437
        self.discard_request_mask.np[:num_reqs] = (
            self.seq_lens.np[:num_reqs] < num_tokens_np
1438
        )
1439
        self.discard_request_mask.copy_to_gpu(num_reqs)
1440

1441
        # Copy the tensors to the GPU.
1442
1443
1444
1445
1446
        self._prepare_input_ids(
            scheduler_output,
            total_num_scheduled_tokens,
            cu_num_tokens,
        )
1447

1448
        if self.uses_mrope:
1449
            # Only relevant for models using M-RoPE (e.g, Qwen2-VL)
1450
1451
            self.mrope_positions.gpu[:, :total_num_scheduled_tokens].copy_(
                self.mrope_positions.cpu[:, :total_num_scheduled_tokens],
1452
1453
                non_blocking=True,
            )
1454
1455
1456
1457
1458
1459
        elif self.uses_xdrope_dim > 0:
            # Only relevant for models using XD-RoPE (e.g, HunYuan-VL)
            self.xdrope_positions.gpu[:, :total_num_scheduled_tokens].copy_(
                self.xdrope_positions.cpu[:, :total_num_scheduled_tokens],
                non_blocking=True,
            )
1460
1461
        else:
            # Common case (1D positions)
1462
            self.positions.copy_to_gpu(total_num_scheduled_tokens)
1463

1464
        use_spec_decode = len(scheduler_output.scheduled_spec_decode_tokens) > 0
1465
1466
1467
1468
1469
1470
1471
        if not use_spec_decode:
            # NOTE(woosuk): Due to chunked prefills, the batch may contain
            # partial requests. While we should not sample any token
            # from these partial requests, we do so for simplicity.
            # We will ignore the sampled tokens from the partial requests.
            # TODO: Support prompt logprobs.
            logits_indices = query_start_loc[1:] - 1
1472
            num_draft_tokens = None
1473
            spec_decode_metadata = None
1474
            num_sampled_tokens = np.ones(num_reqs, dtype=np.int32)
1475
1476
1477
1478
1479
        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)
1480
1481
1482
            # For chunked prefills, use -1 as mask rather than 0, as guided
            # decoding may rollback speculative tokens.
            num_decode_draft_tokens = np.full(num_reqs, -1, dtype=np.int32)
1483
1484
1485
1486
            for (
                req_id,
                draft_token_ids,
            ) in scheduler_output.scheduled_spec_decode_tokens.items():
1487
1488
                req_idx = self.input_batch.req_id_to_index[req_id]
                num_draft_tokens[req_idx] = len(draft_token_ids)
1489
1490
1491
1492
1493
1494
1495
1496
                num_decode_draft_tokens[req_idx] = (
                    len(draft_token_ids)
                    if (
                        self.input_batch.num_computed_tokens_cpu[req_idx]
                        >= self.input_batch.num_prompt_tokens[req_idx]
                    )
                    else -1
                )
1497
            spec_decode_metadata = self._calc_spec_decode_metadata(
1498
1499
                num_draft_tokens, cu_num_tokens
            )
1500
            logits_indices = spec_decode_metadata.logits_indices
1501
            num_sampled_tokens = num_draft_tokens + 1
1502
            # For DECODE only cuda graph of some attention backends (e.g., GDN).
1503
            self.num_decode_draft_tokens.np[:num_reqs] = num_decode_draft_tokens
1504
1505
            self.num_decode_draft_tokens.np[num_reqs:].fill(-1)
            self.num_decode_draft_tokens.copy_to_gpu()
1506

1507
1508
1509
1510
1511
        # Hot-Swap lora model
        if self.lora_config:
            assert (
                np.sum(num_sampled_tokens)
                <= self.vllm_config.scheduler_config.max_num_batched_tokens
1512
            )
1513
1514
1515
1516
1517
1518
1519
1520
1521
1522
1523
            self.set_active_loras(
                self.input_batch, num_scheduled_tokens, num_sampled_tokens
            )

        return (
            logits_indices,
            spec_decode_metadata,
        )

    def _build_attention_metadata(
        self,
1524
        num_tokens: int,
1525
        num_reqs: int,
1526
1527
1528
        max_query_len: int,
        num_tokens_padded: int | None = None,
        num_reqs_padded: int | None = None,
1529
1530
1531
1532
        ubatch_slices: UBatchSlices | None = None,
        logits_indices: torch.Tensor | None = None,
        use_spec_decode: bool = False,
        for_cudagraph_capture: bool = False,
1533
        num_scheduled_tokens: dict[str, int] | None = None,
1534
1535
1536
1537
1538
        cascade_attn_prefix_lens: list[list[int]] | None = None,
    ) -> tuple[PerLayerAttnMetadata, CommonAttentionMetadata | None]:
        """
        :return: tuple[attn_metadata, spec_decode_common_attn_metadata]
        """
1539
1540
1541
1542
        # Attention metadata is not needed for attention free models
        if len(self.kv_cache_config.kv_cache_groups) == 0:
            return {}, None

1543
1544
        num_tokens_padded = num_tokens_padded or num_tokens
        num_reqs_padded = num_reqs_padded or num_reqs
1545
        assert num_reqs_padded is not None and num_tokens_padded is not None
1546

1547
1548
1549
        attn_metadata: PerLayerAttnMetadata = {}
        if ubatch_slices is not None:
            attn_metadata = [dict() for _ in range(len(ubatch_slices))]
1550

1551
1552
1553
1554
1555
1556
1557
1558
        if for_cudagraph_capture:
            # For some attention backends (e.g. FA) with sliding window models we need
            # to make sure the backend see a max_seq_len that is larger to the sliding
            # window size when capturing to make sure the correct kernel is selected.
            max_seq_len = self.max_model_len
        else:
            max_seq_len = self.seq_lens.np[:num_reqs].max().item()

1559
1560
        if use_spec_decode:
            self.num_accepted_tokens.np[:num_reqs] = (
1561
1562
                self.input_batch.num_accepted_tokens_cpu[:num_reqs]
            )
1563
1564
            self.num_accepted_tokens.np[num_reqs:].fill(1)
            self.num_accepted_tokens.copy_to_gpu()
1565

1566
        kv_cache_groups = self.kv_cache_config.kv_cache_groups
1567

1568
1569
1570
1571
        def _get_block_table_and_slot_mapping(kv_cache_gid: int):
            assert num_reqs_padded is not None and num_tokens_padded is not None
            kv_cache_spec = kv_cache_groups[kv_cache_gid].kv_cache_spec
            if isinstance(kv_cache_spec, EncoderOnlyAttentionSpec):
1572
                blk_table_tensor = torch.zeros(
1573
                    (num_reqs_padded, 1),
1574
                    dtype=torch.int32,
1575
1576
1577
                    device=self.device,
                )
                slot_mapping = torch.zeros(
1578
                    (num_tokens_padded,),
1579
1580
1581
                    dtype=torch.int64,
                    device=self.device,
                )
1582
            else:
1583
                blk_table = self.input_batch.block_table[kv_cache_gid]
1584
1585
                blk_table_tensor = blk_table.get_device_tensor(num_reqs_padded)
                slot_mapping = blk_table.slot_mapping.gpu[:num_tokens_padded]
1586

1587
1588
1589
1590
1591
1592
1593
1594
1595
1596
1597
1598
1599
1600
1601
1602
1603
1604
1605
1606
1607
1608
1609
1610
1611
1612
1613
1614
1615
1616
1617
1618
1619
1620
1621
1622
1623
1624
1625
1626
1627
1628
1629
1630
1631
1632
1633
1634
1635
1636
1637
1638
1639
1640
1641
1642
1643
1644
1645
1646
1647
1648
1649
1650
1651
1652
1653
1654
1655
1656
1657
1658
1659
1660
1661
1662
1663
1664
1665
1666
1667
1668
1669
1670
1671
1672
1673
1674
1675
1676
1677
1678
1679
1680
1681
1682
1683
1684
1685
1686
1687
1688
1689
            # Fill unused with -1. Needed for reshape_and_cache in full cuda
            # graph mode. `blk_table_tensor` -1 to match mamba PAD_SLOT_ID
            slot_mapping[num_tokens:num_tokens_padded].fill_(-1)
            blk_table_tensor[num_reqs:num_reqs_padded].fill_(-1)

            return blk_table_tensor, slot_mapping

        block_table_gid_0, slot_mapping_gid_0 = _get_block_table_and_slot_mapping(0)
        cm_base = CommonAttentionMetadata(
            query_start_loc=self.query_start_loc.gpu[: num_reqs_padded + 1],
            query_start_loc_cpu=self.query_start_loc.cpu[: num_reqs_padded + 1],
            seq_lens=self.seq_lens.gpu[:num_reqs_padded],
            _seq_lens_cpu=self.seq_lens.cpu[:num_reqs_padded],
            _num_computed_tokens_cpu=self.input_batch.num_computed_tokens_cpu_tensor[
                :num_reqs_padded
            ],
            num_reqs=num_reqs_padded,
            num_actual_tokens=num_tokens_padded,
            max_query_len=max_query_len,
            max_seq_len=max_seq_len,
            block_table_tensor=block_table_gid_0,
            slot_mapping=slot_mapping_gid_0,
            causal=True,
        )

        if self.dcp_world_size > 1:
            self.dcp_local_seq_lens.cpu[:num_reqs] = get_dcp_local_seq_lens(
                self.seq_lens.cpu[:num_reqs],
                self.dcp_world_size,
                self.dcp_rank,
                self.parallel_config.cp_kv_cache_interleave_size,
            )
            self.dcp_local_seq_lens.cpu[num_reqs:].fill_(0)
            self.dcp_local_seq_lens.copy_to_gpu(num_reqs_padded)

            cm_base.dcp_local_seq_lens = self.dcp_local_seq_lens.gpu[:num_reqs_padded]
            cm_base.dcp_local_seq_lens_cpu = self.dcp_local_seq_lens.cpu[
                :num_reqs_padded
            ]

        if logits_indices is not None and self.cache_config.kv_sharing_fast_prefill:
            cm_base.num_logits_indices = logits_indices.size(0)
            cm_base.logits_indices_padded = self._prepare_kv_sharing_fast_prefill(
                logits_indices
            )

        def _build_attn_group_metadata(
            kv_cache_gid: int,
            attn_gid: int,
            common_attn_metadata: CommonAttentionMetadata,
            ubid: int | None = None,
        ) -> None:
            attn_group = self.attn_groups[kv_cache_gid][attn_gid]
            cascade_attn_prefix_len = (
                cascade_attn_prefix_lens[kv_cache_gid][attn_gid]
                if cascade_attn_prefix_lens
                else 0
            )

            builder = attn_group.get_metadata_builder(ubid or 0)
            extra_attn_metadata_args = {}
            if use_spec_decode and isinstance(builder, GDNAttentionMetadataBuilder):
                assert ubid is None, "UBatching not supported with GDN yet"
                extra_attn_metadata_args = dict(
                    num_accepted_tokens=self.num_accepted_tokens.gpu[:num_reqs_padded],
                    num_decode_draft_tokens_cpu=self.num_decode_draft_tokens.cpu[
                        :num_reqs_padded
                    ],
                )

            if for_cudagraph_capture:
                attn_metadata_i = builder.build_for_cudagraph_capture(
                    common_attn_metadata
                )
            else:
                attn_metadata_i = builder.build(
                    common_prefix_len=cascade_attn_prefix_len,
                    common_attn_metadata=common_attn_metadata,
                    **extra_attn_metadata_args,
                )

            if ubid is None:
                assert isinstance(attn_metadata, dict)
                attn_metadata_dict = attn_metadata
            else:
                assert isinstance(attn_metadata, list)
                attn_metadata_dict = attn_metadata[ubid]

            for layer_name in attn_group.layer_names:
                attn_metadata_dict[layer_name] = attn_metadata_i

        # Prepare the attention metadata for each KV cache group and make layers
        # in the same group share the same metadata.
        spec_decode_common_attn_metadata = None
        for kv_cache_gid, kv_cache_group in enumerate(kv_cache_groups):
            cm = copy(cm_base)  # shallow copy

            # Basically only the encoder seq_lens, block_table and slot_mapping change
            # for each kv_cache_group.
            cm.encoder_seq_lens, cm.encoder_seq_lens_cpu = self._get_encoder_seq_lens(
                num_scheduled_tokens or {},
                kv_cache_group.kv_cache_spec,
                num_reqs_padded,
1690
            )
1691
1692
1693
1694
            if kv_cache_gid > 0:
                cm.block_table_tensor, cm.slot_mapping = (
                    _get_block_table_and_slot_mapping(kv_cache_gid)
                )
1695

1696
            if self.speculative_config and spec_decode_common_attn_metadata is None:
1697
                if isinstance(self.drafter, EagleProposer):
1698
                    if self.drafter.attn_layer_names[0] in kv_cache_group.layer_names:
1699
                        spec_decode_common_attn_metadata = cm
1700
                else:
1701
                    spec_decode_common_attn_metadata = cm
1702

1703
            for attn_gid in range(len(self.attn_groups[kv_cache_gid])):
1704
                if ubatch_slices is not None:
1705
1706
1707
                    for ubid, _cm in enumerate(split_attn_metadata(ubatch_slices, cm)):
                        _build_attn_group_metadata(kv_cache_gid, attn_gid, _cm, ubid)

1708
                else:
1709
                    _build_attn_group_metadata(kv_cache_gid, attn_gid, cm)
1710

1711
1712
1713
1714
1715
1716
1717
1718
1719
1720
1721
1722
1723
1724
1725
1726
1727
1728
1729
1730
        if self.is_mm_prefix_lm:
            req_doc_ranges = {}
            for req_id in self.input_batch.req_ids:
                image_doc_ranges = []
                req_state = self.requests[req_id]
                for mm_feature in req_state.mm_features:
                    pos_info = mm_feature.mm_position
                    img_doc_range = pos_info.extract_embeds_range()
                    image_doc_ranges.extend(img_doc_range)
                req_idx = self.input_batch.req_id_to_index[req_id]
                req_doc_ranges[req_idx] = image_doc_ranges

            if isinstance(attn_metadata, list):
                for ub_metadata in attn_metadata:
                    for _metadata in ub_metadata.values():
                        _metadata.mm_prefix_range = req_doc_ranges  # type: ignore[attr-defined]
            else:
                for _metadata in attn_metadata.values():
                    _metadata.mm_prefix_range = req_doc_ranges  # type: ignore[attr-defined]

1731
1732
1733
1734
1735
1736
1737
1738
1739
1740
        if spec_decode_common_attn_metadata is not None and (
            num_reqs != num_reqs_padded or num_tokens != num_tokens_padded
        ):
            # Currently the drafter still only uses piecewise cudagraphs (and modifies
            # the attention metadata in directly), and therefore does not want to use
            # padded attention metadata.
            spec_decode_common_attn_metadata = (
                spec_decode_common_attn_metadata.unpadded(num_tokens, num_reqs)
            )

1741
        return attn_metadata, spec_decode_common_attn_metadata
1742

1743
1744
1745
    def _compute_cascade_attn_prefix_lens(
        self,
        num_scheduled_tokens: np.ndarray,
1746
        num_computed_tokens: np.ndarray,
1747
1748
1749
1750
1751
1752
1753
        num_common_prefix_blocks: list[int],
    ) -> list[list[int]] | None:
        """
        :return: Optional[cascade_attn_prefix_lens]
            cascade_attn_prefix_lens is 2D: ``[kv_cache_group_id][attn_group_idx]``,
            None if we should not use cascade attention
        """
1754

1755
1756
1757
1758
1759
1760
1761
1762
1763
1764
1765
1766
1767
1768
        use_cascade_attn = False
        num_kv_cache_groups = len(self.kv_cache_config.kv_cache_groups)
        cascade_attn_prefix_lens: list[list[int]] = [
            [] for _ in range(num_kv_cache_groups)
        ]

        for kv_cache_gid in range(num_kv_cache_groups):
            for attn_group in self.attn_groups[kv_cache_gid]:
                if isinstance(attn_group.kv_cache_spec, EncoderOnlyAttentionSpec):
                    cascade_attn_prefix_len = 0
                else:
                    # 0 if cascade attention should not be used
                    cascade_attn_prefix_len = self._compute_cascade_attn_prefix_len(
                        num_scheduled_tokens,
1769
                        num_computed_tokens,
1770
1771
1772
1773
1774
1775
1776
1777
                        num_common_prefix_blocks[kv_cache_gid],
                        attn_group.kv_cache_spec,
                        attn_group.get_metadata_builder(),
                    )
                cascade_attn_prefix_lens[kv_cache_gid].append(cascade_attn_prefix_len)
                use_cascade_attn |= cascade_attn_prefix_len > 0

        return cascade_attn_prefix_lens if use_cascade_attn else None
1778

1779
1780
1781
    def _compute_cascade_attn_prefix_len(
        self,
        num_scheduled_tokens: np.ndarray,
1782
        num_computed_tokens: np.ndarray,
1783
        num_common_prefix_blocks: int,
1784
1785
        kv_cache_spec: KVCacheSpec,
        attn_metadata_builder: AttentionMetadataBuilder,
1786
1787
1788
1789
1790
1791
1792
1793
1794
1795
1796
1797
1798
1799
1800
1801
1802
1803
    ) -> 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.
        """
1804

1805
        common_prefix_len = num_common_prefix_blocks * kv_cache_spec.block_size
1806
1807
1808
1809
1810
1811
1812
1813
1814
1815
1816
1817
1818
1819
1820
1821
1822
1823
1824
1825
1826
1827
1828
1829
1830
1831
1832
1833
1834
1835
1836
1837
1838
1839
1840
1841
1842
        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]
1843
        # Request 3's num_computed_tokens: 3 (i.e., [A, B, C])
1844
1845
1846
1847
1848
        # 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.
1849
        common_prefix_len = min(common_prefix_len, num_computed_tokens.min())
1850
        # common_prefix_len should be a multiple of the block size.
1851
1852
1853
1854
1855
1856
1857
1858
1859
1860
1861
        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
        )
        use_local_attention = isinstance(kv_cache_spec, ChunkedLocalAttentionSpec) or (
            isinstance(kv_cache_spec, FullAttentionSpec)
            and kv_cache_spec.attention_chunk_size is not None
        )
1862
1863
        assert isinstance(kv_cache_spec, AttentionSpec)
        use_cascade = attn_metadata_builder.use_cascade_attention(
1864
1865
1866
            common_prefix_len=common_prefix_len,
            query_lens=num_scheduled_tokens,
            num_query_heads=self.num_query_heads,
1867
            num_kv_heads=kv_cache_spec.num_kv_heads,
1868
            use_alibi=self.use_alibi,
1869
            use_sliding_window=use_sliding_window,
1870
            use_local_attention=use_local_attention,
1871
            num_sms=self.num_sms,
1872
            dcp_world_size=self.dcp_world_size,
1873
1874
1875
        )
        return common_prefix_len if use_cascade else 0

1876
1877
    def _calc_mrope_positions(self, scheduler_output: "SchedulerOutput"):
        mrope_pos_ptr = 0
1878
        for index, req_id in enumerate(self.input_batch.req_ids):
1879
1880
1881
            req = self.requests[req_id]
            assert req.mrope_positions is not None

1882
1883
            num_computed_tokens = self.input_batch.num_computed_tokens_cpu[index]
            num_scheduled_tokens = scheduler_output.num_scheduled_tokens[req_id]
1884
            num_prompt_tokens = length_from_prompt_token_ids_or_embeds(
1885
1886
                req.prompt_token_ids, req.prompt_embeds
            )
1887
1888

            if num_computed_tokens + num_scheduled_tokens > num_prompt_tokens:
1889
1890
                prompt_part_len = max(0, num_prompt_tokens - num_computed_tokens)
                completion_part_len = max(0, num_scheduled_tokens - prompt_part_len)
1891
1892
1893
1894
1895
1896
1897
1898
1899
1900
1901
1902
1903
            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

1904
1905
1906
                self.mrope_positions.cpu[:, dst_start:dst_end] = req.mrope_positions[
                    :, src_start:src_end
                ]
1907
1908
1909
1910
1911
1912
1913
                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

1914
                assert req.mrope_position_delta is not None
1915
                MRotaryEmbedding.get_next_input_positions_tensor(
1916
                    out=self.mrope_positions.np,
1917
1918
1919
1920
1921
                    out_offset=dst_start,
                    mrope_position_delta=req.mrope_position_delta,
                    context_len=num_computed_tokens + prompt_part_len,
                    num_new_tokens=completion_part_len,
                )
1922
1923
1924

                mrope_pos_ptr += completion_part_len

1925
1926
1927
1928
1929
1930
1931
1932
1933
1934
1935
1936
1937
1938
1939
1940
1941
1942
1943
1944
1945
1946
1947
1948
1949
1950
1951
1952
1953
1954
1955
1956
1957
1958
1959
1960
1961
1962
1963
1964
1965
1966
1967
1968
1969
1970
1971
    def _calc_xdrope_positions(self, scheduler_output: "SchedulerOutput"):
        xdrope_pos_ptr = 0
        for index, req_id in enumerate(self.input_batch.req_ids):
            req = self.requests[req_id]
            assert req.xdrope_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 = length_from_prompt_token_ids_or_embeds(
                req.prompt_token_ids, req.prompt_embeds
            )

            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 xdrope_positions are pre-computed
                dst_start = xdrope_pos_ptr
                dst_end = xdrope_pos_ptr + prompt_part_len
                src_start = num_computed_tokens
                src_end = num_computed_tokens + prompt_part_len

                self.xdrope_positions.cpu[:, dst_start:dst_end] = req.xdrope_positions[
                    :, src_start:src_end
                ]
                xdrope_pos_ptr += prompt_part_len

            if completion_part_len > 0:
                # compute completion's xdrope_positions on-the-fly
                dst_start = xdrope_pos_ptr
                dst_end = xdrope_pos_ptr + completion_part_len

                XDRotaryEmbedding.get_next_input_positions_tensor(
                    out=self.xdrope_positions.np,
                    out_offset=dst_start,
                    context_len=num_computed_tokens + prompt_part_len,
                    num_new_tokens=completion_part_len,
                )

                xdrope_pos_ptr += completion_part_len

1972
1973
    def _calc_spec_decode_metadata(
        self,
1974
1975
1976
1977
1978
1979
1980
1981
1982
1983
1984
1985
1986
1987
1988
1989
        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
1990
1991
1992
1993

        # 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(
1994
1995
            num_sampled_tokens, cumsum_dtype=np.int32
        )
1996
        # Step 2. [0, 0, 0, 0, 103, 104, 104, 104, 206, 207, 207]
1997
        logits_indices = np.repeat(
1998
1999
            cu_num_scheduled_tokens - num_sampled_tokens, num_sampled_tokens
        )
2000
        # Step 3. [0, 1, 2, 3, 103, 104, 105, 106, 206, 207, 208]
2001
2002
2003
2004
2005
2006
        logits_indices += arange

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

        # Compute the draft logits indices.
2007
2008
2009
        # 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(
2010
2011
            num_draft_tokens, cumsum_dtype=np.int32
        )
2012
2013
        # [0, 0, 0, 5, 5, 9]
        target_logits_indices = np.repeat(
2014
2015
            cu_num_sampled_tokens - num_sampled_tokens, num_draft_tokens
        )
2016
2017
2018
2019
2020
        # [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(
2021
2022
            self.device, non_blocking=True
        )
2023
2024
2025
        cu_num_sampled_tokens = torch.from_numpy(cu_num_sampled_tokens).to(
            self.device, non_blocking=True
        )
2026
2027
2028
        logits_indices = torch.from_numpy(logits_indices).to(
            self.device, non_blocking=True
        )
2029
        target_logits_indices = torch.from_numpy(target_logits_indices).to(
2030
2031
            self.device, non_blocking=True
        )
2032
        bonus_logits_indices = torch.from_numpy(bonus_logits_indices).to(
2033
2034
            self.device, non_blocking=True
        )
2035

2036
2037
        # Compute the draft token ids.
        # draft_token_indices:      [  1,   2,   3, 105, 106, 208]
2038
        draft_token_ids = self.input_ids.gpu[logits_indices]
2039
2040
        draft_token_ids = draft_token_ids[target_logits_indices + 1]

2041
        return SpecDecodeMetadata(
2042
2043
2044
            draft_token_ids=draft_token_ids,
            num_draft_tokens=num_draft_tokens.tolist(),
            cu_num_draft_tokens=cu_num_draft_tokens,
2045
            cu_num_sampled_tokens=cu_num_sampled_tokens,
2046
2047
2048
2049
2050
            target_logits_indices=target_logits_indices,
            bonus_logits_indices=bonus_logits_indices,
            logits_indices=logits_indices,
        )

2051
2052
2053
2054
2055
2056
2057
    def _prepare_kv_sharing_fast_prefill(
        self,
        logits_indices: torch.Tensor,
    ) -> torch.Tensor:
        assert self.kv_sharing_fast_prefill_logits_indices is not None
        num_logits = logits_indices.shape[0]
        assert num_logits > 0
2058
        self.kv_sharing_fast_prefill_logits_indices[:num_logits].copy_(logits_indices)
2059
2060
2061
2062
2063
        # There might have leftover indices in logits_indices[num_logits:]
        # from previous iterations, whose values may be greater than the
        # batch size in the current iteration. To ensure indices are always
        # valid, we fill the padded indices with the last index.
        self.kv_sharing_fast_prefill_logits_indices[num_logits:].fill_(
2064
2065
2066
2067
2068
2069
            logits_indices[-1].item()
        )
        if (
            self.compilation_config.cudagraph_mode != CUDAGraphMode.NONE
            and num_logits <= self.cudagraph_batch_sizes[-1]
        ):
2070
2071
2072
2073
2074
            # Use piecewise CUDA graphs.
            # Add padding to the batch size.
            num_logits_padded = self.vllm_config.pad_for_cudagraph(num_logits)
        else:
            num_logits_padded = num_logits
2075
2076
2077
        logits_indices_padded = self.kv_sharing_fast_prefill_logits_indices[
            :num_logits_padded
        ]
2078
2079
        return logits_indices_padded

2080
2081
2082
2083
2084
2085
2086
2087
    def _batch_mm_kwargs_from_scheduler(
        self,
        scheduler_output: "SchedulerOutput",
    ) -> tuple[list[MultiModalKwargsItem], list[tuple[str, PlaceholderRange]]]:
        """Batch multimodal kwargs from scheduled encoder inputs.

        Args:
            scheduler_output: The scheduler output containing scheduled encoder
2088
                inputs.
2089
2090
2091
2092
2093
2094

        Returns:
            A tuple of (mm_kwargs, req_ids_pos) where:
            - mm_kwargs: List of multimodal kwargs items to be batched
            - mm_hashes_pos: List of (mm_hash, position_info) tuples
        """
2095
2096
        scheduled_encoder_inputs = scheduler_output.scheduled_encoder_inputs
        if not scheduled_encoder_inputs:
2097
            return [], []
2098
        # Batch the multi-modal inputs.
2099
        mm_kwargs = list[MultiModalKwargsItem]()
2100
2101
        # list of tuple (mm_hash, position_info)
        mm_hashes_pos = list[tuple[str, PlaceholderRange]]()
2102
2103
        for req_id, encoder_input_ids in scheduled_encoder_inputs.items():
            req_state = self.requests[req_id]
2104
2105

            for mm_input_id in encoder_input_ids:
2106
                mm_feature = req_state.mm_features[mm_input_id]
2107
2108
                if mm_feature.data is None:
                    continue
2109
2110
2111
                mm_hash = mm_feature.identifier
                mm_kwargs.append(mm_feature.data)
                mm_hashes_pos.append((mm_hash, mm_feature.mm_position))
2112

2113
2114
        return mm_kwargs, mm_hashes_pos

2115
2116
2117
    def _execute_mm_encoder(
        self, scheduler_output: "SchedulerOutput"
    ) -> list[torch.Tensor]:
2118
2119
        # Batch the multi-modal inputs using the helper method.
        mm_kwargs, mm_hashes_pos = self._batch_mm_kwargs_from_scheduler(
2120
2121
            scheduler_output
        )
2122
2123

        if not mm_kwargs:
2124
            return []
2125

2126
2127
2128
2129
2130
2131
2132
        # 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.
2133
        model = cast(SupportsMultiModal, self.model)
2134
        encoder_outputs: list[torch.Tensor] = []
2135
        for modality, num_items, mm_kwargs_group in group_mm_kwargs_by_modality(
2136
2137
2138
            mm_kwargs,
            device=self.device,
            pin_memory=self.pin_memory,
2139
        ):
2140
            curr_group_outputs: list[torch.Tensor] = []
2141
2142

            # EVS-related change.
2143
            # (ekhvedchenia): Temporary hack to limit peak memory usage when
2144
            # processing multimodal data. This solves the issue with scheduler
2145
2146
2147
2148
            # putting too many video samples into a single batch. Scheduler
            # uses pruned vision tokens count to compare it versus compute
            # budget which is incorrect (Either input media size or non-pruned
            # output vision tokens count should be considered)
2149
2150
2151
2152
2153
2154
2155
2156
2157
2158
2159
2160
2161
2162
2163
2164
            # TODO(ywang96): Fix memory profiling to take EVS into account and
            # remove this hack.
            if (
                self.is_multimodal_pruning_enabled
                and modality == "video"
                and num_items > 1
            ):
                for video_mm_kwargs_item in filter(
                    lambda item: item.modality == "video", mm_kwargs
                ):
                    _, _, micro_batch_mm_inputs = next(
                        group_mm_kwargs_by_modality(
                            [video_mm_kwargs_item],
                            device=self.device,
                            pin_memory=self.pin_memory,
                        )
2165
                    )
2166

2167
                    micro_batch_outputs = model.embed_multimodal(
2168
2169
                        **micro_batch_mm_inputs
                    )
2170
2171
2172
2173
2174
2175
2176
2177
2178
2179

                    curr_group_outputs.extend(micro_batch_outputs)
            else:
                # Run the encoder.
                # `curr_group_outputs` is either of the following:
                # 1. A tensor of shape (num_items, feature_size, hidden_size)
                # in case feature_size is fixed across all multimodal items.
                # 2. A list or tuple (length: num_items) of tensors,
                # each of shape (feature_size, hidden_size) in case the feature
                # size is dynamic depending on the input multimodal items.
2180
                curr_group_outputs = model.embed_multimodal(**mm_kwargs_group)  # type: ignore[assignment]
2181

2182
2183
            sanity_check_mm_encoder_outputs(
                curr_group_outputs,
2184
                expected_num_items=num_items,
2185
            )
2186
            encoder_outputs.extend(curr_group_outputs)
2187

2188
2189
2190
        # Cache the encoder outputs by mm_hash
        for (mm_hash, pos_info), output in zip(mm_hashes_pos, encoder_outputs):
            self.encoder_cache[mm_hash] = scatter_mm_placeholders(
2191
2192
2193
                output,
                is_embed=pos_info.is_embed,
            )
2194
2195
            logger.debug("Finish execute for mm hash %s", mm_hash)
            self.maybe_save_ec_to_connector(self.encoder_cache, mm_hash)
2196

2197
2198
        return encoder_outputs

2199
    def _gather_mm_embeddings(
2200
2201
        self,
        scheduler_output: "SchedulerOutput",
2202
        shift_computed_tokens: int = 0,
2203
2204
2205
2206
2207
2208
2209
2210
    ) -> tuple[list[torch.Tensor], torch.Tensor]:
        total_num_scheduled_tokens = scheduler_output.total_num_scheduled_tokens

        mm_embeds = list[torch.Tensor]()
        is_mm_embed = self.is_mm_embed.cpu
        is_mm_embed[:total_num_scheduled_tokens] = False

        req_start_idx = 0
2211
        should_sync_mrope_positions = False
2212
        should_sync_xdrope_positions = False
2213

2214
        for req_id in self.input_batch.req_ids:
2215
2216
            mm_embeds_req: list[torch.Tensor] = []

2217
            num_scheduled_tokens = scheduler_output.num_scheduled_tokens[req_id]
2218
            req_state = self.requests[req_id]
2219
            num_computed_tokens = req_state.num_computed_tokens + shift_computed_tokens
2220

2221
2222
            for mm_feature in req_state.mm_features:
                pos_info = mm_feature.mm_position
2223
2224
                start_pos = pos_info.offset
                num_encoder_tokens = pos_info.length
2225
2226
2227
2228
2229
2230
2231
2232
2233
2234
2235
2236
2237
2238
2239
2240

                # 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,
2241
2242
                    num_encoder_tokens,
                )
2243
                assert start_idx < end_idx
2244

2245
                mm_hash = mm_feature.identifier
2246
                encoder_output = self.encoder_cache.get(mm_hash, None)
2247
                assert encoder_output is not None, f"Encoder cache miss for {mm_hash}."
2248
2249
2250
2251

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

2252
                req_start_pos = req_start_idx + start_pos - num_computed_tokens
2253
2254
2255
                is_mm_embed[req_start_pos + start_idx : req_start_pos + end_idx] = (
                    True if is_embed is None else is_embed
                )
2256

2257
2258
2259
2260
                mm_embeds_item = gather_mm_placeholders(
                    encoder_output[start_idx:end_idx],
                    is_embed=is_embed,
                )
2261
2262
2263
                mm_embeds_req.append(mm_embeds_item)

            if self.is_multimodal_pruning_enabled and self.uses_mrope:
2264
                assert req_state.mrope_positions is not None
2265
2266
2267
2268
2269
2270
2271
                should_sync_mrope_positions = True
                mm_embeds_req, new_mrope_positions, new_delta = (
                    self.model.recompute_mrope_positions(
                        input_ids=req_state.prompt_token_ids,
                        multimodal_embeddings=mm_embeds_req,
                        mrope_positions=req_state.mrope_positions,
                        num_computed_tokens=req_state.num_computed_tokens,
2272
2273
                    )
                )
2274
2275
2276
2277
                req_state.mrope_positions.copy_(new_mrope_positions)
                req_state.mrope_position_delta = new_delta

            mm_embeds.extend(mm_embeds_req)
2278
2279
2280
            req_start_idx += num_scheduled_tokens

        is_mm_embed = self.is_mm_embed.copy_to_gpu(total_num_scheduled_tokens)
2281
2282
2283

        if should_sync_mrope_positions:
            self._calc_mrope_positions(scheduler_output)
2284
            self.mrope_positions.copy_to_gpu(total_num_scheduled_tokens)
2285

2286
2287
2288
2289
        if should_sync_xdrope_positions:
            self._calc_xdrope_positions(scheduler_output)
            self.xdrope_positions.copy_to_gpu(total_num_scheduled_tokens)

2290
        return mm_embeds, is_mm_embed
2291

2292
    def get_model(self) -> nn.Module:
2293
        # get raw model out of the cudagraph wrapper.
2294
        if isinstance(self.model, (CUDAGraphWrapper, UBatchWrapper)):
2295
            return self.model.unwrap()
2296
2297
        return self.model

2298
2299
2300
2301
2302
2303
2304
2305
2306
2307
2308
2309
2310
2311
2312
    def get_supported_generation_tasks(self) -> list[GenerationTask]:
        model = self.get_model()
        supported_tasks = list[GenerationTask]()

        if is_text_generation_model(model):
            supported_tasks.append("generate")

        if supports_transcription(model):
            if model.supports_transcription_only:
                return ["transcription"]

            supported_tasks.append("transcription")

        return supported_tasks

2313
2314
2315
2316
2317
    def get_supported_pooling_tasks(self) -> list[PoolingTask]:
        model = self.get_model()
        if not is_pooling_model(model):
            return []

2318
2319
        supported_tasks = list(model.pooler.get_supported_tasks())

2320
2321
2322
2323
        if "score" in supported_tasks:
            num_labels = getattr(self.model_config.hf_config, "num_labels", 0)
            if num_labels != 1:
                supported_tasks.remove("score")
2324
                logger.debug_once("Score API is only enabled for num_labels == 1.")
2325
2326

        return supported_tasks
2327

2328
2329
2330
2331
2332
2333
2334
2335
2336
2337
    def get_supported_tasks(self) -> tuple[SupportedTask, ...]:
        tasks = list[SupportedTask]()

        if self.model_config.runner_type == "generate":
            tasks.extend(self.get_supported_generation_tasks())
        if self.model_config.runner_type == "pooling":
            tasks.extend(self.get_supported_pooling_tasks())

        return tuple(tasks)

2338
    def sync_and_slice_intermediate_tensors(
2339
2340
        self,
        num_tokens: int,
2341
        intermediate_tensors: IntermediateTensors | None,
2342
2343
        sync_self: bool,
    ) -> IntermediateTensors:
2344
2345
2346
        assert self.intermediate_tensors is not None

        tp = self.vllm_config.parallel_config.tensor_parallel_size
2347
        is_rs = is_residual_scattered_for_sp(self.vllm_config, num_tokens)
2348
2349
2350
2351
2352
2353

        # 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():
2354
                is_scattered = k == "residual" and is_rs
2355
                copy_len = num_tokens // tp if is_scattered else num_tokens
2356
                self.intermediate_tensors[k][:copy_len].copy_(
2357
2358
2359
2360
2361
2362
2363
2364
2365
2366
2367
2368
2369
                    v[:copy_len], non_blocking=True
                )

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

    def eplb_step(self, is_dummy: bool = False, is_profile: bool = False) -> None:
2370
2371
2372
2373
2374
2375
2376
        """
        Step for the EPLB (Expert Parallelism Load Balancing) state.
        """
        if not self.parallel_config.enable_eplb:
            return

        assert self.eplb_state is not None
2377
2378
        model = self.get_model()
        assert is_mixture_of_experts(model)
2379
2380
2381
        self.eplb_state.step(
            is_dummy,
            is_profile,
2382
            log_stats=self.parallel_config.eplb_config.log_balancedness,
2383
2384
        )

2385
2386
2387
2388
2389
2390
    def _pool(
        self,
        hidden_states: torch.Tensor,
        num_scheduled_tokens: int,
        num_scheduled_tokens_np: np.ndarray,
    ) -> ModelRunnerOutput:
2391
2392
2393
        assert self.input_batch.num_reqs == len(self.input_batch.pooling_params), (
            "Either all or none of the requests in a batch must be pooling request"
        )
2394

2395
        hidden_states = hidden_states[:num_scheduled_tokens]
2396
2397
        seq_lens_cpu = self.seq_lens.cpu[: self.input_batch.num_reqs]

2398
        pooling_metadata = self.input_batch.get_pooling_metadata()
2399
        pooling_metadata.build_pooling_cursor(
2400
            num_scheduled_tokens_np.tolist(), seq_lens_cpu, device=hidden_states.device
2401
        )
2402

2403
2404
2405
2406
2407
2408
        model = cast(VllmModelForPooling, self.model)
        raw_pooler_output: PoolerOutput = model.pooler(
            hidden_states=hidden_states,
            pooling_metadata=pooling_metadata,
        )
        raw_pooler_output = json_map_leaves(
2409
            lambda x: x.to("cpu", non_blocking=True) if x is not None else x,
2410
2411
2412
            raw_pooler_output,
        )
        self._sync_device()
2413

2414
        pooler_output: list[torch.Tensor | None] = []
2415
        for raw_output, seq_len, prompt_len in zip(
2416
2417
            raw_pooler_output, seq_lens_cpu, pooling_metadata.prompt_lens
        ):
2418
            output = raw_output if seq_len == prompt_len else None
2419
            pooler_output.append(output)
2420
2421
2422
2423
2424
2425
2426
2427
2428
2429

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

2430
    def _pad_for_sequence_parallelism(self, num_scheduled_tokens: int) -> int:
2431
2432
2433
        # Pad tokens to multiple of tensor_parallel_size when
        # enabled collective fusion for SP
        tp_size = self.vllm_config.parallel_config.tensor_parallel_size
2434
        if self.compilation_config.pass_config.enable_sp and tp_size > 1:
2435
2436
2437
            return round_up(num_scheduled_tokens, tp_size)
        return num_scheduled_tokens

2438
    def _preprocess(
2439
2440
        self,
        scheduler_output: "SchedulerOutput",
2441
        num_input_tokens: int,  # Padded
2442
        intermediate_tensors: IntermediateTensors | None = None,
2443
    ) -> tuple[
2444
2445
        torch.Tensor | None,
        torch.Tensor | None,
2446
        torch.Tensor,
2447
        IntermediateTensors | None,
2448
        dict[str, Any],
2449
        ECConnectorOutput | None,
2450
    ]:
2451
        num_scheduled_tokens = scheduler_output.total_num_scheduled_tokens
2452
        is_first_rank = get_pp_group().is_first_rank
2453
        is_encoder_decoder = self.model_config.is_encoder_decoder
2454

2455
2456
        # _prepare_inputs may reorder the batch, so we must gather multi
        # modal outputs after that to ensure the correct order
2457
2458
        ec_connector_output = None

2459
        if self.supports_mm_inputs and is_first_rank and not is_encoder_decoder:
2460
            # Run the multimodal encoder if any.
2461
2462
2463
2464
2465
2466
            with self.maybe_get_ec_connector_output(
                scheduler_output,
                encoder_cache=self.encoder_cache,
            ) as ec_connector_output:
                self._execute_mm_encoder(scheduler_output)
                mm_embeds, is_mm_embed = self._gather_mm_embeddings(scheduler_output)
2467

2468
2469
2470
            # 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.
2471
            inputs_embeds_scheduled = self.model.embed_input_ids(
2472
2473
2474
                self.input_ids.gpu[:num_scheduled_tokens],
                multimodal_embeddings=mm_embeds,
                is_multimodal=is_mm_embed,
2475
            )
2476

2477
            # TODO(woosuk): Avoid the copy. Optimize.
2478
            self.inputs_embeds.gpu[:num_scheduled_tokens].copy_(inputs_embeds_scheduled)
2479

2480
            input_ids = None
2481
            inputs_embeds = self.inputs_embeds.gpu[:num_input_tokens]
2482
2483
2484
2485
            model_kwargs = {
                **self._init_model_kwargs(num_scheduled_tokens),
                **self._extract_mm_kwargs(scheduler_output),
            }
2486
        elif self.enable_prompt_embeds and is_first_rank:
2487
2488
2489
2490
2491
2492
2493
2494
2495
2496
2497
2498
            # Get the input embeddings for the tokens that are not input embeds,
            # then put them into the appropriate positions.
            # TODO(qthequartermasterman): Since even when prompt embeds are
            # enabled, (a) not all requests will use prompt embeds, and (b)
            # after the initial prompt is processed, the rest of the generated
            # tokens will be token ids, it is not desirable to have the
            # embedding layer outside of the CUDA graph all the time. The v0
            # engine avoids this by "double compiling" the CUDA graph, once
            # with input_ids and again with inputs_embeds, for all num_tokens.
            # If a batch only has token ids, then including the embedding layer
            # in the CUDA graph will be more performant (like in the else case
            # below).
2499
2500
2501
            token_ids_idx = (
                self.is_token_ids.gpu[:num_scheduled_tokens]
                .nonzero(as_tuple=False)
2502
                .squeeze(1)
2503
            )
2504
2505
2506
            # Some tokens ids may need to become embeds
            if token_ids_idx.numel() > 0:
                token_ids = self.input_ids.gpu[token_ids_idx]
2507
                tokens_to_embeds = self.model.embed_input_ids(input_ids=token_ids)
2508
2509
2510
2511
2512
                self.inputs_embeds.gpu[token_ids_idx] = tokens_to_embeds

            inputs_embeds = self.inputs_embeds.gpu[:num_input_tokens]
            model_kwargs = self._init_model_kwargs(num_input_tokens)
            input_ids = None
2513
        else:
2514
2515
2516
2517
            # 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.
2518
            input_ids = self.input_ids.gpu[:num_input_tokens]
2519
            inputs_embeds = None
2520
            model_kwargs = self._init_model_kwargs(num_input_tokens)
2521

2522
        if self.uses_mrope:
2523
            positions = self.mrope_positions.gpu[:, :num_input_tokens]
2524
2525
        elif self.uses_xdrope_dim > 0:
            positions = self.xdrope_positions.gpu[:, :num_input_tokens]
2526
        else:
2527
            positions = self.positions.gpu[:num_input_tokens]
2528

2529
        if is_first_rank:
2530
2531
            intermediate_tensors = None
        else:
2532
            assert intermediate_tensors is not None
2533
            intermediate_tensors = self.sync_and_slice_intermediate_tensors(
2534
2535
                num_input_tokens, intermediate_tensors, True
            )
2536

2537
        if is_encoder_decoder and scheduler_output.scheduled_encoder_inputs:
2538
2539
2540
2541
2542
2543
2544
            # Run the encoder, just like we do with other multimodal inputs.
            # For an encoder-decoder model, our processing here is a bit
            # simpler, because the outputs are just passed to the decoder.
            # We are not doing any prompt replacement. We also will only
            # ever have a single encoder input.
            encoder_outputs = self._execute_mm_encoder(scheduler_output)
            model_kwargs.update({"encoder_outputs": encoder_outputs})
2545

2546
2547
2548
2549
2550
2551
        return (
            input_ids,
            inputs_embeds,
            positions,
            intermediate_tensors,
            model_kwargs,
2552
            ec_connector_output,
2553
        )
2554

2555
    def _sample(
2556
        self,
2557
2558
        logits: torch.Tensor | None,
        spec_decode_metadata: SpecDecodeMetadata | None,
2559
    ) -> SamplerOutput:
2560
        # Sample the next token and get logprobs if needed.
2561
        sampling_metadata = self.input_batch.sampling_metadata
2562
        if spec_decode_metadata is None:
2563
2564
2565
            # Update output token ids with tokens sampled in last step
            # if async scheduling and required by current sampling params.
            self.input_batch.update_async_output_token_ids()
2566
            return self.sampler(
2567
2568
2569
                logits=logits,
                sampling_metadata=sampling_metadata,
            )
2570

2571
        sampler_output = self.rejection_sampler(
2572
2573
            spec_decode_metadata,
            None,  # draft_probs
2574
            logits,
2575
2576
            sampling_metadata,
        )
2577
        self._update_states_after_model_execute(sampler_output.sampled_token_ids)
2578
2579
2580
        return sampler_output

    def _bookkeeping_sync(
2581
2582
2583
        self,
        scheduler_output: "SchedulerOutput",
        sampler_output: SamplerOutput,
2584
        logits: torch.Tensor | None,
2585
2586
        hidden_states: torch.Tensor,
        num_scheduled_tokens: int,
2587
        spec_decode_metadata: SpecDecodeMetadata | None,
2588
    ) -> tuple[
2589
        dict[str, int],
2590
        LogprobsLists | None,
2591
        list[list[int]],
2592
        dict[str, LogprobsTensors | None],
2593
2594
2595
        list[str],
        dict[str, int],
        list[int],
2596
    ]:
2597
2598
2599
2600
        num_nans_in_logits = {}
        if envs.VLLM_COMPUTE_NANS_IN_LOGITS:
            num_nans_in_logits = self._get_nans_in_logits(logits)

2601
2602
2603
2604
        num_reqs = self.input_batch.num_reqs
        discard_sampled_tokens_req_indices = np.nonzero(
            self.discard_request_mask.np[:num_reqs]
        )[0]
2605
2606
2607
2608
        for i in discard_sampled_tokens_req_indices:
            gen = self.input_batch.generators.get(int(i))
            if gen is not None:
                gen.set_offset(gen.get_offset() - 4)
2609

2610
2611
2612
        # Copy some objects so they don't get modified after returning.
        # This is important when using async scheduling.
        req_ids_output_copy = self.input_batch.req_ids.copy()
2613
        req_id_to_index_output_copy = self.input_batch.req_id_to_index.copy()
2614
2615

        num_sampled_tokens = sampler_output.sampled_token_ids.shape[0]
2616
        sampled_token_ids = sampler_output.sampled_token_ids
2617
        logprobs_tensors = sampler_output.logprobs_tensors
2618
        invalid_req_indices = []
2619
        cu_num_tokens: list[int] | None = None
2620
2621
2622
2623
2624
2625
        if not self.use_async_scheduling:
            # Get the valid generated tokens.
            max_gen_len = sampled_token_ids.shape[-1]
            if max_gen_len == 1:
                # No spec decode tokens.
                valid_sampled_token_ids = self._to_list(sampled_token_ids)
2626
2627
2628
                # Mask out the sampled tokens that should not be sampled.
                for i in discard_sampled_tokens_req_indices:
                    valid_sampled_token_ids[int(i)].clear()
2629
2630
            else:
                # Includes spec decode tokens.
2631
                valid_sampled_token_ids, cu_num_tokens = RejectionSampler.parse_output(
2632
2633
                    sampled_token_ids,
                    self.input_batch.vocab_size,
2634
2635
                    discard_sampled_tokens_req_indices,
                    return_cu_num_tokens=logprobs_tensors is not None,
2636
                )
2637
        else:
2638
            valid_sampled_token_ids = []
2639
            invalid_req_indices = discard_sampled_tokens_req_indices.tolist()
2640
2641
2642
2643
2644
            invalid_req_indices_set = set(invalid_req_indices)

            # Cache the sampled tokens on the GPU and avoid CPU sync.
            # These will be copied into input_ids in the next step
            # when preparing inputs.
2645
2646
2647
2648
            # With spec decoding, this is done in propose_draft_token_ids().
            if self.input_batch.prev_sampled_token_ids is None:
                assert sampled_token_ids.shape[-1] == 1
                self.input_batch.prev_sampled_token_ids = sampled_token_ids
2649
2650
2651
2652
2653
            self.input_batch.prev_req_id_to_index = {
                req_id: i
                for i, req_id in enumerate(self.input_batch.req_ids)
                if i not in invalid_req_indices_set
            }
2654

2655
2656
2657
2658
2659
        # Cache the sampled tokens in the model runner, so that the scheduler
        # doesn't need to send them back.
        # NOTE(woosuk): As an exception, when using PP, the scheduler sends
        # the sampled tokens back, because there's no direct communication
        # between the first-stage worker and the last-stage worker.
2660
        req_ids = self.input_batch.req_ids
2661
2662
        for req_idx in range(num_sampled_tokens):
            if self.use_async_scheduling:
2663
                sampled_ids = [-1] if req_idx not in invalid_req_indices_set else None
2664
2665
            else:
                sampled_ids = valid_sampled_token_ids[req_idx]
2666

2667
            num_sampled_ids: int = len(sampled_ids) if sampled_ids else 0
2668

2669
            if not sampled_ids:
2670
2671
2672
                continue

            start_idx = self.input_batch.num_tokens_no_spec[req_idx]
2673
            end_idx = start_idx + num_sampled_ids
2674
2675
2676
2677
            assert end_idx <= self.max_model_len, (
                "Sampled token IDs exceed the max model length. "
                f"Total number of tokens: {end_idx} > max_model_len: "
                f"{self.max_model_len}"
2678
            )
2679

2680
2681
            self.input_batch.token_ids_cpu[req_idx, start_idx:end_idx] = sampled_ids
            self.input_batch.is_token_ids[req_idx, start_idx:end_idx] = True
2682
2683
            self.input_batch.num_tokens_no_spec[req_idx] = end_idx
            self.input_batch.num_tokens[req_idx] = end_idx
2684

2685
            req_id = req_ids[req_idx]
2686
2687
2688
            req_state = self.requests[req_id]
            req_state.output_token_ids.extend(sampled_ids)

2689
        logprobs_lists = (
2690
            logprobs_tensors.tolists(cu_num_tokens)
2691
            if not self.use_async_scheduling and logprobs_tensors is not None
2692
2693
2694
2695
2696
2697
2698
2699
2700
            else None
        )

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

2701
2702
2703
2704
2705
2706
2707
2708
2709
2710
        return (
            num_nans_in_logits,
            logprobs_lists,
            valid_sampled_token_ids,
            prompt_logprobs_dict,
            req_ids_output_copy,
            req_id_to_index_output_copy,
            invalid_req_indices,
        )

2711
2712
2713
2714
2715
2716
2717
2718
2719
2720
2721
2722
2723
2724
2725
    @contextmanager
    def synchronize_input_prep(self):
        if self.prepare_inputs_event is None:
            yield
            return

        # Ensure prior step has finished with reused CPU tensors.
        # This is required in the async scheduling case because
        # the CPU->GPU transfer happens async.
        self.prepare_inputs_event.synchronize()
        try:
            yield
        finally:
            self.prepare_inputs_event.record()

2726
2727
    def _model_forward(
        self,
2728
2729
2730
2731
        input_ids: torch.Tensor | None = None,
        positions: torch.Tensor | None = None,
        intermediate_tensors: IntermediateTensors | None = None,
        inputs_embeds: torch.Tensor | None = None,
2732
2733
2734
2735
2736
        **model_kwargs: dict[str, Any],
    ) -> Any:
        """Helper method to call the model forward pass.

        This method can be overridden by subclasses for model execution.
2737
        Motivation: We can inspect only this method versus
2738
2739
2740
2741
2742
2743
2744
2745
2746
2747
2748
2749
2750
2751
2752
2753
2754
2755
2756
2757
        the whole execute_model, which has additional logic.

        Args:
            input_ids: Input token IDs
            positions: Token positions
            intermediate_tensors: Tensors from previous pipeline stages
            inputs_embeds: Input embeddings (alternative to input_ids)
            **model_kwargs: Additional model arguments

        Returns:
            Model output tensor
        """
        return self.model(
            input_ids=input_ids,
            positions=positions,
            intermediate_tensors=intermediate_tensors,
            inputs_embeds=inputs_embeds,
            **model_kwargs,
        )

2758
2759
2760
2761
2762
2763
2764
2765
2766
2767
2768
2769
2770
    def _determine_batch_execution_and_padding(
        self,
        num_tokens: int,
        num_reqs: int,
        num_scheduled_tokens_np: np.ndarray,
        max_num_scheduled_tokens: int,
        use_cascade_attn: bool,
        allow_microbatching: bool = True,
        force_eager: bool = False,
        # For cudagraph capture TODO(lucas): Refactor how we capture cudagraphs (will
        # be improved in model runner v2)
        force_uniform_decode: bool | None = None,
        force_has_lora: bool | None = None,
2771
        num_encoder_reqs: int = 0,
2772
    ) -> tuple[
2773
2774
        CUDAGraphMode,
        BatchDescriptor,
2775
        bool,
2776
2777
        torch.Tensor | None,
        CUDAGraphStat | None,
2778
2779
2780
2781
2782
2783
2784
2785
2786
2787
    ]:
        num_tokens_padded = self._pad_for_sequence_parallelism(num_tokens)
        uniform_decode = (
            (
                (max_num_scheduled_tokens == self.uniform_decode_query_len)
                and (num_tokens_padded == max_num_scheduled_tokens * num_reqs)
            )
            if force_uniform_decode is None
            else force_uniform_decode
        )
2788
2789
2790
2791
2792
        # Encoder-decoder models only support CG for decoder_step > 0 (no enc_output
        # is present). Also, chunked-prefill is disabled, so batch are uniform.
        has_encoder_output = (
            self.model_config.is_encoder_decoder and num_encoder_reqs > 0
        )
2793
2794
2795
2796
2797
2798
2799
2800

        has_lora = (
            len(self.input_batch.lora_id_to_lora_request) > 0
            if force_has_lora is None
            else force_has_lora
        )

        dispatch_cudagraph = (
2801
            lambda num_tokens, disable_full: self.cudagraph_dispatcher.dispatch(
2802
2803
2804
                num_tokens=num_tokens,
                has_lora=has_lora,
                uniform_decode=uniform_decode,
2805
                disable_full=disable_full,
2806
2807
2808
2809
2810
            )
            if not force_eager
            else (CUDAGraphMode.NONE, BatchDescriptor(num_tokens_padded))
        )

2811
        cudagraph_mode, batch_descriptor = dispatch_cudagraph(
2812
            num_tokens_padded, use_cascade_attn or has_encoder_output
2813
        )
2814
2815
2816
2817
        num_tokens_padded = batch_descriptor.num_tokens

        # Extra coordination when running data-parallel since we need to coordinate
        # across ranks
2818
        should_ubatch, num_tokens_across_dp = False, None
2819
2820
2821
2822
2823
2824
2825
2826
2827
        if self.vllm_config.parallel_config.data_parallel_size > 1:
            # Disable DP padding when running eager to avoid excessive padding when
            # running prefills. This lets us set cudagraph_mode="NONE" on the prefiller
            # in a P/D setup and still use CUDA graphs (enabled by this padding) on the
            # decoder.
            allow_dp_padding = (
                self.compilation_config.cudagraph_mode != CUDAGraphMode.NONE
            )

2828
2829
2830
2831
2832
2833
2834
2835
2836
2837
2838
            should_ubatch, num_tokens_across_dp, synced_cudagraph_mode = (
                coordinate_batch_across_dp(
                    num_tokens_unpadded=num_tokens,
                    parallel_config=self.parallel_config,
                    allow_microbatching=allow_microbatching,
                    allow_dp_padding=allow_dp_padding,
                    num_tokens_padded=num_tokens_padded,
                    uniform_decode=uniform_decode,
                    num_scheduled_tokens_per_request=num_scheduled_tokens_np,
                    cudagraph_mode=cudagraph_mode.value,
                )
2839
2840
            )

2841
            # Extract DP-synced values
2842
2843
2844
            if num_tokens_across_dp is not None:
                dp_rank = self.parallel_config.data_parallel_rank
                num_tokens_padded = int(num_tokens_across_dp[dp_rank].item())
2845
2846
2847
2848
2849
                # Re-dispatch with DP padding so we have the correct batch_descriptor
                cudagraph_mode, batch_descriptor = dispatch_cudagraph(
                    num_tokens_padded,
                    disable_full=synced_cudagraph_mode <= CUDAGraphMode.PIECEWISE.value,
                )
2850
2851
2852
2853
                # Assert to make sure the agreed upon token count is correct otherwise
                # num_tokens_across_dp will no-longer be valid
                assert batch_descriptor.num_tokens == num_tokens_padded

2854
2855
2856
2857
2858
2859
2860
2861
2862
2863
2864
2865
        cudagraph_stats = None
        if self.vllm_config.observability_config.cudagraph_metrics:
            cudagraph_stats = CUDAGraphStat(
                num_unpadded_tokens=num_tokens,
                num_padded_tokens=batch_descriptor.num_tokens,
                num_paddings=batch_descriptor.num_tokens - num_tokens,
                runtime_mode=str(cudagraph_mode),
            )

        return (
            cudagraph_mode,
            batch_descriptor,
2866
            should_ubatch,
2867
2868
2869
            num_tokens_across_dp,
            cudagraph_stats,
        )
2870

2871
2872
2873
2874
2875
2876
2877
2878
2879
2880
2881
2882
2883
2884
2885
2886
2887
2888
2889
2890
2891
2892
2893
2894
2895
2896
2897
2898
2899
2900
2901
2902
2903
2904
2905
2906
    def _register_layerwise_nvtx_hooks(self) -> None:
        """
        Register layerwise NVTX hooks if --enable-layerwise-nvtx-tracing is enabled
        to trace detailed information of each layer or module in the model.
        """

        if (
            self.vllm_config.observability_config.enable_layerwise_nvtx_tracing
            and not self.layerwise_nvtx_hooks_registered
        ):
            if self.compilation_config.cudagraph_mode != CUDAGraphMode.NONE:
                logger.debug_once(
                    "layerwise NVTX tracing is not supported when CUDA graph is "
                    "turned off; you may observe part or all of the model "
                    "missing NVTX markers"
                )

            # In STOCK_TORCH_COMPILE mode, after registering hooks here,
            # the __call__ function of nn.module will be recompiled with
            # fullgraph=True. Since nvtx.range_push/pop are not traceable
            # by torch dynamo, we can't register hook functions here
            # because hook functions will also be traced by torch dynamo.
            if (
                self.vllm_config.compilation_config.mode
                == CompilationMode.STOCK_TORCH_COMPILE
            ):
                logger.debug_once(
                    "layerwise NVTX tracing is not supported when "
                    "CompilationMode is STOCK_TORCH_COMPILE, skipping "
                    "function hooks registration"
                )
            else:
                pyt_hooks = PytHooks()
                pyt_hooks.register_hooks(self.model, self.model.__class__.__name__)
                self.layerwise_nvtx_hooks_registered = True

2907
2908
2909
2910
    @torch.inference_mode()
    def execute_model(
        self,
        scheduler_output: "SchedulerOutput",
2911
        intermediate_tensors: IntermediateTensors | None = None,
2912
2913
2914
2915
2916
2917
    ) -> ModelRunnerOutput | IntermediateTensors | None:
        if self.execute_model_state is not None:
            raise RuntimeError(
                "State error: sample_tokens() must be called "
                "after execute_model() returns None."
            )
2918
2919
2920
2921
2922
2923
2924
2925
2926
2927
2928
2929
2930
2931
2932

        # self._draft_token_ids is None when `input_fits_in_drafter=False`
        # and there is no draft tokens scheduled. so it need to update the
        # spec_decoding info in scheduler_output with async_scheduling.
        # use deepcopy to avoid the modification has influence on the
        # scheduler_output in engine core process.
        # TODO(Ronald1995): deepcopy is expensive when there is a large
        # number of requests, optimize it later.
        if (
            self.use_async_scheduling
            and self.num_spec_tokens
            and self._draft_token_ids is None
        ):
            scheduler_output = deepcopy(scheduler_output)

2933
        num_scheduled_tokens = scheduler_output.total_num_scheduled_tokens
2934
        with record_function_or_nullcontext("gpu_model_runner: preprocess"):
2935
2936
2937
2938
            with self.synchronize_input_prep():
                # Update persistent batch states.
                self._update_states(scheduler_output)

2939
2940
2941
2942
2943
2944
2945
2946
                if has_ec_transfer() and get_ec_transfer().is_producer:
                    with self.maybe_get_ec_connector_output(
                        scheduler_output,
                        encoder_cache=self.encoder_cache,
                    ) as ec_connector_output:
                        self._execute_mm_encoder(scheduler_output)
                        return make_empty_encoder_model_runner_output(scheduler_output)

2947
                if not num_scheduled_tokens:
2948
2949
2950
2951
2952
2953
2954
2955
2956
2957
2958
                    if (
                        self.parallel_config.distributed_executor_backend
                        == "external_launcher"
                        and self.parallel_config.data_parallel_size > 1
                    ):
                        # this is a corner case when both external launcher
                        # and DP are enabled, num_scheduled_tokens could be
                        # 0, and has_unfinished_requests in the outer loop
                        # returns True. before returning early here we call
                        # dummy run to ensure coordinate_batch_across_dp
                        # is called into to avoid out of sync issues.
2959
                        self._dummy_run(1)
2960
2961
2962
2963
                    if not has_kv_transfer_group():
                        # Return empty ModelRunnerOutput if no work to do.
                        return EMPTY_MODEL_RUNNER_OUTPUT
                    return self.kv_connector_no_forward(
2964
2965
                        scheduler_output, self.vllm_config
                    )
2966
                if self.cache_config.kv_sharing_fast_prefill:
2967
                    assert not self.num_prompt_logprobs, (
2968
2969
                        "--kv-sharing-fast-prefill produces incorrect "
                        "logprobs for prompt tokens, tokens, please disable "
2970
2971
                        "it when the requests need prompt logprobs"
                    )
2972

2973
2974
2975
2976
2977
                num_reqs = self.input_batch.num_reqs
                req_ids = self.input_batch.req_ids
                tokens = [scheduler_output.num_scheduled_tokens[i] for i in req_ids]
                num_scheduled_tokens_np = np.array(tokens, dtype=np.int32)
                max_num_scheduled_tokens = int(num_scheduled_tokens_np.max())
2978
                num_tokens_unpadded = scheduler_output.total_num_scheduled_tokens
2979

2980
2981
2982
                (
                    logits_indices,
                    spec_decode_metadata,
2983
                ) = self._prepare_inputs(
2984
2985
                    scheduler_output,
                    num_scheduled_tokens_np,
2986
2987
2988
2989
                )

                cascade_attn_prefix_lens = None
                # Disable cascade attention when using microbatching (DBO)
2990
                if self.cascade_attn_enabled and not self.parallel_config.enable_dbo:
2991
2992
2993
                    # Pre-compute cascade attention prefix lengths
                    cascade_attn_prefix_lens = self._compute_cascade_attn_prefix_lens(
                        num_scheduled_tokens_np,
2994
                        self.input_batch.num_computed_tokens_cpu[:num_reqs],
2995
2996
2997
                        scheduler_output.num_common_prefix_blocks,
                    )

2998
2999
3000
                (
                    cudagraph_mode,
                    batch_desc,
3001
                    should_ubatch,
3002
                    num_tokens_across_dp,
3003
                    cudagraph_stats,
3004
3005
3006
3007
3008
3009
                ) = self._determine_batch_execution_and_padding(
                    num_tokens=num_tokens_unpadded,
                    num_reqs=num_reqs,
                    num_scheduled_tokens_np=num_scheduled_tokens_np,
                    max_num_scheduled_tokens=max_num_scheduled_tokens,
                    use_cascade_attn=cascade_attn_prefix_lens is not None,
3010
                    num_encoder_reqs=len(scheduler_output.scheduled_encoder_inputs),
3011
3012
3013
3014
                )

                logger.debug(
                    "Running batch with cudagraph_mode: %s, batch_descriptor: %s, "
3015
                    "should_ubatch: %s, num_tokens_across_dp: %s",
3016
3017
                    cudagraph_mode,
                    batch_desc,
3018
                    should_ubatch,
3019
3020
3021
3022
3023
3024
3025
                    num_tokens_across_dp,
                )

                num_tokens_padded = batch_desc.num_tokens
                num_reqs_padded = (
                    batch_desc.num_reqs if batch_desc.num_reqs is not None else num_reqs
                )
3026
3027
3028
3029
3030
3031
                ubatch_slices, ubatch_slices_padded = maybe_create_ubatch_slices(
                    should_ubatch,
                    num_scheduled_tokens_np,
                    num_tokens_padded,
                    num_reqs_padded,
                )
3032

3033
3034
                pad_attn = cudagraph_mode == CUDAGraphMode.FULL

3035
3036
3037
                use_spec_decode = len(scheduler_output.scheduled_spec_decode_tokens) > 0
                ubatch_slices_attn = ubatch_slices_padded if pad_attn else ubatch_slices

3038
                (attn_metadata, spec_decode_common_attn_metadata) = (
3039
                    self._build_attention_metadata(
3040
3041
                        num_tokens=num_tokens_unpadded,
                        num_tokens_padded=num_tokens_padded if pad_attn else None,
3042
                        num_reqs=num_reqs,
3043
3044
                        num_reqs_padded=num_reqs_padded if pad_attn else None,
                        max_query_len=max_num_scheduled_tokens,
3045
                        ubatch_slices=ubatch_slices_attn,
3046
3047
                        logits_indices=logits_indices,
                        use_spec_decode=use_spec_decode,
3048
                        num_scheduled_tokens=scheduler_output.num_scheduled_tokens,
3049
3050
3051
                        cascade_attn_prefix_lens=cascade_attn_prefix_lens,
                    )
                )
3052

3053
3054
3055
3056
3057
3058
3059
3060
3061
            (
                input_ids,
                inputs_embeds,
                positions,
                intermediate_tensors,
                model_kwargs,
                ec_connector_output,
            ) = self._preprocess(
                scheduler_output, num_tokens_padded, intermediate_tensors
3062
            )
3063

3064
        # Set cudagraph mode to none if calc_kv_scales is true.
3065
3066
3067
        # KV scales calculation involves dynamic operations that are incompatible
        # with CUDA graph capture.
        if self.calculate_kv_scales:
3068
            cudagraph_mode = CUDAGraphMode.NONE
3069
3070
            # Mark KV scales as calculated after the first forward pass
            self.calculate_kv_scales = False
3071

3072
3073
        # Run the model.
        # Use persistent buffers for CUDA graphs.
3074
3075
        with (
            set_forward_context(
3076
3077
                attn_metadata,
                self.vllm_config,
3078
                num_tokens=num_tokens_padded,
3079
                num_tokens_across_dp=num_tokens_across_dp,
3080
3081
                cudagraph_runtime_mode=cudagraph_mode,
                batch_descriptor=batch_desc,
3082
                ubatch_slices=ubatch_slices_padded,
3083
            ),
3084
            record_function_or_nullcontext("gpu_model_runner: forward"),
3085
3086
            self.maybe_get_kv_connector_output(scheduler_output) as kv_connector_output,
        ):
3087
            model_output = self._model_forward(
3088
3089
3090
3091
3092
3093
3094
                input_ids=input_ids,
                positions=positions,
                intermediate_tensors=intermediate_tensors,
                inputs_embeds=inputs_embeds,
                **model_kwargs,
            )

3095
        with record_function_or_nullcontext("gpu_model_runner: postprocess"):
3096
            if self.use_aux_hidden_state_outputs:
3097
                # True when EAGLE 3 is used.
3098
3099
                hidden_states, aux_hidden_states = model_output
            else:
3100
                # Common case.
3101
3102
3103
                hidden_states = model_output
                aux_hidden_states = None

3104
3105
3106
3107
3108
            if not self.broadcast_pp_output:
                # Common case.
                if not get_pp_group().is_last_rank:
                    # Return the intermediate tensors.
                    assert isinstance(hidden_states, IntermediateTensors)
3109
                    hidden_states.kv_connector_output = kv_connector_output
3110
                    self.kv_connector_output = kv_connector_output
3111
                    return hidden_states
3112

3113
                if self.is_pooling_model:
3114
                    # Return the pooling output.
3115
3116
3117
                    output = self._pool(
                        hidden_states, num_scheduled_tokens, num_scheduled_tokens_np
                    )
3118
3119
                    output.kv_connector_output = kv_connector_output
                    return output
3120
3121

                sample_hidden_states = hidden_states[logits_indices]
3122
                logits = self.model.compute_logits(sample_hidden_states)
3123
3124
3125
3126
            else:
                # Rare case.
                assert not self.is_pooling_model

3127
                sample_hidden_states = hidden_states[logits_indices]
3128
                if not get_pp_group().is_last_rank:
3129
                    all_gather_tensors = {
3130
                        "residual": not is_residual_scattered_for_sp(
3131
                            self.vllm_config, num_tokens_padded
3132
                        )
3133
                    }
3134
                    get_pp_group().send_tensor_dict(
3135
3136
                        hidden_states.tensors,
                        all_gather_group=get_tp_group(),
3137
3138
                        all_gather_tensors=all_gather_tensors,
                    )
3139
3140
                    logits = None
                else:
3141
                    logits = self.model.compute_logits(sample_hidden_states)
3142

3143
                model_output_broadcast_data: dict[str, Any] = {}
3144
3145
3146
                if logits is not None:
                    model_output_broadcast_data["logits"] = logits.contiguous()

3147
                broadcasted = get_pp_group().broadcast_tensor_dict(
3148
3149
                    model_output_broadcast_data, src=len(get_pp_group().ranks) - 1
                )
3150
3151
                assert broadcasted is not None
                logits = broadcasted["logits"]
3152

3153
3154
3155
3156
3157
3158
3159
3160
        self.execute_model_state = ExecuteModelState(
            scheduler_output,
            logits,
            spec_decode_metadata,
            spec_decode_common_attn_metadata,
            hidden_states,
            sample_hidden_states,
            aux_hidden_states,
3161
            ec_connector_output,
3162
            cudagraph_stats,
3163
        )
3164
        self.kv_connector_output = kv_connector_output
3165
3166
3167
3168
3169
3170
        return None

    @torch.inference_mode
    def sample_tokens(
        self, grammar_output: "GrammarOutput | None"
    ) -> ModelRunnerOutput | AsyncModelRunnerOutput | IntermediateTensors:
3171
3172
3173
        kv_connector_output = self.kv_connector_output
        self.kv_connector_output = None

3174
3175
        if self.execute_model_state is None:
            # Nothing to do (PP non-final rank case), output isn't used.
3176
            if not kv_connector_output:
3177
                return None  # type: ignore[return-value]
3178
3179
3180
3181
3182
3183
3184
3185
3186

            # In case of PP with kv transfer, we need to pass through the
            # kv_connector_output
            if kv_connector_output.is_empty():
                return EMPTY_MODEL_RUNNER_OUTPUT

            output = copy(EMPTY_MODEL_RUNNER_OUTPUT)
            output.kv_connector_output = kv_connector_output
            return output
3187
3188
3189
3190
3191
3192
3193
3194
3195
3196

        # Unpack ephemeral state.
        (
            scheduler_output,
            logits,
            spec_decode_metadata,
            spec_decode_common_attn_metadata,
            hidden_states,
            sample_hidden_states,
            aux_hidden_states,
3197
            ec_connector_output,
3198
            cudagraph_stats,
3199
3200
3201
3202
3203
3204
3205
3206
3207
        ) = self.execute_model_state
        # Clear ephemeral state.
        self.execute_model_state = None

        # Apply structured output bitmasks if present.
        if grammar_output is not None:
            apply_grammar_bitmask(
                scheduler_output, grammar_output, self.input_batch, logits
            )
3208

3209
        with record_function_or_nullcontext("gpu_model_runner: sample"):
3210
3211
            sampler_output = self._sample(logits, spec_decode_metadata)

3212
3213
        self.input_batch.prev_sampled_token_ids = None

3214
        def propose_draft_token_ids(sampled_token_ids):
3215
            assert spec_decode_common_attn_metadata is not None
3216
            with record_function_or_nullcontext("gpu_model_runner: draft"):
3217
3218
3219
3220
3221
3222
3223
3224
3225
3226
3227
                self._draft_token_ids = self.propose_draft_token_ids(
                    scheduler_output,
                    sampled_token_ids,
                    self.input_batch.sampling_metadata,
                    hidden_states,
                    sample_hidden_states,
                    aux_hidden_states,
                    spec_decode_metadata,
                    spec_decode_common_attn_metadata,
                )

3228
        spec_config = self.speculative_config
3229
        use_padded_batch_for_eagle = (
3230
3231
3232
            spec_config is not None
            and spec_config.use_eagle()
            and not spec_config.disable_padded_drafter_batch
3233
        )
3234
3235
3236
        effective_drafter_max_model_len = self.max_model_len
        if effective_drafter_max_model_len is None:
            effective_drafter_max_model_len = self.model_config.max_model_len
3237
        if (
3238
3239
3240
            spec_config is not None
            and spec_config.draft_model_config is not None
            and spec_config.draft_model_config.max_model_len is not None
3241
        ):
3242
            effective_drafter_max_model_len = (
3243
                spec_config.draft_model_config.max_model_len
3244
            )
3245
        input_fits_in_drafter = spec_decode_common_attn_metadata and (
3246
            spec_decode_common_attn_metadata.max_seq_len + self.num_spec_tokens
3247
3248
            <= effective_drafter_max_model_len
        )
3249
        if use_padded_batch_for_eagle:
3250
3251
            assert self.speculative_config is not None
            assert isinstance(self.drafter, EagleProposer)
3252
3253
3254
3255
3256
3257
            sampled_token_ids = sampler_output.sampled_token_ids
            if input_fits_in_drafter:
                # EAGLE speculative decoding can use the GPU sampled tokens
                # as inputs, and does not need to wait for bookkeeping to finish.
                propose_draft_token_ids(sampled_token_ids)
            elif self.valid_sampled_token_count_event is not None:
3258
                assert spec_decode_common_attn_metadata is not None
3259
3260
3261
3262
3263
3264
                next_token_ids, valid_sampled_tokens_count = (
                    self.drafter.prepare_next_token_ids_padded(
                        spec_decode_common_attn_metadata,
                        sampled_token_ids,
                        self.requests,
                        self.input_batch,
3265
                        self.discard_request_mask.gpu,
3266
3267
3268
3269
3270
                    )
                )
                self._copy_valid_sampled_token_count(
                    next_token_ids, valid_sampled_tokens_count
                )
3271

3272
        with record_function_or_nullcontext("gpu_model_runner: bookkeep"):
3273
3274
3275
3276
3277
3278
3279
3280
            (
                num_nans_in_logits,
                logprobs_lists,
                valid_sampled_token_ids,
                prompt_logprobs_dict,
                req_ids_output_copy,
                req_id_to_index_output_copy,
                invalid_req_indices,
3281
3282
3283
3284
3285
            ) = self._bookkeeping_sync(
                scheduler_output,
                sampler_output,
                logits,
                hidden_states,
3286
                scheduler_output.total_num_scheduled_tokens,
3287
                spec_decode_metadata,
3288
            )
3289

3290
3291
3292
3293
3294
        if (
            self.speculative_config
            and not use_padded_batch_for_eagle
            and input_fits_in_drafter
        ):
3295
3296
3297
            # ngram and other speculative decoding methods use the sampled
            # tokens on the CPU, so they are run after bookkeeping.
            propose_draft_token_ids(valid_sampled_token_ids)
3298

3299
        with record_function_or_nullcontext("gpu_model_runner: eplb"):
3300
            self.eplb_step()
3301
3302
3303
3304
3305
3306
3307
3308
3309
        with record_function_or_nullcontext("gpu_model_runner: ModelRunnerOutput"):
            output = ModelRunnerOutput(
                req_ids=req_ids_output_copy,
                req_id_to_index=req_id_to_index_output_copy,
                sampled_token_ids=valid_sampled_token_ids,
                logprobs=logprobs_lists,
                prompt_logprobs_dict=prompt_logprobs_dict,
                pooler_output=[],
                kv_connector_output=kv_connector_output,
3310
3311
3312
                ec_connector_output=ec_connector_output
                if self.supports_mm_inputs
                else None,
3313
                num_nans_in_logits=num_nans_in_logits,
3314
                cudagraph_stats=cudagraph_stats,
3315
            )
3316

3317
3318
        if not self.use_async_scheduling:
            return output
3319
3320
3321
3322
3323
3324
3325
3326
3327
        with record_function_or_nullcontext(
            "gpu_model_runner: AsyncGPUModelRunnerOutput"
        ):
            async_output = AsyncGPUModelRunnerOutput(
                model_runner_output=output,
                sampled_token_ids=sampler_output.sampled_token_ids,
                logprobs_tensors=sampler_output.logprobs_tensors,
                invalid_req_indices=invalid_req_indices,
                async_output_copy_stream=self.async_output_copy_stream,
3328
                vocab_size=self.input_batch.vocab_size,
3329
3330
3331
3332
3333
            )
        with record_function_or_nullcontext(
            "gpu_model_runner: set_async_sampled_token_ids"
        ):
            # Save ref of sampled_token_ids CPU tensor if the batch contains
3334
            # any requests with sampling params that require output ids.
3335
3336
3337
3338
            self.input_batch.set_async_sampled_token_ids(
                async_output.sampled_token_ids_cpu,
                async_output.async_copy_ready_event,
            )
3339
3340
3341

        return async_output

3342
    def take_draft_token_ids(self) -> DraftTokenIds | None:
3343
3344
3345
3346
3347
3348
3349
3350
3351
3352
        if self._draft_token_ids is None:
            return None
        req_ids = self.input_batch.req_ids
        if isinstance(self._draft_token_ids, torch.Tensor):
            draft_token_ids = self._draft_token_ids.tolist()
        else:
            draft_token_ids = self._draft_token_ids
        self._draft_token_ids = None
        return DraftTokenIds(req_ids, draft_token_ids)

3353
3354
3355
3356
3357
3358
3359
3360
3361
3362
3363
3364
3365
3366
3367
3368
3369
3370
3371
3372
3373
3374
3375
3376
3377
3378
3379
3380
3381
3382
3383
    def _copy_valid_sampled_token_count(
        self, next_token_ids: torch.Tensor, valid_sampled_tokens_count: torch.Tensor
    ) -> None:
        if self.valid_sampled_token_count_event is None:
            return

        default_stream = torch.cuda.current_stream()
        # Initialize a new stream to overlap the copy operation with
        # prepare_input of draft model.
        with torch.cuda.stream(self.valid_sampled_token_count_copy_stream):
            self.valid_sampled_token_count_copy_stream.wait_stream(default_stream)  # type: ignore
            counts = valid_sampled_tokens_count
            counts_cpu = self.valid_sampled_token_count_cpu
            counts_cpu[: counts.shape[0]].copy_(counts, non_blocking=True)
            self.valid_sampled_token_count_event.record()

        self.input_batch.prev_sampled_token_ids = next_token_ids.unsqueeze(1)

    def _get_valid_sampled_token_count(self) -> list[int]:
        # Wait until valid_sampled_tokens_count is copied to cpu,
        prev_sampled_token_ids = self.input_batch.prev_sampled_token_ids
        if (
            self.valid_sampled_token_count_event is None
            or prev_sampled_token_ids is None
        ):
            return []

        counts_cpu = self.valid_sampled_token_count_cpu
        self.valid_sampled_token_count_event.synchronize()
        return counts_cpu[: prev_sampled_token_ids.shape[0]].tolist()

3384
3385
3386
    def propose_draft_token_ids(
        self,
        scheduler_output: "SchedulerOutput",
3387
        sampled_token_ids: torch.Tensor | list[list[int]],
3388
3389
3390
        sampling_metadata: SamplingMetadata,
        hidden_states: torch.Tensor,
        sample_hidden_states: torch.Tensor,
3391
3392
        aux_hidden_states: list[torch.Tensor] | None,
        spec_decode_metadata: SpecDecodeMetadata | None,
3393
        common_attn_metadata: CommonAttentionMetadata,
3394
    ) -> list[list[int]] | torch.Tensor:
3395
        num_scheduled_tokens = scheduler_output.total_num_scheduled_tokens
3396
3397
3398
        spec_config = self.speculative_config
        assert spec_config is not None
        if spec_config.method == "ngram":
3399
            assert isinstance(sampled_token_ids, list)
3400
            assert isinstance(self.drafter, NgramProposer)
3401
            draft_token_ids = self.drafter.propose(
3402
3403
                sampled_token_ids,
                self.input_batch.req_ids,
3404
3405
                self.input_batch.num_tokens_no_spec,
                self.input_batch.token_ids_cpu,
3406
3407
                self.input_batch.spec_decode_unsupported_reqs,
            )
3408
        elif spec_config.method == "suffix":
3409
3410
3411
            assert isinstance(sampled_token_ids, list)
            assert isinstance(self.drafter, SuffixDecodingProposer)
            draft_token_ids = self.drafter.propose(self.input_batch, sampled_token_ids)
3412
        elif spec_config.method == "medusa":
3413
            assert isinstance(sampled_token_ids, list)
3414
            assert isinstance(self.drafter, MedusaProposer)
3415

3416
3417
            if sample_hidden_states.shape[0] == len(sampled_token_ids):
                # The input to the target model does not include draft tokens.
3418
3419
3420
3421
                hidden_states = sample_hidden_states
            else:
                indices = []
                offset = 0
3422
3423
3424
                assert spec_decode_metadata is not None, (
                    "No spec decode metadata for medusa"
                )
3425
                for num_draft, tokens in zip(
3426
3427
                    spec_decode_metadata.num_draft_tokens, sampled_token_ids
                ):
3428
                    indices.append(offset + len(tokens) - 1)
3429
                    offset += num_draft + 1
3430
                indices = torch.tensor(indices, device=self.device)
3431
3432
                hidden_states = sample_hidden_states[indices]

3433
            draft_token_ids = self.drafter.propose(
3434
3435
3436
                target_hidden_states=hidden_states,
                sampling_metadata=sampling_metadata,
            )
3437
        elif spec_config.use_eagle():
3438
            assert isinstance(self.drafter, EagleProposer)
3439

3440
            if spec_config.disable_padded_drafter_batch:
3441
3442
3443
                # When padded-batch is disabled, the sampled_token_ids should be
                # the cpu-side list[list[int]] of valid sampled tokens for each
                # request, with invalid requests having empty lists.
3444
3445
                assert isinstance(sampled_token_ids, list), (
                    "sampled_token_ids should be a python list when"
3446
                    "padded-batch is disabled."
3447
                )
3448
                next_token_ids = self.drafter.prepare_next_token_ids_cpu(
3449
3450
3451
3452
3453
                    sampled_token_ids,
                    self.requests,
                    self.input_batch,
                    scheduler_output.num_scheduled_tokens,
                )
3454
3455
3456
3457
3458
            else:
                # When using padded-batch, the sampled_token_ids should be
                # the gpu tensor of sampled tokens for each request, of shape
                # (num_reqs, num_spec_tokens + 1) with rejected tokens having
                # value -1.
3459
3460
                assert isinstance(sampled_token_ids, torch.Tensor), (
                    "sampled_token_ids should be a torch.Tensor when"
3461
                    "padded-batch is enabled."
3462
3463
                )
                next_token_ids, valid_sampled_tokens_count = (
3464
3465
3466
3467
3468
                    self.drafter.prepare_next_token_ids_padded(
                        common_attn_metadata,
                        sampled_token_ids,
                        self.requests,
                        self.input_batch,
3469
                        self.discard_request_mask.gpu,
3470
                    )
3471
                )
3472
3473
3474
                self._copy_valid_sampled_token_count(
                    next_token_ids, valid_sampled_tokens_count
                )
Jiayi Yao's avatar
Jiayi Yao committed
3475

3476
            if spec_decode_metadata is None:
3477
                token_indices_to_sample = None
3478
                # input_ids can be None for multimodal models.
3479
                target_token_ids = self.input_ids.gpu[:num_scheduled_tokens]
3480
                target_positions = self._get_positions(num_scheduled_tokens)
3481
                if self.use_aux_hidden_state_outputs:
Wentao Ye's avatar
Wentao Ye committed
3482
                    assert aux_hidden_states is not None
3483
                    target_hidden_states = torch.cat(
3484
3485
                        [h[:num_scheduled_tokens] for h in aux_hidden_states], dim=-1
                    )
3486
3487
                else:
                    target_hidden_states = hidden_states[:num_scheduled_tokens]
3488
            else:
3489
                if spec_config.disable_padded_drafter_batch:
3490
                    token_indices_to_sample = None
3491
3492
3493
3494
3495
                    common_attn_metadata, token_indices = self.drafter.prepare_inputs(
                        common_attn_metadata,
                        sampled_token_ids,
                        spec_decode_metadata.num_draft_tokens,
                    )
3496
3497
3498
3499
3500
3501
3502
3503
3504
                    target_token_ids = self.input_ids.gpu[token_indices]
                    target_positions = self._get_positions(token_indices)
                    if self.use_aux_hidden_state_outputs:
                        assert aux_hidden_states is not None
                        target_hidden_states = torch.cat(
                            [h[token_indices] for h in aux_hidden_states], dim=-1
                        )
                    else:
                        target_hidden_states = hidden_states[token_indices]
3505
                else:
3506
                    common_attn_metadata, token_indices_to_sample = (
3507
3508
3509
                        self.drafter.prepare_inputs_padded(
                            common_attn_metadata,
                            spec_decode_metadata,
3510
3511
3512
                            valid_sampled_tokens_count,
                        )
                    )
3513
3514
3515
3516
3517
3518
3519
3520
3521
3522
3523
                    total_num_tokens = common_attn_metadata.num_actual_tokens
                    # When padding the batch, token_indices is just a range
                    target_token_ids = self.input_ids.gpu[:total_num_tokens]
                    target_positions = self._get_positions(total_num_tokens)
                    if self.use_aux_hidden_state_outputs:
                        assert aux_hidden_states is not None
                        target_hidden_states = torch.cat(
                            [h[:total_num_tokens] for h in aux_hidden_states], dim=-1
                        )
                    else:
                        target_hidden_states = hidden_states[:total_num_tokens]
3524

3525
            if self.supports_mm_inputs:
3526
3527
3528
3529
3530
3531
                mm_embed_inputs = self._gather_mm_embeddings(
                    scheduler_output,
                    shift_computed_tokens=1,
                )
            else:
                mm_embed_inputs = None
3532

3533
            draft_token_ids = self.drafter.propose(
3534
3535
3536
3537
                target_token_ids=target_token_ids,
                target_positions=target_positions,
                target_hidden_states=target_hidden_states,
                next_token_ids=next_token_ids,
3538
                last_token_indices=token_indices_to_sample,
3539
                sampling_metadata=sampling_metadata,
3540
                common_attn_metadata=common_attn_metadata,
3541
                mm_embed_inputs=mm_embed_inputs,
3542
            )
3543

3544
        return draft_token_ids
3545

3546
3547
3548
    def update_config(self, overrides: dict[str, Any]) -> None:
        allowed_config_names = {"load_config", "model_config"}
        for config_name, config_overrides in overrides.items():
3549
3550
            assert config_name in allowed_config_names, (
                f"Config `{config_name}` not supported. "
3551
                f"Allowed configs: {allowed_config_names}"
3552
            )
3553
3554
3555
3556
            config = getattr(self, config_name)
            new_config = update_config(config, config_overrides)
            setattr(self, config_name, new_config)

3557
3558
3559
3560
3561
    def load_model(self, eep_scale_up: bool = False) -> None:
        """
        Args:
            eep_scale_up: the model loading is for elastic EP scale up.
        """
3562
3563
3564
3565
3566
        logger.info_once(
            "Starting to load model %s...",
            self.model_config.model,
            scope="global",
        )
3567
3568
3569
3570
3571
        global_expert_loads, old_global_expert_indices_per_model, rank_mapping = (
            EplbState.get_eep_state(self.parallel_config)
            if eep_scale_up
            else (None, None, None)
        )
3572

3573
3574
3575
3576
        if self.parallel_config.enable_eplb:
            self.eplb_state = EplbState(self.parallel_config, self.device)
            eplb_models = 0

3577
3578
3579
3580
3581
3582
3583
3584
3585
3586
        try:
            with DeviceMemoryProfiler() as m:
                time_before_load = time.perf_counter()
                model_loader = get_model_loader(self.load_config)
                self.model = model_loader.load_model(
                    vllm_config=self.vllm_config, model_config=self.model_config
                )
                if self.lora_config:
                    self.model = self.load_lora_model(
                        self.model, self.vllm_config, self.device
3587
                    )
3588
3589
3590
3591
3592
3593
3594
3595
3596
3597
3598
3599
3600
3601
3602
                if hasattr(self, "drafter"):
                    logger.info_once("Loading drafter model...")
                    self.drafter.load_model(self.model)
                    if (
                        hasattr(self.drafter, "model")
                        and is_mixture_of_experts(self.drafter.model)
                        and self.parallel_config.enable_eplb
                    ):
                        spec_config = self.vllm_config.speculative_config
                        assert spec_config is not None
                        assert spec_config.draft_model_config is not None
                        logger.info_once(
                            "EPLB is enabled for drafter model %s.",
                            spec_config.draft_model_config.model,
                        )
3603

3604
3605
3606
3607
3608
3609
3610
3611
3612
3613
3614
3615
3616
3617
3618
3619
3620
3621
3622
3623
3624
3625
                        global_expert_load = (
                            global_expert_loads[eplb_models]
                            if global_expert_loads
                            else None
                        )
                        old_global_expert_indices = (
                            old_global_expert_indices_per_model[eplb_models]
                            if old_global_expert_indices_per_model
                            else None
                        )
                        if self.eplb_state is None:
                            self.eplb_state = EplbState(
                                self.parallel_config, self.device
                            )
                        self.eplb_state.add_model(
                            self.drafter.model,
                            spec_config.draft_model_config,
                            global_expert_load,
                            old_global_expert_indices,
                            rank_mapping,
                        )
                        eplb_models += 1
3626

3627
3628
3629
3630
3631
3632
                if self.use_aux_hidden_state_outputs:
                    if not supports_eagle3(self.get_model()):
                        raise RuntimeError(
                            "Model does not support EAGLE3 interface but "
                            "aux_hidden_state_outputs was requested"
                        )
3633

3634
3635
3636
3637
3638
3639
3640
3641
3642
3643
3644
3645
3646
3647
3648
3649
3650
3651
3652
3653
3654
3655
3656
3657
3658
                    # Try to get auxiliary layers from speculative config,
                    # otherwise use model's default layers
                    aux_layers = self._get_eagle3_aux_layers_from_config()
                    if aux_layers:
                        logger.info(
                            "Using auxiliary layers from speculative config: %s",
                            aux_layers,
                        )
                    else:
                        aux_layers = self.model.get_eagle3_aux_hidden_state_layers()

                    self.model.set_aux_hidden_state_layers(aux_layers)
                time_after_load = time.perf_counter()
            self.model_memory_usage = m.consumed_memory
        except torch.cuda.OutOfMemoryError as e:
            msg = (
                "Failed to load model - not enough GPU memory. "
                "Try lowering --gpu-memory-utilization to free memory for weights, "
                "increasing --tensor-parallel-size, or using --quantization. "
                "See https://docs.vllm.ai/en/latest/configuration/conserving_memory/ "
                "for more tips."
            )
            combined_msg = f"{msg} (original error: {e})"
            logger.error(combined_msg)
            raise e
3659
        logger.info_once(
3660
            "Model loading took %.4f GiB memory and %.6f seconds",
3661
3662
            self.model_memory_usage / GiB_bytes,
            time_after_load - time_before_load,
3663
            scope="local",
3664
        )
3665
        prepare_communication_buffer_for_model(self.model)
3666
3667
3668
3669
        if (drafter := getattr(self, "drafter", None)) and (
            drafter_model := getattr(drafter, "model", None)
        ):
            prepare_communication_buffer_for_model(drafter_model)
3670
        mm_config = self.model_config.multimodal_config
3671
        self.is_multimodal_pruning_enabled = (
3672
            supports_multimodal_pruning(self.get_model())
3673
3674
            and mm_config is not None
            and mm_config.is_multimodal_pruning_enabled()
3675
        )
3676

3677
        if is_mixture_of_experts(self.model) and self.parallel_config.enable_eplb:
3678
3679
3680
3681
3682
3683
3684
3685
3686
3687
3688
            logger.info_once("EPLB is enabled for model %s.", self.model_config.model)
            global_expert_load = (
                global_expert_loads[eplb_models] if global_expert_loads else None
            )
            old_global_expert_indices = (
                old_global_expert_indices_per_model[eplb_models]
                if old_global_expert_indices_per_model
                else None
            )
            assert self.eplb_state is not None
            self.eplb_state.add_model(
3689
                self.model,
3690
                self.model_config,
3691
3692
3693
                global_expert_load,
                old_global_expert_indices,
                rank_mapping,
3694
            )
3695
3696
            if self.eplb_state.is_async:
                self.eplb_state.start_async_loop(rank_mapping=rank_mapping)
3697

3698
        if (
3699
3700
            self.vllm_config.compilation_config.mode
            == CompilationMode.STOCK_TORCH_COMPILE
3701
            and supports_dynamo()
3702
        ):
3703
            backend = self.vllm_config.compilation_config.init_backend(self.vllm_config)
3704
            compilation_counter.stock_torch_compile_count += 1
3705
            self.model.compile(fullgraph=True, backend=backend)
3706
            return
3707
        # for other compilation modes, cudagraph behavior is controlled by
3708
3709
3710
        # CudagraphWraper and CudagraphDispatcher of vllm.

        # wrap the model with full cudagraph wrapper if needed.
3711
3712
3713
        cudagraph_mode = self.compilation_config.cudagraph_mode
        assert cudagraph_mode is not None
        if cudagraph_mode.has_full_cudagraphs() and not self.parallel_config.enable_dbo:
3714
3715
3716
            self.model = CUDAGraphWrapper(
                self.model, self.vllm_config, runtime_mode=CUDAGraphMode.FULL
            )
3717
        elif self.parallel_config.enable_dbo:
3718
            if cudagraph_mode.has_full_cudagraphs():
3719
3720
3721
                self.model = UBatchWrapper(
                    self.model, self.vllm_config, CUDAGraphMode.FULL, self.device
                )
3722
            else:
3723
3724
3725
                self.model = UBatchWrapper(
                    self.model, self.vllm_config, CUDAGraphMode.NONE, self.device
                )
3726

3727
    def _get_eagle3_aux_layers_from_config(self) -> tuple[int, ...] | None:
3728
3729
3730
3731
3732
3733
3734
3735
3736
3737
3738
3739
3740
3741
3742
3743
3744
3745
3746
3747
3748
3749
3750
        """Extract Eagle3 auxiliary layer indices from speculative config.

        These indices specify which hidden states from the base model should
        be used as auxiliary inputs for the Eagle3 drafter model during
        speculative decoding.

        Returns:
            Tuple of layer indices if found in draft model config,
            None otherwise.
        """
        if not (self.speculative_config and self.speculative_config.draft_model_config):
            return None

        hf_config = self.speculative_config.draft_model_config.hf_config
        if not hasattr(hf_config, "eagle_aux_hidden_state_layer_ids"):
            return None

        layer_ids = hf_config.eagle_aux_hidden_state_layer_ids
        if layer_ids and isinstance(layer_ids, (list, tuple)):
            return tuple(layer_ids)

        return None

3751
    def reload_weights(self) -> None:
3752
        assert getattr(self, "model", None) is not None, (
3753
            "Cannot reload weights before model is loaded."
3754
        )
3755
3756
        model_loader = get_model_loader(self.load_config)
        logger.info("Reloading weights inplace...")
3757
        model_loader.load_weights(self.get_model(), model_config=self.model_config)
3758

3759
3760
3761
3762
3763
    def save_tensorized_model(
        self,
        tensorizer_config: "TensorizerConfig",
    ) -> None:
        TensorizerLoader.save_model(
3764
            self.get_model(),
3765
            tensorizer_config=tensorizer_config,
3766
            model_config=self.model_config,
3767
3768
        )

3769
3770
3771
    def _get_prompt_logprobs_dict(
        self,
        hidden_states: torch.Tensor,
3772
        num_scheduled_tokens: dict[str, int],
3773
    ) -> dict[str, LogprobsTensors | None]:
3774
        num_prompt_logprobs_dict = self.num_prompt_logprobs
3775
3776
3777
        if not num_prompt_logprobs_dict:
            return {}

3778
        in_progress_dict = self.input_batch.in_progress_prompt_logprobs_cpu
3779
        prompt_logprobs_dict: dict[str, LogprobsTensors | None] = {}
3780
3781
3782
3783
3784

        # 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():
3785
3786
3787
3788
            num_tokens = num_scheduled_tokens.get(req_id)
            if num_tokens is None:
                # This can happen if the request was preempted in prefill stage.
                continue
3789
3790
3791

            # Get metadata for this request.
            request = self.requests[req_id]
3792
3793
3794
3795
            if request.prompt_token_ids is None:
                # Prompt logprobs is incompatible with prompt embeddings
                continue

3796
3797
            num_prompt_tokens = len(request.prompt_token_ids)
            prompt_token_ids = torch.tensor(request.prompt_token_ids).to(
3798
3799
                self.device, non_blocking=True
            )
3800

3801
3802
3803
3804
3805
3806
            # 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(
3807
3808
                    num_prompt_tokens - 1, num_prompt_logprobs + 1
                )
3809
3810
                in_progress_dict[req_id] = logprobs_tensors

3811
            # Determine number of logits to retrieve.
3812
3813
            start_idx = request.num_computed_tokens
            start_tok = start_idx + 1
3814
            num_remaining_tokens = num_prompt_tokens - start_tok
3815
            if num_tokens <= num_remaining_tokens:
3816
                # This is a chunk, more tokens remain.
3817
3818
3819
                # 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.
3820
3821
3822
3823
3824
                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)
3825
3826
3827
3828
3829
3830
3831
                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
3832
3833
3834
3835
3836

            # 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]
3837
            offset = self.query_start_loc.np[req_idx].item()
3838
            prompt_hidden_states = hidden_states[offset : offset + num_logits]
3839
            logits = self.model.compute_logits(prompt_hidden_states)
3840
3841
3842
3843

            # 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.
3844
            tgt_token_ids = prompt_token_ids[start_tok : start_tok + num_logits]
3845
3846

            # Compute prompt logprobs.
3847
3848
            logprobs = self.sampler.compute_logprobs(logits)
            token_ids, logprobs, ranks = self.sampler.gather_logprobs(
3849
3850
                logprobs, num_prompt_logprobs, tgt_token_ids
            )
3851
3852

            # Transfer GPU->CPU async.
3853
3854
            chunk_slice = slice(start_idx, start_idx + num_logits)
            logprobs_tensors.logprob_token_ids[chunk_slice].copy_(
3855
3856
3857
                token_ids, non_blocking=True
            )
            logprobs_tensors.logprobs[chunk_slice].copy_(logprobs, non_blocking=True)
3858
            logprobs_tensors.selected_token_ranks[chunk_slice].copy_(
3859
3860
                ranks, non_blocking=True
            )
3861
3862
3863
3864
3865

        # 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]
3866
            del in_progress_dict[req_id]
3867
3868

        # Must synchronize the non-blocking GPU->CPU transfers.
3869
        if prompt_logprobs_dict:
3870
            self._sync_device()
3871
3872
3873

        return prompt_logprobs_dict

3874
3875
    def _get_nans_in_logits(
        self,
3876
        logits: torch.Tensor | None,
3877
3878
3879
3880
3881
3882
3883
3884
3885
3886
3887
    ) -> dict[str, int]:
        try:
            if logits is None:
                return {req_id: 0 for req_id in self.input_batch.req_ids}

            num_nans_in_logits = {}
            num_nans_for_index = logits.isnan().sum(dim=-1).cpu().numpy()
            for req_id in self.input_batch.req_ids:
                req_index = self.input_batch.req_id_to_index[req_id]
                num_nans_in_logits[req_id] = (
                    int(num_nans_for_index[req_index])
3888
3889
3890
                    if num_nans_for_index is not None and req_index < logits.shape[0]
                    else 0
                )
3891
3892
3893
3894
            return num_nans_in_logits
        except IndexError:
            return {}

3895
    @contextmanager
3896
3897
3898
    def maybe_randomize_inputs(
        self, input_ids: torch.Tensor | None, inputs_embeds: torch.Tensor | None
    ):
3899
3900
3901
3902
        """
        Randomize input_ids if VLLM_RANDOMIZE_DP_DUMMY_INPUTS is set.
        This is to help balance expert-selection
         - during profile_run
3903
         - during DP rank dummy run
3904
        """
3905

3906
3907
3908
3909
        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
3910
        elif input_ids is not None:
3911
3912
3913
3914

            @functools.cache
            def rand_input_ids() -> torch.Tensor:
                return torch.randint_like(
3915
                    self.input_ids.gpu,
3916
3917
                    low=0,
                    high=self.model_config.get_vocab_size(),
3918
                )
3919

3920
            logger.debug_once("Randomizing dummy input_ids for DP Rank")
3921
            input_ids.copy_(rand_input_ids()[: input_ids.size(0)], non_blocking=True)
3922
3923
            yield
            input_ids.fill_(0)
3924
3925
3926
3927
3928
3929
3930
3931
3932
3933
3934
3935
3936
3937
3938
        else:

            @functools.cache
            def rand_inputs_embeds() -> torch.Tensor:
                return torch.randn_like(
                    self.inputs_embeds.gpu,
                )

            assert inputs_embeds is not None
            logger.debug_once("Randomizing dummy inputs_embeds for DP Rank")
            inputs_embeds.copy_(
                rand_inputs_embeds()[: inputs_embeds.size(0)], non_blocking=True
            )
            yield
            inputs_embeds.fill_(0)
3939

3940
3941
3942
3943
3944
3945
    def _get_mm_dummy_batch(
        self,
        modality: str,
        max_items_per_batch: int,
    ) -> BatchedTensorInputs:
        """Dummy data for profiling and precompiling multimodal models."""
3946
3947
        assert self.mm_budget is not None

3948
        dummy_decoder_data = self.mm_registry.get_decoder_dummy_data(
3949
            model_config=self.model_config,
3950
            seq_len=self.max_model_len,
3951
            mm_counts={modality: 1},
3952
            cache=self.mm_budget.cache,
3953
3954
3955
3956
        )
        dummy_mm_data = dummy_decoder_data.multi_modal_data

        # Result in the maximum GPU consumption of the model
3957
3958
        dummy_mm_item = dummy_mm_data[modality][0]
        dummy_mm_items = [dummy_mm_item] * max_items_per_batch
3959

3960
3961
3962
3963
3964
3965
3966
3967
        return next(
            mm_kwargs_group
            for _, _, mm_kwargs_group in group_mm_kwargs_by_modality(
                dummy_mm_items,
                device=self.device,
                pin_memory=self.pin_memory,
            )
        )
3968

3969
3970
3971
3972
    @torch.inference_mode()
    def _dummy_run(
        self,
        num_tokens: int,
3973
        cudagraph_runtime_mode: CUDAGraphMode | None = None,
3974
3975
        force_attention: bool = False,
        uniform_decode: bool = False,
3976
        allow_microbatching: bool = True,
3977
3978
        skip_eplb: bool = False,
        is_profile: bool = False,
3979
        create_mixed_batch: bool = False,
3980
        remove_lora: bool = True,
3981
        activate_lora: bool = False,
Rémi Delacourt's avatar
Rémi Delacourt committed
3982
        is_graph_capturing: bool = False,
3983
    ) -> tuple[torch.Tensor, torch.Tensor]:
3984
3985
3986
3987
3988
3989
3990
        """
        Run a dummy forward pass to warm up/profile run or capture the
        CUDA graph for the model.

        Args:
            num_tokens: Number of tokens to run the dummy forward pass.
            cudagraph_runtime_mode: used to control the behavior.
3991
                - if not set will determine the cudagraph mode based on using
3992
                    the self.cudagraph_dispatcher.
3993
3994
3995
3996
                - CUDAGraphMode.NONE: No cudagraph, for warm up and profile run
                - CUDAGraphMode.PIECEWISE: Piecewise cudagraph.
                - CUDAGraphMode.FULL: Full cudagraph, attention metadata is
                    needed.
3997
            force_attention: If True, always create attention metadata. Used to
3998
3999
4000
4001
                warm up attention backend when mode is NONE.
            uniform_decode: If True, the batch is a uniform decode batch.
            skip_eplb: If True, skip EPLB state update.
            is_profile: If True, this is a profile run.
4002
4003
            create_mixed_batch: If True, create a mixed batch with both decode
                (1 token) and prefill (multiple tokens) requests.
4004
            remove_lora: If False, dummy LoRAs are not destroyed after the run
4005
            activate_lora: If False, dummy_run is performed without LoRAs.
4006
        """
4007
4008
4009
4010
        assert (
            cudagraph_runtime_mode is None
            or cudagraph_runtime_mode.valid_runtime_modes()
        )
4011

4012
        # If cudagraph_mode.decode_mode() == FULL and
4013
        # cudagraph_mode.separate_routine(). This means that we are using
4014
4015
4016
4017
4018
4019
4020
4021
4022
4023
4024
        # different graphs and/or modes for mixed prefill-decode batches vs.
        # uniform decode batches. A uniform decode batch means that all
        # requests have identical query length, except a potential virtual
        # request (shorter) in the batch account for padding.
        # Uniform decode batch could either be common pure decode, where
        # max_query_len == 1, or speculative decode, where
        # max_query_len == 1 + num_spec_decode_tokens.

        # When setting max_query_len = 1, we switch to and capture the optimized
        # routine of FA2 for pure decode, i.e., Flashdecode + an optimization
        # for GQA/MQA.
4025
        max_query_len = self.uniform_decode_query_len if uniform_decode else num_tokens
4026

4027
4028
4029
4030
4031
        # 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
4032
4033
4034
4035
        if create_mixed_batch:
            assert not uniform_decode
            # Create mixed batch:
            # first half decode tokens, second half one prefill
4036
            num_decode_tokens = min(max_num_reqs - 1, num_tokens // 2)
4037
4038
4039
4040
            num_prefill_tokens = num_tokens - num_decode_tokens
            num_reqs = num_decode_tokens + 1

            # Create decode requests (1 token each) followed by prefill request
4041
            num_scheduled_tokens_list = [1] * num_decode_tokens + [num_prefill_tokens]
4042
4043
4044
            # Note: Overriding max_query_len to be the prefill tokens
            max_query_len = num_prefill_tokens
        elif uniform_decode:
4045
            assert not create_mixed_batch
4046
            num_reqs = min(max_num_reqs, cdiv(num_tokens, max_query_len))
4047
4048
            num_scheduled_tokens_list = [max_query_len] * num_reqs
            if num_tokens % max_query_len != 0:
4049
                num_scheduled_tokens_list[-1] = num_tokens % max_query_len
4050
4051
4052
4053
4054
4055
        else:
            num_reqs = min(num_tokens, max_num_reqs)
            min_tokens_per_req = num_tokens // num_reqs
            num_scheduled_tokens_list = [min_tokens_per_req] * num_reqs
            num_scheduled_tokens_list[-1] += num_tokens % num_reqs

4056
4057
        assert sum(num_scheduled_tokens_list) == num_tokens
        assert len(num_scheduled_tokens_list) == num_reqs
4058
        num_scheduled_tokens = np.array(num_scheduled_tokens_list, dtype=np.int32)
4059
4060
        num_tokens_unpadded = int(num_scheduled_tokens.sum())

4061
        num_sampled_tokens = np.ones(num_reqs, dtype=np.int32)
4062

4063
        _cudagraph_mode, batch_desc, should_ubatch, num_tokens_across_dp, _ = (
4064
4065
4066
4067
4068
4069
4070
4071
4072
4073
4074
4075
4076
4077
4078
4079
4080
4081
            self._determine_batch_execution_and_padding(
                num_tokens=num_tokens_unpadded,
                num_reqs=num_reqs,
                num_scheduled_tokens_np=num_scheduled_tokens,
                max_num_scheduled_tokens=max_query_len,
                use_cascade_attn=False,
                allow_microbatching=allow_microbatching,
                force_eager=is_profile
                or (cudagraph_runtime_mode == CUDAGraphMode.NONE),
                # `force_uniform_decode` is used for cudagraph capture; because for
                # capturing mixed prefill-decode batches, we sometimes use
                # num_tokens == num_reqs which looks like a uniform decode batch to the
                # dispatcher; but we actually want to capture a piecewise cudagraph
                force_uniform_decode=uniform_decode,
                # `force_has_lora` is used for cudagraph capture; because LoRA is
                # activated later in the context manager, but we need to know the
                # LoRA state when determining the batch descriptor for capture
                force_has_lora=activate_lora,
4082
4083
            )
        )
4084
4085
4086

        if cudagraph_runtime_mode is None:
            cudagraph_runtime_mode = _cudagraph_mode
4087
        else:
4088
4089
4090
4091
4092
4093
4094
4095
4096
            assert cudagraph_runtime_mode == _cudagraph_mode, (
                f"Cudagraph runtime mode mismatch in dummy_run. "
                f"Expected {_cudagraph_mode}, but got {cudagraph_runtime_mode}."
            )

        num_tokens_padded = batch_desc.num_tokens
        num_reqs_padded = (
            batch_desc.num_reqs if batch_desc.num_reqs is not None else num_reqs
        )
4097
4098
4099
        ubatch_slices, ubatch_slices_padded = maybe_create_ubatch_slices(
            should_ubatch, num_scheduled_tokens, num_tokens_padded, num_reqs_padded
        )
4100

4101
        attn_metadata: PerLayerAttnMetadata | None = None
4102
4103
4104

        # If force_attention is True, we always capture attention. Otherwise,
        # it only happens for cudagraph_runtime_mode=FULL.
4105
        if force_attention or cudagraph_runtime_mode == CUDAGraphMode.FULL:
4106
4107
4108
4109
4110
4111
            if create_mixed_batch:
                # In the mixed batch mode (used for FI warmup), we use
                # shorter sequence lengths to run faster.
                # TODO(luka) better system for describing dummy batches
                seq_lens = [1] * num_decode_tokens + [num_prefill_tokens + 1]
            else:
4112
                seq_lens = max_query_len  # type: ignore[assignment]
4113
            self.seq_lens.np[:num_reqs] = seq_lens
4114
4115
            self.seq_lens.np[num_reqs:] = 0
            self.seq_lens.copy_to_gpu()
4116

4117
4118
            cum_num_tokens, _ = self._get_cumsum_and_arange(num_scheduled_tokens)
            self.query_start_loc.np[1 : num_reqs + 1] = cum_num_tokens
4119
4120
            self.query_start_loc.copy_to_gpu()

4121
            pad_attn = cudagraph_runtime_mode == CUDAGraphMode.FULL
4122
            attn_metadata, _ = self._build_attention_metadata(
4123
4124
4125
                num_tokens=num_tokens_unpadded,
                num_reqs=num_reqs_padded,
                max_query_len=max_query_len,
4126
                ubatch_slices=ubatch_slices_padded if pad_attn else ubatch_slices,
4127
                for_cudagraph_capture=is_graph_capturing,
4128
            )
4129

4130
        with self.maybe_dummy_run_with_lora(
4131
4132
4133
4134
4135
            self.lora_config,
            num_scheduled_tokens,
            num_sampled_tokens,
            activate_lora,
            remove_lora,
4136
        ):
4137
            # Make sure padding doesn't exceed max_num_tokens
4138
4139
            assert num_tokens_padded <= self.max_num_tokens
            model_kwargs = self._init_model_kwargs(num_tokens_padded)
4140
            if self.supports_mm_inputs and not self.model_config.is_encoder_decoder:
4141
                input_ids = None
4142
                inputs_embeds = self.inputs_embeds.gpu[:num_tokens_padded]
4143
                model_kwargs = {
4144
                    **model_kwargs,
4145
4146
                    **self._dummy_mm_kwargs(num_reqs),
                }
4147
4148
            elif self.enable_prompt_embeds:
                input_ids = None
4149
4150
                inputs_embeds = self.inputs_embeds.gpu[:num_tokens_padded]
                model_kwargs = self._init_model_kwargs(num_tokens_padded)
4151
            else:
4152
                input_ids = self.input_ids.gpu[:num_tokens_padded]
4153
                inputs_embeds = None
4154

4155
            if self.uses_mrope:
4156
                positions = self.mrope_positions.gpu[:, :num_tokens_padded]
4157
            elif self.uses_xdrope_dim > 0:
4158
                positions = self.xdrope_positions.gpu[:, :num_tokens_padded]
4159
            else:
4160
                positions = self.positions.gpu[:num_tokens_padded]
4161
4162
4163
4164
4165
4166
4167
4168
4169

            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,
4170
4171
4172
                            device=self.device,
                        )
                    )
4173
4174

                intermediate_tensors = self.sync_and_slice_intermediate_tensors(
4175
                    num_tokens_padded, None, False
4176
                )
4177

4178
            if ubatch_slices_padded is not None:
4179
4180
4181
                # Adjust values to reflect a single ubatch.
                # TODO(sage,lucas): this is cruft that should be addressed in
                #  the padding refactor.
4182
                num_tokens_padded = ubatch_slices_padded[0].num_tokens
4183
                if num_tokens_across_dp is not None:
4184
                    num_tokens_across_dp[:] = num_tokens_padded
4185

4186
            with (
4187
                self.maybe_randomize_inputs(input_ids, inputs_embeds),
4188
                set_forward_context(
4189
4190
                    attn_metadata,
                    self.vllm_config,
4191
                    num_tokens=num_tokens_padded,
4192
4193
                    num_tokens_across_dp=num_tokens_across_dp,
                    cudagraph_runtime_mode=cudagraph_runtime_mode,
4194
                    batch_descriptor=batch_desc,
4195
                    ubatch_slices=ubatch_slices_padded,
4196
4197
                ),
            ):
4198
                outputs = self.model(
4199
4200
4201
4202
                    input_ids=input_ids,
                    positions=positions,
                    intermediate_tensors=intermediate_tensors,
                    inputs_embeds=inputs_embeds,
4203
                    **model_kwargs,
4204
                )
4205

4206
4207
4208
4209
            if self.use_aux_hidden_state_outputs:
                hidden_states, _ = outputs
            else:
                hidden_states = outputs
4210

4211
            if self.speculative_config and self.speculative_config.use_eagle():
4212
                assert isinstance(self.drafter, EagleProposer)
4213
4214
4215
                # Eagle currently only supports PIECEWISE cudagraphs.
                # Therefore only use cudagraphs if the main model uses PIECEWISE
                # NOTE(lucas): this is a hack, need to clean up.
4216
                use_cudagraphs = (
4217
4218
4219
4220
4221
4222
4223
4224
4225
                    (
                        is_graph_capturing
                        and cudagraph_runtime_mode == CUDAGraphMode.PIECEWISE
                    )
                    or (
                        not is_graph_capturing
                        and cudagraph_runtime_mode != CUDAGraphMode.NONE
                    )
                ) and not self.speculative_config.enforce_eager
4226
4227
4228
4229
4230
4231
4232
4233
4234
4235
4236

                # Note(gnovack) - We need to disable cudagraphs for one of the two
                # lora cases when cudagraph_specialize_lora is enabled. This is a
                # short term mitigation for issue mentioned in
                # https://github.com/vllm-project/vllm/issues/28334
                if self.compilation_config.cudagraph_specialize_lora and activate_lora:
                    use_cudagraphs = False

                self.drafter.dummy_run(
                    num_tokens,
                    use_cudagraphs=use_cudagraphs,
Rémi Delacourt's avatar
Rémi Delacourt committed
4237
                    is_graph_capturing=is_graph_capturing,
4238
                )
4239

4240
4241
4242
4243
4244
4245
4246
4247
4248
4249
4250
        # We register layerwise NVTX hooks here after the first dynamo tracing is
        # done to avoid nvtx operations in hook functions being traced by
        # torch dynamo and causing graph breaks.
        # Note that for DYNAMO_ONCE and VLLM_COMPILE mode,
        # compiled model's dynamo tracing is only done once and the compiled model's
        # __call__ function is replaced by calling the compiled function.
        # So it's safe to register hooks here. Hooks will be registered to
        # both compiled and uncompiled models but they will never
        # be called on the compiled model execution path.
        self._register_layerwise_nvtx_hooks()

4251
4252
4253
4254
4255
4256
4257
4258
4259
4260
        # This is necessary to avoid blocking DP.
        # For dummy runs, we typically skip EPLB since we don't have any real
        # requests to process.
        # However, in DP settings, there may be cases when some DP ranks do
        # not have any requests to process, so they're executing dummy batches.
        # In such cases, we still have to trigger EPLB to make sure
        # ranks execute the rearrangement in synchronization.
        if not skip_eplb:
            self.eplb_step(is_dummy=True, is_profile=is_profile)

4261
        logit_indices = np.cumsum(num_scheduled_tokens) - 1
4262
4263
4264
4265
        logit_indices_device = torch.from_numpy(logit_indices).to(
            self.device, non_blocking=True
        )
        return hidden_states, hidden_states[logit_indices_device]
4266
4267
4268
4269
4270
4271

    @torch.inference_mode()
    def _dummy_sampler_run(
        self,
        hidden_states: torch.Tensor,
    ) -> torch.Tensor:
4272
4273
4274
4275
        # 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)
4276

4277
        logits = self.model.compute_logits(hidden_states)
4278
4279
        num_reqs = logits.size(0)

4280
        dummy_tensors = lambda v: torch.full((num_reqs,), v, device=self.device)
4281
4282
4283
4284
4285
4286
4287
4288
4289
4290
4291
4292
4293
4294
4295

        dummy_metadata = SamplingMetadata(
            temperature=dummy_tensors(0.5),
            all_greedy=False,
            all_random=False,
            top_p=dummy_tensors(0.9),
            top_k=dummy_tensors(logits.size(1) - 1),
            generators={},
            max_num_logprobs=None,
            no_penalties=True,
            prompt_token_ids=None,
            frequency_penalties=dummy_tensors(0.1),
            presence_penalties=dummy_tensors(0.1),
            repetition_penalties=dummy_tensors(0.1),
            output_token_ids=[[] for _ in range(num_reqs)],
4296
            spec_token_ids=[[] for _ in range(num_reqs)],
4297
4298
            allowed_token_ids_mask=None,
            bad_words_token_ids={},
4299
            logitsprocs=LogitsProcessors(),
4300
        )
4301
        try:
4302
4303
4304
            sampler_output = self.sampler(
                logits=logits, sampling_metadata=dummy_metadata
            )
4305
        except RuntimeError as e:
4306
            if "out of memory" in str(e):
4307
4308
4309
4310
                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 "
4311
4312
                    "initializing the engine."
                ) from e
4313
4314
            else:
                raise e
4315
        if self.speculative_config:
4316
4317
            draft_token_ids = [[0] for _ in range(num_reqs)]
            dummy_spec_decode_metadata = SpecDecodeMetadata.make_dummy(
4318
4319
                draft_token_ids, self.device
            )
4320
4321
4322
4323
4324
4325

            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
4326
4327
4328
4329
4330
            logits = torch.randn(
                num_tokens + num_reqs,
                logits.shape[-1],
                device=self.device,
                dtype=logits.dtype,
4331
            )
4332
4333
4334
            self.rejection_sampler(
                dummy_spec_decode_metadata,
                draft_probs,
4335
                logits,
4336
4337
                dummy_metadata,
            )
4338
        return sampler_output
4339

4340
    def _dummy_pooler_run_task(
4341
4342
        self,
        hidden_states: torch.Tensor,
4343
4344
        task: PoolingTask,
    ) -> PoolerOutput:
4345
4346
4347
4348
4349
4350
4351
4352
4353
4354
4355
        num_tokens = hidden_states.shape[0]
        max_num_reqs = self.scheduler_config.max_num_seqs
        num_reqs = min(num_tokens, max_num_reqs)
        min_tokens_per_req = num_tokens // num_reqs
        num_scheduled_tokens_list = [min_tokens_per_req] * num_reqs
        num_scheduled_tokens_list[-1] += num_tokens % num_reqs
        assert sum(num_scheduled_tokens_list) == num_tokens
        assert len(num_scheduled_tokens_list) == num_reqs

        req_num_tokens = num_tokens // num_reqs

4356
        dummy_prompt_lens = torch.tensor(
4357
4358
            num_scheduled_tokens_list,
            device="cpu",
4359
        )
4360
4361
4362
        dummy_token_ids = torch.zeros(
            (num_reqs, req_num_tokens), dtype=torch.int32, device=self.device
        )
4363

4364
        model = cast(VllmModelForPooling, self.get_model())
4365
        dummy_pooling_params = PoolingParams(task=task)
4366
        dummy_pooling_params.verify(task=task, model_config=self.model_config)
4367
        to_update = model.pooler.get_pooling_updates(task)
4368
4369
        to_update.apply(dummy_pooling_params)

4370
        dummy_metadata = PoolingMetadata(
4371
4372
4373
            prompt_lens=dummy_prompt_lens,
            prompt_token_ids=dummy_token_ids,
            pooling_params=[dummy_pooling_params] * num_reqs,
4374
            pooling_states=[PoolingStates() for i in range(num_reqs)],
4375
        )
4376

4377
        dummy_metadata.build_pooling_cursor(
4378
4379
4380
            num_scheduled_tokens_list,
            seq_lens_cpu=dummy_prompt_lens,
            device=hidden_states.device,
4381
        )
4382

4383
        try:
4384
4385
4386
            return model.pooler(
                hidden_states=hidden_states, pooling_metadata=dummy_metadata
            )
4387
        except RuntimeError as e:
4388
            if "out of memory" in str(e):
4389
                raise RuntimeError(
4390
4391
4392
                    "CUDA out of memory occurred when warming up pooler "
                    f"({task=}) with {num_reqs} dummy requests. Please try "
                    "lowering `max_num_seqs` or `gpu_memory_utilization` when "
4393
4394
                    "initializing the engine."
                ) from e
4395
4396
            else:
                raise e
4397
4398
4399
4400
4401
4402
4403

    @torch.inference_mode()
    def _dummy_pooler_run(
        self,
        hidden_states: torch.Tensor,
    ) -> PoolerOutput:
        # Find the task that has the largest output for subsequent steps
4404
4405
4406
        supported_pooling_tasks = self.get_supported_pooling_tasks()

        if not supported_pooling_tasks:
4407
4408
4409
4410
4411
4412
            raise RuntimeError(
                f"Model {self.model_config.model} does not support "
                "any pooling tasks. See "
                "https://docs.vllm.ai/en/latest/models/pooling_models.html "
                "to learn more."
            )
4413

4414
        output_size = dict[PoolingTask, float]()
4415
        for task in supported_pooling_tasks:
4416
4417
            # Run a full batch with each task to ensure none of them OOMs
            output = self._dummy_pooler_run_task(hidden_states, task)
4418
            output_size[task] = sum(o.nbytes for o in output)
4419
4420
4421
4422
            del output  # Allow GC

        max_task = max(output_size.items(), key=lambda x: x[1])[0]
        return self._dummy_pooler_run_task(hidden_states, max_task)
4423

4424
    def profile_run(self) -> None:
4425
        # Profile with multimodal encoder & encoder cache.
4426
        if self.supports_mm_inputs:
4427
4428
            mm_config = self.model_config.multimodal_config
            if mm_config is not None and mm_config.skip_mm_profiling:
4429
                logger.info(
4430
                    "Skipping memory profiling for multimodal encoder and "
4431
4432
                    "encoder cache."
                )
4433
4434
4435
4436
4437
4438
4439
4440
            else:
                mm_budget = self.mm_budget
                assert mm_budget is not None

                if (encoder_budget := mm_budget.get_encoder_budget()) > 0:
                    # NOTE: Currently model is profiled with a single non-text
                    # modality with the max possible input tokens even when
                    # it supports multiple.
4441
                    dummy_modality = mm_budget.get_modality_with_max_tokens()
4442
4443
4444
                    max_mm_items_per_batch = mm_budget.max_items_per_batch_by_modality[
                        dummy_modality
                    ]
4445
4446
4447
4448
4449
4450
4451
4452
4453

                    logger.info(
                        "Encoder cache will be initialized with a budget of "
                        "%s tokens, and profiled with %s %s items of the "
                        "maximum feature size.",
                        encoder_budget,
                        max_mm_items_per_batch,
                        dummy_modality,
                    )
4454

4455
4456
4457
4458
4459
                    # Create dummy batch of multimodal inputs.
                    batched_dummy_mm_inputs = self._get_mm_dummy_batch(
                        dummy_modality,
                        max_mm_items_per_batch,
                    )
4460

4461
                    # Run multimodal encoder.
4462
                    dummy_encoder_outputs = self.model.embed_multimodal(
4463
4464
                        **batched_dummy_mm_inputs
                    )
4465

4466
4467
4468
4469
                    sanity_check_mm_encoder_outputs(
                        dummy_encoder_outputs,
                        expected_num_items=max_mm_items_per_batch,
                    )
4470

4471
4472
4473
                    # NOTE: This happens when encoder cache needs to store
                    # the embeddings that encoder outputs are scattered onto.
                    # In this case we create dummy embeddings of size
4474
4475
                    # (max_tokens_for_modality, hidden_size) and scatter
                    # encoder output into it.
4476
                    encoder_output_shape = dummy_encoder_outputs[0].shape
4477
4478
4479
4480
4481
                    max_mm_tokens_per_item = mm_budget.max_tokens_by_modality[
                        dummy_modality
                    ]
                    if encoder_output_shape[0] < max_mm_tokens_per_item:
                        encoder_hidden_size = encoder_output_shape[-1]
4482
4483
4484
                        expanded_outputs = []
                        for output in dummy_encoder_outputs:
                            expanded = output.new_zeros(
4485
                                (max_mm_tokens_per_item, encoder_hidden_size)
4486
                            )
4487
4488
4489
4490
4491
4492
                            num_tokens = output.shape[0]
                            expanded[:num_tokens].copy_(output)
                            expanded_outputs.append(expanded)

                        dummy_encoder_outputs = expanded_outputs

4493
                    # Cache the dummy encoder outputs.
4494
                    self.encoder_cache["tmp"] = dict(enumerate(dummy_encoder_outputs))
4495

4496
        # Add `is_profile` here to pre-allocate communication buffers
4497
4498
4499
        hidden_states, last_hidden_states = self._dummy_run(
            self.max_num_tokens, is_profile=True
        )
4500
        if get_pp_group().is_last_rank:
4501
4502
4503
4504
            if self.is_pooling_model:
                output = self._dummy_pooler_run(hidden_states)
            else:
                output = self._dummy_sampler_run(last_hidden_states)
4505
        else:
4506
            output = None
4507
        self._sync_device()
4508
        del hidden_states, output
4509
        self.encoder_cache.clear()
4510
        gc.collect()
4511

4512
    def capture_model(self) -> int:
4513
        if self.compilation_config.cudagraph_mode == CUDAGraphMode.NONE:
4514
            logger.warning(
4515
                "Skipping CUDA graph capture. To turn on CUDA graph capture, "
4516
4517
                "ensure `cudagraph_mode` was not manually set to `NONE`"
            )
4518
            return 0
4519

4520
4521
        compilation_counter.num_gpu_runner_capture_triggers += 1

4522
4523
        start_time = time.perf_counter()

4524
4525
4526
4527
4528
4529
4530
4531
4532
4533
4534
4535
4536
4537
        @contextmanager
        def freeze_gc():
            # Optimize garbage collection during CUDA graph capture.
            # Clean up, then freeze all remaining objects from being included
            # in future collections.
            gc.collect()
            should_freeze = not envs.VLLM_ENABLE_CUDAGRAPH_GC
            if should_freeze:
                gc.freeze()
            try:
                yield
            finally:
                if should_freeze:
                    gc.unfreeze()
4538
                    gc.collect()
4539

4540
4541
4542
        # 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.
4543
        set_cudagraph_capturing_enabled(True)
4544
        with freeze_gc(), graph_capture(device=self.device):
4545
            start_free_gpu_memory = torch.cuda.mem_get_info()[0]
4546
            cudagraph_mode = self.compilation_config.cudagraph_mode
4547
            assert cudagraph_mode is not None
4548
4549
4550
4551
4552
4553
4554
4555
4556

            if self.lora_config:
                if self.compilation_config.cudagraph_specialize_lora:
                    lora_cases = [True, False]
                else:
                    lora_cases = [True]
            else:
                lora_cases = [False]

4557
4558
            if cudagraph_mode.mixed_mode() != CUDAGraphMode.NONE:
                cudagraph_runtime_mode = cudagraph_mode.mixed_mode()
4559
                # make sure we capture the largest batch size first
4560
4561
4562
                compilation_cases = list(
                    product(reversed(self.cudagraph_batch_sizes), lora_cases)
                )
4563
4564
4565
                self._capture_cudagraphs(
                    compilation_cases,
                    cudagraph_runtime_mode=cudagraph_runtime_mode,
4566
4567
                    uniform_decode=False,
                )
4568

4569
4570
            # Capture full cudagraph for uniform decode batches if we
            # don't already have full mixed prefill-decode cudagraphs.
4571
4572
4573
4574
4575
4576
4577
            if (
                cudagraph_mode.decode_mode() == CUDAGraphMode.FULL
                and cudagraph_mode.separate_routine()
            ):
                max_num_tokens = (
                    self.scheduler_config.max_num_seqs * self.uniform_decode_query_len
                )
4578
                decode_cudagraph_batch_sizes = [
4579
4580
                    x
                    for x in self.cudagraph_batch_sizes
4581
                    if max_num_tokens >= x >= self.uniform_decode_query_len
4582
                ]
4583
4584
4585
                compilation_cases_decode = list(
                    product(reversed(decode_cudagraph_batch_sizes), lora_cases)
                )
4586
4587
4588
                self._capture_cudagraphs(
                    compilation_cases=compilation_cases_decode,
                    cudagraph_runtime_mode=CUDAGraphMode.FULL,
4589
4590
                    uniform_decode=True,
                )
4591

4592
4593
4594
            torch.cuda.synchronize()
            end_free_gpu_memory = torch.cuda.mem_get_info()[0]

4595
4596
4597
        # Disable cudagraph capturing globally, so any unexpected cudagraph
        # capturing will be detected and raise an error after here.
        # Note: We don't put it into graph_capture context manager because
4598
        # we may do lazy capturing in future that still allows capturing
4599
4600
        # after here.
        set_cudagraph_capturing_enabled(False)
4601

4602
4603
4604
4605
        # Lock workspace to prevent resizing during execution.
        # Max workspace sizes should have been captured during warmup/profiling.
        lock_workspace()

4606
4607
4608
4609
        end_time = time.perf_counter()
        elapsed_time = end_time - start_time
        cuda_graph_size = start_free_gpu_memory - end_free_gpu_memory
        # This usually takes 5~20 seconds.
4610
        logger.info_once(
4611
4612
4613
            "Graph capturing finished in %.0f secs, took %.2f GiB",
            elapsed_time,
            cuda_graph_size / (1 << 30),
4614
            scope="local",
4615
        )
4616
        return cuda_graph_size
4617

4618
4619
    def _capture_cudagraphs(
        self,
4620
        compilation_cases: list[tuple[int, bool]],
4621
4622
4623
4624
4625
4626
4627
        cudagraph_runtime_mode: CUDAGraphMode,
        uniform_decode: bool,
    ):
        assert (
            cudagraph_runtime_mode != CUDAGraphMode.NONE
            and cudagraph_runtime_mode.valid_runtime_modes()
        ), f"Invalid cudagraph runtime mode: {cudagraph_runtime_mode}"
4628
4629
4630
4631
4632
4633
4634
4635

        # Only rank 0 should print progress bar during capture
        if is_global_first_rank():
            compilation_cases = tqdm(
                compilation_cases,
                disable=not self.load_config.use_tqdm_on_load,
                desc="Capturing CUDA graphs ({}, {})".format(
                    "decode" if uniform_decode else "mixed prefill-decode",
4636
4637
4638
                    cudagraph_runtime_mode.name,
                ),
            )
4639

4640
        # We skip EPLB here since we don't want to record dummy metrics
4641
        for num_tokens, activate_lora in compilation_cases:
4642
            # We currently only capture ubatched graphs when its a FULL
4643
4644
4645
            # cudagraph, a uniform decode batch, and the number of tokens
            # is above the threshold. Otherwise we just capture a non-ubatched
            # version of the graph
4646
4647
4648
4649
            allow_microbatching = (
                self.parallel_config.enable_dbo
                and cudagraph_runtime_mode == CUDAGraphMode.FULL
                and uniform_decode
4650
4651
4652
4653
4654
                and check_ubatch_thresholds(
                    config=self.vllm_config.parallel_config,
                    num_tokens=num_tokens,
                    uniform_decode=uniform_decode,
                )
4655
            )
4656

4657
4658
4659
4660
4661
4662
            for _ in range(self.compilation_config.cudagraph_num_of_warmups):
                # Use CUDAGraphRuntimeStyle.NONE (default) for warmup.
                # But be careful, warm up with `NONE`is orthogonal to
                # if we want to warm up attention or not. This is
                # different from the case where `FULL` implies capture
                # attention while `PIECEWISE` implies no attention.
4663
4664
4665
4666
4667
4668
4669
4670
4671
                force_attention = cudagraph_runtime_mode == CUDAGraphMode.FULL
                self._dummy_run(
                    num_tokens,
                    cudagraph_runtime_mode=CUDAGraphMode.NONE,
                    force_attention=force_attention,
                    uniform_decode=uniform_decode,
                    allow_microbatching=allow_microbatching,
                    skip_eplb=True,
                    remove_lora=False,
4672
                    activate_lora=activate_lora,
4673
4674
4675
4676
4677
4678
4679
4680
                )
            self._dummy_run(
                num_tokens,
                cudagraph_runtime_mode=cudagraph_runtime_mode,
                uniform_decode=uniform_decode,
                allow_microbatching=allow_microbatching,
                skip_eplb=True,
                remove_lora=False,
4681
                activate_lora=activate_lora,
Rémi Delacourt's avatar
Rémi Delacourt committed
4682
                is_graph_capturing=True,
4683
            )
4684
        self.maybe_remove_all_loras(self.lora_config)
4685

4686
4687
4688
4689
    def initialize_attn_backend(self, kv_cache_config: KVCacheConfig) -> None:
        """
        Initialize the attention backends and attention metadata builders.
        """
4690
        assert len(self.attn_groups) == 0, "Attention backends are already initialized"
4691

4692
4693
4694
4695
4696
4697
        class AttentionGroupKey(NamedTuple):
            attn_backend: type[AttentionBackend]
            kv_cache_spec: KVCacheSpec

        def get_attn_backends_for_group(
            kv_cache_group_spec: KVCacheGroupSpec,
4698
        ) -> tuple[dict[AttentionGroupKey, list[str]], set[type[AttentionBackend]]]:
4699
            layer_type = cast(type[Any], AttentionLayerBase)
4700
            layers = get_layers_from_vllm_config(
4701
                self.vllm_config, layer_type, kv_cache_group_spec.layer_names
4702
            )
4703
4704
            attn_backends = {}
            attn_backend_layers = defaultdict(list)
4705
            # Dedupe based on full class name; this is a bit safer than
4706
4707
4708
4709
            # using the class itself as the key because when we create dynamic
            # attention backend subclasses (e.g. ChunkedLocalAttention) unless
            # they are cached correctly, there will be different objects per
            # layer.
4710
            for layer_name in kv_cache_group_spec.layer_names:
4711
                attn_backend = layers[layer_name].get_attn_backend()
4712
4713
4714
4715

                if layer_name in self.kv_sharing_fast_prefill_eligible_layers:
                    attn_backend = create_fast_prefill_custom_backend(
                        "FastPrefill",
4716
                        attn_backend,  # type: ignore[arg-type]
4717
4718
                    )

4719
4720
4721
                full_cls_name = attn_backend.full_cls_name()
                layer_kv_cache_spec = kv_cache_group_spec.kv_cache_spec
                if isinstance(layer_kv_cache_spec, UniformTypeKVCacheSpecs):
4722
                    layer_kv_cache_spec = layer_kv_cache_spec.kv_cache_specs[layer_name]
4723
                key = (full_cls_name, layer_kv_cache_spec)
4724
4725
4726
                attn_backends[key] = AttentionGroupKey(
                    attn_backend, layer_kv_cache_spec
                )
4727
                attn_backend_layers[key].append(layer_name)
4728
4729
4730
4731
            return (
                {attn_backends[k]: v for k, v in attn_backend_layers.items()},
                set(group_key.attn_backend for group_key in attn_backends.values()),
            )
4732
4733

        def create_attn_groups(
4734
            attn_backends_map: dict[AttentionGroupKey, list[str]],
4735
            kv_cache_group_id: int,
4736
4737
        ) -> list[AttentionGroup]:
            attn_groups: list[AttentionGroup] = []
4738
            for (attn_backend, kv_cache_spec), layer_names in attn_backends_map.items():
4739
                attn_group = AttentionGroup(
4740
                    attn_backend,
4741
                    layer_names,
4742
                    kv_cache_spec,
4743
                    kv_cache_group_id,
4744
4745
                )

4746
4747
4748
                attn_groups.append(attn_group)
            return attn_groups

4749
        attention_backend_maps = []
4750
        attention_backend_list = []
4751
        for kv_cache_group_spec in kv_cache_config.kv_cache_groups:
4752
            attn_backends = get_attn_backends_for_group(kv_cache_group_spec)
4753
            attention_backend_maps.append(attn_backends[0])
4754
            attention_backend_list.append(attn_backends[1])
4755
4756

        # Resolve cudagraph_mode before actually initialize metadata_builders
4757
4758
4759
        self._check_and_update_cudagraph_mode(
            attention_backend_list, kv_cache_config.kv_cache_groups
        )
4760

4761
4762
4763
        # Check if attention backend supports PCP&DCP and related features.
        check_attention_cp_compatibility(self.vllm_config)

4764
4765
        for i, attn_backend_map in enumerate(attention_backend_maps):
            self.attn_groups.append(create_attn_groups(attn_backend_map, i))
4766

4767
4768
4769
4770
4771
4772
4773
4774
4775
4776
4777
4778
4779
4780
4781
4782
4783
4784
    def initialize_metadata_builders(
        self, kv_cache_config: KVCacheConfig, kernel_block_sizes: list[int]
    ) -> None:
        """
        Create the metadata builders for all KV cache groups and attn groups.
        """
        for kv_cache_group_id in range(len(kv_cache_config.kv_cache_groups)):
            for attn_group in self.attn_groups[kv_cache_group_id]:
                attn_group.create_metadata_builders(
                    self.vllm_config,
                    self.device,
                    kernel_block_sizes[kv_cache_group_id]
                    if kv_cache_group_id < len(kernel_block_sizes)
                    else None,
                    num_metadata_builders=1
                    if not self.parallel_config.enable_dbo
                    else 2,
                )
co63oc's avatar
co63oc committed
4785
        # Calculate reorder batch threshold (if needed)
4786
4787
        # Note (tdoublep): do this *after* constructing builders,
        # because some of them change the threshold at init time.
4788
4789
        self.calculate_reorder_batch_threshold()

4790
    def _check_and_update_cudagraph_mode(
4791
4792
4793
        self,
        attention_backends: list[set[type[AttentionBackend]]],
        kv_cache_groups: list[KVCacheGroupSpec],
4794
    ) -> None:
4795
        """
4796
        Resolve the cudagraph_mode when there are multiple attention
4797
        groups with potential conflicting CUDA graph support.
4798
4799
4800
        Then initialize the cudagraph_dispatcher based on the resolved
        cudagraph_mode.
        """
4801
        min_cg_support = AttentionCGSupport.ALWAYS
4802
        min_cg_backend_name = None
4803

4804
4805
4806
4807
4808
4809
4810
4811
4812
4813
4814
4815
        for attn_backend_set, kv_cache_group in zip(
            attention_backends, kv_cache_groups
        ):
            for attn_backend in attn_backend_set:
                builder_cls = attn_backend.get_builder_cls()

                cg_support = builder_cls.get_cudagraph_support(
                    self.vllm_config, kv_cache_group.kv_cache_spec
                )
                if cg_support.value < min_cg_support.value:
                    min_cg_support = cg_support
                    min_cg_backend_name = attn_backend.__name__
4816
4817
        # Flexible resolve the cudagraph mode
        cudagraph_mode = self.compilation_config.cudagraph_mode
4818
        assert cudagraph_mode is not None
4819
        # check cudagraph for mixed batch is supported
4820
4821
4822
4823
4824
4825
        if (
            cudagraph_mode.mixed_mode() == CUDAGraphMode.FULL
            and min_cg_support != AttentionCGSupport.ALWAYS
        ):
            msg = (
                f"CUDAGraphMode.{cudagraph_mode.name} is not supported "
4826
                f"with {min_cg_backend_name} backend (support: "
4827
4828
                f"{min_cg_support})"
            )
4829
4830
            if min_cg_support == AttentionCGSupport.NEVER:
                # if not supported any full cudagraphs, just raise it.
4831
4832
                msg += (
                    "; please try cudagraph_mode=PIECEWISE, and "
4833
                    "make sure compilation mode is VLLM_COMPILE"
4834
                )
4835
4836
4837
4838
4839
                raise ValueError(msg)

            # attempt to resolve the full cudagraph related mode
            if self.compilation_config.splitting_ops_contain_attention():
                msg += "; setting cudagraph_mode=FULL_AND_PIECEWISE"
4840
                cudagraph_mode = self.compilation_config.cudagraph_mode = (
4841
                    CUDAGraphMode.FULL_AND_PIECEWISE
4842
                )
4843
4844
            else:
                msg += "; setting cudagraph_mode=FULL_DECODE_ONLY"
4845
                cudagraph_mode = self.compilation_config.cudagraph_mode = (
4846
                    CUDAGraphMode.FULL_DECODE_ONLY
4847
                )
4848
4849
            logger.warning(msg)

4850
        # check that if we are doing decode full-cudagraphs it is supported
4851
4852
4853
4854
4855
4856
        if (
            cudagraph_mode.decode_mode() == CUDAGraphMode.FULL
            and min_cg_support == AttentionCGSupport.NEVER
        ):
            msg = (
                f"CUDAGraphMode.{cudagraph_mode.name} is not supported "
4857
                f"with {min_cg_backend_name} backend (support: "
4858
4859
                f"{min_cg_support})"
            )
4860
            if self.compilation_config.mode == CompilationMode.VLLM_COMPILE and (
4861
4862
4863
4864
4865
                self.compilation_config.splitting_ops_contain_attention()
                or self.compilation_config.use_inductor_graph_partition
            ):
                msg += (
                    "; setting cudagraph_mode=PIECEWISE because "
4866
                    "attention is compiled piecewise"
4867
4868
                )
                cudagraph_mode = self.compilation_config.cudagraph_mode = (
4869
                    CUDAGraphMode.PIECEWISE
4870
                )
4871
            else:
4872
4873
                msg += (
                    "; setting cudagraph_mode=NONE because "
4874
                    "attention is not compiled piecewise"
4875
4876
                )
                cudagraph_mode = self.compilation_config.cudagraph_mode = (
4877
                    CUDAGraphMode.NONE
4878
                )
4879
4880
            logger.warning(msg)

4881
4882
        # check that if we are doing spec-decode + decode full-cudagraphs it is
        # supported
4883
4884
4885
4886
4887
4888
4889
4890
        if (
            cudagraph_mode.decode_mode() == CUDAGraphMode.FULL
            and self.uniform_decode_query_len > 1
            and min_cg_support.value < AttentionCGSupport.UNIFORM_BATCH.value
        ):
            msg = (
                f"CUDAGraphMode.{cudagraph_mode.name} is not supported"
                f" with spec-decode for attention backend "
4891
                f"{min_cg_backend_name} (support: {min_cg_support})"
4892
            )
4893
4894
            if self.compilation_config.splitting_ops_contain_attention():
                msg += "; setting cudagraph_mode=PIECEWISE"
4895
                cudagraph_mode = self.compilation_config.cudagraph_mode = (
4896
                    CUDAGraphMode.PIECEWISE
4897
                )
4898
4899
            else:
                msg += "; setting cudagraph_mode=NONE"
4900
                cudagraph_mode = self.compilation_config.cudagraph_mode = (
4901
                    CUDAGraphMode.NONE
4902
                )
4903
4904
4905
4906
            logger.warning(msg)

        # double check that we can support full cudagraph if they are requested
        # even after automatic downgrades
4907
4908
4909
4910
4911
4912
        if (
            cudagraph_mode.has_full_cudagraphs()
            and min_cg_support == AttentionCGSupport.NEVER
        ):
            raise ValueError(
                f"CUDAGraphMode.{cudagraph_mode.name} is not "
4913
                f"supported with {min_cg_backend_name} backend ("
4914
4915
                f"support:{min_cg_support}) "
                "; please try cudagraph_mode=PIECEWISE, "
4916
                "and make sure compilation mode is VLLM_COMPILE"
4917
            )
4918

4919
4920
4921
4922
        # if we have dedicated decode cudagraphs, and spec-decode is enabled,
        # we need to adjust the cudagraph sizes to be a multiple of the uniform
        # decode query length to avoid: https://github.com/vllm-project/vllm/issues/28207
        # temp-fix: https://github.com/vllm-project/vllm/issues/28207#issuecomment-3504004536
4923
        # Will be removed in the near future when we have separate cudagraph capture
4924
4925
4926
4927
4928
4929
4930
4931
4932
        # sizes for decode and mixed prefill-decode.
        if (
            cudagraph_mode.decode_mode() == CUDAGraphMode.FULL
            and cudagraph_mode.separate_routine()
            and self.uniform_decode_query_len > 1
        ):
            self.compilation_config.adjust_cudagraph_sizes_for_spec_decode(
                self.uniform_decode_query_len, self.parallel_config.tensor_parallel_size
            )
4933
4934
4935
4936
            capture_sizes = self.compilation_config.cudagraph_capture_sizes
            self.cudagraph_batch_sizes = (
                capture_sizes if capture_sizes is not None else []
            )
4937

4938
4939
        # Trigger cudagraph dispatching keys initialization after
        # resolved cudagraph mode.
4940
        self.compilation_config.cudagraph_mode = cudagraph_mode
4941
        self.cudagraph_dispatcher.initialize_cudagraph_keys(
4942
            cudagraph_mode, self.uniform_decode_query_len
4943
        )
4944

4945
4946
    def calculate_reorder_batch_threshold(self) -> None:
        """
4947
4948
4949
4950
        Choose the minimum reorder batch threshold from all attention groups.
        Backends should be able to support lower threshold then what they request
        just may have a performance penalty due to that backend treating decodes
        as prefills.
4951
        """
4952
4953
        min_none_high = lambda a, b: a if b is None else b if a is None else min(a, b)

4954
        reorder_batch_thresholds: list[int | None] = [
4955
4956
4957
            group.get_metadata_builder().reorder_batch_threshold
            for group in self._attn_group_iterator()
        ]
4958
4959
4960
4961
4962
        # If there are no attention groups (attention-free model) or no backend
        # reports a threshold, leave reordering disabled.
        if len(reorder_batch_thresholds) == 0:
            self.reorder_batch_threshold = None
            return
4963
        self.reorder_batch_threshold = reduce(min_none_high, reorder_batch_thresholds)  # type: ignore[assignment]
4964

4965
4966
4967
    @staticmethod
    def select_common_block_size(
        kv_manager_block_size: int, attn_groups: list[AttentionGroup]
4968
4969
    ) -> int:
        """
4970
4971
4972
4973
4974
        Select a block size that is supported by all backends and is a factor of
        kv_manager_block_size.

        If kv_manager_block_size is supported by all backends, return it directly.
        Otherwise, return the max supported size.
4975
4976
4977
4978
4979
4980

        Args:
            kv_manager_block_size: Block size of KV cache
            attn_groups: List of attention groups

        Returns:
4981
            The selected block size
4982
4983

        Raises:
4984
            ValueError: If no valid block size found
4985
4986
        """

4987
4988
4989
4990
4991
4992
4993
4994
        def block_size_is_supported(
            backends: list[type[AttentionBackend]], block_size: int
        ) -> bool:
            """
            Check if the block size is supported by all backends.
            """
            for backend in backends:
                is_supported = False
4995
                for supported_size in backend.get_supported_kernel_block_sizes():
4996
4997
4998
4999
5000
5001
5002
5003
5004
5005
5006
5007
5008
5009
5010
5011
5012
5013
5014
5015
5016
5017
5018
5019
5020
5021
5022
5023
5024
5025
                    if isinstance(supported_size, int):
                        if block_size == supported_size:
                            is_supported = True
                    elif isinstance(supported_size, MultipleOf):
                        if block_size % supported_size.base == 0:
                            is_supported = True
                    else:
                        raise ValueError(f"Unknown supported size: {supported_size}")
                if not is_supported:
                    return False
            return True

        backends = [group.backend for group in attn_groups]

        # Case 1: if the block_size of kv cache manager is supported by all backends,
        # return it directly
        if block_size_is_supported(backends, kv_manager_block_size):
            return kv_manager_block_size

        # Case 2: otherwise, the block_size must be an `int`-format supported size of
        # at least one backend. Iterate over all `int`-format supported sizes in
        # descending order and return the first one that is supported by all backends.
        # Simple proof:
        # If the supported size b is in MultipleOf(x_i) format for all attention
        # backends i, and b a factor of kv_manager_block_size, then
        # kv_manager_block_size also satisfies MultipleOf(x_i) for all i. We will
        # return kv_manager_block_size in case 1.
        all_int_supported_sizes = set(
            supported_size
            for backend in backends
5026
            for supported_size in backend.get_supported_kernel_block_sizes()
5027
5028
            if isinstance(supported_size, int)
        )
5029

5030
5031
5032
5033
5034
5035
        for supported_size in sorted(all_int_supported_sizes, reverse=True):
            if kv_manager_block_size % supported_size != 0:
                continue
            if block_size_is_supported(backends, supported_size):
                return supported_size
        raise ValueError(f"No common block size for {kv_manager_block_size}. ")
5036

5037
5038
5039
    def may_reinitialize_input_batch(
        self, kv_cache_config: KVCacheConfig, kernel_block_sizes: list[int]
    ) -> None:
5040
5041
5042
5043
5044
5045
5046
        """
        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.
5047
            kernel_block_sizes: The kernel block sizes for each KV cache group.
5048
5049
5050
5051
        """
        block_sizes = [
            kv_cache_group.kv_cache_spec.block_size
            for kv_cache_group in kv_cache_config.kv_cache_groups
5052
            if not isinstance(kv_cache_group.kv_cache_spec, EncoderOnlyAttentionSpec)
5053
        ]
5054
5055
5056
5057

        if block_sizes != [self.cache_config.block_size] or kernel_block_sizes != [
            self.cache_config.block_size
        ]:
5058
5059
5060
            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
5061
5062
                "for more details."
            )
5063
5064
            self.input_batch = InputBatch(
                max_num_reqs=self.max_num_reqs,
5065
                max_model_len=max(self.max_model_len, self.max_encoder_len),
5066
5067
5068
5069
5070
                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,
5071
                kernel_block_sizes=kernel_block_sizes,
5072
                is_spec_decode=bool(self.vllm_config.speculative_config),
5073
                logitsprocs=self.input_batch.logitsprocs,
5074
                logitsprocs_need_output_token_ids=self.input_batch.logitsprocs_need_output_token_ids,
5075
                is_pooling_model=self.is_pooling_model,
5076
                num_speculative_tokens=self.num_spec_tokens,
5077
5078
            )

5079
    def _allocate_kv_cache_tensors(
5080
5081
        self, kv_cache_config: KVCacheConfig
    ) -> dict[str, torch.Tensor]:
5082
        """
5083
5084
5085
        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.

5086
        Args:
5087
            kv_cache_config: The KV cache config
5088
        Returns:
5089
            dict[str, torch.Tensor]: A map between layer names to their
5090
            corresponding memory buffer for KV cache.
5091
        """
5092
5093
        kv_cache_raw_tensors: dict[str, torch.Tensor] = {}
        for kv_cache_tensor in kv_cache_config.kv_cache_tensors:
5094
5095
5096
            tensor = torch.zeros(
                kv_cache_tensor.size, dtype=torch.int8, device=self.device
            )
5097
5098
5099
5100
5101
            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:
5102
5103
5104
5105
            for layer_name in group.layer_names:
                if layer_name in self.runner_only_attn_layers:
                    continue
                layer_names.add(layer_name)
5106
5107
5108
        assert layer_names == set(kv_cache_raw_tensors.keys()), (
            "Some layers are not correctly initialized"
        )
5109
5110
        return kv_cache_raw_tensors

5111
5112
5113
    def _attn_group_iterator(self) -> Iterator[AttentionGroup]:
        return itertools.chain.from_iterable(self.attn_groups)

5114
    def _kv_cache_spec_attn_group_iterator(self) -> Iterator[AttentionGroup]:
5115
5116
        if not self.kv_cache_config.kv_cache_groups:
            return
5117
5118
        for attn_groups in self.attn_groups:
            yield from attn_groups
5119

5120
5121
5122
5123
5124
5125
5126
5127
5128
5129
5130
5131
5132
5133
5134
    def _prepare_kernel_block_sizes(self, kv_cache_config: KVCacheConfig) -> list[int]:
        """
        Generate kernel_block_sizes that matches each block_size.

        For attention backends that support virtual block splitting,
        use the supported block sizes from the backend.
        For other backends (like Mamba), use the same block size (no splitting).

        Args:
            kv_cache_config: The KV cache configuration.

        Returns:
            list[int]: List of kernel block sizes for each cache group.
        """
        kernel_block_sizes = []
5135
        for kv_cache_gid, kv_cache_group in enumerate(kv_cache_config.kv_cache_groups):
5136
5137
5138
5139
5140
5141
            kv_cache_spec = kv_cache_group.kv_cache_spec
            if isinstance(kv_cache_spec, UniformTypeKVCacheSpecs):
                # All layers in the UniformTypeKVCacheSpecs have the same type,
                # Pick an arbitrary one to dispatch.
                kv_cache_spec = next(iter(kv_cache_spec.kv_cache_specs.values()))
            if isinstance(kv_cache_spec, EncoderOnlyAttentionSpec):
5142
                continue
5143
            elif isinstance(kv_cache_spec, AttentionSpec):
5144
5145
5146
                # This is an attention backend that supports virtual
                # block splitting. Get the supported block sizes from
                # all backends in the group.
5147
                attn_groups = self.attn_groups[kv_cache_gid]
5148
                kv_manager_block_size = kv_cache_group.kv_cache_spec.block_size
5149
                selected_kernel_size = self.select_common_block_size(
5150
5151
5152
                    kv_manager_block_size, attn_groups
                )
                kernel_block_sizes.append(selected_kernel_size)
5153
            elif isinstance(kv_cache_spec, MambaSpec):
5154
5155
                # This is likely Mamba or other non-attention cache,
                # no splitting.
5156
                kernel_block_sizes.append(kv_cache_spec.block_size)
5157
5158
5159
5160
5161
5162
            else:
                raise NotImplementedError(
                    f"unknown kv cache spec {kv_cache_group.kv_cache_spec}"
                )
        return kernel_block_sizes

5163
5164
5165
5166
    def _reshape_kv_cache_tensors(
        self,
        kv_cache_config: KVCacheConfig,
        kv_cache_raw_tensors: dict[str, torch.Tensor],
5167
        kernel_block_sizes: list[int],
5168
    ) -> dict[str, torch.Tensor]:
5169
        """
5170
        Reshape the KV cache tensors to the desired shape and dtype.
5171

5172
        Args:
5173
5174
            kv_cache_config: The KV cache config
            kv_cache_raw_tensors: The KV cache buffer of each layer, with
5175
                correct size but uninitialized shape.
5176
            kernel_block_sizes: The kernel block sizes for each KV cache group.
5177
        Returns:
5178
            Dict[str, torch.Tensor]: A map between layer names to their
5179
5180
            corresponding memory buffer for KV cache.
        """
5181
        kv_caches: dict[str, torch.Tensor] = {}
5182
        has_attn, has_mamba = False, False
5183
5184
        for group in self._kv_cache_spec_attn_group_iterator():
            kv_cache_spec = group.kv_cache_spec
5185
            attn_backend = group.backend
5186
5187
5188
5189
            if group.kv_cache_group_id == len(kernel_block_sizes):
                # There may be a last group for layers without kv cache.
                continue
            kernel_block_size = kernel_block_sizes[group.kv_cache_group_id]
5190
            for layer_name in group.layer_names:
5191
5192
                if layer_name in self.runner_only_attn_layers:
                    continue
5193
5194
                raw_tensor = kv_cache_raw_tensors[layer_name]
                assert raw_tensor.numel() % kv_cache_spec.page_size_bytes == 0
5195
                num_blocks = raw_tensor.numel() // kv_cache_spec.page_size_bytes
5196
                if isinstance(kv_cache_spec, AttentionSpec):
5197
                    has_attn = True
5198
5199
                    num_blocks_per_kv_block = (
                        kv_cache_spec.block_size // kernel_block_size
5200
5201
5202
                    )
                    kernel_num_blocks = num_blocks * num_blocks_per_kv_block

5203
                    kv_cache_shape = attn_backend.get_kv_cache_shape(
5204
                        kernel_num_blocks,
5205
                        kernel_block_size,
5206
5207
                        kv_cache_spec.num_kv_heads,
                        kv_cache_spec.head_size,
5208
5209
                        cache_dtype_str=self.cache_config.cache_dtype,
                    )
5210
                    dtype = kv_cache_spec.dtype
5211
                    try:
5212
                        kv_cache_stride_order = attn_backend.get_kv_cache_stride_order()
5213
                        assert len(kv_cache_stride_order) == len(kv_cache_shape)
5214
                    except (AttributeError, NotImplementedError):
5215
                        kv_cache_stride_order = tuple(range(len(kv_cache_shape)))
5216
5217
5218
5219
5220
                    # 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.
5221
5222
5223
                    kv_cache_shape = tuple(
                        kv_cache_shape[i] for i in kv_cache_stride_order
                    )
5224
5225
5226
5227
5228
                    # Maintain original KV shape view.
                    inv_order = [
                        kv_cache_stride_order.index(i)
                        for i in range(len(kv_cache_stride_order))
                    ]
5229
5230
5231
5232
5233
5234
                    kv_caches[layer_name] = (
                        kv_cache_raw_tensors[layer_name]
                        .view(dtype)
                        .view(kv_cache_shape)
                        .permute(*inv_order)
                    )
Chen Zhang's avatar
Chen Zhang committed
5235
                elif isinstance(kv_cache_spec, MambaSpec):
5236
                    has_mamba = True
Chen Zhang's avatar
Chen Zhang committed
5237
5238
                    raw_tensor = kv_cache_raw_tensors[layer_name]
                    state_tensors = []
5239
                    storage_offset_bytes = 0
5240
                    for shape, dtype in zip(kv_cache_spec.shapes, kv_cache_spec.dtypes):
5241
5242
                        dtype_size = get_dtype_size(dtype)
                        num_element_per_page = (
5243
5244
                            kv_cache_spec.page_size_bytes // dtype_size
                        )
Chen Zhang's avatar
Chen Zhang committed
5245
                        target_shape = (num_blocks, *shape)
5246
5247
                        stride = torch.empty(target_shape).stride()
                        target_stride = (num_element_per_page, *stride[1:])
5248
                        assert storage_offset_bytes % dtype_size == 0
5249
5250
5251
5252
                        tensor = torch.as_strided(
                            raw_tensor.view(dtype),
                            size=target_shape,
                            stride=target_stride,
5253
                            storage_offset=storage_offset_bytes // dtype_size,
5254
                        )
Chen Zhang's avatar
Chen Zhang committed
5255
                        state_tensors.append(tensor)
5256
                        storage_offset_bytes += stride[0] * dtype_size
5257
5258

                    kv_caches[layer_name] = state_tensors
5259
                else:
5260
                    raise NotImplementedError
5261
5262

        if has_attn and has_mamba:
5263
            self._update_hybrid_attention_mamba_layout(kv_caches)
5264

5265
5266
        return kv_caches

5267
    def _update_hybrid_attention_mamba_layout(
5268
5269
        self, kv_caches: dict[str, torch.Tensor]
    ) -> None:
5270
        """
5271
5272
        Update the layout of attention layers from (2, num_blocks, ...) to
        (num_blocks, 2, ...).
5273
5274

        Args:
5275
            kv_caches: The KV cache buffer of each layer.
5276
5277
        """

5278
5279
        for group in self._kv_cache_spec_attn_group_iterator():
            kv_cache_spec = group.kv_cache_spec
5280
            for layer_name in group.layer_names:
5281
                kv_cache = kv_caches[layer_name]
5282
5283
5284
5285
                if isinstance(kv_cache_spec, AttentionSpec) and kv_cache.shape[0] == 2:
                    assert kv_cache.shape[1] != 2, (
                        "Fail to determine whether the layout is "
                        "(2, num_blocks, ...) or (num_blocks, 2, ...) for "
5286
                        f"a tensor of shape {kv_cache.shape}"
5287
                    )
5288
                    hidden_size = kv_cache.shape[2:].numel()
5289
5290
5291
5292
                    kv_cache.as_strided_(
                        size=kv_cache.shape,
                        stride=(hidden_size, 2 * hidden_size, *kv_cache.stride()[2:]),
                    )
5293

5294
    def initialize_kv_cache_tensors(
5295
        self, kv_cache_config: KVCacheConfig, kernel_block_sizes: list[int]
5296
    ) -> dict[str, torch.Tensor]:
5297
5298
5299
5300
5301
        """
        Initialize the memory buffer for KV cache.

        Args:
            kv_cache_config: The KV cache config
5302
5303
            kernel_block_sizes: The kernel block sizes for each KV cache group.

5304
        Returns:
5305
            Dict[str, torch.Tensor]: A map between layer names to their
5306
5307
            corresponding memory buffer for KV cache.
        """
5308
5309
5310
5311
5312
5313
5314
5315
5316
5317
5318
5319
5320
5321
5322
5323
5324
5325
5326
5327
5328
5329
5330
5331

        # Try creating KV caches optimized for kv-connector transfers
        cache_dtype = self.cache_config.cache_dtype
        if self.use_uniform_kv_cache(self.attn_groups, cache_dtype):
            kv_caches, cross_layers_kv_cache, attn_backend = (
                self.allocate_uniform_kv_caches(
                    kv_cache_config,
                    self.attn_groups,
                    cache_dtype,
                    self.device,
                    kernel_block_sizes,
                )
            )
            self.cross_layers_kv_cache = cross_layers_kv_cache
            self.cross_layers_attn_backend = attn_backend
        else:
            # Fallback to the general case
            # 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, kernel_block_sizes
            )
5332

5333
        # Set up cross-layer KV cache sharing
5334
5335
        for layer_name, target_layer_name in self.shared_kv_cache_layers.items():
            logger.debug("%s reuses KV cache of %s", layer_name, target_layer_name)
5336
5337
            kv_caches[layer_name] = kv_caches[target_layer_name]

5338
5339
5340
5341
5342
5343
5344
5345
5346
        num_attn_module = (
            2 if self.model_config.hf_config.model_type == "longcat_flash" else 1
        )
        bind_kv_cache(
            kv_caches,
            self.compilation_config.static_forward_context,
            self.kv_caches,
            num_attn_module,
        )
5347
5348
5349
        return kv_caches

    def maybe_add_kv_sharing_layers_to_kv_cache_groups(
5350
5351
        self, kv_cache_config: KVCacheConfig
    ) -> None:
5352
5353
5354
5355
5356
5357
5358
5359
5360
5361
5362
5363
5364
5365
5366
5367
5368
5369
        """
        Add layers that re-use KV cache to KV cache group of its target layer.
        Mapping of KV cache tensors happens in `initialize_kv_cache_tensors()`
        """
        if not self.shared_kv_cache_layers:
            # No cross-layer KV sharing, return
            return

        add_kv_sharing_layers_to_kv_cache_groups(
            self.shared_kv_cache_layers,
            kv_cache_config.kv_cache_groups,
            self.runner_only_attn_layers,
        )

        if self.cache_config.kv_sharing_fast_prefill:
            # In You Only Cache Once (https://arxiv.org/abs/2405.05254) or other
            # similar KV sharing setups, only the layers that generate KV caches
            # are involved in the prefill phase, enabling prefill to early exit.
5370
            attn_layers = get_layers_from_vllm_config(self.vllm_config, Attention)
5371
5372
            for layer_name in reversed(attn_layers):
                if layer_name in self.shared_kv_cache_layers:
5373
                    self.kv_sharing_fast_prefill_eligible_layers.add(layer_name)
5374
5375
                else:
                    break
5376

5377
5378
5379
5380
5381
5382
5383
    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
        """
5384
        kv_cache_config = deepcopy(kv_cache_config)
5385
        self.kv_cache_config = kv_cache_config
5386
        self.may_add_encoder_only_layers_to_kv_cache_config()
5387
        self.maybe_add_kv_sharing_layers_to_kv_cache_groups(kv_cache_config)
5388
        self.initialize_attn_backend(kv_cache_config)
5389
5390
5391
5392
5393
5394
        # The kernel block size for all KV cache groups. For example, if
        # kv_cache_manager uses block_size 256 for a given group, but the attention
        # backends for that group only supports block_size 64, we will return
        # kernel_block_size 64 and split the 256-token-block to 4 blocks with 64
        # tokens each.
        kernel_block_sizes = self._prepare_kernel_block_sizes(kv_cache_config)
5395
5396
5397
5398

        # create metadata builders
        self.initialize_metadata_builders(kv_cache_config, kernel_block_sizes)

5399
        # Reinitialize need to after initialize_attn_backend
5400
5401
5402
5403
        self.may_reinitialize_input_batch(kv_cache_config, kernel_block_sizes)
        kv_caches = self.initialize_kv_cache_tensors(
            kv_cache_config, kernel_block_sizes
        )
5404

5405
5406
5407
5408
5409
5410
        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
5411
        if has_kv_transfer_group():
5412
            kv_transfer_group = get_kv_transfer_group()
5413
5414
5415
5416
5417
5418
5419
            if self.cross_layers_kv_cache is not None:
                assert self.cross_layers_attn_backend is not None
                kv_transfer_group.register_cross_layers_kv_cache(
                    self.cross_layers_kv_cache, self.cross_layers_attn_backend
                )
            else:
                kv_transfer_group.register_kv_caches(kv_caches)
5420
            kv_transfer_group.set_host_xfer_buffer_ops(copy_kv_blocks)
Robert Shaw's avatar
Robert Shaw committed
5421

5422
5423
5424
5425
5426
    def may_add_encoder_only_layers_to_kv_cache_config(self) -> None:
        """
        Add encoder-only layers to the KV cache config.
        """
        block_size = self.vllm_config.cache_config.block_size
5427
        encoder_only_attn_specs: dict[AttentionSpec, list[str]] = defaultdict(list)
5428
5429
5430
        attn_layers = get_layers_from_vllm_config(self.vllm_config, Attention)
        for layer_name, attn_module in attn_layers.items():
            if attn_module.attn_type == AttentionType.ENCODER_ONLY:
5431
                attn_spec: AttentionSpec = EncoderOnlyAttentionSpec(
5432
5433
5434
                    block_size=block_size,
                    num_kv_heads=attn_module.num_kv_heads,
                    head_size=attn_module.head_size,
5435
5436
                    dtype=self.kv_cache_dtype,
                )
5437
5438
5439
                encoder_only_attn_specs[attn_spec].append(layer_name)
                self.runner_only_attn_layers.add(layer_name)
        if len(encoder_only_attn_specs) > 0:
5440
5441
5442
            assert len(encoder_only_attn_specs) == 1, (
                "Only support one encoder-only attention spec now"
            )
5443
5444
            spec, layer_names = encoder_only_attn_specs.popitem()
            self.kv_cache_config.kv_cache_groups.append(
5445
5446
                KVCacheGroupSpec(layer_names=layer_names, kv_cache_spec=spec)
            )
5447

5448
    def get_kv_cache_spec(self) -> dict[str, KVCacheSpec]:
5449
        """
5450
        Generates the KVCacheSpec by parsing the kv cache format from each
5451
5452
        Attention module in the static forward context.
        Returns:
5453
            KVCacheSpec: A dictionary mapping layer names to their KV cache
5454
5455
            format. Layers that do not need KV cache are not included.
        """
5456
5457
        if has_ec_transfer() and get_ec_transfer().is_producer:
            return {}
5458
        kv_cache_spec: dict[str, KVCacheSpec] = {}
5459
5460
        layer_type = cast(type[Any], AttentionLayerBase)
        attn_layers = get_layers_from_vllm_config(self.vllm_config, layer_type)
Chen Zhang's avatar
Chen Zhang committed
5461
        for layer_name, attn_module in attn_layers.items():
5462
5463
5464
5465
5466
5467
5468
5469
5470
5471
5472
5473
5474
5475
5476
            if isinstance(attn_module, Attention) and (
                kv_tgt_layer := attn_module.kv_sharing_target_layer_name
            ):
                # 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
            # Skip modules that don't need KV cache (eg encoder-only attention)
            if spec := attn_module.get_kv_cache_spec(self.vllm_config):
                kv_cache_spec[layer_name] = spec
5477

5478
        return kv_cache_spec
5479

5480
    def _to_list(self, sampled_token_ids: torch.Tensor) -> list[list[int]]:
5481
5482
5483
5484
5485
5486
5487
5488
        # This is a short term mitigation for issue mentioned in
        # https://github.com/vllm-project/vllm/issues/22754.
        # `tolist` would trigger a cuda wise stream sync, which
        # would block other copy ops from other cuda streams.
        # A cuda event sync would avoid such a situation. Since
        # this is in the critical path of every single model
        # forward loop, this has caused perf issue for a disagg
        # setup.
5489
        pinned = self.sampled_token_ids_pinned_cpu[: sampled_token_ids.shape[0]]
5490
5491
5492
        pinned.copy_(sampled_token_ids, non_blocking=True)
        self.transfer_event.record()
        self.transfer_event.synchronize()
5493
        return pinned.tolist()