gpu_model_runner.py 238 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
from .utils import (
    AttentionGroup,
    MultiModalBudget,
    add_kv_sharing_layers_to_kv_cache_groups,
    bind_kv_cache,
    sanity_check_mm_encoder_outputs,
)
174

175
if TYPE_CHECKING:
176
    from vllm.model_executor.model_loader.tensorizer import TensorizerConfig
177
    from vllm.v1.core.sched.output import GrammarOutput, SchedulerOutput
178
179
180

logger = init_logger(__name__)

181
182
AttnMetadataDict: TypeAlias = dict[str, AttentionMetadata]
# list when ubatching is enabled
183
PerLayerAttnMetadata: TypeAlias = list[AttnMetadataDict] | AttnMetadataDict
184

185

186
187
188
189
190
191
# Wrapper for ModelRunnerOutput to support overlapped execution.
class AsyncGPUModelRunnerOutput(AsyncModelRunnerOutput):
    def __init__(
        self,
        model_runner_output: ModelRunnerOutput,
        sampled_token_ids: torch.Tensor,
192
        logprobs_tensors: LogprobsTensors | None,
193
194
        invalid_req_indices: list[int],
        async_output_copy_stream: torch.cuda.Stream,
195
        vocab_size: int,
196
197
198
199
200
    ):
        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.
201
        self.async_copy_ready_event = torch.Event()
202
203
204
205

        # 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
206
        self.vocab_size = vocab_size
207
        self._logprobs_tensors = logprobs_tensors
208
209
210
211
212

        # 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)
213
            self.sampled_token_ids_cpu = self._sampled_token_ids.to(
214
215
                "cpu", non_blocking=True
            )
216
217
218
219
220
            self._logprobs_tensors_cpu = (
                self._logprobs_tensors.to_cpu_nonblocking()
                if self._logprobs_tensors
                else None
            )
221
            self.async_copy_ready_event.record()
222
223
224

    def get_output(self) -> ModelRunnerOutput:
        """Copy the device tensors to the host and return a ModelRunnerOutput.
225

226
227
        This function blocks until the copy is finished.
        """
228
        max_gen_len = self.sampled_token_ids_cpu.shape[-1]
229
        self.async_copy_ready_event.synchronize()
230

231
232
        # Release the device tensors once the copy has completed.
        del self._logprobs_tensors
233
        del self._sampled_token_ids
234
        if max_gen_len == 1:
235
            valid_sampled_token_ids = self.sampled_token_ids_cpu.tolist()
236
237
238
            for i in self._invalid_req_indices:
                valid_sampled_token_ids[i].clear()
            cu_num_tokens = None
239
        else:
240
            valid_sampled_token_ids, cu_num_tokens = RejectionSampler.parse_output(
241
242
                self.sampled_token_ids_cpu,
                self.vocab_size,
243
244
                self._invalid_req_indices,
                return_cu_num_tokens=self._logprobs_tensors_cpu is not None,
245
            )
246
247
248

        output = self._model_runner_output
        output.sampled_token_ids = valid_sampled_token_ids
249
        if self._logprobs_tensors_cpu:
250
            output.logprobs = self._logprobs_tensors_cpu.tolists(cu_num_tokens)
251
252
253
        return output


254
255
256
257
258
259
260
261
262
263
264
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
265
    ec_connector_output: ECConnectorOutput | None
266
    cudagraph_stats: CUDAGraphStat | None
267
268


269
270
271
class GPUModelRunner(
    LoRAModelRunnerMixin, KVConnectorModelRunnerMixin, ECConnectorModelRunnerMixin
):
272
273
    def __init__(
        self,
274
        vllm_config: VllmConfig,
275
        device: torch.device,
276
    ):
277
278
279
        self.vllm_config = vllm_config
        self.model_config = vllm_config.model_config
        self.cache_config = vllm_config.cache_config
280
        self.compilation_config = vllm_config.compilation_config
281
282
283
284
285
286
        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
287

288
        from vllm.model_executor.models.utils import set_cpu_offload_max_bytes
289
290

        set_cpu_offload_max_bytes(int(self.cache_config.cpu_offload_gb * 1024**3))
291

292
293
294
295
        model_config = self.model_config
        cache_config = self.cache_config
        scheduler_config = self.scheduler_config
        parallel_config = self.parallel_config
296
        self.device = device
297
298
        self.pin_memory = is_pin_memory_available()
        self.dtype = self.model_config.dtype
299

300
301
302
        self.kv_cache_dtype = kv_cache_dtype_str_to_dtype(
            cache_config.cache_dtype, self.model_config
        )
303

304
        self.is_pooling_model = model_config.runner_type == "pooling"
305
        self.enable_prompt_embeds = model_config.enable_prompt_embeds
306
        self.is_multimodal_raw_input_only_model = (
307
308
            model_config.is_multimodal_raw_input_only_model
        )
309
310
        # This will be overridden in load_model()
        self.is_multimodal_pruning_enabled = False
311
        self.max_model_len = model_config.max_model_len
312
313
314

        # Always set to false after the first forward pass
        self.calculate_kv_scales = self.cache_config.calculate_kv_scales
315
        self.dcp_world_size = self.parallel_config.decode_context_parallel_size
316
        self.dcp_rank = 0 if self.dcp_world_size <= 1 else get_dcp_group().rank_in_group
317
        self.max_num_tokens = scheduler_config.max_num_batched_tokens
318
        self.max_num_reqs = scheduler_config.max_num_seqs
319

320
321
322
323
324
        # 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 = (
325
326
327
            self.parallel_config.distributed_executor_backend == "external_launcher"
            and len(get_pp_group().ranks) > 0
        )
328

329
        # Model-related.
330
        self.num_query_heads = model_config.get_num_attention_heads(parallel_config)
331
        self.inputs_embeds_size = model_config.get_inputs_embeds_size()
332
        self.attention_chunk_size = model_config.attention_chunk_size
333
        # Only relevant for models using ALiBi (e.g, MPT)
334
        self.use_alibi = model_config.uses_alibi
335

336
        self.cascade_attn_enabled = not self.model_config.disable_cascade_attn
337
        self.is_mm_prefix_lm = self.model_config.is_mm_prefix_lm
338

339
        # Multi-modal data support
340
        self.mm_registry = MULTIMODAL_REGISTRY
341
        self.uses_mrope = model_config.uses_mrope
342
        self.uses_xdrope_dim = model_config.uses_xdrope_dim
343
        self.supports_mm_inputs = self.mm_registry.supports_multimodal_inputs(
344
            model_config
345
        )
346

347
348
349
        if self.model_config.is_encoder_decoder:
            # Maximum length of the encoder input, only for encoder-decoder
            # models.
350
            self.max_encoder_len = scheduler_config.max_num_encoder_input_tokens
351
352
353
        else:
            self.max_encoder_len = 0

354
        # Sampler
355
        self.sampler = Sampler(logprobs_mode=self.model_config.logprobs_mode)
356

357
        self.eplb_state: EplbState | None = None
358
359
360
361
362
363
        """
        State of the expert parallelism load balancer.

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

364
        # Lazy initializations
365
        # self.model: nn.Module  # Set after load_model
366
        # Initialize in initialize_kv_cache
367
        self.kv_caches: list[torch.Tensor] = []
368
369
370
        # Initialize in initialize_kv_cache_tensors
        self.cross_layers_kv_cache: torch.Tensor | None = None
        self.cross_layers_attn_backend: type[AttentionBackend] | None = None
371
372
        # indexes: [kv_cache_group_id][attn_group]
        self.attn_groups: list[list[AttentionGroup]] = []
373
374
        # self.kv_cache_config: KVCacheConfig

375
376
        # mm_hash ->  encoder_output
        self.encoder_cache: dict[str, torch.Tensor] = {}
377

378
        self.use_aux_hidden_state_outputs = False
379
380
381
382
383
        # 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:
384
385
386
            self.drafter: (
                NgramProposer | SuffixDecodingProposer | EagleProposer | MedusaProposer
            )
387
388
            if self.speculative_config.method == "ngram":
                self.drafter = NgramProposer(self.vllm_config)
389
390
            elif self.speculative_config.method == "suffix":
                self.drafter = SuffixDecodingProposer(self.vllm_config)
391
            elif self.speculative_config.use_eagle():
392
                self.drafter = EagleProposer(self.vllm_config, self.device, self)
393
                if self.speculative_config.method == "eagle3":
394
395
396
                    self.use_aux_hidden_state_outputs = (
                        self.drafter.eagle3_use_aux_hidden_state
                    )
397
398
            elif self.speculative_config.method == "medusa":
                self.drafter = MedusaProposer(
399
                    vllm_config=self.vllm_config, device=self.device
400
                )
401
            else:
402
403
404
405
                raise ValueError(
                    "Unknown speculative decoding method: "
                    f"{self.speculative_config.method}"
                )
406
            self.rejection_sampler = RejectionSampler(self.sampler)
407

408
409
410
411
        self.num_spec_tokens = 0
        if self.speculative_config:
            self.num_spec_tokens = self.speculative_config.num_speculative_tokens

412
        # Request states.
413
        self.requests: dict[str, CachedRequestState] = {}
414
415
416
        # NOTE(rob): num_prompt_logprobs only includes reqs
        # that are currently in the prefill phase.
        self.num_prompt_logprobs: dict[str, int] = {}
417
        self.comm_stream = torch.cuda.Stream()
418

419
420
421
422
423
424
425
426
427
        # 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.
428
429
430
431
        logits_processors = model_config.logits_processors
        custom_logitsprocs: Sequence[str | type[LogitsProcessor]] = (
            tuple(logits_processors) if logits_processors is not None else ()
        )
432
433
        self.input_batch = InputBatch(
            max_num_reqs=self.max_num_reqs,
434
435
436
            # 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),
437
438
439
            max_num_batched_tokens=self.max_num_tokens,
            device=self.device,
            pin_memory=self.pin_memory,
440
            vocab_size=self.model_config.get_vocab_size(),
441
            block_sizes=[self.cache_config.block_size],
442
            kernel_block_sizes=[self.cache_config.block_size],
443
            is_spec_decode=bool(self.vllm_config.speculative_config),
444
            logitsprocs=build_logitsprocs(
445
446
447
                self.vllm_config,
                self.device,
                self.pin_memory,
448
                self.is_pooling_model,
449
                custom_logitsprocs,
450
            ),
451
452
453
            # 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),
454
            is_pooling_model=self.is_pooling_model,
455
            cp_kv_cache_interleave_size=self.parallel_config.cp_kv_cache_interleave_size,
456
        )
457

458
        self.use_async_scheduling = self.scheduler_config.async_scheduling
459
460
461
462
463
        # 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.
464
        self.prepare_inputs_event: torch.Event | None = None
465
466
        if self.use_async_scheduling:
            self.async_output_copy_stream = torch.cuda.Stream()
467
            self.prepare_inputs_event = torch.Event()
468

469
        # self.cudagraph_batch_sizes sorts in ascending order.
470
471
472
473
        if (
            self.compilation_config.cudagraph_capture_sizes
            and self.compilation_config.cudagraph_mode != CUDAGraphMode.NONE
        ):
474
475
            self.cudagraph_batch_sizes = sorted(
                self.compilation_config.cudagraph_capture_sizes
476
            )
477

478
        # Cache the device properties.
479
        self._init_device_properties()
480

481
        # Persistent buffers for CUDA graphs.
482
483
484
485
486
        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
        )
487
        self.seq_lens = self._make_buffer(self.max_num_reqs, dtype=torch.int32)
488
        self.encoder_seq_lens = self._make_buffer(self.max_num_reqs, dtype=torch.int32)
489
490
491
492
        if self.dcp_world_size > 1:
            self.dcp_local_seq_lens = self._make_buffer(
                self.max_num_reqs, dtype=torch.int32
            )
493
494
495
        # 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.
496
        self.inputs_embeds = self._make_buffer(
497
            self.max_num_tokens, self.inputs_embeds_size, dtype=self.dtype, numpy=False
498
499
        )
        self.is_token_ids = self._make_buffer(self.max_num_tokens, dtype=torch.bool)
500
501
        self.discard_request_mask = self._make_buffer(
            self.max_num_reqs, dtype=torch.bool
502
503
504
505
506
507
508
        )
        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
        )
509

510
511
        # Only relevant for multimodal models
        if self.supports_mm_inputs:
512
            self.is_mm_embed = self._make_buffer(self.max_num_tokens, dtype=torch.bool)
513

514
        # Only relevant for models using M-RoPE (e.g, Qwen2-VL)
515
        if self.uses_mrope:
Roger Wang's avatar
Roger Wang committed
516
517
518
519
            # 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
520
521
522
523
524
525

            # 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
526
            self.mrope_positions = self._make_buffer(
527
528
                (3, self.max_num_tokens + 1), dtype=torch.int64
            )
529

530
531
532
533
534
535
536
        # 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
            )

537
        # None in the first PP rank. The rest are set after load_model.
538
        self.intermediate_tensors: IntermediateTensors | None = None
539

540
        # OPTIMIZATION: Cache the tensors rather than creating them every step.
541
        # Keep in int64 to avoid overflow with long context
542
543
544
545
        self.arange_np = np.arange(
            max(self.max_num_reqs + 1, self.max_model_len, self.max_num_tokens),
            dtype=np.int64,
        )
546

547
548
549
550
551
        # 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] = {}
552
553
554
555
556
        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(
557
558
                self.max_num_tokens, dtype=torch.int32, device=self.device
            )
559

560
        self.uniform_decode_query_len = 1 + self.num_spec_tokens
561
562
563
564

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

565
566
        self.mm_budget = (
            MultiModalBudget(
567
                self.model_config,
568
569
570
571
572
573
                self.scheduler_config,
                self.mm_registry,
            )
            if self.supports_mm_inputs
            else None
        )
574

575
        self.reorder_batch_threshold: int | None = None
576

577
578
579
580
581
        # 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()

582
        # Cached outputs.
583
        self._draft_token_ids: list[list[int]] | torch.Tensor | None = None
584
        self.transfer_event = torch.Event()
585
        self.sampled_token_ids_pinned_cpu = torch.empty(
586
            (self.max_num_reqs, 1),
587
588
            dtype=torch.int64,
            device="cpu",
589
590
            pin_memory=self.pin_memory,
        )
591

592
593
        # Pre-allocated tensor for copying valid sampled token counts to CPU,
        # with dedicated stream for overlapping and event for coordination.
594
        self.valid_sampled_token_count_event: torch.Event | None = None
595
596
        self.valid_sampled_token_count_copy_stream: torch.cuda.Stream | None = None
        if self.use_async_scheduling and self.num_spec_tokens:
597
            self.valid_sampled_token_count_event = torch.Event()
598
599
600
601
602
603
604
605
            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,
        )

606
607
        # Ephemeral state transferred between execute_model() and sample_tokens().
        self.execute_model_state: ExecuteModelState | None = None
608
        self.kv_connector_output: KVConnectorOutput | None = None
609
        self.layerwise_nvtx_hooks_registered = False
610

611
612
613
614
    def reset_mm_cache(self) -> None:
        if self.mm_budget:
            self.mm_budget.reset_cache()

615
616
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
    @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)

659
660
661
662
    def _get_positions(self, num_tokens: Any):
        if isinstance(num_tokens, int):
            if self.uses_mrope:
                return self.mrope_positions.gpu[:, :num_tokens]
663
664
            if self.uses_xdrope_dim > 0:
                return self.xdrope_positions.gpu[:, :num_tokens]
665
666
667
668
            return self.positions.gpu[:num_tokens]
        else:
            if self.uses_mrope:
                return self.mrope_positions.gpu[:, num_tokens]
669
670
            if self.uses_xdrope_dim > 0:
                return self.xdrope_positions.gpu[:, num_tokens]
671
672
            return self.positions.gpu[num_tokens]

673
    def _make_buffer(
674
        self, *size: int | torch.SymInt, dtype: torch.dtype, numpy: bool = True
675
676
677
678
679
680
681
682
    ) -> CpuGpuBuffer:
        return CpuGpuBuffer(
            *size,
            dtype=dtype,
            device=self.device,
            pin_memory=self.pin_memory,
            with_numpy=numpy,
        )
683

684
685
686
    def _init_model_kwargs(self, num_tokens: int):
        model_kwargs = dict[str, Any]()

687
        if not self.is_pooling_model:
688
689
            return model_kwargs

690
691
        num_reqs = self.input_batch.num_reqs
        pooling_params = self.input_batch.get_pooling_params()
692
693
694

        token_type_id_requests = dict[int, Any]()
        for i, param in enumerate(pooling_params):
695
696
697
698
699
            if (
                param.extra_kwargs is not None
                and (token_types := param.extra_kwargs.get("compressed_token_type_ids"))
                is not None
            ):
700
701
702
703
704
                token_type_id_requests[i] = token_types

        if len(token_type_id_requests) == 0:
            return model_kwargs

705
        seq_lens = self.seq_lens.gpu[:num_reqs]
706
707
708
709
710
711
712
713
        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(
714
715
            device=self.device
        )
716
717
        return model_kwargs

718
    def _may_reorder_batch(self, scheduler_output: "SchedulerOutput") -> None:
719
720
        """
        Update the order of requests in the batch based on the attention
721
        backend's needs. For example, some attention backends (namely MLA) may
722
723
724
725
726
727
        want to separate requests based on if the attention computation will be
        compute-bound or memory-bound.

        Args:
            scheduler_output: The scheduler output.
        """
728
729
730
731
732
733
734
735
        # 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

736
737
738
739
        if self.reorder_batch_threshold is not None:
            reorder_batch_to_split_decodes_and_prefills(
                self.input_batch,
                scheduler_output,
740
741
                decode_threshold=self.reorder_batch_threshold,
            )
742

743
744
    # Note: used for model runner override.
    def _init_device_properties(self) -> None:
745
        """Initialize attributes from torch.cuda.get_device_properties"""
746
747
748
749
750
751
752
        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()

753
    def _update_states(self, scheduler_output: "SchedulerOutput") -> None:
754
755
756
757
758
759
        """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.

760
761
        The SamplingMetadata is updated and copied to the GPU if there is a
        new/resumed/paused/finished request in the batch.
762
763
        """
        # Remove finished requests from the cached states.
764
765
        for req_id in scheduler_output.finished_req_ids:
            self.requests.pop(req_id, None)
766
            self.num_prompt_logprobs.pop(req_id, None)
767
768
769
770
771
772
773
        # 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:
774
            self.input_batch.remove_request(req_id)
775
776

        # Free the cached encoder outputs.
777
778
        for mm_hash in scheduler_output.free_encoder_mm_hashes:
            self.encoder_cache.pop(mm_hash, None)
779

780
781
782
783
784
785
786
        # 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()
787
788
789
790
791
792
793
794
        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)
795
796
797
798
799
        # 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:
800
            self.input_batch.remove_request(req_id)
801

802
        reqs_to_add: list[CachedRequestState] = []
803
        # Add new requests to the cached states.
804
805
806
        for new_req_data in scheduler_output.scheduled_new_reqs:
            req_id = new_req_data.req_id
            sampling_params = new_req_data.sampling_params
807
            pooling_params = new_req_data.pooling_params
808

809
810
811
812
            if (
                sampling_params
                and sampling_params.sampling_type == SamplingType.RANDOM_SEED
            ):
813
814
815
816
817
                generator = torch.Generator(device=self.device)
                generator.manual_seed(sampling_params.seed)
            else:
                generator = None

818
819
            if self.is_pooling_model:
                assert pooling_params is not None
820
821
                task = pooling_params.task
                assert task is not None, "You did not set `task` in the API"
822

823
                model = cast(VllmModelForPooling, self.get_model())
824
                to_update = model.pooler.get_pooling_updates(task)
825
826
                to_update.apply(pooling_params)

827
            req_state = CachedRequestState(
828
                req_id=req_id,
829
                prompt_token_ids=new_req_data.prompt_token_ids,
830
                prompt_embeds=new_req_data.prompt_embeds,
831
                mm_features=new_req_data.mm_features,
832
                sampling_params=sampling_params,
833
                pooling_params=pooling_params,
834
                generator=generator,
835
836
                block_ids=new_req_data.block_ids,
                num_computed_tokens=new_req_data.num_computed_tokens,
837
                output_token_ids=[],
838
                lora_request=new_req_data.lora_request,
839
            )
840
841
            self.requests[req_id] = req_state

842
843
844
845
846
847
848
            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
                )

849
            # Only relevant for models using M-RoPE (e.g, Qwen2-VL)
850
            if self.uses_mrope:
851
                self._init_mrope_positions(req_state)
852

853
854
855
856
            # Only relevant for models using XD-RoPE (e.g, HunYuan-VL)
            if self.uses_xdrope_dim > 0:
                self._init_xdrope_positions(req_state)

857
            reqs_to_add.append(req_state)
858

859
        # Update the states of the running/resumed requests.
860
        is_last_rank = get_pp_group().is_last_rank
861
        req_data = scheduler_output.scheduled_cached_reqs
862
863
864
865
866

        # 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()

867
        for i, req_id in enumerate(req_data.req_ids):
868
            req_state = self.requests[req_id]
869
870
            num_computed_tokens = req_data.num_computed_tokens[i]
            new_block_ids = req_data.new_block_ids[i]
871
            resumed_from_preemption = req_id in req_data.resumed_req_ids
872
            num_output_tokens = req_data.num_output_tokens[i]
873
            req_index = self.input_batch.req_id_to_index.get(req_id)
874

875
876
877
878
879
880
881
882
883
884
885
886
887
888
889
890
891
892
893
894
895
896
897
            # 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)
898

899
            # Update the cached states.
900
            req_state.num_computed_tokens = num_computed_tokens
901
902
903
904
905
906
907
908

            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.
909
910
911
                num_new_tokens = (
                    num_computed_tokens + len(new_token_ids) - req_state.num_tokens
                )
912
913
914
915
                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:
916
                    req_state.output_token_ids.extend(new_token_ids[-num_new_tokens:])
917
918
919
920
921
            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:
922
923
924
925
                    end_idx = (
                        self.input_batch.num_prompt_tokens[req_index]
                        + num_output_tokens
                    )
926
                    self.input_batch.num_tokens_no_spec[req_index] = end_idx
927

928
            # Update the block IDs.
929
            if not resumed_from_preemption:
930
931
                if new_block_ids is not None:
                    # Append the new blocks to the existing block IDs.
932
                    for block_ids, new_ids in zip(req_state.block_ids, new_block_ids):
933
                        block_ids.extend(new_ids)
934
            else:
935
                assert req_index is None
936
                assert new_block_ids is not None
937
938
                # The request is resumed from preemption.
                # Replace the existing block IDs with the new ones.
939
                req_state.block_ids = new_block_ids
940
941
942
943
944

            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.
945
946
947
948
949
950
951

                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:]

952
                reqs_to_add.append(req_state)
953
954
955
                continue

            # Update the persistent batch.
956
            self.input_batch.num_computed_tokens_cpu[req_index] = num_computed_tokens
957
            if new_block_ids is not None:
958
                self.input_batch.block_table.append_row(new_block_ids, req_index)
959
960
961
962
963
964
965

            # 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)
966
                self.input_batch.token_ids_cpu[
967
968
969
                    req_index, start_token_index:end_token_index
                ] = new_token_ids
                self.input_batch.num_tokens_no_spec[req_index] = end_token_index
970

971
            # Add spec_token_ids to token_ids_cpu.
972
            spec_token_ids = scheduler_output.scheduled_spec_decode_tokens.get(
973
                req_id, []
974
            )
975
976
977
978
979
            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:
980
981
982
                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[
983
984
                    req_index, start_index:end_token_index
                ] = spec_token_ids
985
986
987
988
989
990

            # 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.
991
992
            self.input_batch.spec_token_ids[req_index].clear()
            self.input_batch.spec_token_ids[req_index].extend(spec_token_ids)
993

994
995
996
997
998
999
1000
1001
1002
            # 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)
1003
1004
        # Add the new or resumed requests to the persistent batch.
        # The smaller empty indices are filled first.
1005
1006
        for request in reqs_to_add:
            self.input_batch.add_request(request)
1007

1008
1009
1010
1011
1012
1013
        # 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()
1014

1015
    def _update_states_after_model_execute(
1016
1017
        self, output_token_ids: torch.Tensor
    ) -> None:
1018
1019
1020
1021
1022
1023
1024
1025
1026
1027
1028
1029
        """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.
1030
1031
1032
1033
1034
1035
1036
1037
1038
1039
1040
1041
1042
1043
1044
1045
1046
1047
1048
1049
        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()
        )
1050
1051
1052
        for i, num_tokens in enumerate(num_accepted_tokens):
            self.input_batch.num_accepted_tokens_cpu[i] = num_tokens

1053
    def _init_mrope_positions(self, req_state: CachedRequestState):
1054
1055
        model = self.get_model()
        assert supports_mrope(model), "M-RoPE support is not implemented."
1056
1057
1058
1059
        assert req_state.prompt_token_ids is not None, (
            "M-RoPE requires prompt_token_ids to be available."
        )
        mrope_model = cast(SupportsMRoPE, model)
1060
1061

        req_state.mrope_positions, req_state.mrope_position_delta = (
1062
            mrope_model.get_mrope_input_positions(
1063
                req_state.prompt_token_ids,
1064
                req_state.mm_features,
1065
            )
1066
        )
1067

1068
1069
1070
1071
1072
1073
1074
1075
1076
1077
1078
1079
1080
    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,
        )

1081
    def _extract_mm_kwargs(
1082
        self,
1083
1084
        scheduler_output: "SchedulerOutput",
    ) -> BatchedTensorInputs:
1085
        if not scheduler_output or not self.is_multimodal_raw_input_only_model:
1086
            return {}
1087

1088
1089
        mm_kwargs = list[MultiModalKwargsItem]()
        for req in scheduler_output.scheduled_new_reqs:
1090
1091
1092
            for feature in req.mm_features:
                if feature.data is not None:
                    mm_kwargs.append(feature.data)
1093

1094
        # Input all modalities at once
1095
        model = cast(SupportsMultiModal, self.model)
1096
1097
        mm_kwargs_combined: BatchedTensorInputs = {}
        for _, _, mm_kwargs_group in group_mm_kwargs_by_modality(
1098
1099
1100
1101
            mm_kwargs,
            device=self.device,
            pin_memory=self.pin_memory,
            merge_by_field_config=model.merge_by_field_config,
1102
1103
        ):
            mm_kwargs_combined.update(mm_kwargs_group)
1104

1105
        return mm_kwargs_combined
1106

1107
    def _dummy_mm_kwargs(self, num_seqs: int) -> BatchedTensorInputs:
1108
        if not self.is_multimodal_raw_input_only_model:
1109
            return {}
1110

1111
1112
1113
1114
1115
        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)
1116

1117
1118
1119
    def _get_cumsum_and_arange(
        self,
        num_tokens: np.ndarray,
1120
        cumsum_dtype: np.dtype | None = None,
1121
1122
1123
1124
1125
1126
1127
1128
1129
1130
1131
1132
1133
1134
1135
1136
    ) -> 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

1137
    def _prepare_input_ids(
1138
1139
1140
1141
        self,
        scheduler_output: "SchedulerOutput",
        total_num_scheduled_tokens: int,
        cu_num_tokens: np.ndarray,
1142
    ) -> None:
1143
        """Prepare the input IDs for the current batch.
1144

1145
1146
1147
1148
1149
1150
1151
        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)
1152
1153
1154
            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)
1155
1156
1157
1158
1159
1160
1161
            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
1162
1163
1164
1165
        sample_flattened_indices: list[int] = []
        spec_flattened_indices: list[int] = []
        prev_common_req_indices: list[int] = []
        prev_draft_token_indices: list[int] = []
1166
1167
        indices_match = True
        max_flattened_index = -1
1168
1169
1170
        total_num_spec_tokens = 0
        scheduled_spec_tokens = scheduler_output.scheduled_spec_decode_tokens

1171
1172
1173
1174
1175
        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.
1176
1177
                draft_len = len(scheduled_spec_tokens.get(req_id, ()))
                total_num_spec_tokens += draft_len
1178
                flattened_index = cu_num_tokens[cur_index].item() - 1
1179
1180
1181
1182
1183
1184
1185
1186
1187
1188
1189
1190
1191
1192
1193
                # 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))
1194
                indices_match &= prev_index == flattened_index
1195
                max_flattened_index = max(max_flattened_index, flattened_index)
1196
1197
1198
        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:
1199
1200
1201
            # 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)
1202
1203
1204
            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)
1205
1206
        if num_commmon_tokens == 0:
            # No requests in common with the previous iteration
1207
            # So input_ids.cpu will have all the input ids.
1208
1209
1210
1211
1212
1213
1214
            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_(
1215
1216
1217
                self.input_batch.prev_sampled_token_ids[:num_commmon_tokens, 0],
                non_blocking=True,
            )
1218
1219
            if self.enable_prompt_embeds:
                self.is_token_ids.gpu[:num_commmon_tokens] = True
1220
            return
1221
        # Upload the index tensors asynchronously so the scatter can be non-blocking.
1222
1223
        sampled_tokens_index_tensor = torch.tensor(
            sample_flattened_indices, dtype=torch.int64, pin_memory=self.pin_memory
1224
        ).to(self.device, non_blocking=True)
1225
        prev_common_req_indices_tensor = torch.tensor(
1226
1227
            prev_common_req_indices, dtype=torch.int64, pin_memory=self.pin_memory
        ).to(self.device, non_blocking=True)
1228
1229
        self.input_ids.gpu.scatter_(
            dim=0,
1230
            index=sampled_tokens_index_tensor,
1231
            src=self.input_batch.prev_sampled_token_ids[
1232
1233
1234
                prev_common_req_indices_tensor, 0
            ],
        )
1235

1236
1237
1238
1239
1240
1241
1242
1243
1244
1245
1246
1247
1248
1249
1250
1251
1252
1253
1254
1255
1256
1257
1258
        # 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],
        )

1259
1260
    def _get_encoder_seq_lens(
        self,
1261
        num_scheduled_tokens: dict[str, int],
1262
1263
        kv_cache_spec: KVCacheSpec,
        num_reqs: int,
1264
    ) -> tuple[torch.Tensor | None, np.ndarray | None]:
1265
        if not isinstance(kv_cache_spec, CrossAttentionSpec):
1266
            return None, None
1267

1268
1269
        # Zero out buffer for padding requests that are not actually scheduled (CGs)
        self.encoder_seq_lens.np[:num_reqs] = 0
1270
1271
        # Build encoder_seq_lens array mapping request indices to
        # encoder lengths for inputs scheduled in this batch
1272
        for req_id in num_scheduled_tokens:
1273
            req_index = self.input_batch.req_id_to_index[req_id]
1274
1275
1276
1277
1278
1279
1280
1281
1282
1283
1284
1285
1286
1287
1288
1289
            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]
1290

1291
        return encoder_seq_lens, encoder_seq_lens_cpu
1292

1293
    def _prepare_inputs(
1294
1295
1296
        self,
        scheduler_output: "SchedulerOutput",
        num_scheduled_tokens: np.ndarray,
1297
1298
    ) -> tuple[
        torch.Tensor,
1299
        SpecDecodeMetadata | None,
1300
    ]:
1301
1302
        """
        :return: tuple[
1303
            logits_indices, spec_decode_metadata,
1304
1305
        ]
        """
1306
1307
1308
1309
1310
1311
1312
        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.
1313
        self.input_batch.block_table.commit_block_table(num_reqs)
1314
1315
1316

        # Get request indices.
        # E.g., [2, 5, 3] -> [0, 0, 1, 1, 1, 1, 1, 2, 2, 2]
1317
        req_indices = np.repeat(self.arange_np[:num_reqs], num_scheduled_tokens)
1318

1319
1320
        # cu_num_tokens: [2, 5, 3] -> [2, 7, 10]
        # arange: [0, 1, 0, 1, 2, 3, 4, 0, 1, 2]
1321
        cu_num_tokens, arange = self._get_cumsum_and_arange(num_scheduled_tokens)
1322
1323

        # Get positions.
1324
        positions_np = self.positions.np[:total_num_scheduled_tokens]
1325
1326
1327
1328
1329
        np.add(
            self.input_batch.num_computed_tokens_cpu[req_indices],
            arange,
            out=positions_np,
        )
1330

1331
1332
        # Calculate M-RoPE positions.
        # Only relevant for models using M-RoPE (e.g, Qwen2-VL)
1333
        if self.uses_mrope:
1334
1335
            self._calc_mrope_positions(scheduler_output)

1336
1337
1338
1339
1340
        # 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)

1341
1342
1343
1344
        # 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.
1345
1346
1347
        token_indices = (
            positions_np + req_indices * self.input_batch.token_ids_cpu.shape[1]
        )
1348
        token_indices_tensor = torch.from_numpy(token_indices)
1349

1350
1351
1352
        # 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.
1353
1354
1355
1356
1357
1358
        torch.index_select(
            self.input_batch.token_ids_cpu_tensor.flatten(),
            0,
            token_indices_tensor,
            out=self.input_ids.cpu[:total_num_scheduled_tokens],
        )
1359
        if self.enable_prompt_embeds:
1360
            is_token_ids = self.input_batch.is_token_ids_tensor.flatten()
1361
1362
1363
1364
            torch.index_select(
                is_token_ids,
                0,
                token_indices_tensor,
1365
1366
                out=self.is_token_ids.cpu[:total_num_scheduled_tokens],
            )
1367
1368
1369
1370
1371
1372
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

        # 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:
1400
1401
1402
                    self.inputs_embeds.cpu[
                        output_idx : output_idx + actual_num_sched
                    ].copy_(req_embeds[start_pos:actual_end])
1403
1404

                output_idx += num_sched
1405

1406
1407
        self.input_batch.block_table.compute_slot_mapping(req_indices, positions_np)
        self.input_batch.block_table.commit_slot_mapping(total_num_scheduled_tokens)
1408
1409

        # Prepare the attention metadata.
1410
        self.query_start_loc.np[0] = 0
1411
        self.query_start_loc.np[1 : num_reqs + 1] = cu_num_tokens
1412
1413
        # Note: pad query_start_loc to be non-decreasing, as kernels
        # like FlashAttention requires that
1414
        self.query_start_loc.np[num_reqs + 1 :].fill(cu_num_tokens[-1])
1415
        self.query_start_loc.copy_to_gpu()
1416
        query_start_loc = self.query_start_loc.gpu[: num_reqs + 1]
1417

1418
        self.seq_lens.np[:num_reqs] = (
1419
1420
            self.input_batch.num_computed_tokens_cpu[:num_reqs] + num_scheduled_tokens
        )
1421
        # Fill unused with 0 for full cuda graph mode.
1422
1423
        self.seq_lens.np[num_reqs:].fill(0)
        self.seq_lens.copy_to_gpu()
1424

1425
        num_tokens = [self.requests[r].num_tokens for r in self.input_batch.req_ids]
1426
1427
        num_tokens_np = np.array(num_tokens, dtype=np.int32)

1428
        # Record which requests should not be sampled,
1429
        # so that we could clear the sampled tokens before returning
1430
1431
        self.discard_request_mask.np[:num_reqs] = (
            self.seq_lens.np[:num_reqs] < num_tokens_np
1432
        )
1433
        self.discard_request_mask.copy_to_gpu(num_reqs)
1434

1435
        # Copy the tensors to the GPU.
1436
1437
1438
1439
1440
        self._prepare_input_ids(
            scheduler_output,
            total_num_scheduled_tokens,
            cu_num_tokens,
        )
1441

1442
        if self.uses_mrope:
1443
            # Only relevant for models using M-RoPE (e.g, Qwen2-VL)
1444
1445
            self.mrope_positions.gpu[:, :total_num_scheduled_tokens].copy_(
                self.mrope_positions.cpu[:, :total_num_scheduled_tokens],
1446
1447
                non_blocking=True,
            )
1448
1449
1450
1451
1452
1453
        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,
            )
1454
1455
        else:
            # Common case (1D positions)
1456
            self.positions.copy_to_gpu(total_num_scheduled_tokens)
1457

1458
        use_spec_decode = len(scheduler_output.scheduled_spec_decode_tokens) > 0
1459
1460
1461
1462
1463
1464
1465
        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
1466
            num_draft_tokens = None
1467
            spec_decode_metadata = None
1468
            num_sampled_tokens = np.ones(num_reqs, dtype=np.int32)
1469
1470
1471
1472
1473
        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)
1474
1475
1476
            # 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)
1477
1478
1479
1480
            for (
                req_id,
                draft_token_ids,
            ) in scheduler_output.scheduled_spec_decode_tokens.items():
1481
1482
                req_idx = self.input_batch.req_id_to_index[req_id]
                num_draft_tokens[req_idx] = len(draft_token_ids)
1483
1484
1485
1486
1487
1488
1489
1490
                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
                )
1491
            spec_decode_metadata = self._calc_spec_decode_metadata(
1492
1493
                num_draft_tokens, cu_num_tokens
            )
1494
            logits_indices = spec_decode_metadata.logits_indices
1495
            num_sampled_tokens = num_draft_tokens + 1
1496
            # For DECODE only cuda graph of some attention backends (e.g., GDN).
1497
            self.num_decode_draft_tokens.np[:num_reqs] = num_decode_draft_tokens
1498
1499
            self.num_decode_draft_tokens.np[num_reqs:].fill(-1)
            self.num_decode_draft_tokens.copy_to_gpu()
1500

1501
1502
1503
1504
1505
        # Hot-Swap lora model
        if self.lora_config:
            assert (
                np.sum(num_sampled_tokens)
                <= self.vllm_config.scheduler_config.max_num_batched_tokens
1506
            )
1507
1508
1509
1510
1511
1512
1513
1514
1515
1516
1517
            self.set_active_loras(
                self.input_batch, num_scheduled_tokens, num_sampled_tokens
            )

        return (
            logits_indices,
            spec_decode_metadata,
        )

    def _build_attention_metadata(
        self,
1518
        num_tokens: int,
1519
        num_reqs: int,
1520
1521
1522
        max_query_len: int,
        num_tokens_padded: int | None = None,
        num_reqs_padded: int | None = None,
1523
1524
1525
1526
        ubatch_slices: UBatchSlices | None = None,
        logits_indices: torch.Tensor | None = None,
        use_spec_decode: bool = False,
        for_cudagraph_capture: bool = False,
1527
        num_scheduled_tokens: dict[str, int] | None = None,
1528
1529
1530
1531
1532
        cascade_attn_prefix_lens: list[list[int]] | None = None,
    ) -> tuple[PerLayerAttnMetadata, CommonAttentionMetadata | None]:
        """
        :return: tuple[attn_metadata, spec_decode_common_attn_metadata]
        """
1533
1534
1535
1536
        # Attention metadata is not needed for attention free models
        if len(self.kv_cache_config.kv_cache_groups) == 0:
            return {}, None

1537
1538
        num_tokens_padded = num_tokens_padded or num_tokens
        num_reqs_padded = num_reqs_padded or num_reqs
1539
        assert num_reqs_padded is not None and num_tokens_padded is not None
1540

1541
1542
1543
        attn_metadata: PerLayerAttnMetadata = {}
        if ubatch_slices is not None:
            attn_metadata = [dict() for _ in range(len(ubatch_slices))]
1544

1545
1546
1547
1548
1549
1550
1551
1552
        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()

1553
1554
        if use_spec_decode:
            self.num_accepted_tokens.np[:num_reqs] = (
1555
1556
                self.input_batch.num_accepted_tokens_cpu[:num_reqs]
            )
1557
1558
            self.num_accepted_tokens.np[num_reqs:].fill(1)
            self.num_accepted_tokens.copy_to_gpu()
1559

1560
        kv_cache_groups = self.kv_cache_config.kv_cache_groups
1561

1562
1563
1564
1565
        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):
1566
                blk_table_tensor = torch.zeros(
1567
                    (num_reqs_padded, 1),
1568
                    dtype=torch.int32,
1569
1570
1571
                    device=self.device,
                )
                slot_mapping = torch.zeros(
1572
                    (num_tokens_padded,),
1573
1574
1575
                    dtype=torch.int64,
                    device=self.device,
                )
1576
            else:
1577
                blk_table = self.input_batch.block_table[kv_cache_gid]
1578
1579
                blk_table_tensor = blk_table.get_device_tensor(num_reqs_padded)
                slot_mapping = blk_table.slot_mapping.gpu[:num_tokens_padded]
1580

1581
1582
1583
1584
1585
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
            # 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
            )

1627
1628
1629
1630
1631
1632
1633
1634
1635
        # Cache attention metadata builds across hybrid KV-cache groups
        # The only thing that changes between different hybrid KV-cache groups when the
        # same metadata builder and KVCacheSpec is the same is the block table, so we
        # can cache the attention metadata builds and just update the block table using
        # `builder.update_block_table` if the builder supports it.
        cached_attn_metadata: dict[
            tuple[KVCacheSpec, type[AttentionMetadataBuilder]], AttentionMetadata
        ] = {}

1636
1637
1638
1639
1640
1641
1642
        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]
1643
            builder = attn_group.get_metadata_builder(ubid or 0)
1644
1645
1646
1647
            kv_cache_spec = kv_cache_groups[kv_cache_gid].kv_cache_spec
            if isinstance(kv_cache_spec, UniformTypeKVCacheSpecs):
                kv_cache_spec = kv_cache_spec.kv_cache_specs[attn_group.layer_names[0]]
            cache_key = (kv_cache_spec, type(builder))
1648

1649
1650
1651
1652
1653
1654
1655
1656
1657
1658
1659
1660
1661
1662
1663
1664
1665
1666
1667
1668
            cascade_attn_prefix_len = (
                cascade_attn_prefix_lens[kv_cache_gid][attn_gid]
                if cascade_attn_prefix_lens
                else 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
                )
1669
1670
1671
1672
1673
1674
1675
1676
1677
            elif (
                cache_key in cached_attn_metadata
                and builder.supports_update_block_table
            ):
                attn_metadata_i = builder.update_block_table(
                    cached_attn_metadata[cache_key],
                    common_attn_metadata.block_table_tensor,
                    common_attn_metadata.slot_mapping,
                )
1678
1679
1680
1681
1682
1683
            else:
                attn_metadata_i = builder.build(
                    common_prefix_len=cascade_attn_prefix_len,
                    common_attn_metadata=common_attn_metadata,
                    **extra_attn_metadata_args,
                )
1684
1685
                if builder.supports_update_block_table:
                    cached_attn_metadata[cache_key] = attn_metadata_i
1686
1687
1688
1689
1690
1691
1692
1693
1694
1695
1696
1697
1698
1699
1700
1701
1702
1703
1704
1705
1706
1707
1708

            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,
1709
            )
1710
1711
1712
1713
            if kv_cache_gid > 0:
                cm.block_table_tensor, cm.slot_mapping = (
                    _get_block_table_and_slot_mapping(kv_cache_gid)
                )
1714

1715
            if self.speculative_config and spec_decode_common_attn_metadata is None:
1716
                if isinstance(self.drafter, EagleProposer):
1717
                    if self.drafter.attn_layer_names[0] in kv_cache_group.layer_names:
1718
                        spec_decode_common_attn_metadata = cm
1719
                else:
1720
                    spec_decode_common_attn_metadata = cm
1721

1722
            for attn_gid in range(len(self.attn_groups[kv_cache_gid])):
1723
                if ubatch_slices is not None:
1724
1725
1726
                    for ubid, _cm in enumerate(split_attn_metadata(ubatch_slices, cm)):
                        _build_attn_group_metadata(kv_cache_gid, attn_gid, _cm, ubid)

1727
                else:
1728
                    _build_attn_group_metadata(kv_cache_gid, attn_gid, cm)
1729

1730
1731
1732
1733
1734
1735
1736
1737
1738
1739
1740
1741
1742
1743
1744
1745
1746
1747
1748
1749
        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]

1750
1751
1752
1753
1754
1755
1756
1757
1758
1759
        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)
            )

1760
        return attn_metadata, spec_decode_common_attn_metadata
1761

1762
1763
1764
    def _compute_cascade_attn_prefix_lens(
        self,
        num_scheduled_tokens: np.ndarray,
1765
        num_computed_tokens: np.ndarray,
1766
1767
1768
1769
1770
1771
1772
        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
        """
1773

1774
1775
1776
1777
1778
1779
1780
1781
1782
1783
1784
1785
1786
1787
        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,
1788
                        num_computed_tokens,
1789
1790
1791
1792
1793
1794
1795
1796
                        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
1797

1798
1799
1800
    def _compute_cascade_attn_prefix_len(
        self,
        num_scheduled_tokens: np.ndarray,
1801
        num_computed_tokens: np.ndarray,
1802
        num_common_prefix_blocks: int,
1803
1804
        kv_cache_spec: KVCacheSpec,
        attn_metadata_builder: AttentionMetadataBuilder,
1805
1806
1807
1808
1809
1810
1811
1812
1813
1814
1815
1816
1817
1818
1819
1820
1821
1822
    ) -> 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.
        """
1823

1824
        common_prefix_len = num_common_prefix_blocks * kv_cache_spec.block_size
1825
1826
1827
1828
1829
1830
1831
1832
1833
1834
1835
1836
1837
1838
1839
1840
1841
1842
1843
1844
1845
1846
1847
1848
1849
1850
1851
1852
1853
1854
1855
1856
1857
1858
1859
1860
1861
        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]
1862
        # Request 3's num_computed_tokens: 3 (i.e., [A, B, C])
1863
1864
1865
1866
1867
        # 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.
1868
        common_prefix_len = min(common_prefix_len, num_computed_tokens.min())
1869
        # common_prefix_len should be a multiple of the block size.
1870
1871
1872
1873
1874
1875
1876
1877
1878
1879
1880
        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
        )
1881
1882
        assert isinstance(kv_cache_spec, AttentionSpec)
        use_cascade = attn_metadata_builder.use_cascade_attention(
1883
1884
1885
            common_prefix_len=common_prefix_len,
            query_lens=num_scheduled_tokens,
            num_query_heads=self.num_query_heads,
1886
            num_kv_heads=kv_cache_spec.num_kv_heads,
1887
            use_alibi=self.use_alibi,
1888
            use_sliding_window=use_sliding_window,
1889
            use_local_attention=use_local_attention,
1890
            num_sms=self.num_sms,
1891
            dcp_world_size=self.dcp_world_size,
1892
1893
1894
        )
        return common_prefix_len if use_cascade else 0

1895
1896
    def _calc_mrope_positions(self, scheduler_output: "SchedulerOutput"):
        mrope_pos_ptr = 0
1897
        for index, req_id in enumerate(self.input_batch.req_ids):
1898
1899
1900
            req = self.requests[req_id]
            assert req.mrope_positions is not None

1901
1902
            num_computed_tokens = self.input_batch.num_computed_tokens_cpu[index]
            num_scheduled_tokens = scheduler_output.num_scheduled_tokens[req_id]
1903
            num_prompt_tokens = length_from_prompt_token_ids_or_embeds(
1904
1905
                req.prompt_token_ids, req.prompt_embeds
            )
1906
1907

            if num_computed_tokens + num_scheduled_tokens > num_prompt_tokens:
1908
1909
                prompt_part_len = max(0, num_prompt_tokens - num_computed_tokens)
                completion_part_len = max(0, num_scheduled_tokens - prompt_part_len)
1910
1911
1912
1913
1914
1915
1916
1917
1918
1919
1920
1921
1922
            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

1923
1924
1925
                self.mrope_positions.cpu[:, dst_start:dst_end] = req.mrope_positions[
                    :, src_start:src_end
                ]
1926
1927
1928
1929
1930
1931
1932
                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

1933
                assert req.mrope_position_delta is not None
1934
                MRotaryEmbedding.get_next_input_positions_tensor(
1935
                    out=self.mrope_positions.np,
1936
1937
1938
1939
1940
                    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,
                )
1941
1942
1943

                mrope_pos_ptr += completion_part_len

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
1972
1973
1974
1975
1976
1977
1978
1979
1980
1981
1982
1983
1984
1985
1986
1987
1988
1989
1990
    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

1991
1992
    def _calc_spec_decode_metadata(
        self,
1993
1994
1995
1996
1997
1998
1999
2000
2001
2002
2003
2004
2005
2006
2007
2008
        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
2009
2010
2011
2012

        # 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(
2013
2014
            num_sampled_tokens, cumsum_dtype=np.int32
        )
2015
        # Step 2. [0, 0, 0, 0, 103, 104, 104, 104, 206, 207, 207]
2016
        logits_indices = np.repeat(
2017
2018
            cu_num_scheduled_tokens - num_sampled_tokens, num_sampled_tokens
        )
2019
        # Step 3. [0, 1, 2, 3, 103, 104, 105, 106, 206, 207, 208]
2020
2021
2022
2023
2024
2025
        logits_indices += arange

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

        # Compute the draft logits indices.
2026
2027
2028
        # 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(
2029
2030
            num_draft_tokens, cumsum_dtype=np.int32
        )
2031
2032
        # [0, 0, 0, 5, 5, 9]
        target_logits_indices = np.repeat(
2033
2034
            cu_num_sampled_tokens - num_sampled_tokens, num_draft_tokens
        )
2035
2036
2037
2038
2039
        # [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(
2040
2041
            self.device, non_blocking=True
        )
2042
2043
2044
        cu_num_sampled_tokens = torch.from_numpy(cu_num_sampled_tokens).to(
            self.device, non_blocking=True
        )
2045
2046
2047
        logits_indices = torch.from_numpy(logits_indices).to(
            self.device, non_blocking=True
        )
2048
        target_logits_indices = torch.from_numpy(target_logits_indices).to(
2049
2050
            self.device, non_blocking=True
        )
2051
        bonus_logits_indices = torch.from_numpy(bonus_logits_indices).to(
2052
2053
            self.device, non_blocking=True
        )
2054

2055
2056
        # Compute the draft token ids.
        # draft_token_indices:      [  1,   2,   3, 105, 106, 208]
2057
        draft_token_ids = self.input_ids.gpu[logits_indices]
2058
2059
        draft_token_ids = draft_token_ids[target_logits_indices + 1]

2060
        return SpecDecodeMetadata(
2061
2062
2063
            draft_token_ids=draft_token_ids,
            num_draft_tokens=num_draft_tokens.tolist(),
            cu_num_draft_tokens=cu_num_draft_tokens,
2064
            cu_num_sampled_tokens=cu_num_sampled_tokens,
2065
2066
2067
2068
2069
            target_logits_indices=target_logits_indices,
            bonus_logits_indices=bonus_logits_indices,
            logits_indices=logits_indices,
        )

2070
2071
2072
2073
2074
2075
2076
    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
2077
        self.kv_sharing_fast_prefill_logits_indices[:num_logits].copy_(logits_indices)
2078
2079
2080
2081
2082
        # 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_(
2083
2084
2085
2086
2087
2088
            logits_indices[-1].item()
        )
        if (
            self.compilation_config.cudagraph_mode != CUDAGraphMode.NONE
            and num_logits <= self.cudagraph_batch_sizes[-1]
        ):
2089
2090
2091
2092
2093
            # 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
2094
2095
2096
        logits_indices_padded = self.kv_sharing_fast_prefill_logits_indices[
            :num_logits_padded
        ]
2097
2098
        return logits_indices_padded

2099
2100
2101
2102
2103
2104
2105
2106
    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
2107
                inputs.
2108
2109
2110
2111
2112
2113

        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
        """
2114
2115
        scheduled_encoder_inputs = scheduler_output.scheduled_encoder_inputs
        if not scheduled_encoder_inputs:
2116
            return [], []
2117
        # Batch the multi-modal inputs.
2118
        mm_kwargs = list[MultiModalKwargsItem]()
2119
2120
        # list of tuple (mm_hash, position_info)
        mm_hashes_pos = list[tuple[str, PlaceholderRange]]()
2121
2122
        for req_id, encoder_input_ids in scheduled_encoder_inputs.items():
            req_state = self.requests[req_id]
2123
2124

            for mm_input_id in encoder_input_ids:
2125
                mm_feature = req_state.mm_features[mm_input_id]
2126
2127
                if mm_feature.data is None:
                    continue
2128
2129
2130
                mm_hash = mm_feature.identifier
                mm_kwargs.append(mm_feature.data)
                mm_hashes_pos.append((mm_hash, mm_feature.mm_position))
2131

2132
2133
        return mm_kwargs, mm_hashes_pos

2134
2135
2136
    def _execute_mm_encoder(
        self, scheduler_output: "SchedulerOutput"
    ) -> list[torch.Tensor]:
2137
2138
        # Batch the multi-modal inputs using the helper method.
        mm_kwargs, mm_hashes_pos = self._batch_mm_kwargs_from_scheduler(
2139
2140
            scheduler_output
        )
2141
2142

        if not mm_kwargs:
2143
            return []
2144

2145
2146
2147
2148
2149
2150
2151
        # 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.
2152
        model = cast(SupportsMultiModal, self.model)
2153
        encoder_outputs: list[torch.Tensor] = []
2154
        for modality, num_items, mm_kwargs_group in group_mm_kwargs_by_modality(
2155
2156
2157
            mm_kwargs,
            device=self.device,
            pin_memory=self.pin_memory,
2158
        ):
2159
            curr_group_outputs: list[torch.Tensor] = []
2160
2161

            # EVS-related change.
2162
            # (ekhvedchenia): Temporary hack to limit peak memory usage when
2163
            # processing multimodal data. This solves the issue with scheduler
2164
2165
2166
2167
            # 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)
2168
2169
2170
2171
2172
2173
2174
2175
2176
2177
2178
2179
2180
2181
2182
2183
            # 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,
                        )
2184
                    )
2185

2186
                    micro_batch_outputs = model.embed_multimodal(
2187
2188
                        **micro_batch_mm_inputs
                    )
2189
2190
2191
2192
2193
2194
2195
2196
2197
2198

                    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.
2199
                curr_group_outputs = model.embed_multimodal(**mm_kwargs_group)  # type: ignore[assignment]
2200

2201
2202
            sanity_check_mm_encoder_outputs(
                curr_group_outputs,
2203
                expected_num_items=num_items,
2204
            )
2205
            encoder_outputs.extend(curr_group_outputs)
2206

2207
2208
        # Cache the encoder outputs by mm_hash
        for (mm_hash, pos_info), output in zip(mm_hashes_pos, encoder_outputs):
2209
            self.encoder_cache[mm_hash] = output
2210
2211
            logger.debug("Finish execute for mm hash %s", mm_hash)
            self.maybe_save_ec_to_connector(self.encoder_cache, mm_hash)
2212

2213
2214
        return encoder_outputs

2215
    def _gather_mm_embeddings(
2216
2217
        self,
        scheduler_output: "SchedulerOutput",
2218
        shift_computed_tokens: int = 0,
2219
2220
2221
2222
2223
2224
2225
2226
    ) -> 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
2227
        should_sync_mrope_positions = False
2228
        should_sync_xdrope_positions = False
2229

2230
        for req_id in self.input_batch.req_ids:
2231
2232
            mm_embeds_req: list[torch.Tensor] = []

2233
            num_scheduled_tokens = scheduler_output.num_scheduled_tokens[req_id]
2234
            req_state = self.requests[req_id]
2235
            num_computed_tokens = req_state.num_computed_tokens + shift_computed_tokens
2236

2237
2238
            for mm_feature in req_state.mm_features:
                pos_info = mm_feature.mm_position
2239
2240
                start_pos = pos_info.offset
                num_encoder_tokens = pos_info.length
2241
2242
2243
2244
2245
2246
2247
2248
2249
2250
2251
2252
2253
2254
2255
2256

                # 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,
2257
2258
                    num_encoder_tokens,
                )
2259
                assert start_idx < end_idx
2260
2261
2262
2263
2264
2265
2266
                curr_embeds_start, curr_embeds_end = (
                    pos_info.get_embeds_indices_in_range(start_idx, end_idx)
                )
                # If there are no embeddings in the current range, we skip
                # gathering the embeddings.
                if curr_embeds_start == curr_embeds_end:
                    continue
2267

2268
                mm_hash = mm_feature.identifier
2269
                encoder_output = self.encoder_cache.get(mm_hash, None)
2270
                assert encoder_output is not None, f"Encoder cache miss for {mm_hash}."
2271
2272
2273

                if (is_embed := pos_info.is_embed) is not None:
                    is_embed = is_embed[start_idx:end_idx]
2274
2275
2276
                    mm_embeds_item = encoder_output[curr_embeds_start:curr_embeds_end]
                else:
                    mm_embeds_item = encoder_output[start_idx:end_idx]
2277

2278
                req_start_pos = req_start_idx + start_pos - num_computed_tokens
2279
2280
2281
                is_mm_embed[req_start_pos + start_idx : req_start_pos + end_idx] = (
                    True if is_embed is None else is_embed
                )
2282
2283
2284
                mm_embeds_req.append(mm_embeds_item)

            if self.is_multimodal_pruning_enabled and self.uses_mrope:
2285
                assert req_state.mrope_positions is not None
2286
2287
2288
2289
2290
2291
2292
                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,
2293
2294
                    )
                )
2295
2296
2297
2298
                req_state.mrope_positions.copy_(new_mrope_positions)
                req_state.mrope_position_delta = new_delta

            mm_embeds.extend(mm_embeds_req)
2299
2300
2301
            req_start_idx += num_scheduled_tokens

        is_mm_embed = self.is_mm_embed.copy_to_gpu(total_num_scheduled_tokens)
2302
2303
2304

        if should_sync_mrope_positions:
            self._calc_mrope_positions(scheduler_output)
2305
            self.mrope_positions.copy_to_gpu(total_num_scheduled_tokens)
2306

2307
2308
2309
2310
        if should_sync_xdrope_positions:
            self._calc_xdrope_positions(scheduler_output)
            self.xdrope_positions.copy_to_gpu(total_num_scheduled_tokens)

2311
        return mm_embeds, is_mm_embed
2312

2313
    def get_model(self) -> nn.Module:
2314
        # get raw model out of the cudagraph wrapper.
2315
        if isinstance(self.model, (CUDAGraphWrapper, UBatchWrapper)):
2316
            return self.model.unwrap()
2317
2318
        return self.model

2319
2320
2321
2322
2323
2324
2325
2326
2327
2328
2329
2330
2331
2332
2333
    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

2334
2335
2336
2337
2338
    def get_supported_pooling_tasks(self) -> list[PoolingTask]:
        model = self.get_model()
        if not is_pooling_model(model):
            return []

2339
2340
        supported_tasks = list(model.pooler.get_supported_tasks())

2341
2342
2343
2344
        if "score" in supported_tasks:
            num_labels = getattr(self.model_config.hf_config, "num_labels", 0)
            if num_labels != 1:
                supported_tasks.remove("score")
2345
                logger.debug_once("Score API is only enabled for num_labels == 1.")
2346
2347

        return supported_tasks
2348

2349
2350
2351
2352
2353
2354
2355
2356
2357
2358
    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)

2359
    def sync_and_slice_intermediate_tensors(
2360
2361
        self,
        num_tokens: int,
2362
        intermediate_tensors: IntermediateTensors | None,
2363
2364
        sync_self: bool,
    ) -> IntermediateTensors:
2365
2366
2367
        assert self.intermediate_tensors is not None

        tp = self.vllm_config.parallel_config.tensor_parallel_size
2368
        is_rs = is_residual_scattered_for_sp(self.vllm_config, num_tokens)
2369
2370
2371
2372
2373
2374

        # 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():
2375
                is_scattered = k == "residual" and is_rs
2376
                copy_len = num_tokens // tp if is_scattered else num_tokens
2377
                self.intermediate_tensors[k][:copy_len].copy_(
2378
2379
2380
2381
2382
2383
2384
2385
2386
2387
2388
2389
2390
                    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:
2391
2392
2393
2394
2395
2396
2397
        """
        Step for the EPLB (Expert Parallelism Load Balancing) state.
        """
        if not self.parallel_config.enable_eplb:
            return

        assert self.eplb_state is not None
2398
2399
        model = self.get_model()
        assert is_mixture_of_experts(model)
2400
2401
2402
        self.eplb_state.step(
            is_dummy,
            is_profile,
2403
            log_stats=self.parallel_config.eplb_config.log_balancedness,
2404
2405
        )

2406
2407
2408
2409
2410
2411
    def _pool(
        self,
        hidden_states: torch.Tensor,
        num_scheduled_tokens: int,
        num_scheduled_tokens_np: np.ndarray,
    ) -> ModelRunnerOutput:
2412
2413
2414
        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"
        )
2415

2416
        hidden_states = hidden_states[:num_scheduled_tokens]
2417
2418
        seq_lens_cpu = self.seq_lens.cpu[: self.input_batch.num_reqs]

2419
        pooling_metadata = self.input_batch.get_pooling_metadata()
2420
        pooling_metadata.build_pooling_cursor(
2421
            num_scheduled_tokens_np.tolist(), seq_lens_cpu, device=hidden_states.device
2422
        )
2423

2424
2425
2426
2427
2428
2429
        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(
2430
            lambda x: x.to("cpu", non_blocking=True) if x is not None else x,
2431
2432
2433
            raw_pooler_output,
        )
        self._sync_device()
2434

2435
        pooler_output: list[torch.Tensor | None] = []
2436
        for raw_output, seq_len, prompt_len in zip(
2437
2438
            raw_pooler_output, seq_lens_cpu, pooling_metadata.prompt_lens
        ):
2439
            output = raw_output if seq_len == prompt_len else None
2440
            pooler_output.append(output)
2441
2442
2443
2444
2445
2446
2447
2448
2449
2450

        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,
        )

2451
    def _pad_for_sequence_parallelism(self, num_scheduled_tokens: int) -> int:
2452
2453
2454
        # Pad tokens to multiple of tensor_parallel_size when
        # enabled collective fusion for SP
        tp_size = self.vllm_config.parallel_config.tensor_parallel_size
2455
        if self.compilation_config.pass_config.enable_sp and tp_size > 1:
2456
2457
2458
            return round_up(num_scheduled_tokens, tp_size)
        return num_scheduled_tokens

2459
    def _preprocess(
2460
2461
        self,
        scheduler_output: "SchedulerOutput",
2462
        num_input_tokens: int,  # Padded
2463
        intermediate_tensors: IntermediateTensors | None = None,
2464
    ) -> tuple[
2465
2466
        torch.Tensor | None,
        torch.Tensor | None,
2467
        torch.Tensor,
2468
        IntermediateTensors | None,
2469
        dict[str, Any],
2470
        ECConnectorOutput | None,
2471
    ]:
2472
        num_scheduled_tokens = scheduler_output.total_num_scheduled_tokens
2473
        is_first_rank = get_pp_group().is_first_rank
2474
        is_encoder_decoder = self.model_config.is_encoder_decoder
2475

2476
2477
        # _prepare_inputs may reorder the batch, so we must gather multi
        # modal outputs after that to ensure the correct order
2478
2479
        ec_connector_output = None

2480
        if self.supports_mm_inputs and is_first_rank and not is_encoder_decoder:
2481
            # Run the multimodal encoder if any.
2482
2483
2484
2485
2486
2487
            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)
2488

2489
2490
2491
            # 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.
2492
            inputs_embeds_scheduled = self.model.embed_input_ids(
2493
2494
2495
                self.input_ids.gpu[:num_scheduled_tokens],
                multimodal_embeddings=mm_embeds,
                is_multimodal=is_mm_embed,
2496
            )
2497

2498
            # TODO(woosuk): Avoid the copy. Optimize.
2499
            self.inputs_embeds.gpu[:num_scheduled_tokens].copy_(inputs_embeds_scheduled)
2500

2501
            input_ids = None
2502
            inputs_embeds = self.inputs_embeds.gpu[:num_input_tokens]
2503
2504
2505
2506
            model_kwargs = {
                **self._init_model_kwargs(num_scheduled_tokens),
                **self._extract_mm_kwargs(scheduler_output),
            }
2507
        elif self.enable_prompt_embeds and is_first_rank:
2508
2509
2510
2511
2512
2513
2514
2515
2516
2517
2518
2519
            # 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).
2520
2521
2522
            token_ids_idx = (
                self.is_token_ids.gpu[:num_scheduled_tokens]
                .nonzero(as_tuple=False)
2523
                .squeeze(1)
2524
            )
2525
2526
2527
            # Some tokens ids may need to become embeds
            if token_ids_idx.numel() > 0:
                token_ids = self.input_ids.gpu[token_ids_idx]
2528
                tokens_to_embeds = self.model.embed_input_ids(input_ids=token_ids)
2529
2530
2531
2532
2533
                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
2534
        else:
2535
2536
2537
2538
            # 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.
2539
            input_ids = self.input_ids.gpu[:num_input_tokens]
2540
            inputs_embeds = None
2541
            model_kwargs = self._init_model_kwargs(num_input_tokens)
2542

2543
        if self.uses_mrope:
2544
            positions = self.mrope_positions.gpu[:, :num_input_tokens]
2545
2546
        elif self.uses_xdrope_dim > 0:
            positions = self.xdrope_positions.gpu[:, :num_input_tokens]
2547
        else:
2548
            positions = self.positions.gpu[:num_input_tokens]
2549

2550
        if is_first_rank:
2551
2552
            intermediate_tensors = None
        else:
2553
            assert intermediate_tensors is not None
2554
            intermediate_tensors = self.sync_and_slice_intermediate_tensors(
2555
2556
                num_input_tokens, intermediate_tensors, True
            )
2557

2558
        if is_encoder_decoder and scheduler_output.scheduled_encoder_inputs:
2559
2560
2561
2562
2563
2564
2565
            # 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})
2566

2567
2568
2569
2570
2571
2572
        return (
            input_ids,
            inputs_embeds,
            positions,
            intermediate_tensors,
            model_kwargs,
2573
            ec_connector_output,
2574
        )
2575

2576
    def _sample(
2577
        self,
2578
2579
        logits: torch.Tensor | None,
        spec_decode_metadata: SpecDecodeMetadata | None,
2580
    ) -> SamplerOutput:
2581
        # Sample the next token and get logprobs if needed.
2582
        sampling_metadata = self.input_batch.sampling_metadata
2583
        if spec_decode_metadata is None:
2584
2585
2586
            # 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()
2587
            return self.sampler(
2588
2589
2590
                logits=logits,
                sampling_metadata=sampling_metadata,
            )
2591

2592
        sampler_output = self.rejection_sampler(
2593
2594
            spec_decode_metadata,
            None,  # draft_probs
2595
            logits,
2596
2597
            sampling_metadata,
        )
2598
        self._update_states_after_model_execute(sampler_output.sampled_token_ids)
2599
2600
2601
        return sampler_output

    def _bookkeeping_sync(
2602
2603
2604
        self,
        scheduler_output: "SchedulerOutput",
        sampler_output: SamplerOutput,
2605
        logits: torch.Tensor | None,
2606
2607
        hidden_states: torch.Tensor,
        num_scheduled_tokens: int,
2608
        spec_decode_metadata: SpecDecodeMetadata | None,
2609
    ) -> tuple[
2610
        dict[str, int],
2611
        LogprobsLists | None,
2612
        list[list[int]],
2613
        dict[str, LogprobsTensors | None],
2614
2615
2616
        list[str],
        dict[str, int],
        list[int],
2617
    ]:
2618
2619
2620
2621
        num_nans_in_logits = {}
        if envs.VLLM_COMPUTE_NANS_IN_LOGITS:
            num_nans_in_logits = self._get_nans_in_logits(logits)

2622
2623
2624
2625
        num_reqs = self.input_batch.num_reqs
        discard_sampled_tokens_req_indices = np.nonzero(
            self.discard_request_mask.np[:num_reqs]
        )[0]
2626
2627
2628
2629
        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)
2630

2631
2632
2633
        # 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()
2634
        req_id_to_index_output_copy = self.input_batch.req_id_to_index.copy()
2635
2636

        num_sampled_tokens = sampler_output.sampled_token_ids.shape[0]
2637
        sampled_token_ids = sampler_output.sampled_token_ids
2638
        logprobs_tensors = sampler_output.logprobs_tensors
2639
        invalid_req_indices = []
2640
        cu_num_tokens: list[int] | None = None
2641
2642
2643
2644
2645
2646
        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)
2647
2648
2649
                # 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()
2650
2651
            else:
                # Includes spec decode tokens.
2652
                valid_sampled_token_ids, cu_num_tokens = RejectionSampler.parse_output(
2653
2654
                    sampled_token_ids,
                    self.input_batch.vocab_size,
2655
2656
                    discard_sampled_tokens_req_indices,
                    return_cu_num_tokens=logprobs_tensors is not None,
2657
                )
2658
        else:
2659
            valid_sampled_token_ids = []
2660
            invalid_req_indices = discard_sampled_tokens_req_indices.tolist()
2661
2662
2663
2664
2665
            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.
2666
2667
2668
2669
            # 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
2670
2671
2672
2673
2674
            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
            }
2675

2676
2677
2678
2679
2680
        # 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.
2681
        req_ids = self.input_batch.req_ids
2682
2683
        for req_idx in range(num_sampled_tokens):
            if self.use_async_scheduling:
2684
                sampled_ids = [-1] if req_idx not in invalid_req_indices_set else None
2685
2686
            else:
                sampled_ids = valid_sampled_token_ids[req_idx]
2687

2688
            num_sampled_ids: int = len(sampled_ids) if sampled_ids else 0
2689

2690
            if not sampled_ids:
2691
2692
2693
                continue

            start_idx = self.input_batch.num_tokens_no_spec[req_idx]
2694
            end_idx = start_idx + num_sampled_ids
2695
2696
2697
2698
            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}"
2699
            )
2700

2701
2702
            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
2703
            self.input_batch.num_tokens_no_spec[req_idx] = end_idx
2704

2705
            req_id = req_ids[req_idx]
2706
2707
2708
            req_state = self.requests[req_id]
            req_state.output_token_ids.extend(sampled_ids)

2709
        logprobs_lists = (
2710
            logprobs_tensors.tolists(cu_num_tokens)
2711
            if not self.use_async_scheduling and logprobs_tensors is not None
2712
2713
2714
2715
2716
2717
2718
2719
2720
            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,
        )

2721
2722
2723
2724
2725
2726
2727
2728
2729
2730
        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,
        )

2731
2732
2733
2734
2735
2736
2737
2738
2739
2740
2741
2742
2743
2744
2745
    @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()

2746
2747
    def _model_forward(
        self,
2748
2749
2750
2751
        input_ids: torch.Tensor | None = None,
        positions: torch.Tensor | None = None,
        intermediate_tensors: IntermediateTensors | None = None,
        inputs_embeds: torch.Tensor | None = None,
2752
2753
2754
2755
2756
        **model_kwargs: dict[str, Any],
    ) -> Any:
        """Helper method to call the model forward pass.

        This method can be overridden by subclasses for model execution.
2757
        Motivation: We can inspect only this method versus
2758
2759
2760
2761
2762
2763
2764
2765
2766
2767
2768
2769
2770
2771
2772
2773
2774
2775
2776
2777
        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,
        )

2778
2779
2780
2781
2782
2783
2784
2785
2786
2787
2788
2789
2790
2791
2792
2793
2794
2795
2796
2797
2798
    @staticmethod
    def _is_uniform_decode(
        max_num_scheduled_tokens: int,
        uniform_decode_query_len: int,
        num_tokens: int,
        num_reqs: int,
        force_uniform_decode: bool | None = None,
    ) -> bool:
        """
        Checks if it's a decode batch with same amount scheduled tokens
        across all requests.
        """
        return (
            (
                (max_num_scheduled_tokens == uniform_decode_query_len)
                and (num_tokens == max_num_scheduled_tokens * num_reqs)
            )
            if force_uniform_decode is None
            else force_uniform_decode
        )

2799
2800
2801
2802
2803
2804
2805
2806
2807
2808
2809
2810
2811
    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,
2812
        num_encoder_reqs: int = 0,
2813
    ) -> tuple[
2814
2815
        CUDAGraphMode,
        BatchDescriptor,
2816
        bool,
2817
2818
        torch.Tensor | None,
        CUDAGraphStat | None,
2819
    ]:
2820
2821
2822
2823
2824
2825
        uniform_decode = self._is_uniform_decode(
            max_num_scheduled_tokens=max_num_scheduled_tokens,
            uniform_decode_query_len=self.uniform_decode_query_len,
            num_tokens=num_tokens,
            num_reqs=num_reqs,
            force_uniform_decode=force_uniform_decode,
2826
        )
2827
2828
2829
2830
2831
        # 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
        )
2832
2833
2834
2835
2836
2837
2838

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

2839
        num_tokens_padded = self._pad_for_sequence_parallelism(num_tokens)
2840
        dispatch_cudagraph = (
2841
            lambda num_tokens, disable_full: self.cudagraph_dispatcher.dispatch(
2842
2843
2844
                num_tokens=num_tokens,
                has_lora=has_lora,
                uniform_decode=uniform_decode,
2845
                disable_full=disable_full,
2846
2847
2848
2849
2850
            )
            if not force_eager
            else (CUDAGraphMode.NONE, BatchDescriptor(num_tokens_padded))
        )

2851
        cudagraph_mode, batch_descriptor = dispatch_cudagraph(
2852
            num_tokens_padded, use_cascade_attn or has_encoder_output
2853
        )
2854
        num_tokens_padded = batch_descriptor.num_tokens
2855
2856
2857
2858
2859
2860
2861
2862
2863
        if self.compilation_config.pass_config.enable_sp:
            assert (
                batch_descriptor.num_tokens
                % self.vllm_config.parallel_config.tensor_parallel_size
                == 0
            ), (
                "Sequence parallelism requires num_tokens to be "
                "a multiple of tensor parallel size"
            )
2864
2865
2866

        # Extra coordination when running data-parallel since we need to coordinate
        # across ranks
2867
        should_ubatch, num_tokens_across_dp = False, None
2868
2869
2870
2871
2872
2873
2874
2875
2876
        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
            )

2877
2878
2879
2880
2881
2882
2883
2884
2885
2886
2887
            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,
                )
2888
2889
            )

2890
            # Extract DP-synced values
2891
2892
2893
            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())
2894
2895
2896
2897
2898
                # 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,
                )
2899
2900
2901
2902
                # 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

2903
2904
2905
2906
2907
2908
2909
2910
2911
2912
2913
2914
        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,
2915
            should_ubatch,
2916
2917
2918
            num_tokens_across_dp,
            cudagraph_stats,
        )
2919

2920
2921
2922
2923
2924
2925
2926
2927
2928
2929
2930
2931
2932
2933
2934
2935
2936
2937
2938
2939
2940
2941
2942
2943
2944
2945
2946
2947
2948
2949
2950
2951
2952
2953
2954
2955
    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

2956
2957
2958
2959
    @torch.inference_mode()
    def execute_model(
        self,
        scheduler_output: "SchedulerOutput",
2960
        intermediate_tensors: IntermediateTensors | None = None,
2961
2962
2963
2964
2965
2966
    ) -> 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."
            )
2967
2968
2969
2970
2971
2972
2973
2974
2975
2976
2977
2978
2979
2980
2981

        # 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)

2982
        num_scheduled_tokens = scheduler_output.total_num_scheduled_tokens
2983
        with record_function_or_nullcontext("gpu_model_runner: preprocess"):
2984
2985
2986
2987
            with self.synchronize_input_prep():
                # Update persistent batch states.
                self._update_states(scheduler_output)

2988
2989
2990
2991
2992
2993
2994
2995
                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)

2996
                if not num_scheduled_tokens:
2997
2998
2999
3000
3001
3002
3003
3004
3005
3006
3007
                    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.
3008
                        self._dummy_run(1)
3009
3010
3011
3012
                    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(
3013
3014
                        scheduler_output, self.vllm_config
                    )
3015
                if self.cache_config.kv_sharing_fast_prefill:
3016
                    assert not self.num_prompt_logprobs, (
3017
3018
                        "--kv-sharing-fast-prefill produces incorrect "
                        "logprobs for prompt tokens, tokens, please disable "
3019
3020
                        "it when the requests need prompt logprobs"
                    )
3021

3022
3023
3024
3025
3026
                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())
3027
                num_tokens_unpadded = scheduler_output.total_num_scheduled_tokens
3028

3029
3030
3031
                (
                    logits_indices,
                    spec_decode_metadata,
3032
                ) = self._prepare_inputs(
3033
3034
                    scheduler_output,
                    num_scheduled_tokens_np,
3035
3036
3037
3038
                )

                cascade_attn_prefix_lens = None
                # Disable cascade attention when using microbatching (DBO)
3039
                if self.cascade_attn_enabled and not self.parallel_config.use_ubatching:
3040
3041
3042
                    # Pre-compute cascade attention prefix lengths
                    cascade_attn_prefix_lens = self._compute_cascade_attn_prefix_lens(
                        num_scheduled_tokens_np,
3043
                        self.input_batch.num_computed_tokens_cpu[:num_reqs],
3044
3045
3046
                        scheduler_output.num_common_prefix_blocks,
                    )

3047
3048
3049
                (
                    cudagraph_mode,
                    batch_desc,
3050
                    should_ubatch,
3051
                    num_tokens_across_dp,
3052
                    cudagraph_stats,
3053
3054
3055
3056
3057
3058
                ) = 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,
3059
                    num_encoder_reqs=len(scheduler_output.scheduled_encoder_inputs),
3060
3061
3062
3063
                )

                logger.debug(
                    "Running batch with cudagraph_mode: %s, batch_descriptor: %s, "
3064
                    "should_ubatch: %s, num_tokens_across_dp: %s",
3065
3066
                    cudagraph_mode,
                    batch_desc,
3067
                    should_ubatch,
3068
3069
3070
3071
3072
3073
3074
                    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
                )
3075
3076
3077
3078
3079
                ubatch_slices, ubatch_slices_padded = maybe_create_ubatch_slices(
                    should_ubatch,
                    num_scheduled_tokens_np,
                    num_tokens_padded,
                    num_reqs_padded,
3080
3081
3082
3083
3084
3085
3086
                    self.parallel_config.num_ubatches,
                )

                logger.debug(
                    "ubatch_slices: %s, ubatch_slices_padded: %s",
                    ubatch_slices,
                    ubatch_slices_padded,
3087
                )
3088

3089
3090
                pad_attn = cudagraph_mode == CUDAGraphMode.FULL

3091
3092
3093
                use_spec_decode = len(scheduler_output.scheduled_spec_decode_tokens) > 0
                ubatch_slices_attn = ubatch_slices_padded if pad_attn else ubatch_slices

3094
                (attn_metadata, spec_decode_common_attn_metadata) = (
3095
                    self._build_attention_metadata(
3096
3097
                        num_tokens=num_tokens_unpadded,
                        num_tokens_padded=num_tokens_padded if pad_attn else None,
3098
                        num_reqs=num_reqs,
3099
3100
                        num_reqs_padded=num_reqs_padded if pad_attn else None,
                        max_query_len=max_num_scheduled_tokens,
3101
                        ubatch_slices=ubatch_slices_attn,
3102
3103
                        logits_indices=logits_indices,
                        use_spec_decode=use_spec_decode,
3104
                        num_scheduled_tokens=scheduler_output.num_scheduled_tokens,
3105
3106
3107
                        cascade_attn_prefix_lens=cascade_attn_prefix_lens,
                    )
                )
3108

3109
3110
3111
3112
3113
3114
3115
3116
3117
            (
                input_ids,
                inputs_embeds,
                positions,
                intermediate_tensors,
                model_kwargs,
                ec_connector_output,
            ) = self._preprocess(
                scheduler_output, num_tokens_padded, intermediate_tensors
3118
            )
3119

3120
        # Set cudagraph mode to none if calc_kv_scales is true.
3121
3122
3123
        # KV scales calculation involves dynamic operations that are incompatible
        # with CUDA graph capture.
        if self.calculate_kv_scales:
3124
            cudagraph_mode = CUDAGraphMode.NONE
3125
3126
            # Mark KV scales as calculated after the first forward pass
            self.calculate_kv_scales = False
3127

3128
3129
        # Run the model.
        # Use persistent buffers for CUDA graphs.
3130
3131
        with (
            set_forward_context(
3132
3133
                attn_metadata,
                self.vllm_config,
3134
                num_tokens=num_tokens_padded,
3135
                num_tokens_across_dp=num_tokens_across_dp,
3136
3137
                cudagraph_runtime_mode=cudagraph_mode,
                batch_descriptor=batch_desc,
3138
                ubatch_slices=ubatch_slices_padded,
3139
            ),
3140
            record_function_or_nullcontext("gpu_model_runner: forward"),
3141
3142
            self.maybe_get_kv_connector_output(scheduler_output) as kv_connector_output,
        ):
3143
            model_output = self._model_forward(
3144
3145
3146
3147
3148
3149
3150
                input_ids=input_ids,
                positions=positions,
                intermediate_tensors=intermediate_tensors,
                inputs_embeds=inputs_embeds,
                **model_kwargs,
            )

3151
        with record_function_or_nullcontext("gpu_model_runner: postprocess"):
3152
            if self.use_aux_hidden_state_outputs:
3153
                # True when EAGLE 3 is used.
3154
3155
                hidden_states, aux_hidden_states = model_output
            else:
3156
                # Common case.
3157
3158
3159
                hidden_states = model_output
                aux_hidden_states = None

3160
3161
3162
3163
3164
            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)
3165
                    hidden_states.kv_connector_output = kv_connector_output
3166
                    self.kv_connector_output = kv_connector_output
3167
                    return hidden_states
3168

3169
                if self.is_pooling_model:
3170
                    # Return the pooling output.
3171
3172
3173
                    output = self._pool(
                        hidden_states, num_scheduled_tokens, num_scheduled_tokens_np
                    )
3174
3175
                    output.kv_connector_output = kv_connector_output
                    return output
3176
3177

                sample_hidden_states = hidden_states[logits_indices]
3178
                logits = self.model.compute_logits(sample_hidden_states)
3179
3180
3181
3182
            else:
                # Rare case.
                assert not self.is_pooling_model

3183
                sample_hidden_states = hidden_states[logits_indices]
3184
                if not get_pp_group().is_last_rank:
3185
                    all_gather_tensors = {
3186
                        "residual": not is_residual_scattered_for_sp(
3187
                            self.vllm_config, num_tokens_padded
3188
                        )
3189
                    }
3190
                    get_pp_group().send_tensor_dict(
3191
3192
                        hidden_states.tensors,
                        all_gather_group=get_tp_group(),
3193
3194
                        all_gather_tensors=all_gather_tensors,
                    )
3195
3196
                    logits = None
                else:
3197
                    logits = self.model.compute_logits(sample_hidden_states)
3198

3199
                model_output_broadcast_data: dict[str, Any] = {}
3200
3201
3202
                if logits is not None:
                    model_output_broadcast_data["logits"] = logits.contiguous()

3203
                broadcasted = get_pp_group().broadcast_tensor_dict(
3204
3205
                    model_output_broadcast_data, src=len(get_pp_group().ranks) - 1
                )
3206
3207
                assert broadcasted is not None
                logits = broadcasted["logits"]
3208

3209
3210
3211
3212
3213
3214
3215
3216
        self.execute_model_state = ExecuteModelState(
            scheduler_output,
            logits,
            spec_decode_metadata,
            spec_decode_common_attn_metadata,
            hidden_states,
            sample_hidden_states,
            aux_hidden_states,
3217
            ec_connector_output,
3218
            cudagraph_stats,
3219
        )
3220
        self.kv_connector_output = kv_connector_output
3221
3222
3223
3224
3225
3226
        return None

    @torch.inference_mode
    def sample_tokens(
        self, grammar_output: "GrammarOutput | None"
    ) -> ModelRunnerOutput | AsyncModelRunnerOutput | IntermediateTensors:
3227
3228
3229
        kv_connector_output = self.kv_connector_output
        self.kv_connector_output = None

3230
3231
        if self.execute_model_state is None:
            # Nothing to do (PP non-final rank case), output isn't used.
3232
            if not kv_connector_output:
3233
                return None  # type: ignore[return-value]
3234
3235
3236
3237
3238
3239
3240
3241
3242

            # 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
3243
3244
3245
3246
3247
3248
3249
3250
3251
3252

        # Unpack ephemeral state.
        (
            scheduler_output,
            logits,
            spec_decode_metadata,
            spec_decode_common_attn_metadata,
            hidden_states,
            sample_hidden_states,
            aux_hidden_states,
3253
            ec_connector_output,
3254
            cudagraph_stats,
3255
3256
3257
3258
3259
3260
3261
3262
3263
        ) = 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
            )
3264

3265
        with record_function_or_nullcontext("gpu_model_runner: sample"):
3266
3267
            sampler_output = self._sample(logits, spec_decode_metadata)

3268
3269
        self.input_batch.prev_sampled_token_ids = None

3270
        def propose_draft_token_ids(sampled_token_ids):
3271
            assert spec_decode_common_attn_metadata is not None
3272
            with record_function_or_nullcontext("gpu_model_runner: draft"):
3273
3274
3275
3276
3277
3278
3279
3280
3281
3282
3283
                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,
                )

3284
        spec_config = self.speculative_config
3285
        use_padded_batch_for_eagle = (
3286
3287
3288
            spec_config is not None
            and spec_config.use_eagle()
            and not spec_config.disable_padded_drafter_batch
3289
        )
3290
3291
3292
        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
3293
        if (
3294
3295
3296
            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
3297
        ):
3298
            effective_drafter_max_model_len = (
3299
                spec_config.draft_model_config.max_model_len
3300
            )
3301
        input_fits_in_drafter = spec_decode_common_attn_metadata and (
3302
            spec_decode_common_attn_metadata.max_seq_len + self.num_spec_tokens
3303
3304
            <= effective_drafter_max_model_len
        )
3305
        if use_padded_batch_for_eagle:
3306
3307
            assert self.speculative_config is not None
            assert isinstance(self.drafter, EagleProposer)
3308
3309
3310
3311
3312
3313
            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:
3314
                assert spec_decode_common_attn_metadata is not None
3315
3316
3317
3318
3319
3320
                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,
3321
                        self.discard_request_mask.gpu,
3322
3323
3324
3325
3326
                    )
                )
                self._copy_valid_sampled_token_count(
                    next_token_ids, valid_sampled_tokens_count
                )
3327

3328
        with record_function_or_nullcontext("gpu_model_runner: bookkeep"):
3329
3330
3331
3332
3333
3334
3335
3336
            (
                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,
3337
3338
3339
3340
3341
            ) = self._bookkeeping_sync(
                scheduler_output,
                sampler_output,
                logits,
                hidden_states,
3342
                scheduler_output.total_num_scheduled_tokens,
3343
                spec_decode_metadata,
3344
            )
3345

3346
3347
3348
3349
3350
        if (
            self.speculative_config
            and not use_padded_batch_for_eagle
            and input_fits_in_drafter
        ):
3351
3352
3353
            # 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)
3354

3355
        with record_function_or_nullcontext("gpu_model_runner: eplb"):
3356
            self.eplb_step()
3357
3358
3359
3360
3361
3362
3363
3364
3365
        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,
3366
3367
3368
                ec_connector_output=ec_connector_output
                if self.supports_mm_inputs
                else None,
3369
                num_nans_in_logits=num_nans_in_logits,
3370
                cudagraph_stats=cudagraph_stats,
3371
            )
3372

3373
3374
        if not self.use_async_scheduling:
            return output
3375
3376
3377
3378
3379
3380
3381
3382
3383
        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,
3384
                vocab_size=self.input_batch.vocab_size,
3385
3386
3387
3388
3389
            )
        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
3390
            # any requests with sampling params that require output ids.
3391
3392
3393
3394
            self.input_batch.set_async_sampled_token_ids(
                async_output.sampled_token_ids_cpu,
                async_output.async_copy_ready_event,
            )
3395
3396
3397

        return async_output

3398
    def take_draft_token_ids(self) -> DraftTokenIds | None:
3399
        if not self.num_spec_tokens:
3400
            return None
3401

3402
        req_ids = self.input_batch.req_ids
3403
3404
3405
        if self._draft_token_ids is None:
            return DraftTokenIds(req_ids, [[] for _ in req_ids])

3406
3407
3408
3409
3410
3411
3412
        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)

3413
3414
3415
3416
3417
3418
3419
3420
3421
3422
3423
3424
3425
3426
3427
3428
3429
3430
3431
3432
3433
3434
3435
3436
3437
3438
3439
3440
3441
3442
3443
    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()

3444
3445
3446
    def propose_draft_token_ids(
        self,
        scheduler_output: "SchedulerOutput",
3447
        sampled_token_ids: torch.Tensor | list[list[int]],
3448
3449
3450
        sampling_metadata: SamplingMetadata,
        hidden_states: torch.Tensor,
        sample_hidden_states: torch.Tensor,
3451
3452
        aux_hidden_states: list[torch.Tensor] | None,
        spec_decode_metadata: SpecDecodeMetadata | None,
3453
        common_attn_metadata: CommonAttentionMetadata,
3454
    ) -> list[list[int]] | torch.Tensor:
3455
        num_scheduled_tokens = scheduler_output.total_num_scheduled_tokens
3456
3457
3458
        spec_config = self.speculative_config
        assert spec_config is not None
        if spec_config.method == "ngram":
3459
            assert isinstance(sampled_token_ids, list)
3460
            assert isinstance(self.drafter, NgramProposer)
3461
            draft_token_ids = self.drafter.propose(
3462
3463
                sampled_token_ids,
                self.input_batch.req_ids,
3464
3465
                self.input_batch.num_tokens_no_spec,
                self.input_batch.token_ids_cpu,
3466
3467
                self.input_batch.spec_decode_unsupported_reqs,
            )
3468
        elif spec_config.method == "suffix":
3469
3470
3471
            assert isinstance(sampled_token_ids, list)
            assert isinstance(self.drafter, SuffixDecodingProposer)
            draft_token_ids = self.drafter.propose(self.input_batch, sampled_token_ids)
3472
        elif spec_config.method == "medusa":
3473
            assert isinstance(sampled_token_ids, list)
3474
            assert isinstance(self.drafter, MedusaProposer)
3475

3476
3477
            if sample_hidden_states.shape[0] == len(sampled_token_ids):
                # The input to the target model does not include draft tokens.
3478
3479
3480
3481
                hidden_states = sample_hidden_states
            else:
                indices = []
                offset = 0
3482
3483
3484
                assert spec_decode_metadata is not None, (
                    "No spec decode metadata for medusa"
                )
3485
                for num_draft, tokens in zip(
3486
3487
                    spec_decode_metadata.num_draft_tokens, sampled_token_ids
                ):
3488
                    indices.append(offset + len(tokens) - 1)
3489
                    offset += num_draft + 1
3490
                indices = torch.tensor(indices, device=self.device)
3491
3492
                hidden_states = sample_hidden_states[indices]

3493
            draft_token_ids = self.drafter.propose(
3494
3495
3496
                target_hidden_states=hidden_states,
                sampling_metadata=sampling_metadata,
            )
3497
        elif spec_config.use_eagle():
3498
            assert isinstance(self.drafter, EagleProposer)
3499

3500
            if spec_config.disable_padded_drafter_batch:
3501
3502
3503
                # 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.
3504
3505
                assert isinstance(sampled_token_ids, list), (
                    "sampled_token_ids should be a python list when"
3506
                    "padded-batch is disabled."
3507
                )
3508
                next_token_ids = self.drafter.prepare_next_token_ids_cpu(
3509
3510
3511
3512
3513
                    sampled_token_ids,
                    self.requests,
                    self.input_batch,
                    scheduler_output.num_scheduled_tokens,
                )
3514
3515
3516
3517
3518
            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.
3519
3520
                assert isinstance(sampled_token_ids, torch.Tensor), (
                    "sampled_token_ids should be a torch.Tensor when"
3521
                    "padded-batch is enabled."
3522
3523
                )
                next_token_ids, valid_sampled_tokens_count = (
3524
3525
3526
3527
3528
                    self.drafter.prepare_next_token_ids_padded(
                        common_attn_metadata,
                        sampled_token_ids,
                        self.requests,
                        self.input_batch,
3529
                        self.discard_request_mask.gpu,
3530
                    )
3531
                )
3532
3533
3534
                self._copy_valid_sampled_token_count(
                    next_token_ids, valid_sampled_tokens_count
                )
Jiayi Yao's avatar
Jiayi Yao committed
3535

3536
            if spec_decode_metadata is None:
3537
                token_indices_to_sample = None
3538
                # input_ids can be None for multimodal models.
3539
                target_token_ids = self.input_ids.gpu[:num_scheduled_tokens]
3540
                target_positions = self._get_positions(num_scheduled_tokens)
3541
                if self.use_aux_hidden_state_outputs:
Wentao Ye's avatar
Wentao Ye committed
3542
                    assert aux_hidden_states is not None
3543
                    target_hidden_states = torch.cat(
3544
3545
                        [h[:num_scheduled_tokens] for h in aux_hidden_states], dim=-1
                    )
3546
3547
                else:
                    target_hidden_states = hidden_states[:num_scheduled_tokens]
3548
            else:
3549
                if spec_config.disable_padded_drafter_batch:
3550
                    token_indices_to_sample = None
3551
3552
3553
3554
3555
                    common_attn_metadata, token_indices = self.drafter.prepare_inputs(
                        common_attn_metadata,
                        sampled_token_ids,
                        spec_decode_metadata.num_draft_tokens,
                    )
3556
3557
3558
3559
3560
3561
3562
3563
3564
                    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]
3565
                else:
3566
                    common_attn_metadata, token_indices_to_sample = (
3567
3568
3569
                        self.drafter.prepare_inputs_padded(
                            common_attn_metadata,
                            spec_decode_metadata,
3570
3571
3572
                            valid_sampled_tokens_count,
                        )
                    )
3573
3574
3575
3576
3577
3578
3579
3580
3581
3582
3583
                    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]
3584

3585
            if self.supports_mm_inputs:
3586
3587
3588
3589
3590
3591
                mm_embed_inputs = self._gather_mm_embeddings(
                    scheduler_output,
                    shift_computed_tokens=1,
                )
            else:
                mm_embed_inputs = None
3592

3593
            draft_token_ids = self.drafter.propose(
3594
3595
3596
3597
                target_token_ids=target_token_ids,
                target_positions=target_positions,
                target_hidden_states=target_hidden_states,
                next_token_ids=next_token_ids,
3598
                last_token_indices=token_indices_to_sample,
3599
                sampling_metadata=sampling_metadata,
3600
                common_attn_metadata=common_attn_metadata,
3601
                mm_embed_inputs=mm_embed_inputs,
3602
            )
3603

3604
        return draft_token_ids
3605

3606
3607
3608
    def update_config(self, overrides: dict[str, Any]) -> None:
        allowed_config_names = {"load_config", "model_config"}
        for config_name, config_overrides in overrides.items():
3609
3610
            assert config_name in allowed_config_names, (
                f"Config `{config_name}` not supported. "
3611
                f"Allowed configs: {allowed_config_names}"
3612
            )
3613
3614
3615
3616
            config = getattr(self, config_name)
            new_config = update_config(config, config_overrides)
            setattr(self, config_name, new_config)

3617
3618
3619
3620
3621
    def load_model(self, eep_scale_up: bool = False) -> None:
        """
        Args:
            eep_scale_up: the model loading is for elastic EP scale up.
        """
3622
3623
3624
3625
3626
        logger.info_once(
            "Starting to load model %s...",
            self.model_config.model,
            scope="global",
        )
3627
3628
3629
3630
3631
        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)
        )
3632

3633
3634
3635
3636
        if self.parallel_config.enable_eplb:
            self.eplb_state = EplbState(self.parallel_config, self.device)
            eplb_models = 0

3637
3638
3639
3640
3641
3642
3643
3644
3645
3646
        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
3647
                    )
3648
3649
3650
3651
3652
3653
3654
3655
3656
3657
3658
3659
3660
3661
3662
                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,
                        )
3663

3664
3665
3666
3667
3668
3669
3670
3671
3672
3673
3674
3675
3676
3677
3678
3679
3680
3681
3682
3683
3684
3685
                        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
3686

3687
3688
3689
3690
3691
3692
                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"
                        )
3693

3694
3695
3696
3697
3698
3699
3700
3701
3702
3703
3704
3705
3706
3707
3708
3709
3710
3711
3712
3713
3714
3715
3716
3717
3718
                    # 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
3719
        logger.info_once(
3720
            "Model loading took %.4f GiB memory and %.6f seconds",
3721
3722
            self.model_memory_usage / GiB_bytes,
            time_after_load - time_before_load,
3723
            scope="local",
3724
        )
3725
        prepare_communication_buffer_for_model(self.model)
3726
3727
3728
3729
        if (drafter := getattr(self, "drafter", None)) and (
            drafter_model := getattr(drafter, "model", None)
        ):
            prepare_communication_buffer_for_model(drafter_model)
3730
        mm_config = self.model_config.multimodal_config
3731
        self.is_multimodal_pruning_enabled = (
3732
            supports_multimodal_pruning(self.get_model())
3733
3734
            and mm_config is not None
            and mm_config.is_multimodal_pruning_enabled()
3735
        )
3736

3737
        if is_mixture_of_experts(self.model) and self.parallel_config.enable_eplb:
3738
3739
3740
3741
3742
3743
3744
3745
3746
3747
3748
            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(
3749
                self.model,
3750
                self.model_config,
3751
3752
3753
                global_expert_load,
                old_global_expert_indices,
                rank_mapping,
3754
            )
3755
3756
            if self.eplb_state.is_async:
                self.eplb_state.start_async_loop(rank_mapping=rank_mapping)
3757

3758
        if (
3759
3760
            self.vllm_config.compilation_config.mode
            == CompilationMode.STOCK_TORCH_COMPILE
3761
            and supports_dynamo()
3762
        ):
3763
            backend = self.vllm_config.compilation_config.init_backend(self.vllm_config)
3764
            compilation_counter.stock_torch_compile_count += 1
3765
            self.model.compile(fullgraph=True, backend=backend)
3766
            return
3767
        # for other compilation modes, cudagraph behavior is controlled by
3768
3769
3770
        # CudagraphWraper and CudagraphDispatcher of vllm.

        # wrap the model with full cudagraph wrapper if needed.
3771
3772
        cudagraph_mode = self.compilation_config.cudagraph_mode
        assert cudagraph_mode is not None
3773
3774
3775
3776
        if (
            cudagraph_mode.has_full_cudagraphs()
            and not self.parallel_config.use_ubatching
        ):
3777
3778
3779
            self.model = CUDAGraphWrapper(
                self.model, self.vllm_config, runtime_mode=CUDAGraphMode.FULL
            )
3780
        elif self.parallel_config.use_ubatching:
3781
            if cudagraph_mode.has_full_cudagraphs():
3782
3783
3784
                self.model = UBatchWrapper(
                    self.model, self.vllm_config, CUDAGraphMode.FULL, self.device
                )
3785
            else:
3786
3787
3788
                self.model = UBatchWrapper(
                    self.model, self.vllm_config, CUDAGraphMode.NONE, self.device
                )
3789

3790
    def _get_eagle3_aux_layers_from_config(self) -> tuple[int, ...] | None:
3791
3792
3793
3794
3795
3796
3797
3798
3799
3800
3801
3802
3803
3804
3805
3806
3807
3808
3809
3810
3811
3812
3813
        """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

3814
    def reload_weights(self) -> None:
3815
        assert getattr(self, "model", None) is not None, (
3816
            "Cannot reload weights before model is loaded."
3817
        )
3818
3819
        model_loader = get_model_loader(self.load_config)
        logger.info("Reloading weights inplace...")
3820
        model_loader.load_weights(self.get_model(), model_config=self.model_config)
3821

3822
3823
3824
3825
3826
    def save_tensorized_model(
        self,
        tensorizer_config: "TensorizerConfig",
    ) -> None:
        TensorizerLoader.save_model(
3827
            self.get_model(),
3828
            tensorizer_config=tensorizer_config,
3829
            model_config=self.model_config,
3830
3831
        )

3832
3833
3834
    def _get_prompt_logprobs_dict(
        self,
        hidden_states: torch.Tensor,
3835
        num_scheduled_tokens: dict[str, int],
3836
    ) -> dict[str, LogprobsTensors | None]:
3837
        num_prompt_logprobs_dict = self.num_prompt_logprobs
3838
3839
3840
        if not num_prompt_logprobs_dict:
            return {}

3841
        in_progress_dict = self.input_batch.in_progress_prompt_logprobs_cpu
3842
        prompt_logprobs_dict: dict[str, LogprobsTensors | None] = {}
3843
3844
3845
3846
3847

        # 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():
3848
3849
3850
3851
            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
3852
3853
3854

            # Get metadata for this request.
            request = self.requests[req_id]
3855
3856
3857
3858
            if request.prompt_token_ids is None:
                # Prompt logprobs is incompatible with prompt embeddings
                continue

3859
3860
            num_prompt_tokens = len(request.prompt_token_ids)
            prompt_token_ids = torch.tensor(request.prompt_token_ids).to(
3861
3862
                self.device, non_blocking=True
            )
3863

3864
3865
3866
3867
3868
3869
            # 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(
3870
3871
                    num_prompt_tokens - 1, num_prompt_logprobs + 1
                )
3872
3873
                in_progress_dict[req_id] = logprobs_tensors

3874
            # Determine number of logits to retrieve.
3875
3876
            start_idx = request.num_computed_tokens
            start_tok = start_idx + 1
3877
            num_remaining_tokens = num_prompt_tokens - start_tok
3878
            if num_tokens <= num_remaining_tokens:
3879
                # This is a chunk, more tokens remain.
3880
3881
3882
                # 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.
3883
3884
3885
3886
3887
                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)
3888
3889
3890
3891
3892
3893
3894
                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
3895
3896
3897
3898
3899

            # 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]
3900
            offset = self.query_start_loc.np[req_idx].item()
3901
            prompt_hidden_states = hidden_states[offset : offset + num_logits]
3902
            logits = self.model.compute_logits(prompt_hidden_states)
3903
3904
3905
3906

            # 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.
3907
            tgt_token_ids = prompt_token_ids[start_tok : start_tok + num_logits]
3908
3909

            # Compute prompt logprobs.
3910
3911
            logprobs = self.sampler.compute_logprobs(logits)
            token_ids, logprobs, ranks = self.sampler.gather_logprobs(
3912
3913
                logprobs, num_prompt_logprobs, tgt_token_ids
            )
3914
3915

            # Transfer GPU->CPU async.
3916
3917
            chunk_slice = slice(start_idx, start_idx + num_logits)
            logprobs_tensors.logprob_token_ids[chunk_slice].copy_(
3918
3919
3920
                token_ids, non_blocking=True
            )
            logprobs_tensors.logprobs[chunk_slice].copy_(logprobs, non_blocking=True)
3921
            logprobs_tensors.selected_token_ranks[chunk_slice].copy_(
3922
3923
                ranks, non_blocking=True
            )
3924
3925
3926
3927
3928

        # 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]
3929
            del in_progress_dict[req_id]
3930
3931

        # Must synchronize the non-blocking GPU->CPU transfers.
3932
        if prompt_logprobs_dict:
3933
            self._sync_device()
3934
3935
3936

        return prompt_logprobs_dict

3937
3938
    def _get_nans_in_logits(
        self,
3939
        logits: torch.Tensor | None,
3940
3941
3942
3943
3944
3945
3946
3947
3948
3949
3950
    ) -> 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])
3951
3952
3953
                    if num_nans_for_index is not None and req_index < logits.shape[0]
                    else 0
                )
3954
3955
3956
3957
            return num_nans_in_logits
        except IndexError:
            return {}

3958
    @contextmanager
3959
3960
3961
    def maybe_randomize_inputs(
        self, input_ids: torch.Tensor | None, inputs_embeds: torch.Tensor | None
    ):
3962
3963
3964
3965
        """
        Randomize input_ids if VLLM_RANDOMIZE_DP_DUMMY_INPUTS is set.
        This is to help balance expert-selection
         - during profile_run
3966
         - during DP rank dummy run
3967
        """
3968

3969
3970
3971
3972
        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
3973
        elif input_ids is not None:
3974
3975
3976
3977

            @functools.cache
            def rand_input_ids() -> torch.Tensor:
                return torch.randint_like(
3978
                    self.input_ids.gpu,
3979
3980
                    low=0,
                    high=self.model_config.get_vocab_size(),
3981
                )
3982

3983
            logger.debug_once("Randomizing dummy input_ids for DP Rank")
3984
            input_ids.copy_(rand_input_ids()[: input_ids.size(0)], non_blocking=True)
3985
3986
            yield
            input_ids.fill_(0)
3987
3988
3989
3990
3991
3992
3993
3994
3995
3996
3997
3998
3999
4000
4001
        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)
4002

4003
4004
4005
4006
4007
4008
    def _get_mm_dummy_batch(
        self,
        modality: str,
        max_items_per_batch: int,
    ) -> BatchedTensorInputs:
        """Dummy data for profiling and precompiling multimodal models."""
4009
4010
        assert self.mm_budget is not None

4011
        dummy_decoder_data = self.mm_registry.get_decoder_dummy_data(
4012
            model_config=self.model_config,
4013
            seq_len=self.max_model_len,
4014
            mm_counts={modality: 1},
4015
            cache=self.mm_budget.cache,
4016
4017
4018
4019
        )
        dummy_mm_data = dummy_decoder_data.multi_modal_data

        # Result in the maximum GPU consumption of the model
4020
4021
        dummy_mm_item = dummy_mm_data[modality][0]
        dummy_mm_items = [dummy_mm_item] * max_items_per_batch
4022

4023
4024
4025
4026
4027
4028
4029
4030
        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,
            )
        )
4031

4032
4033
4034
4035
    @torch.inference_mode()
    def _dummy_run(
        self,
        num_tokens: int,
4036
        cudagraph_runtime_mode: CUDAGraphMode | None = None,
4037
4038
        force_attention: bool = False,
        uniform_decode: bool = False,
4039
        allow_microbatching: bool = True,
4040
4041
        skip_eplb: bool = False,
        is_profile: bool = False,
4042
        create_mixed_batch: bool = False,
4043
        remove_lora: bool = True,
4044
        activate_lora: bool = False,
Rémi Delacourt's avatar
Rémi Delacourt committed
4045
        is_graph_capturing: bool = False,
4046
    ) -> tuple[torch.Tensor, torch.Tensor]:
4047
4048
4049
4050
4051
4052
4053
        """
        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.
4054
                - if not set will determine the cudagraph mode based on using
4055
                    the self.cudagraph_dispatcher.
4056
4057
4058
4059
                - CUDAGraphMode.NONE: No cudagraph, for warm up and profile run
                - CUDAGraphMode.PIECEWISE: Piecewise cudagraph.
                - CUDAGraphMode.FULL: Full cudagraph, attention metadata is
                    needed.
4060
            force_attention: If True, always create attention metadata. Used to
4061
4062
4063
4064
                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.
4065
4066
            create_mixed_batch: If True, create a mixed batch with both decode
                (1 token) and prefill (multiple tokens) requests.
4067
            remove_lora: If False, dummy LoRAs are not destroyed after the run
4068
            activate_lora: If False, dummy_run is performed without LoRAs.
4069
        """
4070
4071
4072
4073
        assert (
            cudagraph_runtime_mode is None
            or cudagraph_runtime_mode.valid_runtime_modes()
        )
4074

4075
        # If cudagraph_mode.decode_mode() == FULL and
4076
        # cudagraph_mode.separate_routine(). This means that we are using
4077
4078
4079
4080
4081
4082
4083
4084
4085
4086
4087
        # 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.
4088
        max_query_len = self.uniform_decode_query_len if uniform_decode else num_tokens
4089

4090
4091
4092
4093
4094
        # 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
4095
4096
4097
4098
        if create_mixed_batch:
            assert not uniform_decode
            # Create mixed batch:
            # first half decode tokens, second half one prefill
4099
            num_decode_tokens = min(max_num_reqs - 1, num_tokens // 2)
4100
4101
4102
4103
            num_prefill_tokens = num_tokens - num_decode_tokens
            num_reqs = num_decode_tokens + 1

            # Create decode requests (1 token each) followed by prefill request
4104
            num_scheduled_tokens_list = [1] * num_decode_tokens + [num_prefill_tokens]
4105
4106
4107
            # Note: Overriding max_query_len to be the prefill tokens
            max_query_len = num_prefill_tokens
        elif uniform_decode:
4108
            assert not create_mixed_batch
4109
            num_reqs = min(max_num_reqs, cdiv(num_tokens, max_query_len))
4110
4111
            num_scheduled_tokens_list = [max_query_len] * num_reqs
            if num_tokens % max_query_len != 0:
4112
                num_scheduled_tokens_list[-1] = num_tokens % max_query_len
4113
4114
4115
4116
4117
4118
        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

4119
4120
        assert sum(num_scheduled_tokens_list) == num_tokens
        assert len(num_scheduled_tokens_list) == num_reqs
4121
        num_scheduled_tokens = np.array(num_scheduled_tokens_list, dtype=np.int32)
4122
4123
        num_tokens_unpadded = int(num_scheduled_tokens.sum())

4124
        num_sampled_tokens = np.ones(num_reqs, dtype=np.int32)
4125

4126
        _cudagraph_mode, batch_desc, should_ubatch, num_tokens_across_dp, _ = (
4127
4128
4129
4130
4131
4132
4133
4134
4135
4136
4137
4138
4139
4140
4141
4142
4143
4144
            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,
4145
4146
            )
        )
4147
4148
4149

        if cudagraph_runtime_mode is None:
            cudagraph_runtime_mode = _cudagraph_mode
4150
        else:
4151
4152
4153
4154
4155
4156
4157
4158
4159
            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
        )
4160
        ubatch_slices, ubatch_slices_padded = maybe_create_ubatch_slices(
4161
4162
4163
4164
4165
4166
4167
4168
4169
4170
            should_ubatch,
            num_scheduled_tokens,
            num_tokens_padded,
            num_reqs_padded,
            self.vllm_config.parallel_config.num_ubatches,
        )
        logger.debug(
            "ubatch_slices: %s, ubatch_slices_padded: %s",
            ubatch_slices,
            ubatch_slices_padded,
4171
        )
4172

4173
        attn_metadata: PerLayerAttnMetadata | None = None
4174
4175
4176

        # If force_attention is True, we always capture attention. Otherwise,
        # it only happens for cudagraph_runtime_mode=FULL.
4177
        if force_attention or cudagraph_runtime_mode == CUDAGraphMode.FULL:
4178
4179
4180
4181
4182
4183
            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:
4184
                seq_lens = max_query_len  # type: ignore[assignment]
4185
            self.seq_lens.np[:num_reqs] = seq_lens
4186
4187
            self.seq_lens.np[num_reqs:] = 0
            self.seq_lens.copy_to_gpu()
4188

4189
4190
            cum_num_tokens, _ = self._get_cumsum_and_arange(num_scheduled_tokens)
            self.query_start_loc.np[1 : num_reqs + 1] = cum_num_tokens
4191
4192
            self.query_start_loc.copy_to_gpu()

4193
            pad_attn = cudagraph_runtime_mode == CUDAGraphMode.FULL
4194
            attn_metadata, _ = self._build_attention_metadata(
4195
4196
4197
                num_tokens=num_tokens_unpadded,
                num_reqs=num_reqs_padded,
                max_query_len=max_query_len,
4198
                ubatch_slices=ubatch_slices_padded if pad_attn else ubatch_slices,
4199
                for_cudagraph_capture=is_graph_capturing,
4200
            )
4201

4202
        with self.maybe_dummy_run_with_lora(
4203
4204
4205
4206
4207
            self.lora_config,
            num_scheduled_tokens,
            num_sampled_tokens,
            activate_lora,
            remove_lora,
4208
        ):
4209
            # Make sure padding doesn't exceed max_num_tokens
4210
4211
            assert num_tokens_padded <= self.max_num_tokens
            model_kwargs = self._init_model_kwargs(num_tokens_padded)
4212
            if self.supports_mm_inputs and not self.model_config.is_encoder_decoder:
4213
                input_ids = None
4214
                inputs_embeds = self.inputs_embeds.gpu[:num_tokens_padded]
4215
                model_kwargs = {
4216
                    **model_kwargs,
4217
4218
                    **self._dummy_mm_kwargs(num_reqs),
                }
4219
4220
            elif self.enable_prompt_embeds:
                input_ids = None
4221
4222
                inputs_embeds = self.inputs_embeds.gpu[:num_tokens_padded]
                model_kwargs = self._init_model_kwargs(num_tokens_padded)
4223
            else:
4224
                input_ids = self.input_ids.gpu[:num_tokens_padded]
4225
                inputs_embeds = None
4226

4227
            if self.uses_mrope:
4228
                positions = self.mrope_positions.gpu[:, :num_tokens_padded]
4229
            elif self.uses_xdrope_dim > 0:
4230
                positions = self.xdrope_positions.gpu[:, :num_tokens_padded]
4231
            else:
4232
                positions = self.positions.gpu[:num_tokens_padded]
4233
4234
4235
4236
4237
4238
4239
4240
4241

            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,
4242
4243
4244
                            device=self.device,
                        )
                    )
4245
4246

                intermediate_tensors = self.sync_and_slice_intermediate_tensors(
4247
                    num_tokens_padded, None, False
4248
                )
4249

4250
            if ubatch_slices_padded is not None:
4251
4252
4253
                # Adjust values to reflect a single ubatch.
                # TODO(sage,lucas): this is cruft that should be addressed in
                #  the padding refactor.
4254
                num_tokens_padded = ubatch_slices_padded[0].num_tokens
4255
                if num_tokens_across_dp is not None:
4256
                    num_tokens_across_dp[:] = num_tokens_padded
4257

4258
            with (
4259
                self.maybe_randomize_inputs(input_ids, inputs_embeds),
4260
                set_forward_context(
4261
4262
                    attn_metadata,
                    self.vllm_config,
4263
                    num_tokens=num_tokens_padded,
4264
4265
                    num_tokens_across_dp=num_tokens_across_dp,
                    cudagraph_runtime_mode=cudagraph_runtime_mode,
4266
                    batch_descriptor=batch_desc,
4267
                    ubatch_slices=ubatch_slices_padded,
4268
4269
                ),
            ):
4270
                outputs = self.model(
4271
4272
4273
4274
                    input_ids=input_ids,
                    positions=positions,
                    intermediate_tensors=intermediate_tensors,
                    inputs_embeds=inputs_embeds,
4275
                    **model_kwargs,
4276
                )
4277

4278
4279
4280
4281
            if self.use_aux_hidden_state_outputs:
                hidden_states, _ = outputs
            else:
                hidden_states = outputs
4282

4283
            if self.speculative_config and self.speculative_config.use_eagle():
4284
                assert isinstance(self.drafter, EagleProposer)
4285
4286
4287
                # 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.
4288
                use_cudagraphs = (
4289
4290
4291
4292
4293
4294
4295
4296
4297
                    (
                        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
4298
4299
4300
4301
4302
4303
4304
4305
4306
4307
4308

                # 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
4309
                    is_graph_capturing=is_graph_capturing,
4310
                )
4311

4312
4313
4314
4315
4316
4317
4318
4319
4320
4321
4322
        # 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()

4323
4324
4325
4326
4327
4328
4329
4330
4331
4332
        # 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)

4333
        logit_indices = np.cumsum(num_scheduled_tokens) - 1
4334
4335
4336
4337
        logit_indices_device = torch.from_numpy(logit_indices).to(
            self.device, non_blocking=True
        )
        return hidden_states, hidden_states[logit_indices_device]
4338
4339
4340
4341
4342
4343

    @torch.inference_mode()
    def _dummy_sampler_run(
        self,
        hidden_states: torch.Tensor,
    ) -> torch.Tensor:
4344
4345
4346
4347
        # 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)
4348

4349
        logits = self.model.compute_logits(hidden_states)
4350
4351
        num_reqs = logits.size(0)

4352
        dummy_tensors = lambda v: torch.full((num_reqs,), v, device=self.device)
4353
4354
4355
4356
4357
4358
4359
4360
4361
4362
4363
4364
4365
4366
4367

        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)],
4368
            spec_token_ids=[[] for _ in range(num_reqs)],
4369
4370
            allowed_token_ids_mask=None,
            bad_words_token_ids={},
4371
            logitsprocs=LogitsProcessors(),
4372
        )
4373
        try:
4374
4375
4376
            sampler_output = self.sampler(
                logits=logits, sampling_metadata=dummy_metadata
            )
4377
        except RuntimeError as e:
4378
            if "out of memory" in str(e):
4379
4380
4381
4382
                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 "
4383
4384
                    "initializing the engine."
                ) from e
4385
4386
            else:
                raise e
4387
        if self.speculative_config:
4388
4389
            draft_token_ids = [[0] for _ in range(num_reqs)]
            dummy_spec_decode_metadata = SpecDecodeMetadata.make_dummy(
4390
4391
                draft_token_ids, self.device
            )
4392
4393
4394
4395
4396
4397

            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
4398
4399
4400
4401
4402
            logits = torch.randn(
                num_tokens + num_reqs,
                logits.shape[-1],
                device=self.device,
                dtype=logits.dtype,
4403
            )
4404
4405
4406
            self.rejection_sampler(
                dummy_spec_decode_metadata,
                draft_probs,
4407
                logits,
4408
4409
                dummy_metadata,
            )
4410
        return sampler_output
4411

4412
    def _dummy_pooler_run_task(
4413
4414
        self,
        hidden_states: torch.Tensor,
4415
4416
        task: PoolingTask,
    ) -> PoolerOutput:
4417
4418
4419
4420
4421
4422
4423
4424
4425
4426
4427
        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

4428
        dummy_prompt_lens = torch.tensor(
4429
4430
            num_scheduled_tokens_list,
            device="cpu",
4431
        )
4432
4433
4434
        dummy_token_ids = torch.zeros(
            (num_reqs, req_num_tokens), dtype=torch.int32, device=self.device
        )
4435

4436
        model = cast(VllmModelForPooling, self.get_model())
4437
        dummy_pooling_params = PoolingParams(task=task)
4438
        dummy_pooling_params.verify(task=task, model_config=self.model_config)
4439
        to_update = model.pooler.get_pooling_updates(task)
4440
4441
        to_update.apply(dummy_pooling_params)

4442
        dummy_metadata = PoolingMetadata(
4443
4444
4445
            prompt_lens=dummy_prompt_lens,
            prompt_token_ids=dummy_token_ids,
            pooling_params=[dummy_pooling_params] * num_reqs,
4446
            pooling_states=[PoolingStates() for i in range(num_reqs)],
4447
        )
4448

4449
        dummy_metadata.build_pooling_cursor(
4450
4451
4452
            num_scheduled_tokens_list,
            seq_lens_cpu=dummy_prompt_lens,
            device=hidden_states.device,
4453
        )
4454

4455
        try:
4456
4457
4458
            return model.pooler(
                hidden_states=hidden_states, pooling_metadata=dummy_metadata
            )
4459
        except RuntimeError as e:
4460
            if "out of memory" in str(e):
4461
                raise RuntimeError(
4462
4463
4464
                    "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 "
4465
4466
                    "initializing the engine."
                ) from e
4467
4468
            else:
                raise e
4469
4470
4471
4472
4473
4474
4475

    @torch.inference_mode()
    def _dummy_pooler_run(
        self,
        hidden_states: torch.Tensor,
    ) -> PoolerOutput:
        # Find the task that has the largest output for subsequent steps
4476
4477
4478
        supported_pooling_tasks = self.get_supported_pooling_tasks()

        if not supported_pooling_tasks:
4479
4480
4481
4482
4483
4484
            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."
            )
4485

4486
        output_size = dict[PoolingTask, float]()
4487
        for task in supported_pooling_tasks:
4488
4489
            # Run a full batch with each task to ensure none of them OOMs
            output = self._dummy_pooler_run_task(hidden_states, task)
4490
            output_size[task] = sum(o.nbytes for o in output)
4491
4492
4493
4494
            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)
4495

4496
    def profile_run(self) -> None:
4497
        # Profile with multimodal encoder & encoder cache.
4498
        if self.supports_mm_inputs:
4499
4500
            mm_config = self.model_config.multimodal_config
            if mm_config is not None and mm_config.skip_mm_profiling:
4501
                logger.info(
4502
                    "Skipping memory profiling for multimodal encoder and "
4503
4504
                    "encoder cache."
                )
4505
4506
4507
4508
4509
4510
4511
4512
            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.
4513
                    dummy_modality = mm_budget.get_modality_with_max_tokens()
4514
4515
4516
                    max_mm_items_per_batch = mm_budget.max_items_per_batch_by_modality[
                        dummy_modality
                    ]
4517
4518
4519
4520
4521
4522
4523
4524
4525

                    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,
                    )
4526

4527
4528
4529
4530
4531
                    # Create dummy batch of multimodal inputs.
                    batched_dummy_mm_inputs = self._get_mm_dummy_batch(
                        dummy_modality,
                        max_mm_items_per_batch,
                    )
4532

4533
                    # Run multimodal encoder.
4534
                    dummy_encoder_outputs = self.model.embed_multimodal(
4535
4536
                        **batched_dummy_mm_inputs
                    )
4537

4538
4539
4540
4541
                    sanity_check_mm_encoder_outputs(
                        dummy_encoder_outputs,
                        expected_num_items=max_mm_items_per_batch,
                    )
4542
4543
                    for i, output in enumerate(dummy_encoder_outputs):
                        self.encoder_cache[f"tmp_{i}"] = output
4544

4545
        # Add `is_profile` here to pre-allocate communication buffers
4546
4547
4548
        hidden_states, last_hidden_states = self._dummy_run(
            self.max_num_tokens, is_profile=True
        )
4549
        if get_pp_group().is_last_rank:
4550
4551
4552
4553
            if self.is_pooling_model:
                output = self._dummy_pooler_run(hidden_states)
            else:
                output = self._dummy_sampler_run(last_hidden_states)
4554
        else:
4555
            output = None
4556
        self._sync_device()
4557
        del hidden_states, output
4558
        self.encoder_cache.clear()
4559
        gc.collect()
4560

4561
    def capture_model(self) -> int:
4562
        if self.compilation_config.cudagraph_mode == CUDAGraphMode.NONE:
4563
            logger.warning(
4564
                "Skipping CUDA graph capture. To turn on CUDA graph capture, "
4565
4566
                "ensure `cudagraph_mode` was not manually set to `NONE`"
            )
4567
            return 0
4568

4569
4570
        compilation_counter.num_gpu_runner_capture_triggers += 1

4571
4572
        start_time = time.perf_counter()

4573
4574
4575
4576
4577
4578
4579
4580
4581
4582
4583
4584
4585
4586
        @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()
4587
                    gc.collect()
4588

4589
4590
4591
        # 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.
4592
        set_cudagraph_capturing_enabled(True)
4593
        with freeze_gc(), graph_capture(device=self.device):
4594
            start_free_gpu_memory = torch.cuda.mem_get_info()[0]
4595
            cudagraph_mode = self.compilation_config.cudagraph_mode
4596
            assert cudagraph_mode is not None
4597
4598
4599
4600
4601
4602
4603
4604
4605

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

4606
4607
            if cudagraph_mode.mixed_mode() != CUDAGraphMode.NONE:
                cudagraph_runtime_mode = cudagraph_mode.mixed_mode()
4608
                # make sure we capture the largest batch size first
4609
4610
4611
                compilation_cases = list(
                    product(reversed(self.cudagraph_batch_sizes), lora_cases)
                )
4612
4613
4614
                self._capture_cudagraphs(
                    compilation_cases,
                    cudagraph_runtime_mode=cudagraph_runtime_mode,
4615
4616
                    uniform_decode=False,
                )
4617

4618
4619
            # Capture full cudagraph for uniform decode batches if we
            # don't already have full mixed prefill-decode cudagraphs.
4620
4621
4622
4623
4624
4625
4626
            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
                )
4627
                decode_cudagraph_batch_sizes = [
4628
4629
                    x
                    for x in self.cudagraph_batch_sizes
4630
                    if max_num_tokens >= x >= self.uniform_decode_query_len
4631
                ]
4632
4633
4634
                compilation_cases_decode = list(
                    product(reversed(decode_cudagraph_batch_sizes), lora_cases)
                )
4635
4636
4637
                self._capture_cudagraphs(
                    compilation_cases=compilation_cases_decode,
                    cudagraph_runtime_mode=CUDAGraphMode.FULL,
4638
4639
                    uniform_decode=True,
                )
4640

4641
4642
4643
            torch.cuda.synchronize()
            end_free_gpu_memory = torch.cuda.mem_get_info()[0]

4644
4645
4646
        # 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
4647
        # we may do lazy capturing in future that still allows capturing
4648
4649
        # after here.
        set_cudagraph_capturing_enabled(False)
4650

4651
4652
4653
4654
        # Lock workspace to prevent resizing during execution.
        # Max workspace sizes should have been captured during warmup/profiling.
        lock_workspace()

4655
4656
4657
4658
        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.
4659
        logger.info_once(
4660
4661
4662
            "Graph capturing finished in %.0f secs, took %.2f GiB",
            elapsed_time,
            cuda_graph_size / (1 << 30),
4663
            scope="local",
4664
        )
4665
        return cuda_graph_size
4666

4667
4668
    def _capture_cudagraphs(
        self,
4669
        compilation_cases: list[tuple[int, bool]],
4670
4671
4672
4673
4674
4675
4676
        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}"
4677
4678
4679
4680
4681
4682
4683
4684

        # 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",
4685
4686
4687
                    cudagraph_runtime_mode.name,
                ),
            )
4688

4689
        # We skip EPLB here since we don't want to record dummy metrics
4690
        for num_tokens, activate_lora in compilation_cases:
4691
            # We currently only capture ubatched graphs when its a FULL
4692
4693
4694
            # 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
4695
            allow_microbatching = (
4696
                self.parallel_config.use_ubatching
4697
4698
                and cudagraph_runtime_mode == CUDAGraphMode.FULL
                and uniform_decode
4699
4700
4701
4702
4703
                and check_ubatch_thresholds(
                    config=self.vllm_config.parallel_config,
                    num_tokens=num_tokens,
                    uniform_decode=uniform_decode,
                )
4704
            )
4705

4706
4707
4708
4709
4710
4711
            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.
4712
4713
4714
4715
4716
4717
4718
4719
4720
                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,
4721
                    activate_lora=activate_lora,
4722
4723
4724
4725
4726
4727
4728
4729
                )
            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,
4730
                activate_lora=activate_lora,
Rémi Delacourt's avatar
Rémi Delacourt committed
4731
                is_graph_capturing=True,
4732
            )
4733
        self.maybe_remove_all_loras(self.lora_config)
4734

4735
4736
4737
4738
    def initialize_attn_backend(self, kv_cache_config: KVCacheConfig) -> None:
        """
        Initialize the attention backends and attention metadata builders.
        """
4739
        assert len(self.attn_groups) == 0, "Attention backends are already initialized"
4740

4741
4742
4743
4744
4745
4746
        class AttentionGroupKey(NamedTuple):
            attn_backend: type[AttentionBackend]
            kv_cache_spec: KVCacheSpec

        def get_attn_backends_for_group(
            kv_cache_group_spec: KVCacheGroupSpec,
4747
        ) -> tuple[dict[AttentionGroupKey, list[str]], set[type[AttentionBackend]]]:
4748
            layer_type = cast(type[Any], AttentionLayerBase)
4749
            layers = get_layers_from_vllm_config(
4750
                self.vllm_config, layer_type, kv_cache_group_spec.layer_names
4751
            )
4752
4753
            attn_backends = {}
            attn_backend_layers = defaultdict(list)
4754
            # Dedupe based on full class name; this is a bit safer than
4755
4756
4757
4758
            # 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.
4759
            for layer_name in kv_cache_group_spec.layer_names:
4760
                attn_backend = layers[layer_name].get_attn_backend()
4761
4762
4763
4764

                if layer_name in self.kv_sharing_fast_prefill_eligible_layers:
                    attn_backend = create_fast_prefill_custom_backend(
                        "FastPrefill",
4765
                        attn_backend,  # type: ignore[arg-type]
4766
4767
                    )

4768
4769
4770
                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):
4771
                    layer_kv_cache_spec = layer_kv_cache_spec.kv_cache_specs[layer_name]
4772
                key = (full_cls_name, layer_kv_cache_spec)
4773
4774
4775
                attn_backends[key] = AttentionGroupKey(
                    attn_backend, layer_kv_cache_spec
                )
4776
                attn_backend_layers[key].append(layer_name)
4777
4778
4779
4780
            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()),
            )
4781
4782

        def create_attn_groups(
4783
            attn_backends_map: dict[AttentionGroupKey, list[str]],
4784
            kv_cache_group_id: int,
4785
4786
        ) -> list[AttentionGroup]:
            attn_groups: list[AttentionGroup] = []
4787
            for (attn_backend, kv_cache_spec), layer_names in attn_backends_map.items():
4788
                attn_group = AttentionGroup(
4789
                    attn_backend,
4790
                    layer_names,
4791
                    kv_cache_spec,
4792
                    kv_cache_group_id,
4793
4794
                )

4795
4796
4797
                attn_groups.append(attn_group)
            return attn_groups

4798
        attention_backend_maps = []
4799
        attention_backend_list = []
4800
        for kv_cache_group_spec in kv_cache_config.kv_cache_groups:
4801
            attn_backends = get_attn_backends_for_group(kv_cache_group_spec)
4802
            attention_backend_maps.append(attn_backends[0])
4803
            attention_backend_list.append(attn_backends[1])
4804
4805

        # Resolve cudagraph_mode before actually initialize metadata_builders
4806
4807
4808
        self._check_and_update_cudagraph_mode(
            attention_backend_list, kv_cache_config.kv_cache_groups
        )
4809

4810
4811
4812
        # Check if attention backend supports PCP&DCP and related features.
        check_attention_cp_compatibility(self.vllm_config)

4813
4814
        for i, attn_backend_map in enumerate(attention_backend_maps):
            self.attn_groups.append(create_attn_groups(attn_backend_map, i))
4815

4816
4817
4818
4819
4820
4821
4822
4823
4824
4825
4826
4827
4828
4829
4830
    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
4831
4832
                    if not self.parallel_config.use_ubatching
                    else self.parallel_config.num_ubatches,
4833
                )
co63oc's avatar
co63oc committed
4834
        # Calculate reorder batch threshold (if needed)
4835
4836
        # Note (tdoublep): do this *after* constructing builders,
        # because some of them change the threshold at init time.
4837
4838
        self.calculate_reorder_batch_threshold()

4839
    def _check_and_update_cudagraph_mode(
4840
4841
4842
        self,
        attention_backends: list[set[type[AttentionBackend]]],
        kv_cache_groups: list[KVCacheGroupSpec],
4843
    ) -> None:
4844
        """
4845
        Resolve the cudagraph_mode when there are multiple attention
4846
        groups with potential conflicting CUDA graph support.
4847
4848
4849
        Then initialize the cudagraph_dispatcher based on the resolved
        cudagraph_mode.
        """
4850
        min_cg_support = AttentionCGSupport.ALWAYS
4851
        min_cg_backend_name = None
4852

4853
4854
4855
4856
4857
4858
4859
4860
4861
4862
4863
4864
        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__
4865
4866
        # Flexible resolve the cudagraph mode
        cudagraph_mode = self.compilation_config.cudagraph_mode
4867
        assert cudagraph_mode is not None
4868
        # check cudagraph for mixed batch is supported
4869
4870
4871
4872
4873
4874
        if (
            cudagraph_mode.mixed_mode() == CUDAGraphMode.FULL
            and min_cg_support != AttentionCGSupport.ALWAYS
        ):
            msg = (
                f"CUDAGraphMode.{cudagraph_mode.name} is not supported "
4875
                f"with {min_cg_backend_name} backend (support: "
4876
4877
                f"{min_cg_support})"
            )
4878
4879
            if min_cg_support == AttentionCGSupport.NEVER:
                # if not supported any full cudagraphs, just raise it.
4880
4881
                msg += (
                    "; please try cudagraph_mode=PIECEWISE, and "
4882
                    "make sure compilation mode is VLLM_COMPILE"
4883
                )
4884
4885
4886
4887
4888
                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"
4889
                cudagraph_mode = self.compilation_config.cudagraph_mode = (
4890
                    CUDAGraphMode.FULL_AND_PIECEWISE
4891
                )
4892
4893
            else:
                msg += "; setting cudagraph_mode=FULL_DECODE_ONLY"
4894
                cudagraph_mode = self.compilation_config.cudagraph_mode = (
4895
                    CUDAGraphMode.FULL_DECODE_ONLY
4896
                )
4897
4898
            logger.warning(msg)

4899
        # check that if we are doing decode full-cudagraphs it is supported
4900
4901
4902
4903
4904
4905
        if (
            cudagraph_mode.decode_mode() == CUDAGraphMode.FULL
            and min_cg_support == AttentionCGSupport.NEVER
        ):
            msg = (
                f"CUDAGraphMode.{cudagraph_mode.name} is not supported "
4906
                f"with {min_cg_backend_name} backend (support: "
4907
4908
                f"{min_cg_support})"
            )
4909
            if self.compilation_config.mode == CompilationMode.VLLM_COMPILE and (
4910
4911
4912
4913
4914
                self.compilation_config.splitting_ops_contain_attention()
                or self.compilation_config.use_inductor_graph_partition
            ):
                msg += (
                    "; setting cudagraph_mode=PIECEWISE because "
4915
                    "attention is compiled piecewise"
4916
4917
                )
                cudagraph_mode = self.compilation_config.cudagraph_mode = (
4918
                    CUDAGraphMode.PIECEWISE
4919
                )
4920
            else:
4921
4922
                msg += (
                    "; setting cudagraph_mode=NONE because "
4923
                    "attention is not compiled piecewise"
4924
4925
                )
                cudagraph_mode = self.compilation_config.cudagraph_mode = (
4926
                    CUDAGraphMode.NONE
4927
                )
4928
4929
            logger.warning(msg)

4930
4931
        # check that if we are doing spec-decode + decode full-cudagraphs it is
        # supported
4932
4933
4934
4935
4936
4937
4938
4939
        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 "
4940
                f"{min_cg_backend_name} (support: {min_cg_support})"
4941
            )
4942
4943
            if self.compilation_config.splitting_ops_contain_attention():
                msg += "; setting cudagraph_mode=PIECEWISE"
4944
                cudagraph_mode = self.compilation_config.cudagraph_mode = (
4945
                    CUDAGraphMode.PIECEWISE
4946
                )
4947
4948
            else:
                msg += "; setting cudagraph_mode=NONE"
4949
                cudagraph_mode = self.compilation_config.cudagraph_mode = (
4950
                    CUDAGraphMode.NONE
4951
                )
4952
4953
4954
4955
            logger.warning(msg)

        # double check that we can support full cudagraph if they are requested
        # even after automatic downgrades
4956
4957
4958
4959
4960
4961
        if (
            cudagraph_mode.has_full_cudagraphs()
            and min_cg_support == AttentionCGSupport.NEVER
        ):
            raise ValueError(
                f"CUDAGraphMode.{cudagraph_mode.name} is not "
4962
                f"supported with {min_cg_backend_name} backend ("
4963
4964
                f"support:{min_cg_support}) "
                "; please try cudagraph_mode=PIECEWISE, "
4965
                "and make sure compilation mode is VLLM_COMPILE"
4966
            )
4967

4968
4969
4970
4971
        # 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
4972
        # Will be removed in the near future when we have separate cudagraph capture
4973
4974
4975
4976
4977
4978
4979
4980
4981
        # 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
            )
4982
4983
4984
4985
            capture_sizes = self.compilation_config.cudagraph_capture_sizes
            self.cudagraph_batch_sizes = (
                capture_sizes if capture_sizes is not None else []
            )
4986

4987
4988
        # Trigger cudagraph dispatching keys initialization after
        # resolved cudagraph mode.
4989
        self.compilation_config.cudagraph_mode = cudagraph_mode
4990
        self.cudagraph_dispatcher.initialize_cudagraph_keys(
4991
            cudagraph_mode, self.uniform_decode_query_len
4992
        )
4993

4994
4995
    def calculate_reorder_batch_threshold(self) -> None:
        """
4996
4997
4998
4999
        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.
5000
        """
5001
5002
        min_none_high = lambda a, b: a if b is None else b if a is None else min(a, b)

5003
        reorder_batch_thresholds: list[int | None] = [
5004
5005
5006
            group.get_metadata_builder().reorder_batch_threshold
            for group in self._attn_group_iterator()
        ]
5007
5008
5009
5010
5011
        # 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
5012
        self.reorder_batch_threshold = reduce(min_none_high, reorder_batch_thresholds)  # type: ignore[assignment]
5013

5014
5015
5016
    @staticmethod
    def select_common_block_size(
        kv_manager_block_size: int, attn_groups: list[AttentionGroup]
5017
5018
    ) -> int:
        """
5019
5020
5021
5022
5023
        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.
5024
5025
5026
5027
5028
5029

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

        Returns:
5030
            The selected block size
5031
5032

        Raises:
5033
            ValueError: If no valid block size found
5034
5035
        """

5036
5037
5038
5039
5040
5041
5042
5043
        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
5044
                for supported_size in backend.get_supported_kernel_block_sizes():
5045
5046
5047
5048
5049
5050
5051
5052
5053
5054
5055
5056
5057
5058
5059
5060
5061
5062
5063
5064
5065
5066
5067
5068
5069
5070
5071
5072
5073
5074
                    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
5075
            for supported_size in backend.get_supported_kernel_block_sizes()
5076
5077
            if isinstance(supported_size, int)
        )
5078

5079
5080
5081
5082
5083
5084
        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}. ")
5085

5086
5087
5088
    def may_reinitialize_input_batch(
        self, kv_cache_config: KVCacheConfig, kernel_block_sizes: list[int]
    ) -> None:
5089
5090
5091
5092
5093
5094
5095
        """
        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.
5096
            kernel_block_sizes: The kernel block sizes for each KV cache group.
5097
5098
5099
5100
        """
        block_sizes = [
            kv_cache_group.kv_cache_spec.block_size
            for kv_cache_group in kv_cache_config.kv_cache_groups
5101
            if not isinstance(kv_cache_group.kv_cache_spec, EncoderOnlyAttentionSpec)
5102
        ]
5103
5104
5105
5106

        if block_sizes != [self.cache_config.block_size] or kernel_block_sizes != [
            self.cache_config.block_size
        ]:
5107
5108
5109
            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
5110
5111
                "for more details."
            )
5112
5113
            self.input_batch = InputBatch(
                max_num_reqs=self.max_num_reqs,
5114
                max_model_len=max(self.max_model_len, self.max_encoder_len),
5115
5116
5117
5118
5119
                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,
5120
                kernel_block_sizes=kernel_block_sizes,
5121
                is_spec_decode=bool(self.vllm_config.speculative_config),
5122
                logitsprocs=self.input_batch.logitsprocs,
5123
                logitsprocs_need_output_token_ids=self.input_batch.logitsprocs_need_output_token_ids,
5124
                is_pooling_model=self.is_pooling_model,
5125
                num_speculative_tokens=self.num_spec_tokens,
5126
5127
            )

5128
    def _allocate_kv_cache_tensors(
5129
5130
        self, kv_cache_config: KVCacheConfig
    ) -> dict[str, torch.Tensor]:
5131
        """
5132
5133
5134
        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.

5135
        Args:
5136
            kv_cache_config: The KV cache config
5137
        Returns:
5138
            dict[str, torch.Tensor]: A map between layer names to their
5139
            corresponding memory buffer for KV cache.
5140
        """
5141
5142
        kv_cache_raw_tensors: dict[str, torch.Tensor] = {}
        for kv_cache_tensor in kv_cache_config.kv_cache_tensors:
5143
5144
5145
            tensor = torch.zeros(
                kv_cache_tensor.size, dtype=torch.int8, device=self.device
            )
5146
5147
5148
5149
5150
            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:
5151
5152
5153
5154
            for layer_name in group.layer_names:
                if layer_name in self.runner_only_attn_layers:
                    continue
                layer_names.add(layer_name)
5155
5156
5157
        assert layer_names == set(kv_cache_raw_tensors.keys()), (
            "Some layers are not correctly initialized"
        )
5158
5159
        return kv_cache_raw_tensors

5160
5161
5162
    def _attn_group_iterator(self) -> Iterator[AttentionGroup]:
        return itertools.chain.from_iterable(self.attn_groups)

5163
    def _kv_cache_spec_attn_group_iterator(self) -> Iterator[AttentionGroup]:
5164
5165
        if not self.kv_cache_config.kv_cache_groups:
            return
5166
5167
        for attn_groups in self.attn_groups:
            yield from attn_groups
5168

5169
5170
5171
5172
5173
5174
5175
5176
5177
5178
5179
5180
5181
5182
5183
    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 = []
5184
        for kv_cache_gid, kv_cache_group in enumerate(kv_cache_config.kv_cache_groups):
5185
5186
5187
5188
5189
5190
            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):
5191
                continue
5192
            elif isinstance(kv_cache_spec, AttentionSpec):
5193
5194
5195
                # This is an attention backend that supports virtual
                # block splitting. Get the supported block sizes from
                # all backends in the group.
5196
                attn_groups = self.attn_groups[kv_cache_gid]
5197
                kv_manager_block_size = kv_cache_group.kv_cache_spec.block_size
5198
                selected_kernel_size = self.select_common_block_size(
5199
5200
5201
                    kv_manager_block_size, attn_groups
                )
                kernel_block_sizes.append(selected_kernel_size)
5202
            elif isinstance(kv_cache_spec, MambaSpec):
5203
5204
                # This is likely Mamba or other non-attention cache,
                # no splitting.
5205
                kernel_block_sizes.append(kv_cache_spec.block_size)
5206
5207
5208
5209
5210
5211
            else:
                raise NotImplementedError(
                    f"unknown kv cache spec {kv_cache_group.kv_cache_spec}"
                )
        return kernel_block_sizes

5212
5213
5214
5215
    def _reshape_kv_cache_tensors(
        self,
        kv_cache_config: KVCacheConfig,
        kv_cache_raw_tensors: dict[str, torch.Tensor],
5216
        kernel_block_sizes: list[int],
5217
    ) -> dict[str, torch.Tensor]:
5218
        """
5219
        Reshape the KV cache tensors to the desired shape and dtype.
5220

5221
        Args:
5222
5223
            kv_cache_config: The KV cache config
            kv_cache_raw_tensors: The KV cache buffer of each layer, with
5224
                correct size but uninitialized shape.
5225
            kernel_block_sizes: The kernel block sizes for each KV cache group.
5226
        Returns:
5227
            Dict[str, torch.Tensor]: A map between layer names to their
5228
5229
            corresponding memory buffer for KV cache.
        """
5230
        kv_caches: dict[str, torch.Tensor] = {}
5231
        has_attn, has_mamba = False, False
5232
5233
        for group in self._kv_cache_spec_attn_group_iterator():
            kv_cache_spec = group.kv_cache_spec
5234
            attn_backend = group.backend
5235
5236
5237
5238
            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]
5239
            for layer_name in group.layer_names:
5240
5241
                if layer_name in self.runner_only_attn_layers:
                    continue
5242
5243
                raw_tensor = kv_cache_raw_tensors[layer_name]
                assert raw_tensor.numel() % kv_cache_spec.page_size_bytes == 0
5244
                num_blocks = raw_tensor.numel() // kv_cache_spec.page_size_bytes
5245
                if isinstance(kv_cache_spec, AttentionSpec):
5246
                    has_attn = True
5247
5248
                    num_blocks_per_kv_block = (
                        kv_cache_spec.block_size // kernel_block_size
5249
5250
5251
                    )
                    kernel_num_blocks = num_blocks * num_blocks_per_kv_block

5252
                    kv_cache_shape = attn_backend.get_kv_cache_shape(
5253
                        kernel_num_blocks,
5254
                        kernel_block_size,
5255
5256
                        kv_cache_spec.num_kv_heads,
                        kv_cache_spec.head_size,
5257
5258
                        cache_dtype_str=self.cache_config.cache_dtype,
                    )
5259
                    dtype = kv_cache_spec.dtype
5260
                    try:
5261
                        kv_cache_stride_order = attn_backend.get_kv_cache_stride_order()
5262
                        assert len(kv_cache_stride_order) == len(kv_cache_shape)
5263
                    except (AttributeError, NotImplementedError):
5264
                        kv_cache_stride_order = tuple(range(len(kv_cache_shape)))
5265
5266
5267
5268
5269
                    # 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.
5270
5271
5272
                    kv_cache_shape = tuple(
                        kv_cache_shape[i] for i in kv_cache_stride_order
                    )
5273
5274
5275
5276
5277
                    # Maintain original KV shape view.
                    inv_order = [
                        kv_cache_stride_order.index(i)
                        for i in range(len(kv_cache_stride_order))
                    ]
5278
5279
5280
5281
5282
5283
                    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
5284
                elif isinstance(kv_cache_spec, MambaSpec):
5285
                    has_mamba = True
Chen Zhang's avatar
Chen Zhang committed
5286
5287
                    raw_tensor = kv_cache_raw_tensors[layer_name]
                    state_tensors = []
5288
                    storage_offset_bytes = 0
5289
                    for shape, dtype in zip(kv_cache_spec.shapes, kv_cache_spec.dtypes):
5290
5291
                        dtype_size = get_dtype_size(dtype)
                        num_element_per_page = (
5292
5293
                            kv_cache_spec.page_size_bytes // dtype_size
                        )
Chen Zhang's avatar
Chen Zhang committed
5294
                        target_shape = (num_blocks, *shape)
5295
5296
                        stride = torch.empty(target_shape).stride()
                        target_stride = (num_element_per_page, *stride[1:])
5297
                        assert storage_offset_bytes % dtype_size == 0
5298
5299
5300
5301
                        tensor = torch.as_strided(
                            raw_tensor.view(dtype),
                            size=target_shape,
                            stride=target_stride,
5302
                            storage_offset=storage_offset_bytes // dtype_size,
5303
                        )
Chen Zhang's avatar
Chen Zhang committed
5304
                        state_tensors.append(tensor)
5305
                        storage_offset_bytes += stride[0] * dtype_size
5306
5307

                    kv_caches[layer_name] = state_tensors
5308
                else:
5309
                    raise NotImplementedError
5310
5311

        if has_attn and has_mamba:
5312
            self._update_hybrid_attention_mamba_layout(kv_caches)
5313

5314
5315
        return kv_caches

5316
    def _update_hybrid_attention_mamba_layout(
5317
5318
        self, kv_caches: dict[str, torch.Tensor]
    ) -> None:
5319
        """
5320
5321
        Update the layout of attention layers from (2, num_blocks, ...) to
        (num_blocks, 2, ...).
5322
5323

        Args:
5324
            kv_caches: The KV cache buffer of each layer.
5325
5326
        """

5327
5328
        for group in self._kv_cache_spec_attn_group_iterator():
            kv_cache_spec = group.kv_cache_spec
5329
            for layer_name in group.layer_names:
5330
                kv_cache = kv_caches[layer_name]
5331
5332
5333
5334
                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 "
5335
                        f"a tensor of shape {kv_cache.shape}"
5336
                    )
5337
                    hidden_size = kv_cache.shape[2:].numel()
5338
5339
5340
5341
                    kv_cache.as_strided_(
                        size=kv_cache.shape,
                        stride=(hidden_size, 2 * hidden_size, *kv_cache.stride()[2:]),
                    )
5342

5343
    def initialize_kv_cache_tensors(
5344
        self, kv_cache_config: KVCacheConfig, kernel_block_sizes: list[int]
5345
    ) -> dict[str, torch.Tensor]:
5346
5347
5348
5349
5350
        """
        Initialize the memory buffer for KV cache.

        Args:
            kv_cache_config: The KV cache config
5351
5352
            kernel_block_sizes: The kernel block sizes for each KV cache group.

5353
        Returns:
5354
            Dict[str, torch.Tensor]: A map between layer names to their
5355
5356
            corresponding memory buffer for KV cache.
        """
5357
5358
5359
5360
5361
5362
5363
5364
5365
5366
5367
5368
5369
5370
5371
5372
5373
5374
5375
5376
5377
5378
5379
5380

        # 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
            )
5381

5382
        # Set up cross-layer KV cache sharing
5383
5384
        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)
5385
5386
            kv_caches[layer_name] = kv_caches[target_layer_name]

5387
5388
5389
5390
5391
5392
5393
5394
5395
        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,
        )
5396
5397
5398
        return kv_caches

    def maybe_add_kv_sharing_layers_to_kv_cache_groups(
5399
5400
        self, kv_cache_config: KVCacheConfig
    ) -> None:
5401
5402
5403
5404
5405
5406
5407
5408
5409
5410
5411
5412
5413
5414
5415
5416
5417
5418
        """
        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.
5419
            attn_layers = get_layers_from_vllm_config(self.vllm_config, Attention)
5420
5421
            for layer_name in reversed(attn_layers):
                if layer_name in self.shared_kv_cache_layers:
5422
                    self.kv_sharing_fast_prefill_eligible_layers.add(layer_name)
5423
5424
                else:
                    break
5425

5426
5427
5428
5429
5430
5431
5432
    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
        """
5433
        kv_cache_config = deepcopy(kv_cache_config)
5434
        self.kv_cache_config = kv_cache_config
5435
        self.may_add_encoder_only_layers_to_kv_cache_config()
5436
        self.maybe_add_kv_sharing_layers_to_kv_cache_groups(kv_cache_config)
5437
        self.initialize_attn_backend(kv_cache_config)
5438
5439
5440
5441
5442
5443
        # 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)
5444
5445
5446
5447

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

5448
        # Reinitialize need to after initialize_attn_backend
5449
5450
5451
5452
        self.may_reinitialize_input_batch(kv_cache_config, kernel_block_sizes)
        kv_caches = self.initialize_kv_cache_tensors(
            kv_cache_config, kernel_block_sizes
        )
5453

5454
5455
5456
5457
5458
5459
        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
5460
        if has_kv_transfer_group():
5461
            kv_transfer_group = get_kv_transfer_group()
5462
5463
5464
5465
5466
5467
5468
            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)
5469
            kv_transfer_group.set_host_xfer_buffer_ops(copy_kv_blocks)
Robert Shaw's avatar
Robert Shaw committed
5470

5471
5472
5473
5474
5475
    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
5476
        encoder_only_attn_specs: dict[AttentionSpec, list[str]] = defaultdict(list)
5477
5478
5479
        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:
5480
                attn_spec: AttentionSpec = EncoderOnlyAttentionSpec(
5481
5482
5483
                    block_size=block_size,
                    num_kv_heads=attn_module.num_kv_heads,
                    head_size=attn_module.head_size,
5484
5485
                    dtype=self.kv_cache_dtype,
                )
5486
5487
5488
                encoder_only_attn_specs[attn_spec].append(layer_name)
                self.runner_only_attn_layers.add(layer_name)
        if len(encoder_only_attn_specs) > 0:
5489
5490
5491
            assert len(encoder_only_attn_specs) == 1, (
                "Only support one encoder-only attention spec now"
            )
5492
5493
            spec, layer_names = encoder_only_attn_specs.popitem()
            self.kv_cache_config.kv_cache_groups.append(
5494
5495
                KVCacheGroupSpec(layer_names=layer_names, kv_cache_spec=spec)
            )
5496

5497
    def get_kv_cache_spec(self) -> dict[str, KVCacheSpec]:
5498
        """
5499
        Generates the KVCacheSpec by parsing the kv cache format from each
5500
5501
        Attention module in the static forward context.
        Returns:
5502
            KVCacheSpec: A dictionary mapping layer names to their KV cache
5503
5504
            format. Layers that do not need KV cache are not included.
        """
5505
5506
        if has_ec_transfer() and get_ec_transfer().is_producer:
            return {}
5507
        kv_cache_spec: dict[str, KVCacheSpec] = {}
5508
5509
        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
5510
        for layer_name, attn_module in attn_layers.items():
5511
5512
5513
5514
5515
5516
5517
5518
5519
5520
5521
5522
5523
5524
5525
            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
5526

5527
        return kv_cache_spec
5528

5529
    def _to_list(self, sampled_token_ids: torch.Tensor) -> list[list[int]]:
5530
5531
5532
5533
5534
5535
5536
5537
        # 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.
5538
        pinned = self.sampled_token_ids_pinned_cpu[: sampled_token_ids.shape[0]]
5539
5540
5541
        pinned.copy_(sampled_token_ids, non_blocking=True)
        self.transfer_event.record()
        self.transfer_event.synchronize()
5542
        return pinned.tolist()