gpu_model_runner.py 262 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 threading
8
import time
9
from collections import defaultdict
10
from collections.abc import Iterator, Sequence
11
from contextlib import contextmanager
12
from copy import copy, deepcopy
13
from dataclasses import dataclass
14
from functools import reduce
15
from typing import TYPE_CHECKING, Any, NamedTuple, TypeAlias, cast
16
17
18
19
20

import numpy as np
import torch
import torch.distributed
import torch.nn as nn
21
from tqdm import tqdm
22

23
import vllm.envs as envs
24
from vllm.compilation.counter import compilation_counter
25
from vllm.compilation.cuda_graph import CUDAGraphStat, CUDAGraphWrapper
26
from vllm.compilation.monitor import set_cudagraph_capturing_enabled
27
from vllm.config import (
28
    CompilationMode,
29
30
31
32
33
    CUDAGraphMode,
    VllmConfig,
    get_layers_from_vllm_config,
    update_config,
)
34
from vllm.distributed.ec_transfer import get_ec_transfer, has_ec_transfer
35
from vllm.distributed.eplb.eplb_state import EplbState
36
from vllm.distributed.kv_transfer import get_kv_transfer_group, has_kv_transfer_group
37
from vllm.distributed.kv_transfer.kv_connector.utils import copy_kv_blocks
38
from vllm.distributed.parallel_state import (
39
    get_dcp_group,
40
41
42
43
44
45
    get_pp_group,
    get_tp_group,
    graph_capture,
    is_global_first_rank,
    prepare_communication_buffer_for_model,
)
46
47
48
49
from vllm.forward_context import (
    BatchDescriptor,
    set_forward_context,
)
50
from vllm.logger import init_logger
51
from vllm.lora.layers import LoRAMapping, LoRAMappingType
52
from vllm.model_executor.layers.attention import Attention, MLAAttention
53
from vllm.model_executor.layers.attention_layer_base import AttentionLayerBase
54
55
56
from vllm.model_executor.layers.fused_moe.routed_experts_capturer import (
    RoutedExpertsCapturer,
)
57
58
59
60
from vllm.model_executor.layers.rotary_embedding import (
    MRotaryEmbedding,
    XDRotaryEmbedding,
)
61
from vllm.model_executor.model_loader import TensorizerLoader, get_model_loader
62
from vllm.model_executor.models.interfaces import (
63
    MultiModalEmbeddings,
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
from vllm.utils.mem_utils import DeviceMemoryProfiler, format_gib
94
from vllm.utils.nvtx_pytorch_hooks import PytHooks
95
from vllm.utils.platform_utils import is_pin_memory_available
96
97
98
99
from vllm.utils.torch_utils import (
    get_dtype_size,
    kv_cache_dtype_str_to_dtype,
)
100
101
from vllm.v1.attention.backend import (
    AttentionBackend,
102
    AttentionCGSupport,
103
    AttentionMetadata,
104
    AttentionMetadataBuilder,
105
    AttentionType,
106
    CommonAttentionMetadata,
107
108
    MultipleOf,
)
109
from vllm.v1.attention.backends.gdn_attn import GDNAttentionMetadataBuilder
110
from vllm.v1.attention.backends.utils import (
111
    create_fast_prefill_custom_backend,
112
    get_dcp_local_seq_lens,
113
114
    reorder_batch_to_split_decodes_and_prefills,
)
115
from vllm.v1.core.sched.output import NewRequestData
116
from vllm.v1.cudagraph_dispatcher import CudagraphDispatcher
117
118
119
120
121
122
123
124
125
126
127
128
129
130
131
132
133
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,
134
    ECConnectorOutput,
135
    KVConnectorOutput,
136
137
138
139
140
    LogprobsLists,
    LogprobsTensors,
    ModelRunnerOutput,
    PoolerOutput,
    SamplerOutput,
141
    make_empty_encoder_model_runner_output,
142
)
143
from vllm.v1.pool.metadata import PoolingMetadata, PoolingStates
144
from vllm.v1.sample.logits_processor import LogitsProcessors, build_logitsprocs
145
from vllm.v1.sample.logits_processor.interface import LogitsProcessor
146
from vllm.v1.sample.metadata import SamplingMetadata
147
from vllm.v1.sample.rejection_sampler import RejectionSampler
148
from vllm.v1.sample.sampler import Sampler
149
from vllm.v1.spec_decode.draft_model import DraftModelProposer
150
from vllm.v1.spec_decode.eagle import EagleProposer
151
from vllm.v1.spec_decode.medusa import MedusaProposer
152
from vllm.v1.spec_decode.metadata import SpecDecodeMetadata
153
from vllm.v1.spec_decode.suffix_decoding import SuffixDecodingProposer
154
from vllm.v1.structured_output.utils import apply_grammar_bitmask
155
from vllm.v1.utils import CpuGpuBuffer, record_function_or_nullcontext
156
157
158
159
160
from vllm.v1.worker import mamba_utils
from vllm.v1.worker.cp_utils import (
    check_attention_cp_compatibility,
    get_total_cp_world_size,
)
161
from vllm.v1.worker.dp_utils import coordinate_batch_across_dp
162
from vllm.v1.worker.ec_connector_model_runner_mixin import ECConnectorModelRunnerMixin
163
from vllm.v1.worker.gpu_input_batch import CachedRequestState, InputBatch
164
from vllm.v1.worker.gpu_ubatch_wrapper import UBatchWrapper
165
from vllm.v1.worker.kv_connector_model_runner_mixin import KVConnectorModelRunnerMixin
166
from vllm.v1.worker.lora_model_runner_mixin import LoRAModelRunnerMixin
167
168
169
from vllm.v1.worker.ubatch_utils import (
    UBatchSlices,
    check_ubatch_thresholds,
170
    maybe_create_ubatch_slices,
171
    split_attn_metadata,
172
)
173
from vllm.v1.worker.utils import is_residual_scattered_for_sp
174
from vllm.v1.worker.workspace import lock_workspace
175

176
177
178
179
180
181
182
from .utils import (
    AttentionGroup,
    MultiModalBudget,
    add_kv_sharing_layers_to_kv_cache_groups,
    bind_kv_cache,
    sanity_check_mm_encoder_outputs,
)
183

184
if TYPE_CHECKING:
185
    from vllm.model_executor.model_loader.tensorizer import TensorizerConfig
186
    from vllm.v1.core.sched.output import GrammarOutput, SchedulerOutput
187
    from vllm.v1.spec_decode.ngram_proposer import NgramProposer
188
189
190

logger = init_logger(__name__)

191
192
AttnMetadataDict: TypeAlias = dict[str, AttentionMetadata]
# list when ubatching is enabled
193
PerLayerAttnMetadata: TypeAlias = list[AttnMetadataDict] | AttnMetadataDict
194

195

196
197
198
199
200
201
# Wrapper for ModelRunnerOutput to support overlapped execution.
class AsyncGPUModelRunnerOutput(AsyncModelRunnerOutput):
    def __init__(
        self,
        model_runner_output: ModelRunnerOutput,
        sampled_token_ids: torch.Tensor,
202
        logprobs_tensors: LogprobsTensors | None,
203
204
        invalid_req_indices: list[int],
        async_output_copy_stream: torch.cuda.Stream,
205
        vocab_size: int,
206
207
208
209
210
    ):
        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.
211
        self.async_copy_ready_event = torch.Event()
212
213
214
215

        # 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
216
        self.vocab_size = vocab_size
217
        self._logprobs_tensors = logprobs_tensors
218
219
220
221
222

        # 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)
223
            self.sampled_token_ids_cpu = self._sampled_token_ids.to(
224
225
                "cpu", non_blocking=True
            )
226
227
228
229
230
            self._logprobs_tensors_cpu = (
                self._logprobs_tensors.to_cpu_nonblocking()
                if self._logprobs_tensors
                else None
            )
231
            self.async_copy_ready_event.record()
232
233
234

    def get_output(self) -> ModelRunnerOutput:
        """Copy the device tensors to the host and return a ModelRunnerOutput.
235

236
237
        This function blocks until the copy is finished.
        """
238
        max_gen_len = self.sampled_token_ids_cpu.shape[-1]
239
        self.async_copy_ready_event.synchronize()
240

241
242
        # Release the device tensors once the copy has completed.
        del self._logprobs_tensors
243
        del self._sampled_token_ids
244
        if max_gen_len == 1:
245
            valid_sampled_token_ids = self.sampled_token_ids_cpu.tolist()
246
247
            for i in self._invalid_req_indices:
                valid_sampled_token_ids[i].clear()
248
249
250
            logprobs_lists = None
            if self._logprobs_tensors_cpu is not None:
                logprobs_lists = self._logprobs_tensors_cpu.tolists()
251
        else:
252
            valid_sampled_token_ids, logprobs_lists = RejectionSampler.parse_output(
253
254
                self.sampled_token_ids_cpu,
                self.vocab_size,
255
                self._invalid_req_indices,
256
                logprobs_tensors=self._logprobs_tensors_cpu,
257
            )
258
259
260

        output = self._model_runner_output
        output.sampled_token_ids = valid_sampled_token_ids
261
        output.logprobs = logprobs_lists
262
263
264
        return output


265
266
267
268
269
270
271
272
273
274
275
276
277
278
279
280
281
282
283
284
285
class AsyncGPUPoolingModelRunnerOutput(AsyncModelRunnerOutput):
    def __init__(
        self,
        model_runner_output: ModelRunnerOutput,
        raw_pooler_output: PoolerOutput,
        finished_mask: list[bool],
        async_output_copy_stream: torch.cuda.Stream,
    ):
        self._model_runner_output = model_runner_output

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

        # Keep a reference to the device tensors to avoid them being
        # deallocated until we finish copying it to the host.
        self._raw_pooler_output = raw_pooler_output

        # 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)
286
            raw_pooler_output_cpu = json_map_leaves(
287
288
289
290
                lambda x: None if x is None else x.to("cpu", non_blocking=True),
                self._raw_pooler_output,
            )
            self.async_copy_ready_event.record()
291
292
293
294
            self._model_runner_output.pooler_output = [
                out if include else None
                for out, include in zip(raw_pooler_output_cpu, finished_mask)
            ]
295
296
297
298
299
300
301
302
303
304
305
306

    def get_output(self) -> ModelRunnerOutput:
        """Copy the device tensors to the host and return a ModelRunnerOutput.
        This function blocks until the copy is finished.
        """
        self.async_copy_ready_event.synchronize()

        # Release the device tensors once the copy has completed.
        del self._raw_pooler_output
        return self._model_runner_output


307
308
309
310
311
312
313
314
315
316
317
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
318
    ec_connector_output: ECConnectorOutput | None
319
    cudagraph_stats: CUDAGraphStat | None
320
    slot_mappings: dict[str, torch.Tensor] | list[dict[str, torch.Tensor]] | None
321
322


323
324
325
class GPUModelRunner(
    LoRAModelRunnerMixin, KVConnectorModelRunnerMixin, ECConnectorModelRunnerMixin
):
326
327
    def __init__(
        self,
328
        vllm_config: VllmConfig,
329
        device: torch.device,
330
    ):
331
332
333
        self.vllm_config = vllm_config
        self.model_config = vllm_config.model_config
        self.cache_config = vllm_config.cache_config
334
        self.compilation_config = vllm_config.compilation_config
335
336
337
338
339
340
        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
341

342
        from vllm.model_executor.models.utils import set_cpu_offload_max_bytes
343
344

        set_cpu_offload_max_bytes(int(self.cache_config.cpu_offload_gb * 1024**3))
345

346
347
348
349
        model_config = self.model_config
        cache_config = self.cache_config
        scheduler_config = self.scheduler_config
        parallel_config = self.parallel_config
350
        self.device = device
351
352
        self.pin_memory = is_pin_memory_available()
        self.dtype = self.model_config.dtype
353

354
355
356
        self.kv_cache_dtype = kv_cache_dtype_str_to_dtype(
            cache_config.cache_dtype, self.model_config
        )
357

358
        self.is_pooling_model = model_config.runner_type == "pooling"
359
        self.enable_prompt_embeds = model_config.enable_prompt_embeds
360
        self.is_multimodal_raw_input_only_model = (
361
362
            model_config.is_multimodal_raw_input_only_model
        )
363
364
        # This will be overridden in load_model()
        self.is_multimodal_pruning_enabled = False
365
        self.max_model_len = model_config.max_model_len
366
367
368

        # Always set to false after the first forward pass
        self.calculate_kv_scales = self.cache_config.calculate_kv_scales
369
        self.dcp_world_size = self.parallel_config.decode_context_parallel_size
370
        self.dcp_rank = 0 if self.dcp_world_size <= 1 else get_dcp_group().rank_in_group
371
        self.max_num_tokens = scheduler_config.max_num_batched_tokens
372
        self.max_num_reqs = scheduler_config.max_num_seqs
373

374
375
376
377
378
        # 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 = (
379
            self.parallel_config.distributed_executor_backend == "external_launcher"
380
            and len(get_pp_group().ranks) > 1
381
        )
382

383
        # Model-related.
384
        self.num_query_heads = model_config.get_num_attention_heads(parallel_config)
385
        self.inputs_embeds_size = model_config.get_inputs_embeds_size()
386
        self.attention_chunk_size = model_config.attention_chunk_size
387
        # Only relevant for models using ALiBi (e.g, MPT)
388
        self.use_alibi = model_config.uses_alibi
389

390
        self.cascade_attn_enabled = not self.model_config.disable_cascade_attn
391
        self.is_mm_prefix_lm = self.model_config.is_mm_prefix_lm
392

393
        # Multi-modal data support
394
        self.mm_registry = MULTIMODAL_REGISTRY
395
        self.uses_mrope = model_config.uses_mrope
396
        self.uses_xdrope_dim = model_config.uses_xdrope_dim
397
        self.supports_mm_inputs = self.mm_registry.supports_multimodal_inputs(
398
            model_config
399
        )
400

401
402
403
        if self.model_config.is_encoder_decoder:
            # Maximum length of the encoder input, only for encoder-decoder
            # models.
404
            self.max_encoder_len = scheduler_config.max_num_encoder_input_tokens
405
406
407
        else:
            self.max_encoder_len = 0

408
409
410
        # Async scheduling
        self.use_async_scheduling = self.scheduler_config.async_scheduling

411
        # Sampler
412
        self.sampler = Sampler(logprobs_mode=self.model_config.logprobs_mode)
413

414
        self.eplb_state: EplbState | None = None
415
416
417
418
419
420
        """
        State of the expert parallelism load balancer.

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

421
        # Lazy initializations
422
        # self.model: nn.Module  # Set after load_model
423
        # Initialize in initialize_kv_cache
424
        self.kv_caches: list[torch.Tensor] = []
425
426
427
        # Initialize in initialize_kv_cache_tensors
        self.cross_layers_kv_cache: torch.Tensor | None = None
        self.cross_layers_attn_backend: type[AttentionBackend] | None = None
428
429
        # indexes: [kv_cache_group_id][attn_group]
        self.attn_groups: list[list[AttentionGroup]] = []
430
431
        # self.kv_cache_config: KVCacheConfig

432
433
        # mm_hash ->  encoder_output
        self.encoder_cache: dict[str, torch.Tensor] = {}
434

435
        self.use_aux_hidden_state_outputs = False
436
437
438
439
440
        # 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:
441
            self.drafter: (
442
                NgramProposer  # noqa: F823
443
444
445
446
                | SuffixDecodingProposer
                | EagleProposer
                | DraftModelProposer
                | MedusaProposer
447
            )
448
            if self.speculative_config.method == "ngram":
449
450
                from vllm.v1.spec_decode.ngram_proposer import NgramProposer

451
                self.drafter = NgramProposer(self.vllm_config)
452
453
454
455
456
457
            elif self.speculative_config.uses_draft_model():
                self.drafter = DraftModelProposer(
                    vllm_config=self.vllm_config,
                    device=self.device,
                    runner=self,
                )
458
459
            elif self.speculative_config.method == "suffix":
                self.drafter = SuffixDecodingProposer(self.vllm_config)
460
            elif self.speculative_config.use_eagle():
461
                self.drafter = EagleProposer(self.vllm_config, self.device, self)
462
                if self.speculative_config.method == "eagle3":
463
464
465
                    self.use_aux_hidden_state_outputs = (
                        self.drafter.eagle3_use_aux_hidden_state
                    )
466
467
            elif self.speculative_config.method == "medusa":
                self.drafter = MedusaProposer(
468
                    vllm_config=self.vllm_config, device=self.device
469
                )
470
            else:
471
472
473
474
                raise ValueError(
                    "Unknown speculative decoding method: "
                    f"{self.speculative_config.method}"
                )
475
            self.rejection_sampler = RejectionSampler(self.sampler)
476

477
478
479
        self.num_spec_tokens = 0
        if self.speculative_config:
            self.num_spec_tokens = self.speculative_config.num_speculative_tokens
480
481
482
483
484
            draft_config = self.speculative_config.draft_model_config
            if draft_config is not None and draft_config.max_model_len is not None:
                self.effective_drafter_max_model_len = draft_config.max_model_len
            else:
                self.effective_drafter_max_model_len = self.max_model_len
485

486
        # Request states.
487
        self.requests: dict[str, CachedRequestState] = {}
488
489
490
        # NOTE(rob): num_prompt_logprobs only includes reqs
        # that are currently in the prefill phase.
        self.num_prompt_logprobs: dict[str, int] = {}
491
        self.comm_stream = torch.cuda.Stream()
492

493
494
495
496
497
498
499
500
501
        # 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.
502
503
504
505
        logits_processors = model_config.logits_processors
        custom_logitsprocs: Sequence[str | type[LogitsProcessor]] = (
            tuple(logits_processors) if logits_processors is not None else ()
        )
506
507
        self.input_batch = InputBatch(
            max_num_reqs=self.max_num_reqs,
508
509
510
            # 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),
511
512
513
            max_num_batched_tokens=self.max_num_tokens,
            device=self.device,
            pin_memory=self.pin_memory,
514
            vocab_size=self.model_config.get_vocab_size(),
515
            block_sizes=[self.cache_config.block_size],
516
            kernel_block_sizes=[self.cache_config.block_size],
517
            is_spec_decode=bool(self.vllm_config.speculative_config),
518
            logitsprocs=build_logitsprocs(
519
520
521
                self.vllm_config,
                self.device,
                self.pin_memory,
522
                self.is_pooling_model,
523
                custom_logitsprocs,
524
            ),
525
526
527
            # 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),
528
            is_pooling_model=self.is_pooling_model,
529
            cp_kv_cache_interleave_size=self.parallel_config.cp_kv_cache_interleave_size,
530
        )
531

532
533
534
535
536
        # 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.
537
        self.prepare_inputs_event: torch.Event | None = None
538
539
        if self.use_async_scheduling:
            self.async_output_copy_stream = torch.cuda.Stream()
540
            self.prepare_inputs_event = torch.Event()
541

542
        # self.cudagraph_batch_sizes sorts in ascending order.
543
544
545
546
        if (
            self.compilation_config.cudagraph_capture_sizes
            and self.compilation_config.cudagraph_mode != CUDAGraphMode.NONE
        ):
547
548
            self.cudagraph_batch_sizes = sorted(
                self.compilation_config.cudagraph_capture_sizes
549
            )
550

551
        # Cache the device properties.
552
        self._init_device_properties()
553

554
555
556
557
        # Encoder timing registry for observability
        self.encoder_timing_registry: dict[str, EncoderTimingStats] = {}
        self._encoder_timing_lock = threading.Lock()

558
        # Persistent buffers for CUDA graphs.
559
560
561
562
563
        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
        )
564
        self.seq_lens = self._make_buffer(self.max_num_reqs, dtype=torch.int32)
565
        self.encoder_seq_lens = self._make_buffer(self.max_num_reqs, dtype=torch.int32)
566
567
568
569
        if self.dcp_world_size > 1:
            self.dcp_local_seq_lens = self._make_buffer(
                self.max_num_reqs, dtype=torch.int32
            )
570
571
572
        # 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.
573
        self.inputs_embeds = self._make_buffer(
574
            self.max_num_tokens, self.inputs_embeds_size, dtype=self.dtype, numpy=False
575
576
        )
        self.is_token_ids = self._make_buffer(self.max_num_tokens, dtype=torch.bool)
577
578
        self.discard_request_mask = self._make_buffer(
            self.max_num_reqs, dtype=torch.bool
579
580
581
582
583
584
585
        )
        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
        )
586

587
588
        # Only relevant for multimodal models
        if self.supports_mm_inputs:
589
590
591
592
593
594
595
            # Double buffer to avoid race condition: previous iteration's async
            # copy may still be reading from CPU while current iteration writes.
            self.is_mm_embed_buffers = [
                self._make_buffer(self.max_num_tokens, dtype=torch.bool),
                self._make_buffer(self.max_num_tokens, dtype=torch.bool),
            ]
            self.is_mm_embed_idx = 0
596

597
        # Only relevant for models using M-RoPE (e.g, Qwen2-VL)
598
        if self.uses_mrope:
Roger Wang's avatar
Roger Wang committed
599
600
601
602
            # 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
603
604
605
606
607
608

            # 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
609
            self.mrope_positions = self._make_buffer(
610
611
                (3, self.max_num_tokens + 1), dtype=torch.int64
            )
612

613
614
615
616
617
618
619
        # 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
            )

620
        # None in the first PP rank. The rest are set after load_model.
621
        self.intermediate_tensors: IntermediateTensors | None = None
622

623
        # OPTIMIZATION: Cache the tensors rather than creating them every step.
624
        # Keep in int64 to avoid overflow with long context
625
626
627
628
        self.arange_np = np.arange(
            max(self.max_num_reqs + 1, self.max_model_len, self.max_num_tokens),
            dtype=np.int64,
        )
629

630
631
632
633
634
        # 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] = {}
635
636
637
638
639
        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(
640
641
                self.max_num_tokens, dtype=torch.int32, device=self.device
            )
642

643
        self.uniform_decode_query_len = 1 + self.num_spec_tokens
644
645
646
647

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

648
        self.mm_budget = (
649
            MultiModalBudget(self.vllm_config, self.mm_registry)
650
651
652
            if self.supports_mm_inputs
            else None
        )
653

654
        self.reorder_batch_threshold: int | None = None
655

656
657
658
659
660
        # 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()

661
        # Cached outputs.
662
        self._draft_token_ids: list[list[int]] | torch.Tensor | None = None
663
        self._draft_token_req_ids: list[str] | None = None
664
        self.transfer_event = torch.Event()
665
        self.sampled_token_ids_pinned_cpu = torch.empty(
666
            (self.max_num_reqs, 1),
667
668
            dtype=torch.int64,
            device="cpu",
669
670
            pin_memory=self.pin_memory,
        )
671

672
673
        # Pre-allocated tensor for copying valid sampled token counts to CPU,
        # with dedicated stream for overlapping and event for coordination.
674
        self.valid_sampled_token_count_event: torch.Event | None = None
675
        self.valid_sampled_token_count_copy_stream: torch.cuda.Stream | None = None
676
677
678
679
680
681
682
683
684
685
686
687
688
689
690
691
692
693
694
695
696
697
698
699
        # We also copy the drafted tokens to the CPU asynchronously,
        # in case we need them for structured outputs.
        self.draft_token_ids_event: torch.Event | None = None
        self.draft_token_ids_copy_stream: torch.cuda.Stream | None = None
        self.valid_sampled_token_count_cpu: torch.Tensor | None = None
        self.draft_token_ids_cpu: torch.Tensor | None = None
        if self.num_spec_tokens:
            self.draft_token_ids_event = torch.Event()
            self.draft_token_ids_copy_stream = torch.cuda.Stream()
            self.draft_token_ids_cpu = torch.empty(
                (self.max_num_reqs, self.num_spec_tokens),
                dtype=torch.int64,
                device="cpu",
                pin_memory=self.pin_memory,
            )
            if self.use_async_scheduling:
                self.valid_sampled_token_count_event = torch.Event()
                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,
                )
700

701
702
        # Ephemeral state transferred between execute_model() and sample_tokens().
        self.execute_model_state: ExecuteModelState | None = None
703
        self.kv_connector_output: KVConnectorOutput | None = None
704
        self.mamba_state_idx: dict[str, int] = {}
705
        self.layerwise_nvtx_hooks_registered = False
706

707
708
709
710
711
712
713
    def update_max_model_len(self, max_model_len: int) -> None:
        self.max_model_len = max_model_len
        if self.speculative_config:
            draft_config = self.speculative_config.draft_model_config
            if draft_config is None or draft_config.max_model_len is None:
                self.effective_drafter_max_model_len = self.max_model_len

714
715
716
717
    def reset_mm_cache(self) -> None:
        if self.mm_budget:
            self.mm_budget.reset_cache()

718
719
720
721
722
723
724
725
726
727
728
729
730
731
732
733
734
735
736
737
738
739
740
741
742
743
744
745
746
747
748
749
750
751
752
753
754
755
756
757
758
759
760
761
    @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)

762
763
764
765
    def _get_positions(self, num_tokens: Any):
        if isinstance(num_tokens, int):
            if self.uses_mrope:
                return self.mrope_positions.gpu[:, :num_tokens]
766
767
            if self.uses_xdrope_dim > 0:
                return self.xdrope_positions.gpu[:, :num_tokens]
768
769
770
771
            return self.positions.gpu[:num_tokens]
        else:
            if self.uses_mrope:
                return self.mrope_positions.gpu[:, num_tokens]
772
773
            if self.uses_xdrope_dim > 0:
                return self.xdrope_positions.gpu[:, num_tokens]
774
775
            return self.positions.gpu[num_tokens]

776
    def _make_buffer(
777
        self, *size: int | torch.SymInt, dtype: torch.dtype, numpy: bool = True
778
779
780
781
782
783
784
785
    ) -> CpuGpuBuffer:
        return CpuGpuBuffer(
            *size,
            dtype=dtype,
            device=self.device,
            pin_memory=self.pin_memory,
            with_numpy=numpy,
        )
786

787
    def _init_model_kwargs(self):
788
789
        model_kwargs = dict[str, Any]()

790
        if not self.is_pooling_model:
791
792
            return model_kwargs

793
794
        num_reqs = self.input_batch.num_reqs
        pooling_params = self.input_batch.get_pooling_params()
795
796
797

        token_type_id_requests = dict[int, Any]()
        for i, param in enumerate(pooling_params):
798
799
800
801
802
            if (
                param.extra_kwargs is not None
                and (token_types := param.extra_kwargs.get("compressed_token_type_ids"))
                is not None
            ):
803
804
805
806
807
                token_type_id_requests[i] = token_types

        if len(token_type_id_requests) == 0:
            return model_kwargs

808
        seq_lens = self.seq_lens.gpu[:num_reqs]
809
810
811
812
813
814
815
816
        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(
817
818
            device=self.device
        )
819
820
        return model_kwargs

821
    def _may_reorder_batch(self, scheduler_output: "SchedulerOutput") -> None:
822
823
        """
        Update the order of requests in the batch based on the attention
824
        backend's needs. For example, some attention backends (namely MLA) may
825
826
827
828
829
830
        want to separate requests based on if the attention computation will be
        compute-bound or memory-bound.

        Args:
            scheduler_output: The scheduler output.
        """
831
832
833
834
835
836
837
838
        # 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

839
840
841
842
        if self.reorder_batch_threshold is not None:
            reorder_batch_to_split_decodes_and_prefills(
                self.input_batch,
                scheduler_output,
843
844
                decode_threshold=self.reorder_batch_threshold,
            )
845

846
847
    # Note: used for model runner override.
    def _init_device_properties(self) -> None:
848
        """Initialize attributes from torch.cuda.get_device_properties"""
849
850
851
852
853
854
855
        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()

856
    def _update_states(self, scheduler_output: "SchedulerOutput") -> None:
857
858
859
860
861
862
        """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.

863
864
        The SamplingMetadata is updated and copied to the GPU if there is a
        new/resumed/paused/finished request in the batch.
865
866
        """
        # Remove finished requests from the cached states.
867
868
        for req_id in scheduler_output.finished_req_ids:
            self.requests.pop(req_id, None)
869
            self.num_prompt_logprobs.pop(req_id, None)
870
871
872
873
874
875
876
        # 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:
877
            self.input_batch.remove_request(req_id)
878
879

        # Free the cached encoder outputs.
880
881
        for mm_hash in scheduler_output.free_encoder_mm_hashes:
            self.encoder_cache.pop(mm_hash, None)
882

883
884
885
886
887
888
889
        # 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()
890
891
892
893
894
895
896
897
        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)
898
899
900
901
902
        # 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:
903
            self.input_batch.remove_request(req_id)
904

905
        reqs_to_add: list[CachedRequestState] = []
906
        # Add new requests to the cached states.
907
908
        for new_req_data in scheduler_output.scheduled_new_reqs:
            req_id = new_req_data.req_id
909
910
911
912
913
914
            if req_id in self.requests:
                # For streaming case only.
                req_state = self._update_streaming_request(req_id, new_req_data)
                reqs_to_add.append(req_state)
                continue

915
            sampling_params = new_req_data.sampling_params
916
            pooling_params = new_req_data.pooling_params
917

918
919
920
921
            if (
                sampling_params
                and sampling_params.sampling_type == SamplingType.RANDOM_SEED
            ):
922
923
924
925
926
                generator = torch.Generator(device=self.device)
                generator.manual_seed(sampling_params.seed)
            else:
                generator = None

927
928
            if self.is_pooling_model:
                assert pooling_params is not None
929
930
                task = pooling_params.task
                assert task is not None, "You did not set `task` in the API"
931

932
                model = cast(VllmModelForPooling, self.get_model())
933
                to_update = model.pooler.get_pooling_updates(task)
934
935
                to_update.apply(pooling_params)

936
            req_state = CachedRequestState(
937
                req_id=req_id,
938
                prompt_token_ids=new_req_data.prompt_token_ids,
939
                prompt_embeds=new_req_data.prompt_embeds,
940
                mm_features=new_req_data.mm_features,
941
                sampling_params=sampling_params,
942
                pooling_params=pooling_params,
943
                generator=generator,
944
945
                block_ids=new_req_data.block_ids,
                num_computed_tokens=new_req_data.num_computed_tokens,
946
                output_token_ids=[],
947
                lora_request=new_req_data.lora_request,
948
            )
949
950
            self.requests[req_id] = req_state

951
952
953
954
955
956
957
            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
                )

958
            # Only relevant for models using M-RoPE (e.g, Qwen2-VL)
959
            if self.uses_mrope:
960
                self._init_mrope_positions(req_state)
961

962
963
964
965
            # Only relevant for models using XD-RoPE (e.g, HunYuan-VL)
            if self.uses_xdrope_dim > 0:
                self._init_xdrope_positions(req_state)

966
            reqs_to_add.append(req_state)
967

968
        # Update the states of the running/resumed requests.
969
        is_last_rank = get_pp_group().is_last_rank
970
        req_data = scheduler_output.scheduled_cached_reqs
971
        scheduled_spec_tokens = scheduler_output.scheduled_spec_decode_tokens
972
973
974
975
976

        # 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()

977
        for i, req_id in enumerate(req_data.req_ids):
978
            req_state = self.requests[req_id]
979
980
            num_computed_tokens = req_data.num_computed_tokens[i]
            new_block_ids = req_data.new_block_ids[i]
981
            resumed_from_preemption = req_id in req_data.resumed_req_ids
982
            num_output_tokens = req_data.num_output_tokens[i]
983
            req_index = self.input_batch.req_id_to_index.get(req_id)
984

985
986
987
988
989
990
991
992
993
994
995
996
997
998
            if req_state.prev_num_draft_len and self.use_async_scheduling:
                # 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.
999
1000
1001
1002
1003
1004
1005
1006
1007
                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)
1008

1009
            # Update the cached states.
1010
            req_state.num_computed_tokens = num_computed_tokens
1011
1012

            if not is_last_rank:
1013
1014
1015
1016
1017
1018
1019
1020
1021
1022
1023
1024
1025
1026
1027
1028
1029
1030
1031
1032
                if not req_data.new_token_ids:
                    # Async scheduled PP: Sampled tokens propagated via GPU broadcast.
                    new_token_ids: list[int] = []
                else:
                    # Non-async scheduling with PP: The scheduler sends
                    # sampled token ids back because there's no direct communication
                    # between the first-stage worker and the last-stage worker.
                    new_token_ids = req_data.new_token_ids[i]
                    # Add the sampled token(s) from the previous step (if any).
                    # This doesn't include "unverified" tokens like spec tokens.
                    num_new_tokens = (
                        num_computed_tokens + len(new_token_ids) - req_state.num_tokens
                    )
                    if num_new_tokens == 1:
                        # Avoid slicing list in most common case.
                        req_state.output_token_ids.append(new_token_ids[-1])
                    elif num_new_tokens > 0:
                        req_state.output_token_ids.extend(
                            new_token_ids[-num_new_tokens:]
                        )
1033
1034
1035
1036
1037
            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:
1038
1039
1040
1041
                    end_idx = (
                        self.input_batch.num_prompt_tokens[req_index]
                        + num_output_tokens
                    )
1042
                    self.input_batch.num_tokens_no_spec[req_index] = end_idx
1043

1044
            # Update the block IDs.
1045
            if not resumed_from_preemption:
1046
1047
                if new_block_ids is not None:
                    # Append the new blocks to the existing block IDs.
1048
                    for block_ids, new_ids in zip(req_state.block_ids, new_block_ids):
1049
                        block_ids.extend(new_ids)
1050
            else:
1051
                assert req_index is None
1052
                assert new_block_ids is not None
1053
1054
                # The request is resumed from preemption.
                # Replace the existing block IDs with the new ones.
1055
                req_state.block_ids = new_block_ids
1056
1057
1058
1059
1060

            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.
1061
1062
1063
1064
1065
1066
1067

                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:]

1068
                reqs_to_add.append(req_state)
1069
1070
1071
                continue

            # Update the persistent batch.
1072
            self.input_batch.num_computed_tokens_cpu[req_index] = num_computed_tokens
1073
            if new_block_ids is not None:
1074
                self.input_batch.block_table.append_row(new_block_ids, req_index)
1075
1076
1077
1078
1079
1080
1081

            # 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)
1082
                self.input_batch.token_ids_cpu[
1083
1084
1085
                    req_index, start_token_index:end_token_index
                ] = new_token_ids
                self.input_batch.num_tokens_no_spec[req_index] = end_token_index
1086

1087
            # Add spec_token_ids to token_ids_cpu.
1088
            self.input_batch.update_req_spec_token_ids(req_state, scheduled_spec_tokens)
1089

1090
1091
        # Add the new or resumed requests to the persistent batch.
        # The smaller empty indices are filled first.
1092
1093
        for request in reqs_to_add:
            self.input_batch.add_request(request)
1094
            self.input_batch.update_req_spec_token_ids(request, scheduled_spec_tokens)
1095

1096
1097
1098
1099
1100
1101
        # 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()
1102

1103
    def _update_states_after_model_execute(
1104
        self, output_token_ids: torch.Tensor, scheduler_output: "SchedulerOutput"
1105
    ) -> None:
1106
1107
1108
1109
1110
1111
1112
1113
        """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.
        """
1114
        if not self.speculative_config or not self.model_config.is_hybrid:
1115
1116
1117
            return

        # Find the number of accepted tokens for each sequence.
1118
1119
1120
1121
1122
1123
1124
1125
1126
1127
1128
1129
1130
1131
1132
1133
1134
1135
1136
1137
        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()
        )
1138
1139
        for i, num_tokens in enumerate(num_accepted_tokens):
            self.input_batch.num_accepted_tokens_cpu[i] = num_tokens
1140
1141
1142
1143
1144
1145
1146
1147
1148
1149
        if self.cache_config.mamba_cache_mode == "align":
            mamba_utils.postprocess_mamba(
                scheduler_output,
                self.kv_cache_config,
                self.input_batch,
                self.requests,
                self.mamba_state_idx,
                self.compilation_config.static_forward_context,
                self.model.get_mamba_state_copy_func(),
            )
1150

1151
1152
1153
1154
1155
1156
1157
1158
1159
1160
1161
1162
1163
1164
1165
1166
1167
1168
1169
1170
1171
1172
1173
1174
1175
1176
1177
1178
1179
1180
1181
1182
1183
1184
    def _update_streaming_request(
        self, req_id: str, new_req_data: NewRequestData
    ) -> CachedRequestState:
        """Updates streaming session request from `scheduled_new_reqs`.

        Removes the request from InputBatch (if present), updates the cached
        state, and prepares it for re-addition to the batch.

        NOTE: prompt_token_ids includes intermediate output tokens - tokens
        previously generated but now are input context (part of the prompt).
        """
        self.input_batch.remove_request(req_id)
        req_state = self.requests[req_id]

        req_state.prompt_token_ids = new_req_data.prompt_token_ids
        req_state.mm_features = new_req_data.mm_features
        req_state.prompt_embeds = new_req_data.prompt_embeds
        req_state.sampling_params = new_req_data.sampling_params
        req_state.pooling_params = new_req_data.pooling_params
        req_state.block_ids = new_req_data.block_ids
        req_state.num_computed_tokens = new_req_data.num_computed_tokens
        req_state.num_prompt_tokens = length_from_prompt_token_ids_or_embeds(
            req_state.prompt_token_ids, req_state.prompt_embeds
        )

        # Clear `output_token_ids` as previous output tokens are now part of
        # `prompt_token_ids`.
        req_state.output_token_ids.clear()

        if self.uses_mrope:
            self._init_mrope_positions(req_state)

        return req_state

1185
    def _init_mrope_positions(self, req_state: CachedRequestState):
1186
1187
        model = self.get_model()
        assert supports_mrope(model), "M-RoPE support is not implemented."
1188
1189
1190
1191
        assert req_state.prompt_token_ids is not None, (
            "M-RoPE requires prompt_token_ids to be available."
        )
        mrope_model = cast(SupportsMRoPE, model)
1192
1193

        req_state.mrope_positions, req_state.mrope_position_delta = (
1194
            mrope_model.get_mrope_input_positions(
1195
                req_state.prompt_token_ids,
1196
                req_state.mm_features,
1197
            )
1198
        )
1199

1200
1201
1202
1203
1204
1205
1206
1207
1208
1209
1210
1211
1212
    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,
        )

1213
    def _extract_mm_kwargs(
1214
        self,
1215
1216
        scheduler_output: "SchedulerOutput",
    ) -> BatchedTensorInputs:
1217
        if not scheduler_output or not self.is_multimodal_raw_input_only_model:
1218
            return {}
1219

1220
        mm_kwargs = list[tuple[str, MultiModalKwargsItem]]()
1221
        for req in scheduler_output.scheduled_new_reqs:
1222
1223
            for feature in req.mm_features:
                if feature.data is not None:
1224
                    mm_kwargs.append((feature.modality, feature.data))
1225

1226
1227
1228
        # Input all modalities at once
        mm_kwargs_combined: BatchedTensorInputs = {}
        for _, _, mm_kwargs_group in group_mm_kwargs_by_modality(
1229
1230
1231
            mm_kwargs,
            device=self.device,
            pin_memory=self.pin_memory,
1232
1233
        ):
            mm_kwargs_combined.update(mm_kwargs_group)
1234

1235
        return mm_kwargs_combined
1236

1237
    def _dummy_mm_kwargs(self, num_seqs: int) -> BatchedTensorInputs:
1238
        if not self.is_multimodal_raw_input_only_model:
1239
            return {}
1240

1241
1242
1243
1244
1245
        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)
1246

1247
1248
1249
    def _get_cumsum_and_arange(
        self,
        num_tokens: np.ndarray,
1250
        cumsum_dtype: np.dtype | None = None,
1251
1252
1253
1254
1255
1256
1257
1258
1259
1260
1261
1262
1263
1264
1265
1266
    ) -> 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

1267
    def _prepare_input_ids(
1268
1269
1270
1271
        self,
        scheduler_output: "SchedulerOutput",
        total_num_scheduled_tokens: int,
        cu_num_tokens: np.ndarray,
1272
    ) -> None:
1273
        """Prepare the input IDs for the current batch.
1274

1275
1276
1277
1278
1279
1280
1281
        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)
1282
1283
1284
            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)
1285
1286
1287
1288
1289
1290
1291
            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
1292
1293
1294
1295
        sample_flattened_indices: list[int] = []
        spec_flattened_indices: list[int] = []
        prev_common_req_indices: list[int] = []
        prev_draft_token_indices: list[int] = []
1296
1297
        indices_match = True
        max_flattened_index = -1
1298
1299
1300
        total_num_spec_tokens = 0
        scheduled_spec_tokens = scheduler_output.scheduled_spec_decode_tokens

1301
1302
1303
1304
1305
        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.
1306
1307
                draft_len = len(scheduled_spec_tokens.get(req_id, ()))
                total_num_spec_tokens += draft_len
1308
                flattened_index = cu_num_tokens[cur_index].item() - 1
1309
1310
1311
1312
1313
1314
1315
1316
1317
1318
1319
1320
1321
1322
1323
                # 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))
1324
                indices_match &= prev_index == flattened_index
1325
                max_flattened_index = max(max_flattened_index, flattened_index)
1326
1327
1328
        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:
1329
1330
1331
            # 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)
1332
1333
1334
            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)
1335
1336
        if num_commmon_tokens == 0:
            # No requests in common with the previous iteration
1337
            # So input_ids.cpu will have all the input ids.
1338
1339
1340
1341
1342
1343
1344
            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_(
1345
1346
1347
                self.input_batch.prev_sampled_token_ids[:num_commmon_tokens, 0],
                non_blocking=True,
            )
1348
1349
            if self.enable_prompt_embeds:
                self.is_token_ids.gpu[:num_commmon_tokens] = True
1350
            return
1351
        # Upload the index tensors asynchronously so the scatter can be non-blocking.
1352
1353
        sampled_tokens_index_tensor = torch.tensor(
            sample_flattened_indices, dtype=torch.int64, pin_memory=self.pin_memory
1354
        ).to(self.device, non_blocking=True)
1355
        prev_common_req_indices_tensor = torch.tensor(
1356
1357
            prev_common_req_indices, dtype=torch.int64, pin_memory=self.pin_memory
        ).to(self.device, non_blocking=True)
1358
1359
        self.input_ids.gpu.scatter_(
            dim=0,
1360
            index=sampled_tokens_index_tensor,
1361
            src=self.input_batch.prev_sampled_token_ids[
1362
1363
1364
                prev_common_req_indices_tensor, 0
            ],
        )
1365

1366
1367
1368
1369
1370
1371
1372
1373
1374
1375
1376
1377
1378
1379
1380
1381
1382
1383
1384
1385
1386
1387
        # 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.input_ids.gpu.scatter_(
            dim=0,
            index=draft_tokens_index_tensor,
            src=draft_token_ids.flatten()[prev_draft_token_indices_tensor],
        )

1388
1389
    def _get_encoder_seq_lens(
        self,
1390
        num_scheduled_tokens: dict[str, int],
1391
1392
        kv_cache_spec: KVCacheSpec,
        num_reqs: int,
1393
    ) -> tuple[torch.Tensor | None, np.ndarray | None]:
1394
        if not isinstance(kv_cache_spec, CrossAttentionSpec):
1395
            return None, None
1396

1397
1398
        # Zero out buffer for padding requests that are not actually scheduled (CGs)
        self.encoder_seq_lens.np[:num_reqs] = 0
1399
1400
        # Build encoder_seq_lens array mapping request indices to
        # encoder lengths for inputs scheduled in this batch
1401
        for req_id in num_scheduled_tokens:
1402
            req_index = self.input_batch.req_id_to_index[req_id]
1403
1404
1405
1406
1407
1408
1409
1410
1411
1412
1413
1414
1415
1416
1417
1418
            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]
1419

1420
        return encoder_seq_lens, encoder_seq_lens_cpu
1421

1422
    def _prepare_inputs(
1423
1424
1425
        self,
        scheduler_output: "SchedulerOutput",
        num_scheduled_tokens: np.ndarray,
1426
1427
    ) -> tuple[
        torch.Tensor,
1428
        SpecDecodeMetadata | None,
1429
    ]:
1430
1431
        """
        :return: tuple[
1432
            logits_indices, spec_decode_metadata,
1433
1434
        ]
        """
1435
1436
1437
1438
1439
1440
1441
        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.
1442
        self.input_batch.block_table.commit_block_table(num_reqs)
1443
1444
1445

        # Get request indices.
        # E.g., [2, 5, 3] -> [0, 0, 1, 1, 1, 1, 1, 2, 2, 2]
1446
        req_indices = np.repeat(self.arange_np[:num_reqs], num_scheduled_tokens)
1447

1448
1449
        # cu_num_tokens: [2, 5, 3] -> [2, 7, 10]
        # arange: [0, 1, 0, 1, 2, 3, 4, 0, 1, 2]
1450
        cu_num_tokens, arange = self._get_cumsum_and_arange(num_scheduled_tokens)
1451
1452

        # Get positions.
1453
        positions_np = self.positions.np[:total_num_scheduled_tokens]
1454
1455
1456
1457
1458
        np.add(
            self.input_batch.num_computed_tokens_cpu[req_indices],
            arange,
            out=positions_np,
        )
1459

1460
1461
        # Calculate M-RoPE positions.
        # Only relevant for models using M-RoPE (e.g, Qwen2-VL)
1462
        if self.uses_mrope:
1463
1464
            self._calc_mrope_positions(scheduler_output)

1465
1466
1467
1468
1469
        # 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)

1470
1471
1472
1473
        # 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.
1474
1475
1476
        token_indices = (
            positions_np + req_indices * self.input_batch.token_ids_cpu.shape[1]
        )
1477
        token_indices_tensor = torch.from_numpy(token_indices)
1478

1479
1480
1481
        # 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.
1482
1483
1484
1485
1486
1487
        torch.index_select(
            self.input_batch.token_ids_cpu_tensor.flatten(),
            0,
            token_indices_tensor,
            out=self.input_ids.cpu[:total_num_scheduled_tokens],
        )
1488
        if self.enable_prompt_embeds:
1489
            is_token_ids = self.input_batch.is_token_ids_tensor.flatten()
1490
1491
1492
1493
            torch.index_select(
                is_token_ids,
                0,
                token_indices_tensor,
1494
1495
                out=self.is_token_ids.cpu[:total_num_scheduled_tokens],
            )
1496
1497
1498
1499
1500
1501
1502
1503
1504
1505
1506
1507
1508
1509
1510
1511
1512
1513
1514
1515
1516
1517
1518
1519
1520
1521
1522
1523
1524
1525
1526
1527
1528

        # 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:
1529
1530
1531
                    self.inputs_embeds.cpu[
                        output_idx : output_idx + actual_num_sched
                    ].copy_(req_embeds[start_pos:actual_end])
1532
1533

                output_idx += num_sched
1534

1535
1536
        self.input_batch.block_table.compute_slot_mapping(req_indices, positions_np)
        self.input_batch.block_table.commit_slot_mapping(total_num_scheduled_tokens)
1537
1538

        # Prepare the attention metadata.
1539
        self.query_start_loc.np[0] = 0
1540
        self.query_start_loc.np[1 : num_reqs + 1] = cu_num_tokens
1541
1542
        # Note: pad query_start_loc to be non-decreasing, as kernels
        # like FlashAttention requires that
1543
        self.query_start_loc.np[num_reqs + 1 :].fill(cu_num_tokens[-1])
1544
        self.query_start_loc.copy_to_gpu()
1545
        query_start_loc = self.query_start_loc.gpu[: num_reqs + 1]
1546

1547
        self.seq_lens.np[:num_reqs] = (
1548
1549
            self.input_batch.num_computed_tokens_cpu[:num_reqs] + num_scheduled_tokens
        )
1550
        # Fill unused with 0 for full cuda graph mode.
1551
1552
        self.seq_lens.np[num_reqs:].fill(0)
        self.seq_lens.copy_to_gpu()
1553

1554
        num_tokens = [self.requests[r].num_tokens for r in self.input_batch.req_ids]
1555
1556
        num_tokens_np = np.array(num_tokens, dtype=np.int32)

1557
        # Record which requests should not be sampled,
1558
        # so that we could clear the sampled tokens before returning
1559
1560
        self.discard_request_mask.np[:num_reqs] = (
            self.seq_lens.np[:num_reqs] < num_tokens_np
1561
        )
1562
        self.discard_request_mask.copy_to_gpu(num_reqs)
1563

1564
        # Copy the tensors to the GPU.
1565
1566
1567
1568
1569
        self._prepare_input_ids(
            scheduler_output,
            total_num_scheduled_tokens,
            cu_num_tokens,
        )
1570

1571
        if self.uses_mrope:
1572
            # Only relevant for models using M-RoPE (e.g, Qwen2-VL)
1573
1574
            self.mrope_positions.gpu[:, :total_num_scheduled_tokens].copy_(
                self.mrope_positions.cpu[:, :total_num_scheduled_tokens],
1575
1576
                non_blocking=True,
            )
1577
1578
1579
1580
1581
1582
        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,
            )
1583
1584
        else:
            # Common case (1D positions)
1585
            self.positions.copy_to_gpu(total_num_scheduled_tokens)
1586

1587
        use_spec_decode = len(scheduler_output.scheduled_spec_decode_tokens) > 0
1588
1589
1590
1591
1592
1593
1594
1595
        if not use_spec_decode:
            # NOTE(woosuk): Due to chunked prefills, the batch may contain
            # partial requests. While we should not sample any token
            # from these partial requests, we do so for simplicity.
            # We will ignore the sampled tokens from the partial requests.
            # TODO: Support prompt logprobs.
            logits_indices = query_start_loc[1:] - 1
            spec_decode_metadata = None
1596
            num_sampled_tokens = np.ones(num_reqs, dtype=np.int32)
1597
1598
1599
1600
1601
        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)
1602
1603
1604
            # 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)
1605
1606
1607
1608
            for (
                req_id,
                draft_token_ids,
            ) in scheduler_output.scheduled_spec_decode_tokens.items():
1609
1610
                req_idx = self.input_batch.req_id_to_index[req_id]
                num_draft_tokens[req_idx] = len(draft_token_ids)
1611
1612
1613
1614
1615
                if (
                    self.input_batch.num_computed_tokens_cpu[req_idx]
                    >= self.input_batch.num_prompt_tokens[req_idx]
                ):
                    num_decode_draft_tokens[req_idx] = len(draft_token_ids)
1616
            spec_decode_metadata = self._calc_spec_decode_metadata(
1617
1618
                num_draft_tokens, cu_num_tokens
            )
1619
            logits_indices = spec_decode_metadata.logits_indices
1620
            num_sampled_tokens = num_draft_tokens + 1
1621
            # For DECODE only cuda graph of some attention backends (e.g., GDN).
1622
            self.num_decode_draft_tokens.np[:num_reqs] = num_decode_draft_tokens
1623
1624
            self.num_decode_draft_tokens.np[num_reqs:].fill(-1)
            self.num_decode_draft_tokens.copy_to_gpu()
1625

1626
1627
1628
1629
1630
        # Hot-Swap lora model
        if self.lora_config:
            assert (
                np.sum(num_sampled_tokens)
                <= self.vllm_config.scheduler_config.max_num_batched_tokens
1631
            )
1632
1633
1634
1635
1636
1637
1638
1639
1640
1641
1642
            self.set_active_loras(
                self.input_batch, num_scheduled_tokens, num_sampled_tokens
            )

        return (
            logits_indices,
            spec_decode_metadata,
        )

    def _build_attention_metadata(
        self,
1643
        num_tokens: int,
1644
        num_reqs: int,
1645
1646
1647
        max_query_len: int,
        num_tokens_padded: int | None = None,
        num_reqs_padded: int | None = None,
1648
1649
1650
1651
        ubatch_slices: UBatchSlices | None = None,
        logits_indices: torch.Tensor | None = None,
        use_spec_decode: bool = False,
        for_cudagraph_capture: bool = False,
1652
        num_scheduled_tokens: dict[str, int] | None = None,
1653
        cascade_attn_prefix_lens: list[list[int]] | None = None,
1654
        slot_mappings: dict[int, torch.Tensor] | None = None,
1655
1656
1657
1658
    ) -> tuple[PerLayerAttnMetadata, CommonAttentionMetadata | None]:
        """
        :return: tuple[attn_metadata, spec_decode_common_attn_metadata]
        """
1659
1660
1661
1662
        # Attention metadata is not needed for attention free models
        if len(self.kv_cache_config.kv_cache_groups) == 0:
            return {}, None

1663
1664
        num_tokens_padded = num_tokens_padded or num_tokens
        num_reqs_padded = num_reqs_padded or num_reqs
1665
        assert num_reqs_padded is not None and num_tokens_padded is not None
1666

1667
1668
1669
        attn_metadata: PerLayerAttnMetadata = {}
        if ubatch_slices is not None:
            attn_metadata = [dict() for _ in range(len(ubatch_slices))]
1670

1671
1672
1673
1674
1675
1676
1677
1678
        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()

1679
1680
        if use_spec_decode:
            self.num_accepted_tokens.np[:num_reqs] = (
1681
1682
                self.input_batch.num_accepted_tokens_cpu[:num_reqs]
            )
1683
1684
            self.num_accepted_tokens.np[num_reqs:].fill(1)
            self.num_accepted_tokens.copy_to_gpu()
1685

1686
        kv_cache_groups = self.kv_cache_config.kv_cache_groups
1687

1688
        def _get_block_table(kv_cache_gid: int):
1689
1690
1691
            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):
1692
                blk_table_tensor = torch.zeros(
1693
                    (num_reqs_padded, 1),
1694
                    dtype=torch.int32,
1695
1696
                    device=self.device,
                )
1697
            else:
1698
                blk_table = self.input_batch.block_table[kv_cache_gid]
1699
                blk_table_tensor = blk_table.get_device_tensor(num_reqs_padded)
1700

1701
1702
1703
            # Fill unused with -1. Needed for reshape_and_cache in full cuda
            # graph mode. `blk_table_tensor` -1 to match mamba PAD_SLOT_ID
            blk_table_tensor[num_reqs:num_reqs_padded].fill_(-1)
1704
            return blk_table_tensor
1705

1706
1707
1708
        assert slot_mappings is not None
        block_table_gid_0 = _get_block_table(0)
        slot_mapping_gid_0 = slot_mappings[0]
1709

1710
1711
        if self.model_config.enable_return_routed_experts:
            self.slot_mapping = slot_mapping_gid_0[:num_tokens].cpu().numpy()
1712
1713
1714
1715
1716
1717
1718
1719
1720
1721
1722
1723
1724
1725
1726
1727
1728
1729
1730
1731
1732
1733
1734
1735
1736
1737
1738
1739
1740
1741
1742
1743
1744
1745
1746
1747
1748
1749
        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
            )

1750
1751
1752
1753
1754
1755
1756
1757
1758
        # 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
        ] = {}

1759
1760
1761
1762
1763
1764
1765
        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]
1766
            builder = attn_group.get_metadata_builder(ubid or 0)
1767
1768
1769
1770
            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))
1771

1772
1773
1774
1775
1776
1777
1778
1779
1780
1781
1782
1783
1784
1785
1786
1787
1788
1789
1790
1791
            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
                )
1792
1793
1794
1795
1796
1797
1798
1799
1800
            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,
                )
1801
1802
1803
1804
1805
1806
            else:
                attn_metadata_i = builder.build(
                    common_prefix_len=cascade_attn_prefix_len,
                    common_attn_metadata=common_attn_metadata,
                    **extra_attn_metadata_args,
                )
1807
1808
                if builder.supports_update_block_table:
                    cached_attn_metadata[cache_key] = attn_metadata_i
1809
1810
1811
1812
1813
1814
1815
1816
1817
1818
1819
1820
1821
1822
1823
1824
1825
1826
1827
1828
1829
1830
1831

            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,
1832
            )
1833
            if kv_cache_gid > 0:
1834
1835
                cm.block_table_tensor = _get_block_table(kv_cache_gid)
                cm.slot_mapping = slot_mappings[kv_cache_gid]
1836

1837
            if self.speculative_config and spec_decode_common_attn_metadata is None:
1838
                if isinstance(self.drafter, EagleProposer):
1839
                    if self.drafter.attn_layer_names[0] in kv_cache_group.layer_names:
1840
                        spec_decode_common_attn_metadata = cm
1841
                else:
1842
                    spec_decode_common_attn_metadata = cm
1843

1844
            for attn_gid in range(len(self.attn_groups[kv_cache_gid])):
1845
                if ubatch_slices is not None:
1846
1847
1848
                    for ubid, _cm in enumerate(split_attn_metadata(ubatch_slices, cm)):
                        _build_attn_group_metadata(kv_cache_gid, attn_gid, _cm, ubid)

1849
                else:
1850
                    _build_attn_group_metadata(kv_cache_gid, attn_gid, cm)
1851

1852
1853
1854
1855
1856
1857
1858
1859
1860
1861
1862
1863
1864
1865
1866
1867
1868
1869
1870
1871
        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]

1872
1873
1874
1875
1876
1877
1878
1879
1880
1881
        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)
            )

1882
        return attn_metadata, spec_decode_common_attn_metadata
1883

1884
1885
1886
    def _compute_cascade_attn_prefix_lens(
        self,
        num_scheduled_tokens: np.ndarray,
1887
        num_computed_tokens: np.ndarray,
1888
1889
1890
1891
1892
1893
1894
        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
        """
1895

1896
1897
1898
1899
1900
1901
1902
1903
1904
1905
1906
1907
1908
1909
        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,
1910
                        num_computed_tokens,
1911
1912
1913
1914
1915
1916
1917
1918
                        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
1919

1920
1921
1922
    def _compute_cascade_attn_prefix_len(
        self,
        num_scheduled_tokens: np.ndarray,
1923
        num_computed_tokens: np.ndarray,
1924
        num_common_prefix_blocks: int,
1925
1926
        kv_cache_spec: KVCacheSpec,
        attn_metadata_builder: AttentionMetadataBuilder,
1927
1928
1929
1930
1931
1932
1933
1934
1935
1936
1937
1938
1939
1940
1941
1942
1943
1944
    ) -> 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.
        """
1945

1946
        common_prefix_len = num_common_prefix_blocks * kv_cache_spec.block_size
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
        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]
1984
        # Request 3's num_computed_tokens: 3 (i.e., [A, B, C])
1985
1986
1987
1988
1989
        # 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.
1990
        common_prefix_len = min(common_prefix_len, num_computed_tokens.min())
1991
        # common_prefix_len should be a multiple of the block size.
1992
1993
1994
1995
1996
1997
1998
1999
2000
2001
2002
        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
        )
2003
2004
        assert isinstance(kv_cache_spec, AttentionSpec)
        use_cascade = attn_metadata_builder.use_cascade_attention(
2005
2006
2007
            common_prefix_len=common_prefix_len,
            query_lens=num_scheduled_tokens,
            num_query_heads=self.num_query_heads,
2008
            num_kv_heads=kv_cache_spec.num_kv_heads,
2009
            use_alibi=self.use_alibi,
2010
            use_sliding_window=use_sliding_window,
2011
            use_local_attention=use_local_attention,
2012
            num_sms=self.num_sms,
2013
            dcp_world_size=self.dcp_world_size,
2014
2015
2016
        )
        return common_prefix_len if use_cascade else 0

2017
2018
    def _calc_mrope_positions(self, scheduler_output: "SchedulerOutput"):
        mrope_pos_ptr = 0
2019
        for index, req_id in enumerate(self.input_batch.req_ids):
2020
2021
2022
            req = self.requests[req_id]
            assert req.mrope_positions is not None

2023
2024
            num_computed_tokens = self.input_batch.num_computed_tokens_cpu[index]
            num_scheduled_tokens = scheduler_output.num_scheduled_tokens[req_id]
2025
            num_prompt_tokens = length_from_prompt_token_ids_or_embeds(
2026
2027
                req.prompt_token_ids, req.prompt_embeds
            )
2028
2029

            if num_computed_tokens + num_scheduled_tokens > num_prompt_tokens:
2030
2031
                prompt_part_len = max(0, num_prompt_tokens - num_computed_tokens)
                completion_part_len = max(0, num_scheduled_tokens - prompt_part_len)
2032
2033
2034
2035
2036
2037
2038
2039
2040
2041
2042
2043
2044
            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

2045
2046
2047
                self.mrope_positions.cpu[:, dst_start:dst_end] = req.mrope_positions[
                    :, src_start:src_end
                ]
2048
2049
2050
2051
2052
2053
2054
                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

2055
                assert req.mrope_position_delta is not None
2056
                MRotaryEmbedding.get_next_input_positions_tensor(
2057
                    out=self.mrope_positions.np,
2058
2059
2060
2061
2062
                    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,
                )
2063
2064
2065

                mrope_pos_ptr += completion_part_len

2066
2067
2068
2069
2070
2071
2072
2073
2074
2075
2076
2077
2078
2079
2080
2081
2082
2083
2084
2085
2086
2087
2088
2089
2090
2091
2092
2093
2094
2095
2096
2097
2098
2099
2100
2101
2102
2103
2104
2105
2106
2107
2108
2109
2110
2111
2112
    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

2113
2114
    def _calc_spec_decode_metadata(
        self,
2115
2116
2117
2118
2119
2120
2121
2122
2123
2124
2125
2126
2127
2128
2129
2130
        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
2131
2132
2133
2134

        # 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(
2135
2136
            num_sampled_tokens, cumsum_dtype=np.int32
        )
2137
        # Step 2. [0, 0, 0, 0, 103, 104, 104, 104, 206, 207, 207]
2138
        logits_indices = np.repeat(
2139
2140
            cu_num_scheduled_tokens - num_sampled_tokens, num_sampled_tokens
        )
2141
        # Step 3. [0, 1, 2, 3, 103, 104, 105, 106, 206, 207, 208]
2142
2143
2144
2145
2146
2147
        logits_indices += arange

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

        # Compute the draft logits indices.
2148
2149
2150
        # 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(
2151
2152
            num_draft_tokens, cumsum_dtype=np.int32
        )
2153
2154
        # [0, 0, 0, 5, 5, 9]
        target_logits_indices = np.repeat(
2155
2156
            cu_num_sampled_tokens - num_sampled_tokens, num_draft_tokens
        )
2157
2158
2159
2160
2161
        # [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(
2162
2163
            self.device, non_blocking=True
        )
2164
2165
2166
        cu_num_sampled_tokens = torch.from_numpy(cu_num_sampled_tokens).to(
            self.device, non_blocking=True
        )
2167
2168
2169
        logits_indices = torch.from_numpy(logits_indices).to(
            self.device, non_blocking=True
        )
2170
        target_logits_indices = torch.from_numpy(target_logits_indices).to(
2171
2172
            self.device, non_blocking=True
        )
2173
        bonus_logits_indices = torch.from_numpy(bonus_logits_indices).to(
2174
2175
            self.device, non_blocking=True
        )
2176

2177
2178
        # Compute the draft token ids.
        # draft_token_indices:      [  1,   2,   3, 105, 106, 208]
2179
        draft_token_ids = self.input_ids.gpu[logits_indices]
2180
2181
        draft_token_ids = draft_token_ids[target_logits_indices + 1]

2182
        return SpecDecodeMetadata(
2183
2184
2185
            draft_token_ids=draft_token_ids,
            num_draft_tokens=num_draft_tokens.tolist(),
            cu_num_draft_tokens=cu_num_draft_tokens,
2186
            cu_num_sampled_tokens=cu_num_sampled_tokens,
2187
2188
2189
2190
2191
            target_logits_indices=target_logits_indices,
            bonus_logits_indices=bonus_logits_indices,
            logits_indices=logits_indices,
        )

2192
2193
2194
2195
2196
2197
2198
    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
2199
        self.kv_sharing_fast_prefill_logits_indices[:num_logits].copy_(logits_indices)
2200
2201
2202
2203
2204
        # 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_(
2205
2206
            logits_indices[-1].item()
        )
2207
2208
2209
2210
2211
        # Dispatch for the decoder portion of the model.
        _, batch_desc = self.cudagraph_dispatcher.dispatch(
            num_logits, disable_full=True
        )
        num_logits_padded = batch_desc.num_tokens
2212
2213
2214
        logits_indices_padded = self.kv_sharing_fast_prefill_logits_indices[
            :num_logits_padded
        ]
2215
2216
        return logits_indices_padded

2217
    def _batch_mm_inputs_from_scheduler(
2218
2219
        self,
        scheduler_output: "SchedulerOutput",
2220
2221
    ) -> tuple[
        list[str],
2222
        list[tuple[str, MultiModalKwargsItem]],
2223
2224
        list[tuple[str, PlaceholderRange]],
    ]:
2225
        """Batch multimodal inputs from scheduled encoder inputs.
2226
2227
2228

        Args:
            scheduler_output: The scheduler output containing scheduled encoder
2229
                inputs.
2230
2231

        Returns:
2232
            A tuple of (mm_hashes, mm_kwargs, mm_lora_refs) where:
2233
2234
            - mm_hashes: List of multimodal hashes for each item
            - mm_kwargs: List of multimodal kwargs for each item
2235
            - mm_lora_refs: List of (req_id, placeholder_range) for each item
2236
        """
2237
2238
        scheduled_encoder_inputs = scheduler_output.scheduled_encoder_inputs
        if not scheduled_encoder_inputs:
2239
            return [], [], []
2240
2241

        mm_hashes = list[str]()
2242
        mm_kwargs = list[tuple[str, MultiModalKwargsItem]]()
2243
2244
2245
        # Multimodal LoRA reference info to map each multimodal item
        # back to its request & position
        mm_lora_refs = list[tuple[str, PlaceholderRange]]()
2246
2247
        for req_id, encoder_input_ids in scheduled_encoder_inputs.items():
            req_state = self.requests[req_id]
2248
2249

            for mm_input_id in encoder_input_ids:
2250
                mm_feature = req_state.mm_features[mm_input_id]
2251
2252
                if mm_feature.data is None:
                    continue
2253
2254

                mm_hashes.append(mm_feature.identifier)
2255
                mm_kwargs.append((mm_feature.modality, mm_feature.data))
2256
                mm_lora_refs.append((req_id, mm_feature.mm_position))
2257

2258
        return mm_hashes, mm_kwargs, mm_lora_refs
2259

2260
2261
2262
    def _execute_mm_encoder(
        self, scheduler_output: "SchedulerOutput"
    ) -> list[torch.Tensor]:
2263
2264
2265
        mm_hashes, mm_kwargs, mm_lora_refs = self._batch_mm_inputs_from_scheduler(
            scheduler_output
        )
2266
2267

        if not mm_kwargs:
2268
            return []
2269

2270
2271
2272
2273
2274
2275
        should_time = bool(
            self.observability_config
            and self.observability_config.enable_mm_processor_stats
            and scheduler_output.scheduled_encoder_inputs
        )

2276
2277
2278
2279
2280
2281
2282
        # 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.
2283
        model = cast(SupportsMultiModal, self.model)
2284
2285
2286
2287
2288
2289
2290
2291
2292
2293
2294
2295
2296
2297
2298
2299
2300
2301
2302
2303
2304
2305
2306
2307
2308
2309
2310
2311
2312
2313
2314
2315
2316
2317
2318
2319
2320
2321
2322
2323
2324
2325
2326
2327
2328
2329
2330
2331
2332
2333
2334
2335
2336
2337
2338
2339
2340

        if self.lora_config and self.lora_manager.supports_tower_connector_lora():
            # Build LoRA mappings independently for encoder inputs
            # (encoder batch structure is different from main batch)
            prompt_lora_mapping = []
            token_lora_mapping = []
            lora_requests = set()
            encoder_token_counts = []

            for req_id, pos_info in mm_lora_refs:
                req_idx = self.input_batch.req_id_to_index[req_id]
                lora_id = int(self.input_batch.request_lora_mapping[req_idx])

                # Prefer pos_info.get_num_embeds to count precise MM embedding tokens.
                num_tokens = self.model.get_num_mm_encoder_tokens(  # type: ignore[attr-defined]
                    pos_info.get_num_embeds
                )
                prompt_lora_mapping.append(lora_id)
                token_lora_mapping.extend([lora_id] * num_tokens)
                encoder_token_counts.append(num_tokens)

                if lora_id > 0:
                    lora_request = self.input_batch.lora_id_to_lora_request.get(lora_id)
                    if lora_request is not None:
                        lora_requests.add(lora_request)

            # Set tower adapter mapping
            tower_mapping = LoRAMapping(
                tuple(token_lora_mapping),
                tuple(prompt_lora_mapping),
                is_prefill=True,
                type=LoRAMappingType.TOWER,
            )
            self.lora_manager.set_active_adapters(lora_requests, tower_mapping)

            if hasattr(self.model, "get_num_mm_connector_tokens"):
                post_op_counts = [
                    self.model.get_num_mm_connector_tokens(num_tokens)  # type: ignore[attr-defined]
                    for num_tokens in encoder_token_counts
                ]

                connector_token_mapping = np.repeat(
                    np.array(prompt_lora_mapping, dtype=np.int32),
                    np.array(post_op_counts, dtype=np.int32),
                )
                connector_mapping = LoRAMapping(
                    index_mapping=tuple(connector_token_mapping.tolist()),
                    prompt_mapping=tuple(prompt_lora_mapping),
                    is_prefill=True,
                    type=LoRAMappingType.CONNECTOR,
                )

                self.lora_manager.set_active_adapters(
                    lora_requests,
                    connector_mapping,
                )

2341
        encoder_outputs: list[torch.Tensor] = []
2342
2343
        # Track the current index in mm_kwargs/mm_lora_refs to map groups to request IDs
        current_item_idx = 0
2344
        for modality, num_items, mm_kwargs_group in group_mm_kwargs_by_modality(
2345
2346
2347
            mm_kwargs,
            device=self.device,
            pin_memory=self.pin_memory,
2348
        ):
2349
            curr_group_outputs: MultiModalEmbeddings
2350
2351

            # EVS-related change.
2352
            # (ekhvedchenia): Temporary hack to limit peak memory usage when
2353
            # processing multimodal data. This solves the issue with scheduler
2354
2355
2356
2357
            # 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)
2358
2359
2360
2361
2362
2363
2364
            # 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
            ):
2365
                curr_group_outputs_lst = list[torch.Tensor]()
2366
2367
2368
2369
2370
2371
2372
2373
2374
2375
2376
                for video_idx in range(num_items):
                    video_mm_kwargs_item = mm_kwargs[current_item_idx + video_idx]
                    with self.timed_encoder_operation(
                        should_time, mm_lora_refs, current_item_idx + video_idx, 1
                    ):
                        _, _, micro_batch_mm_inputs = next(
                            group_mm_kwargs_by_modality(
                                [video_mm_kwargs_item],
                                device=self.device,
                                pin_memory=self.pin_memory,
                            )
2377
                        )
2378

2379
2380
2381
                        micro_batch_outputs = model.embed_multimodal(
                            **micro_batch_mm_inputs
                        )
2382

2383
                        curr_group_outputs_lst.extend(micro_batch_outputs)
2384
2385

                curr_group_outputs = curr_group_outputs_lst
2386
2387
2388
2389
2390
2391
2392
2393
            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.
2394
2395
2396
2397
2398

                with self.timed_encoder_operation(
                    should_time, mm_lora_refs, current_item_idx, num_items
                ):
                    curr_group_outputs = model.embed_multimodal(**mm_kwargs_group)
2399

2400
2401
            sanity_check_mm_encoder_outputs(
                curr_group_outputs,
2402
                expected_num_items=num_items,
2403
            )
2404
            encoder_outputs.extend(curr_group_outputs)
2405

2406
2407
            current_item_idx += num_items

2408
        # Cache the encoder outputs by mm_hash
2409
        for mm_hash, output in zip(mm_hashes, encoder_outputs):
2410
            self.encoder_cache[mm_hash] = output
2411
2412
            logger.debug("Finish execute for mm hash %s", mm_hash)
            self.maybe_save_ec_to_connector(self.encoder_cache, mm_hash)
2413

2414
2415
        return encoder_outputs

2416
    def _gather_mm_embeddings(
2417
2418
        self,
        scheduler_output: "SchedulerOutput",
2419
        shift_computed_tokens: int = 0,
2420
2421
2422
    ) -> tuple[list[torch.Tensor], torch.Tensor]:
        total_num_scheduled_tokens = scheduler_output.total_num_scheduled_tokens

2423
2424
2425
2426
2427
        # Swap to the other buffer to avoid race condition with previous
        # iteration's async copy that may still be reading from CPU.
        self.is_mm_embed_idx = 1 - self.is_mm_embed_idx
        is_mm_embed_buf = self.is_mm_embed_buffers[self.is_mm_embed_idx]

2428
        mm_embeds = list[torch.Tensor]()
2429
        is_mm_embed = is_mm_embed_buf.cpu
2430
2431
2432
        is_mm_embed[:total_num_scheduled_tokens] = False

        req_start_idx = 0
2433
        should_sync_mrope_positions = False
2434
        should_sync_xdrope_positions = False
2435

2436
        for req_id in self.input_batch.req_ids:
2437
2438
            mm_embeds_req: list[torch.Tensor] = []

2439
            num_scheduled_tokens = scheduler_output.num_scheduled_tokens[req_id]
2440
            req_state = self.requests[req_id]
2441
            num_computed_tokens = req_state.num_computed_tokens + shift_computed_tokens
2442

2443
2444
            for mm_feature in req_state.mm_features:
                pos_info = mm_feature.mm_position
2445
2446
                start_pos = pos_info.offset
                num_encoder_tokens = pos_info.length
2447
2448
2449
2450
2451
2452
2453
2454
2455
2456
2457
2458
2459
2460
2461
2462

                # 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,
2463
2464
                    num_encoder_tokens,
                )
2465
                assert start_idx < end_idx
2466
2467
2468
2469
2470
2471
2472
                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
2473

2474
                mm_hash = mm_feature.identifier
2475
                encoder_output = self.encoder_cache.get(mm_hash, None)
2476
                assert encoder_output is not None, f"Encoder cache miss for {mm_hash}."
2477
2478
2479

                if (is_embed := pos_info.is_embed) is not None:
                    is_embed = is_embed[start_idx:end_idx]
2480
2481
2482
                    mm_embeds_item = encoder_output[curr_embeds_start:curr_embeds_end]
                else:
                    mm_embeds_item = encoder_output[start_idx:end_idx]
2483

2484
                req_start_pos = req_start_idx + start_pos - num_computed_tokens
2485
2486
2487
2488
2489
2490
2491
2492
2493
                # OR mask for overlapping mm_features (use_audio_in_video)
                if is_embed is None:
                    is_mm_embed[req_start_pos + start_idx : req_start_pos + end_idx] = (
                        True
                    )
                else:
                    is_mm_embed[
                        req_start_pos + start_idx : req_start_pos + end_idx
                    ] |= is_embed
2494
2495
2496
                mm_embeds_req.append(mm_embeds_item)

            if self.is_multimodal_pruning_enabled and self.uses_mrope:
2497
                assert req_state.mrope_positions is not None
2498
2499
2500
2501
2502
2503
2504
                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,
2505
2506
                    )
                )
2507
2508
2509
2510
                req_state.mrope_positions.copy_(new_mrope_positions)
                req_state.mrope_position_delta = new_delta

            mm_embeds.extend(mm_embeds_req)
2511
2512
            req_start_idx += num_scheduled_tokens

2513
        is_mm_embed = is_mm_embed_buf.copy_to_gpu(total_num_scheduled_tokens)
2514
2515
2516

        if should_sync_mrope_positions:
            self._calc_mrope_positions(scheduler_output)
2517
            self.mrope_positions.copy_to_gpu(total_num_scheduled_tokens)
2518

2519
2520
2521
2522
        if should_sync_xdrope_positions:
            self._calc_xdrope_positions(scheduler_output)
            self.xdrope_positions.copy_to_gpu(total_num_scheduled_tokens)

2523
        return mm_embeds, is_mm_embed
2524

2525
    def get_model(self) -> nn.Module:
2526
        # get raw model out of the cudagraph wrapper.
2527
        if isinstance(self.model, (CUDAGraphWrapper, UBatchWrapper)):
2528
            return self.model.unwrap()
2529
2530
        return self.model

2531
2532
2533
2534
2535
2536
2537
2538
2539
2540
2541
2542
2543
2544
2545
    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

2546
2547
2548
2549
2550
    def get_supported_pooling_tasks(self) -> list[PoolingTask]:
        model = self.get_model()
        if not is_pooling_model(model):
            return []

2551
2552
        supported_tasks = list(model.pooler.get_supported_tasks())

2553
2554
2555
2556
        if "score" in supported_tasks:
            num_labels = getattr(self.model_config.hf_config, "num_labels", 0)
            if num_labels != 1:
                supported_tasks.remove("score")
2557
                logger.debug_once("Score API is only enabled for num_labels == 1.")
2558
2559

        return supported_tasks
2560

2561
2562
2563
2564
2565
2566
2567
2568
2569
2570
    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)

2571
    def sync_and_slice_intermediate_tensors(
2572
2573
        self,
        num_tokens: int,
2574
        intermediate_tensors: IntermediateTensors | None,
2575
2576
        sync_self: bool,
    ) -> IntermediateTensors:
2577
2578
2579
        assert self.intermediate_tensors is not None

        tp = self.vllm_config.parallel_config.tensor_parallel_size
2580
        is_rs = is_residual_scattered_for_sp(self.vllm_config, num_tokens)
2581
2582
2583
2584
2585
2586

        # 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():
2587
                is_scattered = k == "residual" and is_rs
2588
                copy_len = num_tokens // tp if is_scattered else num_tokens
2589
                self.intermediate_tensors[k][:copy_len].copy_(
2590
2591
2592
2593
2594
2595
2596
2597
2598
2599
2600
2601
2602
                    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:
2603
2604
2605
2606
2607
2608
2609
        """
        Step for the EPLB (Expert Parallelism Load Balancing) state.
        """
        if not self.parallel_config.enable_eplb:
            return

        assert self.eplb_state is not None
2610
2611
        model = self.get_model()
        assert is_mixture_of_experts(model)
2612
2613
2614
        self.eplb_state.step(
            is_dummy,
            is_profile,
2615
            log_stats=self.parallel_config.eplb_config.log_balancedness,
2616
2617
        )

2618
2619
2620
2621
2622
    def _pool(
        self,
        hidden_states: torch.Tensor,
        num_scheduled_tokens: int,
        num_scheduled_tokens_np: np.ndarray,
2623
2624
2625
2626
        kv_connector_output: KVConnectorOutput | None,
    ) -> ModelRunnerOutput | AsyncModelRunnerOutput:
        num_reqs = self.input_batch.num_reqs
        assert num_reqs == len(self.input_batch.pooling_params), (
2627
2628
            "Either all or none of the requests in a batch must be pooling request"
        )
2629

2630
        hidden_states = hidden_states[:num_scheduled_tokens]
2631
        seq_lens_cpu = self.seq_lens.cpu[:num_reqs]
2632

2633
        pooling_metadata = self.input_batch.get_pooling_metadata()
2634
        pooling_metadata.build_pooling_cursor(
2635
            num_scheduled_tokens_np, seq_lens_cpu, device=hidden_states.device
2636
        )
2637

2638
2639
        model = cast(VllmModelForPooling, self.model)
        raw_pooler_output: PoolerOutput = model.pooler(
2640
            hidden_states=hidden_states, pooling_metadata=pooling_metadata
2641
        )
2642
2643
2644
2645
2646
2647
2648
2649
2650
2651
2652
2653
2654
2655
2656
2657
2658
2659
2660
2661
2662
2663
2664
2665

        finished_mask = [
            seq_len == prompt_len
            for seq_len, prompt_len in zip(seq_lens_cpu, pooling_metadata.prompt_lens)
        ]

        model_runner_output = ModelRunnerOutput(
            req_ids=self.input_batch.req_ids.copy(),
            req_id_to_index=self.input_batch.req_id_to_index.copy(),
            kv_connector_output=kv_connector_output,
        )

        if raw_pooler_output is None or not any(finished_mask):
            model_runner_output.pooler_output = [None] * num_reqs
            return model_runner_output

        if self.use_async_scheduling:
            return AsyncGPUPoolingModelRunnerOutput(
                model_runner_output=model_runner_output,
                raw_pooler_output=raw_pooler_output,
                finished_mask=finished_mask,
                async_output_copy_stream=self.async_output_copy_stream,
            )

2666
        raw_pooler_output = json_map_leaves(
2667
            lambda x: None if x is None else x.to("cpu", non_blocking=True),
2668
2669
            raw_pooler_output,
        )
2670
2671
2672
2673
        model_runner_output.pooler_output = [
            out if include else None
            for out, include in zip(raw_pooler_output, finished_mask)
        ]
2674
2675
        self._sync_device()

2676
        return model_runner_output
2677

2678
    def _pad_for_sequence_parallelism(self, num_scheduled_tokens: int) -> int:
2679
2680
2681
        # Pad tokens to multiple of tensor_parallel_size when
        # enabled collective fusion for SP
        tp_size = self.vllm_config.parallel_config.tensor_parallel_size
2682
        if self.compilation_config.pass_config.enable_sp and tp_size > 1:
2683
2684
2685
            return round_up(num_scheduled_tokens, tp_size)
        return num_scheduled_tokens

Patrick von Platen's avatar
Patrick von Platen committed
2686
2687
2688
2689
2690
2691
2692
2693
2694
2695
2696
    def _prepare_mm_inputs(
        self, num_tokens: int
    ) -> tuple[torch.Tensor | None, torch.Tensor]:
        if self.model.requires_raw_input_tokens:
            input_ids = self.input_ids.gpu[:num_tokens]
        else:
            input_ids = None

        inputs_embeds = self.inputs_embeds.gpu[:num_tokens]
        return input_ids, inputs_embeds

2697
    def _preprocess(
2698
2699
        self,
        scheduler_output: "SchedulerOutput",
2700
        num_input_tokens: int,  # Padded
2701
        intermediate_tensors: IntermediateTensors | None = None,
2702
    ) -> tuple[
2703
2704
        torch.Tensor | None,
        torch.Tensor | None,
2705
        torch.Tensor,
2706
        IntermediateTensors | None,
2707
        dict[str, Any],
2708
        ECConnectorOutput | None,
2709
    ]:
2710
        num_scheduled_tokens = scheduler_output.total_num_scheduled_tokens
2711
        is_first_rank = get_pp_group().is_first_rank
2712
        is_encoder_decoder = self.model_config.is_encoder_decoder
2713

2714
2715
        # _prepare_inputs may reorder the batch, so we must gather multi
        # modal outputs after that to ensure the correct order
2716
2717
        ec_connector_output = None

2718
        if self.supports_mm_inputs and is_first_rank and not is_encoder_decoder:
2719
            # Run the multimodal encoder if any.
2720
2721
2722
2723
2724
2725
            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)
2726

2727
2728
2729
            # 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.
2730
            inputs_embeds_scheduled = self.model.embed_input_ids(
2731
2732
2733
                self.input_ids.gpu[:num_scheduled_tokens],
                multimodal_embeddings=mm_embeds,
                is_multimodal=is_mm_embed,
2734
            )
2735

2736
            # TODO(woosuk): Avoid the copy. Optimize.
2737
            self.inputs_embeds.gpu[:num_scheduled_tokens].copy_(inputs_embeds_scheduled)
2738

Patrick von Platen's avatar
Patrick von Platen committed
2739
            input_ids, inputs_embeds = self._prepare_mm_inputs(num_input_tokens)
2740
            model_kwargs = {
2741
                **self._init_model_kwargs(),
2742
2743
                **self._extract_mm_kwargs(scheduler_output),
            }
2744
        elif self.enable_prompt_embeds and is_first_rank:
2745
2746
2747
2748
2749
2750
2751
2752
2753
2754
2755
2756
            # 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).
2757
2758
2759
            token_ids_idx = (
                self.is_token_ids.gpu[:num_scheduled_tokens]
                .nonzero(as_tuple=False)
2760
                .squeeze(1)
2761
            )
2762
2763
2764
            # Some tokens ids may need to become embeds
            if token_ids_idx.numel() > 0:
                token_ids = self.input_ids.gpu[token_ids_idx]
2765
                tokens_to_embeds = self.model.embed_input_ids(input_ids=token_ids)
2766
2767
2768
                self.inputs_embeds.gpu[token_ids_idx] = tokens_to_embeds

            inputs_embeds = self.inputs_embeds.gpu[:num_input_tokens]
2769
            model_kwargs = self._init_model_kwargs()
2770
            input_ids = None
2771
        else:
2772
2773
2774
2775
            # 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.
2776
            input_ids = self.input_ids.gpu[:num_input_tokens]
2777
            inputs_embeds = None
2778
            model_kwargs = self._init_model_kwargs()
2779

2780
        if self.uses_mrope:
2781
            positions = self.mrope_positions.gpu[:, :num_input_tokens]
2782
2783
        elif self.uses_xdrope_dim > 0:
            positions = self.xdrope_positions.gpu[:, :num_input_tokens]
2784
        else:
2785
            positions = self.positions.gpu[:num_input_tokens]
2786

2787
        if is_first_rank:
2788
2789
            intermediate_tensors = None
        else:
2790
            assert intermediate_tensors is not None
2791
            intermediate_tensors = self.sync_and_slice_intermediate_tensors(
2792
2793
                num_input_tokens, intermediate_tensors, True
            )
2794

2795
        if is_encoder_decoder and scheduler_output.scheduled_encoder_inputs:
2796
2797
2798
2799
2800
2801
2802
            # 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})
2803

2804
2805
2806
2807
2808
2809
        return (
            input_ids,
            inputs_embeds,
            positions,
            intermediate_tensors,
            model_kwargs,
2810
            ec_connector_output,
2811
        )
2812

2813
    def _sample(
2814
        self,
2815
2816
        logits: torch.Tensor | None,
        spec_decode_metadata: SpecDecodeMetadata | None,
2817
    ) -> SamplerOutput:
2818
        # Sample the next token and get logprobs if needed.
2819
        sampling_metadata = self.input_batch.sampling_metadata
2820
2821
2822
        # 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()
2823
        if spec_decode_metadata is None:
2824
            return self.sampler(
2825
2826
2827
                logits=logits,
                sampling_metadata=sampling_metadata,
            )
2828

2829
2830
2831
2832
2833
2834
        # Update spec_token_ids with real draft tokens from pre step only when
        # output_token_ids is needed (penalties or bad_words are in use).
        if self.use_async_scheduling and self._draft_token_req_ids is not None:
            draft_token_ids_cpu, _ = self._get_draft_token_ids_cpu()
            self.input_batch.update_async_spec_token_ids(draft_token_ids_cpu)

2835
        sampler_output = self.rejection_sampler(
2836
2837
            spec_decode_metadata,
            None,  # draft_probs
2838
            logits,
2839
2840
            sampling_metadata,
        )
2841
2842
2843
        return sampler_output

    def _bookkeeping_sync(
2844
2845
2846
        self,
        scheduler_output: "SchedulerOutput",
        sampler_output: SamplerOutput,
2847
        logits: torch.Tensor | None,
2848
2849
        hidden_states: torch.Tensor,
        num_scheduled_tokens: int,
2850
        spec_decode_metadata: SpecDecodeMetadata | None,
2851
    ) -> tuple[
2852
        dict[str, int],
2853
        LogprobsLists | None,
2854
        list[list[int]],
2855
        dict[str, LogprobsTensors | None],
2856
2857
2858
        list[str],
        dict[str, int],
        list[int],
2859
    ]:
2860
2861
2862
2863
        num_nans_in_logits = {}
        if envs.VLLM_COMPUTE_NANS_IN_LOGITS:
            num_nans_in_logits = self._get_nans_in_logits(logits)

2864
2865
2866
2867
        num_reqs = self.input_batch.num_reqs
        discard_sampled_tokens_req_indices = np.nonzero(
            self.discard_request_mask.np[:num_reqs]
        )[0]
2868
2869
2870
2871
        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)
2872

2873
2874
2875
        # 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()
2876
        req_id_to_index_output_copy = self.input_batch.req_id_to_index.copy()
2877
2878

        num_sampled_tokens = sampler_output.sampled_token_ids.shape[0]
2879
        sampled_token_ids = sampler_output.sampled_token_ids
2880
        logprobs_tensors = sampler_output.logprobs_tensors
2881
        invalid_req_indices = []
2882
        logprobs_lists = None
2883
2884
2885
2886
2887
2888
        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)
2889
2890
2891
                # 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()
2892
2893
2894

                if logprobs_tensors is not None:
                    logprobs_lists = logprobs_tensors.tolists()
2895
2896
            else:
                # Includes spec decode tokens.
2897
                valid_sampled_token_ids, logprobs_lists = RejectionSampler.parse_output(
2898
2899
                    sampled_token_ids,
                    self.input_batch.vocab_size,
2900
                    discard_sampled_tokens_req_indices,
2901
                    logprobs_tensors=logprobs_tensors,
2902
                )
2903
        else:
2904
            valid_sampled_token_ids = []
2905
            invalid_req_indices = discard_sampled_tokens_req_indices.tolist()
2906
2907
2908
2909
2910
            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.
2911
2912
2913
2914
            # 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
2915
2916
2917
2918
2919
            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
            }
2920

2921
2922
2923
2924
2925
        # 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.
2926
        req_ids = self.input_batch.req_ids
2927
2928
        for req_idx in range(num_sampled_tokens):
            if self.use_async_scheduling:
2929
                sampled_ids = [-1] if req_idx not in invalid_req_indices_set else None
2930
2931
            else:
                sampled_ids = valid_sampled_token_ids[req_idx]
2932

2933
            num_sampled_ids: int = len(sampled_ids) if sampled_ids else 0
2934

2935
            if not sampled_ids:
2936
2937
2938
                continue

            start_idx = self.input_batch.num_tokens_no_spec[req_idx]
2939
            end_idx = start_idx + num_sampled_ids
2940
2941
2942
2943
            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}"
2944
            )
2945

2946
2947
            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
2948
            self.input_batch.num_tokens_no_spec[req_idx] = end_idx
2949

2950
            req_id = req_ids[req_idx]
2951
2952
2953
            req_state = self.requests[req_id]
            req_state.output_token_ids.extend(sampled_ids)

2954
2955
2956
2957
2958
2959
        # Compute prompt logprobs if needed.
        prompt_logprobs_dict = self._get_prompt_logprobs_dict(
            hidden_states[:num_scheduled_tokens],
            scheduler_output.num_scheduled_tokens,
        )

2960
2961
2962
2963
2964
2965
2966
2967
2968
2969
        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,
        )

2970
2971
2972
2973
2974
2975
2976
2977
2978
2979
2980
2981
2982
2983
2984
    @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()

2985
2986
    def _model_forward(
        self,
2987
2988
2989
2990
        input_ids: torch.Tensor | None = None,
        positions: torch.Tensor | None = None,
        intermediate_tensors: IntermediateTensors | None = None,
        inputs_embeds: torch.Tensor | None = None,
2991
2992
2993
2994
2995
        **model_kwargs: dict[str, Any],
    ) -> Any:
        """Helper method to call the model forward pass.

        This method can be overridden by subclasses for model execution.
2996
        Motivation: We can inspect only this method versus
2997
2998
2999
3000
3001
3002
3003
3004
3005
3006
3007
3008
3009
3010
3011
3012
3013
3014
3015
3016
        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,
        )

3017
3018
3019
3020
3021
3022
3023
3024
3025
3026
3027
3028
3029
3030
3031
3032
3033
3034
3035
3036
3037
    @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
        )

3038
3039
3040
3041
3042
3043
3044
3045
3046
3047
3048
3049
3050
    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,
3051
        num_encoder_reqs: int = 0,
3052
    ) -> tuple[
3053
3054
        CUDAGraphMode,
        BatchDescriptor,
3055
        bool,
3056
3057
        torch.Tensor | None,
        CUDAGraphStat | None,
3058
    ]:
3059
3060
3061
3062
3063
3064
        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,
3065
        )
3066
3067
3068
3069
3070
        # 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
        )
3071
3072
3073
3074
3075
3076
3077

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

3078
        num_tokens_padded = self._pad_for_sequence_parallelism(num_tokens)
3079
        dispatch_cudagraph = (
3080
            lambda num_tokens, disable_full: self.cudagraph_dispatcher.dispatch(
3081
3082
3083
                num_tokens=num_tokens,
                has_lora=has_lora,
                uniform_decode=uniform_decode,
3084
                disable_full=disable_full,
3085
3086
3087
3088
3089
            )
            if not force_eager
            else (CUDAGraphMode.NONE, BatchDescriptor(num_tokens_padded))
        )

3090
        cudagraph_mode, batch_descriptor = dispatch_cudagraph(
3091
            num_tokens_padded, use_cascade_attn or has_encoder_output
3092
        )
3093
        num_tokens_padded = batch_descriptor.num_tokens
3094
3095
3096
3097
3098
3099
3100
3101
3102
        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"
            )
3103
3104
3105

        # Extra coordination when running data-parallel since we need to coordinate
        # across ranks
3106
        should_ubatch, num_tokens_across_dp = False, None
3107
3108
3109
3110
3111
3112
3113
3114
3115
        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
            )

3116
3117
3118
3119
3120
3121
3122
3123
3124
3125
3126
            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,
                )
3127
3128
            )

3129
            # Extract DP-synced values
3130
3131
3132
            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())
3133
3134
3135
3136
3137
                # 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,
                )
3138
3139
3140
3141
                # 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

3142
3143
3144
3145
3146
3147
3148
3149
3150
3151
3152
3153
        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,
3154
            should_ubatch,
3155
3156
3157
            num_tokens_across_dp,
            cudagraph_stats,
        )
3158

3159
3160
3161
3162
3163
3164
3165
3166
3167
3168
3169
3170
3171
3172
3173
3174
3175
3176
3177
3178
3179
3180
3181
3182
3183
3184
3185
3186
3187
3188
3189
3190
3191
3192
3193
3194
    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

3195
3196
3197
3198
3199
3200
3201
3202
3203
3204
3205
3206
3207
3208
3209
3210
3211
3212
3213
3214
3215
3216
3217
3218
3219
3220
3221
3222
3223
3224
3225
3226
3227
3228
3229
3230
3231
3232
3233
3234
3235
3236
3237
3238
3239
3240
3241
3242
3243
3244
3245
3246
3247
3248
3249
3250
3251
3252
3253
3254
3255
3256
3257
3258
3259
3260
3261
3262
3263
3264
3265
3266
3267
3268
    def _get_slot_mappings(
        self,
        num_tokens_padded: int,
        num_reqs_padded: int,
        num_tokens_unpadded: int,
        ubatch_slices: "UBatchSlices | None" = None,
    ) -> tuple[
        dict[int, torch.Tensor] | None,
        dict[str, torch.Tensor] | list[dict[str, torch.Tensor]] | None,
    ]:
        """
        Build slot mappings in both formats needed by the system.

        Args:
            num_tokens_padded: Total number of tokens (padded)
            num_reqs_padded: Total number of requests (padded)
            num_tokens_unpadded: Actual number of tokens (unpadded)
            ubatch_slices: Optional ubatch slicing info for DBO

        Returns:
            A tuple of:
            - slot_mappings_by_gid: dict[int, torch.Tensor] for attention metadata
            - slot_mappings_by_layer: dict[str, torch.Tensor] or list for ForwardContext
        """
        if not (
            hasattr(self, "kv_cache_config")
            and self.kv_cache_config is not None
            and len(self.kv_cache_config.kv_cache_groups) > 0
        ):
            return None, None

        def _get_slot_mapping(kv_cache_gid: int):
            assert num_reqs_padded is not None and num_tokens_padded is not None
            kv_cache_spec = self.kv_cache_config.kv_cache_groups[
                kv_cache_gid
            ].kv_cache_spec
            if isinstance(kv_cache_spec, EncoderOnlyAttentionSpec):
                slot_mapping = torch.zeros(
                    (num_tokens_padded,),
                    dtype=torch.int64,
                    device=self.device,
                )
            else:
                blk_table = self.input_batch.block_table[kv_cache_gid]
                slot_mapping = blk_table.slot_mapping.gpu[:num_tokens_padded]

            # 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_unpadded:num_tokens_padded].fill_(-1)

            return slot_mapping

        slot_mappings_by_gid = {
            gid: _get_slot_mapping(gid)
            for gid, _ in enumerate(self.kv_cache_config.kv_cache_groups)
        }

        slot_mappings_by_layer: dict[str, torch.Tensor] = {}
        for gid, kv_cache_group in enumerate(self.kv_cache_config.kv_cache_groups):
            slot_mapping = slot_mappings_by_gid[gid]
            for layer_name in kv_cache_group.layer_names:
                slot_mappings_by_layer[layer_name] = slot_mapping

        if ubatch_slices is not None:
            result: list[dict[str, torch.Tensor]] = []
            for ubatch in ubatch_slices:
                sliced_mappings: dict[str, torch.Tensor] = {}
                for layer_name, slot_mapping in slot_mappings_by_layer.items():
                    sliced_mappings[layer_name] = slot_mapping[ubatch.token_slice]
                result.append(sliced_mappings)
            return slot_mappings_by_gid, result

        return slot_mappings_by_gid, slot_mappings_by_layer

3269
3270
3271
3272
    @torch.inference_mode()
    def execute_model(
        self,
        scheduler_output: "SchedulerOutput",
3273
        intermediate_tensors: IntermediateTensors | None = None,
3274
    ) -> ModelRunnerOutput | AsyncModelRunnerOutput | IntermediateTensors | None:
3275
3276
3277
3278
3279
        if self.execute_model_state is not None:
            raise RuntimeError(
                "State error: sample_tokens() must be called "
                "after execute_model() returns None."
            )
3280

3281
3282
3283
3284
3285
3286
3287
        if self.vllm_config.model_config.enable_return_routed_experts:
            capturer = RoutedExpertsCapturer.get_instance()
            if capturer is not None:
                capturer.clear_buffer()  # noqa
            else:
                logger.error("RoutedExpertsCapturer not initialized.")

3288
3289
3290
3291
3292
        if scheduler_output.preempted_req_ids and has_kv_transfer_group():
            get_kv_transfer_group().handle_preemptions(
                scheduler_output.preempted_req_ids
            )

3293
        num_scheduled_tokens = scheduler_output.total_num_scheduled_tokens
3294
3295
3296
3297
3298
3299
        with (
            record_function_or_nullcontext("gpu_model_runner: preprocess"),
            self.synchronize_input_prep(),
        ):
            # Update persistent batch states.
            self._update_states(scheduler_output)
3300

3301
3302
            if has_ec_transfer() and get_ec_transfer().is_producer:
                with self.maybe_get_ec_connector_output(
3303
                    scheduler_output,
3304
3305
3306
3307
3308
3309
3310
3311
3312
3313
3314
3315
3316
3317
3318
3319
3320
3321
3322
3323
3324
3325
3326
3327
3328
3329
3330
3331
                    encoder_cache=self.encoder_cache,
                ) as ec_connector_output:
                    self._execute_mm_encoder(scheduler_output)
                    return make_empty_encoder_model_runner_output(scheduler_output)

            if not num_scheduled_tokens:
                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.
                    self._dummy_run(1)
                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(scheduler_output, self.vllm_config)

            if self.cache_config.kv_sharing_fast_prefill:
                assert not self.num_prompt_logprobs, (
                    "--kv-sharing-fast-prefill produces incorrect "
                    "logprobs for prompt tokens, tokens, please disable "
                    "it when the requests need prompt logprobs"
3332
3333
                )

3334
3335
3336
3337
3338
3339
            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())
            num_tokens_unpadded = scheduler_output.total_num_scheduled_tokens
3340

3341
3342
3343
3344
            logits_indices, spec_decode_metadata = self._prepare_inputs(
                scheduler_output,
                num_scheduled_tokens_np,
            )
3345

3346
3347
3348
3349
3350
            cascade_attn_prefix_lens = None
            # Disable cascade attention when using microbatching (DBO)
            if self.cascade_attn_enabled and not self.parallel_config.use_ubatching:
                # Pre-compute cascade attention prefix lengths
                cascade_attn_prefix_lens = self._compute_cascade_attn_prefix_lens(
3351
                    num_scheduled_tokens_np,
3352
3353
                    self.input_batch.num_computed_tokens_cpu[:num_reqs],
                    scheduler_output.num_common_prefix_blocks,
3354
3355
                )

3356
3357
3358
3359
3360
3361
3362
3363
3364
3365
3366
3367
3368
3369
            (
                cudagraph_mode,
                batch_desc,
                should_ubatch,
                num_tokens_across_dp,
                cudagraph_stats,
            ) = 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,
                num_encoder_reqs=len(scheduler_output.scheduled_encoder_inputs),
            )
3370

3371
3372
3373
3374
3375
3376
3377
3378
3379
3380
3381
3382
3383
3384
3385
3386
3387
3388
3389
3390
3391
3392
3393
3394
3395
3396
3397
            logger.debug(
                "Running batch with cudagraph_mode: %s, batch_descriptor: %s, "
                "should_ubatch: %s, num_tokens_across_dp: %s",
                cudagraph_mode,
                batch_desc,
                should_ubatch,
                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
            )
            ubatch_slices, ubatch_slices_padded = maybe_create_ubatch_slices(
                should_ubatch,
                num_scheduled_tokens_np,
                num_tokens_padded,
                num_reqs_padded,
                self.parallel_config.num_ubatches,
            )

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

3398
3399
3400
3401
3402
3403
3404
3405
3406
3407
3408
            # True if any attention backend handles KV cache update separately
            # from forward() (i.e., forward_includes_kv_cache_update=False). When true,
            # slot_mappings must use padded dimensions to match the key/value tensors.
            has_separate_kv_update = not all(
                all(
                    g.backend.forward_includes_kv_cache_update
                    for g in self.attn_groups[id]
                )
                for id, spec in enumerate(self.kv_cache_config.kv_cache_groups)
                if not isinstance(spec.kv_cache_spec, EncoderOnlyAttentionSpec)
            )
3409
3410
            pad_attn = cudagraph_mode == CUDAGraphMode.FULL

3411
3412
3413
3414
3415
3416
3417
3418
3419
3420
3421
3422
            if self.cache_config.mamba_cache_mode == "align":
                mamba_utils.preprocess_mamba(
                    scheduler_output,
                    self.kv_cache_config,
                    self.cache_config,
                    self.mamba_state_idx,
                    self.input_batch,
                    self.requests,
                    self.compilation_config.static_forward_context,
                    self.model.get_mamba_state_copy_func(),
                )

3423
3424
3425
            use_spec_decode = len(scheduler_output.scheduled_spec_decode_tokens) > 0
            ubatch_slices_attn = ubatch_slices_padded if pad_attn else ubatch_slices

3426
3427
3428
3429
3430
3431
3432
3433
3434
3435
3436
            slot_mappings_by_group, slot_mappings = self._get_slot_mappings(
                num_tokens_padded=num_tokens_padded
                if pad_attn or has_separate_kv_update
                else num_tokens_unpadded,
                num_reqs_padded=(
                    num_reqs_padded if pad_attn or has_separate_kv_update else num_reqs
                ),
                num_tokens_unpadded=num_tokens_unpadded,
                ubatch_slices=ubatch_slices_padded,
            )

3437
3438
3439
3440
3441
3442
3443
3444
3445
3446
3447
3448
            attn_metadata, spec_decode_common_attn_metadata = (
                self._build_attention_metadata(
                    num_tokens=num_tokens_unpadded,
                    num_tokens_padded=num_tokens_padded if pad_attn else None,
                    num_reqs=num_reqs,
                    num_reqs_padded=num_reqs_padded if pad_attn else None,
                    max_query_len=max_num_scheduled_tokens,
                    ubatch_slices=ubatch_slices_attn,
                    logits_indices=logits_indices,
                    use_spec_decode=use_spec_decode,
                    num_scheduled_tokens=scheduler_output.num_scheduled_tokens,
                    cascade_attn_prefix_lens=cascade_attn_prefix_lens,
3449
                    slot_mappings=slot_mappings_by_group,
3450
                )
3451
            )
3452

3453
3454
3455
3456
3457
3458
3459
3460
3461
            (
                input_ids,
                inputs_embeds,
                positions,
                intermediate_tensors,
                model_kwargs,
                ec_connector_output,
            ) = self._preprocess(
                scheduler_output, num_tokens_padded, intermediate_tensors
3462
            )
3463

3464
        # Set cudagraph mode to none if calc_kv_scales is true.
3465
3466
3467
        # KV scales calculation involves dynamic operations that are incompatible
        # with CUDA graph capture.
        if self.calculate_kv_scales:
3468
            cudagraph_mode = CUDAGraphMode.NONE
3469
3470
            # Mark KV scales as calculated after the first forward pass
            self.calculate_kv_scales = False
3471

3472
3473
3474
3475
3476
3477
3478
        # Encoder-decoder models can only compile the pure decode steps where no
        # encoder inputs are present. Use eager for the first pass.
        num_encoder_reqs = len(scheduler_output.scheduled_encoder_inputs)
        has_encoder_input = (
            self.model_config.is_encoder_decoder and num_encoder_reqs > 0
        )

3479
3480
        # Run the model.
        # Use persistent buffers for CUDA graphs.
3481
3482
        with (
            set_forward_context(
3483
3484
                attn_metadata,
                self.vllm_config,
3485
                num_tokens=num_tokens_padded,
3486
                num_tokens_across_dp=num_tokens_across_dp,
3487
3488
                cudagraph_runtime_mode=cudagraph_mode,
                batch_descriptor=batch_desc,
3489
                ubatch_slices=ubatch_slices_padded,
3490
                slot_mapping=slot_mappings,
3491
                skip_compiled=has_encoder_input,
3492
            ),
3493
            record_function_or_nullcontext("gpu_model_runner: forward"),
3494
3495
            self.maybe_get_kv_connector_output(scheduler_output) as kv_connector_output,
        ):
3496
            model_output = self._model_forward(
3497
3498
3499
3500
3501
3502
3503
                input_ids=input_ids,
                positions=positions,
                intermediate_tensors=intermediate_tensors,
                inputs_embeds=inputs_embeds,
                **model_kwargs,
            )

3504
        with record_function_or_nullcontext("gpu_model_runner: postprocess"):
3505
            if self.use_aux_hidden_state_outputs:
3506
                # True when EAGLE 3 is used.
3507
3508
                hidden_states, aux_hidden_states = model_output
            else:
3509
                # Common case.
3510
3511
3512
                hidden_states = model_output
                aux_hidden_states = None

3513
3514
3515
3516
3517
            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)
3518
                    hidden_states.kv_connector_output = kv_connector_output
3519
                    self.kv_connector_output = kv_connector_output
3520
                    return hidden_states
3521

3522
                if self.is_pooling_model:
3523
                    # Return the pooling output.
3524
3525
3526
3527
3528
                    return self._pool(
                        hidden_states,
                        num_scheduled_tokens,
                        num_scheduled_tokens_np,
                        kv_connector_output,
3529
                    )
3530
3531

                sample_hidden_states = hidden_states[logits_indices]
3532
                logits = self.model.compute_logits(sample_hidden_states)
3533
3534
3535
3536
            else:
                # Rare case.
                assert not self.is_pooling_model

3537
                sample_hidden_states = hidden_states[logits_indices]
3538
                if not get_pp_group().is_last_rank:
3539
                    all_gather_tensors = {
3540
                        "residual": not is_residual_scattered_for_sp(
3541
                            self.vllm_config, num_tokens_padded
3542
                        )
3543
                    }
3544
                    get_pp_group().send_tensor_dict(
3545
3546
                        hidden_states.tensors,
                        all_gather_group=get_tp_group(),
3547
3548
                        all_gather_tensors=all_gather_tensors,
                    )
3549
3550
                    logits = None
                else:
3551
                    logits = self.model.compute_logits(sample_hidden_states)
3552

3553
                model_output_broadcast_data: dict[str, Any] = {}
3554
3555
3556
                if logits is not None:
                    model_output_broadcast_data["logits"] = logits.contiguous()

3557
                broadcasted = get_pp_group().broadcast_tensor_dict(
3558
3559
                    model_output_broadcast_data, src=len(get_pp_group().ranks) - 1
                )
3560
3561
                assert broadcasted is not None
                logits = broadcasted["logits"]
3562

3563
3564
3565
3566
3567
3568
3569
3570
        self.execute_model_state = ExecuteModelState(
            scheduler_output,
            logits,
            spec_decode_metadata,
            spec_decode_common_attn_metadata,
            hidden_states,
            sample_hidden_states,
            aux_hidden_states,
3571
            ec_connector_output,
3572
            cudagraph_stats,
3573
            slot_mappings,
3574
        )
3575
        self.kv_connector_output = kv_connector_output
3576
3577
3578
3579
3580
3581
        return None

    @torch.inference_mode
    def sample_tokens(
        self, grammar_output: "GrammarOutput | None"
    ) -> ModelRunnerOutput | AsyncModelRunnerOutput | IntermediateTensors:
3582
3583
3584
        kv_connector_output = self.kv_connector_output
        self.kv_connector_output = None

3585
        if self.execute_model_state is None:
3586
3587
3588
            # receive sampled token ids from the last PP rank.
            if self.use_async_scheduling and get_pp_group().world_size > 1:
                self._pp_receive_prev_sampled_token_ids_to_input_batch()
3589
            if not kv_connector_output:
3590
                return None  # type: ignore[return-value]
3591
3592
3593
3594
3595
3596
3597
3598
3599

            # 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
3600
3601
3602
3603
3604
3605
3606
3607
3608
3609

        # Unpack ephemeral state.
        (
            scheduler_output,
            logits,
            spec_decode_metadata,
            spec_decode_common_attn_metadata,
            hidden_states,
            sample_hidden_states,
            aux_hidden_states,
3610
            ec_connector_output,
3611
            cudagraph_stats,
3612
            slot_mappings,
3613
3614
3615
3616
3617
3618
3619
3620
3621
        ) = 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
            )
3622

3623
        with record_function_or_nullcontext("gpu_model_runner: sample"):
3624
3625
            sampler_output = self._sample(logits, spec_decode_metadata)

3626
3627
3628
        self._update_states_after_model_execute(
            sampler_output.sampled_token_ids, scheduler_output
        )
3629
3630
3631
3632
3633
3634
        if self.use_async_scheduling:
            pp = get_pp_group()
            if pp.world_size > 1 and pp.is_last_rank:
                self._pp_broadcast_prev_sampled_token_ids(
                    sampler_output.sampled_token_ids
                )
3635

3636
3637
        self._draft_token_ids = None
        self._draft_token_req_ids = None
3638
3639
        self.input_batch.prev_sampled_token_ids = None

3640
        def propose_draft_token_ids(sampled_token_ids):
3641
            assert spec_decode_common_attn_metadata is not None
3642
            with record_function_or_nullcontext("gpu_model_runner: draft"):
3643
3644
3645
3646
3647
3648
3649
3650
3651
                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,
3652
                    slot_mappings,
3653
                )
3654
                self._copy_draft_token_ids_to_cpu(scheduler_output)
3655

3656
        spec_config = self.speculative_config
3657
3658
3659
3660
3661
        propose_drafts_after_bookkeeping = False
        if spec_config is not None:
            input_fits_in_drafter = spec_decode_common_attn_metadata is not None and (
                spec_decode_common_attn_metadata.max_seq_len + self.num_spec_tokens
                <= self.effective_drafter_max_model_len
3662
            )
3663
3664
3665
3666
3667
            use_gpu_toks = (
                spec_config.use_eagle() or spec_config.uses_draft_model()
            ) and not spec_config.disable_padded_drafter_batch
            if use_gpu_toks:
                # EAGLE/DraftModel speculative decoding can use the GPU sampled tokens
3668
                # as inputs, and does not need to wait for bookkeeping to finish.
3669
                assert isinstance(self.drafter, EagleProposer | DraftModelProposer)
3670
3671
3672
3673
3674
3675
3676
3677
3678
3679
3680
3681
3682
                sampled_token_ids = sampler_output.sampled_token_ids
                if input_fits_in_drafter:
                    propose_draft_token_ids(sampled_token_ids)
                elif self.valid_sampled_token_count_event is not None:
                    assert spec_decode_common_attn_metadata is not None
                    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,
                            self.discard_request_mask.gpu,
                        )
3683
                    )
3684
3685
3686
3687
3688
3689
3690
3691
3692
3693
3694
                    self._copy_valid_sampled_token_count(
                        next_token_ids, valid_sampled_tokens_count
                    )
                    # Since we couldn't run the drafter,
                    # just use zeros for the draft tokens.
                    self._draft_token_ids = torch.zeros(
                        1, device=self.device, dtype=torch.int32
                    ).expand(len(self.input_batch.req_ids), self.num_spec_tokens)
                    self._copy_draft_token_ids_to_cpu(scheduler_output, zeros_only=True)
            else:
                propose_drafts_after_bookkeeping = input_fits_in_drafter
3695

3696
        with record_function_or_nullcontext("gpu_model_runner: bookkeep"):
3697
3698
3699
3700
3701
3702
3703
3704
            (
                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,
3705
3706
3707
3708
3709
            ) = self._bookkeeping_sync(
                scheduler_output,
                sampler_output,
                logits,
                hidden_states,
3710
                scheduler_output.total_num_scheduled_tokens,
3711
                spec_decode_metadata,
3712
            )
3713

3714
        if propose_drafts_after_bookkeeping:
3715
3716
3717
            # 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)
3718

3719
        with record_function_or_nullcontext("gpu_model_runner: eplb"):
3720
            self.eplb_step()
3721

3722
        with record_function_or_nullcontext("gpu_model_runner: ModelRunnerOutput"):
3723
3724
3725
3726
3727
3728
3729
            if self.model_config.enable_return_routed_experts:
                capturer = RoutedExpertsCapturer.get_instance()
                if capturer is not None:
                    capturer.save_captured_experts(indices=self.slot_mapping)  # noqa
                else:
                    logger.error("RoutedExpertsCapturer not initialized.")

3730
3731
3732
3733
3734
3735
3736
            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,
                kv_connector_output=kv_connector_output,
3737
3738
3739
                ec_connector_output=ec_connector_output
                if self.supports_mm_inputs
                else None,
3740
                num_nans_in_logits=num_nans_in_logits,
3741
                cudagraph_stats=cudagraph_stats,
3742
            )
3743

3744
3745
        if not self.use_async_scheduling:
            return output
3746

3747
3748
3749
3750
3751
3752
3753
3754
3755
        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,
3756
                vocab_size=self.input_batch.vocab_size,
3757
3758
3759
3760
3761
            )
        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
3762
            # any requests with sampling params that require output ids.
3763
3764
3765
3766
            self.input_batch.set_async_sampled_token_ids(
                async_output.sampled_token_ids_cpu,
                async_output.async_copy_ready_event,
            )
3767
3768
3769

        return async_output

3770
3771
3772
3773
3774
3775
3776
3777
3778
3779
3780
3781
3782
3783
3784
3785
3786
3787
3788
3789
3790
3791
3792
3793
3794
3795
3796
3797
3798
3799
3800
3801
3802
3803
3804
3805
3806
3807
3808
    def _pp_broadcast_prev_sampled_token_ids(
        self, sampled_token_ids: torch.Tensor
    ) -> None:
        """Broadcast sampled token ids (GPU) from last PP stage"""
        pp = get_pp_group()
        assert pp.is_last_rank
        # `prev_sampled_token_ids` is expected to have shape [num_reqs, 1].
        assert sampled_token_ids.dim() == 2 and sampled_token_ids.shape[-1] == 1, (
            "PP+async expects sampled_token_ids to have shape [num_reqs, 1]"
        )
        torch.distributed.broadcast(
            sampled_token_ids, src=pp.rank, group=pp.device_group
        )

    def _pp_receive_prev_sampled_token_ids_to_input_batch(self) -> None:
        """Receive sampled token ids broadcast from last PP stage"""
        pp = get_pp_group()
        assert not pp.is_last_rank
        num_reqs = self.input_batch.num_reqs
        # `prev_sampled_token_ids` is expected to have shape [num_reqs, 1].
        recv = torch.empty((num_reqs, 1), dtype=torch.int32, device=self.device)
        torch.distributed.broadcast(recv, src=pp.last_rank, group=pp.device_group)
        self.input_batch.prev_sampled_token_ids = recv

        # construct `prev_req_id_to_index` here so `_prepare_input_ids`
        # can map req_id -> previous batch row
        discard_req_indices = np.nonzero(self.discard_request_mask.np[:num_reqs])[0]
        discard_req_indices_set = set(discard_req_indices)
        prev_req_id_to_index: dict[str, int] = {}
        for i, req_id in enumerate(self.input_batch.req_ids):
            if i in discard_req_indices_set:
                continue
            prev_req_id_to_index[req_id] = i
            # PP+async scheduling: advance per-request local cached output length by
            # appending a placeholder (-1) token id.
            if (req_state := self.requests.get(req_id)) is not None:
                req_state.output_token_ids.append(-1)
        self.input_batch.prev_req_id_to_index = prev_req_id_to_index

3809
    def take_draft_token_ids(self) -> DraftTokenIds | None:
3810
        if not self.num_spec_tokens or not self._draft_token_req_ids:
3811
            return None
3812
        draft_token_ids, req_ids = self._get_draft_token_ids_cpu()
3813
        return DraftTokenIds(req_ids, draft_token_ids)
3814

3815
3816
3817
    def _copy_draft_token_ids_to_cpu(
        self, scheduler_output: "SchedulerOutput", zeros_only: bool = False
    ) -> None:
3818
3819
3820
3821
3822
3823
        # Check if we need to copy draft tokens to CPU. In async scheduling,
        # we only copy when needed for structured output, penalties or bad_words.
        if self.use_async_scheduling and not (
            scheduler_output.has_structured_output_requests
            or self.input_batch.sampling_metadata.output_token_ids
        ):
3824
3825
3826
            return
        # We must also set the corresponding request ids.
        self._draft_token_req_ids = self.input_batch.req_ids.copy()
3827

3828
3829
3830
3831
3832
3833
3834
3835
3836
3837
3838
3839
3840
3841
3842
3843
3844
3845
3846
3847
        draft_token_ids: torch.Tensor = self._draft_token_ids
        if not torch.is_tensor(draft_token_ids):
            return
        assert self.draft_token_ids_event is not None
        assert self.draft_token_ids_copy_stream is not None
        assert self.draft_token_ids_cpu is not None
        default_stream = torch.cuda.current_stream()
        num_reqs = draft_token_ids.shape[0]
        with torch.cuda.stream(self.draft_token_ids_copy_stream):
            if not zeros_only:
                # Trigger async copy of draft token ids to cpu.
                self.draft_token_ids_copy_stream.wait_stream(default_stream)
                self.draft_token_ids_cpu[:num_reqs].copy_(
                    draft_token_ids, non_blocking=True
                )
            else:
                # No copy needed, just zero-out cpu tensor.
                self.draft_token_ids_cpu[:num_reqs] = 0
            self.draft_token_ids_event.record()

3848
    def _get_draft_token_ids_cpu(self) -> tuple[list[list[int]], list[str]]:
3849
        if isinstance(self._draft_token_ids, list):
3850
3851
3852
3853
            return self._draft_token_ids, self.input_batch.req_ids
        req_ids = self._draft_token_req_ids
        if req_ids is None:
            return [], []
3854
3855
3856
        assert self.draft_token_ids_event is not None
        assert self.draft_token_ids_cpu is not None
        self.draft_token_ids_event.synchronize()
3857
        return self.draft_token_ids_cpu[: len(req_ids)].tolist(), req_ids
3858

3859
3860
3861
3862
3863
3864
3865
3866
3867
3868
3869
3870
3871
    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
3872
            assert counts_cpu is not None
3873
3874
3875
3876
3877
3878
3879
3880
            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
3881
3882
        sampled_count_event = self.valid_sampled_token_count_event
        if sampled_count_event is None or prev_sampled_token_ids is None:
3883
3884
3885
            return []

        counts_cpu = self.valid_sampled_token_count_cpu
3886
3887
        assert counts_cpu is not None
        sampled_count_event.synchronize()
3888
3889
        return counts_cpu[: prev_sampled_token_ids.shape[0]].tolist()

3890
3891
3892
    def propose_draft_token_ids(
        self,
        scheduler_output: "SchedulerOutput",
3893
        sampled_token_ids: torch.Tensor | list[list[int]],
3894
3895
3896
        sampling_metadata: SamplingMetadata,
        hidden_states: torch.Tensor,
        sample_hidden_states: torch.Tensor,
3897
3898
        aux_hidden_states: list[torch.Tensor] | None,
        spec_decode_metadata: SpecDecodeMetadata | None,
3899
        common_attn_metadata: CommonAttentionMetadata,
3900
        slot_mappings: dict[str, torch.Tensor] | list[dict[str, torch.Tensor]] | None,
3901
    ) -> list[list[int]] | torch.Tensor:
3902
        num_scheduled_tokens = scheduler_output.total_num_scheduled_tokens
3903
3904
3905
        spec_config = self.speculative_config
        assert spec_config is not None
        if spec_config.method == "ngram":
3906
3907
            from vllm.v1.spec_decode.ngram_proposer import NgramProposer

3908
            assert isinstance(sampled_token_ids, list)
3909
            assert isinstance(self.drafter, NgramProposer)
3910
            draft_token_ids = self.drafter.propose(
3911
                sampled_token_ids,
3912
3913
                self.input_batch.num_tokens_no_spec,
                self.input_batch.token_ids_cpu,
3914
                slot_mappings=slot_mappings,
3915
            )
3916
        elif spec_config.method == "suffix":
3917
3918
            assert isinstance(sampled_token_ids, list)
            assert isinstance(self.drafter, SuffixDecodingProposer)
3919
3920
3921
            draft_token_ids = self.drafter.propose(
                self.input_batch, sampled_token_ids, slot_mappings=slot_mappings
            )
3922
        elif spec_config.method == "medusa":
3923
            assert isinstance(sampled_token_ids, list)
3924
            assert isinstance(self.drafter, MedusaProposer)
3925

3926
3927
            if sample_hidden_states.shape[0] == len(sampled_token_ids):
                # The input to the target model does not include draft tokens.
3928
3929
3930
3931
                hidden_states = sample_hidden_states
            else:
                indices = []
                offset = 0
3932
3933
3934
                assert spec_decode_metadata is not None, (
                    "No spec decode metadata for medusa"
                )
3935
                for num_draft, tokens in zip(
3936
3937
                    spec_decode_metadata.num_draft_tokens, sampled_token_ids
                ):
3938
                    indices.append(offset + len(tokens) - 1)
3939
                    offset += num_draft + 1
3940
                indices = torch.tensor(indices, device=self.device)
3941
3942
                hidden_states = sample_hidden_states[indices]

3943
            draft_token_ids = self.drafter.propose(
3944
3945
                target_hidden_states=hidden_states,
                sampling_metadata=sampling_metadata,
3946
                slot_mappings=slot_mappings,
3947
            )
3948
3949
        elif spec_config.use_eagle() or spec_config.uses_draft_model():
            assert isinstance(self.drafter, EagleProposer | DraftModelProposer)
3950

3951
            if spec_config.disable_padded_drafter_batch:
3952
3953
3954
                # 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.
3955
3956
                assert isinstance(sampled_token_ids, list), (
                    "sampled_token_ids should be a python list when"
3957
                    "padded-batch is disabled."
3958
                )
3959
                next_token_ids = self.drafter.prepare_next_token_ids_cpu(
3960
3961
3962
3963
3964
                    sampled_token_ids,
                    self.requests,
                    self.input_batch,
                    scheduler_output.num_scheduled_tokens,
                )
3965
3966
3967
3968
3969
            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.
3970
3971
                assert isinstance(sampled_token_ids, torch.Tensor), (
                    "sampled_token_ids should be a torch.Tensor when"
3972
                    "padded-batch is enabled."
3973
3974
                )
                next_token_ids, valid_sampled_tokens_count = (
3975
3976
3977
3978
3979
                    self.drafter.prepare_next_token_ids_padded(
                        common_attn_metadata,
                        sampled_token_ids,
                        self.requests,
                        self.input_batch,
3980
                        self.discard_request_mask.gpu,
3981
                    )
3982
                )
3983
3984
3985
                self._copy_valid_sampled_token_count(
                    next_token_ids, valid_sampled_tokens_count
                )
Jiayi Yao's avatar
Jiayi Yao committed
3986

3987
            num_rejected_tokens_gpu = None
3988
            if spec_decode_metadata is None:
3989
                token_indices_to_sample = None
3990
                # input_ids can be None for multimodal models.
3991
                target_token_ids = self.input_ids.gpu[:num_scheduled_tokens]
3992
                target_positions = self._get_positions(num_scheduled_tokens)
3993
                if self.use_aux_hidden_state_outputs:
Wentao Ye's avatar
Wentao Ye committed
3994
                    assert aux_hidden_states is not None
3995
                    target_hidden_states = torch.cat(
3996
3997
                        [h[:num_scheduled_tokens] for h in aux_hidden_states], dim=-1
                    )
3998
3999
                else:
                    target_hidden_states = hidden_states[:num_scheduled_tokens]
4000
            else:
4001
                if spec_config.disable_padded_drafter_batch:
4002
                    token_indices_to_sample = None
4003
4004
4005
4006
4007
                    common_attn_metadata, token_indices = self.drafter.prepare_inputs(
                        common_attn_metadata,
                        sampled_token_ids,
                        spec_decode_metadata.num_draft_tokens,
                    )
4008
4009
4010
4011
4012
4013
4014
4015
4016
                    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]
4017
                else:
4018
4019
4020
4021
4022
4023
4024
4025
                    (
                        common_attn_metadata,
                        token_indices_to_sample,
                        num_rejected_tokens_gpu,
                    ) = self.drafter.prepare_inputs_padded(
                        common_attn_metadata,
                        spec_decode_metadata,
                        valid_sampled_tokens_count,
4026
                    )
4027
4028
4029
4030
4031
4032
4033
4034
4035
4036
4037
                    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]
4038

4039
            if self.supports_mm_inputs:
4040
4041
4042
4043
4044
4045
                mm_embed_inputs = self._gather_mm_embeddings(
                    scheduler_output,
                    shift_computed_tokens=1,
                )
            else:
                mm_embed_inputs = None
4046

4047
            draft_token_ids = self.drafter.propose(
4048
4049
4050
4051
                target_token_ids=target_token_ids,
                target_positions=target_positions,
                target_hidden_states=target_hidden_states,
                next_token_ids=next_token_ids,
4052
                last_token_indices=token_indices_to_sample,
4053
                sampling_metadata=sampling_metadata,
4054
                common_attn_metadata=common_attn_metadata,
4055
                mm_embed_inputs=mm_embed_inputs,
4056
                num_rejected_tokens_gpu=num_rejected_tokens_gpu,
4057
                slot_mappings=slot_mappings,
4058
            )
4059

4060
        return draft_token_ids
4061

4062
4063
4064
    def update_config(self, overrides: dict[str, Any]) -> None:
        allowed_config_names = {"load_config", "model_config"}
        for config_name, config_overrides in overrides.items():
4065
4066
            assert config_name in allowed_config_names, (
                f"Config `{config_name}` not supported. "
4067
                f"Allowed configs: {allowed_config_names}"
4068
            )
4069
4070
4071
4072
            config = getattr(self, config_name)
            new_config = update_config(config, config_overrides)
            setattr(self, config_name, new_config)

4073
4074
4075
4076
4077
    def load_model(self, eep_scale_up: bool = False) -> None:
        """
        Args:
            eep_scale_up: the model loading is for elastic EP scale up.
        """
4078
4079
4080
4081
4082
        logger.info_once(
            "Starting to load model %s...",
            self.model_config.model,
            scope="global",
        )
4083
4084
4085
4086
4087
        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)
        )
4088

4089
4090
4091
4092
        if self.parallel_config.enable_eplb:
            self.eplb_state = EplbState(self.parallel_config, self.device)
            eplb_models = 0

4093
4094
4095
4096
4097
4098
4099
4100
4101
4102
        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
4103
                    )
4104
4105
4106
4107
4108
4109
4110
4111
4112
4113
4114
4115
4116
4117
4118
                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,
                        )
4119

4120
4121
4122
4123
4124
4125
4126
4127
4128
4129
4130
4131
4132
4133
4134
4135
4136
4137
4138
4139
4140
4141
                        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
4142

4143
4144
4145
4146
4147
4148
                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"
                        )
4149

4150
4151
4152
4153
4154
4155
4156
4157
4158
4159
4160
4161
4162
4163
4164
4165
4166
4167
4168
4169
4170
4171
4172
4173
4174
                    # 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
4175
        logger.info_once(
4176
4177
            "Model loading took %s GiB memory and %.6f seconds",
            format_gib(self.model_memory_usage),
4178
            time_after_load - time_before_load,
4179
            scope="local",
4180
        )
4181
        prepare_communication_buffer_for_model(self.model)
4182
4183
4184
4185
        if (drafter := getattr(self, "drafter", None)) and (
            drafter_model := getattr(drafter, "model", None)
        ):
            prepare_communication_buffer_for_model(drafter_model)
4186
        mm_config = self.model_config.multimodal_config
4187
        self.is_multimodal_pruning_enabled = (
4188
            supports_multimodal_pruning(self.get_model())
4189
4190
            and mm_config is not None
            and mm_config.is_multimodal_pruning_enabled()
4191
        )
4192

4193
        if is_mixture_of_experts(self.model) and self.parallel_config.enable_eplb:
4194
4195
4196
4197
4198
4199
4200
4201
4202
4203
4204
            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(
4205
                self.model,
4206
                self.model_config,
4207
4208
4209
                global_expert_load,
                old_global_expert_indices,
                rank_mapping,
4210
            )
4211
4212
            if self.eplb_state.is_async:
                self.eplb_state.start_async_loop(rank_mapping=rank_mapping)
4213

4214
        if (
4215
4216
            self.vllm_config.compilation_config.mode
            == CompilationMode.STOCK_TORCH_COMPILE
4217
        ):
4218
            backend = self.vllm_config.compilation_config.init_backend(self.vllm_config)
4219
            compilation_counter.stock_torch_compile_count += 1
4220
            self.model.compile(fullgraph=True, backend=backend)
4221
            return
4222
        # for other compilation modes, cudagraph behavior is controlled by
4223
4224
4225
        # CudagraphWraper and CudagraphDispatcher of vllm.

        # wrap the model with full cudagraph wrapper if needed.
4226
4227
        cudagraph_mode = self.compilation_config.cudagraph_mode
        assert cudagraph_mode is not None
4228
4229
4230
4231
        if (
            cudagraph_mode.has_full_cudagraphs()
            and not self.parallel_config.use_ubatching
        ):
4232
4233
4234
            self.model = CUDAGraphWrapper(
                self.model, self.vllm_config, runtime_mode=CUDAGraphMode.FULL
            )
4235
        elif self.parallel_config.use_ubatching:
4236
            if cudagraph_mode.has_full_cudagraphs():
4237
4238
4239
                self.model = UBatchWrapper(
                    self.model, self.vllm_config, CUDAGraphMode.FULL, self.device
                )
4240
            else:
4241
4242
4243
                self.model = UBatchWrapper(
                    self.model, self.vllm_config, CUDAGraphMode.NONE, self.device
                )
4244

4245
    def _get_eagle3_aux_layers_from_config(self) -> tuple[int, ...] | None:
4246
4247
4248
4249
4250
4251
4252
4253
4254
4255
4256
4257
4258
4259
4260
4261
4262
4263
4264
4265
4266
4267
4268
        """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

4269
    def reload_weights(self) -> None:
4270
        assert getattr(self, "model", None) is not None, (
4271
            "Cannot reload weights before model is loaded."
4272
        )
4273
4274
        model_loader = get_model_loader(self.load_config)
        logger.info("Reloading weights inplace...")
4275
        model_loader.load_weights(self.get_model(), model_config=self.model_config)
4276

4277
4278
4279
4280
4281
    def save_tensorized_model(
        self,
        tensorizer_config: "TensorizerConfig",
    ) -> None:
        TensorizerLoader.save_model(
4282
            self.get_model(),
4283
            tensorizer_config=tensorizer_config,
4284
            model_config=self.model_config,
4285
4286
        )

4287
4288
4289
    def _get_prompt_logprobs_dict(
        self,
        hidden_states: torch.Tensor,
4290
        num_scheduled_tokens: dict[str, int],
4291
    ) -> dict[str, LogprobsTensors | None]:
4292
        num_prompt_logprobs_dict = self.num_prompt_logprobs
4293
4294
4295
        if not num_prompt_logprobs_dict:
            return {}

4296
        in_progress_dict = self.input_batch.in_progress_prompt_logprobs_cpu
4297
        prompt_logprobs_dict: dict[str, LogprobsTensors | None] = {}
4298
4299
4300
4301
4302

        # 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():
4303
4304
4305
4306
            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
4307
4308
4309

            # Get metadata for this request.
            request = self.requests[req_id]
4310
4311
4312
4313
            if request.prompt_token_ids is None:
                # Prompt logprobs is incompatible with prompt embeddings
                continue

4314
4315
            num_prompt_tokens = len(request.prompt_token_ids)
            prompt_token_ids = torch.tensor(request.prompt_token_ids).to(
4316
4317
                self.device, non_blocking=True
            )
4318

4319
4320
4321
4322
4323
4324
            # 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(
4325
4326
                    num_prompt_tokens - 1, num_prompt_logprobs + 1
                )
4327
4328
                in_progress_dict[req_id] = logprobs_tensors

4329
            # Determine number of logits to retrieve.
4330
4331
            start_idx = request.num_computed_tokens
            start_tok = start_idx + 1
4332
            num_remaining_tokens = num_prompt_tokens - start_tok
4333
            if num_tokens <= num_remaining_tokens:
4334
                # This is a chunk, more tokens remain.
4335
4336
4337
                # 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.
4338
4339
4340
4341
4342
                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)
4343
4344
4345
4346
4347
4348
4349
                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
4350
4351
4352
4353
4354

            # 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]
4355
            offset = self.query_start_loc.np[req_idx].item()
4356
            prompt_hidden_states = hidden_states[offset : offset + num_logits]
4357
            logits = self.model.compute_logits(prompt_hidden_states)
4358
4359
4360
4361

            # 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.
4362
            tgt_token_ids = prompt_token_ids[start_tok : start_tok + num_logits]
4363
4364

            # Compute prompt logprobs.
4365
4366
            logprobs = self.sampler.compute_logprobs(logits)
            token_ids, logprobs, ranks = self.sampler.gather_logprobs(
4367
4368
                logprobs, num_prompt_logprobs, tgt_token_ids
            )
4369
4370

            # Transfer GPU->CPU async.
4371
4372
            chunk_slice = slice(start_idx, start_idx + num_logits)
            logprobs_tensors.logprob_token_ids[chunk_slice].copy_(
4373
4374
4375
                token_ids, non_blocking=True
            )
            logprobs_tensors.logprobs[chunk_slice].copy_(logprobs, non_blocking=True)
4376
            logprobs_tensors.selected_token_ranks[chunk_slice].copy_(
4377
4378
                ranks, non_blocking=True
            )
4379
4380
4381
4382
4383

        # 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]
4384
            del in_progress_dict[req_id]
4385
4386

        # Must synchronize the non-blocking GPU->CPU transfers.
4387
        if prompt_logprobs_dict:
4388
            self._sync_device()
4389
4390
4391

        return prompt_logprobs_dict

4392
4393
    def _get_nans_in_logits(
        self,
4394
        logits: torch.Tensor | None,
4395
4396
4397
4398
4399
4400
4401
4402
4403
4404
4405
    ) -> 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])
4406
4407
4408
                    if num_nans_for_index is not None and req_index < logits.shape[0]
                    else 0
                )
4409
4410
4411
4412
            return num_nans_in_logits
        except IndexError:
            return {}

4413
    @contextmanager
4414
4415
4416
    def maybe_randomize_inputs(
        self, input_ids: torch.Tensor | None, inputs_embeds: torch.Tensor | None
    ):
4417
4418
4419
4420
        """
        Randomize input_ids if VLLM_RANDOMIZE_DP_DUMMY_INPUTS is set.
        This is to help balance expert-selection
         - during profile_run
4421
         - during DP rank dummy run
4422
        """
4423

4424
4425
4426
4427
        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
4428
        elif input_ids is not None:
4429
4430
4431
4432

            @functools.cache
            def rand_input_ids() -> torch.Tensor:
                return torch.randint_like(
4433
                    self.input_ids.gpu,
4434
4435
                    low=0,
                    high=self.model_config.get_vocab_size(),
4436
                )
4437

4438
            logger.debug_once("Randomizing dummy input_ids for DP Rank")
4439
            input_ids.copy_(rand_input_ids()[: input_ids.size(0)], non_blocking=True)
4440
4441
            yield
            input_ids.fill_(0)
4442
4443
4444
4445
4446
4447
4448
4449
4450
4451
4452
4453
4454
4455
4456
        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)
4457

4458
4459
4460
4461
4462
4463
    def _get_mm_dummy_batch(
        self,
        modality: str,
        max_items_per_batch: int,
    ) -> BatchedTensorInputs:
        """Dummy data for profiling and precompiling multimodal models."""
4464
4465
        assert self.mm_budget is not None

4466
4467
4468
        # Don't use `max_items_per_batch` here to avoid redundant computation
        dummy_mm_inputs = self.mm_registry.get_dummy_mm_inputs(
            self.model_config,
4469
            mm_counts={modality: 1},
4470
            cache=self.mm_budget.cache,
4471
        )
4472
4473
4474
4475
4476
        dummy_mm_item = dummy_mm_inputs["mm_kwargs"][modality][0]

        # We use the cache so that the item is saved to the cache,
        # but not read from the cache
        assert dummy_mm_item is not None, "Item should not already be cached"
4477

4478
4479
4480
        return next(
            mm_kwargs_group
            for _, _, mm_kwargs_group in group_mm_kwargs_by_modality(
4481
                [(modality, dummy_mm_item)] * max_items_per_batch,
4482
4483
4484
4485
                device=self.device,
                pin_memory=self.pin_memory,
            )
        )
4486

4487
4488
4489
4490
    @torch.inference_mode()
    def _dummy_run(
        self,
        num_tokens: int,
4491
        cudagraph_runtime_mode: CUDAGraphMode | None = None,
4492
4493
        force_attention: bool = False,
        uniform_decode: bool = False,
4494
        allow_microbatching: bool = True,
4495
4496
        skip_eplb: bool = False,
        is_profile: bool = False,
4497
        create_mixed_batch: bool = False,
4498
        remove_lora: bool = True,
4499
        activate_lora: bool = False,
Rémi Delacourt's avatar
Rémi Delacourt committed
4500
        is_graph_capturing: bool = False,
4501
    ) -> tuple[torch.Tensor, torch.Tensor]:
4502
4503
4504
4505
4506
4507
4508
        """
        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.
4509
                - if not set will determine the cudagraph mode based on using
4510
                    the self.cudagraph_dispatcher.
4511
4512
4513
4514
                - CUDAGraphMode.NONE: No cudagraph, for warm up and profile run
                - CUDAGraphMode.PIECEWISE: Piecewise cudagraph.
                - CUDAGraphMode.FULL: Full cudagraph, attention metadata is
                    needed.
4515
            force_attention: If True, always create attention metadata. Used to
4516
4517
4518
4519
                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.
4520
4521
            create_mixed_batch: If True, create a mixed batch with both decode
                (1 token) and prefill (multiple tokens) requests.
4522
            remove_lora: If False, dummy LoRAs are not destroyed after the run
4523
            activate_lora: If False, dummy_run is performed without LoRAs.
4524
        """
4525
4526
        mm_config = self.vllm_config.model_config.multimodal_config
        if mm_config and mm_config.mm_encoder_only:
4527
4528
4529
4530
            # The current dummy run only covers LM execution, so we can skip it.
            # mm encoder dummy run may need to add in the future.
            return torch.tensor([]), torch.tensor([])

4531
4532
4533
4534
        assert (
            cudagraph_runtime_mode is None
            or cudagraph_runtime_mode.valid_runtime_modes()
        )
4535

4536
        # If cudagraph_mode.decode_mode() == FULL and
4537
        # cudagraph_mode.separate_routine(). This means that we are using
4538
4539
4540
4541
4542
4543
4544
4545
4546
4547
4548
        # 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.
4549
        max_query_len = self.uniform_decode_query_len if uniform_decode else num_tokens
4550

4551
4552
4553
4554
4555
        # 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
4556
4557
4558
4559
        if create_mixed_batch:
            assert not uniform_decode
            # Create mixed batch:
            # first half decode tokens, second half one prefill
4560
            num_decode_tokens = min(max_num_reqs - 1, num_tokens // 2)
4561
4562
4563
4564
            num_prefill_tokens = num_tokens - num_decode_tokens
            num_reqs = num_decode_tokens + 1

            # Create decode requests (1 token each) followed by prefill request
4565
            num_scheduled_tokens_list = [1] * num_decode_tokens + [num_prefill_tokens]
4566
4567
4568
            # Note: Overriding max_query_len to be the prefill tokens
            max_query_len = num_prefill_tokens
        elif uniform_decode:
4569
            assert not create_mixed_batch
4570
            num_reqs = min(max_num_reqs, cdiv(num_tokens, max_query_len))
4571
4572
            num_scheduled_tokens_list = [max_query_len] * num_reqs
            if num_tokens % max_query_len != 0:
4573
                num_scheduled_tokens_list[-1] = num_tokens % max_query_len
4574
4575
4576
4577
4578
4579
        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

4580
4581
        assert sum(num_scheduled_tokens_list) == num_tokens
        assert len(num_scheduled_tokens_list) == num_reqs
4582
        num_scheduled_tokens = np.array(num_scheduled_tokens_list, dtype=np.int32)
4583
4584
        num_tokens_unpadded = int(num_scheduled_tokens.sum())

4585
        num_sampled_tokens = np.ones(num_reqs, dtype=np.int32)
4586

4587
        _cudagraph_mode, batch_desc, should_ubatch, num_tokens_across_dp, _ = (
4588
4589
4590
4591
4592
4593
4594
4595
4596
4597
4598
4599
4600
4601
4602
4603
4604
4605
            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,
4606
4607
            )
        )
4608
4609
4610

        if cudagraph_runtime_mode is None:
            cudagraph_runtime_mode = _cudagraph_mode
4611
        else:
4612
4613
4614
4615
4616
4617
4618
4619
4620
            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
        )
4621
        ubatch_slices, ubatch_slices_padded = maybe_create_ubatch_slices(
4622
4623
4624
4625
4626
4627
4628
4629
4630
4631
            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,
4632
        )
4633

4634
        attn_metadata: PerLayerAttnMetadata | None = None
4635

4636
4637
4638
4639
4640
4641
4642
        slot_mappings_by_group, slot_mappings = self._get_slot_mappings(
            num_tokens_padded=num_tokens,
            num_reqs_padded=num_reqs_padded,
            num_tokens_unpadded=num_tokens_unpadded,
            ubatch_slices=ubatch_slices_padded,
        )

4643
4644
        # If force_attention is True, we always capture attention. Otherwise,
        # it only happens for cudagraph_runtime_mode=FULL.
4645
        if force_attention or cudagraph_runtime_mode == CUDAGraphMode.FULL:
4646
4647
4648
4649
4650
4651
            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:
4652
                seq_lens = max_query_len  # type: ignore[assignment]
4653
            self.seq_lens.np[:num_reqs] = seq_lens
4654
4655
            self.seq_lens.np[num_reqs:] = 0
            self.seq_lens.copy_to_gpu()
4656

4657
4658
            cum_num_tokens, _ = self._get_cumsum_and_arange(num_scheduled_tokens)
            self.query_start_loc.np[1 : num_reqs + 1] = cum_num_tokens
4659
4660
            self.query_start_loc.copy_to_gpu()

4661
            pad_attn = cudagraph_runtime_mode == CUDAGraphMode.FULL
4662
            attn_metadata, _ = self._build_attention_metadata(
4663
4664
4665
                num_tokens=num_tokens_unpadded,
                num_reqs=num_reqs_padded,
                max_query_len=max_query_len,
4666
                ubatch_slices=ubatch_slices_padded if pad_attn else ubatch_slices,
4667
                for_cudagraph_capture=is_graph_capturing,
4668
                slot_mappings=slot_mappings_by_group,
4669
            )
4670

4671
        with self.maybe_dummy_run_with_lora(
4672
4673
4674
4675
4676
            self.lora_config,
            num_scheduled_tokens,
            num_sampled_tokens,
            activate_lora,
            remove_lora,
4677
        ):
4678
            # Make sure padding doesn't exceed max_num_tokens
4679
            assert num_tokens_padded <= self.max_num_tokens
4680
            model_kwargs = self._init_model_kwargs()
4681
            if self.supports_mm_inputs and not self.model_config.is_encoder_decoder:
Patrick von Platen's avatar
Patrick von Platen committed
4682
4683
                input_ids, inputs_embeds = self._prepare_mm_inputs(num_tokens_padded)

4684
                model_kwargs = {
4685
                    **model_kwargs,
4686
4687
                    **self._dummy_mm_kwargs(num_reqs),
                }
4688
4689
            elif self.enable_prompt_embeds:
                input_ids = None
4690
                inputs_embeds = self.inputs_embeds.gpu[:num_tokens_padded]
4691
                model_kwargs = self._init_model_kwargs()
4692
            else:
4693
                input_ids = self.input_ids.gpu[:num_tokens_padded]
4694
                inputs_embeds = None
4695

4696
            if self.uses_mrope:
4697
                positions = self.mrope_positions.gpu[:, :num_tokens_padded]
4698
            elif self.uses_xdrope_dim > 0:
4699
                positions = self.xdrope_positions.gpu[:, :num_tokens_padded]
4700
            else:
4701
                positions = self.positions.gpu[:num_tokens_padded]
4702
4703
4704
4705
4706
4707
4708
4709
4710

            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,
4711
4712
4713
                            device=self.device,
                        )
                    )
4714
4715

                intermediate_tensors = self.sync_and_slice_intermediate_tensors(
4716
                    num_tokens_padded, None, False
4717
                )
4718

4719
            if ubatch_slices_padded is not None:
4720
4721
4722
                # Adjust values to reflect a single ubatch.
                # TODO(sage,lucas): this is cruft that should be addressed in
                #  the padding refactor.
4723
                num_tokens_padded = ubatch_slices_padded[0].num_tokens
4724
                if num_tokens_across_dp is not None:
4725
                    num_tokens_across_dp[:] = num_tokens_padded
4726

4727
            with (
4728
                self.maybe_randomize_inputs(input_ids, inputs_embeds),
4729
                set_forward_context(
4730
4731
                    attn_metadata,
                    self.vllm_config,
4732
                    num_tokens=num_tokens_padded,
4733
4734
                    num_tokens_across_dp=num_tokens_across_dp,
                    cudagraph_runtime_mode=cudagraph_runtime_mode,
4735
                    batch_descriptor=batch_desc,
4736
                    ubatch_slices=ubatch_slices_padded,
4737
                    slot_mapping=slot_mappings,
4738
4739
                ),
            ):
4740
                outputs = self.model(
4741
4742
4743
4744
                    input_ids=input_ids,
                    positions=positions,
                    intermediate_tensors=intermediate_tensors,
                    inputs_embeds=inputs_embeds,
4745
                    **model_kwargs,
4746
                )
4747

4748
4749
4750
4751
            if self.use_aux_hidden_state_outputs:
                hidden_states, _ = outputs
            else:
                hidden_states = outputs
4752

4753
4754
4755
4756
4757
4758
            if self.speculative_config and (
                self.speculative_config.use_eagle()
                or self.speculative_config.uses_draft_model()
            ):
                assert isinstance(self.drafter, EagleProposer | DraftModelProposer)
                assert self.speculative_config is not None
4759
4760
4761
                # 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.
4762
                use_cudagraphs = (
4763
4764
4765
4766
4767
4768
4769
4770
4771
                    (
                        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
4772
4773
4774
4775
4776
4777
4778
4779
4780
4781
4782

                # 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
4783
                    is_graph_capturing=is_graph_capturing,
4784
                    slot_mappings=slot_mappings,
4785
                )
4786

4787
4788
4789
4790
4791
4792
4793
4794
4795
4796
4797
        # 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()

4798
4799
4800
4801
4802
4803
4804
4805
4806
4807
        # 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)

4808
        logit_indices = np.cumsum(num_scheduled_tokens) - 1
4809
4810
4811
4812
        logit_indices_device = torch.from_numpy(logit_indices).to(
            self.device, non_blocking=True
        )
        return hidden_states, hidden_states[logit_indices_device]
4813
4814
4815
4816
4817
4818

    @torch.inference_mode()
    def _dummy_sampler_run(
        self,
        hidden_states: torch.Tensor,
    ) -> torch.Tensor:
4819
4820
4821
        # The dummy hidden states may contain special values,
        # like `inf` or `nan`.
        # To avoid breaking the sampler, we use a random tensor here instead.
4822

4823
4824
        mm_config = self.vllm_config.model_config.multimodal_config
        if mm_config and mm_config.mm_encoder_only:
4825
4826
4827
            # MM Encoder only model no need to run sampler.
            return torch.tensor([])

4828
        hidden_states = torch.rand_like(hidden_states)
4829

4830
        logits = self.model.compute_logits(hidden_states)
4831
4832
        num_reqs = logits.size(0)

4833
        dummy_tensors = lambda v: torch.full((num_reqs,), v, device=self.device)
4834
4835
4836
4837
4838
4839
4840
4841
4842
4843
4844
4845
4846
4847
4848

        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)],
4849
            spec_token_ids=[[] for _ in range(num_reqs)],
4850
4851
            allowed_token_ids_mask=None,
            bad_words_token_ids={},
4852
            logitsprocs=LogitsProcessors(),
4853
        )
4854
        try:
4855
4856
4857
            sampler_output = self.sampler(
                logits=logits, sampling_metadata=dummy_metadata
            )
4858
        except RuntimeError as e:
4859
            if "out of memory" in str(e):
4860
4861
4862
4863
                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 "
4864
4865
                    "initializing the engine."
                ) from e
4866
4867
            else:
                raise e
4868
        if self.speculative_config:
4869
4870
            draft_token_ids = [[0] for _ in range(num_reqs)]
            dummy_spec_decode_metadata = SpecDecodeMetadata.make_dummy(
4871
4872
                draft_token_ids, self.device
            )
4873
4874
4875
4876
4877
4878

            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
4879
4880
4881
4882
4883
            logits = torch.randn(
                num_tokens + num_reqs,
                logits.shape[-1],
                device=self.device,
                dtype=logits.dtype,
4884
            )
4885
4886
4887
            self.rejection_sampler(
                dummy_spec_decode_metadata,
                draft_probs,
4888
                logits,
4889
4890
                dummy_metadata,
            )
4891
        return sampler_output
4892

4893
    def _dummy_pooler_run_task(
4894
4895
        self,
        hidden_states: torch.Tensor,
4896
4897
        task: PoolingTask,
    ) -> PoolerOutput:
4898
4899
4900
4901
        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
4902
4903
4904
4905
        num_scheduled_tokens_np = np.full(num_reqs, min_tokens_per_req)
        num_scheduled_tokens_np[-1] += num_tokens % num_reqs
        assert np.sum(num_scheduled_tokens_np) == num_tokens
        assert len(num_scheduled_tokens_np) == num_reqs
4906
4907
4908

        req_num_tokens = num_tokens // num_reqs

4909
        dummy_prompt_lens = torch.from_numpy(num_scheduled_tokens_np)
4910
4911
4912
        dummy_token_ids = torch.zeros(
            (num_reqs, req_num_tokens), dtype=torch.int32, device=self.device
        )
4913

4914
        model = cast(VllmModelForPooling, self.get_model())
4915
        dummy_pooling_params = PoolingParams(task=task)
4916
        dummy_pooling_params.verify(task=task, model_config=self.model_config)
4917
        to_update = model.pooler.get_pooling_updates(task)
4918
4919
        to_update.apply(dummy_pooling_params)

4920
        dummy_metadata = PoolingMetadata(
4921
4922
4923
            prompt_lens=dummy_prompt_lens,
            prompt_token_ids=dummy_token_ids,
            pooling_params=[dummy_pooling_params] * num_reqs,
4924
            pooling_states=[PoolingStates() for i in range(num_reqs)],
4925
        )
4926

4927
        dummy_metadata.build_pooling_cursor(
4928
            num_scheduled_tokens_np,
4929
4930
            seq_lens_cpu=dummy_prompt_lens,
            device=hidden_states.device,
4931
        )
4932

4933
        try:
4934
4935
4936
            return model.pooler(
                hidden_states=hidden_states, pooling_metadata=dummy_metadata
            )
4937
        except RuntimeError as e:
4938
            if "out of memory" in str(e):
4939
                raise RuntimeError(
4940
4941
4942
                    "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 "
4943
4944
                    "initializing the engine."
                ) from e
4945
4946
            else:
                raise e
4947
4948
4949
4950
4951
4952

    @torch.inference_mode()
    def _dummy_pooler_run(
        self,
        hidden_states: torch.Tensor,
    ) -> PoolerOutput:
4953
4954
        mm_config = self.vllm_config.model_config.multimodal_config
        if mm_config and mm_config.mm_encoder_only:
4955
4956
4957
            # MM Encoder only model not need to run pooler.
            return torch.tensor([])

4958
        # Find the task that has the largest output for subsequent steps
4959
4960
4961
        supported_pooling_tasks = self.get_supported_pooling_tasks()

        if not supported_pooling_tasks:
4962
4963
4964
4965
4966
4967
            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."
            )
4968

4969
        output_size = dict[PoolingTask, float]()
4970
        for task in supported_pooling_tasks:
4971
4972
            # Run a full batch with each task to ensure none of them OOMs
            output = self._dummy_pooler_run_task(hidden_states, task)
4973
            output_size[task] = sum(o.nbytes for o in output if o is not None)
4974
4975
4976
4977
            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)
4978

4979
    def profile_run(self) -> None:
4980
        # Profile with multimodal encoder & encoder cache.
4981
        if self.supports_mm_inputs:
4982
4983
            mm_config = self.model_config.multimodal_config
            if mm_config is not None and mm_config.skip_mm_profiling:
4984
                logger.info(
4985
                    "Skipping memory profiling for multimodal encoder and "
4986
4987
                    "encoder cache."
                )
4988
4989
4990
4991
4992
4993
4994
4995
            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.
4996
                    dummy_modality = mm_budget.get_modality_with_max_tokens()
4997
4998
4999
                    max_mm_items_per_batch = mm_budget.max_items_per_batch_by_modality[
                        dummy_modality
                    ]
5000
5001
5002
5003
5004
5005
5006
5007
5008

                    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,
                    )
5009

5010
5011
5012
5013
5014
                    # Create dummy batch of multimodal inputs.
                    batched_dummy_mm_inputs = self._get_mm_dummy_batch(
                        dummy_modality,
                        max_mm_items_per_batch,
                    )
5015

5016
                    # Run multimodal encoder.
5017
                    dummy_encoder_outputs = self.model.embed_multimodal(
5018
5019
                        **batched_dummy_mm_inputs
                    )
5020

5021
5022
5023
5024
                    sanity_check_mm_encoder_outputs(
                        dummy_encoder_outputs,
                        expected_num_items=max_mm_items_per_batch,
                    )
5025
5026
                    for i, output in enumerate(dummy_encoder_outputs):
                        self.encoder_cache[f"tmp_{i}"] = output
5027

5028
        # Add `is_profile` here to pre-allocate communication buffers
5029
5030
5031
        hidden_states, last_hidden_states = self._dummy_run(
            self.max_num_tokens, is_profile=True
        )
5032
        if get_pp_group().is_last_rank:
5033
5034
5035
5036
            if self.is_pooling_model:
                output = self._dummy_pooler_run(hidden_states)
            else:
                output = self._dummy_sampler_run(last_hidden_states)
5037
        else:
5038
            output = None
5039
        self._sync_device()
5040
        del hidden_states, output
5041
        self.encoder_cache.clear()
5042
        gc.collect()
5043

5044
    def capture_model(self) -> int:
5045
        if self.compilation_config.cudagraph_mode == CUDAGraphMode.NONE:
5046
            logger.warning(
5047
                "Skipping CUDA graph capture. To turn on CUDA graph capture, "
5048
5049
                "ensure `cudagraph_mode` was not manually set to `NONE`"
            )
5050
            return 0
5051

5052
5053
        compilation_counter.num_gpu_runner_capture_triggers += 1

5054
5055
        start_time = time.perf_counter()

5056
5057
5058
5059
5060
5061
5062
5063
5064
5065
5066
5067
5068
5069
        @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()
5070
                    gc.collect()
5071

5072
5073
5074
        # 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.
5075
        set_cudagraph_capturing_enabled(True)
5076
        with freeze_gc(), graph_capture(device=self.device):
5077
            start_free_gpu_memory = torch.cuda.mem_get_info()[0]
5078

5079
5080
5081
5082
            for (
                runtime_mode,
                batch_descs,
            ) in self.cudagraph_dispatcher.get_capture_descs():
5083
                self._capture_cudagraphs(
5084
5085
                    batch_descriptors=batch_descs,
                    cudagraph_runtime_mode=runtime_mode,
5086
                )
5087

5088
5089
5090
            torch.cuda.synchronize()
            end_free_gpu_memory = torch.cuda.mem_get_info()[0]

5091
5092
5093
        # 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
5094
        # we may do lazy capturing in future that still allows capturing
5095
5096
        # after here.
        set_cudagraph_capturing_enabled(False)
5097

5098
5099
5100
5101
        # Lock workspace to prevent resizing during execution.
        # Max workspace sizes should have been captured during warmup/profiling.
        lock_workspace()

5102
5103
5104
5105
        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.
5106
        logger.info_once(
5107
5108
5109
            "Graph capturing finished in %.0f secs, took %.2f GiB",
            elapsed_time,
            cuda_graph_size / (1 << 30),
5110
            scope="local",
5111
        )
5112
        return cuda_graph_size
5113

5114
5115
    def _capture_cudagraphs(
        self,
5116
        batch_descriptors: list[BatchDescriptor],
5117
5118
5119
5120
5121
5122
        cudagraph_runtime_mode: CUDAGraphMode,
    ):
        assert (
            cudagraph_runtime_mode != CUDAGraphMode.NONE
            and cudagraph_runtime_mode.valid_runtime_modes()
        ), f"Invalid cudagraph runtime mode: {cudagraph_runtime_mode}"
5123

5124
5125
5126
5127
5128
5129
5130
5131
5132
5133
5134
5135
5136
5137
        if not batch_descriptors:
            return

        uniform_decode = batch_descriptors[0].uniform
        force_attention = cudagraph_runtime_mode == CUDAGraphMode.FULL

        dummy_run = functools.partial(
            self._dummy_run,
            uniform_decode=uniform_decode,
            skip_eplb=True,
            remove_lora=False,
            force_attention=force_attention,
        )

5138
5139
        # Only rank 0 should print progress bar during capture
        if is_global_first_rank():
5140
5141
            batch_descriptors = tqdm(
                batch_descriptors,
5142
5143
5144
                disable=not self.load_config.use_tqdm_on_load,
                desc="Capturing CUDA graphs ({}, {})".format(
                    "decode" if uniform_decode else "mixed prefill-decode",
5145
5146
5147
                    cudagraph_runtime_mode.name,
                ),
            )
5148

5149
        # We skip EPLB here since we don't want to record dummy metrics
5150
5151
5152
5153
        for batch_desc in batch_descriptors:
            num_tokens = batch_desc.num_tokens
            activate_lora = batch_desc.has_lora

5154
            # We currently only capture ubatched graphs when its a FULL
5155
5156
5157
            # 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
5158
            allow_microbatching = (
5159
                self.parallel_config.use_ubatching
5160
5161
                and cudagraph_runtime_mode == CUDAGraphMode.FULL
                and uniform_decode
5162
5163
5164
5165
5166
                and check_ubatch_thresholds(
                    config=self.vllm_config.parallel_config,
                    num_tokens=num_tokens,
                    uniform_decode=uniform_decode,
                )
5167
            )
5168

5169
5170
            for _ in range(self.compilation_config.cudagraph_num_of_warmups):
                # Use CUDAGraphRuntimeStyle.NONE (default) for warmup.
5171
                # But be careful, warm up with `NONE` is orthogonal to
5172
5173
5174
                # 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.
5175
                dummy_run(
5176
5177
5178
                    num_tokens,
                    cudagraph_runtime_mode=CUDAGraphMode.NONE,
                    allow_microbatching=allow_microbatching,
5179
                    activate_lora=activate_lora,
5180
                )
5181
5182
5183

            # Capture run
            dummy_run(
5184
5185
5186
                num_tokens,
                cudagraph_runtime_mode=cudagraph_runtime_mode,
                allow_microbatching=allow_microbatching,
5187
                activate_lora=activate_lora,
Rémi Delacourt's avatar
Rémi Delacourt committed
5188
                is_graph_capturing=True,
5189
            )
5190
        self.maybe_remove_all_loras(self.lora_config)
5191

5192
5193
5194
5195
    def initialize_attn_backend(self, kv_cache_config: KVCacheConfig) -> None:
        """
        Initialize the attention backends and attention metadata builders.
        """
5196
        assert len(self.attn_groups) == 0, "Attention backends are already initialized"
5197

5198
5199
5200
5201
5202
5203
        class AttentionGroupKey(NamedTuple):
            attn_backend: type[AttentionBackend]
            kv_cache_spec: KVCacheSpec

        def get_attn_backends_for_group(
            kv_cache_group_spec: KVCacheGroupSpec,
5204
        ) -> tuple[dict[AttentionGroupKey, list[str]], set[type[AttentionBackend]]]:
5205
            layer_type = cast(type[Any], AttentionLayerBase)
5206
            layers = get_layers_from_vllm_config(
5207
                self.vllm_config, layer_type, kv_cache_group_spec.layer_names
5208
            )
5209
5210
            attn_backends = {}
            attn_backend_layers = defaultdict(list)
5211
            # Dedupe based on full class name; this is a bit safer than
5212
5213
5214
5215
            # 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.
5216
            for layer_name in kv_cache_group_spec.layer_names:
5217
                attn_backend = layers[layer_name].get_attn_backend()
5218
5219
5220
5221

                if layer_name in self.kv_sharing_fast_prefill_eligible_layers:
                    attn_backend = create_fast_prefill_custom_backend(
                        "FastPrefill",
5222
                        attn_backend,  # type: ignore[arg-type]
5223
5224
                    )

5225
5226
5227
                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):
5228
                    layer_kv_cache_spec = layer_kv_cache_spec.kv_cache_specs[layer_name]
5229
                key = (full_cls_name, layer_kv_cache_spec)
5230
5231
5232
                attn_backends[key] = AttentionGroupKey(
                    attn_backend, layer_kv_cache_spec
                )
5233
                attn_backend_layers[key].append(layer_name)
5234
5235
5236
5237
            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()),
            )
5238
5239

        def create_attn_groups(
5240
            attn_backends_map: dict[AttentionGroupKey, list[str]],
5241
            kv_cache_group_id: int,
5242
5243
        ) -> list[AttentionGroup]:
            attn_groups: list[AttentionGroup] = []
5244
            for (attn_backend, kv_cache_spec), layer_names in attn_backends_map.items():
5245
                attn_group = AttentionGroup(
5246
                    attn_backend,
5247
                    layer_names,
5248
                    kv_cache_spec,
5249
                    kv_cache_group_id,
5250
5251
                )

5252
5253
5254
                attn_groups.append(attn_group)
            return attn_groups

5255
        attention_backend_maps = []
5256
        attention_backend_list = []
5257
        for kv_cache_group_spec in kv_cache_config.kv_cache_groups:
5258
            attn_backends = get_attn_backends_for_group(kv_cache_group_spec)
5259
            attention_backend_maps.append(attn_backends[0])
5260
            attention_backend_list.append(attn_backends[1])
5261
5262

        # Resolve cudagraph_mode before actually initialize metadata_builders
5263
5264
5265
        self._check_and_update_cudagraph_mode(
            attention_backend_list, kv_cache_config.kv_cache_groups
        )
5266

5267
5268
5269
        # Check if attention backend supports PCP&DCP and related features.
        check_attention_cp_compatibility(self.vllm_config)

5270
5271
        for i, attn_backend_map in enumerate(attention_backend_maps):
            self.attn_groups.append(create_attn_groups(attn_backend_map, i))
5272

5273
5274
5275
5276
5277
5278
5279
5280
5281
5282
5283
5284
5285
5286
5287
    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
5288
5289
                    if not self.parallel_config.use_ubatching
                    else self.parallel_config.num_ubatches,
5290
                )
co63oc's avatar
co63oc committed
5291
        # Calculate reorder batch threshold (if needed)
5292
5293
        # Note (tdoublep): do this *after* constructing builders,
        # because some of them change the threshold at init time.
5294
5295
        self.calculate_reorder_batch_threshold()

5296
    def _check_and_update_cudagraph_mode(
5297
5298
5299
        self,
        attention_backends: list[set[type[AttentionBackend]]],
        kv_cache_groups: list[KVCacheGroupSpec],
5300
    ) -> None:
5301
        """
5302
        Resolve the cudagraph_mode when there are multiple attention
5303
        groups with potential conflicting CUDA graph support.
5304
5305
5306
        Then initialize the cudagraph_dispatcher based on the resolved
        cudagraph_mode.
        """
5307
        min_cg_support = AttentionCGSupport.ALWAYS
5308
        min_cg_backend_name = None
5309

5310
5311
5312
5313
5314
5315
5316
5317
5318
5319
5320
5321
        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__
5322
5323
        # Flexible resolve the cudagraph mode
        cudagraph_mode = self.compilation_config.cudagraph_mode
5324
        assert cudagraph_mode is not None
5325
        # check cudagraph for mixed batch is supported
5326
5327
5328
5329
5330
5331
        if (
            cudagraph_mode.mixed_mode() == CUDAGraphMode.FULL
            and min_cg_support != AttentionCGSupport.ALWAYS
        ):
            msg = (
                f"CUDAGraphMode.{cudagraph_mode.name} is not supported "
5332
                f"with {min_cg_backend_name} backend (support: "
5333
5334
                f"{min_cg_support})"
            )
5335
5336
            if min_cg_support == AttentionCGSupport.NEVER:
                # if not supported any full cudagraphs, just raise it.
5337
5338
                msg += (
                    "; please try cudagraph_mode=PIECEWISE, and "
5339
                    "make sure compilation mode is VLLM_COMPILE"
5340
                )
5341
5342
5343
5344
5345
                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"
5346
                cudagraph_mode = self.compilation_config.cudagraph_mode = (
5347
                    CUDAGraphMode.FULL_AND_PIECEWISE
5348
                )
5349
5350
            else:
                msg += "; setting cudagraph_mode=FULL_DECODE_ONLY"
5351
                cudagraph_mode = self.compilation_config.cudagraph_mode = (
5352
                    CUDAGraphMode.FULL_DECODE_ONLY
5353
                )
5354
5355
            logger.warning(msg)

5356
        # check that if we are doing decode full-cudagraphs it is supported
5357
5358
5359
5360
5361
5362
        if (
            cudagraph_mode.decode_mode() == CUDAGraphMode.FULL
            and min_cg_support == AttentionCGSupport.NEVER
        ):
            msg = (
                f"CUDAGraphMode.{cudagraph_mode.name} is not supported "
5363
                f"with {min_cg_backend_name} backend (support: "
5364
5365
                f"{min_cg_support})"
            )
5366
            if self.compilation_config.mode == CompilationMode.VLLM_COMPILE and (
5367
5368
5369
5370
5371
                self.compilation_config.splitting_ops_contain_attention()
                or self.compilation_config.use_inductor_graph_partition
            ):
                msg += (
                    "; setting cudagraph_mode=PIECEWISE because "
5372
                    "attention is compiled piecewise"
5373
5374
                )
                cudagraph_mode = self.compilation_config.cudagraph_mode = (
5375
                    CUDAGraphMode.PIECEWISE
5376
                )
5377
            else:
5378
5379
                msg += (
                    "; setting cudagraph_mode=NONE because "
5380
                    "attention is not compiled piecewise"
5381
5382
                )
                cudagraph_mode = self.compilation_config.cudagraph_mode = (
5383
                    CUDAGraphMode.NONE
5384
                )
5385
5386
            logger.warning(msg)

5387
5388
        # check that if we are doing spec-decode + decode full-cudagraphs it is
        # supported
5389
5390
5391
5392
5393
5394
5395
5396
        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 "
5397
                f"{min_cg_backend_name} (support: {min_cg_support})"
5398
            )
5399
5400
            if self.compilation_config.splitting_ops_contain_attention():
                msg += "; setting cudagraph_mode=PIECEWISE"
5401
                cudagraph_mode = self.compilation_config.cudagraph_mode = (
5402
                    CUDAGraphMode.PIECEWISE
5403
                )
5404
5405
            else:
                msg += "; setting cudagraph_mode=NONE"
5406
                cudagraph_mode = self.compilation_config.cudagraph_mode = (
5407
                    CUDAGraphMode.NONE
5408
                )
5409
5410
5411
5412
            logger.warning(msg)

        # double check that we can support full cudagraph if they are requested
        # even after automatic downgrades
5413
5414
5415
5416
5417
5418
        if (
            cudagraph_mode.has_full_cudagraphs()
            and min_cg_support == AttentionCGSupport.NEVER
        ):
            raise ValueError(
                f"CUDAGraphMode.{cudagraph_mode.name} is not "
5419
                f"supported with {min_cg_backend_name} backend ("
5420
5421
                f"support:{min_cg_support}) "
                "; please try cudagraph_mode=PIECEWISE, "
5422
                "and make sure compilation mode is VLLM_COMPILE"
5423
            )
5424

5425
5426
5427
5428
        # 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
5429
        # Will be removed in the near future when we have separate cudagraph capture
5430
5431
5432
5433
5434
5435
5436
5437
5438
        # 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
            )
5439
5440
5441
5442
            capture_sizes = self.compilation_config.cudagraph_capture_sizes
            self.cudagraph_batch_sizes = (
                capture_sizes if capture_sizes is not None else []
            )
5443

5444
5445
        # Trigger cudagraph dispatching keys initialization after
        # resolved cudagraph mode.
5446
        self.compilation_config.cudagraph_mode = cudagraph_mode
5447
        self.cudagraph_dispatcher.initialize_cudagraph_keys(
5448
            cudagraph_mode, self.uniform_decode_query_len
5449
        )
5450

5451
5452
5453
5454
5455
        # Initialize eagle's cudagraph dispatcher if using eagle spec decode.
        if self.speculative_config and self.speculative_config.use_eagle():
            assert isinstance(self.drafter, EagleProposer)
            self.drafter.initialize_cudagraph_keys(cudagraph_mode)

5456
5457
    def calculate_reorder_batch_threshold(self) -> None:
        """
5458
5459
5460
5461
        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.
5462
        """
5463
5464
        min_none_high = lambda a, b: a if b is None else b if a is None else min(a, b)

5465
        reorder_batch_thresholds: list[int | None] = [
5466
5467
5468
            group.get_metadata_builder().reorder_batch_threshold
            for group in self._attn_group_iterator()
        ]
5469
5470
5471
5472
5473
        # 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
5474
        self.reorder_batch_threshold = reduce(min_none_high, reorder_batch_thresholds)  # type: ignore[assignment]
5475

5476
5477
5478
    @staticmethod
    def select_common_block_size(
        kv_manager_block_size: int, attn_groups: list[AttentionGroup]
5479
5480
    ) -> int:
        """
5481
5482
5483
5484
5485
        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.
5486
5487
5488
5489
5490
5491

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

        Returns:
5492
            The selected block size
5493
5494

        Raises:
5495
            ValueError: If no valid block size found
5496
5497
        """

5498
5499
5500
5501
5502
5503
5504
5505
        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
5506
                for supported_size in backend.get_supported_kernel_block_sizes():
5507
5508
5509
5510
5511
5512
5513
5514
5515
5516
5517
5518
5519
5520
5521
5522
5523
5524
5525
5526
5527
5528
5529
5530
5531
5532
5533
5534
5535
5536
                    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
5537
            for supported_size in backend.get_supported_kernel_block_sizes()
5538
5539
            if isinstance(supported_size, int)
        )
5540

5541
5542
5543
5544
5545
5546
        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}. ")
5547

5548
5549
5550
    def may_reinitialize_input_batch(
        self, kv_cache_config: KVCacheConfig, kernel_block_sizes: list[int]
    ) -> None:
5551
5552
5553
5554
5555
5556
5557
        """
        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.
5558
            kernel_block_sizes: The kernel block sizes for each KV cache group.
5559
5560
5561
5562
        """
        block_sizes = [
            kv_cache_group.kv_cache_spec.block_size
            for kv_cache_group in kv_cache_config.kv_cache_groups
5563
            if not isinstance(kv_cache_group.kv_cache_spec, EncoderOnlyAttentionSpec)
5564
        ]
5565
5566
5567
5568
5569
5570
5571
5572
5573
5574
5575
5576
5577
5578
5579
5580
5581
5582
        max_num_blocks = []
        max_model_len = max(self.max_model_len, self.max_encoder_len)
        for i, kv_cache_group in enumerate(kv_cache_config.kv_cache_groups):
            if isinstance(kv_cache_group.kv_cache_spec, EncoderOnlyAttentionSpec):
                continue
            max_num_blocks_per_req = cdiv(
                max_model_len, block_sizes[i] * get_total_cp_world_size()
            )
            if isinstance(kv_cache_group.kv_cache_spec, MambaSpec):
                mamba_blocks_per_req = (
                    max_num_blocks_per_req
                    if self.cache_config.enable_prefix_caching
                    else 1
                ) + kv_cache_group.kv_cache_spec.num_speculative_blocks
                max_num_blocks_per_req = max(
                    max_num_blocks_per_req, mamba_blocks_per_req
                )
            max_num_blocks.append(max_num_blocks_per_req)
5583
5584
5585
5586

        if block_sizes != [self.cache_config.block_size] or kernel_block_sizes != [
            self.cache_config.block_size
        ]:
5587
5588
5589
            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
5590
5591
                "for more details."
            )
5592
5593
            self.input_batch = InputBatch(
                max_num_reqs=self.max_num_reqs,
5594
                max_model_len=max_model_len,
5595
5596
5597
5598
5599
                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,
5600
                kernel_block_sizes=kernel_block_sizes,
5601
                max_num_blocks_per_req=max_num_blocks,
5602
                is_spec_decode=bool(self.vllm_config.speculative_config),
5603
                logitsprocs=self.input_batch.logitsprocs,
5604
                logitsprocs_need_output_token_ids=self.input_batch.logitsprocs_need_output_token_ids,
5605
                is_pooling_model=self.is_pooling_model,
5606
5607
            )

5608
    def _allocate_kv_cache_tensors(
5609
5610
        self, kv_cache_config: KVCacheConfig
    ) -> dict[str, torch.Tensor]:
5611
        """
5612
5613
5614
        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.

5615
        Args:
5616
            kv_cache_config: The KV cache config
5617
        Returns:
5618
            dict[str, torch.Tensor]: A map between layer names to their
5619
            corresponding memory buffer for KV cache.
5620
        """
5621
5622
        kv_cache_raw_tensors: dict[str, torch.Tensor] = {}
        for kv_cache_tensor in kv_cache_config.kv_cache_tensors:
5623
5624
5625
            tensor = torch.zeros(
                kv_cache_tensor.size, dtype=torch.int8, device=self.device
            )
5626
5627
5628
5629
5630
            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:
5631
5632
5633
5634
            for layer_name in group.layer_names:
                if layer_name in self.runner_only_attn_layers:
                    continue
                layer_names.add(layer_name)
5635
5636
5637
        assert layer_names == set(kv_cache_raw_tensors.keys()), (
            "Some layers are not correctly initialized"
        )
5638
5639
        return kv_cache_raw_tensors

5640
5641
5642
    def _attn_group_iterator(self) -> Iterator[AttentionGroup]:
        return itertools.chain.from_iterable(self.attn_groups)

5643
    def _kv_cache_spec_attn_group_iterator(self) -> Iterator[AttentionGroup]:
5644
5645
        if not self.kv_cache_config.kv_cache_groups:
            return
5646
5647
        for attn_groups in self.attn_groups:
            yield from attn_groups
5648

5649
5650
5651
5652
5653
5654
5655
5656
5657
5658
5659
5660
5661
5662
5663
    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 = []
5664
        for kv_cache_gid, kv_cache_group in enumerate(kv_cache_config.kv_cache_groups):
5665
5666
5667
5668
5669
5670
            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):
5671
                continue
5672
            elif isinstance(kv_cache_spec, AttentionSpec):
5673
5674
5675
                # This is an attention backend that supports virtual
                # block splitting. Get the supported block sizes from
                # all backends in the group.
5676
                attn_groups = self.attn_groups[kv_cache_gid]
5677
                kv_manager_block_size = kv_cache_group.kv_cache_spec.block_size
5678
                selected_kernel_size = self.select_common_block_size(
5679
5680
5681
                    kv_manager_block_size, attn_groups
                )
                kernel_block_sizes.append(selected_kernel_size)
5682
            elif isinstance(kv_cache_spec, MambaSpec):
5683
5684
                # This is likely Mamba or other non-attention cache,
                # no splitting.
5685
                kernel_block_sizes.append(kv_cache_spec.block_size)
5686
5687
5688
5689
5690
5691
            else:
                raise NotImplementedError(
                    f"unknown kv cache spec {kv_cache_group.kv_cache_spec}"
                )
        return kernel_block_sizes

5692
5693
5694
5695
    def _reshape_kv_cache_tensors(
        self,
        kv_cache_config: KVCacheConfig,
        kv_cache_raw_tensors: dict[str, torch.Tensor],
5696
        kernel_block_sizes: list[int],
5697
    ) -> dict[str, torch.Tensor]:
5698
        """
5699
        Reshape the KV cache tensors to the desired shape and dtype.
5700

5701
        Args:
5702
5703
            kv_cache_config: The KV cache config
            kv_cache_raw_tensors: The KV cache buffer of each layer, with
5704
                correct size but uninitialized shape.
5705
            kernel_block_sizes: The kernel block sizes for each KV cache group.
5706
        Returns:
5707
            Dict[str, torch.Tensor]: A map between layer names to their
5708
5709
            corresponding memory buffer for KV cache.
        """
5710
        kv_caches: dict[str, torch.Tensor] = {}
5711
        has_attn, has_mamba = False, False
5712
5713
        for group in self._kv_cache_spec_attn_group_iterator():
            kv_cache_spec = group.kv_cache_spec
5714
            attn_backend = group.backend
5715
5716
5717
5718
            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]
5719
            for layer_name in group.layer_names:
5720
5721
                if layer_name in self.runner_only_attn_layers:
                    continue
5722
5723
                raw_tensor = kv_cache_raw_tensors[layer_name]
                assert raw_tensor.numel() % kv_cache_spec.page_size_bytes == 0
5724
                num_blocks = raw_tensor.numel() // kv_cache_spec.page_size_bytes
5725
                if isinstance(kv_cache_spec, AttentionSpec):
5726
                    has_attn = True
5727
5728
                    num_blocks_per_kv_block = (
                        kv_cache_spec.block_size // kernel_block_size
5729
5730
5731
                    )
                    kernel_num_blocks = num_blocks * num_blocks_per_kv_block

5732
                    kv_cache_shape = attn_backend.get_kv_cache_shape(
5733
                        kernel_num_blocks,
5734
                        kernel_block_size,
5735
5736
                        kv_cache_spec.num_kv_heads,
                        kv_cache_spec.head_size,
5737
5738
                        cache_dtype_str=self.cache_config.cache_dtype,
                    )
5739
                    dtype = kv_cache_spec.dtype
5740
                    try:
5741
                        kv_cache_stride_order = attn_backend.get_kv_cache_stride_order()
5742
                        assert len(kv_cache_stride_order) == len(kv_cache_shape)
5743
                    except (AttributeError, NotImplementedError):
5744
                        kv_cache_stride_order = tuple(range(len(kv_cache_shape)))
5745
5746
5747
5748
5749
                    # 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.
5750
5751
5752
                    kv_cache_shape = tuple(
                        kv_cache_shape[i] for i in kv_cache_stride_order
                    )
5753
5754
5755
5756
5757
                    # Maintain original KV shape view.
                    inv_order = [
                        kv_cache_stride_order.index(i)
                        for i in range(len(kv_cache_stride_order))
                    ]
5758
5759
5760
5761
5762
5763
                    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
5764
                elif isinstance(kv_cache_spec, MambaSpec):
5765
                    has_mamba = True
Chen Zhang's avatar
Chen Zhang committed
5766
5767
                    raw_tensor = kv_cache_raw_tensors[layer_name]
                    state_tensors = []
5768
                    storage_offset_bytes = 0
5769
                    for shape, dtype in zip(kv_cache_spec.shapes, kv_cache_spec.dtypes):
5770
5771
                        dtype_size = get_dtype_size(dtype)
                        num_element_per_page = (
5772
5773
                            kv_cache_spec.page_size_bytes // dtype_size
                        )
Chen Zhang's avatar
Chen Zhang committed
5774
                        target_shape = (num_blocks, *shape)
5775
5776
                        stride = torch.empty(target_shape).stride()
                        target_stride = (num_element_per_page, *stride[1:])
5777
                        assert storage_offset_bytes % dtype_size == 0
5778
5779
5780
5781
                        tensor = torch.as_strided(
                            raw_tensor.view(dtype),
                            size=target_shape,
                            stride=target_stride,
5782
                            storage_offset=storage_offset_bytes // dtype_size,
5783
                        )
Chen Zhang's avatar
Chen Zhang committed
5784
                        state_tensors.append(tensor)
5785
                        storage_offset_bytes += stride[0] * dtype_size
5786
5787

                    kv_caches[layer_name] = state_tensors
5788
                else:
5789
                    raise NotImplementedError
5790
5791

        if has_attn and has_mamba:
5792
            self._update_hybrid_attention_mamba_layout(kv_caches)
5793

5794
5795
        return kv_caches

5796
    def _update_hybrid_attention_mamba_layout(
5797
5798
        self, kv_caches: dict[str, torch.Tensor]
    ) -> None:
5799
        """
5800
5801
        Update the layout of attention layers from (2, num_blocks, ...) to
        (num_blocks, 2, ...).
5802
5803

        Args:
5804
            kv_caches: The KV cache buffer of each layer.
5805
5806
        """

5807
5808
        for group in self._kv_cache_spec_attn_group_iterator():
            kv_cache_spec = group.kv_cache_spec
5809
            for layer_name in group.layer_names:
5810
                kv_cache = kv_caches[layer_name]
5811
5812
5813
5814
                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 "
5815
                        f"a tensor of shape {kv_cache.shape}"
5816
                    )
5817
                    hidden_size = kv_cache.shape[2:].numel()
5818
5819
5820
5821
                    kv_cache.as_strided_(
                        size=kv_cache.shape,
                        stride=(hidden_size, 2 * hidden_size, *kv_cache.stride()[2:]),
                    )
5822

5823
    def initialize_kv_cache_tensors(
5824
        self, kv_cache_config: KVCacheConfig, kernel_block_sizes: list[int]
5825
    ) -> dict[str, torch.Tensor]:
5826
5827
5828
5829
5830
        """
        Initialize the memory buffer for KV cache.

        Args:
            kv_cache_config: The KV cache config
5831
5832
            kernel_block_sizes: The kernel block sizes for each KV cache group.

5833
        Returns:
5834
            Dict[str, torch.Tensor]: A map between layer names to their
5835
5836
            corresponding memory buffer for KV cache.
        """
5837
5838
5839
5840
5841
5842
5843
5844
5845
5846
5847
5848
5849
5850
5851
5852
5853
5854
5855
5856
5857
5858
5859
5860

        # 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
            )
5861

5862
        # Set up cross-layer KV cache sharing
5863
5864
        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)
5865
5866
            kv_caches[layer_name] = kv_caches[target_layer_name]

5867
5868
5869
5870
5871
5872
5873
5874
5875
        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,
        )
5876
5877
5878
        return kv_caches

    def maybe_add_kv_sharing_layers_to_kv_cache_groups(
5879
5880
        self, kv_cache_config: KVCacheConfig
    ) -> None:
5881
5882
5883
5884
5885
5886
5887
5888
5889
5890
5891
5892
5893
5894
5895
5896
5897
5898
        """
        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.
5899
            attn_layers = get_layers_from_vllm_config(self.vllm_config, Attention)
5900
5901
            for layer_name in reversed(attn_layers):
                if layer_name in self.shared_kv_cache_layers:
5902
                    self.kv_sharing_fast_prefill_eligible_layers.add(layer_name)
5903
5904
                else:
                    break
5905

5906
5907
5908
5909
5910
5911
5912
    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
        """
5913
        kv_cache_config = deepcopy(kv_cache_config)
5914
        self.kv_cache_config = kv_cache_config
5915
        self.may_add_encoder_only_layers_to_kv_cache_config()
5916
        self.maybe_add_kv_sharing_layers_to_kv_cache_groups(kv_cache_config)
5917
        self.initialize_attn_backend(kv_cache_config)
5918
5919
5920
5921
5922
5923
        # 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)
5924
5925
5926
5927

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

5928
        # Reinitialize need to after initialize_attn_backend
5929
5930
5931
5932
        self.may_reinitialize_input_batch(kv_cache_config, kernel_block_sizes)
        kv_caches = self.initialize_kv_cache_tensors(
            kv_cache_config, kernel_block_sizes
        )
5933

5934
5935
5936
5937
5938
        if self.speculative_config and (
            self.speculative_config.use_eagle()
            or self.speculative_config.uses_draft_model()
        ):
            assert isinstance(self.drafter, EagleProposer | DraftModelProposer)
5939
5940
5941
5942
            # 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
5943
        if has_kv_transfer_group():
5944
            kv_transfer_group = get_kv_transfer_group()
5945
5946
5947
5948
5949
5950
5951
            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)
5952
            kv_transfer_group.set_host_xfer_buffer_ops(copy_kv_blocks)
Robert Shaw's avatar
Robert Shaw committed
5953

5954
5955
5956
5957
5958
5959
5960
5961
5962
5963
5964
5965
5966
5967
5968
5969
5970
        if self.model_config.enable_return_routed_experts:
            self.init_routed_experts_capturer()

    def init_routed_experts_capturer(self):
        logger.info(
            "Initializing routed experts capturer, enable_return_routed_experts: %s",
            self.model_config.enable_return_routed_experts,
        )
        routed_experts_capturer = RoutedExpertsCapturer.create()
        block_size = self.cache_config.block_size
        self.max_num_kv_tokens = (
            self.kv_cache_config.num_blocks // len(self.kv_cache_config.kv_cache_groups)
            + 1
        ) * block_size
        routed_experts_capturer.init_buffer(
            max_num_batched_tokens=self.scheduler_config.max_num_batched_tokens,
            max_num_kv_tokens=self.max_num_kv_tokens,
5971
            vllm_config=self.vllm_config,
5972
5973
        )

5974
5975
5976
5977
5978
    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
5979
        encoder_only_attn_specs: dict[AttentionSpec, list[str]] = defaultdict(list)
5980
5981
5982
        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:
5983
                attn_spec: AttentionSpec = EncoderOnlyAttentionSpec(
5984
5985
5986
                    block_size=block_size,
                    num_kv_heads=attn_module.num_kv_heads,
                    head_size=attn_module.head_size,
5987
5988
                    dtype=self.kv_cache_dtype,
                )
5989
5990
5991
                encoder_only_attn_specs[attn_spec].append(layer_name)
                self.runner_only_attn_layers.add(layer_name)
        if len(encoder_only_attn_specs) > 0:
5992
5993
5994
            assert len(encoder_only_attn_specs) == 1, (
                "Only support one encoder-only attention spec now"
            )
5995
5996
            spec, layer_names = encoder_only_attn_specs.popitem()
            self.kv_cache_config.kv_cache_groups.append(
5997
5998
                KVCacheGroupSpec(layer_names=layer_names, kv_cache_spec=spec)
            )
5999

6000
    def get_kv_cache_spec(self) -> dict[str, KVCacheSpec]:
6001
        """
6002
        Generates the KVCacheSpec by parsing the kv cache format from each
6003
6004
        Attention module in the static forward context.
        Returns:
6005
            KVCacheSpec: A dictionary mapping layer names to their KV cache
6006
6007
            format. Layers that do not need KV cache are not included.
        """
6008
6009
        if has_ec_transfer() and get_ec_transfer().is_producer:
            return {}
6010
        kv_cache_spec: dict[str, KVCacheSpec] = {}
6011
6012
        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
6013
        for layer_name, attn_module in attn_layers.items():
6014
6015
6016
6017
6018
6019
6020
6021
6022
6023
6024
6025
6026
6027
6028
            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
6029

6030
        return kv_cache_spec
6031

6032
    def _to_list(self, sampled_token_ids: torch.Tensor) -> list[list[int]]:
6033
6034
6035
6036
6037
6038
6039
6040
        # 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.
6041
        pinned = self.sampled_token_ids_pinned_cpu[: sampled_token_ids.shape[0]]
6042
6043
6044
        pinned.copy_(sampled_token_ids, non_blocking=True)
        self.transfer_event.record()
        self.transfer_event.synchronize()
6045
        return pinned.tolist()
6046
6047
6048
6049
6050
6051
6052
6053
6054
6055
6056
6057
6058
6059
6060
6061
6062
6063
6064
6065
6066
6067
6068
6069
6070
6071
6072
6073
6074
6075
6076
6077
6078
6079
6080
6081
6082
6083
6084
6085
6086
6087
6088
6089
6090
6091
6092
6093
6094
6095
6096
6097
6098
6099
6100
6101
6102
6103
6104
6105
6106
6107
6108
6109
6110
6111
6112
6113
6114
6115
6116
6117
6118
6119
6120
6121

    def get_encoder_timing_stats(self) -> dict[str, dict[str, float | int]]:
        """
        Get encoder timing stats for all requests and clear the registry.

        Returns:
            Dictionary mapping request_id to stats dict.
        """
        with self._encoder_timing_lock:
            stats = {
                req_id: stats_obj.to_dict()
                for req_id, stats_obj in self.encoder_timing_registry.items()
            }
            self.encoder_timing_registry.clear()
            return stats

    @contextmanager
    def timed_encoder_operation(
        self,
        should_time: bool,
        group_lora_refs: list[tuple[str, Any]],
        current_item_idx: int,
        num_items: int,
    ):
        """
        Context manager to time encoder forward operations.

        Args:
            should_time: Whether timing is enabled
            group_lora_refs: Full list of (request_id, pos_info) tuples
            current_item_idx: Starting index for this group
            num_items: Number of items in this group
        """
        if not should_time:
            yield
            return

        group_refs = group_lora_refs[current_item_idx : current_item_idx + num_items]
        group_request_ids = {req_id for req_id, _ in group_refs}

        torch.cuda.synchronize()
        start_time = time.perf_counter()

        try:
            yield
        finally:
            torch.cuda.synchronize()
            elapsed = time.perf_counter() - start_time

            per_request_time = elapsed / max(len(group_request_ids), 1)

            with self._encoder_timing_lock:
                for req_id in group_request_ids:
                    if req_id not in self.encoder_timing_registry:
                        self.encoder_timing_registry[req_id] = EncoderTimingStats()

                    stats = self.encoder_timing_registry[req_id]
                    stats.encoder_forward_time += per_request_time
                    stats.num_encoder_calls += 1


@dataclass
class EncoderTimingStats:
    """Per-request timing statistics for encoder forward pass."""

    encoder_forward_time: float = 0.0
    """Time spent in vision encoder forward pass (seconds)."""

    num_encoder_calls: int = 0
    """Number of times encoder was called for this request."""

    def to_dict(self) -> dict[str, float | int]:
        return {
            "encoder_forward_time": self.encoder_forward_time,
            "num_encoder_calls": self.num_encoder_calls,
        }