gpu_model_runner.py 231 KB
Newer Older
1
# SPDX-License-Identifier: Apache-2.0
2
# SPDX-FileCopyrightText: Copyright contributors to the vLLM project
3

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

import numpy as np
import torch
import torch.distributed
import torch.nn as nn
19
from tqdm import tqdm
20

21
import vllm.envs as envs
22
23
24
from vllm.attention.backends.abstract import (
    AttentionBackend,
    AttentionMetadata,
25
    AttentionType,
26
27
    MultipleOf,
)
28
from vllm.attention.layer import Attention, MLAAttention
29
from vllm.compilation.counter import compilation_counter
30
from vllm.compilation.cuda_graph import CUDAGraphStat, CUDAGraphWrapper
31
from vllm.compilation.monitor import set_cudagraph_capturing_enabled
32
from vllm.config import (
33
    CompilationMode,
34
35
36
37
38
    CUDAGraphMode,
    VllmConfig,
    get_layers_from_vllm_config,
    update_config,
)
39
from vllm.distributed.ec_transfer import get_ec_transfer, has_ec_transfer
40
from vllm.distributed.eplb.eplb_state import EplbState
41
from vllm.distributed.kv_transfer import get_kv_transfer_group, has_kv_transfer_group
42
from vllm.distributed.kv_transfer.kv_connector.utils import copy_kv_blocks
43
from vllm.distributed.parallel_state import (
44
    get_dcp_group,
45
46
47
48
49
50
    get_pp_group,
    get_tp_group,
    graph_capture,
    is_global_first_rank,
    prepare_communication_buffer_for_model,
)
51
from vllm.forward_context import BatchDescriptor, set_forward_context
52
from vllm.logger import init_logger
53
from vllm.model_executor.layers.attention_layer_base import AttentionLayerBase
54
55
56
57
from vllm.model_executor.layers.rotary_embedding import (
    MRotaryEmbedding,
    XDRotaryEmbedding,
)
58
from vllm.model_executor.model_loader import TensorizerLoader, get_model_loader
59
from vllm.model_executor.models.interfaces import (
60
    SupportsMRoPE,
61
    SupportsMultiModal,
62
    SupportsXDRoPE,
63
64
65
66
67
    is_mixture_of_experts,
    supports_eagle3,
    supports_mrope,
    supports_multimodal_pruning,
    supports_transcription,
68
    supports_xdrope,
69
)
70
from vllm.model_executor.models.interfaces_base import (
71
72
73
74
    VllmModelForPooling,
    is_pooling_model,
    is_text_generation_model,
)
75
from vllm.multimodal import MULTIMODAL_REGISTRY
76
77
78
79
80
from vllm.multimodal.inputs import (
    BatchedTensorInputs,
    MultiModalKwargsItem,
    PlaceholderRange,
)
81
from vllm.multimodal.utils import group_mm_kwargs_by_modality
82
from vllm.pooling_params import PoolingParams
83
from vllm.sampling_params import SamplingType
84
from vllm.sequence import IntermediateTensors
85
from vllm.tasks import GenerationTask, PoolingTask, SupportedTask
86
from vllm.utils import length_from_prompt_token_ids_or_embeds
87
from vllm.utils.jsontree import json_map_leaves
88
from vllm.utils.math_utils import cdiv, round_up
89
90
from vllm.utils.mem_constants import GiB_bytes
from vllm.utils.mem_utils import DeviceMemoryProfiler
91
from vllm.utils.platform_utils import is_pin_memory_available
92
93
94
95
96
from vllm.utils.torch_utils import (
    get_dtype_size,
    kv_cache_dtype_str_to_dtype,
    supports_dynamo,
)
97
from vllm.v1.attention.backends.gdn_attn import GDNAttentionMetadataBuilder
98
from vllm.v1.attention.backends.utils import (
99
100
101
    AttentionCGSupport,
    AttentionMetadataBuilder,
    CommonAttentionMetadata,
102
    create_fast_prefill_custom_backend,
103
    get_dcp_local_seq_lens,
104
105
106
    reorder_batch_to_split_decodes_and_prefills,
    split_attn_metadata,
)
107
from vllm.v1.cudagraph_dispatcher import CudagraphDispatcher
108
109
110
111
112
113
114
115
116
117
118
119
120
121
122
123
124
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,
125
    ECConnectorOutput,
126
    KVConnectorOutput,
127
128
129
130
131
    LogprobsLists,
    LogprobsTensors,
    ModelRunnerOutput,
    PoolerOutput,
    SamplerOutput,
132
    make_empty_encoder_model_runner_output,
133
)
134
from vllm.v1.pool.metadata import PoolingMetadata, PoolingStates
135
from vllm.v1.sample.logits_processor import LogitsProcessors, build_logitsprocs
136
from vllm.v1.sample.logits_processor.interface import LogitsProcessor
137
from vllm.v1.sample.metadata import SamplingMetadata
138
from vllm.v1.sample.rejection_sampler import RejectionSampler
139
from vllm.v1.sample.sampler import Sampler
140
from vllm.v1.spec_decode.eagle import EagleProposer
141
from vllm.v1.spec_decode.medusa import MedusaProposer
142
from vllm.v1.spec_decode.metadata import SpecDecodeMetadata
143
from vllm.v1.spec_decode.ngram_proposer import NgramProposer
144
from vllm.v1.spec_decode.suffix_decoding import SuffixDecodingProposer
145
from vllm.v1.structured_output.utils import apply_grammar_bitmask
146
from vllm.v1.utils import CpuGpuBuffer, record_function_or_nullcontext
147
from vllm.v1.worker.dp_utils import coordinate_batch_across_dp
148
from vllm.v1.worker.ec_connector_model_runner_mixin import ECConnectorModelRunnerMixin
149
from vllm.v1.worker.gpu_input_batch import CachedRequestState, InputBatch
150
from vllm.v1.worker.gpu_ubatch_wrapper import UBatchWrapper
151
from vllm.v1.worker.kv_connector_model_runner_mixin import KVConnectorModelRunnerMixin
152
from vllm.v1.worker.lora_model_runner_mixin import LoRAModelRunnerMixin
153
154
155
156
from vllm.v1.worker.ubatch_utils import (
    UBatchSlices,
    check_ubatch_thresholds,
)
157
from vllm.v1.worker.utils import is_residual_scattered_for_sp
158

159
160
161
162
163
164
165
166
167
from .utils import (
    AttentionGroup,
    MultiModalBudget,
    add_kv_sharing_layers_to_kv_cache_groups,
    bind_kv_cache,
    gather_mm_placeholders,
    sanity_check_mm_encoder_outputs,
    scatter_mm_placeholders,
)
168

169
if TYPE_CHECKING:
170
    from vllm.model_executor.model_loader.tensorizer import TensorizerConfig
171
    from vllm.v1.core.sched.output import GrammarOutput, SchedulerOutput
172
173
174

logger = init_logger(__name__)

175
176
AttnMetadataDict: TypeAlias = dict[str, AttentionMetadata]
# list when ubatching is enabled
177
PerLayerAttnMetadata: TypeAlias = list[AttnMetadataDict] | AttnMetadataDict
178

179

180
181
182
183
184
185
# Wrapper for ModelRunnerOutput to support overlapped execution.
class AsyncGPUModelRunnerOutput(AsyncModelRunnerOutput):
    def __init__(
        self,
        model_runner_output: ModelRunnerOutput,
        sampled_token_ids: torch.Tensor,
186
        logprobs_tensors: LogprobsTensors | None,
187
188
        invalid_req_indices: list[int],
        async_output_copy_stream: torch.cuda.Stream,
189
        vocab_size: int,
190
191
192
193
194
    ):
        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.
195
        self.async_copy_ready_event = torch.Event()
196
197
198
199

        # 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
200
        self.vocab_size = vocab_size
201
        self._logprobs_tensors = logprobs_tensors
202
203
204
205
206

        # 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)
207
            self.sampled_token_ids_cpu = self._sampled_token_ids.to(
208
209
                "cpu", non_blocking=True
            )
210
211
212
213
214
            self._logprobs_tensors_cpu = (
                self._logprobs_tensors.to_cpu_nonblocking()
                if self._logprobs_tensors
                else None
            )
215
            self.async_copy_ready_event.record()
216
217
218

    def get_output(self) -> ModelRunnerOutput:
        """Copy the device tensors to the host and return a ModelRunnerOutput.
219

220
221
        This function blocks until the copy is finished.
        """
222
        max_gen_len = self.sampled_token_ids_cpu.shape[-1]
223
        self.async_copy_ready_event.synchronize()
224

225
226
        # Release the device tensors once the copy has completed.
        del self._logprobs_tensors
227
        del self._sampled_token_ids
228
        if max_gen_len == 1:
229
            valid_sampled_token_ids = self.sampled_token_ids_cpu.tolist()
230
231
232
            for i in self._invalid_req_indices:
                valid_sampled_token_ids[i].clear()
            cu_num_tokens = None
233
        else:
234
            valid_sampled_token_ids, cu_num_tokens = RejectionSampler.parse_output(
235
236
                self.sampled_token_ids_cpu,
                self.vocab_size,
237
238
                self._invalid_req_indices,
                return_cu_num_tokens=self._logprobs_tensors_cpu is not None,
239
            )
240
241
242

        output = self._model_runner_output
        output.sampled_token_ids = valid_sampled_token_ids
243
        if self._logprobs_tensors_cpu:
244
            output.logprobs = self._logprobs_tensors_cpu.tolists(cu_num_tokens)
245
246
247
        return output


248
249
250
251
252
253
254
255
256
257
258
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
259
    ec_connector_output: ECConnectorOutput | None
260
    cudagraph_stats: CUDAGraphStat | None
261
262


263
264
265
class GPUModelRunner(
    LoRAModelRunnerMixin, KVConnectorModelRunnerMixin, ECConnectorModelRunnerMixin
):
266
267
    def __init__(
        self,
268
        vllm_config: VllmConfig,
269
        device: torch.device,
270
    ):
271
272
273
        self.vllm_config = vllm_config
        self.model_config = vllm_config.model_config
        self.cache_config = vllm_config.cache_config
274
        self.compilation_config = vllm_config.compilation_config
275
276
277
278
279
280
        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
281

282
        from vllm.model_executor.models.utils import set_cpu_offload_max_bytes
283
284

        set_cpu_offload_max_bytes(int(self.cache_config.cpu_offload_gb * 1024**3))
285

286
287
288
289
        model_config = self.model_config
        cache_config = self.cache_config
        scheduler_config = self.scheduler_config
        parallel_config = self.parallel_config
290
        self.device = device
291
292
        self.pin_memory = is_pin_memory_available()
        self.dtype = self.model_config.dtype
293
294
295
        self.kv_cache_dtype = kv_cache_dtype_str_to_dtype(
            cache_config.cache_dtype, self.model_config
        )
296

297
        self.is_pooling_model = model_config.runner_type == "pooling"
298
        self.enable_prompt_embeds = model_config.enable_prompt_embeds
299
        self.is_multimodal_raw_input_only_model = (
300
301
            model_config.is_multimodal_raw_input_only_model
        )
302
303
        # This will be overridden in load_model()
        self.is_multimodal_pruning_enabled = False
304
        self.max_model_len = model_config.max_model_len
305
306
307

        # Always set to false after the first forward pass
        self.calculate_kv_scales = self.cache_config.calculate_kv_scales
308
        self.dcp_world_size = self.parallel_config.decode_context_parallel_size
309
        self.dcp_rank = 0 if self.dcp_world_size <= 1 else get_dcp_group().rank_in_group
310
        self.max_num_tokens = scheduler_config.max_num_batched_tokens
311
        self.max_num_reqs = scheduler_config.max_num_seqs
312

313
314
315
316
317
        # 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 = (
318
319
320
            self.parallel_config.distributed_executor_backend == "external_launcher"
            and len(get_pp_group().ranks) > 0
        )
321

322
        # Model-related.
323
        self.num_query_heads = model_config.get_num_attention_heads(parallel_config)
324
        self.inputs_embeds_size = model_config.get_inputs_embeds_size()
325
        self.attention_chunk_size = model_config.attention_chunk_size
326
        # Only relevant for models using ALiBi (e.g, MPT)
327
        self.use_alibi = model_config.uses_alibi
328

329
        self.cascade_attn_enabled = not self.model_config.disable_cascade_attn
330

331
        # Multi-modal data support
332
        self.mm_registry = MULTIMODAL_REGISTRY
333
        self.uses_mrope = model_config.uses_mrope
334
        self.uses_xdrope_dim = model_config.uses_xdrope_dim
335
        self.supports_mm_inputs = self.mm_registry.supports_multimodal_inputs(
336
337
            model_config
        )
338

339
340
341
        if self.model_config.is_encoder_decoder:
            # Maximum length of the encoder input, only for encoder-decoder
            # models.
342
            self.max_encoder_len = scheduler_config.max_num_encoder_input_tokens
343
344
345
        else:
            self.max_encoder_len = 0

346
        # Sampler
347
        self.sampler = Sampler(logprobs_mode=self.model_config.logprobs_mode)
348

349
        self.eplb_state: EplbState | None = None
350
351
352
353
354
355
        """
        State of the expert parallelism load balancer.

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

356
        # Lazy initializations
357
        # self.model: nn.Module  # Set after load_model
358
        # Initialize in initialize_kv_cache
359
        self.kv_caches: list[torch.Tensor] = []
360
361
362
        # Initialize in initialize_kv_cache_tensors
        self.cross_layers_kv_cache: torch.Tensor | None = None
        self.cross_layers_attn_backend: type[AttentionBackend] | None = None
363
364
        # indexes: [kv_cache_group_id][attn_group]
        self.attn_groups: list[list[AttentionGroup]] = []
365
366
        # self.kv_cache_config: KVCacheConfig

367
368
        # mm_hash ->  encoder_output
        self.encoder_cache: dict[str, torch.Tensor] = {}
369

370
        self.use_aux_hidden_state_outputs = False
371
372
373
374
375
        # 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:
376
377
378
            self.drafter: (
                NgramProposer | SuffixDecodingProposer | EagleProposer | MedusaProposer
            )
379
380
            if self.speculative_config.method == "ngram":
                self.drafter = NgramProposer(self.vllm_config)
381
382
            elif self.speculative_config.method == "suffix":
                self.drafter = SuffixDecodingProposer(self.vllm_config)
383
            elif self.speculative_config.use_eagle():
384
                self.drafter = EagleProposer(self.vllm_config, self.device, self)
385
                if self.speculative_config.method == "eagle3":
386
387
388
                    self.use_aux_hidden_state_outputs = (
                        self.drafter.eagle3_use_aux_hidden_state
                    )
389
390
            elif self.speculative_config.method == "medusa":
                self.drafter = MedusaProposer(
391
                    vllm_config=self.vllm_config, device=self.device
392
                )
393
            else:
394
395
396
397
                raise ValueError(
                    "Unknown speculative decoding method: "
                    f"{self.speculative_config.method}"
                )
398
            self.rejection_sampler = RejectionSampler(self.sampler)
399

400
401
402
403
        self.num_spec_tokens = 0
        if self.speculative_config:
            self.num_spec_tokens = self.speculative_config.num_speculative_tokens

404
        # Request states.
405
        self.requests: dict[str, CachedRequestState] = {}
406
407
408
        # NOTE(rob): num_prompt_logprobs only includes reqs
        # that are currently in the prefill phase.
        self.num_prompt_logprobs: dict[str, int] = {}
409
        self.comm_stream = torch.cuda.Stream()
410

411
412
413
414
415
416
417
418
419
        # 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.
420
421
422
423
        logits_processors = model_config.logits_processors
        custom_logitsprocs: Sequence[str | type[LogitsProcessor]] = (
            tuple(logits_processors) if logits_processors is not None else ()
        )
424
425
        self.input_batch = InputBatch(
            max_num_reqs=self.max_num_reqs,
426
427
428
            # 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),
429
430
431
            max_num_batched_tokens=self.max_num_tokens,
            device=self.device,
            pin_memory=self.pin_memory,
432
            vocab_size=self.model_config.get_vocab_size(),
433
            block_sizes=[self.cache_config.block_size],
434
            kernel_block_sizes=[self.cache_config.block_size],
435
            is_spec_decode=bool(self.vllm_config.speculative_config),
436
            logitsprocs=build_logitsprocs(
437
438
439
                self.vllm_config,
                self.device,
                self.pin_memory,
440
                self.is_pooling_model,
441
                custom_logitsprocs,
442
            ),
443
444
445
            # 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),
446
            is_pooling_model=self.is_pooling_model,
447
            cp_kv_cache_interleave_size=self.parallel_config.cp_kv_cache_interleave_size,
448
        )
449

450
        self.use_async_scheduling = self.scheduler_config.async_scheduling
451
452
453
454
455
        # 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.
456
        self.prepare_inputs_event: torch.Event | None = None
457
458
        if self.use_async_scheduling:
            self.async_output_copy_stream = torch.cuda.Stream()
459
            self.prepare_inputs_event = torch.Event()
460

461
        # self.cudagraph_batch_sizes sorts in ascending order.
462
463
464
465
        if (
            self.compilation_config.cudagraph_capture_sizes
            and self.compilation_config.cudagraph_mode != CUDAGraphMode.NONE
        ):
466
467
            self.cudagraph_batch_sizes = sorted(
                self.compilation_config.cudagraph_capture_sizes
468
            )
469

470
        # Cache the device properties.
471
        self._init_device_properties()
472

473
        # Persistent buffers for CUDA graphs.
474
475
476
477
478
        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
        )
479
        self.seq_lens = self._make_buffer(self.max_num_reqs, dtype=torch.int32)
480
        self.encoder_seq_lens = self._make_buffer(self.max_num_reqs, dtype=torch.int32)
481
482
483
484
        if self.dcp_world_size > 1:
            self.dcp_local_seq_lens = self._make_buffer(
                self.max_num_reqs, dtype=torch.int32
            )
485
486
487
        # 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.
488
        self.inputs_embeds = self._make_buffer(
489
            self.max_num_tokens, self.inputs_embeds_size, dtype=self.dtype, numpy=False
490
491
        )
        self.is_token_ids = self._make_buffer(self.max_num_tokens, dtype=torch.bool)
492
493
        self.discard_request_mask = self._make_buffer(
            self.max_num_reqs, dtype=torch.bool
494
495
496
497
498
499
500
        )
        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
        )
501

502
503
        # Only relevant for multimodal models
        if self.supports_mm_inputs:
504
            self.is_mm_embed = self._make_buffer(self.max_num_tokens, dtype=torch.bool)
505

506
        # Only relevant for models using M-RoPE (e.g, Qwen2-VL)
507
        if self.uses_mrope:
Roger Wang's avatar
Roger Wang committed
508
509
510
511
            # 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
512
513
514
515
516
517

            # 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
518
            self.mrope_positions = self._make_buffer(
519
520
                (3, self.max_num_tokens + 1), dtype=torch.int64
            )
521

522
523
524
525
526
527
528
        # 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
            )

529
        # None in the first PP rank. The rest are set after load_model.
530
        self.intermediate_tensors: IntermediateTensors | None = None
531

532
        # OPTIMIZATION: Cache the tensors rather than creating them every step.
533
        # Keep in int64 to avoid overflow with long context
534
535
536
537
        self.arange_np = np.arange(
            max(self.max_num_reqs + 1, self.max_model_len, self.max_num_tokens),
            dtype=np.int64,
        )
538

539
540
541
542
543
        # 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] = {}
544
545
546
547
548
        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(
549
550
                self.max_num_tokens, dtype=torch.int32, device=self.device
            )
551

552
        self.uniform_decode_query_len = 1 + self.num_spec_tokens
553
554
555
556

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

557
558
559
560
561
562
563
564
565
        self.mm_budget = (
            MultiModalBudget(
                self.model_config,
                self.scheduler_config,
                self.mm_registry,
            )
            if self.supports_mm_inputs
            else None
        )
566

567
        self.reorder_batch_threshold: int | None = None
568

569
570
571
572
573
        # 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()

574
        # Cached outputs.
575
        self._draft_token_ids: list[list[int]] | torch.Tensor | None = None
576
        self.transfer_event = torch.Event()
577
        self.sampled_token_ids_pinned_cpu = torch.empty(
578
            (self.max_num_reqs, 1),
579
580
            dtype=torch.int64,
            device="cpu",
581
582
            pin_memory=self.pin_memory,
        )
583

584
585
        # Pre-allocated tensor for copying valid sampled token counts to CPU,
        # with dedicated stream for overlapping and event for coordination.
586
        self.valid_sampled_token_count_event: torch.Event | None = None
587
588
        self.valid_sampled_token_count_copy_stream: torch.cuda.Stream | None = None
        if self.use_async_scheduling and self.num_spec_tokens:
589
            self.valid_sampled_token_count_event = torch.Event()
590
591
592
593
594
595
596
597
            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,
        )

598
599
        # Ephemeral state transferred between execute_model() and sample_tokens().
        self.execute_model_state: ExecuteModelState | None = None
600
        self.kv_connector_output: KVConnectorOutput | None = None
601

602
603
604
605
    def reset_mm_cache(self) -> None:
        if self.mm_budget:
            self.mm_budget.reset_cache()

606
607
608
609
610
611
612
613
614
615
616
617
618
619
620
621
622
623
624
625
626
627
628
629
630
631
632
633
634
635
636
637
638
639
640
641
642
643
644
645
646
647
648
649
    @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)

650
651
652
653
    def _get_positions(self, num_tokens: Any):
        if isinstance(num_tokens, int):
            if self.uses_mrope:
                return self.mrope_positions.gpu[:, :num_tokens]
654
655
            if self.uses_xdrope_dim > 0:
                return self.xdrope_positions.gpu[:, :num_tokens]
656
657
658
659
            return self.positions.gpu[:num_tokens]
        else:
            if self.uses_mrope:
                return self.mrope_positions.gpu[:, num_tokens]
660
661
            if self.uses_xdrope_dim > 0:
                return self.xdrope_positions.gpu[:, num_tokens]
662
663
            return self.positions.gpu[num_tokens]

664
    def _make_buffer(
665
        self, *size: int | torch.SymInt, dtype: torch.dtype, numpy: bool = True
666
667
668
669
670
671
672
673
    ) -> CpuGpuBuffer:
        return CpuGpuBuffer(
            *size,
            dtype=dtype,
            device=self.device,
            pin_memory=self.pin_memory,
            with_numpy=numpy,
        )
674

675
676
677
    def _init_model_kwargs(self, num_tokens: int):
        model_kwargs = dict[str, Any]()

678
        if not self.is_pooling_model:
679
680
            return model_kwargs

681
682
        num_reqs = self.input_batch.num_reqs
        pooling_params = self.input_batch.get_pooling_params()
683
684
685

        token_type_id_requests = dict[int, Any]()
        for i, param in enumerate(pooling_params):
686
687
688
689
690
            if (
                param.extra_kwargs is not None
                and (token_types := param.extra_kwargs.get("compressed_token_type_ids"))
                is not None
            ):
691
692
693
694
695
                token_type_id_requests[i] = token_types

        if len(token_type_id_requests) == 0:
            return model_kwargs

696
        seq_lens = self.seq_lens.gpu[:num_reqs]
697
698
699
700
701
702
703
704
        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(
705
706
            device=self.device
        )
707
708
        return model_kwargs

709
    def _may_reorder_batch(self, scheduler_output: "SchedulerOutput") -> None:
710
711
        """
        Update the order of requests in the batch based on the attention
712
        backend's needs. For example, some attention backends (namely MLA) may
713
714
715
716
717
718
        want to separate requests based on if the attention computation will be
        compute-bound or memory-bound.

        Args:
            scheduler_output: The scheduler output.
        """
719
720
721
722
723
724
725
726
        # 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

727
728
729
730
        if self.reorder_batch_threshold is not None:
            reorder_batch_to_split_decodes_and_prefills(
                self.input_batch,
                scheduler_output,
731
732
                decode_threshold=self.reorder_batch_threshold,
            )
733

734
735
    # Note: used for model runner override.
    def _init_device_properties(self) -> None:
736
        """Initialize attributes from torch.cuda.get_device_properties"""
737
738
739
740
741
742
743
        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()

744
    def _update_states(self, scheduler_output: "SchedulerOutput") -> None:
745
746
747
748
749
750
        """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.

751
752
        The SamplingMetadata is updated and copied to the GPU if there is a
        new/resumed/paused/finished request in the batch.
753
754
        """
        # Remove finished requests from the cached states.
755
756
        for req_id in scheduler_output.finished_req_ids:
            self.requests.pop(req_id, None)
757
            self.num_prompt_logprobs.pop(req_id, None)
758
759
760
761
762
763
764
        # 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:
765
            self.input_batch.remove_request(req_id)
766
767

        # Free the cached encoder outputs.
768
769
        for mm_hash in scheduler_output.free_encoder_mm_hashes:
            self.encoder_cache.pop(mm_hash, None)
770

771
772
773
774
775
776
777
        # 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()
778
779
780
781
782
783
784
785
        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)
786
787
788
789
790
        # 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:
791
            self.input_batch.remove_request(req_id)
792

793
        reqs_to_add: list[CachedRequestState] = []
794
        # Add new requests to the cached states.
795
796
797
        for new_req_data in scheduler_output.scheduled_new_reqs:
            req_id = new_req_data.req_id
            sampling_params = new_req_data.sampling_params
798
            pooling_params = new_req_data.pooling_params
799

800
801
802
803
            if (
                sampling_params
                and sampling_params.sampling_type == SamplingType.RANDOM_SEED
            ):
804
805
806
807
808
                generator = torch.Generator(device=self.device)
                generator.manual_seed(sampling_params.seed)
            else:
                generator = None

809
810
            if self.is_pooling_model:
                assert pooling_params is not None
811
812
                task = pooling_params.task
                assert task is not None, "You did not set `task` in the API"
813

814
                model = cast(VllmModelForPooling, self.get_model())
815
                to_update = model.pooler.get_pooling_updates(task)
816
817
                to_update.apply(pooling_params)

818
            req_state = CachedRequestState(
819
                req_id=req_id,
820
                prompt_token_ids=new_req_data.prompt_token_ids,
821
                prompt_embeds=new_req_data.prompt_embeds,
822
                mm_features=new_req_data.mm_features,
823
                sampling_params=sampling_params,
824
                pooling_params=pooling_params,
825
                generator=generator,
826
827
                block_ids=new_req_data.block_ids,
                num_computed_tokens=new_req_data.num_computed_tokens,
828
                output_token_ids=[],
829
                lora_request=new_req_data.lora_request,
830
            )
831
832
            self.requests[req_id] = req_state

833
834
835
836
837
838
839
            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
                )

840
            # Only relevant for models using M-RoPE (e.g, Qwen2-VL)
841
            if self.uses_mrope:
842
                self._init_mrope_positions(req_state)
843

844
845
846
847
            # Only relevant for models using XD-RoPE (e.g, HunYuan-VL)
            if self.uses_xdrope_dim > 0:
                self._init_xdrope_positions(req_state)

848
            reqs_to_add.append(req_state)
849

850
        # Update the states of the running/resumed requests.
851
        is_last_rank = get_pp_group().is_last_rank
852
        req_data = scheduler_output.scheduled_cached_reqs
853
854
855
856
857

        # 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()

858
        for i, req_id in enumerate(req_data.req_ids):
859
            req_state = self.requests[req_id]
860
861
            num_computed_tokens = req_data.num_computed_tokens[i]
            new_block_ids = req_data.new_block_ids[i]
862
            resumed_from_preemption = req_id in req_data.resumed_req_ids
863
            num_output_tokens = req_data.num_output_tokens[i]
864
            req_index = self.input_batch.req_id_to_index.get(req_id)
865

866
867
868
869
870
871
872
873
874
875
876
877
878
879
880
881
882
883
884
885
886
887
888
            # prev_num_draft_len is used in async scheduling mode with
            # spec decode. it indicates if need to update num_computed_tokens
            # of the request. for example:
            # fist step: num_computed_tokens = 0, spec_tokens = [],
            # prev_num_draft_len = 0.
            # second step: num_computed_tokens = 100(prompt lenth),
            # spec_tokens = [a,b], prev_num_draft_len = 0.
            # third step: num_computed_tokens = 100 + 2, spec_tokens = [c,d],
            # prev_num_draft_len = 2.
            # num_computed_tokens in first step and second step does't contain
            # the spec tokens length, but in third step it contains the
            # spec tokens length. we only need to update num_computed_tokens
            # when prev_num_draft_len > 0.
            if req_state.prev_num_draft_len:
                if req_index is None:
                    req_state.prev_num_draft_len = 0
                else:
                    assert self.input_batch.prev_req_id_to_index is not None
                    prev_req_index = self.input_batch.prev_req_id_to_index[req_id]
                    num_accepted = valid_sampled_token_count[prev_req_index] - 1
                    num_rejected = req_state.prev_num_draft_len - num_accepted
                    num_computed_tokens -= num_rejected
                    req_state.output_token_ids.extend([-1] * num_accepted)
889

890
            # Update the cached states.
891
            req_state.num_computed_tokens = num_computed_tokens
892
893
894
895
896
897
898
899

            if not is_last_rank:
                # When using PP, the scheduler sends the sampled tokens back,
                # because there's no direct communication between the first-
                # stage worker and the last-stage worker.
                new_token_ids = req_data.new_token_ids[i]
                # Add the sampled token(s) from the previous step (if any).
                # This doesn't include "unverified" tokens like spec tokens.
900
901
902
                num_new_tokens = (
                    num_computed_tokens + len(new_token_ids) - req_state.num_tokens
                )
903
904
905
906
                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:
907
                    req_state.output_token_ids.extend(new_token_ids[-num_new_tokens:])
908
909
910
911
912
            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:
913
914
915
916
                    end_idx = (
                        self.input_batch.num_prompt_tokens[req_index]
                        + num_output_tokens
                    )
917
918
                    self.input_batch.num_tokens[req_index] = end_idx
                    self.input_batch.num_tokens_no_spec[req_index] = end_idx
919

920
            # Update the block IDs.
921
            if not resumed_from_preemption:
922
923
                if new_block_ids is not None:
                    # Append the new blocks to the existing block IDs.
924
                    for block_ids, new_ids in zip(req_state.block_ids, new_block_ids):
925
                        block_ids.extend(new_ids)
926
            else:
927
                assert req_index is None
928
                assert new_block_ids is not None
929
930
                # The request is resumed from preemption.
                # Replace the existing block IDs with the new ones.
931
                req_state.block_ids = new_block_ids
932
933
934
935
936

            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.
937
938
939
940
941
942
943

                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:]

944
                reqs_to_add.append(req_state)
945
946
947
                continue

            # Update the persistent batch.
948
            self.input_batch.num_computed_tokens_cpu[req_index] = num_computed_tokens
949
            if new_block_ids is not None:
950
                self.input_batch.block_table.append_row(new_block_ids, req_index)
951
952
953
954
955
956
957

            # 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)
958
                self.input_batch.token_ids_cpu[
959
960
961
                    req_index, start_token_index:end_token_index
                ] = new_token_ids
                self.input_batch.num_tokens_no_spec[req_index] = end_token_index
962
                self.input_batch.num_tokens[req_index] = end_token_index
963

964
            # Add spec_token_ids to token_ids_cpu.
965
            spec_token_ids = scheduler_output.scheduled_spec_decode_tokens.get(
966
                req_id, []
967
            )
968
969
970
971
972
            num_spec_tokens = len(spec_token_ids)
            # For async scheduling, token_ids_cpu assigned from
            # spec_token_ids are placeholders and will be overwritten in
            # _prepare_input_ids.
            if num_spec_tokens:
973
974
975
                start_index = self.input_batch.num_tokens_no_spec[req_index]
                end_token_index = start_index + num_spec_tokens
                self.input_batch.token_ids_cpu[
976
977
                    req_index, start_index:end_token_index
                ] = spec_token_ids
978
979
                # NOTE(woosuk): `num_tokens` here may include spec tokens.
                self.input_batch.num_tokens[req_index] += num_spec_tokens
980
981
982
983
984
985

            # When speculative decoding is used with structured output,
            # the scheduler can drop draft tokens that do not
            # conform to the schema. This can result in
            # scheduler_output.scheduled_spec_decode_tokens being empty,
            # even when speculative decoding is enabled.
986
987
            self.input_batch.spec_token_ids[req_index].clear()
            self.input_batch.spec_token_ids[req_index].extend(spec_token_ids)
988

989
990
991
992
993
994
995
996
997
            # there are no draft tokens with async scheduling,
            # we clear the spec_decoding info in scheduler_output and
            # use normal sampling but rejection_sampling.
            if self.use_async_scheduling:
                req_state.prev_num_draft_len = num_spec_tokens
                if num_spec_tokens and self._draft_token_ids is None:
                    scheduler_output.total_num_scheduled_tokens -= num_spec_tokens
                    scheduler_output.num_scheduled_tokens[req_id] -= num_spec_tokens
                    scheduler_output.scheduled_spec_decode_tokens.pop(req_id, None)
998
999
        # Add the new or resumed requests to the persistent batch.
        # The smaller empty indices are filled first.
1000
1001
        for request in reqs_to_add:
            self.input_batch.add_request(request)
1002

1003
1004
1005
1006
1007
1008
        # 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()
1009

1010
    def _update_states_after_model_execute(
1011
1012
        self, output_token_ids: torch.Tensor
    ) -> None:
1013
1014
1015
1016
1017
1018
1019
1020
1021
1022
1023
1024
        """Update the cached states after model execution.

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

        # Find the number of accepted tokens for each sequence.
1025
1026
1027
1028
1029
1030
1031
1032
1033
1034
1035
1036
1037
1038
1039
1040
1041
1042
1043
1044
        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()
        )
1045
1046
1047
        for i, num_tokens in enumerate(num_accepted_tokens):
            self.input_batch.num_accepted_tokens_cpu[i] = num_tokens

1048
    def _init_mrope_positions(self, req_state: CachedRequestState):
1049
1050
        model = self.get_model()
        assert supports_mrope(model), "M-RoPE support is not implemented."
1051
1052
1053
1054
        assert req_state.prompt_token_ids is not None, (
            "M-RoPE requires prompt_token_ids to be available."
        )
        mrope_model = cast(SupportsMRoPE, model)
1055
1056

        req_state.mrope_positions, req_state.mrope_position_delta = (
1057
            mrope_model.get_mrope_input_positions(
1058
                req_state.prompt_token_ids,
1059
                req_state.mm_features,
1060
            )
1061
        )
1062

1063
1064
1065
1066
1067
1068
1069
1070
1071
1072
1073
1074
1075
    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,
        )

1076
    def _extract_mm_kwargs(
1077
        self,
1078
1079
        scheduler_output: "SchedulerOutput",
    ) -> BatchedTensorInputs:
1080
        if not scheduler_output or not self.is_multimodal_raw_input_only_model:
1081
            return {}
1082

1083
1084
        mm_kwargs = list[MultiModalKwargsItem]()
        for req in scheduler_output.scheduled_new_reqs:
1085
1086
1087
            for feature in req.mm_features:
                if feature.data is not None:
                    mm_kwargs.append(feature.data)
1088

1089
        # Input all modalities at once
1090
        model = cast(SupportsMultiModal, self.model)
1091
1092
        mm_kwargs_combined: BatchedTensorInputs = {}
        for _, _, mm_kwargs_group in group_mm_kwargs_by_modality(
1093
1094
1095
1096
            mm_kwargs,
            device=self.device,
            pin_memory=self.pin_memory,
            merge_by_field_config=model.merge_by_field_config,
1097
            multimodal_cpu_fields=model.multimodal_cpu_fields,
1098
1099
        ):
            mm_kwargs_combined.update(mm_kwargs_group)
1100

1101
        return mm_kwargs_combined
1102

1103
    def _dummy_mm_kwargs(self, num_seqs: int) -> BatchedTensorInputs:
1104
        if not self.is_multimodal_raw_input_only_model:
1105
            return {}
1106

1107
1108
1109
1110
1111
        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)
1112

1113
1114
1115
    def _get_cumsum_and_arange(
        self,
        num_tokens: np.ndarray,
1116
        cumsum_dtype: np.dtype | None = None,
1117
1118
1119
1120
1121
1122
1123
1124
1125
1126
1127
1128
1129
1130
1131
1132
    ) -> 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

1133
    def _prepare_input_ids(
1134
1135
1136
1137
        self,
        scheduler_output: "SchedulerOutput",
        total_num_scheduled_tokens: int,
        cu_num_tokens: np.ndarray,
1138
    ) -> None:
1139
        """Prepare the input IDs for the current batch.
1140

1141
1142
1143
1144
1145
1146
1147
        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)
1148
1149
1150
            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)
1151
1152
1153
1154
1155
1156
1157
            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
1158
1159
1160
1161
        sample_flattened_indices: list[int] = []
        spec_flattened_indices: list[int] = []
        prev_common_req_indices: list[int] = []
        prev_draft_token_indices: list[int] = []
1162
1163
        indices_match = True
        max_flattened_index = -1
1164
1165
1166
        total_num_spec_tokens = 0
        scheduled_spec_tokens = scheduler_output.scheduled_spec_decode_tokens

1167
1168
1169
1170
1171
        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.
1172
1173
                draft_len = len(scheduled_spec_tokens.get(req_id, ()))
                total_num_spec_tokens += draft_len
1174
                flattened_index = cu_num_tokens[cur_index].item() - 1
1175
1176
1177
1178
1179
1180
1181
1182
1183
1184
1185
1186
1187
1188
1189
                # 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))
1190
                indices_match &= prev_index == flattened_index
1191
                max_flattened_index = max(max_flattened_index, flattened_index)
1192
1193
1194
        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:
1195
1196
1197
            # 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)
1198
1199
1200
            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)
1201
1202
        if num_commmon_tokens == 0:
            # No requests in common with the previous iteration
1203
            # So input_ids.cpu will have all the input ids.
1204
1205
1206
1207
1208
1209
1210
            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_(
1211
1212
1213
                self.input_batch.prev_sampled_token_ids[:num_commmon_tokens, 0],
                non_blocking=True,
            )
1214
1215
            if self.enable_prompt_embeds:
                self.is_token_ids.gpu[:num_commmon_tokens] = True
1216
            return
1217
        # Upload the index tensors asynchronously so the scatter can be non-blocking.
1218
1219
        sampled_tokens_index_tensor = torch.tensor(
            sample_flattened_indices, dtype=torch.int64, pin_memory=self.pin_memory
1220
        ).to(self.device, non_blocking=True)
1221
        prev_common_req_indices_tensor = torch.tensor(
1222
1223
            prev_common_req_indices, dtype=torch.int64, pin_memory=self.pin_memory
        ).to(self.device, non_blocking=True)
1224
1225
        self.input_ids.gpu.scatter_(
            dim=0,
1226
            index=sampled_tokens_index_tensor,
1227
            src=self.input_batch.prev_sampled_token_ids[
1228
1229
1230
                prev_common_req_indices_tensor, 0
            ],
        )
1231

1232
1233
1234
1235
1236
1237
1238
1239
1240
1241
1242
1243
1244
1245
1246
1247
1248
1249
1250
1251
1252
1253
1254
        # Scatter the draft tokens after the sampled tokens are scattered.
        if self._draft_token_ids is None or not spec_flattened_indices:
            return

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

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

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

1255
1256
    def _get_encoder_seq_lens(
        self,
1257
        num_scheduled_tokens: dict[str, int],
1258
1259
        kv_cache_spec: KVCacheSpec,
        num_reqs: int,
1260
    ) -> tuple[torch.Tensor | None, np.ndarray | None]:
1261
        if not isinstance(kv_cache_spec, CrossAttentionSpec):
1262
            return None, None
1263
1264
1265

        # Build encoder_seq_lens array mapping request indices to
        # encoder lengths for inputs scheduled in this batch
1266
        for req_id in num_scheduled_tokens:
1267
            req_index = self.input_batch.req_id_to_index[req_id]
1268
1269
1270
1271
1272
1273
1274
1275
1276
1277
1278
1279
1280
1281
1282
1283
            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]
1284

1285
        return encoder_seq_lens, encoder_seq_lens_cpu
1286

1287
    def _prepare_inputs(
1288
1289
1290
        self,
        scheduler_output: "SchedulerOutput",
        num_scheduled_tokens: np.ndarray,
1291
1292
    ) -> tuple[
        torch.Tensor,
1293
        SpecDecodeMetadata | None,
1294
    ]:
1295
1296
        """
        :return: tuple[
1297
            logits_indices, spec_decode_metadata,
1298
1299
        ]
        """
1300
1301
1302
1303
1304
1305
1306
        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.
1307
        self.input_batch.block_table.commit_block_table(num_reqs)
1308
1309
1310

        # Get request indices.
        # E.g., [2, 5, 3] -> [0, 0, 1, 1, 1, 1, 1, 2, 2, 2]
1311
        req_indices = np.repeat(self.arange_np[:num_reqs], num_scheduled_tokens)
1312

1313
1314
        # cu_num_tokens: [2, 5, 3] -> [2, 7, 10]
        # arange: [0, 1, 0, 1, 2, 3, 4, 0, 1, 2]
1315
        cu_num_tokens, arange = self._get_cumsum_and_arange(num_scheduled_tokens)
1316
1317

        # Get positions.
1318
        positions_np = self.positions.np[:total_num_scheduled_tokens]
1319
1320
1321
1322
1323
        np.add(
            self.input_batch.num_computed_tokens_cpu[req_indices],
            arange,
            out=positions_np,
        )
1324

1325
1326
        # Calculate M-RoPE positions.
        # Only relevant for models using M-RoPE (e.g, Qwen2-VL)
1327
        if self.uses_mrope:
1328
1329
            self._calc_mrope_positions(scheduler_output)

1330
1331
1332
1333
1334
        # 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)

1335
1336
1337
1338
        # 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.
1339
1340
1341
        token_indices = (
            positions_np + req_indices * self.input_batch.token_ids_cpu.shape[1]
        )
1342
        token_indices_tensor = torch.from_numpy(token_indices)
1343

1344
1345
1346
        # 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.
1347
1348
1349
1350
1351
1352
        torch.index_select(
            self.input_batch.token_ids_cpu_tensor.flatten(),
            0,
            token_indices_tensor,
            out=self.input_ids.cpu[:total_num_scheduled_tokens],
        )
1353
        if self.enable_prompt_embeds:
1354
            is_token_ids = self.input_batch.is_token_ids_tensor.flatten()
1355
1356
1357
1358
            torch.index_select(
                is_token_ids,
                0,
                token_indices_tensor,
1359
1360
                out=self.is_token_ids.cpu[:total_num_scheduled_tokens],
            )
1361
1362
1363
1364
1365
1366
1367
1368
1369
1370
1371
1372
1373
1374
1375
1376
1377
1378
1379
1380
1381
1382
1383
1384
1385
1386
1387
1388
1389
1390
1391
1392
1393

        # 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:
1394
1395
1396
                    self.inputs_embeds.cpu[
                        output_idx : output_idx + actual_num_sched
                    ].copy_(req_embeds[start_pos:actual_end])
1397
1398

                output_idx += num_sched
1399

1400
1401
        self.input_batch.block_table.compute_slot_mapping(req_indices, positions_np)
        self.input_batch.block_table.commit_slot_mapping(total_num_scheduled_tokens)
1402
1403

        # Prepare the attention metadata.
1404
        self.query_start_loc.np[0] = 0
1405
        self.query_start_loc.np[1 : num_reqs + 1] = cu_num_tokens
1406
1407
        # Note: pad query_start_loc to be non-decreasing, as kernels
        # like FlashAttention requires that
1408
        self.query_start_loc.np[num_reqs + 1 :].fill(cu_num_tokens[-1])
1409
        self.query_start_loc.copy_to_gpu()
1410
        query_start_loc = self.query_start_loc.gpu[: num_reqs + 1]
1411

1412
        self.seq_lens.np[:num_reqs] = (
1413
1414
            self.input_batch.num_computed_tokens_cpu[:num_reqs] + num_scheduled_tokens
        )
1415
        # Fill unused with 0 for full cuda graph mode.
1416
1417
        self.seq_lens.np[num_reqs:].fill(0)
        self.seq_lens.copy_to_gpu()
1418

1419
        num_tokens = [self.requests[r].num_tokens for r in self.input_batch.req_ids]
1420
1421
        num_tokens_np = np.array(num_tokens, dtype=np.int32)

1422
        # Record which requests should not be sampled,
1423
        # so that we could clear the sampled tokens before returning
1424
1425
        self.discard_request_mask.np[:num_reqs] = (
            self.seq_lens.np[:num_reqs] < num_tokens_np
1426
        )
1427
        self.discard_request_mask.copy_to_gpu(num_reqs)
1428

1429
        # Copy the tensors to the GPU.
1430
1431
1432
1433
1434
        self._prepare_input_ids(
            scheduler_output,
            total_num_scheduled_tokens,
            cu_num_tokens,
        )
1435

1436
        if self.uses_mrope:
1437
            # Only relevant for models using M-RoPE (e.g, Qwen2-VL)
1438
1439
            self.mrope_positions.gpu[:, :total_num_scheduled_tokens].copy_(
                self.mrope_positions.cpu[:, :total_num_scheduled_tokens],
1440
1441
                non_blocking=True,
            )
1442
1443
1444
1445
1446
1447
        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,
            )
1448
1449
        else:
            # Common case (1D positions)
1450
            self.positions.copy_to_gpu(total_num_scheduled_tokens)
1451

1452
        use_spec_decode = len(scheduler_output.scheduled_spec_decode_tokens) > 0
1453
1454
1455
1456
1457
1458
1459
        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
1460
            num_draft_tokens = None
1461
            spec_decode_metadata = None
1462
            num_sampled_tokens = np.ones(num_reqs, dtype=np.int32)
1463
1464
1465
1466
1467
        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)
1468
1469
1470
            # 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)
1471
1472
1473
1474
            for (
                req_id,
                draft_token_ids,
            ) in scheduler_output.scheduled_spec_decode_tokens.items():
1475
1476
                req_idx = self.input_batch.req_id_to_index[req_id]
                num_draft_tokens[req_idx] = len(draft_token_ids)
1477
1478
1479
1480
1481
1482
1483
1484
                num_decode_draft_tokens[req_idx] = (
                    len(draft_token_ids)
                    if (
                        self.input_batch.num_computed_tokens_cpu[req_idx]
                        >= self.input_batch.num_prompt_tokens[req_idx]
                    )
                    else -1
                )
1485
            spec_decode_metadata = self._calc_spec_decode_metadata(
1486
1487
                num_draft_tokens, cu_num_tokens
            )
1488
            logits_indices = spec_decode_metadata.logits_indices
1489
            num_sampled_tokens = num_draft_tokens + 1
1490
            # For DECODE only cuda graph of some attention backends (e.g., GDN).
1491
            self.num_decode_draft_tokens.np[:num_reqs] = num_decode_draft_tokens
1492
1493
            self.num_decode_draft_tokens.np[num_reqs:].fill(-1)
            self.num_decode_draft_tokens.copy_to_gpu()
1494

1495
1496
1497
1498
1499
        # Hot-Swap lora model
        if self.lora_config:
            assert (
                np.sum(num_sampled_tokens)
                <= self.vllm_config.scheduler_config.max_num_batched_tokens
1500
            )
1501
1502
1503
1504
1505
1506
1507
1508
1509
1510
1511
            self.set_active_loras(
                self.input_batch, num_scheduled_tokens, num_sampled_tokens
            )

        return (
            logits_indices,
            spec_decode_metadata,
        )

    def _build_attention_metadata(
        self,
1512
        num_tokens: int,
1513
        num_reqs: int,
1514
1515
1516
        max_query_len: int,
        num_tokens_padded: int | None = None,
        num_reqs_padded: int | None = None,
1517
1518
1519
1520
        ubatch_slices: UBatchSlices | None = None,
        logits_indices: torch.Tensor | None = None,
        use_spec_decode: bool = False,
        for_cudagraph_capture: bool = False,
1521
        num_scheduled_tokens: dict[str, int] | None = None,
1522
1523
1524
1525
1526
        cascade_attn_prefix_lens: list[list[int]] | None = None,
    ) -> tuple[PerLayerAttnMetadata, CommonAttentionMetadata | None]:
        """
        :return: tuple[attn_metadata, spec_decode_common_attn_metadata]
        """
1527
1528
1529
        num_tokens_padded = num_tokens_padded or num_tokens
        num_reqs_padded = num_reqs_padded or num_reqs

1530
        logits_indices_padded = None
1531
        num_logits_indices = None
1532
1533
1534
1535
1536
1537
        if logits_indices is not None:
            num_logits_indices = logits_indices.size(0)
            if self.cache_config.kv_sharing_fast_prefill:
                logits_indices_padded = self._prepare_kv_sharing_fast_prefill(
                    logits_indices
                )
1538

1539
1540
1541
1542
1543
1544
        # update seq_lens of decode reqs under DCP.
        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,
1545
                self.parallel_config.cp_kv_cache_interleave_size,
1546
            )
1547
1548
            self.dcp_local_seq_lens.cpu[num_reqs:].fill_(0)
            self.dcp_local_seq_lens.copy_to_gpu(num_reqs_padded)
1549

1550
1551
1552
        attn_metadata: PerLayerAttnMetadata = {}
        if ubatch_slices is not None:
            attn_metadata = [dict() for _ in range(len(ubatch_slices))]
1553

1554
1555
1556
1557
1558
1559
1560
1561
        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()

1562
1563
        if use_spec_decode:
            self.num_accepted_tokens.np[:num_reqs] = (
1564
1565
                self.input_batch.num_accepted_tokens_cpu[:num_reqs]
            )
1566
1567
            self.num_accepted_tokens.np[num_reqs:].fill(1)
            self.num_accepted_tokens.copy_to_gpu()
1568

1569
1570
1571
1572
1573
1574
1575
1576
1577
1578
1579
1580
1581
1582
1583
1584
        # Used in the below loop, uses padded shapes
        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
        ]

        dcp_local_seq_lens, dcp_local_seq_lens_cpu = None, None
        if self.dcp_world_size > 1:
            dcp_local_seq_lens = self.dcp_local_seq_lens.gpu[:num_reqs_padded]
            dcp_local_seq_lens_cpu = self.dcp_local_seq_lens.cpu[:num_reqs_padded]

        spec_decode_common_attn_metadata = None

1585
1586
        # Prepare the attention metadata for each KV cache group and make layers
        # in the same group share the same metadata.
1587
        for kv_cache_gid, kv_cache_group in enumerate(
1588
1589
            self.kv_cache_config.kv_cache_groups
        ):
1590
1591
            encoder_seq_lens, encoder_seq_lens_cpu = self._get_encoder_seq_lens(
                num_scheduled_tokens or {},
1592
                kv_cache_group.kv_cache_spec,
1593
                num_reqs_padded,
1594
            )
1595

1596
            if isinstance(kv_cache_group.kv_cache_spec, EncoderOnlyAttentionSpec):
1597
1598
1599
                # Encoder-only layers do not have KV cache, so we need to
                # create a dummy block table and slot mapping for them.
                blk_table_tensor = torch.zeros(
1600
                    (num_reqs_padded, 1),
1601
                    dtype=torch.int32,
1602
1603
1604
                    device=self.device,
                )
                slot_mapping = torch.zeros(
1605
                    (num_tokens_padded,),
1606
1607
1608
                    dtype=torch.int64,
                    device=self.device,
                )
1609
            else:
1610
                blk_table = self.input_batch.block_table[kv_cache_gid]
1611
1612
                blk_table_tensor = blk_table.get_device_tensor(num_reqs_padded)
                slot_mapping = blk_table.slot_mapping.gpu[:num_tokens_padded]
1613
1614

                # Fill unused with -1. Needed for reshape_and_cache in full cuda
1615
1616
1617
                # graph mode. `blk_table_tensor` -1 to match mamba PAD_SLOT_ID
                slot_mapping[num_tokens:num_tokens_padded].fill_(-1)
                blk_table_tensor[num_reqs:num_reqs_padded].fill_(-1)
1618

1619
            common_attn_metadata = CommonAttentionMetadata(
1620
1621
1622
1623
1624
                query_start_loc=query_start_loc,
                query_start_loc_cpu=query_start_loc_cpu,
                seq_lens=seq_lens,
                seq_lens_cpu=seq_lens_cpu,
                num_computed_tokens_cpu=num_computed_tokens_cpu,
1625
1626
1627
                num_actual_tokens=num_tokens_padded,
                num_reqs=num_reqs_padded,
                max_query_len=max_query_len,
1628
                max_seq_len=max_seq_len,
1629
1630
                block_table_tensor=blk_table_tensor,
                slot_mapping=slot_mapping,
1631
                logits_indices_padded=logits_indices_padded,
1632
                num_logits_indices=num_logits_indices,
1633
                causal=True,
1634
                encoder_seq_lens=encoder_seq_lens,
1635
                encoder_seq_lens_cpu=encoder_seq_lens_cpu,
1636
                dcp_local_seq_lens=dcp_local_seq_lens,
1637
                dcp_local_seq_lens_cpu=dcp_local_seq_lens_cpu,
1638
1639
            )

1640
            if self.speculative_config and spec_decode_common_attn_metadata is None:
1641
                if isinstance(self.drafter, EagleProposer):
1642
                    if self.drafter.attn_layer_names[0] in kv_cache_group.layer_names:
1643
1644
1645
                        spec_decode_common_attn_metadata = common_attn_metadata
                else:
                    spec_decode_common_attn_metadata = common_attn_metadata
1646

1647
1648
1649
1650
1651
1652
            for attn_gid, attn_group in enumerate(self.attn_groups[kv_cache_gid]):
                cascade_attn_prefix_len = (
                    cascade_attn_prefix_lens[kv_cache_gid][attn_gid]
                    if cascade_attn_prefix_lens
                    else 0
                )
1653
                builder = attn_group.get_metadata_builder()
1654

1655
                extra_attn_metadata_args = {}
1656
                if use_spec_decode and isinstance(builder, GDNAttentionMetadataBuilder):
1657
                    extra_attn_metadata_args = dict(
1658
1659
1660
                        num_accepted_tokens=self.num_accepted_tokens.gpu[
                            :num_reqs_padded
                        ],
1661
                        num_decode_draft_tokens_cpu=self.num_decode_draft_tokens.cpu[
1662
                            :num_reqs_padded
1663
                        ],
1664
1665
                    )

1666
1667
                if ubatch_slices is not None:
                    common_attn_metadata_list = split_attn_metadata(
1668
1669
                        ubatch_slices, common_attn_metadata
                    )
1670
                    for ubid, common_attn_metadata in enumerate(
1671
1672
                        common_attn_metadata_list
                    ):
1673
1674
1675
1676
1677
1678
1679
1680
1681
1682
1683
                        builder = attn_group.get_metadata_builder(ubatch_id=ubid)
                        if for_cudagraph_capture:
                            attn_metadata_i = builder.build_for_cudagraph_capture(
                                common_attn_metadata
                            )
                        else:
                            attn_metadata_i = builder.build(
                                common_prefix_len=cascade_attn_prefix_len,
                                common_attn_metadata=common_attn_metadata,
                            )
                        for layer_name in kv_cache_group.layer_names:
1684
1685
1686
1687
                            assert type(attn_metadata) is list
                            attn_metadata[ubid][layer_name] = attn_metadata_i
                else:
                    assert isinstance(attn_metadata, dict)
1688
1689
1690
1691
1692
1693
1694
1695
1696
1697
                    if for_cudagraph_capture:
                        attn_metadata_i = builder.build_for_cudagraph_capture(
                            common_attn_metadata
                        )
                    else:
                        attn_metadata_i = builder.build(
                            common_prefix_len=cascade_attn_prefix_len,
                            common_attn_metadata=common_attn_metadata,
                            **extra_attn_metadata_args,
                        )
1698
1699
                    for layer_name in attn_group.layer_names:
                        attn_metadata[layer_name] = attn_metadata_i
1700

1701
1702
1703
1704
1705
1706
1707
1708
1709
1710
        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)
            )

1711
        return attn_metadata, spec_decode_common_attn_metadata
1712

1713
1714
1715
    def _compute_cascade_attn_prefix_lens(
        self,
        num_scheduled_tokens: np.ndarray,
1716
        num_computed_tokens: np.ndarray,
1717
1718
1719
1720
1721
1722
1723
        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
        """
1724

1725
1726
1727
1728
1729
1730
1731
1732
1733
1734
1735
1736
1737
1738
        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,
1739
                        num_computed_tokens,
1740
1741
1742
1743
1744
1745
1746
1747
                        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
1748

1749
1750
1751
    def _compute_cascade_attn_prefix_len(
        self,
        num_scheduled_tokens: np.ndarray,
1752
        num_computed_tokens: np.ndarray,
1753
        num_common_prefix_blocks: int,
1754
1755
        kv_cache_spec: KVCacheSpec,
        attn_metadata_builder: AttentionMetadataBuilder,
1756
1757
1758
1759
1760
1761
1762
1763
1764
1765
1766
1767
1768
1769
1770
1771
1772
1773
    ) -> 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.
        """
1774

1775
        common_prefix_len = num_common_prefix_blocks * kv_cache_spec.block_size
1776
1777
1778
1779
1780
1781
1782
1783
1784
1785
1786
1787
1788
1789
1790
1791
1792
1793
1794
1795
1796
1797
1798
1799
1800
1801
1802
1803
1804
1805
1806
1807
1808
1809
1810
1811
1812
        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]
1813
        # Request 3's num_computed_tokens: 3 (i.e., [A, B, C])
1814
1815
1816
1817
1818
        # 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.
1819
        common_prefix_len = min(common_prefix_len, num_computed_tokens.min())
1820
        # common_prefix_len should be a multiple of the block size.
1821
1822
1823
1824
1825
1826
1827
1828
1829
1830
1831
        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
        )
1832
1833
        assert isinstance(kv_cache_spec, AttentionSpec)
        use_cascade = attn_metadata_builder.use_cascade_attention(
1834
1835
1836
            common_prefix_len=common_prefix_len,
            query_lens=num_scheduled_tokens,
            num_query_heads=self.num_query_heads,
1837
            num_kv_heads=kv_cache_spec.num_kv_heads,
1838
            use_alibi=self.use_alibi,
1839
            use_sliding_window=use_sliding_window,
1840
            use_local_attention=use_local_attention,
1841
            num_sms=self.num_sms,
1842
            dcp_world_size=self.dcp_world_size,
1843
1844
1845
        )
        return common_prefix_len if use_cascade else 0

1846
1847
    def _calc_mrope_positions(self, scheduler_output: "SchedulerOutput"):
        mrope_pos_ptr = 0
1848
        for index, req_id in enumerate(self.input_batch.req_ids):
1849
1850
1851
            req = self.requests[req_id]
            assert req.mrope_positions is not None

1852
1853
            num_computed_tokens = self.input_batch.num_computed_tokens_cpu[index]
            num_scheduled_tokens = scheduler_output.num_scheduled_tokens[req_id]
1854
            num_prompt_tokens = length_from_prompt_token_ids_or_embeds(
1855
1856
                req.prompt_token_ids, req.prompt_embeds
            )
1857
1858

            if num_computed_tokens + num_scheduled_tokens > num_prompt_tokens:
1859
1860
                prompt_part_len = max(0, num_prompt_tokens - num_computed_tokens)
                completion_part_len = max(0, num_scheduled_tokens - prompt_part_len)
1861
1862
1863
1864
1865
1866
1867
1868
1869
1870
1871
1872
1873
            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

1874
1875
1876
                self.mrope_positions.cpu[:, dst_start:dst_end] = req.mrope_positions[
                    :, src_start:src_end
                ]
1877
1878
1879
1880
1881
1882
1883
                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

1884
                assert req.mrope_position_delta is not None
1885
                MRotaryEmbedding.get_next_input_positions_tensor(
1886
                    out=self.mrope_positions.np,
1887
1888
1889
1890
1891
                    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,
                )
1892
1893
1894

                mrope_pos_ptr += completion_part_len

1895
1896
1897
1898
1899
1900
1901
1902
1903
1904
1905
1906
1907
1908
1909
1910
1911
1912
1913
1914
1915
1916
1917
1918
1919
1920
1921
1922
1923
1924
1925
1926
1927
1928
1929
1930
1931
1932
1933
1934
1935
1936
1937
1938
1939
1940
1941
    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

1942
1943
    def _calc_spec_decode_metadata(
        self,
1944
1945
1946
1947
1948
1949
1950
1951
1952
1953
1954
1955
1956
1957
1958
1959
        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
1960
1961
1962
1963

        # 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(
1964
1965
            num_sampled_tokens, cumsum_dtype=np.int32
        )
1966
        # Step 2. [0, 0, 0, 0, 103, 104, 104, 104, 206, 207, 207]
1967
        logits_indices = np.repeat(
1968
1969
            cu_num_scheduled_tokens - num_sampled_tokens, num_sampled_tokens
        )
1970
        # Step 3. [0, 1, 2, 3, 103, 104, 105, 106, 206, 207, 208]
1971
1972
1973
1974
1975
1976
        logits_indices += arange

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

        # Compute the draft logits indices.
1977
1978
1979
        # 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(
1980
1981
            num_draft_tokens, cumsum_dtype=np.int32
        )
1982
1983
        # [0, 0, 0, 5, 5, 9]
        target_logits_indices = np.repeat(
1984
1985
            cu_num_sampled_tokens - num_sampled_tokens, num_draft_tokens
        )
1986
1987
1988
1989
1990
        # [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(
1991
1992
            self.device, non_blocking=True
        )
1993
1994
1995
        cu_num_sampled_tokens = torch.from_numpy(cu_num_sampled_tokens).to(
            self.device, non_blocking=True
        )
1996
1997
1998
        logits_indices = torch.from_numpy(logits_indices).to(
            self.device, non_blocking=True
        )
1999
        target_logits_indices = torch.from_numpy(target_logits_indices).to(
2000
2001
            self.device, non_blocking=True
        )
2002
        bonus_logits_indices = torch.from_numpy(bonus_logits_indices).to(
2003
2004
            self.device, non_blocking=True
        )
2005

2006
2007
        # Compute the draft token ids.
        # draft_token_indices:      [  1,   2,   3, 105, 106, 208]
2008
        draft_token_ids = self.input_ids.gpu[logits_indices]
2009
2010
        draft_token_ids = draft_token_ids[target_logits_indices + 1]

2011
        return SpecDecodeMetadata(
2012
2013
2014
            draft_token_ids=draft_token_ids,
            num_draft_tokens=num_draft_tokens.tolist(),
            cu_num_draft_tokens=cu_num_draft_tokens,
2015
            cu_num_sampled_tokens=cu_num_sampled_tokens,
2016
2017
2018
2019
2020
            target_logits_indices=target_logits_indices,
            bonus_logits_indices=bonus_logits_indices,
            logits_indices=logits_indices,
        )

2021
2022
2023
2024
2025
2026
2027
    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
2028
        self.kv_sharing_fast_prefill_logits_indices[:num_logits].copy_(logits_indices)
2029
2030
2031
2032
2033
        # 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_(
2034
2035
2036
2037
2038
2039
            logits_indices[-1].item()
        )
        if (
            self.compilation_config.cudagraph_mode != CUDAGraphMode.NONE
            and num_logits <= self.cudagraph_batch_sizes[-1]
        ):
2040
2041
2042
2043
2044
            # Use piecewise CUDA graphs.
            # Add padding to the batch size.
            num_logits_padded = self.vllm_config.pad_for_cudagraph(num_logits)
        else:
            num_logits_padded = num_logits
2045
2046
2047
        logits_indices_padded = self.kv_sharing_fast_prefill_logits_indices[
            :num_logits_padded
        ]
2048
2049
        return logits_indices_padded

2050
2051
2052
2053
2054
2055
2056
2057
    def _batch_mm_kwargs_from_scheduler(
        self,
        scheduler_output: "SchedulerOutput",
    ) -> tuple[list[MultiModalKwargsItem], list[tuple[str, PlaceholderRange]]]:
        """Batch multimodal kwargs from scheduled encoder inputs.

        Args:
            scheduler_output: The scheduler output containing scheduled encoder
2058
                inputs.
2059
2060
2061
2062
2063
2064

        Returns:
            A tuple of (mm_kwargs, req_ids_pos) where:
            - mm_kwargs: List of multimodal kwargs items to be batched
            - mm_hashes_pos: List of (mm_hash, position_info) tuples
        """
2065
2066
        scheduled_encoder_inputs = scheduler_output.scheduled_encoder_inputs
        if not scheduled_encoder_inputs:
2067
            return [], []
2068
        # Batch the multi-modal inputs.
2069
        mm_kwargs = list[MultiModalKwargsItem]()
2070
2071
        # list of tuple (mm_hash, position_info)
        mm_hashes_pos = list[tuple[str, PlaceholderRange]]()
2072
2073
        for req_id, encoder_input_ids in scheduled_encoder_inputs.items():
            req_state = self.requests[req_id]
2074
2075

            for mm_input_id in encoder_input_ids:
2076
                mm_feature = req_state.mm_features[mm_input_id]
2077
2078
                if mm_feature.data is None:
                    continue
2079
2080
2081
                mm_hash = mm_feature.identifier
                mm_kwargs.append(mm_feature.data)
                mm_hashes_pos.append((mm_hash, mm_feature.mm_position))
2082

2083
2084
        return mm_kwargs, mm_hashes_pos

2085
2086
2087
    def _execute_mm_encoder(
        self, scheduler_output: "SchedulerOutput"
    ) -> list[torch.Tensor]:
2088
2089
        # Batch the multi-modal inputs using the helper method.
        mm_kwargs, mm_hashes_pos = self._batch_mm_kwargs_from_scheduler(
2090
2091
            scheduler_output
        )
2092
2093

        if not mm_kwargs:
2094
            return []
2095

2096
2097
2098
2099
2100
2101
2102
        # 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.
2103
        model = cast(SupportsMultiModal, self.model)
2104
        encoder_outputs: list[torch.Tensor] = []
2105
        for modality, num_items, mm_kwargs_group in group_mm_kwargs_by_modality(
2106
2107
2108
2109
            mm_kwargs,
            device=self.device,
            pin_memory=self.pin_memory,
            merge_by_field_config=model.merge_by_field_config,
2110
            multimodal_cpu_fields=model.multimodal_cpu_fields,
2111
        ):
2112
            curr_group_outputs: list[torch.Tensor] = []
2113
2114

            # EVS-related change.
2115
            # (ekhvedchenia): Temporary hack to limit peak memory usage when
2116
            # processing multimodal data. This solves the issue with scheduler
2117
2118
2119
2120
            # 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)
2121
2122
2123
2124
2125
2126
2127
2128
2129
2130
2131
2132
2133
2134
2135
2136
            # TODO(ywang96): Fix memory profiling to take EVS into account and
            # remove this hack.
            if (
                self.is_multimodal_pruning_enabled
                and modality == "video"
                and num_items > 1
            ):
                for video_mm_kwargs_item in filter(
                    lambda item: item.modality == "video", mm_kwargs
                ):
                    _, _, micro_batch_mm_inputs = next(
                        group_mm_kwargs_by_modality(
                            [video_mm_kwargs_item],
                            device=self.device,
                            pin_memory=self.pin_memory,
                            merge_by_field_config=model.merge_by_field_config,
2137
                            multimodal_cpu_fields=model.multimodal_cpu_fields,
2138
                        )
2139
                    )
2140

2141
                    micro_batch_outputs = model.embed_multimodal(
2142
2143
                        **micro_batch_mm_inputs
                    )
2144
2145
2146
2147
2148
2149
2150
2151
2152
2153

                    curr_group_outputs.extend(micro_batch_outputs)
            else:
                # Run the encoder.
                # `curr_group_outputs` is either of the following:
                # 1. A tensor of shape (num_items, feature_size, hidden_size)
                # in case feature_size is fixed across all multimodal items.
                # 2. A list or tuple (length: num_items) of tensors,
                # each of shape (feature_size, hidden_size) in case the feature
                # size is dynamic depending on the input multimodal items.
2154
                curr_group_outputs = model.embed_multimodal(**mm_kwargs_group)  # type: ignore[assignment]
2155

2156
2157
            sanity_check_mm_encoder_outputs(
                curr_group_outputs,
2158
                expected_num_items=num_items,
2159
            )
2160
            encoder_outputs.extend(curr_group_outputs)
2161

2162
2163
2164
        # Cache the encoder outputs by mm_hash
        for (mm_hash, pos_info), output in zip(mm_hashes_pos, encoder_outputs):
            self.encoder_cache[mm_hash] = scatter_mm_placeholders(
2165
2166
2167
                output,
                is_embed=pos_info.is_embed,
            )
2168
2169
            logger.debug("Finish execute for mm hash %s", mm_hash)
            self.maybe_save_ec_to_connector(self.encoder_cache, mm_hash)
2170

2171
2172
        return encoder_outputs

2173
    def _gather_mm_embeddings(
2174
2175
        self,
        scheduler_output: "SchedulerOutput",
2176
        shift_computed_tokens: int = 0,
2177
2178
2179
2180
2181
2182
2183
2184
    ) -> tuple[list[torch.Tensor], torch.Tensor]:
        total_num_scheduled_tokens = scheduler_output.total_num_scheduled_tokens

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

        req_start_idx = 0
2185
        should_sync_mrope_positions = False
2186
        should_sync_xdrope_positions = False
2187

2188
        for req_id in self.input_batch.req_ids:
2189
2190
            mm_embeds_req: list[torch.Tensor] = []

2191
            num_scheduled_tokens = scheduler_output.num_scheduled_tokens[req_id]
2192
            req_state = self.requests[req_id]
2193
            num_computed_tokens = req_state.num_computed_tokens + shift_computed_tokens
2194

2195
2196
            for mm_feature in req_state.mm_features:
                pos_info = mm_feature.mm_position
2197
2198
                start_pos = pos_info.offset
                num_encoder_tokens = pos_info.length
2199
2200
2201
2202
2203
2204
2205
2206
2207
2208
2209
2210
2211
2212
2213
2214

                # 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,
2215
2216
                    num_encoder_tokens,
                )
2217
                assert start_idx < end_idx
2218

2219
                mm_hash = mm_feature.identifier
2220
                encoder_output = self.encoder_cache.get(mm_hash, None)
2221
                assert encoder_output is not None, f"Encoder cache miss for {mm_hash}."
2222
2223
2224
2225

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

2226
                req_start_pos = req_start_idx + start_pos - num_computed_tokens
2227
2228
2229
                is_mm_embed[req_start_pos + start_idx : req_start_pos + end_idx] = (
                    True if is_embed is None else is_embed
                )
2230

2231
2232
2233
2234
                mm_embeds_item = gather_mm_placeholders(
                    encoder_output[start_idx:end_idx],
                    is_embed=is_embed,
                )
2235
2236
2237
                mm_embeds_req.append(mm_embeds_item)

            if self.is_multimodal_pruning_enabled and self.uses_mrope:
2238
                assert req_state.mrope_positions is not None
2239
2240
2241
2242
2243
2244
2245
                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,
2246
2247
                    )
                )
2248
2249
2250
2251
                req_state.mrope_positions.copy_(new_mrope_positions)
                req_state.mrope_position_delta = new_delta

            mm_embeds.extend(mm_embeds_req)
2252
2253
2254
            req_start_idx += num_scheduled_tokens

        is_mm_embed = self.is_mm_embed.copy_to_gpu(total_num_scheduled_tokens)
2255
2256
2257

        if should_sync_mrope_positions:
            self._calc_mrope_positions(scheduler_output)
2258
            self.mrope_positions.copy_to_gpu(total_num_scheduled_tokens)
2259

2260
2261
2262
2263
        if should_sync_xdrope_positions:
            self._calc_xdrope_positions(scheduler_output)
            self.xdrope_positions.copy_to_gpu(total_num_scheduled_tokens)

2264
        return mm_embeds, is_mm_embed
2265

2266
    def get_model(self) -> nn.Module:
2267
        # get raw model out of the cudagraph wrapper.
2268
        if isinstance(self.model, (CUDAGraphWrapper, UBatchWrapper)):
2269
            return self.model.unwrap()
2270
2271
        return self.model

2272
2273
2274
2275
2276
2277
2278
2279
2280
2281
2282
2283
2284
2285
2286
    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

2287
2288
2289
2290
2291
    def get_supported_pooling_tasks(self) -> list[PoolingTask]:
        model = self.get_model()
        if not is_pooling_model(model):
            return []

2292
2293
        supported_tasks = list(model.pooler.get_supported_tasks())

2294
2295
2296
2297
        if "score" in supported_tasks:
            num_labels = getattr(self.model_config.hf_config, "num_labels", 0)
            if num_labels != 1:
                supported_tasks.remove("score")
2298
                logger.debug_once("Score API is only enabled for num_labels == 1.")
2299
2300

        return supported_tasks
2301

2302
2303
2304
2305
2306
2307
2308
2309
2310
2311
    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)

2312
    def sync_and_slice_intermediate_tensors(
2313
2314
        self,
        num_tokens: int,
2315
        intermediate_tensors: IntermediateTensors | None,
2316
2317
        sync_self: bool,
    ) -> IntermediateTensors:
2318
2319
2320
        assert self.intermediate_tensors is not None

        tp = self.vllm_config.parallel_config.tensor_parallel_size
2321
        is_rs = is_residual_scattered_for_sp(self.vllm_config, num_tokens)
2322
2323
2324
2325
2326
2327

        # 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():
2328
                is_scattered = k == "residual" and is_rs
2329
                copy_len = num_tokens // tp if is_scattered else num_tokens
2330
                self.intermediate_tensors[k][:copy_len].copy_(
2331
2332
2333
2334
2335
2336
2337
2338
2339
2340
2341
2342
2343
                    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:
2344
2345
2346
2347
2348
2349
2350
        """
        Step for the EPLB (Expert Parallelism Load Balancing) state.
        """
        if not self.parallel_config.enable_eplb:
            return

        assert self.eplb_state is not None
2351
2352
        model = self.get_model()
        assert is_mixture_of_experts(model)
2353
2354
2355
        self.eplb_state.step(
            is_dummy,
            is_profile,
2356
            log_stats=self.parallel_config.eplb_config.log_balancedness,
2357
2358
        )

2359
2360
2361
2362
2363
2364
    def _pool(
        self,
        hidden_states: torch.Tensor,
        num_scheduled_tokens: int,
        num_scheduled_tokens_np: np.ndarray,
    ) -> ModelRunnerOutput:
2365
2366
2367
        assert self.input_batch.num_reqs == len(self.input_batch.pooling_params), (
            "Either all or none of the requests in a batch must be pooling request"
        )
2368

2369
        hidden_states = hidden_states[:num_scheduled_tokens]
2370
2371
        seq_lens_cpu = self.seq_lens.cpu[: self.input_batch.num_reqs]

2372
        pooling_metadata = self.input_batch.get_pooling_metadata()
2373
        pooling_metadata.build_pooling_cursor(
2374
            num_scheduled_tokens_np.tolist(), seq_lens_cpu, device=hidden_states.device
2375
        )
2376

2377
2378
2379
2380
2381
2382
        model = cast(VllmModelForPooling, self.model)
        raw_pooler_output: PoolerOutput = model.pooler(
            hidden_states=hidden_states,
            pooling_metadata=pooling_metadata,
        )
        raw_pooler_output = json_map_leaves(
2383
            lambda x: x.to("cpu", non_blocking=True) if x is not None else x,
2384
2385
2386
            raw_pooler_output,
        )
        self._sync_device()
2387

2388
        pooler_output: list[torch.Tensor | None] = []
2389
        for raw_output, seq_len, prompt_len in zip(
2390
2391
            raw_pooler_output, seq_lens_cpu, pooling_metadata.prompt_lens
        ):
2392
            output = raw_output if seq_len == prompt_len else None
2393
            pooler_output.append(output)
2394
2395
2396
2397
2398
2399
2400
2401
2402
2403

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

2404
    def _pad_for_sequence_parallelism(self, num_scheduled_tokens: int) -> int:
2405
2406
2407
        # Pad tokens to multiple of tensor_parallel_size when
        # enabled collective fusion for SP
        tp_size = self.vllm_config.parallel_config.tensor_parallel_size
2408
        if self.compilation_config.pass_config.enable_sp and tp_size > 1:
2409
2410
2411
            return round_up(num_scheduled_tokens, tp_size)
        return num_scheduled_tokens

2412
    def _preprocess(
2413
2414
        self,
        scheduler_output: "SchedulerOutput",
2415
        num_input_tokens: int,  # Padded
2416
        intermediate_tensors: IntermediateTensors | None = None,
2417
    ) -> tuple[
2418
2419
        torch.Tensor | None,
        torch.Tensor | None,
2420
        torch.Tensor,
2421
        IntermediateTensors | None,
2422
        dict[str, Any],
2423
        ECConnectorOutput | None,
2424
    ]:
2425
        num_scheduled_tokens = scheduler_output.total_num_scheduled_tokens
2426
        is_first_rank = get_pp_group().is_first_rank
2427
        is_encoder_decoder = self.model_config.is_encoder_decoder
2428

2429
2430
        # _prepare_inputs may reorder the batch, so we must gather multi
        # modal outputs after that to ensure the correct order
2431
2432
        ec_connector_output = None

2433
        if self.supports_mm_inputs and is_first_rank and not is_encoder_decoder:
2434
            # Run the multimodal encoder if any.
2435
2436
2437
2438
2439
2440
            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)
2441

2442
2443
2444
            # 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.
2445
            inputs_embeds_scheduled = self.model.embed_input_ids(
2446
2447
2448
                self.input_ids.gpu[:num_scheduled_tokens],
                multimodal_embeddings=mm_embeds,
                is_multimodal=is_mm_embed,
2449
            )
2450

2451
            # TODO(woosuk): Avoid the copy. Optimize.
2452
            self.inputs_embeds.gpu[:num_scheduled_tokens].copy_(inputs_embeds_scheduled)
2453

2454
            input_ids = None
2455
            inputs_embeds = self.inputs_embeds.gpu[:num_input_tokens]
2456
2457
2458
2459
            model_kwargs = {
                **self._init_model_kwargs(num_scheduled_tokens),
                **self._extract_mm_kwargs(scheduler_output),
            }
2460
        elif self.enable_prompt_embeds and is_first_rank:
2461
2462
2463
2464
2465
2466
2467
2468
2469
2470
2471
2472
            # 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).
2473
2474
2475
            token_ids_idx = (
                self.is_token_ids.gpu[:num_scheduled_tokens]
                .nonzero(as_tuple=False)
2476
                .squeeze(1)
2477
            )
2478
2479
2480
            # Some tokens ids may need to become embeds
            if token_ids_idx.numel() > 0:
                token_ids = self.input_ids.gpu[token_ids_idx]
2481
                tokens_to_embeds = self.model.embed_input_ids(input_ids=token_ids)
2482
2483
2484
2485
2486
                self.inputs_embeds.gpu[token_ids_idx] = tokens_to_embeds

            inputs_embeds = self.inputs_embeds.gpu[:num_input_tokens]
            model_kwargs = self._init_model_kwargs(num_input_tokens)
            input_ids = None
2487
        else:
2488
2489
2490
2491
            # 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.
2492
            input_ids = self.input_ids.gpu[:num_input_tokens]
2493
            inputs_embeds = None
2494
            model_kwargs = self._init_model_kwargs(num_input_tokens)
2495

2496
        if self.uses_mrope:
2497
            positions = self.mrope_positions.gpu[:, :num_input_tokens]
2498
2499
        elif self.uses_xdrope_dim > 0:
            positions = self.xdrope_positions.gpu[:, :num_input_tokens]
2500
        else:
2501
            positions = self.positions.gpu[:num_input_tokens]
2502

2503
        if is_first_rank:
2504
2505
            intermediate_tensors = None
        else:
2506
            assert intermediate_tensors is not None
2507
            intermediate_tensors = self.sync_and_slice_intermediate_tensors(
2508
2509
                num_input_tokens, intermediate_tensors, True
            )
2510

2511
        if is_encoder_decoder and scheduler_output.scheduled_encoder_inputs:
2512
2513
2514
2515
2516
2517
2518
            # 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})
2519

2520
2521
2522
2523
2524
2525
        return (
            input_ids,
            inputs_embeds,
            positions,
            intermediate_tensors,
            model_kwargs,
2526
            ec_connector_output,
2527
        )
2528

2529
    def _sample(
2530
        self,
2531
2532
        logits: torch.Tensor | None,
        spec_decode_metadata: SpecDecodeMetadata | None,
2533
    ) -> SamplerOutput:
2534
        # Sample the next token and get logprobs if needed.
2535
        sampling_metadata = self.input_batch.sampling_metadata
2536
        if spec_decode_metadata is None:
2537
2538
2539
            # 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()
2540
            return self.sampler(
2541
2542
2543
                logits=logits,
                sampling_metadata=sampling_metadata,
            )
2544

2545
        sampler_output = self.rejection_sampler(
2546
2547
            spec_decode_metadata,
            None,  # draft_probs
2548
            logits,
2549
2550
            sampling_metadata,
        )
2551
        self._update_states_after_model_execute(sampler_output.sampled_token_ids)
2552
2553
2554
        return sampler_output

    def _bookkeeping_sync(
2555
2556
2557
        self,
        scheduler_output: "SchedulerOutput",
        sampler_output: SamplerOutput,
2558
        logits: torch.Tensor | None,
2559
2560
        hidden_states: torch.Tensor,
        num_scheduled_tokens: int,
2561
        spec_decode_metadata: SpecDecodeMetadata | None,
2562
    ) -> tuple[
2563
        dict[str, int],
2564
        LogprobsLists | None,
2565
        list[list[int]],
2566
        dict[str, LogprobsTensors | None],
2567
2568
2569
        list[str],
        dict[str, int],
        list[int],
2570
    ]:
2571
2572
2573
2574
        num_nans_in_logits = {}
        if envs.VLLM_COMPUTE_NANS_IN_LOGITS:
            num_nans_in_logits = self._get_nans_in_logits(logits)

2575
2576
2577
2578
        num_reqs = self.input_batch.num_reqs
        discard_sampled_tokens_req_indices = np.nonzero(
            self.discard_request_mask.np[:num_reqs]
        )[0]
2579
2580
2581
2582
        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)
2583

2584
2585
2586
        # 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()
2587
        req_id_to_index_output_copy = self.input_batch.req_id_to_index.copy()
2588
2589

        num_sampled_tokens = sampler_output.sampled_token_ids.shape[0]
2590
        sampled_token_ids = sampler_output.sampled_token_ids
2591
        logprobs_tensors = sampler_output.logprobs_tensors
2592
        invalid_req_indices = []
2593
        cu_num_tokens: list[int] | None = None
2594
2595
2596
2597
2598
2599
        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)
2600
2601
2602
                # 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()
2603
2604
            else:
                # Includes spec decode tokens.
2605
                valid_sampled_token_ids, cu_num_tokens = RejectionSampler.parse_output(
2606
2607
                    sampled_token_ids,
                    self.input_batch.vocab_size,
2608
2609
                    discard_sampled_tokens_req_indices,
                    return_cu_num_tokens=logprobs_tensors is not None,
2610
                )
2611
        else:
2612
            valid_sampled_token_ids = []
2613
            invalid_req_indices = discard_sampled_tokens_req_indices.tolist()
2614
2615
2616
2617
2618
            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.
2619
2620
2621
2622
            # 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
2623
2624
2625
2626
2627
            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
            }
2628

2629
2630
2631
2632
2633
        # 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.
2634
        req_ids = self.input_batch.req_ids
2635
2636
        for req_idx in range(num_sampled_tokens):
            if self.use_async_scheduling:
2637
                sampled_ids = [-1] if req_idx not in invalid_req_indices_set else None
2638
2639
            else:
                sampled_ids = valid_sampled_token_ids[req_idx]
2640

2641
            num_sampled_ids: int = len(sampled_ids) if sampled_ids else 0
2642

2643
            if not sampled_ids:
2644
2645
2646
                continue

            start_idx = self.input_batch.num_tokens_no_spec[req_idx]
2647
            end_idx = start_idx + num_sampled_ids
2648
2649
2650
2651
            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}"
2652
            )
2653

2654
2655
            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
2656
2657
            self.input_batch.num_tokens_no_spec[req_idx] = end_idx
            self.input_batch.num_tokens[req_idx] = end_idx
2658

2659
            req_id = req_ids[req_idx]
2660
2661
2662
            req_state = self.requests[req_id]
            req_state.output_token_ids.extend(sampled_ids)

2663
        logprobs_lists = (
2664
            logprobs_tensors.tolists(cu_num_tokens)
2665
            if not self.use_async_scheduling and logprobs_tensors is not None
2666
2667
2668
2669
2670
2671
2672
2673
2674
            else None
        )

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

2675
2676
2677
2678
2679
2680
2681
2682
2683
2684
        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,
        )

2685
2686
2687
2688
2689
2690
2691
2692
2693
2694
2695
2696
2697
2698
2699
    @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()

2700
2701
    def _model_forward(
        self,
2702
2703
2704
2705
        input_ids: torch.Tensor | None = None,
        positions: torch.Tensor | None = None,
        intermediate_tensors: IntermediateTensors | None = None,
        inputs_embeds: torch.Tensor | None = None,
2706
2707
2708
2709
2710
        **model_kwargs: dict[str, Any],
    ) -> Any:
        """Helper method to call the model forward pass.

        This method can be overridden by subclasses for model execution.
2711
        Motivation: We can inspect only this method versus
2712
2713
2714
2715
2716
2717
2718
2719
2720
2721
2722
2723
2724
2725
2726
2727
2728
2729
2730
2731
        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,
        )

2732
2733
2734
2735
2736
2737
2738
2739
2740
2741
2742
2743
2744
2745
    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,
    ) -> tuple[
2746
2747
2748
2749
2750
        CUDAGraphMode,
        BatchDescriptor,
        UBatchSlices | None,
        torch.Tensor | None,
        CUDAGraphStat | None,
2751
2752
2753
2754
2755
2756
2757
2758
2759
2760
2761
2762
2763
2764
2765
2766
2767
2768
2769
2770
2771
2772
2773
2774
2775
2776
2777
2778
2779
2780
2781
2782
2783
2784
2785
2786
2787
2788
2789
2790
2791
2792
2793
2794
2795
2796
2797
2798
2799
2800
2801
2802
2803
2804
2805
2806
2807
2808
2809
2810
2811
2812
2813
2814
    ]:
        num_tokens_padded = self._pad_for_sequence_parallelism(num_tokens)
        uniform_decode = (
            (
                (max_num_scheduled_tokens == self.uniform_decode_query_len)
                and (num_tokens_padded == max_num_scheduled_tokens * num_reqs)
            )
            if force_uniform_decode is None
            else force_uniform_decode
        )

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

        dispatch_cudagraph = (
            lambda num_tokens: self.cudagraph_dispatcher.dispatch(
                num_tokens=num_tokens,
                has_lora=has_lora,
                use_cascade_attn=use_cascade_attn,
                uniform_decode=uniform_decode,
            )
            if not force_eager
            else (CUDAGraphMode.NONE, BatchDescriptor(num_tokens_padded))
        )

        cudagraph_mode, batch_descriptor = dispatch_cudagraph(num_tokens_padded)
        num_tokens_padded = batch_descriptor.num_tokens

        # Extra coordination when running data-parallel since we need to coordinate
        # across ranks
        ubatch_slices, num_tokens_across_dp = None, None
        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
            )

            ubatch_slices, num_tokens_across_dp = coordinate_batch_across_dp(
                num_tokens_unpadded=num_tokens_padded,
                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,
            )

            # Extract DP padding if there is any
            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())

                # Re-dispatch with DP padding
                cudagraph_mode, batch_descriptor = dispatch_cudagraph(num_tokens_padded)
                # 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

2815
2816
2817
2818
2819
2820
2821
2822
2823
2824
2825
2826
2827
2828
2829
2830
        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,
            ubatch_slices,
            num_tokens_across_dp,
            cudagraph_stats,
        )
2831

2832
2833
2834
2835
    @torch.inference_mode()
    def execute_model(
        self,
        scheduler_output: "SchedulerOutput",
2836
        intermediate_tensors: IntermediateTensors | None = None,
2837
2838
2839
2840
2841
2842
    ) -> ModelRunnerOutput | IntermediateTensors | None:
        if self.execute_model_state is not None:
            raise RuntimeError(
                "State error: sample_tokens() must be called "
                "after execute_model() returns None."
            )
2843
2844
2845
2846
2847
2848
2849
2850
2851
2852
2853
2854
2855
2856
2857

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

2858
        num_scheduled_tokens = scheduler_output.total_num_scheduled_tokens
2859
        with record_function_or_nullcontext("gpu_model_runner: preprocess"):
2860
2861
2862
2863
            with self.synchronize_input_prep():
                # Update persistent batch states.
                self._update_states(scheduler_output)

2864
2865
2866
2867
2868
2869
2870
2871
                if has_ec_transfer() and get_ec_transfer().is_producer:
                    with self.maybe_get_ec_connector_output(
                        scheduler_output,
                        encoder_cache=self.encoder_cache,
                    ) as ec_connector_output:
                        self._execute_mm_encoder(scheduler_output)
                        return make_empty_encoder_model_runner_output(scheduler_output)

2872
                if not num_scheduled_tokens:
2873
2874
2875
2876
2877
2878
2879
2880
2881
2882
2883
                    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.
2884
                        self._dummy_run(1)
2885
2886
2887
2888
                    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(
2889
2890
                        scheduler_output, self.vllm_config
                    )
2891
                if self.cache_config.kv_sharing_fast_prefill:
2892
                    assert not self.num_prompt_logprobs, (
2893
2894
                        "--kv-sharing-fast-prefill produces incorrect "
                        "logprobs for prompt tokens, tokens, please disable "
2895
2896
                        "it when the requests need prompt logprobs"
                    )
2897

2898
2899
2900
2901
2902
                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())
2903
                num_tokens_unpadded = scheduler_output.total_num_scheduled_tokens
2904

2905
2906
2907
                (
                    logits_indices,
                    spec_decode_metadata,
2908
                ) = self._prepare_inputs(
2909
2910
                    scheduler_output,
                    num_scheduled_tokens_np,
2911
2912
2913
2914
                )

                cascade_attn_prefix_lens = None
                # Disable cascade attention when using microbatching (DBO)
2915
                if self.cascade_attn_enabled and not self.parallel_config.enable_dbo:
2916
2917
2918
                    # Pre-compute cascade attention prefix lengths
                    cascade_attn_prefix_lens = self._compute_cascade_attn_prefix_lens(
                        num_scheduled_tokens_np,
2919
                        self.input_batch.num_computed_tokens_cpu[:num_reqs],
2920
2921
2922
                        scheduler_output.num_common_prefix_blocks,
                    )

2923
2924
2925
2926
2927
                (
                    cudagraph_mode,
                    batch_desc,
                    ubatch_slices,
                    num_tokens_across_dp,
2928
                    cudagraph_stats,
2929
2930
2931
2932
2933
2934
2935
2936
2937
2938
2939
2940
2941
2942
2943
2944
2945
2946
2947
2948
2949
                ) = 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,
                )

                logger.debug(
                    "Running batch with cudagraph_mode: %s, batch_descriptor: %s, "
                    "ubatch_slices: %s, num_tokens_across_dp: %s",
                    cudagraph_mode,
                    batch_desc,
                    ubatch_slices,
                    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
                )
2950
2951

                use_spec_decode = len(scheduler_output.scheduled_spec_decode_tokens) > 0
2952
2953
2954
                pad_attn = cudagraph_mode == CUDAGraphMode.FULL

                (attn_metadata, spec_decode_common_attn_metadata) = (
2955
                    self._build_attention_metadata(
2956
2957
                        num_tokens=num_tokens_unpadded,
                        num_tokens_padded=num_tokens_padded if pad_attn else None,
2958
                        num_reqs=num_reqs,
2959
2960
                        num_reqs_padded=num_reqs_padded if pad_attn else None,
                        max_query_len=max_num_scheduled_tokens,
2961
2962
2963
                        ubatch_slices=ubatch_slices,
                        logits_indices=logits_indices,
                        use_spec_decode=use_spec_decode,
2964
                        num_scheduled_tokens=scheduler_output.num_scheduled_tokens,
2965
2966
2967
                        cascade_attn_prefix_lens=cascade_attn_prefix_lens,
                    )
                )
2968

2969
2970
2971
2972
2973
2974
2975
2976
2977
            (
                input_ids,
                inputs_embeds,
                positions,
                intermediate_tensors,
                model_kwargs,
                ec_connector_output,
            ) = self._preprocess(
                scheduler_output, num_tokens_padded, intermediate_tensors
2978
            )
2979

2980
        # Set cudagraph mode to none if calc_kv_scales is true.
2981
2982
2983
        # KV scales calculation involves dynamic operations that are incompatible
        # with CUDA graph capture.
        if self.calculate_kv_scales:
2984
            cudagraph_mode = CUDAGraphMode.NONE
2985
2986
            # Mark KV scales as calculated after the first forward pass
            self.calculate_kv_scales = False
2987

2988
2989
        # Run the model.
        # Use persistent buffers for CUDA graphs.
2990
2991
        with (
            set_forward_context(
2992
2993
                attn_metadata,
                self.vllm_config,
2994
                num_tokens=num_tokens_padded,
2995
                num_tokens_across_dp=num_tokens_across_dp,
2996
2997
                cudagraph_runtime_mode=cudagraph_mode,
                batch_descriptor=batch_desc,
2998
                ubatch_slices=ubatch_slices,
2999
            ),
3000
            record_function_or_nullcontext("gpu_model_runner: forward"),
3001
3002
            self.maybe_get_kv_connector_output(scheduler_output) as kv_connector_output,
        ):
3003
            model_output = self._model_forward(
3004
3005
3006
3007
3008
3009
3010
                input_ids=input_ids,
                positions=positions,
                intermediate_tensors=intermediate_tensors,
                inputs_embeds=inputs_embeds,
                **model_kwargs,
            )

3011
        with record_function_or_nullcontext("gpu_model_runner: postprocess"):
3012
            if self.use_aux_hidden_state_outputs:
3013
                # True when EAGLE 3 is used.
3014
3015
                hidden_states, aux_hidden_states = model_output
            else:
3016
                # Common case.
3017
3018
3019
                hidden_states = model_output
                aux_hidden_states = None

3020
3021
3022
3023
3024
            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)
3025
                    hidden_states.kv_connector_output = kv_connector_output
3026
                    self.kv_connector_output = kv_connector_output
3027
                    return hidden_states
3028

3029
                if self.is_pooling_model:
3030
                    # Return the pooling output.
3031
3032
3033
                    output = self._pool(
                        hidden_states, num_scheduled_tokens, num_scheduled_tokens_np
                    )
3034
3035
                    output.kv_connector_output = kv_connector_output
                    return output
3036
3037

                sample_hidden_states = hidden_states[logits_indices]
3038
                logits = self.model.compute_logits(sample_hidden_states)
3039
3040
3041
3042
            else:
                # Rare case.
                assert not self.is_pooling_model

3043
                sample_hidden_states = hidden_states[logits_indices]
3044
                if not get_pp_group().is_last_rank:
3045
                    all_gather_tensors = {
3046
                        "residual": not is_residual_scattered_for_sp(
3047
                            self.vllm_config, num_tokens_padded
3048
                        )
3049
                    }
3050
                    get_pp_group().send_tensor_dict(
3051
3052
                        hidden_states.tensors,
                        all_gather_group=get_tp_group(),
3053
3054
                        all_gather_tensors=all_gather_tensors,
                    )
3055
3056
                    logits = None
                else:
3057
                    logits = self.model.compute_logits(sample_hidden_states)
3058

3059
                model_output_broadcast_data: dict[str, Any] = {}
3060
3061
3062
                if logits is not None:
                    model_output_broadcast_data["logits"] = logits.contiguous()

3063
                broadcasted = get_pp_group().broadcast_tensor_dict(
3064
3065
                    model_output_broadcast_data, src=len(get_pp_group().ranks) - 1
                )
3066
3067
                assert broadcasted is not None
                logits = broadcasted["logits"]
3068

3069
3070
3071
3072
3073
3074
3075
3076
        self.execute_model_state = ExecuteModelState(
            scheduler_output,
            logits,
            spec_decode_metadata,
            spec_decode_common_attn_metadata,
            hidden_states,
            sample_hidden_states,
            aux_hidden_states,
3077
            ec_connector_output,
3078
            cudagraph_stats,
3079
        )
3080
        self.kv_connector_output = kv_connector_output
3081
3082
3083
3084
3085
3086
        return None

    @torch.inference_mode
    def sample_tokens(
        self, grammar_output: "GrammarOutput | None"
    ) -> ModelRunnerOutput | AsyncModelRunnerOutput | IntermediateTensors:
3087
3088
3089
        kv_connector_output = self.kv_connector_output
        self.kv_connector_output = None

3090
3091
        if self.execute_model_state is None:
            # Nothing to do (PP non-final rank case), output isn't used.
3092
            if not kv_connector_output:
3093
                return None  # type: ignore[return-value]
3094
3095
3096
3097
3098
3099
3100
3101
3102

            # 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
3103
3104
3105
3106
3107
3108
3109
3110
3111
3112

        # Unpack ephemeral state.
        (
            scheduler_output,
            logits,
            spec_decode_metadata,
            spec_decode_common_attn_metadata,
            hidden_states,
            sample_hidden_states,
            aux_hidden_states,
3113
            ec_connector_output,
3114
            cudagraph_stats,
3115
3116
3117
3118
3119
3120
3121
3122
3123
        ) = 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
            )
3124

3125
        with record_function_or_nullcontext("gpu_model_runner: sample"):
3126
3127
            sampler_output = self._sample(logits, spec_decode_metadata)

3128
3129
        self.input_batch.prev_sampled_token_ids = None

3130
        def propose_draft_token_ids(sampled_token_ids):
3131
            assert spec_decode_common_attn_metadata is not None
3132
            with record_function_or_nullcontext("gpu_model_runner: draft"):
3133
3134
3135
3136
3137
3138
3139
3140
3141
3142
3143
                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,
                )

3144
        spec_config = self.speculative_config
3145
        use_padded_batch_for_eagle = (
3146
3147
3148
            spec_config is not None
            and spec_config.use_eagle()
            and not spec_config.disable_padded_drafter_batch
3149
        )
3150
3151
3152
        effective_drafter_max_model_len = self.max_model_len
        if effective_drafter_max_model_len is None:
            effective_drafter_max_model_len = self.model_config.max_model_len
3153
        if (
3154
3155
3156
            spec_config is not None
            and spec_config.draft_model_config is not None
            and spec_config.draft_model_config.max_model_len is not None
3157
        ):
3158
            effective_drafter_max_model_len = (
3159
                spec_config.draft_model_config.max_model_len
3160
            )
3161
        input_fits_in_drafter = spec_decode_common_attn_metadata and (
3162
            spec_decode_common_attn_metadata.max_seq_len + self.num_spec_tokens
3163
3164
            <= effective_drafter_max_model_len
        )
3165
        if use_padded_batch_for_eagle:
3166
3167
            assert self.speculative_config is not None
            assert isinstance(self.drafter, EagleProposer)
3168
3169
3170
3171
3172
3173
            sampled_token_ids = sampler_output.sampled_token_ids
            if input_fits_in_drafter:
                # EAGLE speculative decoding can use the GPU sampled tokens
                # as inputs, and does not need to wait for bookkeeping to finish.
                propose_draft_token_ids(sampled_token_ids)
            elif self.valid_sampled_token_count_event is not None:
3174
                assert spec_decode_common_attn_metadata is not None
3175
3176
3177
3178
3179
3180
                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,
3181
                        self.discard_request_mask.gpu,
3182
3183
3184
3185
3186
                    )
                )
                self._copy_valid_sampled_token_count(
                    next_token_ids, valid_sampled_tokens_count
                )
3187

3188
        with record_function_or_nullcontext("gpu_model_runner: bookkeep"):
3189
3190
3191
3192
3193
3194
3195
3196
            (
                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,
3197
3198
3199
3200
3201
            ) = self._bookkeeping_sync(
                scheduler_output,
                sampler_output,
                logits,
                hidden_states,
3202
                scheduler_output.total_num_scheduled_tokens,
3203
                spec_decode_metadata,
3204
            )
3205

3206
3207
3208
3209
3210
        if (
            self.speculative_config
            and not use_padded_batch_for_eagle
            and input_fits_in_drafter
        ):
3211
3212
3213
            # 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)
3214

3215
        with record_function_or_nullcontext("gpu_model_runner: eplb"):
3216
            self.eplb_step()
3217
3218
3219
3220
3221
3222
3223
3224
3225
        with record_function_or_nullcontext("gpu_model_runner: ModelRunnerOutput"):
            output = ModelRunnerOutput(
                req_ids=req_ids_output_copy,
                req_id_to_index=req_id_to_index_output_copy,
                sampled_token_ids=valid_sampled_token_ids,
                logprobs=logprobs_lists,
                prompt_logprobs_dict=prompt_logprobs_dict,
                pooler_output=[],
                kv_connector_output=kv_connector_output,
3226
3227
3228
                ec_connector_output=ec_connector_output
                if self.supports_mm_inputs
                else None,
3229
                num_nans_in_logits=num_nans_in_logits,
3230
                cudagraph_stats=cudagraph_stats,
3231
            )
3232

3233
3234
        if not self.use_async_scheduling:
            return output
3235
3236
3237
3238
3239
3240
3241
3242
3243
        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,
3244
                vocab_size=self.input_batch.vocab_size,
3245
3246
3247
3248
3249
            )
        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
3250
            # any requests with sampling params that require output ids.
3251
3252
3253
3254
            self.input_batch.set_async_sampled_token_ids(
                async_output.sampled_token_ids_cpu,
                async_output.async_copy_ready_event,
            )
3255
3256
3257

        return async_output

3258
    def take_draft_token_ids(self) -> DraftTokenIds | None:
3259
3260
3261
3262
3263
3264
3265
3266
3267
3268
        if self._draft_token_ids is None:
            return None
        req_ids = self.input_batch.req_ids
        if isinstance(self._draft_token_ids, torch.Tensor):
            draft_token_ids = self._draft_token_ids.tolist()
        else:
            draft_token_ids = self._draft_token_ids
        self._draft_token_ids = None
        return DraftTokenIds(req_ids, draft_token_ids)

3269
3270
3271
3272
3273
3274
3275
3276
3277
3278
3279
3280
3281
3282
3283
3284
3285
3286
3287
3288
3289
3290
3291
3292
3293
3294
3295
3296
3297
3298
3299
    def _copy_valid_sampled_token_count(
        self, next_token_ids: torch.Tensor, valid_sampled_tokens_count: torch.Tensor
    ) -> None:
        if self.valid_sampled_token_count_event is None:
            return

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

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

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

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

3300
3301
3302
    def propose_draft_token_ids(
        self,
        scheduler_output: "SchedulerOutput",
3303
        sampled_token_ids: torch.Tensor | list[list[int]],
3304
3305
3306
        sampling_metadata: SamplingMetadata,
        hidden_states: torch.Tensor,
        sample_hidden_states: torch.Tensor,
3307
3308
        aux_hidden_states: list[torch.Tensor] | None,
        spec_decode_metadata: SpecDecodeMetadata | None,
3309
        common_attn_metadata: CommonAttentionMetadata,
3310
    ) -> list[list[int]] | torch.Tensor:
3311
        num_scheduled_tokens = scheduler_output.total_num_scheduled_tokens
3312
3313
3314
        spec_config = self.speculative_config
        assert spec_config is not None
        if spec_config.method == "ngram":
3315
            assert isinstance(sampled_token_ids, list)
3316
            assert isinstance(self.drafter, NgramProposer)
3317
            draft_token_ids = self.drafter.propose(
3318
3319
                sampled_token_ids,
                self.input_batch.req_ids,
3320
3321
                self.input_batch.num_tokens_no_spec,
                self.input_batch.token_ids_cpu,
3322
3323
                self.input_batch.spec_decode_unsupported_reqs,
            )
3324
        elif spec_config.method == "suffix":
3325
3326
3327
            assert isinstance(sampled_token_ids, list)
            assert isinstance(self.drafter, SuffixDecodingProposer)
            draft_token_ids = self.drafter.propose(self.input_batch, sampled_token_ids)
3328
        elif spec_config.method == "medusa":
3329
            assert isinstance(sampled_token_ids, list)
3330
            assert isinstance(self.drafter, MedusaProposer)
3331

3332
3333
            if sample_hidden_states.shape[0] == len(sampled_token_ids):
                # The input to the target model does not include draft tokens.
3334
3335
3336
3337
                hidden_states = sample_hidden_states
            else:
                indices = []
                offset = 0
3338
3339
3340
                assert spec_decode_metadata is not None, (
                    "No spec decode metadata for medusa"
                )
3341
                for num_draft, tokens in zip(
3342
3343
                    spec_decode_metadata.num_draft_tokens, sampled_token_ids
                ):
3344
                    indices.append(offset + len(tokens) - 1)
3345
                    offset += num_draft + 1
3346
                indices = torch.tensor(indices, device=self.device)
3347
3348
                hidden_states = sample_hidden_states[indices]

3349
            draft_token_ids = self.drafter.propose(
3350
3351
3352
                target_hidden_states=hidden_states,
                sampling_metadata=sampling_metadata,
            )
3353
        elif spec_config.use_eagle():
3354
            assert isinstance(self.drafter, EagleProposer)
3355

3356
            if spec_config.disable_padded_drafter_batch:
3357
3358
3359
                # 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.
3360
3361
                assert isinstance(sampled_token_ids, list), (
                    "sampled_token_ids should be a python list when"
3362
                    "padded-batch is disabled."
3363
                )
3364
                next_token_ids = self.drafter.prepare_next_token_ids_cpu(
3365
3366
3367
3368
3369
                    sampled_token_ids,
                    self.requests,
                    self.input_batch,
                    scheduler_output.num_scheduled_tokens,
                )
3370
3371
3372
3373
3374
            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.
3375
3376
                assert isinstance(sampled_token_ids, torch.Tensor), (
                    "sampled_token_ids should be a torch.Tensor when"
3377
                    "padded-batch is enabled."
3378
3379
                )
                next_token_ids, valid_sampled_tokens_count = (
3380
3381
3382
3383
3384
                    self.drafter.prepare_next_token_ids_padded(
                        common_attn_metadata,
                        sampled_token_ids,
                        self.requests,
                        self.input_batch,
3385
                        self.discard_request_mask.gpu,
3386
                    )
3387
                )
3388
3389
3390
                self._copy_valid_sampled_token_count(
                    next_token_ids, valid_sampled_tokens_count
                )
Jiayi Yao's avatar
Jiayi Yao committed
3391

3392
            if spec_decode_metadata is None:
3393
                token_indices_to_sample = None
3394
                # input_ids can be None for multimodal models.
3395
                target_token_ids = self.input_ids.gpu[:num_scheduled_tokens]
3396
                target_positions = self._get_positions(num_scheduled_tokens)
3397
                if self.use_aux_hidden_state_outputs:
Wentao Ye's avatar
Wentao Ye committed
3398
                    assert aux_hidden_states is not None
3399
                    target_hidden_states = torch.cat(
3400
3401
                        [h[:num_scheduled_tokens] for h in aux_hidden_states], dim=-1
                    )
3402
3403
                else:
                    target_hidden_states = hidden_states[:num_scheduled_tokens]
3404
            else:
3405
                if spec_config.disable_padded_drafter_batch:
3406
                    token_indices_to_sample = None
3407
3408
3409
3410
3411
                    common_attn_metadata, token_indices = self.drafter.prepare_inputs(
                        common_attn_metadata,
                        sampled_token_ids,
                        spec_decode_metadata.num_draft_tokens,
                    )
3412
3413
3414
3415
3416
3417
3418
3419
3420
                    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]
3421
                else:
3422
                    common_attn_metadata, token_indices_to_sample = (
3423
3424
3425
                        self.drafter.prepare_inputs_padded(
                            common_attn_metadata,
                            spec_decode_metadata,
3426
3427
3428
                            valid_sampled_tokens_count,
                        )
                    )
3429
3430
3431
3432
3433
3434
3435
3436
3437
3438
3439
                    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]
3440

3441
            if self.supports_mm_inputs:
3442
3443
3444
3445
3446
3447
                mm_embed_inputs = self._gather_mm_embeddings(
                    scheduler_output,
                    shift_computed_tokens=1,
                )
            else:
                mm_embed_inputs = None
3448

3449
            draft_token_ids = self.drafter.propose(
3450
3451
3452
3453
                target_token_ids=target_token_ids,
                target_positions=target_positions,
                target_hidden_states=target_hidden_states,
                next_token_ids=next_token_ids,
3454
                last_token_indices=token_indices_to_sample,
3455
                sampling_metadata=sampling_metadata,
3456
                common_attn_metadata=common_attn_metadata,
3457
                mm_embed_inputs=mm_embed_inputs,
3458
            )
3459

3460
        return draft_token_ids
3461

3462
3463
3464
    def update_config(self, overrides: dict[str, Any]) -> None:
        allowed_config_names = {"load_config", "model_config"}
        for config_name, config_overrides in overrides.items():
3465
3466
            assert config_name in allowed_config_names, (
                f"Config `{config_name}` not supported. "
3467
                f"Allowed configs: {allowed_config_names}"
3468
            )
3469
3470
3471
3472
            config = getattr(self, config_name)
            new_config = update_config(config, config_overrides)
            setattr(self, config_name, new_config)

3473
3474
3475
3476
3477
    def load_model(self, eep_scale_up: bool = False) -> None:
        """
        Args:
            eep_scale_up: the model loading is for elastic EP scale up.
        """
3478
3479
3480
3481
3482
        logger.info_once(
            "Starting to load model %s...",
            self.model_config.model,
            scope="global",
        )
3483
3484
3485
3486
3487
        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)
        )
3488

3489
3490
3491
        if self.parallel_config.enable_eplb:
            self.eplb_state = EplbState(self.parallel_config, self.device)
            eplb_models = 0
3492
        with DeviceMemoryProfiler() as m:
3493
            time_before_load = time.perf_counter()
3494
            model_loader = get_model_loader(self.load_config)
3495
            self.model = model_loader.load_model(
3496
3497
                vllm_config=self.vllm_config, model_config=self.model_config
            )
3498
            if self.lora_config:
3499
3500
3501
                self.model = self.load_lora_model(
                    self.model, self.vllm_config, self.device
                )
3502
            if hasattr(self, "drafter"):
3503
                logger.info_once("Loading drafter model...")
3504
                self.drafter.load_model(self.model)
3505
3506
3507
3508
3509
                if (
                    hasattr(self.drafter, "model")
                    and is_mixture_of_experts(self.drafter.model)
                    and self.parallel_config.enable_eplb
                ):
3510
3511
3512
                    spec_config = self.vllm_config.speculative_config
                    assert spec_config is not None
                    assert spec_config.draft_model_config is not None
3513
3514
                    logger.info_once(
                        "EPLB is enabled for drafter model %s.",
3515
                        spec_config.draft_model_config.model,
3516
3517
3518
3519
3520
3521
3522
3523
3524
3525
3526
3527
3528
3529
3530
3531
                    )

                    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,
3532
                        spec_config.draft_model_config,
3533
3534
3535
3536
3537
3538
                        global_expert_load,
                        old_global_expert_indices,
                        rank_mapping,
                    )
                    eplb_models += 1

3539
            if self.use_aux_hidden_state_outputs:
3540
                if not supports_eagle3(self.get_model()):
3541
3542
                    raise RuntimeError(
                        "Model does not support EAGLE3 interface but "
3543
3544
                        "aux_hidden_state_outputs was requested"
                    )
3545
3546
3547
3548
3549
3550
3551
3552
3553
3554
3555
3556
3557

                # 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)
3558
            time_after_load = time.perf_counter()
3559
        self.model_memory_usage = m.consumed_memory
3560
        logger.info_once(
3561
            "Model loading took %.4f GiB memory and %.6f seconds",
3562
3563
            self.model_memory_usage / GiB_bytes,
            time_after_load - time_before_load,
3564
            scope="local",
3565
        )
3566
        prepare_communication_buffer_for_model(self.model)
3567
3568
3569
3570
        if (drafter := getattr(self, "drafter", None)) and (
            drafter_model := getattr(drafter, "model", None)
        ):
            prepare_communication_buffer_for_model(drafter_model)
3571
        mm_config = self.model_config.multimodal_config
3572
        self.is_multimodal_pruning_enabled = (
3573
            supports_multimodal_pruning(self.get_model())
3574
3575
            and mm_config is not None
            and mm_config.is_multimodal_pruning_enabled()
3576
        )
3577

3578
        if is_mixture_of_experts(self.model) and self.parallel_config.enable_eplb:
3579
3580
3581
3582
3583
3584
3585
3586
3587
3588
3589
            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(
3590
                self.model,
3591
                self.model_config,
3592
3593
3594
                global_expert_load,
                old_global_expert_indices,
                rank_mapping,
3595
            )
3596
3597
            if self.eplb_state.is_async:
                self.eplb_state.start_async_loop(rank_mapping=rank_mapping)
3598

3599
        if (
3600
3601
            self.vllm_config.compilation_config.mode
            == CompilationMode.STOCK_TORCH_COMPILE
3602
            and supports_dynamo()
3603
        ):
3604
            backend = self.vllm_config.compilation_config.init_backend(self.vllm_config)
3605
            compilation_counter.stock_torch_compile_count += 1
3606
            self.model.compile(fullgraph=True, backend=backend)
3607
            return
3608
        # for other compilation modes, cudagraph behavior is controlled by
3609
3610
3611
        # CudagraphWraper and CudagraphDispatcher of vllm.

        # wrap the model with full cudagraph wrapper if needed.
3612
3613
3614
        cudagraph_mode = self.compilation_config.cudagraph_mode
        assert cudagraph_mode is not None
        if cudagraph_mode.has_full_cudagraphs() and not self.parallel_config.enable_dbo:
3615
3616
3617
            self.model = CUDAGraphWrapper(
                self.model, self.vllm_config, runtime_mode=CUDAGraphMode.FULL
            )
3618
        elif self.parallel_config.enable_dbo:
3619
            if cudagraph_mode.has_full_cudagraphs():
3620
3621
3622
                self.model = UBatchWrapper(
                    self.model, self.vllm_config, CUDAGraphMode.FULL, self.device
                )
3623
            else:
3624
3625
3626
                self.model = UBatchWrapper(
                    self.model, self.vllm_config, CUDAGraphMode.NONE, self.device
                )
3627

3628
    def _get_eagle3_aux_layers_from_config(self) -> tuple[int, ...] | None:
3629
3630
3631
3632
3633
3634
3635
3636
3637
3638
3639
3640
3641
3642
3643
3644
3645
3646
3647
3648
3649
3650
3651
        """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

3652
    def reload_weights(self) -> None:
3653
        assert getattr(self, "model", None) is not None, (
3654
            "Cannot reload weights before model is loaded."
3655
        )
3656
3657
        model_loader = get_model_loader(self.load_config)
        logger.info("Reloading weights inplace...")
3658
        model_loader.load_weights(self.get_model(), model_config=self.model_config)
3659

3660
3661
3662
3663
3664
    def save_tensorized_model(
        self,
        tensorizer_config: "TensorizerConfig",
    ) -> None:
        TensorizerLoader.save_model(
3665
            self.get_model(),
3666
            tensorizer_config=tensorizer_config,
3667
            model_config=self.model_config,
3668
3669
        )

3670
3671
3672
    def _get_prompt_logprobs_dict(
        self,
        hidden_states: torch.Tensor,
3673
        num_scheduled_tokens: dict[str, int],
3674
    ) -> dict[str, LogprobsTensors | None]:
3675
        num_prompt_logprobs_dict = self.num_prompt_logprobs
3676
3677
3678
        if not num_prompt_logprobs_dict:
            return {}

3679
        in_progress_dict = self.input_batch.in_progress_prompt_logprobs_cpu
3680
        prompt_logprobs_dict: dict[str, LogprobsTensors | None] = {}
3681
3682
3683
3684
3685

        # 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():
3686
3687
3688
3689
            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
3690
3691
3692

            # Get metadata for this request.
            request = self.requests[req_id]
3693
3694
3695
3696
            if request.prompt_token_ids is None:
                # Prompt logprobs is incompatible with prompt embeddings
                continue

3697
3698
            num_prompt_tokens = len(request.prompt_token_ids)
            prompt_token_ids = torch.tensor(request.prompt_token_ids).to(
3699
3700
                self.device, non_blocking=True
            )
3701

3702
3703
3704
3705
3706
3707
            # 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(
3708
3709
                    num_prompt_tokens - 1, num_prompt_logprobs + 1
                )
3710
3711
                in_progress_dict[req_id] = logprobs_tensors

3712
            # Determine number of logits to retrieve.
3713
3714
            start_idx = request.num_computed_tokens
            start_tok = start_idx + 1
3715
            num_remaining_tokens = num_prompt_tokens - start_tok
3716
            if num_tokens <= num_remaining_tokens:
3717
                # This is a chunk, more tokens remain.
3718
3719
3720
                # 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.
3721
3722
3723
3724
3725
                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)
3726
3727
3728
3729
3730
3731
3732
                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
3733
3734
3735
3736
3737

            # 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]
3738
            offset = self.query_start_loc.np[req_idx].item()
3739
            prompt_hidden_states = hidden_states[offset : offset + num_logits]
3740
            logits = self.model.compute_logits(prompt_hidden_states)
3741
3742
3743
3744

            # 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.
3745
            tgt_token_ids = prompt_token_ids[start_tok : start_tok + num_logits]
3746
3747

            # Compute prompt logprobs.
3748
3749
            logprobs = self.sampler.compute_logprobs(logits)
            token_ids, logprobs, ranks = self.sampler.gather_logprobs(
3750
3751
                logprobs, num_prompt_logprobs, tgt_token_ids
            )
3752
3753

            # Transfer GPU->CPU async.
3754
3755
            chunk_slice = slice(start_idx, start_idx + num_logits)
            logprobs_tensors.logprob_token_ids[chunk_slice].copy_(
3756
3757
3758
                token_ids, non_blocking=True
            )
            logprobs_tensors.logprobs[chunk_slice].copy_(logprobs, non_blocking=True)
3759
            logprobs_tensors.selected_token_ranks[chunk_slice].copy_(
3760
3761
                ranks, non_blocking=True
            )
3762
3763
3764
3765
3766

        # 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]
3767
            del in_progress_dict[req_id]
3768
3769

        # Must synchronize the non-blocking GPU->CPU transfers.
3770
        if prompt_logprobs_dict:
3771
            self._sync_device()
3772
3773
3774

        return prompt_logprobs_dict

3775
3776
    def _get_nans_in_logits(
        self,
3777
        logits: torch.Tensor | None,
3778
3779
3780
3781
3782
3783
3784
3785
3786
3787
3788
    ) -> 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])
3789
3790
3791
                    if num_nans_for_index is not None and req_index < logits.shape[0]
                    else 0
                )
3792
3793
3794
3795
            return num_nans_in_logits
        except IndexError:
            return {}

3796
3797
3798
3799
3800
3801
    @contextmanager
    def maybe_randomize_inputs(self, input_ids: torch.Tensor):
        """
        Randomize input_ids if VLLM_RANDOMIZE_DP_DUMMY_INPUTS is set.
        This is to help balance expert-selection
         - during profile_run
3802
         - during DP rank dummy run
3803
3804
3805
3806
3807
3808
3809
3810
3811
3812
3813
        """
        dp_size = self.vllm_config.parallel_config.data_parallel_size
        randomize_inputs = envs.VLLM_RANDOMIZE_DP_DUMMY_INPUTS and dp_size > 1
        if not randomize_inputs:
            yield
        else:
            import functools

            @functools.cache
            def rand_input_ids() -> torch.Tensor:
                return torch.randint_like(
3814
                    self.input_ids.gpu,
3815
3816
                    low=0,
                    high=self.model_config.get_vocab_size(),
3817
3818
                    dtype=input_ids.dtype,
                )
3819

3820
            logger.debug_once("Randomizing dummy data for DP Rank")
3821
            input_ids.copy_(rand_input_ids()[: input_ids.size(0)], non_blocking=True)
3822
3823
3824
            yield
            input_ids.fill_(0)

3825
3826
3827
3828
3829
3830
    def _get_mm_dummy_batch(
        self,
        modality: str,
        max_items_per_batch: int,
    ) -> BatchedTensorInputs:
        """Dummy data for profiling and precompiling multimodal models."""
3831
3832
        assert self.mm_budget is not None

3833
3834
        dummy_decoder_data = self.mm_registry.get_decoder_dummy_data(
            model_config=self.model_config,
3835
            seq_len=self.max_model_len,
3836
            mm_counts={modality: 1},
3837
            cache=self.mm_budget.cache,
3838
3839
3840
3841
        )
        dummy_mm_data = dummy_decoder_data.multi_modal_data

        # Result in the maximum GPU consumption of the model
3842
3843
        dummy_mm_item = dummy_mm_data[modality][0]
        dummy_mm_items = [dummy_mm_item] * max_items_per_batch
3844

3845
        model = cast(SupportsMultiModal, self.model)
3846
3847
3848
3849
3850
3851
3852
        return next(
            mm_kwargs_group
            for _, _, mm_kwargs_group in group_mm_kwargs_by_modality(
                dummy_mm_items,
                device=self.device,
                pin_memory=self.pin_memory,
                merge_by_field_config=model.merge_by_field_config,
3853
                multimodal_cpu_fields=model.multimodal_cpu_fields,
3854
3855
            )
        )
3856

3857
3858
3859
3860
    @torch.inference_mode()
    def _dummy_run(
        self,
        num_tokens: int,
3861
        cudagraph_runtime_mode: CUDAGraphMode | None = None,
3862
3863
        force_attention: bool = False,
        uniform_decode: bool = False,
3864
        allow_microbatching: bool = True,
3865
3866
        skip_eplb: bool = False,
        is_profile: bool = False,
3867
        create_mixed_batch: bool = False,
3868
        remove_lora: bool = True,
3869
        activate_lora: bool = False,
Rémi Delacourt's avatar
Rémi Delacourt committed
3870
        is_graph_capturing: bool = False,
3871
    ) -> tuple[torch.Tensor, torch.Tensor]:
3872
3873
3874
3875
3876
3877
3878
        """
        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.
3879
                - if not set will determine the cudagraph mode based on using
3880
                    the self.cudagraph_dispatcher.
3881
3882
3883
3884
                - CUDAGraphMode.NONE: No cudagraph, for warm up and profile run
                - CUDAGraphMode.PIECEWISE: Piecewise cudagraph.
                - CUDAGraphMode.FULL: Full cudagraph, attention metadata is
                    needed.
3885
            force_attention: If True, always create attention metadata. Used to
3886
3887
3888
3889
                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.
3890
3891
            create_mixed_batch: If True, create a mixed batch with both decode
                (1 token) and prefill (multiple tokens) requests.
3892
            remove_lora: If False, dummy LoRAs are not destroyed after the run
3893
            activate_lora: If False, dummy_run is performed without LoRAs.
3894
        """
3895
3896
3897
3898
        assert (
            cudagraph_runtime_mode is None
            or cudagraph_runtime_mode.valid_runtime_modes()
        )
3899

3900
        # If cudagraph_mode.decode_mode() == FULL and
3901
        # cudagraph_mode.separate_routine(). This means that we are using
3902
3903
3904
3905
3906
3907
3908
3909
3910
3911
3912
        # 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.
3913
        max_query_len = self.uniform_decode_query_len if uniform_decode else num_tokens
3914

3915
3916
3917
3918
3919
        # 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
3920
3921
3922
3923
        if create_mixed_batch:
            assert not uniform_decode
            # Create mixed batch:
            # first half decode tokens, second half one prefill
3924
            num_decode_tokens = min(max_num_reqs - 1, num_tokens // 2)
3925
3926
3927
3928
            num_prefill_tokens = num_tokens - num_decode_tokens
            num_reqs = num_decode_tokens + 1

            # Create decode requests (1 token each) followed by prefill request
3929
            num_scheduled_tokens_list = [1] * num_decode_tokens + [num_prefill_tokens]
3930
3931
3932
            # Note: Overriding max_query_len to be the prefill tokens
            max_query_len = num_prefill_tokens
        elif uniform_decode:
3933
            assert not create_mixed_batch
3934
            num_reqs = min(max_num_reqs, cdiv(num_tokens, max_query_len))
3935
3936
            num_scheduled_tokens_list = [max_query_len] * num_reqs
            if num_tokens % max_query_len != 0:
3937
                num_scheduled_tokens_list[-1] = num_tokens % max_query_len
3938
3939
3940
3941
3942
3943
        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

3944
3945
        assert sum(num_scheduled_tokens_list) == num_tokens
        assert len(num_scheduled_tokens_list) == num_reqs
3946
        num_scheduled_tokens = np.array(num_scheduled_tokens_list, dtype=np.int32)
3947
3948
        num_tokens_unpadded = int(num_scheduled_tokens.sum())

3949
        num_sampled_tokens = np.ones(num_reqs, dtype=np.int32)
3950

3951
        _cudagraph_mode, batch_desc, ubatch_slices, num_tokens_across_dp, _ = (
3952
3953
3954
3955
3956
3957
3958
3959
3960
3961
3962
3963
3964
3965
3966
3967
3968
3969
            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,
3970
3971
            )
        )
3972
3973
3974

        if cudagraph_runtime_mode is None:
            cudagraph_runtime_mode = _cudagraph_mode
3975
        else:
3976
3977
3978
3979
3980
3981
3982
3983
3984
            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
        )
3985

3986
        attn_metadata: PerLayerAttnMetadata | None = None
3987
3988
3989

        # If force_attention is True, we always capture attention. Otherwise,
        # it only happens for cudagraph_runtime_mode=FULL.
3990
        if force_attention or cudagraph_runtime_mode == CUDAGraphMode.FULL:
3991
3992
3993
3994
3995
3996
            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:
3997
                seq_lens = max_query_len  # type: ignore[assignment]
3998
            self.seq_lens.np[:num_reqs] = seq_lens
3999
4000
            self.seq_lens.np[num_reqs:] = 0
            self.seq_lens.copy_to_gpu()
4001

4002
4003
            cum_num_tokens, _ = self._get_cumsum_and_arange(num_scheduled_tokens)
            self.query_start_loc.np[1 : num_reqs + 1] = cum_num_tokens
4004
4005
            self.query_start_loc.copy_to_gpu()

4006
            attn_metadata, _ = self._build_attention_metadata(
4007
4008
4009
                num_tokens=num_tokens_unpadded,
                num_reqs=num_reqs_padded,
                max_query_len=max_query_len,
4010
                ubatch_slices=ubatch_slices,
4011
                for_cudagraph_capture=is_graph_capturing,
4012
            )
4013

4014
        with self.maybe_dummy_run_with_lora(
4015
4016
4017
4018
4019
            self.lora_config,
            num_scheduled_tokens,
            num_sampled_tokens,
            activate_lora,
            remove_lora,
4020
        ):
4021
            # Make sure padding doesn't exceed max_num_tokens
4022
4023
            assert num_tokens_padded <= self.max_num_tokens
            model_kwargs = self._init_model_kwargs(num_tokens_padded)
4024
            if self.supports_mm_inputs and not self.model_config.is_encoder_decoder:
4025
                input_ids = None
4026
                inputs_embeds = self.inputs_embeds.gpu[:num_tokens_padded]
4027
                model_kwargs = {
4028
                    **model_kwargs,
4029
4030
                    **self._dummy_mm_kwargs(num_reqs),
                }
4031
4032
            elif self.enable_prompt_embeds:
                input_ids = None
4033
4034
                inputs_embeds = self.inputs_embeds.gpu[:num_tokens_padded]
                model_kwargs = self._init_model_kwargs(num_tokens_padded)
4035
            else:
4036
                input_ids = self.input_ids.gpu[:num_tokens_padded]
4037
                inputs_embeds = None
4038

4039
            if self.uses_mrope:
4040
                positions = self.mrope_positions.gpu[:, :num_tokens_padded]
4041
            elif self.uses_xdrope_dim > 0:
4042
                positions = self.xdrope_positions.gpu[:, :num_tokens_padded]
4043
            else:
4044
                positions = self.positions.gpu[:num_tokens_padded]
4045
4046
4047
4048
4049
4050
4051
4052
4053

            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,
4054
4055
4056
                            device=self.device,
                        )
                    )
4057
4058

                intermediate_tensors = self.sync_and_slice_intermediate_tensors(
4059
                    num_tokens_padded, None, False
4060
                )
4061

4062
            if ubatch_slices is not None:
4063
4064
4065
                # Adjust values to reflect a single ubatch.
                # TODO(sage,lucas): this is cruft that should be addressed in
                #  the padding refactor.
4066
                num_tokens_padded = ubatch_slices[0].num_tokens
4067
                if num_tokens_across_dp is not None:
4068
                    num_tokens_across_dp[:] = num_tokens_padded
4069

4070
4071
4072
            with (
                self.maybe_randomize_inputs(input_ids),
                set_forward_context(
4073
4074
                    attn_metadata,
                    self.vllm_config,
4075
                    num_tokens=num_tokens_padded,
4076
4077
                    num_tokens_across_dp=num_tokens_across_dp,
                    cudagraph_runtime_mode=cudagraph_runtime_mode,
4078
                    batch_descriptor=batch_desc,
4079
4080
4081
                    ubatch_slices=ubatch_slices,
                ),
            ):
4082
                outputs = self.model(
4083
4084
4085
4086
                    input_ids=input_ids,
                    positions=positions,
                    intermediate_tensors=intermediate_tensors,
                    inputs_embeds=inputs_embeds,
4087
                    **model_kwargs,
4088
                )
4089

4090
4091
4092
4093
            if self.use_aux_hidden_state_outputs:
                hidden_states, _ = outputs
            else:
                hidden_states = outputs
4094

4095
            if self.speculative_config and self.speculative_config.use_eagle():
4096
                assert isinstance(self.drafter, EagleProposer)
4097
                use_cudagraphs = (
Rémi Delacourt's avatar
Rémi Delacourt committed
4098
                    cudagraph_runtime_mode.has_mode(CUDAGraphMode.PIECEWISE)
4099
4100
                    and not self.speculative_config.enforce_eager
                )
4101
4102
4103
4104
4105
4106
4107
4108
4109
4110
4111

                # 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
4112
                    is_graph_capturing=is_graph_capturing,
4113
                )
4114

4115
4116
4117
4118
4119
4120
4121
4122
4123
4124
        # 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)

4125
        logit_indices = np.cumsum(num_scheduled_tokens) - 1
4126
4127
4128
4129
        logit_indices_device = torch.from_numpy(logit_indices).to(
            self.device, non_blocking=True
        )
        return hidden_states, hidden_states[logit_indices_device]
4130
4131
4132
4133
4134
4135

    @torch.inference_mode()
    def _dummy_sampler_run(
        self,
        hidden_states: torch.Tensor,
    ) -> torch.Tensor:
4136
4137
4138
4139
        # The dummy hidden states may contain special values,
        # like `inf` or `nan`.
        # To avoid breaking the sampler, we use a random tensor here instead.
        hidden_states = torch.rand_like(hidden_states)
4140

4141
        logits = self.model.compute_logits(hidden_states)
4142
4143
        num_reqs = logits.size(0)

4144
        dummy_tensors = lambda v: torch.full((num_reqs,), v, device=self.device)
4145
4146
4147
4148
4149
4150
4151
4152
4153
4154
4155
4156
4157
4158
4159

        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)],
4160
            spec_token_ids=[[] for _ in range(num_reqs)],
4161
4162
            allowed_token_ids_mask=None,
            bad_words_token_ids={},
4163
            logitsprocs=LogitsProcessors(),
4164
        )
4165
        try:
4166
4167
4168
            sampler_output = self.sampler(
                logits=logits, sampling_metadata=dummy_metadata
            )
4169
        except RuntimeError as e:
4170
            if "out of memory" in str(e):
4171
4172
4173
4174
                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 "
4175
4176
                    "initializing the engine."
                ) from e
4177
4178
            else:
                raise e
4179
        if self.speculative_config:
4180
4181
            draft_token_ids = [[0] for _ in range(num_reqs)]
            dummy_spec_decode_metadata = SpecDecodeMetadata.make_dummy(
4182
4183
                draft_token_ids, self.device
            )
4184
4185
4186
4187
4188
4189

            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
4190
4191
4192
4193
4194
            logits = torch.randn(
                num_tokens + num_reqs,
                logits.shape[-1],
                device=self.device,
                dtype=logits.dtype,
4195
            )
4196
4197
4198
            self.rejection_sampler(
                dummy_spec_decode_metadata,
                draft_probs,
4199
                logits,
4200
4201
                dummy_metadata,
            )
4202
        return sampler_output
4203

4204
    def _dummy_pooler_run_task(
4205
4206
        self,
        hidden_states: torch.Tensor,
4207
4208
        task: PoolingTask,
    ) -> PoolerOutput:
4209
4210
4211
4212
4213
4214
4215
4216
4217
4218
4219
        num_tokens = hidden_states.shape[0]
        max_num_reqs = self.scheduler_config.max_num_seqs
        num_reqs = min(num_tokens, max_num_reqs)
        min_tokens_per_req = num_tokens // num_reqs
        num_scheduled_tokens_list = [min_tokens_per_req] * num_reqs
        num_scheduled_tokens_list[-1] += num_tokens % num_reqs
        assert sum(num_scheduled_tokens_list) == num_tokens
        assert len(num_scheduled_tokens_list) == num_reqs

        req_num_tokens = num_tokens // num_reqs

4220
        dummy_prompt_lens = torch.tensor(
4221
4222
            num_scheduled_tokens_list,
            device="cpu",
4223
        )
4224
4225
4226
        dummy_token_ids = torch.zeros(
            (num_reqs, req_num_tokens), dtype=torch.int32, device=self.device
        )
4227

4228
        model = cast(VllmModelForPooling, self.get_model())
4229
        dummy_pooling_params = PoolingParams(task=task)
4230
        dummy_pooling_params.verify(task=task, model_config=self.model_config)
4231
        to_update = model.pooler.get_pooling_updates(task)
4232
4233
        to_update.apply(dummy_pooling_params)

4234
        dummy_metadata = PoolingMetadata(
4235
4236
4237
            prompt_lens=dummy_prompt_lens,
            prompt_token_ids=dummy_token_ids,
            pooling_params=[dummy_pooling_params] * num_reqs,
4238
            pooling_states=[PoolingStates() for i in range(num_reqs)],
4239
        )
4240

4241
        dummy_metadata.build_pooling_cursor(
4242
4243
4244
            num_scheduled_tokens_list,
            seq_lens_cpu=dummy_prompt_lens,
            device=hidden_states.device,
4245
        )
4246

4247
        try:
4248
4249
4250
            return model.pooler(
                hidden_states=hidden_states, pooling_metadata=dummy_metadata
            )
4251
        except RuntimeError as e:
4252
            if "out of memory" in str(e):
4253
                raise RuntimeError(
4254
4255
4256
                    "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 "
4257
4258
                    "initializing the engine."
                ) from e
4259
4260
            else:
                raise e
4261
4262
4263
4264
4265
4266
4267

    @torch.inference_mode()
    def _dummy_pooler_run(
        self,
        hidden_states: torch.Tensor,
    ) -> PoolerOutput:
        # Find the task that has the largest output for subsequent steps
4268
4269
4270
        supported_pooling_tasks = self.get_supported_pooling_tasks()

        if not supported_pooling_tasks:
4271
4272
4273
4274
4275
4276
            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."
            )
4277

4278
        output_size = dict[PoolingTask, float]()
4279
        for task in supported_pooling_tasks:
4280
4281
            # Run a full batch with each task to ensure none of them OOMs
            output = self._dummy_pooler_run_task(hidden_states, task)
4282
            output_size[task] = sum(o.nbytes for o in output)
4283
4284
4285
4286
            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)
4287

4288
    def profile_run(self) -> None:
4289
        # Profile with multimodal encoder & encoder cache.
4290
        if self.supports_mm_inputs:
4291
4292
            mm_config = self.model_config.multimodal_config
            if mm_config is not None and mm_config.skip_mm_profiling:
4293
                logger.info(
4294
                    "Skipping memory profiling for multimodal encoder and "
4295
4296
                    "encoder cache."
                )
4297
4298
4299
4300
4301
4302
4303
4304
            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.
4305
                    dummy_modality = mm_budget.get_modality_with_max_tokens()
4306
4307
4308
                    max_mm_items_per_batch = mm_budget.max_items_per_batch_by_modality[
                        dummy_modality
                    ]
4309
4310
4311
4312
4313
4314
4315
4316
4317

                    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,
                    )
4318

4319
4320
4321
4322
4323
                    # Create dummy batch of multimodal inputs.
                    batched_dummy_mm_inputs = self._get_mm_dummy_batch(
                        dummy_modality,
                        max_mm_items_per_batch,
                    )
4324

4325
                    # Run multimodal encoder.
4326
                    dummy_encoder_outputs = self.model.embed_multimodal(
4327
4328
                        **batched_dummy_mm_inputs
                    )
4329

4330
4331
4332
4333
                    sanity_check_mm_encoder_outputs(
                        dummy_encoder_outputs,
                        expected_num_items=max_mm_items_per_batch,
                    )
4334

4335
4336
4337
                    # NOTE: This happens when encoder cache needs to store
                    # the embeddings that encoder outputs are scattered onto.
                    # In this case we create dummy embeddings of size
4338
4339
                    # (max_tokens_for_modality, hidden_size) and scatter
                    # encoder output into it.
4340
                    encoder_output_shape = dummy_encoder_outputs[0].shape
4341
4342
4343
4344
4345
                    max_mm_tokens_per_item = mm_budget.max_tokens_by_modality[
                        dummy_modality
                    ]
                    if encoder_output_shape[0] < max_mm_tokens_per_item:
                        encoder_hidden_size = encoder_output_shape[-1]
4346
4347
4348
                        expanded_outputs = []
                        for output in dummy_encoder_outputs:
                            expanded = output.new_zeros(
4349
                                (max_mm_tokens_per_item, encoder_hidden_size)
4350
                            )
4351
4352
4353
4354
4355
4356
                            num_tokens = output.shape[0]
                            expanded[:num_tokens].copy_(output)
                            expanded_outputs.append(expanded)

                        dummy_encoder_outputs = expanded_outputs

4357
                    # Cache the dummy encoder outputs.
4358
                    self.encoder_cache["tmp"] = dict(enumerate(dummy_encoder_outputs))
4359

4360
        # Add `is_profile` here to pre-allocate communication buffers
4361
4362
4363
        hidden_states, last_hidden_states = self._dummy_run(
            self.max_num_tokens, is_profile=True
        )
4364
        if get_pp_group().is_last_rank:
4365
4366
4367
4368
            if self.is_pooling_model:
                output = self._dummy_pooler_run(hidden_states)
            else:
                output = self._dummy_sampler_run(last_hidden_states)
4369
        else:
4370
            output = None
4371
        self._sync_device()
4372
        del hidden_states, output
4373
        self.encoder_cache.clear()
4374
        gc.collect()
4375

4376
    def capture_model(self) -> int:
4377
        if self.compilation_config.cudagraph_mode == CUDAGraphMode.NONE:
4378
            logger.warning(
4379
                "Skipping CUDA graph capture. To turn on CUDA graph capture, "
4380
4381
                "ensure `cudagraph_mode` was not manually set to `NONE`"
            )
4382
            return 0
4383

4384
4385
        compilation_counter.num_gpu_runner_capture_triggers += 1

4386
4387
        start_time = time.perf_counter()

4388
4389
4390
4391
4392
4393
4394
4395
4396
4397
4398
4399
4400
4401
        @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()
4402
                    gc.collect()
4403

4404
4405
4406
        # 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.
4407
        set_cudagraph_capturing_enabled(True)
4408
        with freeze_gc(), graph_capture(device=self.device):
4409
            start_free_gpu_memory = torch.cuda.mem_get_info()[0]
4410
            cudagraph_mode = self.compilation_config.cudagraph_mode
4411
            assert cudagraph_mode is not None
4412
4413
4414
4415
4416
4417
4418
4419
4420

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

4421
4422
            if cudagraph_mode.mixed_mode() != CUDAGraphMode.NONE:
                cudagraph_runtime_mode = cudagraph_mode.mixed_mode()
4423
                # make sure we capture the largest batch size first
4424
4425
4426
                compilation_cases = list(
                    product(reversed(self.cudagraph_batch_sizes), lora_cases)
                )
4427
4428
4429
                self._capture_cudagraphs(
                    compilation_cases,
                    cudagraph_runtime_mode=cudagraph_runtime_mode,
4430
4431
                    uniform_decode=False,
                )
4432

4433
4434
            # Capture full cudagraph for uniform decode batches if we
            # don't already have full mixed prefill-decode cudagraphs.
4435
4436
4437
4438
4439
4440
4441
            if (
                cudagraph_mode.decode_mode() == CUDAGraphMode.FULL
                and cudagraph_mode.separate_routine()
            ):
                max_num_tokens = (
                    self.scheduler_config.max_num_seqs * self.uniform_decode_query_len
                )
4442
                decode_cudagraph_batch_sizes = [
4443
4444
                    x
                    for x in self.cudagraph_batch_sizes
4445
                    if max_num_tokens >= x >= self.uniform_decode_query_len
4446
                ]
4447
4448
4449
                compilation_cases_decode = list(
                    product(reversed(decode_cudagraph_batch_sizes), lora_cases)
                )
4450
4451
4452
                self._capture_cudagraphs(
                    compilation_cases=compilation_cases_decode,
                    cudagraph_runtime_mode=CUDAGraphMode.FULL,
4453
4454
                    uniform_decode=True,
                )
4455

4456
4457
4458
            torch.cuda.synchronize()
            end_free_gpu_memory = torch.cuda.mem_get_info()[0]

4459
4460
4461
        # 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
4462
        # we may do lazy capturing in future that still allows capturing
4463
4464
        # after here.
        set_cudagraph_capturing_enabled(False)
4465
4466
4467
4468
4469

        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.
4470
        logger.info_once(
4471
4472
4473
            "Graph capturing finished in %.0f secs, took %.2f GiB",
            elapsed_time,
            cuda_graph_size / (1 << 30),
4474
            scope="local",
4475
        )
4476
        return cuda_graph_size
4477

4478
4479
    def _capture_cudagraphs(
        self,
4480
        compilation_cases: list[tuple[int, bool]],
4481
4482
4483
4484
4485
4486
4487
        cudagraph_runtime_mode: CUDAGraphMode,
        uniform_decode: bool,
    ):
        assert (
            cudagraph_runtime_mode != CUDAGraphMode.NONE
            and cudagraph_runtime_mode.valid_runtime_modes()
        ), f"Invalid cudagraph runtime mode: {cudagraph_runtime_mode}"
4488
4489
4490
4491
4492
4493
4494
4495

        # Only rank 0 should print progress bar during capture
        if is_global_first_rank():
            compilation_cases = tqdm(
                compilation_cases,
                disable=not self.load_config.use_tqdm_on_load,
                desc="Capturing CUDA graphs ({}, {})".format(
                    "decode" if uniform_decode else "mixed prefill-decode",
4496
4497
4498
                    cudagraph_runtime_mode.name,
                ),
            )
4499

4500
        # We skip EPLB here since we don't want to record dummy metrics
4501
        for num_tokens, activate_lora in compilation_cases:
4502
            # We currently only capture ubatched graphs when its a FULL
4503
4504
4505
            # 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
4506
4507
4508
4509
            allow_microbatching = (
                self.parallel_config.enable_dbo
                and cudagraph_runtime_mode == CUDAGraphMode.FULL
                and uniform_decode
4510
4511
4512
4513
4514
                and check_ubatch_thresholds(
                    config=self.vllm_config.parallel_config,
                    num_tokens=num_tokens,
                    uniform_decode=uniform_decode,
                )
4515
            )
4516

4517
4518
4519
4520
4521
4522
            for _ in range(self.compilation_config.cudagraph_num_of_warmups):
                # Use CUDAGraphRuntimeStyle.NONE (default) for warmup.
                # But be careful, warm up with `NONE`is orthogonal to
                # if we want to warm up attention or not. This is
                # different from the case where `FULL` implies capture
                # attention while `PIECEWISE` implies no attention.
4523
4524
4525
4526
4527
4528
4529
4530
4531
                force_attention = cudagraph_runtime_mode == CUDAGraphMode.FULL
                self._dummy_run(
                    num_tokens,
                    cudagraph_runtime_mode=CUDAGraphMode.NONE,
                    force_attention=force_attention,
                    uniform_decode=uniform_decode,
                    allow_microbatching=allow_microbatching,
                    skip_eplb=True,
                    remove_lora=False,
4532
                    activate_lora=activate_lora,
4533
4534
4535
4536
4537
4538
4539
4540
                )
            self._dummy_run(
                num_tokens,
                cudagraph_runtime_mode=cudagraph_runtime_mode,
                uniform_decode=uniform_decode,
                allow_microbatching=allow_microbatching,
                skip_eplb=True,
                remove_lora=False,
4541
                activate_lora=activate_lora,
Rémi Delacourt's avatar
Rémi Delacourt committed
4542
                is_graph_capturing=True,
4543
            )
4544
        self.maybe_remove_all_loras(self.lora_config)
4545

4546
4547
4548
4549
    def initialize_attn_backend(self, kv_cache_config: KVCacheConfig) -> None:
        """
        Initialize the attention backends and attention metadata builders.
        """
4550
        assert len(self.attn_groups) == 0, "Attention backends are already initialized"
4551

4552
4553
4554
4555
4556
4557
        class AttentionGroupKey(NamedTuple):
            attn_backend: type[AttentionBackend]
            kv_cache_spec: KVCacheSpec

        def get_attn_backends_for_group(
            kv_cache_group_spec: KVCacheGroupSpec,
4558
        ) -> tuple[dict[AttentionGroupKey, list[str]], set[type[AttentionBackend]]]:
4559
            layer_type = cast(type[Any], AttentionLayerBase)
4560
            layers = get_layers_from_vllm_config(
4561
                self.vllm_config, layer_type, kv_cache_group_spec.layer_names
4562
            )
4563
4564
            attn_backends = {}
            attn_backend_layers = defaultdict(list)
4565
            # Dedupe based on full class name; this is a bit safer than
4566
4567
4568
4569
            # 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.
4570
            for layer_name in kv_cache_group_spec.layer_names:
4571
                attn_backend = layers[layer_name].get_attn_backend()
4572
4573
4574
4575

                if layer_name in self.kv_sharing_fast_prefill_eligible_layers:
                    attn_backend = create_fast_prefill_custom_backend(
                        "FastPrefill",
4576
                        attn_backend,  # type: ignore[arg-type]
4577
4578
                    )

4579
4580
4581
                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):
4582
                    layer_kv_cache_spec = layer_kv_cache_spec.kv_cache_specs[layer_name]
4583
                key = (full_cls_name, layer_kv_cache_spec)
4584
4585
4586
                attn_backends[key] = AttentionGroupKey(
                    attn_backend, layer_kv_cache_spec
                )
4587
                attn_backend_layers[key].append(layer_name)
4588
4589
4590
4591
            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()),
            )
4592
4593

        def create_attn_groups(
4594
            attn_backends_map: dict[AttentionGroupKey, list[str]],
4595
            kv_cache_group_id: int,
4596
4597
        ) -> list[AttentionGroup]:
            attn_groups: list[AttentionGroup] = []
4598
            for (attn_backend, kv_cache_spec), layer_names in attn_backends_map.items():
4599
                attn_group = AttentionGroup(
4600
                    attn_backend,
4601
                    layer_names,
4602
                    kv_cache_spec,
4603
                    kv_cache_group_id,
4604
4605
                )

4606
4607
4608
                attn_groups.append(attn_group)
            return attn_groups

4609
        attention_backend_maps = []
4610
        attention_backend_list = []
4611
        for kv_cache_group_spec in kv_cache_config.kv_cache_groups:
4612
            attn_backends = get_attn_backends_for_group(kv_cache_group_spec)
4613
            attention_backend_maps.append(attn_backends[0])
4614
            attention_backend_list.append(attn_backends[1])
4615
4616

        # Resolve cudagraph_mode before actually initialize metadata_builders
4617
4618
4619
        self._check_and_update_cudagraph_mode(
            attention_backend_list, kv_cache_config.kv_cache_groups
        )
4620

4621
4622
        for i, attn_backend_map in enumerate(attention_backend_maps):
            self.attn_groups.append(create_attn_groups(attn_backend_map, i))
4623

4624
4625
4626
4627
4628
4629
4630
4631
4632
4633
4634
4635
4636
4637
4638
4639
4640
4641
    def initialize_metadata_builders(
        self, kv_cache_config: KVCacheConfig, kernel_block_sizes: list[int]
    ) -> None:
        """
        Create the metadata builders for all KV cache groups and attn groups.
        """
        for kv_cache_group_id in range(len(kv_cache_config.kv_cache_groups)):
            for attn_group in self.attn_groups[kv_cache_group_id]:
                attn_group.create_metadata_builders(
                    self.vllm_config,
                    self.device,
                    kernel_block_sizes[kv_cache_group_id]
                    if kv_cache_group_id < len(kernel_block_sizes)
                    else None,
                    num_metadata_builders=1
                    if not self.parallel_config.enable_dbo
                    else 2,
                )
co63oc's avatar
co63oc committed
4642
        # Calculate reorder batch threshold (if needed)
4643
4644
        # Note (tdoublep): do this *after* constructing builders,
        # because some of them change the threshold at init time.
4645
4646
        self.calculate_reorder_batch_threshold()

4647
    def _check_and_update_cudagraph_mode(
4648
4649
4650
        self,
        attention_backends: list[set[type[AttentionBackend]]],
        kv_cache_groups: list[KVCacheGroupSpec],
4651
    ) -> None:
4652
        """
4653
        Resolve the cudagraph_mode when there are multiple attention
4654
        groups with potential conflicting CUDA graph support.
4655
4656
4657
        Then initialize the cudagraph_dispatcher based on the resolved
        cudagraph_mode.
        """
4658
        min_cg_support = AttentionCGSupport.ALWAYS
4659
        min_cg_backend_name = None
4660

4661
4662
4663
4664
4665
4666
4667
4668
4669
4670
4671
4672
        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__
4673
4674
        # Flexible resolve the cudagraph mode
        cudagraph_mode = self.compilation_config.cudagraph_mode
4675
        assert cudagraph_mode is not None
4676
        # check cudagraph for mixed batch is supported
4677
4678
4679
4680
4681
4682
        if (
            cudagraph_mode.mixed_mode() == CUDAGraphMode.FULL
            and min_cg_support != AttentionCGSupport.ALWAYS
        ):
            msg = (
                f"CUDAGraphMode.{cudagraph_mode.name} is not supported "
4683
                f"with {min_cg_backend_name} backend (support: "
4684
4685
                f"{min_cg_support})"
            )
4686
4687
            if min_cg_support == AttentionCGSupport.NEVER:
                # if not supported any full cudagraphs, just raise it.
4688
4689
                msg += (
                    "; please try cudagraph_mode=PIECEWISE, and "
4690
                    "make sure compilation mode is VLLM_COMPILE"
4691
                )
4692
4693
4694
4695
4696
                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"
4697
                cudagraph_mode = self.compilation_config.cudagraph_mode = (
4698
                    CUDAGraphMode.FULL_AND_PIECEWISE
4699
                )
4700
4701
            else:
                msg += "; setting cudagraph_mode=FULL_DECODE_ONLY"
4702
                cudagraph_mode = self.compilation_config.cudagraph_mode = (
4703
                    CUDAGraphMode.FULL_DECODE_ONLY
4704
                )
4705
4706
            logger.warning(msg)

4707
        # check that if we are doing decode full-cudagraphs it is supported
4708
4709
4710
4711
4712
4713
        if (
            cudagraph_mode.decode_mode() == CUDAGraphMode.FULL
            and min_cg_support == AttentionCGSupport.NEVER
        ):
            msg = (
                f"CUDAGraphMode.{cudagraph_mode.name} is not supported "
4714
                f"with {min_cg_backend_name} backend (support: "
4715
4716
                f"{min_cg_support})"
            )
4717
            if self.compilation_config.mode == CompilationMode.VLLM_COMPILE and (
4718
4719
4720
4721
4722
                self.compilation_config.splitting_ops_contain_attention()
                or self.compilation_config.use_inductor_graph_partition
            ):
                msg += (
                    "; setting cudagraph_mode=PIECEWISE because "
4723
                    "attention is compiled piecewise"
4724
4725
                )
                cudagraph_mode = self.compilation_config.cudagraph_mode = (
4726
                    CUDAGraphMode.PIECEWISE
4727
                )
4728
            else:
4729
4730
                msg += (
                    "; setting cudagraph_mode=NONE because "
4731
                    "attention is not compiled piecewise"
4732
4733
                )
                cudagraph_mode = self.compilation_config.cudagraph_mode = (
4734
                    CUDAGraphMode.NONE
4735
                )
4736
4737
            logger.warning(msg)

4738
4739
        # check that if we are doing spec-decode + decode full-cudagraphs it is
        # supported
4740
4741
4742
4743
4744
4745
4746
4747
        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 "
4748
                f"{min_cg_backend_name} (support: {min_cg_support})"
4749
            )
4750
4751
            if self.compilation_config.splitting_ops_contain_attention():
                msg += "; setting cudagraph_mode=PIECEWISE"
4752
                cudagraph_mode = self.compilation_config.cudagraph_mode = (
4753
                    CUDAGraphMode.PIECEWISE
4754
                )
4755
4756
            else:
                msg += "; setting cudagraph_mode=NONE"
4757
                cudagraph_mode = self.compilation_config.cudagraph_mode = (
4758
                    CUDAGraphMode.NONE
4759
                )
4760
4761
4762
4763
            logger.warning(msg)

        # double check that we can support full cudagraph if they are requested
        # even after automatic downgrades
4764
4765
4766
4767
4768
4769
        if (
            cudagraph_mode.has_full_cudagraphs()
            and min_cg_support == AttentionCGSupport.NEVER
        ):
            raise ValueError(
                f"CUDAGraphMode.{cudagraph_mode.name} is not "
4770
                f"supported with {min_cg_backend_name} backend ("
4771
4772
                f"support:{min_cg_support}) "
                "; please try cudagraph_mode=PIECEWISE, "
4773
                "and make sure compilation mode is VLLM_COMPILE"
4774
            )
4775

4776
4777
4778
4779
4780
4781
4782
4783
4784
4785
4786
4787
4788
4789
        # 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
        # Will be removed in the near future when we have seperate cudagraph capture
        # 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
            )
4790
4791
4792
4793
            capture_sizes = self.compilation_config.cudagraph_capture_sizes
            self.cudagraph_batch_sizes = (
                capture_sizes if capture_sizes is not None else []
            )
4794

4795
4796
        # Trigger cudagraph dispatching keys initialization after
        # resolved cudagraph mode.
4797
        self.compilation_config.cudagraph_mode = cudagraph_mode
4798
        self.cudagraph_dispatcher.initialize_cudagraph_keys(
4799
            cudagraph_mode, self.uniform_decode_query_len
4800
        )
4801

4802
4803
    def calculate_reorder_batch_threshold(self) -> None:
        """
4804
4805
4806
4807
        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.
4808
        """
4809
4810
        min_none_high = lambda a, b: a if b is None else b if a is None else min(a, b)

4811
        reorder_batch_thresholds: list[int | None] = [
4812
4813
4814
            group.get_metadata_builder().reorder_batch_threshold
            for group in self._attn_group_iterator()
        ]
4815
4816
4817
4818
4819
        # 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
4820
        self.reorder_batch_threshold = reduce(min_none_high, reorder_batch_thresholds)  # type: ignore[assignment]
4821

4822
4823
4824
    @staticmethod
    def select_common_block_size(
        kv_manager_block_size: int, attn_groups: list[AttentionGroup]
4825
4826
    ) -> int:
        """
4827
4828
4829
4830
4831
        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.
4832
4833
4834
4835
4836
4837

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

        Returns:
4838
            The selected block size
4839
4840

        Raises:
4841
            ValueError: If no valid block size found
4842
4843
        """

4844
4845
4846
4847
4848
4849
4850
4851
        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
4852
                for supported_size in backend.get_supported_kernel_block_sizes():
4853
4854
4855
4856
4857
4858
4859
4860
4861
4862
4863
4864
4865
4866
4867
4868
4869
4870
4871
4872
4873
4874
4875
4876
4877
4878
4879
4880
4881
4882
                    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
4883
            for supported_size in backend.get_supported_kernel_block_sizes()
4884
4885
            if isinstance(supported_size, int)
        )
4886

4887
4888
4889
4890
4891
4892
        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}. ")
4893

4894
4895
4896
    def may_reinitialize_input_batch(
        self, kv_cache_config: KVCacheConfig, kernel_block_sizes: list[int]
    ) -> None:
4897
4898
4899
4900
4901
4902
4903
        """
        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.
4904
            kernel_block_sizes: The kernel block sizes for each KV cache group.
4905
4906
4907
4908
        """
        block_sizes = [
            kv_cache_group.kv_cache_spec.block_size
            for kv_cache_group in kv_cache_config.kv_cache_groups
4909
            if not isinstance(kv_cache_group.kv_cache_spec, EncoderOnlyAttentionSpec)
4910
        ]
4911
4912
4913
4914

        if block_sizes != [self.cache_config.block_size] or kernel_block_sizes != [
            self.cache_config.block_size
        ]:
4915
4916
4917
            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
4918
4919
                "for more details."
            )
4920
4921
            self.input_batch = InputBatch(
                max_num_reqs=self.max_num_reqs,
4922
                max_model_len=max(self.max_model_len, self.max_encoder_len),
4923
4924
4925
4926
4927
                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,
4928
                kernel_block_sizes=kernel_block_sizes,
4929
                is_spec_decode=bool(self.vllm_config.speculative_config),
4930
                logitsprocs=self.input_batch.logitsprocs,
4931
                logitsprocs_need_output_token_ids=self.input_batch.logitsprocs_need_output_token_ids,
4932
                is_pooling_model=self.is_pooling_model,
4933
                num_speculative_tokens=self.num_spec_tokens,
4934
4935
            )

4936
    def _allocate_kv_cache_tensors(
4937
4938
        self, kv_cache_config: KVCacheConfig
    ) -> dict[str, torch.Tensor]:
4939
        """
4940
4941
4942
        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.

4943
        Args:
4944
            kv_cache_config: The KV cache config
4945
        Returns:
4946
            dict[str, torch.Tensor]: A map between layer names to their
4947
            corresponding memory buffer for KV cache.
4948
        """
4949
4950
        kv_cache_raw_tensors: dict[str, torch.Tensor] = {}
        for kv_cache_tensor in kv_cache_config.kv_cache_tensors:
4951
4952
4953
            tensor = torch.zeros(
                kv_cache_tensor.size, dtype=torch.int8, device=self.device
            )
4954
4955
4956
4957
4958
            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:
4959
4960
4961
4962
            for layer_name in group.layer_names:
                if layer_name in self.runner_only_attn_layers:
                    continue
                layer_names.add(layer_name)
4963
4964
4965
        assert layer_names == set(kv_cache_raw_tensors.keys()), (
            "Some layers are not correctly initialized"
        )
4966
4967
        return kv_cache_raw_tensors

4968
4969
4970
    def _attn_group_iterator(self) -> Iterator[AttentionGroup]:
        return itertools.chain.from_iterable(self.attn_groups)

4971
    def _kv_cache_spec_attn_group_iterator(self) -> Iterator[AttentionGroup]:
4972
4973
        if not self.kv_cache_config.kv_cache_groups:
            return
4974
4975
        for attn_groups in self.attn_groups:
            yield from attn_groups
4976

4977
4978
4979
4980
4981
4982
4983
4984
4985
4986
4987
4988
4989
4990
4991
    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 = []
4992
        for kv_cache_gid, kv_cache_group in enumerate(kv_cache_config.kv_cache_groups):
4993
4994
4995
4996
4997
4998
            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):
4999
                continue
5000
            elif isinstance(kv_cache_spec, AttentionSpec):
5001
5002
5003
                # This is an attention backend that supports virtual
                # block splitting. Get the supported block sizes from
                # all backends in the group.
5004
                attn_groups = self.attn_groups[kv_cache_gid]
5005
                kv_manager_block_size = kv_cache_group.kv_cache_spec.block_size
5006
                selected_kernel_size = self.select_common_block_size(
5007
5008
5009
                    kv_manager_block_size, attn_groups
                )
                kernel_block_sizes.append(selected_kernel_size)
5010
            elif isinstance(kv_cache_spec, MambaSpec):
5011
5012
                # This is likely Mamba or other non-attention cache,
                # no splitting.
5013
                kernel_block_sizes.append(kv_cache_spec.block_size)
5014
5015
5016
5017
5018
5019
            else:
                raise NotImplementedError(
                    f"unknown kv cache spec {kv_cache_group.kv_cache_spec}"
                )
        return kernel_block_sizes

5020
5021
5022
5023
    def _reshape_kv_cache_tensors(
        self,
        kv_cache_config: KVCacheConfig,
        kv_cache_raw_tensors: dict[str, torch.Tensor],
5024
        kernel_block_sizes: list[int],
5025
    ) -> dict[str, torch.Tensor]:
5026
        """
5027
        Reshape the KV cache tensors to the desired shape and dtype.
5028

5029
        Args:
5030
5031
            kv_cache_config: The KV cache config
            kv_cache_raw_tensors: The KV cache buffer of each layer, with
5032
                correct size but uninitialized shape.
5033
            kernel_block_sizes: The kernel block sizes for each KV cache group.
5034
        Returns:
5035
            Dict[str, torch.Tensor]: A map between layer names to their
5036
5037
            corresponding memory buffer for KV cache.
        """
5038
        kv_caches: dict[str, torch.Tensor] = {}
5039
        has_attn, has_mamba = False, False
5040
5041
        for group in self._kv_cache_spec_attn_group_iterator():
            kv_cache_spec = group.kv_cache_spec
5042
            attn_backend = group.backend
5043
5044
5045
5046
            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]
5047
            for layer_name in group.layer_names:
5048
5049
                if layer_name in self.runner_only_attn_layers:
                    continue
5050
5051
                raw_tensor = kv_cache_raw_tensors[layer_name]
                assert raw_tensor.numel() % kv_cache_spec.page_size_bytes == 0
5052
                num_blocks = raw_tensor.numel() // kv_cache_spec.page_size_bytes
5053
                if isinstance(kv_cache_spec, AttentionSpec):
5054
                    has_attn = True
5055
5056
                    num_blocks_per_kv_block = (
                        kv_cache_spec.block_size // kernel_block_size
5057
5058
5059
                    )
                    kernel_num_blocks = num_blocks * num_blocks_per_kv_block

5060
                    kv_cache_shape = attn_backend.get_kv_cache_shape(
5061
                        kernel_num_blocks,
5062
                        kernel_block_size,
5063
5064
                        kv_cache_spec.num_kv_heads,
                        kv_cache_spec.head_size,
5065
5066
                        cache_dtype_str=self.cache_config.cache_dtype,
                    )
5067
                    dtype = kv_cache_spec.dtype
5068
                    try:
5069
                        kv_cache_stride_order = attn_backend.get_kv_cache_stride_order()
5070
                        assert len(kv_cache_stride_order) == len(kv_cache_shape)
5071
                    except (AttributeError, NotImplementedError):
5072
                        kv_cache_stride_order = tuple(range(len(kv_cache_shape)))
5073
5074
5075
5076
5077
                    # 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.
5078
5079
5080
                    kv_cache_shape = tuple(
                        kv_cache_shape[i] for i in kv_cache_stride_order
                    )
5081
5082
5083
5084
5085
                    # Maintain original KV shape view.
                    inv_order = [
                        kv_cache_stride_order.index(i)
                        for i in range(len(kv_cache_stride_order))
                    ]
5086
5087
5088
5089
5090
5091
                    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
5092
                elif isinstance(kv_cache_spec, MambaSpec):
5093
                    has_mamba = True
Chen Zhang's avatar
Chen Zhang committed
5094
5095
                    raw_tensor = kv_cache_raw_tensors[layer_name]
                    state_tensors = []
5096
                    storage_offset_bytes = 0
5097
                    for shape, dtype in zip(kv_cache_spec.shapes, kv_cache_spec.dtypes):
5098
5099
                        dtype_size = get_dtype_size(dtype)
                        num_element_per_page = (
5100
5101
                            kv_cache_spec.page_size_bytes // dtype_size
                        )
Chen Zhang's avatar
Chen Zhang committed
5102
                        target_shape = (num_blocks, *shape)
5103
5104
                        stride = torch.empty(target_shape).stride()
                        target_stride = (num_element_per_page, *stride[1:])
5105
                        assert storage_offset_bytes % dtype_size == 0
5106
5107
5108
5109
                        tensor = torch.as_strided(
                            raw_tensor.view(dtype),
                            size=target_shape,
                            stride=target_stride,
5110
                            storage_offset=storage_offset_bytes // dtype_size,
5111
                        )
Chen Zhang's avatar
Chen Zhang committed
5112
                        state_tensors.append(tensor)
5113
                        storage_offset_bytes += stride[0] * dtype_size
5114
5115

                    kv_caches[layer_name] = state_tensors
5116
                else:
5117
                    raise NotImplementedError
5118
5119

        if has_attn and has_mamba:
5120
            self._update_hybrid_attention_mamba_layout(kv_caches)
5121

5122
5123
        return kv_caches

5124
    def _update_hybrid_attention_mamba_layout(
5125
5126
        self, kv_caches: dict[str, torch.Tensor]
    ) -> None:
5127
        """
5128
5129
        Update the layout of attention layers from (2, num_blocks, ...) to
        (num_blocks, 2, ...).
5130
5131

        Args:
5132
            kv_caches: The KV cache buffer of each layer.
5133
5134
        """

5135
5136
        for group in self._kv_cache_spec_attn_group_iterator():
            kv_cache_spec = group.kv_cache_spec
5137
            for layer_name in group.layer_names:
5138
                kv_cache = kv_caches[layer_name]
5139
5140
5141
5142
                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 "
5143
                        f"a tensor of shape {kv_cache.shape}"
5144
                    )
5145
                    hidden_size = kv_cache.shape[2:].numel()
5146
5147
5148
5149
                    kv_cache.as_strided_(
                        size=kv_cache.shape,
                        stride=(hidden_size, 2 * hidden_size, *kv_cache.stride()[2:]),
                    )
5150

5151
    def initialize_kv_cache_tensors(
5152
        self, kv_cache_config: KVCacheConfig, kernel_block_sizes: list[int]
5153
    ) -> dict[str, torch.Tensor]:
5154
5155
5156
5157
5158
        """
        Initialize the memory buffer for KV cache.

        Args:
            kv_cache_config: The KV cache config
5159
5160
            kernel_block_sizes: The kernel block sizes for each KV cache group.

5161
        Returns:
5162
            Dict[str, torch.Tensor]: A map between layer names to their
5163
5164
            corresponding memory buffer for KV cache.
        """
5165
5166
5167
5168
5169
5170
5171
5172
5173
5174
5175
5176
5177
5178
5179
5180
5181
5182
5183
5184
5185
5186
5187
5188

        # 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
            )
5189

5190
        # Set up cross-layer KV cache sharing
5191
5192
        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)
5193
5194
            kv_caches[layer_name] = kv_caches[target_layer_name]

5195
5196
5197
5198
5199
5200
5201
5202
5203
        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,
        )
5204
5205
5206
        return kv_caches

    def maybe_add_kv_sharing_layers_to_kv_cache_groups(
5207
5208
        self, kv_cache_config: KVCacheConfig
    ) -> None:
5209
5210
5211
5212
5213
5214
5215
5216
5217
5218
5219
5220
5221
5222
5223
5224
5225
5226
        """
        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.
5227
            attn_layers = get_layers_from_vllm_config(self.vllm_config, Attention)
5228
5229
            for layer_name in reversed(attn_layers):
                if layer_name in self.shared_kv_cache_layers:
5230
                    self.kv_sharing_fast_prefill_eligible_layers.add(layer_name)
5231
5232
                else:
                    break
5233

5234
5235
5236
5237
5238
5239
5240
    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
        """
5241
        kv_cache_config = deepcopy(kv_cache_config)
5242
        self.kv_cache_config = kv_cache_config
5243
        self.may_add_encoder_only_layers_to_kv_cache_config()
5244
        self.maybe_add_kv_sharing_layers_to_kv_cache_groups(kv_cache_config)
5245
        self.initialize_attn_backend(kv_cache_config)
5246
5247
5248
5249
5250
5251
        # 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)
5252
5253
5254
5255

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

5256
        # Reinitialize need to after initialize_attn_backend
5257
5258
5259
5260
        self.may_reinitialize_input_batch(kv_cache_config, kernel_block_sizes)
        kv_caches = self.initialize_kv_cache_tensors(
            kv_cache_config, kernel_block_sizes
        )
5261

5262
5263
5264
5265
5266
5267
        if self.speculative_config and self.speculative_config.use_eagle():
            assert isinstance(self.drafter, EagleProposer)
            # validate all draft model layers belong to the same kv cache
            # group
            self.drafter.validate_same_kv_cache_group(kv_cache_config)

Robert Shaw's avatar
Robert Shaw committed
5268
        if has_kv_transfer_group():
5269
            kv_transfer_group = get_kv_transfer_group()
5270
5271
5272
5273
5274
5275
5276
            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)
5277
            kv_transfer_group.set_host_xfer_buffer_ops(copy_kv_blocks)
Robert Shaw's avatar
Robert Shaw committed
5278

5279
        if self.dcp_world_size > 1:
5280
5281
            layer_type = cast(type[Any], AttentionLayerBase)
            layers = get_layers_from_vllm_config(self.vllm_config, layer_type)
5282
            for layer in layers.values():
5283
5284
5285
5286
                layer_impl = getattr(layer, "impl", None)
                if layer_impl is None:
                    continue
                assert layer_impl.need_to_return_lse_for_decode, (
5287
5288
                    "DCP requires attention impls to return"
                    " the softmax lse for decode, but the impl "
5289
                    f"{layer_impl.__class__.__name__} "
5290
5291
                    "does not return the softmax lse for decode."
                )
5292

5293
5294
5295
5296
5297
    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
5298
        encoder_only_attn_specs: dict[AttentionSpec, list[str]] = defaultdict(list)
5299
5300
5301
        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:
5302
                attn_spec: AttentionSpec = EncoderOnlyAttentionSpec(
5303
5304
5305
                    block_size=block_size,
                    num_kv_heads=attn_module.num_kv_heads,
                    head_size=attn_module.head_size,
5306
5307
                    dtype=self.kv_cache_dtype,
                )
5308
5309
5310
                encoder_only_attn_specs[attn_spec].append(layer_name)
                self.runner_only_attn_layers.add(layer_name)
        if len(encoder_only_attn_specs) > 0:
5311
5312
5313
            assert len(encoder_only_attn_specs) == 1, (
                "Only support one encoder-only attention spec now"
            )
5314
5315
            spec, layer_names = encoder_only_attn_specs.popitem()
            self.kv_cache_config.kv_cache_groups.append(
5316
5317
                KVCacheGroupSpec(layer_names=layer_names, kv_cache_spec=spec)
            )
5318

5319
    def get_kv_cache_spec(self) -> dict[str, KVCacheSpec]:
5320
        """
5321
        Generates the KVCacheSpec by parsing the kv cache format from each
5322
5323
        Attention module in the static forward context.
        Returns:
5324
            KVCacheSpec: A dictionary mapping layer names to their KV cache
5325
5326
            format. Layers that do not need KV cache are not included.
        """
5327
5328
        if has_ec_transfer() and get_ec_transfer().is_producer:
            return {}
5329
        kv_cache_spec: dict[str, KVCacheSpec] = {}
5330
5331
        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
5332
        for layer_name, attn_module in attn_layers.items():
5333
5334
5335
5336
5337
5338
5339
5340
5341
5342
5343
5344
5345
5346
5347
            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
5348

5349
        return kv_cache_spec
5350

5351
    def _to_list(self, sampled_token_ids: torch.Tensor) -> list[list[int]]:
5352
5353
5354
5355
5356
5357
5358
5359
        # 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.
5360
        pinned = self.sampled_token_ids_pinned_cpu[: sampled_token_ids.shape[0]]
5361
5362
5363
        pinned.copy_(sampled_token_ids, non_blocking=True)
        self.transfer_event.record()
        self.transfer_event.synchronize()
5364
        return pinned.tolist()