gpu_model_runner.py 221 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
from vllm.attention import Attention, AttentionType
23
24
25
26
27
from vllm.attention.backends.abstract import (
    AttentionBackend,
    AttentionMetadata,
    MultipleOf,
)
28
from vllm.compilation.counter import compilation_counter
29
30
from vllm.compilation.cuda_graph import CUDAGraphWrapper
from vllm.compilation.monitor import set_cudagraph_capturing_enabled
31
from vllm.config import (
32
    CompilationMode,
33
34
35
36
37
    CUDAGraphMode,
    VllmConfig,
    get_layers_from_vllm_config,
    update_config,
)
38
from vllm.distributed.ec_transfer import get_ec_transfer, has_ec_transfer
39
from vllm.distributed.eplb.eplb_state import EplbState
40
from vllm.distributed.kv_transfer import get_kv_transfer_group, has_kv_transfer_group
41
from vllm.distributed.kv_transfer.kv_connector.utils import copy_kv_blocks
42
from vllm.distributed.parallel_state import (
43
    get_dcp_group,
44
45
46
47
48
49
    get_pp_group,
    get_tp_group,
    graph_capture,
    is_global_first_rank,
    prepare_communication_buffer_for_model,
)
50
from vllm.forward_context import BatchDescriptor, set_forward_context
51
from vllm.logger import init_logger
52
from vllm.model_executor.layers.attention_layer_base import AttentionLayerBase
53
from vllm.model_executor.layers.rotary_embedding import MRotaryEmbedding
54
from vllm.model_executor.model_loader import TensorizerLoader, get_model_loader
55
from vllm.model_executor.models.interfaces import (
56
    SupportsMRoPE,
57
58
59
60
61
62
63
    SupportsMultiModal,
    is_mixture_of_experts,
    supports_eagle3,
    supports_mrope,
    supports_multimodal_pruning,
    supports_transcription,
)
64
from vllm.model_executor.models.interfaces_base import (
65
66
67
68
    VllmModelForPooling,
    is_pooling_model,
    is_text_generation_model,
)
69
from vllm.multimodal import MULTIMODAL_REGISTRY
70
71
72
73
74
from vllm.multimodal.inputs import (
    BatchedTensorInputs,
    MultiModalKwargsItem,
    PlaceholderRange,
)
75
from vllm.multimodal.utils import group_mm_kwargs_by_modality
76
from vllm.pooling_params import PoolingParams
77
from vllm.sampling_params import SamplingType
78
from vllm.sequence import IntermediateTensors
79
from vllm.tasks import GenerationTask, PoolingTask, SupportedTask
80
from vllm.utils import length_from_prompt_token_ids_or_embeds
81
from vllm.utils.jsontree import json_map_leaves
82
from vllm.utils.math_utils import cdiv, round_up
83
84
from vllm.utils.mem_constants import GiB_bytes
from vllm.utils.mem_utils import DeviceMemoryProfiler
85
from vllm.utils.platform_utils import is_pin_memory_available
86
87
88
89
90
from vllm.utils.torch_utils import (
    get_dtype_size,
    kv_cache_dtype_str_to_dtype,
    supports_dynamo,
)
91
from vllm.v1.attention.backends.gdn_attn import GDNAttentionMetadataBuilder
92
from vllm.v1.attention.backends.utils import (
93
94
95
    AttentionCGSupport,
    AttentionMetadataBuilder,
    CommonAttentionMetadata,
96
    create_fast_prefill_custom_backend,
97
    get_dcp_local_seq_lens,
98
99
100
    reorder_batch_to_split_decodes_and_prefills,
    split_attn_metadata,
)
101
from vllm.v1.cudagraph_dispatcher import CudagraphDispatcher
102
103
104
105
106
107
108
109
110
111
112
113
114
115
116
117
118
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,
119
    ECConnectorOutput,
120
    KVConnectorOutput,
121
122
123
124
125
    LogprobsLists,
    LogprobsTensors,
    ModelRunnerOutput,
    PoolerOutput,
    SamplerOutput,
126
    make_empty_encoder_model_runner_output,
127
)
128
from vllm.v1.pool.metadata import PoolingMetadata
129
from vllm.v1.sample.logits_processor import LogitsProcessors, build_logitsprocs
130
from vllm.v1.sample.logits_processor.interface import LogitsProcessor
131
from vllm.v1.sample.metadata import SamplingMetadata
132
from vllm.v1.sample.rejection_sampler import RejectionSampler
133
from vllm.v1.sample.sampler import Sampler
134
from vllm.v1.spec_decode.eagle import EagleProposer
135
from vllm.v1.spec_decode.medusa import MedusaProposer
136
from vllm.v1.spec_decode.metadata import SpecDecodeMetadata
137
from vllm.v1.spec_decode.ngram_proposer import NgramProposer
138
from vllm.v1.spec_decode.suffix_decoding import SuffixDecodingProposer
139
from vllm.v1.structured_output.utils import apply_grammar_bitmask
140
from vllm.v1.utils import CpuGpuBuffer, record_function_or_nullcontext
141
from vllm.v1.worker.dp_utils import coordinate_batch_across_dp
142
from vllm.v1.worker.ec_connector_model_runner_mixin import ECConnectorModelRunnerMixin
143
from vllm.v1.worker.gpu_input_batch import CachedRequestState, InputBatch
144
from vllm.v1.worker.gpu_ubatch_wrapper import UBatchWrapper
145
from vllm.v1.worker.kv_connector_model_runner_mixin import KVConnectorModelRunnerMixin
146
from vllm.v1.worker.lora_model_runner_mixin import LoRAModelRunnerMixin
147
148
149
150
151
from vllm.v1.worker.ubatch_utils import (
    UBatchSlice,
    UBatchSlices,
    check_ubatch_thresholds,
)
152
from vllm.v1.worker.utils import is_residual_scattered_for_sp
153

154
155
156
157
158
159
160
161
162
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,
)
163

164
if TYPE_CHECKING:
165
    from vllm.model_executor.model_loader.tensorizer import TensorizerConfig
166
    from vllm.v1.core.sched.output import GrammarOutput, SchedulerOutput
167
168
169

logger = init_logger(__name__)

170
171
AttnMetadataDict: TypeAlias = dict[str, AttentionMetadata]
# list when ubatching is enabled
172
PerLayerAttnMetadata: TypeAlias = list[AttnMetadataDict] | AttnMetadataDict
173

174

175
176
177
178
179
180
# Wrapper for ModelRunnerOutput to support overlapped execution.
class AsyncGPUModelRunnerOutput(AsyncModelRunnerOutput):
    def __init__(
        self,
        model_runner_output: ModelRunnerOutput,
        sampled_token_ids: torch.Tensor,
181
        logprobs_tensors: torch.Tensor | None,
182
183
        invalid_req_indices: list[int],
        async_output_copy_stream: torch.cuda.Stream,
184
        vocab_size: int,
185
186
187
188
189
    ):
        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.
190
        self.async_copy_ready_event = torch.Event()
191
192
193
194

        # 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
195
        self.vocab_size = vocab_size
196
        self._logprobs_tensors = logprobs_tensors
197
198
199
200
201

        # 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)
202
            self.sampled_token_ids_cpu = self._sampled_token_ids.to(
203
204
                "cpu", non_blocking=True
            )
205
206
207
208
209
            self._logprobs_tensors_cpu = (
                self._logprobs_tensors.to_cpu_nonblocking()
                if self._logprobs_tensors
                else None
            )
210
            self.async_copy_ready_event.record()
211
212
213

    def get_output(self) -> ModelRunnerOutput:
        """Copy the device tensors to the host and return a ModelRunnerOutput.
214

215
216
        This function blocks until the copy is finished.
        """
217
        self.async_copy_ready_event.synchronize()
218

219
220
        # Release the device tensors once the copy has completed.
        del self._logprobs_tensors
221
        del self._sampled_token_ids
222
223
        max_gen_len = self.sampled_token_ids_cpu.shape[-1]
        if max_gen_len == 1:
224
            valid_sampled_token_ids = self.sampled_token_ids_cpu.tolist()
225
226
227
228
229
        else:
            valid_sampled_token_ids = RejectionSampler.parse_output(
                self.sampled_token_ids_cpu,
                self.vocab_size,
            )
230
        for i in self._invalid_req_indices:
231
            valid_sampled_token_ids[i].clear()
232
233
234

        output = self._model_runner_output
        output.sampled_token_ids = valid_sampled_token_ids
235
236
237
238
        if self._logprobs_tensors_cpu:
            # NOTE(nick): this will need to be updated to use cu_num_accepted_tokens
            # for async sched + spec decode + logprobs compatibility.
            output.logprobs = self._logprobs_tensors_cpu.tolists()
239
240
241
        return output


242
243
244
245
246
247
248
249
250
251
252
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
253
    ec_connector_output: ECConnectorOutput | None
254
255


256
257
258
class GPUModelRunner(
    LoRAModelRunnerMixin, KVConnectorModelRunnerMixin, ECConnectorModelRunnerMixin
):
259
260
    def __init__(
        self,
261
        vllm_config: VllmConfig,
262
        device: torch.device,
263
    ):
264
265
266
        self.vllm_config = vllm_config
        self.model_config = vllm_config.model_config
        self.cache_config = vllm_config.cache_config
267
        self.compilation_config = vllm_config.compilation_config
268
269
270
271
272
273
        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
274

275
        from vllm.model_executor.models.utils import set_cpu_offload_max_bytes
276
277

        set_cpu_offload_max_bytes(int(self.cache_config.cpu_offload_gb * 1024**3))
278

279
280
281
282
        model_config = self.model_config
        cache_config = self.cache_config
        scheduler_config = self.scheduler_config
        parallel_config = self.parallel_config
283
        self.device = device
284
285
        self.pin_memory = is_pin_memory_available()
        self.dtype = self.model_config.dtype
286
287
288
        self.kv_cache_dtype = kv_cache_dtype_str_to_dtype(
            cache_config.cache_dtype, self.model_config
        )
289

290
        self.is_pooling_model = model_config.runner_type == "pooling"
291
        self.enable_prompt_embeds = model_config.enable_prompt_embeds
292
        self.is_multimodal_raw_input_only_model = (
293
294
            model_config.is_multimodal_raw_input_only_model
        )
295
296
        # This will be overridden in load_model()
        self.is_multimodal_pruning_enabled = False
297
        self.max_model_len = model_config.max_model_len
298
299
300

        # Always set to false after the first forward pass
        self.calculate_kv_scales = self.cache_config.calculate_kv_scales
301
        self.dcp_world_size = self.parallel_config.decode_context_parallel_size
302
        self.dcp_rank = 0 if self.dcp_world_size <= 1 else get_dcp_group().rank_in_group
303
        self.max_num_tokens = scheduler_config.max_num_batched_tokens
304
        self.max_num_reqs = scheduler_config.max_num_seqs
305

306
307
308
309
310
        # 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 = (
311
312
313
            self.parallel_config.distributed_executor_backend == "external_launcher"
            and len(get_pp_group().ranks) > 0
        )
314

315
        # Model-related.
316
        self.num_query_heads = model_config.get_num_attention_heads(parallel_config)
317
        self.hidden_size = model_config.get_hidden_size()
318
        self.attention_chunk_size = model_config.attention_chunk_size
319
        # Only relevant for models using ALiBi (e.g, MPT)
320
        self.use_alibi = model_config.uses_alibi
321

322
        self.cascade_attn_enabled = not self.model_config.disable_cascade_attn
323

324
        # Multi-modal data support
325
        self.mm_registry = MULTIMODAL_REGISTRY
326
        self.uses_mrope = model_config.uses_mrope
327
        self.supports_mm_inputs = self.mm_registry.supports_multimodal_inputs(
328
329
            model_config
        )
330

331
332
333
        if self.model_config.is_encoder_decoder:
            # Maximum length of the encoder input, only for encoder-decoder
            # models.
334
            self.max_encoder_len = scheduler_config.max_num_encoder_input_tokens
335
336
337
        else:
            self.max_encoder_len = 0

338
        # Sampler
339
        self.sampler = Sampler(logprobs_mode=self.model_config.logprobs_mode)
340

341
        self.eplb_state: EplbState | None = None
342
343
344
345
346
347
        """
        State of the expert parallelism load balancer.

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

348
        # Lazy initializations
349
        # self.model: nn.Module  # Set after load_model
350
        # Initialize in initialize_kv_cache
351
        self.kv_caches: list[torch.Tensor] = []
352
353
354
        # Initialize in initialize_kv_cache_tensors
        self.cross_layers_kv_cache: torch.Tensor | None = None
        self.cross_layers_attn_backend: type[AttentionBackend] | None = None
355
356
        # indexes: [kv_cache_group_id][attn_group]
        self.attn_groups: list[list[AttentionGroup]] = []
357
358
        # self.kv_cache_config: KVCacheConfig

359
360
        # mm_hash ->  encoder_output
        self.encoder_cache: dict[str, torch.Tensor] = {}
361

362
        self.use_aux_hidden_state_outputs = False
363
364
365
366
367
        # 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:
368
369
370
            self.drafter: (
                NgramProposer | SuffixDecodingProposer | EagleProposer | MedusaProposer
            )
371
372
            if self.speculative_config.method == "ngram":
                self.drafter = NgramProposer(self.vllm_config)
373
374
            elif self.speculative_config.method == "suffix":
                self.drafter = SuffixDecodingProposer(self.vllm_config)
375
            elif self.speculative_config.use_eagle():
376
                self.drafter = EagleProposer(self.vllm_config, self.device, self)
377
                if self.speculative_config.method == "eagle3":
378
379
380
                    self.use_aux_hidden_state_outputs = (
                        self.drafter.eagle3_use_aux_hidden_state
                    )
381
382
            elif self.speculative_config.method == "medusa":
                self.drafter = MedusaProposer(
383
                    vllm_config=self.vllm_config, device=self.device
384
                )
385
            else:
386
387
388
389
                raise ValueError(
                    "Unknown speculative decoding method: "
                    f"{self.speculative_config.method}"
                )
390
            self.rejection_sampler = RejectionSampler(self.sampler)
391

392
393
394
395
        self.num_spec_tokens = 0
        if self.speculative_config:
            self.num_spec_tokens = self.speculative_config.num_speculative_tokens

396
        # Request states.
397
        self.requests: dict[str, CachedRequestState] = {}
398
399
400
        # NOTE(rob): num_prompt_logprobs only includes reqs
        # that are currently in the prefill phase.
        self.num_prompt_logprobs: dict[str, int] = {}
401
        self.comm_stream = torch.cuda.Stream()
402

403
404
405
406
407
408
409
410
411
        # 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.
412
413
414
415
        logits_processors = model_config.logits_processors
        custom_logitsprocs: Sequence[str | type[LogitsProcessor]] = (
            tuple(logits_processors) if logits_processors is not None else ()
        )
416
417
        self.input_batch = InputBatch(
            max_num_reqs=self.max_num_reqs,
418
419
420
            # 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),
421
422
423
            max_num_batched_tokens=self.max_num_tokens,
            device=self.device,
            pin_memory=self.pin_memory,
424
            vocab_size=self.model_config.get_vocab_size(),
425
            block_sizes=[self.cache_config.block_size],
426
            kernel_block_sizes=[self.cache_config.block_size],
427
            is_spec_decode=bool(self.vllm_config.speculative_config),
428
            logitsprocs=build_logitsprocs(
429
430
431
                self.vllm_config,
                self.device,
                self.pin_memory,
432
                self.is_pooling_model,
433
                custom_logitsprocs,
434
            ),
435
436
437
            # 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),
438
            is_pooling_model=self.is_pooling_model,
439
            cp_kv_cache_interleave_size=self.parallel_config.cp_kv_cache_interleave_size,
440
        )
441

442
        self.use_async_scheduling = self.scheduler_config.async_scheduling
443
444
445
446
447
        # 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.
448
        self.prepare_inputs_event: torch.Event | None = None
449
450
        if self.use_async_scheduling:
            self.async_output_copy_stream = torch.cuda.Stream()
451
            self.prepare_inputs_event = torch.Event()
452

453
        # self.cudagraph_batch_sizes sorts in ascending order.
454
455
456
457
        if (
            self.compilation_config.cudagraph_capture_sizes
            and self.compilation_config.cudagraph_mode != CUDAGraphMode.NONE
        ):
458
459
            self.cudagraph_batch_sizes = sorted(
                self.compilation_config.cudagraph_capture_sizes
460
            )
461

462
        # Cache the device properties.
463
        self._init_device_properties()
464

465
        # Persistent buffers for CUDA graphs.
466
467
468
469
470
        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
        )
471
        self.seq_lens = self._make_buffer(self.max_num_reqs, dtype=torch.int32)
472
473
474
475
        if self.dcp_world_size > 1:
            self.dcp_local_seq_lens = self._make_buffer(
                self.max_num_reqs, dtype=torch.int32
            )
476
477
478
        # 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.
479
480
481
482
483
484
485
        self.inputs_embeds = self._make_buffer(
            self.max_num_tokens, self.hidden_size, dtype=self.dtype, numpy=False
        )
        self.is_token_ids = self._make_buffer(self.max_num_tokens, dtype=torch.bool)
        self.discard_request_indices = self._make_buffer(
            self.max_num_reqs, dtype=torch.int64
        )
486
487
        self.num_discarded_requests = 0

488
489
490
491
492
493
        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
        )
494

495
496
        # Only relevant for multimodal models
        if self.supports_mm_inputs:
497
            self.is_mm_embed = self._make_buffer(self.max_num_tokens, dtype=torch.bool)
498

499
        # Only relevant for models using M-RoPE (e.g, Qwen2-VL)
500
        if self.uses_mrope:
Roger Wang's avatar
Roger Wang committed
501
502
503
504
            # 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
505
506
507
508
509
510

            # 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
511
            self.mrope_positions = self._make_buffer(
512
513
                (3, self.max_num_tokens + 1), dtype=torch.int64
            )
514

515
        # None in the first PP rank. The rest are set after load_model.
516
        self.intermediate_tensors: IntermediateTensors | None = None
517

518
        # OPTIMIZATION: Cache the tensors rather than creating them every step.
519
        # Keep in int64 to avoid overflow with long context
520
521
522
523
        self.arange_np = np.arange(
            max(self.max_num_reqs + 1, self.max_model_len, self.max_num_tokens),
            dtype=np.int64,
        )
524

525
526
527
528
529
        # 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] = {}
530
531
532
533
534
        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(
535
536
                self.max_num_tokens, dtype=torch.int32, device=self.device
            )
537

538
        self.uniform_decode_query_len = 1 + self.num_spec_tokens
539
540
541
542

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

543
544
545
546
547
548
549
550
551
        self.mm_budget = (
            MultiModalBudget(
                self.model_config,
                self.scheduler_config,
                self.mm_registry,
            )
            if self.supports_mm_inputs
            else None
        )
552

553
        self.reorder_batch_threshold: int | None = None
554

555
556
557
558
559
        # 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()

560
        # Cached outputs.
561
        self._draft_token_ids: list[list[int]] | torch.Tensor | None = None
562
        self.transfer_event = torch.Event()
563
        self.sampled_token_ids_pinned_cpu = torch.empty(
564
            (self.max_num_reqs, 1),
565
566
            dtype=torch.int64,
            device="cpu",
567
568
            pin_memory=self.pin_memory,
        )
569

570
571
        # Pre-allocated tensor for copying valid sampled token counts to CPU,
        # with dedicated stream for overlapping and event for coordination.
572
        self.valid_sampled_token_count_event: torch.Event | None = None
573
574
        self.valid_sampled_token_count_copy_stream: torch.cuda.Stream | None = None
        if self.use_async_scheduling and self.num_spec_tokens:
575
            self.valid_sampled_token_count_event = torch.Event()
576
577
578
579
580
581
582
583
            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,
        )

584
585
        # Ephemeral state transferred between execute_model() and sample_tokens().
        self.execute_model_state: ExecuteModelState | None = None
586
        self.kv_connector_output: KVConnectorOutput | None = None
587

588
589
590
591
    def reset_mm_cache(self) -> None:
        if self.mm_budget:
            self.mm_budget.reset_cache()

592
593
594
595
596
597
598
599
600
601
    def _get_positions(self, num_tokens: Any):
        if isinstance(num_tokens, int):
            if self.uses_mrope:
                return self.mrope_positions.gpu[:, :num_tokens]
            return self.positions.gpu[:num_tokens]
        else:
            if self.uses_mrope:
                return self.mrope_positions.gpu[:, num_tokens]
            return self.positions.gpu[num_tokens]

602
    def _make_buffer(
603
        self, *size: int | torch.SymInt, dtype: torch.dtype, numpy: bool = True
604
605
606
607
608
609
610
611
    ) -> CpuGpuBuffer:
        return CpuGpuBuffer(
            *size,
            dtype=dtype,
            device=self.device,
            pin_memory=self.pin_memory,
            with_numpy=numpy,
        )
612

613
614
615
    def _init_model_kwargs(self, num_tokens: int):
        model_kwargs = dict[str, Any]()

616
        if not self.is_pooling_model:
617
618
            return model_kwargs

619
620
        num_reqs = self.input_batch.num_reqs
        pooling_params = self.input_batch.get_pooling_params()
621
622
623

        token_type_id_requests = dict[int, Any]()
        for i, param in enumerate(pooling_params):
624
625
626
627
628
            if (
                param.extra_kwargs is not None
                and (token_types := param.extra_kwargs.get("compressed_token_type_ids"))
                is not None
            ):
629
630
631
632
633
                token_type_id_requests[i] = token_types

        if len(token_type_id_requests) == 0:
            return model_kwargs

634
        seq_lens = self.seq_lens.gpu[:num_reqs]
635
636
637
638
639
640
641
642
        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(
643
644
            device=self.device
        )
645
646
        return model_kwargs

647
    def _may_reorder_batch(self, scheduler_output: "SchedulerOutput") -> None:
648
649
        """
        Update the order of requests in the batch based on the attention
650
        backend's needs. For example, some attention backends (namely MLA) may
651
652
653
654
655
656
        want to separate requests based on if the attention computation will be
        compute-bound or memory-bound.

        Args:
            scheduler_output: The scheduler output.
        """
657
658
659
660
661
662
663
664
        # 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

665
666
667
668
        if self.reorder_batch_threshold is not None:
            reorder_batch_to_split_decodes_and_prefills(
                self.input_batch,
                scheduler_output,
669
670
                decode_threshold=self.reorder_batch_threshold,
            )
671

672
673
    # Note: used for model runner override.
    def _init_device_properties(self) -> None:
674
        """Initialize attributes from torch.cuda.get_device_properties"""
675
676
677
678
679
680
681
        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()

682
    def _update_states(self, scheduler_output: "SchedulerOutput") -> None:
683
684
685
686
687
688
        """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.

689
690
        The SamplingMetadata is updated and copied to the GPU if there is a
        new/resumed/paused/finished request in the batch.
691
692
        """
        # Remove finished requests from the cached states.
693
694
        for req_id in scheduler_output.finished_req_ids:
            self.requests.pop(req_id, None)
695
            self.num_prompt_logprobs.pop(req_id, None)
696
697
698
699
700
701
702
        # 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:
703
            self.input_batch.remove_request(req_id)
704
705

        # Free the cached encoder outputs.
706
707
        for mm_hash in scheduler_output.free_encoder_mm_hashes:
            self.encoder_cache.pop(mm_hash, None)
708

709
710
711
712
713
714
715
716
717
718
719
720
721
        # Remove the unscheduled requests from the persistent batch.
        # NOTE(woosuk): The unscheduled requests are either preempted requests
        # or running requests that are not scheduled in this step. We remove
        # them from the persistent batch but keep their cached states since
        # they will be scheduled again sometime in the future.
        scheduled_req_ids = scheduler_output.num_scheduled_tokens.keys()
        cached_req_ids = self.input_batch.req_id_to_index.keys()
        unscheduled_req_ids = cached_req_ids - scheduled_req_ids
        # NOTE(woosuk): The persistent batch optimization assumes that
        # consecutive batches contain mostly the same requests. If batches
        # have low request overlap (e.g., alternating between two distinct
        # sets of requests), this optimization becomes very inefficient.
        for req_id in unscheduled_req_ids:
722
            self.input_batch.remove_request(req_id)
723

724
        reqs_to_add: list[CachedRequestState] = []
725
        # Add new requests to the cached states.
726
727
728
        for new_req_data in scheduler_output.scheduled_new_reqs:
            req_id = new_req_data.req_id
            sampling_params = new_req_data.sampling_params
729
            pooling_params = new_req_data.pooling_params
730

731
732
733
734
            if (
                sampling_params
                and sampling_params.sampling_type == SamplingType.RANDOM_SEED
            ):
735
736
737
738
739
                generator = torch.Generator(device=self.device)
                generator.manual_seed(sampling_params.seed)
            else:
                generator = None

740
741
            if self.is_pooling_model:
                assert pooling_params is not None
742
743
                task = pooling_params.task
                assert task is not None, "You did not set `task` in the API"
744

745
                model = cast(VllmModelForPooling, self.get_model())
746
                to_update = model.pooler.get_pooling_updates(task)
747
748
                to_update.apply(pooling_params)

749
            req_state = CachedRequestState(
750
                req_id=req_id,
751
                prompt_token_ids=new_req_data.prompt_token_ids,
752
                prompt_embeds=new_req_data.prompt_embeds,
753
                mm_features=new_req_data.mm_features,
754
                sampling_params=sampling_params,
755
                pooling_params=pooling_params,
756
                generator=generator,
757
758
                block_ids=new_req_data.block_ids,
                num_computed_tokens=new_req_data.num_computed_tokens,
759
                output_token_ids=[],
760
                lora_request=new_req_data.lora_request,
761
            )
762
763
            self.requests[req_id] = req_state

764
765
766
767
768
769
770
            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
                )

771
            # Only relevant for models using M-RoPE (e.g, Qwen2-VL)
772
            if self.uses_mrope:
773
                self._init_mrope_positions(req_state)
774

775
            reqs_to_add.append(req_state)
776

777
        # Update the states of the running/resumed requests.
778
        is_last_rank = get_pp_group().is_last_rank
779
        req_data = scheduler_output.scheduled_cached_reqs
780
781
782
783
784

        # 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()

785
        for i, req_id in enumerate(req_data.req_ids):
786
            req_state = self.requests[req_id]
787
788
            num_computed_tokens = req_data.num_computed_tokens[i]
            new_block_ids = req_data.new_block_ids[i]
789
            resumed_from_preemption = req_id in req_data.resumed_req_ids
790
            num_output_tokens = req_data.num_output_tokens[i]
791
            req_index = self.input_batch.req_id_to_index.get(req_id)
792

793
794
795
796
797
798
799
800
801
802
803
804
805
806
807
808
809
810
811
812
813
814
815
            # 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)
816

817
            # Update the cached states.
818
            req_state.num_computed_tokens = num_computed_tokens
819
820
821
822
823
824
825
826

            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.
827
828
829
                num_new_tokens = (
                    num_computed_tokens + len(new_token_ids) - req_state.num_tokens
                )
830
831
832
833
                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:
834
                    req_state.output_token_ids.extend(new_token_ids[-num_new_tokens:])
835
836
837
838
839
            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:
840
841
842
843
                    end_idx = (
                        self.input_batch.num_prompt_tokens[req_index]
                        + num_output_tokens
                    )
844
845
                    self.input_batch.num_tokens[req_index] = end_idx
                    self.input_batch.num_tokens_no_spec[req_index] = end_idx
846

847
            # Update the block IDs.
848
            if not resumed_from_preemption:
849
850
                if new_block_ids is not None:
                    # Append the new blocks to the existing block IDs.
851
                    for block_ids, new_ids in zip(req_state.block_ids, new_block_ids):
852
                        block_ids.extend(new_ids)
853
            else:
854
                assert req_index is None
855
                assert new_block_ids is not None
856
857
                # The request is resumed from preemption.
                # Replace the existing block IDs with the new ones.
858
                req_state.block_ids = new_block_ids
859
860
861
862
863

            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.
864
865
866
867
868
869
870

                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:]

871
                reqs_to_add.append(req_state)
872
873
874
                continue

            # Update the persistent batch.
875
            self.input_batch.num_computed_tokens_cpu[req_index] = num_computed_tokens
876
            if new_block_ids is not None:
877
                self.input_batch.block_table.append_row(new_block_ids, req_index)
878
879
880
881
882
883
884

            # 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)
885
                self.input_batch.token_ids_cpu[
886
887
888
                    req_index, start_token_index:end_token_index
                ] = new_token_ids
                self.input_batch.num_tokens_no_spec[req_index] = end_token_index
889
                self.input_batch.num_tokens[req_index] = end_token_index
890

891
            # Add spec_token_ids to token_ids_cpu.
892
            spec_token_ids = scheduler_output.scheduled_spec_decode_tokens.get(
893
                req_id, []
894
            )
895
896
897
898
899
            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:
900
901
902
                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[
903
904
                    req_index, start_index:end_token_index
                ] = spec_token_ids
905
906
                # NOTE(woosuk): `num_tokens` here may include spec tokens.
                self.input_batch.num_tokens[req_index] += num_spec_tokens
907
908
909
910
911
912

            # 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.
913
914
            self.input_batch.spec_token_ids[req_index].clear()
            self.input_batch.spec_token_ids[req_index].extend(spec_token_ids)
915

916
917
918
919
920
921
922
923
924
            # 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)
925
926
        # Add the new or resumed requests to the persistent batch.
        # The smaller empty indices are filled first.
927
928
        for request in reqs_to_add:
            self.input_batch.add_request(request)
929

930
931
932
933
934
935
        # 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()
936

937
    def _update_states_after_model_execute(
938
939
        self, output_token_ids: torch.Tensor
    ) -> None:
940
941
942
943
944
945
946
947
948
949
950
951
        """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.
952
953
954
955
956
957
958
959
960
961
962
963
964
965
966
967
968
969
970
971
        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()
        )
972
973
974
        for i, num_tokens in enumerate(num_accepted_tokens):
            self.input_batch.num_accepted_tokens_cpu[i] = num_tokens

975
    def _init_mrope_positions(self, req_state: CachedRequestState):
976
977
        model = self.get_model()
        assert supports_mrope(model), "M-RoPE support is not implemented."
978
979
980
981
        assert req_state.prompt_token_ids is not None, (
            "M-RoPE requires prompt_token_ids to be available."
        )
        mrope_model = cast(SupportsMRoPE, model)
982
983

        req_state.mrope_positions, req_state.mrope_position_delta = (
984
            mrope_model.get_mrope_input_positions(
985
                req_state.prompt_token_ids,
986
                req_state.mm_features,
987
            )
988
        )
989

990
    def _extract_mm_kwargs(
991
        self,
992
993
        scheduler_output: "SchedulerOutput",
    ) -> BatchedTensorInputs:
994
        if not scheduler_output or not self.is_multimodal_raw_input_only_model:
995
            return {}
996

997
998
        mm_kwargs = list[MultiModalKwargsItem]()
        for req in scheduler_output.scheduled_new_reqs:
999
1000
1001
            for feature in req.mm_features:
                if feature.data is not None:
                    mm_kwargs.append(feature.data)
1002

1003
        # Input all modalities at once
1004
        model = cast(SupportsMultiModal, self.model)
1005
1006
        mm_kwargs_combined: BatchedTensorInputs = {}
        for _, _, mm_kwargs_group in group_mm_kwargs_by_modality(
1007
1008
1009
1010
            mm_kwargs,
            device=self.device,
            pin_memory=self.pin_memory,
            merge_by_field_config=model.merge_by_field_config,
1011
            multimodal_cpu_fields=model.multimodal_cpu_fields,
1012
1013
        ):
            mm_kwargs_combined.update(mm_kwargs_group)
1014

1015
        return mm_kwargs_combined
1016

1017
    def _dummy_mm_kwargs(self, num_seqs: int) -> BatchedTensorInputs:
1018
        if not self.is_multimodal_raw_input_only_model:
1019
            return {}
1020

1021
1022
1023
1024
1025
        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)
1026

1027
1028
1029
    def _get_cumsum_and_arange(
        self,
        num_tokens: np.ndarray,
1030
        cumsum_dtype: np.dtype | None = None,
1031
1032
1033
1034
1035
1036
1037
1038
1039
1040
1041
1042
1043
1044
1045
1046
    ) -> 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

1047
    def _prepare_input_ids(
1048
1049
1050
1051
        self,
        scheduler_output: "SchedulerOutput",
        total_num_scheduled_tokens: int,
        cu_num_tokens: np.ndarray,
1052
    ) -> None:
1053
        """Prepare the input IDs for the current batch.
1054

1055
1056
1057
1058
1059
1060
1061
        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)
1062
1063
1064
            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)
1065
1066
1067
1068
1069
1070
1071
            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
1072
1073
1074
1075
        sample_flattened_indices: list[int] = []
        spec_flattened_indices: list[int] = []
        prev_common_req_indices: list[int] = []
        prev_draft_token_indices: list[int] = []
1076
1077
        indices_match = True
        max_flattened_index = -1
1078
1079
1080
        total_num_spec_tokens = 0
        scheduled_spec_tokens = scheduler_output.scheduled_spec_decode_tokens

1081
1082
1083
1084
1085
        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.
1086
1087
                draft_len = len(scheduled_spec_tokens.get(req_id, ()))
                total_num_spec_tokens += draft_len
1088
                flattened_index = cu_num_tokens[cur_index].item() - 1
1089
1090
1091
1092
1093
1094
1095
1096
1097
1098
1099
1100
1101
1102
1103
                # 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))
1104
                indices_match &= prev_index == flattened_index
1105
                max_flattened_index = max(max_flattened_index, flattened_index)
1106
1107
1108
        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:
1109
1110
1111
            # 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)
1112
1113
1114
            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)
1115
1116
        if num_commmon_tokens == 0:
            # No requests in common with the previous iteration
1117
            # So input_ids.cpu will have all the input ids.
1118
1119
1120
1121
1122
1123
1124
            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_(
1125
1126
1127
                self.input_batch.prev_sampled_token_ids[:num_commmon_tokens, 0],
                non_blocking=True,
            )
1128
1129
            if self.enable_prompt_embeds:
                self.is_token_ids.gpu[:num_commmon_tokens] = True
1130
            return
1131
        # Upload the index tensors asynchronously so the scatter can be non-blocking.
1132
1133
        sampled_tokens_index_tensor = torch.tensor(
            sample_flattened_indices, dtype=torch.int64, pin_memory=self.pin_memory
1134
        ).to(self.device, non_blocking=True)
1135
        prev_common_req_indices_tensor = torch.tensor(
1136
1137
            prev_common_req_indices, dtype=torch.int64, pin_memory=self.pin_memory
        ).to(self.device, non_blocking=True)
1138
1139
        self.input_ids.gpu.scatter_(
            dim=0,
1140
            index=sampled_tokens_index_tensor,
1141
            src=self.input_batch.prev_sampled_token_ids[
1142
1143
1144
                prev_common_req_indices_tensor, 0
            ],
        )
1145

1146
1147
1148
1149
1150
1151
1152
1153
1154
1155
1156
1157
1158
1159
1160
1161
1162
1163
1164
1165
1166
1167
1168
        # 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],
        )

1169
1170
    def _get_encoder_seq_lens(
        self,
1171
        scheduled_encoder_inputs: dict[str, list[int]],
1172
1173
        kv_cache_spec: KVCacheSpec,
        num_reqs: int,
1174
    ) -> np.ndarray | None:
1175
1176
1177
1178
1179
1180
        if not isinstance(kv_cache_spec, CrossAttentionSpec):
            return None

        # Build encoder_seq_lens array mapping request indices to
        # encoder lengths for inputs scheduled in this batch
        encoder_seq_lens = np.zeros(num_reqs, dtype=np.int32)
1181
        for req_id in scheduled_encoder_inputs:
1182
1183
1184
1185
1186
            req_index = self.input_batch.req_id_to_index[req_id]
            encoder_seq_lens[req_index] = self.max_encoder_len

        return encoder_seq_lens

1187
    def _prepare_inputs(
1188
1189
1190
1191
        self,
        scheduler_output: "SchedulerOutput",
        num_scheduled_tokens: np.ndarray,
        max_num_scheduled_tokens: int,
1192
1193
    ) -> tuple[
        torch.Tensor,
1194
1195
1196
        SpecDecodeMetadata | None,
        UBatchSlices | None,
        torch.Tensor | None,
1197
    ]:
1198
1199
        """
        :return: tuple[
1200
            logits_indices, spec_decode_metadata,
1201
            ubatch_slices, num_tokens_across_dp,
1202
1203
        ]
        """
1204
1205
1206
1207
1208
1209
1210
        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.
1211
        self.input_batch.block_table.commit_block_table(num_reqs)
1212
1213
1214

        # Get request indices.
        # E.g., [2, 5, 3] -> [0, 0, 1, 1, 1, 1, 1, 2, 2, 2]
1215
        req_indices = np.repeat(self.arange_np[:num_reqs], num_scheduled_tokens)
1216

1217
1218
        # cu_num_tokens: [2, 5, 3] -> [2, 7, 10]
        # arange: [0, 1, 0, 1, 2, 3, 4, 0, 1, 2]
1219
        cu_num_tokens, arange = self._get_cumsum_and_arange(num_scheduled_tokens)
1220
1221

        # Get positions.
1222
        positions_np = self.positions.np[:total_num_scheduled_tokens]
1223
1224
1225
1226
1227
        np.add(
            self.input_batch.num_computed_tokens_cpu[req_indices],
            arange,
            out=positions_np,
        )
1228

1229
1230
        # Calculate M-RoPE positions.
        # Only relevant for models using M-RoPE (e.g, Qwen2-VL)
1231
        if self.uses_mrope:
1232
1233
            self._calc_mrope_positions(scheduler_output)

1234
1235
1236
1237
        # 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.
1238
1239
1240
        token_indices = (
            positions_np + req_indices * self.input_batch.token_ids_cpu.shape[1]
        )
1241
        token_indices_tensor = torch.from_numpy(token_indices)
1242

1243
1244
1245
        # 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.
1246
1247
1248
1249
1250
1251
        torch.index_select(
            self.input_batch.token_ids_cpu_tensor.flatten(),
            0,
            token_indices_tensor,
            out=self.input_ids.cpu[:total_num_scheduled_tokens],
        )
1252
        if self.enable_prompt_embeds:
1253
            is_token_ids = self.input_batch.is_token_ids_tensor.flatten()
1254
1255
1256
1257
            torch.index_select(
                is_token_ids,
                0,
                token_indices_tensor,
1258
1259
                out=self.is_token_ids.cpu[:total_num_scheduled_tokens],
            )
1260
1261
1262
1263
1264
1265
1266
1267
1268
1269
1270
1271
1272
1273
1274
1275
1276
1277
1278
1279
1280
1281
1282
1283
1284
1285
1286
1287
1288
1289
1290
1291
1292

        # 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:
1293
1294
1295
                    self.inputs_embeds.cpu[
                        output_idx : output_idx + actual_num_sched
                    ].copy_(req_embeds[start_pos:actual_end])
1296
1297

                output_idx += num_sched
1298

1299
1300
        self.input_batch.block_table.compute_slot_mapping(req_indices, positions_np)
        self.input_batch.block_table.commit_slot_mapping(total_num_scheduled_tokens)
1301
1302

        # Prepare the attention metadata.
1303
        self.query_start_loc.np[0] = 0
1304
        self.query_start_loc.np[1 : num_reqs + 1] = cu_num_tokens
1305
1306
        # Note: pad query_start_loc to be non-decreasing, as kernels
        # like FlashAttention requires that
1307
        self.query_start_loc.np[num_reqs + 1 :].fill(cu_num_tokens[-1])
1308
        self.query_start_loc.copy_to_gpu()
1309
        query_start_loc = self.query_start_loc.gpu[: num_reqs + 1]
1310

1311
        num_tokens_unpadded = scheduler_output.total_num_scheduled_tokens
1312
        num_tokens_padded = self._get_num_input_tokens(num_tokens_unpadded)
1313
1314
1315
        uniform_decode = (
            max_num_scheduled_tokens == self.uniform_decode_query_len
        ) and (total_num_scheduled_tokens == num_reqs * max_num_scheduled_tokens)
1316
1317
1318
1319
1320
1321
1322

        # Disable DP padding when running eager to avoid excessive padding when
        # running prefills. This lets us set enforce_eager on the prefiller in
        # a P/D setup and still use CUDA graphs (enabled by this padding) on the
        # decoder.
        allow_dp_padding = self.compilation_config.cudagraph_mode != CUDAGraphMode.NONE

1323
        ubatch_slices, num_tokens_across_dp = coordinate_batch_across_dp(
1324
1325
1326
1327
1328
1329
1330
            num_tokens_unpadded=num_tokens_unpadded,
            parallel_config=self.parallel_config,
            allow_microbatching=True,
            allow_dp_padding=allow_dp_padding,
            num_tokens_padded=num_tokens_padded,
            uniform_decode=uniform_decode,
            num_scheduled_tokens_per_request=num_scheduled_tokens,
1331
        )
1332

1333
        self.seq_lens.np[:num_reqs] = (
1334
1335
            self.input_batch.num_computed_tokens_cpu[:num_reqs] + num_scheduled_tokens
        )
1336
        # Fill unused with 0 for full cuda graph mode.
1337
1338
        self.seq_lens.np[num_reqs:].fill(0)
        self.seq_lens.copy_to_gpu()
1339

1340
        num_tokens = [self.requests[r].num_tokens for r in self.input_batch.req_ids]
1341
1342
1343
1344
1345
1346
1347
        num_tokens_np = np.array(num_tokens, dtype=np.int32)

        # Record the index of requests that should not be sampled,
        # so that we could clear the sampled tokens before returning
        discard_requests_mask = self.seq_lens.np[:num_reqs] < num_tokens_np
        discard_request_indices = np.nonzero(discard_requests_mask)[0]
        self.num_discarded_requests = len(discard_request_indices)
1348
1349
1350
        self.discard_request_indices.np[: self.num_discarded_requests] = (
            discard_request_indices
        )
1351
1352
1353

        self.discard_request_indices.copy_to_gpu(self.num_discarded_requests)

1354
        # Copy the tensors to the GPU.
1355
1356
1357
1358
1359
        self._prepare_input_ids(
            scheduler_output,
            total_num_scheduled_tokens,
            cu_num_tokens,
        )
1360

1361
        if self.uses_mrope:
1362
            # Only relevant for models using M-RoPE (e.g, Qwen2-VL)
1363
1364
            self.mrope_positions.gpu[:, :total_num_scheduled_tokens].copy_(
                self.mrope_positions.cpu[:, :total_num_scheduled_tokens],
1365
1366
                non_blocking=True,
            )
1367
1368
        else:
            # Common case (1D positions)
1369
            self.positions.copy_to_gpu(total_num_scheduled_tokens)
1370

1371
        use_spec_decode = len(scheduler_output.scheduled_spec_decode_tokens) > 0
1372
1373
1374
1375
1376
1377
1378
        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
1379
            num_draft_tokens = None
1380
            spec_decode_metadata = None
1381
            num_sampled_tokens = np.ones(num_reqs, dtype=np.int32)
1382
1383
1384
1385
1386
        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)
1387
1388
1389
            # 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)
1390
1391
1392
1393
            for (
                req_id,
                draft_token_ids,
            ) in scheduler_output.scheduled_spec_decode_tokens.items():
1394
1395
                req_idx = self.input_batch.req_id_to_index[req_id]
                num_draft_tokens[req_idx] = len(draft_token_ids)
1396
1397
1398
1399
1400
1401
1402
1403
                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
                )
1404
            spec_decode_metadata = self._calc_spec_decode_metadata(
1405
1406
                num_draft_tokens, cu_num_tokens
            )
1407
            logits_indices = spec_decode_metadata.logits_indices
1408
            num_sampled_tokens = num_draft_tokens + 1
1409
            # For DECODE only cuda graph of some attention backends (e.g., GDN).
1410
            self.num_decode_draft_tokens.np[:num_reqs] = num_decode_draft_tokens
1411
1412
            self.num_decode_draft_tokens.np[num_reqs:].fill(-1)
            self.num_decode_draft_tokens.copy_to_gpu()
1413

1414
1415
1416
1417
1418
        # Hot-Swap lora model
        if self.lora_config:
            assert (
                np.sum(num_sampled_tokens)
                <= self.vllm_config.scheduler_config.max_num_batched_tokens
1419
            )
1420
1421
1422
1423
1424
1425
1426
1427
1428
1429
1430
1431
1432
1433
1434
1435
1436
1437
1438
1439
1440
1441
1442
1443
1444
1445
1446
            self.set_active_loras(
                self.input_batch, num_scheduled_tokens, num_sampled_tokens
            )

        return (
            logits_indices,
            spec_decode_metadata,
            ubatch_slices,
            num_tokens_across_dp,
        )

    def _build_attention_metadata(
        self,
        total_num_scheduled_tokens: int,
        max_num_scheduled_tokens: int,
        num_reqs: int,
        ubatch_slices: UBatchSlices | None = None,
        logits_indices: torch.Tensor | None = None,
        use_spec_decode: bool = False,
        for_cudagraph_capture: bool = False,
        scheduled_encoder_inputs: dict[str, list[int]] | None = None,
        cascade_attn_prefix_lens: list[list[int]] | None = None,
    ) -> tuple[PerLayerAttnMetadata, CommonAttentionMetadata | None]:
        """
        :return: tuple[attn_metadata, spec_decode_common_attn_metadata]
        """
        logits_indices_padded = None
1447
        num_logits_indices = None
1448
1449
1450
1451
1452
1453
        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
                )
1454

1455
1456
1457
1458
1459
1460
        # 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,
1461
                self.parallel_config.cp_kv_cache_interleave_size,
1462
1463
1464
            )
            self.dcp_local_seq_lens.copy_to_gpu(num_reqs)

1465
1466
1467
        attn_metadata: PerLayerAttnMetadata = {}
        if ubatch_slices is not None:
            attn_metadata = [dict() for _ in range(len(ubatch_slices))]
1468

1469
1470
        # Used in the below loop
        query_start_loc = self.query_start_loc.gpu[: num_reqs + 1]
1471
        query_start_loc_cpu = self.query_start_loc.cpu[: num_reqs + 1]
1472
        seq_lens = self.seq_lens.gpu[:num_reqs]
1473
        seq_lens_cpu = self.seq_lens.cpu[:num_reqs]
1474
1475
1476
        num_computed_tokens_cpu = self.input_batch.num_computed_tokens_cpu_tensor[
            :num_reqs
        ]
1477
1478
1479
1480
1481
1482

        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]
            dcp_local_seq_lens_cpu = self.dcp_local_seq_lens.cpu[:num_reqs]

1483
        spec_decode_common_attn_metadata = None
1484
1485
1486
1487
1488
1489
1490
1491
1492

        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()

1493
1494
        if use_spec_decode:
            self.num_accepted_tokens.np[:num_reqs] = (
1495
1496
                self.input_batch.num_accepted_tokens_cpu[:num_reqs]
            )
1497
1498
            self.num_accepted_tokens.np[num_reqs:].fill(1)
            self.num_accepted_tokens.copy_to_gpu()
1499

1500
1501
        # Prepare the attention metadata for each KV cache group and make layers
        # in the same group share the same metadata.
1502
        for kv_cache_gid, kv_cache_group in enumerate(
1503
1504
            self.kv_cache_config.kv_cache_groups
        ):
1505
            encoder_seq_lens = self._get_encoder_seq_lens(
1506
1507
1508
                scheduled_encoder_inputs or {},
                kv_cache_group.kv_cache_spec,
                num_reqs,
1509
            )
1510

1511
            if isinstance(kv_cache_group.kv_cache_spec, EncoderOnlyAttentionSpec):
1512
1513
1514
1515
1516
                # 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(
                    (num_reqs, 1),
                    dtype=torch.int32,
1517
1518
1519
                    device=self.device,
                )
                slot_mapping = torch.zeros(
1520
                    (total_num_scheduled_tokens,),
1521
1522
1523
                    dtype=torch.int64,
                    device=self.device,
                )
1524
            else:
1525
                blk_table = self.input_batch.block_table[kv_cache_gid]
1526
                blk_table_tensor = blk_table.get_device_tensor(num_reqs)
1527
                slot_mapping = blk_table.slot_mapping.gpu[:total_num_scheduled_tokens]
1528
1529
1530

                # Fill unused with -1. Needed for reshape_and_cache in full cuda
                # graph mode.
1531
                blk_table.slot_mapping.gpu[total_num_scheduled_tokens:].fill_(-1)
1532

1533
            common_attn_metadata = CommonAttentionMetadata(
1534
1535
1536
1537
1538
                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,
1539
1540
1541
                num_reqs=num_reqs,
                num_actual_tokens=total_num_scheduled_tokens,
                max_query_len=max_num_scheduled_tokens,
1542
                max_seq_len=max_seq_len,
1543
1544
                block_table_tensor=blk_table_tensor,
                slot_mapping=slot_mapping,
1545
                logits_indices_padded=logits_indices_padded,
1546
                num_logits_indices=num_logits_indices,
1547
                causal=True,
1548
                encoder_seq_lens=encoder_seq_lens,
1549
                dcp_local_seq_lens=dcp_local_seq_lens,
1550
                dcp_local_seq_lens_cpu=dcp_local_seq_lens_cpu,
1551
1552
            )

1553
            if self.speculative_config and spec_decode_common_attn_metadata is None:
1554
                if isinstance(self.drafter, EagleProposer):
1555
                    if self.drafter.attn_layer_names[0] in kv_cache_group.layer_names:
1556
1557
1558
                        spec_decode_common_attn_metadata = common_attn_metadata
                else:
                    spec_decode_common_attn_metadata = common_attn_metadata
1559

1560
1561
1562
1563
1564
1565
            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
                )
1566
                builder = attn_group.get_metadata_builder()
1567

1568
                extra_attn_metadata_args = {}
1569
                if use_spec_decode and isinstance(builder, GDNAttentionMetadataBuilder):
1570
                    extra_attn_metadata_args = dict(
1571
1572
1573
1574
                        num_accepted_tokens=self.num_accepted_tokens.gpu[:num_reqs],
                        num_decode_draft_tokens_cpu=self.num_decode_draft_tokens.cpu[
                            :num_reqs
                        ],
1575
1576
                    )

1577
1578
                if ubatch_slices is not None:
                    common_attn_metadata_list = split_attn_metadata(
1579
1580
                        ubatch_slices, common_attn_metadata
                    )
1581
                    for ubid, common_attn_metadata in enumerate(
1582
1583
                        common_attn_metadata_list
                    ):
1584
1585
1586
1587
1588
1589
1590
1591
1592
1593
1594
                        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:
1595
1596
1597
1598
                            assert type(attn_metadata) is list
                            attn_metadata[ubid][layer_name] = attn_metadata_i
                else:
                    assert isinstance(attn_metadata, dict)
1599
1600
1601
1602
1603
1604
1605
1606
1607
1608
                    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,
                        )
1609
1610
                    for layer_name in attn_group.layer_names:
                        attn_metadata[layer_name] = attn_metadata_i
1611

1612
        return attn_metadata, spec_decode_common_attn_metadata
1613

1614
1615
1616
1617
1618
1619
1620
1621
1622
1623
    def _compute_cascade_attn_prefix_lens(
        self,
        num_scheduled_tokens: np.ndarray,
        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
        """
1624

1625
1626
1627
1628
1629
1630
1631
1632
1633
1634
1635
1636
1637
1638
1639
1640
1641
1642
1643
1644
1645
1646
        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,
                        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
1647

1648
1649
1650
1651
    def _compute_cascade_attn_prefix_len(
        self,
        num_scheduled_tokens: np.ndarray,
        num_common_prefix_blocks: int,
1652
1653
        kv_cache_spec: KVCacheSpec,
        attn_metadata_builder: AttentionMetadataBuilder,
1654
1655
1656
1657
1658
1659
1660
1661
1662
1663
1664
1665
1666
1667
1668
1669
1670
1671
    ) -> 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.
        """
1672

1673
        common_prefix_len = num_common_prefix_blocks * kv_cache_spec.block_size
1674
1675
1676
1677
1678
1679
1680
1681
1682
1683
1684
1685
1686
1687
1688
1689
1690
1691
1692
1693
1694
1695
1696
1697
1698
1699
1700
1701
1702
1703
1704
1705
1706
1707
1708
1709
1710
        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]
1711
        # Request 3's num_computed_tokens: 3 (i.e., [A, B, C])
1712
1713
1714
1715
1716
1717
1718
        # If we use [A, B, C, D] as the common prefix for Request 1-3,
        # then Request 3 will be processed only by the first kernel,
        # and the second kernel will get an empty input. While this is not
        # a fundamental problem, our current implementation does not support
        # this case.
        num_reqs = len(num_scheduled_tokens)
        common_prefix_len = min(
1719
1720
            common_prefix_len, self.input_batch.num_computed_tokens_cpu[:num_reqs].min()
        )
1721
        # common_prefix_len should be a multiple of the block size.
1722
1723
1724
1725
1726
1727
1728
1729
1730
1731
1732
        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
        )
1733
1734
        assert isinstance(kv_cache_spec, AttentionSpec)
        use_cascade = attn_metadata_builder.use_cascade_attention(
1735
1736
1737
            common_prefix_len=common_prefix_len,
            query_lens=num_scheduled_tokens,
            num_query_heads=self.num_query_heads,
1738
            num_kv_heads=kv_cache_spec.num_kv_heads,
1739
            use_alibi=self.use_alibi,
1740
            use_sliding_window=use_sliding_window,
1741
            use_local_attention=use_local_attention,
1742
            num_sms=self.num_sms,
1743
            dcp_world_size=self.dcp_world_size,
1744
1745
1746
        )
        return common_prefix_len if use_cascade else 0

1747
1748
    def _calc_mrope_positions(self, scheduler_output: "SchedulerOutput"):
        mrope_pos_ptr = 0
1749
        for index, req_id in enumerate(self.input_batch.req_ids):
1750
1751
1752
            req = self.requests[req_id]
            assert req.mrope_positions is not None

1753
1754
            num_computed_tokens = self.input_batch.num_computed_tokens_cpu[index]
            num_scheduled_tokens = scheduler_output.num_scheduled_tokens[req_id]
1755
            num_prompt_tokens = length_from_prompt_token_ids_or_embeds(
1756
1757
                req.prompt_token_ids, req.prompt_embeds
            )
1758
1759

            if num_computed_tokens + num_scheduled_tokens > num_prompt_tokens:
1760
1761
                prompt_part_len = max(0, num_prompt_tokens - num_computed_tokens)
                completion_part_len = max(0, num_scheduled_tokens - prompt_part_len)
1762
1763
1764
1765
1766
1767
1768
1769
1770
1771
1772
1773
1774
            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

1775
1776
1777
                self.mrope_positions.cpu[:, dst_start:dst_end] = req.mrope_positions[
                    :, src_start:src_end
                ]
1778
1779
1780
1781
1782
1783
1784
                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

1785
                assert req.mrope_position_delta is not None
1786
                MRotaryEmbedding.get_next_input_positions_tensor(
1787
                    out=self.mrope_positions.np,
1788
1789
1790
1791
1792
                    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,
                )
1793
1794
1795

                mrope_pos_ptr += completion_part_len

1796
1797
    def _calc_spec_decode_metadata(
        self,
1798
1799
1800
1801
1802
1803
1804
1805
1806
1807
1808
1809
1810
1811
1812
1813
        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
1814
1815
1816
1817

        # 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(
1818
1819
            num_sampled_tokens, cumsum_dtype=np.int32
        )
1820
        # Step 2. [0, 0, 0, 0, 103, 104, 104, 104, 206, 207, 207]
1821
        logits_indices = np.repeat(
1822
1823
            cu_num_scheduled_tokens - num_sampled_tokens, num_sampled_tokens
        )
1824
        # Step 3. [0, 1, 2, 3, 103, 104, 105, 106, 206, 207, 208]
1825
1826
1827
1828
1829
1830
        logits_indices += arange

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

        # Compute the draft logits indices.
1831
1832
1833
        # 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(
1834
1835
            num_draft_tokens, cumsum_dtype=np.int32
        )
1836
1837
        # [0, 0, 0, 5, 5, 9]
        target_logits_indices = np.repeat(
1838
1839
            cu_num_sampled_tokens - num_sampled_tokens, num_draft_tokens
        )
1840
1841
1842
1843
1844
        # [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(
1845
1846
            self.device, non_blocking=True
        )
1847
1848
1849
        cu_num_sampled_tokens = torch.from_numpy(cu_num_sampled_tokens).to(
            self.device, non_blocking=True
        )
1850
1851
1852
        logits_indices = torch.from_numpy(logits_indices).to(
            self.device, non_blocking=True
        )
1853
        target_logits_indices = torch.from_numpy(target_logits_indices).to(
1854
1855
            self.device, non_blocking=True
        )
1856
        bonus_logits_indices = torch.from_numpy(bonus_logits_indices).to(
1857
1858
            self.device, non_blocking=True
        )
1859

1860
1861
        # Compute the draft token ids.
        # draft_token_indices:      [  1,   2,   3, 105, 106, 208]
1862
        draft_token_ids = self.input_ids.gpu[logits_indices]
1863
1864
        draft_token_ids = draft_token_ids[target_logits_indices + 1]

1865
        return SpecDecodeMetadata(
1866
1867
1868
            draft_token_ids=draft_token_ids,
            num_draft_tokens=num_draft_tokens.tolist(),
            cu_num_draft_tokens=cu_num_draft_tokens,
1869
            cu_num_sampled_tokens=cu_num_sampled_tokens,
1870
1871
1872
1873
1874
            target_logits_indices=target_logits_indices,
            bonus_logits_indices=bonus_logits_indices,
            logits_indices=logits_indices,
        )

1875
1876
1877
1878
1879
1880
1881
    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
1882
        self.kv_sharing_fast_prefill_logits_indices[:num_logits].copy_(logits_indices)
1883
1884
1885
1886
1887
        # 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_(
1888
1889
1890
1891
1892
1893
            logits_indices[-1].item()
        )
        if (
            self.compilation_config.cudagraph_mode != CUDAGraphMode.NONE
            and num_logits <= self.cudagraph_batch_sizes[-1]
        ):
1894
1895
1896
1897
1898
            # 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
1899
1900
1901
        logits_indices_padded = self.kv_sharing_fast_prefill_logits_indices[
            :num_logits_padded
        ]
1902
1903
        return logits_indices_padded

1904
1905
1906
1907
1908
1909
1910
1911
    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
1912
                inputs.
1913
1914
1915
1916
1917
1918

        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
        """
1919
1920
        scheduled_encoder_inputs = scheduler_output.scheduled_encoder_inputs
        if not scheduled_encoder_inputs:
1921
            return [], []
1922
        # Batch the multi-modal inputs.
1923
        mm_kwargs = list[MultiModalKwargsItem]()
1924
1925
        # list of tuple (mm_hash, position_info)
        mm_hashes_pos = list[tuple[str, PlaceholderRange]]()
1926
1927
        for req_id, encoder_input_ids in scheduled_encoder_inputs.items():
            req_state = self.requests[req_id]
1928
1929

            for mm_input_id in encoder_input_ids:
1930
                mm_feature = req_state.mm_features[mm_input_id]
1931
1932
                if mm_feature.data is None:
                    continue
1933
1934
1935
                mm_hash = mm_feature.identifier
                mm_kwargs.append(mm_feature.data)
                mm_hashes_pos.append((mm_hash, mm_feature.mm_position))
1936

1937
1938
        return mm_kwargs, mm_hashes_pos

1939
1940
1941
    def _execute_mm_encoder(
        self, scheduler_output: "SchedulerOutput"
    ) -> list[torch.Tensor]:
1942
1943
        # Batch the multi-modal inputs using the helper method.
        mm_kwargs, mm_hashes_pos = self._batch_mm_kwargs_from_scheduler(
1944
1945
            scheduler_output
        )
1946
1947

        if not mm_kwargs:
1948
            return []
1949

1950
1951
1952
1953
1954
1955
1956
        # 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.
1957
        model = cast(SupportsMultiModal, self.model)
1958
        encoder_outputs: list[torch.Tensor] = []
1959
        for modality, num_items, mm_kwargs_group in group_mm_kwargs_by_modality(
1960
1961
1962
1963
            mm_kwargs,
            device=self.device,
            pin_memory=self.pin_memory,
            merge_by_field_config=model.merge_by_field_config,
1964
            multimodal_cpu_fields=model.multimodal_cpu_fields,
1965
        ):
1966
            curr_group_outputs: list[torch.Tensor] = []
1967
1968

            # EVS-related change.
1969
            # (ekhvedchenia): Temporary hack to limit peak memory usage when
1970
            # processing multimodal data. This solves the issue with scheduler
1971
1972
1973
1974
            # 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)
1975
1976
1977
1978
1979
1980
1981
1982
1983
1984
1985
1986
1987
1988
1989
1990
            # 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,
1991
                            multimodal_cpu_fields=model.multimodal_cpu_fields,
1992
                        )
1993
                    )
1994

1995
                    micro_batch_outputs = model.embed_multimodal(
1996
1997
                        **micro_batch_mm_inputs
                    )
1998
1999
2000
2001
2002
2003
2004
2005
2006
2007

                    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.
2008
                curr_group_outputs = model.embed_multimodal(**mm_kwargs_group)  # type: ignore[assignment]
2009

2010
2011
            sanity_check_mm_encoder_outputs(
                curr_group_outputs,
2012
                expected_num_items=num_items,
2013
            )
2014
            encoder_outputs.extend(curr_group_outputs)
2015

2016
2017
2018
        # 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(
2019
2020
2021
                output,
                is_embed=pos_info.is_embed,
            )
2022
2023
            logger.debug("Finish execute for mm hash %s", mm_hash)
            self.maybe_save_ec_to_connector(self.encoder_cache, mm_hash)
2024

2025
2026
        return encoder_outputs

2027
    def _gather_mm_embeddings(
2028
2029
        self,
        scheduler_output: "SchedulerOutput",
2030
        shift_computed_tokens: int = 0,
2031
2032
2033
2034
2035
2036
2037
2038
    ) -> 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
2039
        should_sync_mrope_positions = False
2040

2041
        for req_id in self.input_batch.req_ids:
2042
2043
            mm_embeds_req: list[torch.Tensor] = []

2044
            num_scheduled_tokens = scheduler_output.num_scheduled_tokens[req_id]
2045
            req_state = self.requests[req_id]
2046
            num_computed_tokens = req_state.num_computed_tokens + shift_computed_tokens
2047

2048
2049
            for mm_feature in req_state.mm_features:
                pos_info = mm_feature.mm_position
2050
2051
                start_pos = pos_info.offset
                num_encoder_tokens = pos_info.length
2052
2053
2054
2055
2056
2057
2058
2059
2060
2061
2062
2063
2064
2065
2066
2067

                # 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,
2068
2069
                    num_encoder_tokens,
                )
2070
                assert start_idx < end_idx
2071

2072
                mm_hash = mm_feature.identifier
2073
                encoder_output = self.encoder_cache.get(mm_hash, None)
2074
                assert encoder_output is not None, f"Encoder cache miss for {mm_hash}."
2075
2076
2077
2078

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

2079
                req_start_pos = req_start_idx + start_pos - num_computed_tokens
2080
2081
2082
                is_mm_embed[req_start_pos + start_idx : req_start_pos + end_idx] = (
                    True if is_embed is None else is_embed
                )
2083

2084
2085
2086
2087
                mm_embeds_item = gather_mm_placeholders(
                    encoder_output[start_idx:end_idx],
                    is_embed=is_embed,
                )
2088
2089
2090
                mm_embeds_req.append(mm_embeds_item)

            if self.is_multimodal_pruning_enabled and self.uses_mrope:
2091
                assert req_state.mrope_positions is not None
2092
2093
2094
2095
2096
2097
2098
                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,
2099
2100
                    )
                )
2101
2102
2103
2104
                req_state.mrope_positions.copy_(new_mrope_positions)
                req_state.mrope_position_delta = new_delta

            mm_embeds.extend(mm_embeds_req)
2105
2106
2107
            req_start_idx += num_scheduled_tokens

        is_mm_embed = self.is_mm_embed.copy_to_gpu(total_num_scheduled_tokens)
2108
2109
2110

        if should_sync_mrope_positions:
            self._calc_mrope_positions(scheduler_output)
2111
            self.mrope_positions.copy_to_gpu(total_num_scheduled_tokens)
2112

2113
        return mm_embeds, is_mm_embed
2114

2115
    def get_model(self) -> nn.Module:
2116
        # get raw model out of the cudagraph wrapper.
2117
        if isinstance(self.model, (CUDAGraphWrapper, UBatchWrapper)):
2118
            return self.model.unwrap()
2119
2120
        return self.model

2121
2122
2123
2124
2125
2126
2127
2128
2129
2130
2131
2132
2133
2134
2135
    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

2136
2137
2138
2139
2140
    def get_supported_pooling_tasks(self) -> list[PoolingTask]:
        model = self.get_model()
        if not is_pooling_model(model):
            return []

2141
2142
        supported_tasks = list(model.pooler.get_supported_tasks())

2143
        if self.scheduler_config.enable_chunked_prefill:
2144
2145
2146
2147
            if "token_embed" in supported_tasks:
                supported_tasks.remove("token_embed")
            if "token_classify" in supported_tasks:
                supported_tasks.remove("token_classify")
2148

2149
2150
            logger.debug_once(
                "Chunked prefill is not supported with "
2151
2152
                "token_embed and token_classify tasks "
                "which using ALL pooling. "
2153
2154
2155
                "Please turn off chunked prefill by "
                "`--no-enable-chunked-prefill` before using it."
            )
2156
2157
2158
2159
2160

        if "score" in supported_tasks:
            num_labels = getattr(self.model_config.hf_config, "num_labels", 0)
            if num_labels != 1:
                supported_tasks.remove("score")
2161
                logger.debug_once("Score API is only enabled for num_labels == 1.")
2162
2163

        return supported_tasks
2164

2165
2166
2167
2168
2169
2170
2171
2172
2173
2174
    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)

2175
    def sync_and_slice_intermediate_tensors(
2176
2177
        self,
        num_tokens: int,
2178
        intermediate_tensors: IntermediateTensors | None,
2179
2180
        sync_self: bool,
    ) -> IntermediateTensors:
2181
2182
2183
        assert self.intermediate_tensors is not None

        tp = self.vllm_config.parallel_config.tensor_parallel_size
2184
        is_rs = is_residual_scattered_for_sp(self.vllm_config, num_tokens)
2185
2186
2187
2188
2189
2190

        # 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():
2191
                is_scattered = k == "residual" and is_rs
2192
                copy_len = num_tokens // tp if is_scattered else num_tokens
2193
                self.intermediate_tensors[k][:copy_len].copy_(
2194
2195
2196
2197
2198
2199
2200
2201
2202
2203
2204
2205
2206
                    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:
2207
2208
2209
2210
2211
2212
2213
        """
        Step for the EPLB (Expert Parallelism Load Balancing) state.
        """
        if not self.parallel_config.enable_eplb:
            return

        assert self.eplb_state is not None
2214
2215
        model = self.get_model()
        assert is_mixture_of_experts(model)
2216
2217
2218
        self.eplb_state.step(
            is_dummy,
            is_profile,
2219
            log_stats=self.parallel_config.eplb_config.log_balancedness,
2220
2221
        )

2222
2223
2224
2225
    # This is where the second ubatch is adjusted to account for the padding.
    # Should be called after attention metadata creation. This just pads
    # the second ubatch slice out to the total number of tokens
    # (num_tokens + padding)
2226
2227
    @staticmethod
    def pad_out_ubatch_slice(ubatch_slices: UBatchSlices, num_total_tokens: int):
2228
2229
2230
2231
2232
2233
        padded_second_ubatch_slice = slice(
            ubatch_slices[1].token_slice.start, num_total_tokens
        )
        ubatch_slices[1] = UBatchSlice(
            padded_second_ubatch_slice, padded_second_ubatch_slice
        )
2234

2235
2236
2237
2238
2239
2240
    def _pool(
        self,
        hidden_states: torch.Tensor,
        num_scheduled_tokens: int,
        num_scheduled_tokens_np: np.ndarray,
    ) -> ModelRunnerOutput:
2241
2242
2243
        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"
        )
2244

2245
        hidden_states = hidden_states[:num_scheduled_tokens]
2246
        pooling_metadata = self.input_batch.get_pooling_metadata()
2247
2248
2249
2250
        pooling_metadata.build_pooling_cursor(
            num_scheduled_tokens_np.tolist(), device=hidden_states.device
        )
        seq_lens_cpu = self.seq_lens.cpu[: self.input_batch.num_reqs]
2251

2252
2253
2254
2255
2256
2257
2258
2259
2260
2261
        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(
            lambda x: x.to("cpu", non_blocking=True),
            raw_pooler_output,
        )
        self._sync_device()
2262

2263
        pooler_output: list[torch.Tensor | None] = []
2264
        for raw_output, seq_len, prompt_len in zip(
2265
2266
            raw_pooler_output, seq_lens_cpu, pooling_metadata.prompt_lens
        ):
2267
            output = raw_output if seq_len == prompt_len else None
2268
            pooler_output.append(output)
2269
2270
2271
2272
2273
2274
2275
2276
2277
2278

        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,
        )

2279
    def _get_num_input_tokens(self, num_scheduled_tokens: int) -> int:
2280
2281
2282
2283
2284
2285
        if (
            self.compilation_config.cudagraph_mode != CUDAGraphMode.NONE
            and hasattr(self, "cudagraph_batch_sizes")
            and self.cudagraph_batch_sizes
            and num_scheduled_tokens <= self.cudagraph_batch_sizes[-1]
        ):
2286
2287
2288
2289
2290
2291
2292
2293
            # Use CUDA graphs.
            # Add padding to the batch size.
            return self.vllm_config.pad_for_cudagraph(num_scheduled_tokens)

        # Eager mode.
        # Pad tokens to multiple of tensor_parallel_size when
        # enabled collective fusion for SP
        tp_size = self.vllm_config.parallel_config.tensor_parallel_size
2294
2295
2296
2297
        if (
            self.compilation_config.pass_config.enable_sequence_parallelism
            and tp_size > 1
        ):
2298
2299
2300
            return round_up(num_scheduled_tokens, tp_size)
        return num_scheduled_tokens

2301
    def _preprocess(
2302
2303
        self,
        scheduler_output: "SchedulerOutput",
2304
        num_input_tokens: int,  # Padded
2305
        intermediate_tensors: IntermediateTensors | None = None,
2306
    ) -> tuple[
2307
2308
        torch.Tensor | None,
        torch.Tensor | None,
2309
        torch.Tensor,
2310
        IntermediateTensors | None,
2311
        dict[str, Any],
2312
        ECConnectorOutput | None,
2313
    ]:
2314
        num_scheduled_tokens = scheduler_output.total_num_scheduled_tokens
2315
        is_first_rank = get_pp_group().is_first_rank
2316

2317
2318
        # _prepare_inputs may reorder the batch, so we must gather multi
        # modal outputs after that to ensure the correct order
2319
2320
        ec_connector_output = None

2321
2322
        if (
            self.supports_mm_inputs
2323
            and is_first_rank
2324
2325
            and not self.model_config.is_encoder_decoder
        ):
2326
            # Run the multimodal encoder if any.
2327
2328
2329
2330
2331
2332
            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)
2333

2334
2335
2336
            # 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.
2337
            inputs_embeds_scheduled = self.model.embed_input_ids(
2338
2339
2340
                self.input_ids.gpu[:num_scheduled_tokens],
                multimodal_embeddings=mm_embeds,
                is_multimodal=is_mm_embed,
2341
            )
2342

2343
            # TODO(woosuk): Avoid the copy. Optimize.
2344
            self.inputs_embeds.gpu[:num_scheduled_tokens].copy_(inputs_embeds_scheduled)
2345

2346
            input_ids = None
2347
            inputs_embeds = self.inputs_embeds.gpu[:num_input_tokens]
2348
2349
2350
2351
            model_kwargs = {
                **self._init_model_kwargs(num_scheduled_tokens),
                **self._extract_mm_kwargs(scheduler_output),
            }
2352
        elif self.enable_prompt_embeds and is_first_rank:
2353
2354
2355
2356
2357
2358
2359
2360
2361
2362
2363
2364
            # 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).
2365
2366
2367
            token_ids_idx = (
                self.is_token_ids.gpu[:num_scheduled_tokens]
                .nonzero(as_tuple=False)
2368
                .squeeze(1)
2369
            )
2370
2371
2372
            # Some tokens ids may need to become embeds
            if token_ids_idx.numel() > 0:
                token_ids = self.input_ids.gpu[token_ids_idx]
2373
                tokens_to_embeds = self.model.embed_input_ids(input_ids=token_ids)
2374
2375
2376
2377
2378
                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
2379
        else:
2380
2381
2382
2383
            # 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.
2384
            input_ids = self.input_ids.gpu[:num_input_tokens]
2385
            inputs_embeds = None
2386
            model_kwargs = self._init_model_kwargs(num_input_tokens)
2387
        if self.uses_mrope:
2388
            positions = self.mrope_positions.gpu[:, :num_input_tokens]
2389
        else:
2390
            positions = self.positions.gpu[:num_input_tokens]
2391

2392
        if is_first_rank:
2393
2394
            intermediate_tensors = None
        else:
2395
            assert intermediate_tensors is not None
2396
            intermediate_tensors = self.sync_and_slice_intermediate_tensors(
2397
2398
                num_input_tokens, intermediate_tensors, True
            )
2399

2400
2401
2402
2403
        if (
            self.model_config.is_encoder_decoder
            and scheduler_output.scheduled_encoder_inputs
        ):
2404
2405
2406
2407
2408
2409
2410
            # 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})
2411

2412
2413
2414
2415
2416
2417
        return (
            input_ids,
            inputs_embeds,
            positions,
            intermediate_tensors,
            model_kwargs,
2418
            ec_connector_output,
2419
        )
2420

2421
    def _sample(
2422
        self,
2423
2424
        logits: torch.Tensor | None,
        spec_decode_metadata: SpecDecodeMetadata | None,
2425
    ) -> SamplerOutput:
2426
        # Sample the next token and get logprobs if needed.
2427
        sampling_metadata = self.input_batch.sampling_metadata
2428
        if spec_decode_metadata is None:
2429
2430
2431
            # 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()
2432
            return self.sampler(
2433
2434
2435
                logits=logits,
                sampling_metadata=sampling_metadata,
            )
2436

2437
        sampler_output = self.rejection_sampler(
2438
2439
            spec_decode_metadata,
            None,  # draft_probs
2440
            logits,
2441
2442
            sampling_metadata,
        )
2443
        self._update_states_after_model_execute(sampler_output.sampled_token_ids)
2444
2445
2446
        return sampler_output

    def _bookkeeping_sync(
2447
2448
2449
        self,
        scheduler_output: "SchedulerOutput",
        sampler_output: SamplerOutput,
2450
        logits: torch.Tensor | None,
2451
2452
        hidden_states: torch.Tensor,
        num_scheduled_tokens: int,
2453
        spec_decode_metadata: SpecDecodeMetadata | None,
2454
    ) -> tuple[
2455
        dict[str, int],
2456
        LogprobsLists | None,
2457
        list[list[int]],
2458
        dict[str, LogprobsTensors | None],
2459
2460
2461
        list[str],
        dict[str, int],
        list[int],
2462
    ]:
2463
2464
2465
2466
        num_nans_in_logits = {}
        if envs.VLLM_COMPUTE_NANS_IN_LOGITS:
            num_nans_in_logits = self._get_nans_in_logits(logits)

2467
2468
2469
        discard_sampled_tokens_req_indices = self.discard_request_indices.np[
            : self.num_discarded_requests
        ]
2470
2471
2472
2473
        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)
2474

2475
2476
2477
        # 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()
2478
        req_id_to_index_output_copy = self.input_batch.req_id_to_index.copy()
2479
2480

        num_sampled_tokens = sampler_output.sampled_token_ids.shape[0]
2481
        sampled_token_ids = sampler_output.sampled_token_ids
2482
        logprobs_tensors = sampler_output.logprobs_tensors
2483
        invalid_req_indices = []
2484
        cu_num_new_tokens: list[int] | None = None
2485
2486
2487
2488
2489
2490
2491
2492
2493
2494
2495
2496
        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)
            else:
                # Includes spec decode tokens.
                valid_sampled_token_ids = self.rejection_sampler.parse_output(
                    sampled_token_ids,
                    self.input_batch.vocab_size,
                )
2497
2498
2499
2500
2501
2502
                if logprobs_tensors:
                    # Needed for extracting logprobs when spec decoding.
                    # This must be done prior to discarding sampled tokens.
                    cu_num_new_tokens = [0]
                    for toks in valid_sampled_token_ids:
                        cu_num_new_tokens.append(cu_num_new_tokens[-1] + len(toks))
2503
2504
            # Mask out the sampled tokens that should not be sampled.
            for i in discard_sampled_tokens_req_indices:
2505
                valid_sampled_token_ids[int(i)].clear()
2506
        else:
2507
            valid_sampled_token_ids = []
2508
            invalid_req_indices = discard_sampled_tokens_req_indices.tolist()
2509
2510
2511
2512
2513
            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.
2514
2515
2516
2517
            # 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
2518
2519
2520
2521
2522
            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
            }
2523

2524
2525
2526
2527
2528
        # 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.
2529
        req_ids = self.input_batch.req_ids
2530
2531
        for req_idx in range(num_sampled_tokens):
            if self.use_async_scheduling:
2532
                sampled_ids = [-1] if req_idx not in invalid_req_indices_set else None
2533
2534
            else:
                sampled_ids = valid_sampled_token_ids[req_idx]
2535

2536
            num_sampled_ids: int = len(sampled_ids) if sampled_ids else 0
2537

2538
            if not sampled_ids:
2539
2540
2541
                continue

            start_idx = self.input_batch.num_tokens_no_spec[req_idx]
2542
            end_idx = start_idx + num_sampled_ids
2543
2544
2545
2546
            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}"
2547
            )
2548

2549
2550
            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
2551
2552
            self.input_batch.num_tokens_no_spec[req_idx] = end_idx
            self.input_batch.num_tokens[req_idx] = end_idx
2553

2554
            req_id = req_ids[req_idx]
2555
2556
2557
            req_state = self.requests[req_id]
            req_state.output_token_ids.extend(sampled_ids)

2558
        logprobs_lists = (
2559
            logprobs_tensors.tolists(cu_num_new_tokens)
2560
            if not self.use_async_scheduling and logprobs_tensors is not None
2561
2562
2563
2564
2565
2566
2567
2568
2569
            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,
        )

2570
2571
2572
2573
2574
2575
2576
2577
2578
2579
        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,
        )

2580
2581
2582
2583
2584
2585
2586
2587
2588
2589
2590
2591
2592
2593
2594
    @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()

2595
2596
    def _model_forward(
        self,
2597
2598
2599
2600
        input_ids: torch.Tensor | None = None,
        positions: torch.Tensor | None = None,
        intermediate_tensors: IntermediateTensors | None = None,
        inputs_embeds: torch.Tensor | None = None,
2601
2602
2603
2604
2605
        **model_kwargs: dict[str, Any],
    ) -> Any:
        """Helper method to call the model forward pass.

        This method can be overridden by subclasses for model execution.
2606
        Motivation: We can inspect only this method versus
2607
2608
2609
2610
2611
2612
2613
2614
2615
2616
2617
2618
2619
2620
2621
2622
2623
2624
2625
2626
        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,
        )

2627
2628
2629
2630
    @torch.inference_mode()
    def execute_model(
        self,
        scheduler_output: "SchedulerOutput",
2631
        intermediate_tensors: IntermediateTensors | None = None,
2632
2633
2634
2635
2636
2637
    ) -> 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."
            )
2638
2639
2640
2641
2642
2643
2644
2645
2646
2647
2648
2649
2650
2651
2652

        # 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)

2653
        num_scheduled_tokens = scheduler_output.total_num_scheduled_tokens
2654
        with record_function_or_nullcontext("gpu_model_runner: preprocess"):
2655
2656
2657
2658
            with self.synchronize_input_prep():
                # Update persistent batch states.
                self._update_states(scheduler_output)

2659
2660
2661
2662
2663
2664
2665
2666
                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)

2667
                if not num_scheduled_tokens:
2668
2669
2670
2671
2672
2673
2674
2675
2676
2677
2678
2679
                    if (
                        self.parallel_config.distributed_executor_backend
                        == "external_launcher"
                        and self.parallel_config.data_parallel_size > 1
                    ):
                        # this is a corner case when both external launcher
                        # and DP are enabled, num_scheduled_tokens could be
                        # 0, and has_unfinished_requests in the outer loop
                        # returns True. before returning early here we call
                        # dummy run to ensure coordinate_batch_across_dp
                        # is called into to avoid out of sync issues.
                        self._dummy_run(1)
2680
2681
2682
2683
                    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(
2684
2685
                        scheduler_output, self.vllm_config
                    )
2686
                if self.cache_config.kv_sharing_fast_prefill:
2687
                    assert not self.num_prompt_logprobs, (
2688
2689
                        "--kv-sharing-fast-prefill produces incorrect "
                        "logprobs for prompt tokens, tokens, please disable "
2690
2691
                        "it when the requests need prompt logprobs"
                    )
2692

2693
2694
2695
2696
2697
2698
                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())

2699
2700
2701
2702
                (
                    logits_indices,
                    spec_decode_metadata,
                    ubatch_slices,
2703
                    num_tokens_across_dp,
2704
2705
2706
2707
2708
2709
2710
2711
2712
2713
2714
2715
2716
2717
2718
2719
2720
2721
2722
2723
2724
2725
2726
2727
2728
2729
2730
2731
2732
2733
2734
                ) = self._prepare_inputs(
                    scheduler_output, num_scheduled_tokens_np, max_num_scheduled_tokens
                )

                cascade_attn_prefix_lens = None
                # Disable cascade attention when using microbatching (DBO)
                if self.cascade_attn_enabled and ubatch_slices is None:
                    # Pre-compute cascade attention prefix lengths
                    # NOTE: Must be AFTER _prepare_inputs uses self.input_batch state
                    cascade_attn_prefix_lens = self._compute_cascade_attn_prefix_lens(
                        num_scheduled_tokens_np,
                        scheduler_output.num_common_prefix_blocks,
                    )

                # TODO(lucas): move cudagraph dispatching here:
                #   https://github.com/vllm-project/vllm/issues/23789

                total_num_scheduled_tokens = scheduler_output.total_num_scheduled_tokens
                use_spec_decode = len(scheduler_output.scheduled_spec_decode_tokens) > 0
                attn_metadata, spec_decode_common_attn_metadata = (
                    self._build_attention_metadata(
                        total_num_scheduled_tokens=total_num_scheduled_tokens,
                        max_num_scheduled_tokens=max_num_scheduled_tokens,
                        num_reqs=num_reqs,
                        ubatch_slices=ubatch_slices,
                        logits_indices=logits_indices,
                        use_spec_decode=use_spec_decode,
                        scheduled_encoder_inputs=scheduler_output.scheduled_encoder_inputs,
                        cascade_attn_prefix_lens=cascade_attn_prefix_lens,
                    )
                )
2735

2736
2737
2738
2739
2740
2741
2742
2743
2744
2745
2746
                dp_rank = self.parallel_config.data_parallel_rank
                if ubatch_slices:
                    assert num_tokens_across_dp is not None
                    num_input_tokens = int(num_tokens_across_dp[dp_rank].item())
                    self.pad_out_ubatch_slice(ubatch_slices, num_input_tokens)
                elif num_tokens_across_dp is not None:
                    num_input_tokens = int(num_tokens_across_dp[dp_rank].item())
                else:
                    num_input_tokens = self._get_num_input_tokens(
                        scheduler_output.total_num_scheduled_tokens
                    )
2747

2748
2749
2750
2751
2752
2753
2754
2755
2756
2757
                (
                    input_ids,
                    inputs_embeds,
                    positions,
                    intermediate_tensors,
                    model_kwargs,
                    ec_connector_output,
                ) = self._preprocess(
                    scheduler_output, num_input_tokens, intermediate_tensors
                )
2758

2759
2760
2761
            uniform_decode = (
                max_num_scheduled_tokens == self.uniform_decode_query_len
            ) and (num_scheduled_tokens == num_reqs * max_num_scheduled_tokens)
2762
            batch_desc = BatchDescriptor(
2763
2764
2765
                num_tokens=num_input_tokens,
                uniform_decode=uniform_decode,
                has_lora=len(self.input_batch.lora_id_to_lora_request) > 0,
2766
2767
            )
            cudagraph_runtime_mode, batch_descriptor = (
2768
                self.cudagraph_dispatcher.dispatch(
2769
                    batch_desc,
2770
2771
                    use_cascade_attn=cascade_attn_prefix_lens is not None,
                )
2772
            )
2773

2774
        # Set cudagraph mode to none if calc_kv_scales is true.
2775
2776
2777
2778
2779
2780
        # KV scales calculation involves dynamic operations that are incompatible
        # with CUDA graph capture.
        if self.calculate_kv_scales:
            cudagraph_runtime_mode = CUDAGraphMode.NONE
            # Mark KV scales as calculated after the first forward pass
            self.calculate_kv_scales = False
2781

2782
2783
        # Run the model.
        # Use persistent buffers for CUDA graphs.
2784
2785
        with (
            set_forward_context(
2786
2787
2788
2789
2790
2791
                attn_metadata,
                self.vllm_config,
                num_tokens=num_input_tokens,
                num_tokens_across_dp=num_tokens_across_dp,
                cudagraph_runtime_mode=cudagraph_runtime_mode,
                batch_descriptor=batch_descriptor,
2792
                ubatch_slices=ubatch_slices,
2793
            ),
2794
            record_function_or_nullcontext("gpu_model_runner: forward"),
2795
2796
            self.maybe_get_kv_connector_output(scheduler_output) as kv_connector_output,
        ):
2797
            model_output = self._model_forward(
2798
2799
2800
2801
2802
2803
2804
                input_ids=input_ids,
                positions=positions,
                intermediate_tensors=intermediate_tensors,
                inputs_embeds=inputs_embeds,
                **model_kwargs,
            )

2805
        with record_function_or_nullcontext("gpu_model_runner: postprocess"):
2806
            if self.use_aux_hidden_state_outputs:
2807
                # True when EAGLE 3 is used.
2808
2809
                hidden_states, aux_hidden_states = model_output
            else:
2810
                # Common case.
2811
2812
2813
                hidden_states = model_output
                aux_hidden_states = None

2814
2815
2816
2817
2818
            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)
2819
                    hidden_states.kv_connector_output = kv_connector_output
2820
                    self.kv_connector_output = kv_connector_output
2821
                    return hidden_states
2822

2823
                if self.is_pooling_model:
2824
                    # Return the pooling output.
2825
2826
2827
                    output = self._pool(
                        hidden_states, num_scheduled_tokens, num_scheduled_tokens_np
                    )
2828
2829
                    output.kv_connector_output = kv_connector_output
                    return output
2830
2831

                sample_hidden_states = hidden_states[logits_indices]
2832
                logits = self.model.compute_logits(sample_hidden_states)
2833
2834
2835
2836
            else:
                # Rare case.
                assert not self.is_pooling_model

2837
                sample_hidden_states = hidden_states[logits_indices]
2838
                if not get_pp_group().is_last_rank:
2839
                    all_gather_tensors = {
2840
2841
2842
                        "residual": not is_residual_scattered_for_sp(
                            self.vllm_config, num_input_tokens
                        )
2843
                    }
2844
                    get_pp_group().send_tensor_dict(
2845
2846
                        hidden_states.tensors,
                        all_gather_group=get_tp_group(),
2847
2848
                        all_gather_tensors=all_gather_tensors,
                    )
2849
2850
                    logits = None
                else:
2851
                    logits = self.model.compute_logits(sample_hidden_states)
2852

2853
                model_output_broadcast_data: dict[str, Any] = {}
2854
2855
2856
                if logits is not None:
                    model_output_broadcast_data["logits"] = logits.contiguous()

2857
                broadcasted = get_pp_group().broadcast_tensor_dict(
2858
2859
                    model_output_broadcast_data, src=len(get_pp_group().ranks) - 1
                )
2860
2861
                assert broadcasted is not None
                logits = broadcasted["logits"]
2862

2863
2864
2865
2866
2867
2868
2869
2870
        self.execute_model_state = ExecuteModelState(
            scheduler_output,
            logits,
            spec_decode_metadata,
            spec_decode_common_attn_metadata,
            hidden_states,
            sample_hidden_states,
            aux_hidden_states,
2871
            ec_connector_output,
2872
        )
2873
        self.kv_connector_output = kv_connector_output
2874
2875
2876
2877
2878
2879
        return None

    @torch.inference_mode
    def sample_tokens(
        self, grammar_output: "GrammarOutput | None"
    ) -> ModelRunnerOutput | AsyncModelRunnerOutput | IntermediateTensors:
2880
2881
2882
        kv_connector_output = self.kv_connector_output
        self.kv_connector_output = None

2883
2884
        if self.execute_model_state is None:
            # Nothing to do (PP non-final rank case), output isn't used.
2885
            if not kv_connector_output:
2886
                return None  # type: ignore[return-value]
2887
2888
2889
2890
2891
2892
2893
2894
2895

            # 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
2896
2897
2898
2899
2900
2901
2902
2903
2904
2905

        # Unpack ephemeral state.
        (
            scheduler_output,
            logits,
            spec_decode_metadata,
            spec_decode_common_attn_metadata,
            hidden_states,
            sample_hidden_states,
            aux_hidden_states,
2906
            ec_connector_output,
2907
2908
2909
2910
2911
2912
2913
2914
2915
        ) = 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
            )
2916

2917
        with record_function_or_nullcontext("gpu_model_runner: sample"):
2918
2919
            sampler_output = self._sample(logits, spec_decode_metadata)

2920
2921
        self.input_batch.prev_sampled_token_ids = None

2922
        def propose_draft_token_ids(sampled_token_ids):
2923
            assert spec_decode_common_attn_metadata is not None
2924
            with record_function_or_nullcontext("gpu_model_runner: draft"):
2925
2926
2927
2928
2929
2930
2931
2932
2933
2934
2935
                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,
                )

2936
        spec_config = self.speculative_config
2937
        use_padded_batch_for_eagle = (
2938
2939
2940
            spec_config is not None
            and spec_config.use_eagle()
            and not spec_config.disable_padded_drafter_batch
2941
        )
2942
2943
2944
        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
2945
        if (
2946
2947
2948
            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
2949
        ):
2950
            effective_drafter_max_model_len = (
2951
                spec_config.draft_model_config.max_model_len
2952
            )
2953
        input_fits_in_drafter = spec_decode_common_attn_metadata and (
2954
            spec_decode_common_attn_metadata.max_seq_len + self.num_spec_tokens
2955
2956
            <= effective_drafter_max_model_len
        )
2957
        if use_padded_batch_for_eagle:
2958
2959
            assert self.speculative_config is not None
            assert isinstance(self.drafter, EagleProposer)
2960
2961
2962
2963
2964
2965
            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:
2966
                assert spec_decode_common_attn_metadata is not None
2967
2968
2969
2970
2971
2972
2973
2974
2975
2976
2977
2978
2979
                next_token_ids, valid_sampled_tokens_count = (
                    self.drafter.prepare_next_token_ids_padded(
                        spec_decode_common_attn_metadata,
                        sampled_token_ids,
                        self.requests,
                        self.input_batch,
                        self.discard_request_indices.gpu,
                        self.num_discarded_requests,
                    )
                )
                self._copy_valid_sampled_token_count(
                    next_token_ids, valid_sampled_tokens_count
                )
2980

2981
        with record_function_or_nullcontext("gpu_model_runner: bookkeep"):
2982
2983
2984
2985
2986
2987
2988
2989
            (
                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,
2990
2991
2992
2993
2994
            ) = self._bookkeeping_sync(
                scheduler_output,
                sampler_output,
                logits,
                hidden_states,
2995
                scheduler_output.total_num_scheduled_tokens,
2996
                spec_decode_metadata,
2997
            )
2998

2999
3000
3001
3002
3003
        if (
            self.speculative_config
            and not use_padded_batch_for_eagle
            and input_fits_in_drafter
        ):
3004
3005
3006
            # 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)
3007

3008
        with record_function_or_nullcontext("gpu_model_runner: eplb"):
3009
            self.eplb_step()
3010
3011
3012
3013
3014
3015
3016
3017
3018
        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,
3019
3020
3021
                ec_connector_output=ec_connector_output
                if self.supports_mm_inputs
                else None,
3022
3023
                num_nans_in_logits=num_nans_in_logits,
            )
3024

3025
3026
        if not self.use_async_scheduling:
            return output
3027
3028
3029
3030
3031
3032
3033
3034
3035
        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,
3036
                vocab_size=self.input_batch.vocab_size,
3037
3038
3039
3040
3041
            )
        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
3042
            # any requests with sampling params that require output ids.
3043
3044
3045
3046
            self.input_batch.set_async_sampled_token_ids(
                async_output.sampled_token_ids_cpu,
                async_output.async_copy_ready_event,
            )
3047
3048
3049

        return async_output

3050
    def take_draft_token_ids(self) -> DraftTokenIds | None:
3051
3052
3053
3054
3055
3056
3057
3058
3059
3060
        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)

3061
3062
3063
3064
3065
3066
3067
3068
3069
3070
3071
3072
3073
3074
3075
3076
3077
3078
3079
3080
3081
3082
3083
3084
3085
3086
3087
3088
3089
3090
3091
    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()

3092
3093
3094
    def propose_draft_token_ids(
        self,
        scheduler_output: "SchedulerOutput",
3095
        sampled_token_ids: torch.Tensor | list[list[int]],
3096
3097
3098
        sampling_metadata: SamplingMetadata,
        hidden_states: torch.Tensor,
        sample_hidden_states: torch.Tensor,
3099
3100
        aux_hidden_states: list[torch.Tensor] | None,
        spec_decode_metadata: SpecDecodeMetadata | None,
3101
        common_attn_metadata: CommonAttentionMetadata,
3102
    ) -> list[list[int]] | torch.Tensor:
3103
        num_scheduled_tokens = scheduler_output.total_num_scheduled_tokens
3104
3105
3106
        spec_config = self.speculative_config
        assert spec_config is not None
        if spec_config.method == "ngram":
3107
            assert isinstance(sampled_token_ids, list)
3108
            assert isinstance(self.drafter, NgramProposer)
3109
            draft_token_ids = self.drafter.propose(
3110
3111
                sampled_token_ids,
                self.input_batch.req_ids,
3112
3113
                self.input_batch.num_tokens_no_spec,
                self.input_batch.token_ids_cpu,
3114
3115
                self.input_batch.spec_decode_unsupported_reqs,
            )
3116
        elif spec_config.method == "suffix":
3117
3118
3119
            assert isinstance(sampled_token_ids, list)
            assert isinstance(self.drafter, SuffixDecodingProposer)
            draft_token_ids = self.drafter.propose(self.input_batch, sampled_token_ids)
3120
        elif spec_config.method == "medusa":
3121
            assert isinstance(sampled_token_ids, list)
3122
            assert isinstance(self.drafter, MedusaProposer)
3123

3124
3125
            if sample_hidden_states.shape[0] == len(sampled_token_ids):
                # The input to the target model does not include draft tokens.
3126
3127
3128
3129
                hidden_states = sample_hidden_states
            else:
                indices = []
                offset = 0
3130
3131
3132
                assert spec_decode_metadata is not None, (
                    "No spec decode metadata for medusa"
                )
3133
                for num_draft, tokens in zip(
3134
3135
                    spec_decode_metadata.num_draft_tokens, sampled_token_ids
                ):
3136
                    indices.append(offset + len(tokens) - 1)
3137
                    offset += num_draft + 1
3138
                indices = torch.tensor(indices, device=self.device)
3139
3140
                hidden_states = sample_hidden_states[indices]

3141
            draft_token_ids = self.drafter.propose(
3142
3143
3144
                target_hidden_states=hidden_states,
                sampling_metadata=sampling_metadata,
            )
3145
        elif spec_config.use_eagle():
3146
            assert isinstance(self.drafter, EagleProposer)
3147

3148
            if spec_config.disable_padded_drafter_batch:
3149
3150
3151
                # 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.
3152
3153
                assert isinstance(sampled_token_ids, list), (
                    "sampled_token_ids should be a python list when"
3154
                    "padded-batch is disabled."
3155
                )
3156
                next_token_ids = self.drafter.prepare_next_token_ids_cpu(
3157
3158
3159
3160
3161
                    sampled_token_ids,
                    self.requests,
                    self.input_batch,
                    scheduler_output.num_scheduled_tokens,
                )
3162
3163
3164
3165
3166
            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.
3167
3168
                assert isinstance(sampled_token_ids, torch.Tensor), (
                    "sampled_token_ids should be a torch.Tensor when"
3169
                    "padded-batch is enabled."
3170
3171
                )
                next_token_ids, valid_sampled_tokens_count = (
3172
3173
3174
3175
3176
3177
                    self.drafter.prepare_next_token_ids_padded(
                        common_attn_metadata,
                        sampled_token_ids,
                        self.requests,
                        self.input_batch,
                        self.discard_request_indices.gpu,
3178
                        self.num_discarded_requests,
3179
                    )
3180
                )
3181
3182
3183
                self._copy_valid_sampled_token_count(
                    next_token_ids, valid_sampled_tokens_count
                )
Jiayi Yao's avatar
Jiayi Yao committed
3184

3185
            if spec_decode_metadata is None:
3186
                token_indices_to_sample = None
3187
                # input_ids can be None for multimodal models.
3188
                target_token_ids = self.input_ids.gpu[:num_scheduled_tokens]
3189
                target_positions = self._get_positions(num_scheduled_tokens)
3190
                if self.use_aux_hidden_state_outputs:
Wentao Ye's avatar
Wentao Ye committed
3191
                    assert aux_hidden_states is not None
3192
                    target_hidden_states = torch.cat(
3193
3194
                        [h[:num_scheduled_tokens] for h in aux_hidden_states], dim=-1
                    )
3195
3196
                else:
                    target_hidden_states = hidden_states[:num_scheduled_tokens]
3197
            else:
3198
                if spec_config.disable_padded_drafter_batch:
3199
                    token_indices_to_sample = None
3200
3201
3202
3203
3204
                    common_attn_metadata, token_indices = self.drafter.prepare_inputs(
                        common_attn_metadata,
                        sampled_token_ids,
                        spec_decode_metadata.num_draft_tokens,
                    )
3205
                else:
3206
                    common_attn_metadata, token_indices, token_indices_to_sample = (
3207
3208
3209
                        self.drafter.prepare_inputs_padded(
                            common_attn_metadata,
                            spec_decode_metadata,
3210
3211
3212
                            valid_sampled_tokens_count,
                        )
                    )
3213

3214
                target_token_ids = self.input_ids.gpu[token_indices]
3215
                target_positions = self._get_positions(token_indices)
3216
                if self.use_aux_hidden_state_outputs:
Wentao Ye's avatar
Wentao Ye committed
3217
                    assert aux_hidden_states is not None
3218
                    target_hidden_states = torch.cat(
3219
3220
                        [h[token_indices] for h in aux_hidden_states], dim=-1
                    )
3221
3222
                else:
                    target_hidden_states = hidden_states[token_indices]
3223

3224
            if self.supports_mm_inputs:
3225
3226
3227
3228
3229
3230
                mm_embed_inputs = self._gather_mm_embeddings(
                    scheduler_output,
                    shift_computed_tokens=1,
                )
            else:
                mm_embed_inputs = None
3231

3232
            draft_token_ids = self.drafter.propose(
3233
3234
3235
3236
                target_token_ids=target_token_ids,
                target_positions=target_positions,
                target_hidden_states=target_hidden_states,
                next_token_ids=next_token_ids,
3237
                last_token_indices=token_indices_to_sample,
3238
                sampling_metadata=sampling_metadata,
3239
                common_attn_metadata=common_attn_metadata,
3240
                mm_embed_inputs=mm_embed_inputs,
3241
            )
3242

3243
        return draft_token_ids
3244

3245
3246
3247
    def update_config(self, overrides: dict[str, Any]) -> None:
        allowed_config_names = {"load_config", "model_config"}
        for config_name, config_overrides in overrides.items():
3248
3249
            assert config_name in allowed_config_names, (
                f"Config `{config_name}` not supported. "
3250
                f"Allowed configs: {allowed_config_names}"
3251
            )
3252
3253
3254
3255
            config = getattr(self, config_name)
            new_config = update_config(config, config_overrides)
            setattr(self, config_name, new_config)

3256
3257
3258
3259
3260
    def load_model(self, eep_scale_up: bool = False) -> None:
        """
        Args:
            eep_scale_up: the model loading is for elastic EP scale up.
        """
3261
3262
3263
3264
3265
        logger.info_once(
            "Starting to load model %s...",
            self.model_config.model,
            scope="global",
        )
3266
3267
3268
3269
3270
        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)
        )
3271

3272
3273
3274
        if self.parallel_config.enable_eplb:
            self.eplb_state = EplbState(self.parallel_config, self.device)
            eplb_models = 0
3275
        with DeviceMemoryProfiler() as m:
3276
            time_before_load = time.perf_counter()
3277
            model_loader = get_model_loader(self.load_config)
3278
            self.model = model_loader.load_model(
3279
3280
                vllm_config=self.vllm_config, model_config=self.model_config
            )
3281
            if self.lora_config:
3282
3283
3284
                self.model = self.load_lora_model(
                    self.model, self.vllm_config, self.device
                )
3285
            if hasattr(self, "drafter"):
3286
                logger.info_once("Loading drafter model...")
3287
                self.drafter.load_model(self.model)
3288
3289
3290
3291
3292
                if (
                    hasattr(self.drafter, "model")
                    and is_mixture_of_experts(self.drafter.model)
                    and self.parallel_config.enable_eplb
                ):
3293
3294
3295
                    spec_config = self.vllm_config.speculative_config
                    assert spec_config is not None
                    assert spec_config.draft_model_config is not None
3296
3297
                    logger.info_once(
                        "EPLB is enabled for drafter model %s.",
3298
                        spec_config.draft_model_config.model,
3299
3300
3301
3302
3303
3304
3305
3306
3307
3308
3309
3310
3311
3312
3313
3314
                    )

                    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,
3315
                        spec_config.draft_model_config,
3316
3317
3318
3319
3320
3321
                        global_expert_load,
                        old_global_expert_indices,
                        rank_mapping,
                    )
                    eplb_models += 1

3322
            if self.use_aux_hidden_state_outputs:
3323
                if not supports_eagle3(self.get_model()):
3324
3325
                    raise RuntimeError(
                        "Model does not support EAGLE3 interface but "
3326
3327
                        "aux_hidden_state_outputs was requested"
                    )
3328
3329
3330
3331
3332
3333
3334
3335
3336
3337
3338
3339
3340

                # 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)
3341
            time_after_load = time.perf_counter()
3342
        self.model_memory_usage = m.consumed_memory
3343
        logger.info_once(
3344
            "Model loading took %.4f GiB memory and %.6f seconds",
3345
3346
            self.model_memory_usage / GiB_bytes,
            time_after_load - time_before_load,
3347
            scope="local",
3348
        )
3349
        prepare_communication_buffer_for_model(self.model)
3350
        mm_config = self.model_config.multimodal_config
3351
        self.is_multimodal_pruning_enabled = (
3352
            supports_multimodal_pruning(self.get_model())
3353
3354
            and mm_config is not None
            and mm_config.is_multimodal_pruning_enabled()
3355
        )
3356

3357
        if is_mixture_of_experts(self.model) and self.parallel_config.enable_eplb:
3358
3359
3360
3361
3362
3363
3364
3365
3366
3367
3368
            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(
3369
                self.model,
3370
                self.model_config,
3371
3372
3373
                global_expert_load,
                old_global_expert_indices,
                rank_mapping,
3374
            )
3375
3376
            if self.eplb_state.is_async:
                self.eplb_state.start_async_loop(rank_mapping=rank_mapping)
3377

3378
        if (
3379
3380
            self.vllm_config.compilation_config.mode
            == CompilationMode.STOCK_TORCH_COMPILE
3381
            and supports_dynamo()
3382
        ):
3383
            backend = self.vllm_config.compilation_config.init_backend(self.vllm_config)
3384
            compilation_counter.stock_torch_compile_count += 1
3385
            self.model.compile(fullgraph=True, backend=backend)
3386
            return
3387
        # for other compilation modes, cudagraph behavior is controlled by
3388
3389
3390
        # CudagraphWraper and CudagraphDispatcher of vllm.

        # wrap the model with full cudagraph wrapper if needed.
3391
3392
3393
        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:
3394
3395
3396
            self.model = CUDAGraphWrapper(
                self.model, self.vllm_config, runtime_mode=CUDAGraphMode.FULL
            )
3397
        elif self.parallel_config.enable_dbo:
3398
            if cudagraph_mode.has_full_cudagraphs():
3399
3400
3401
                self.model = UBatchWrapper(
                    self.model, self.vllm_config, CUDAGraphMode.FULL, self.device
                )
3402
            else:
3403
3404
3405
                self.model = UBatchWrapper(
                    self.model, self.vllm_config, CUDAGraphMode.NONE, self.device
                )
3406

3407
    def _get_eagle3_aux_layers_from_config(self) -> tuple[int, ...] | None:
3408
3409
3410
3411
3412
3413
3414
3415
3416
3417
3418
3419
3420
3421
3422
3423
3424
3425
3426
3427
3428
3429
3430
        """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

3431
    def reload_weights(self) -> None:
3432
        assert getattr(self, "model", None) is not None, (
3433
            "Cannot reload weights before model is loaded."
3434
        )
3435
3436
        model_loader = get_model_loader(self.load_config)
        logger.info("Reloading weights inplace...")
3437
        model_loader.load_weights(self.get_model(), model_config=self.model_config)
3438

3439
3440
3441
3442
3443
    def save_tensorized_model(
        self,
        tensorizer_config: "TensorizerConfig",
    ) -> None:
        TensorizerLoader.save_model(
3444
            self.get_model(),
3445
            tensorizer_config=tensorizer_config,
3446
            model_config=self.model_config,
3447
3448
        )

3449
3450
3451
    def _get_prompt_logprobs_dict(
        self,
        hidden_states: torch.Tensor,
3452
        num_scheduled_tokens: dict[str, int],
3453
    ) -> dict[str, LogprobsTensors | None]:
3454
        num_prompt_logprobs_dict = self.num_prompt_logprobs
3455
3456
3457
        if not num_prompt_logprobs_dict:
            return {}

3458
        in_progress_dict = self.input_batch.in_progress_prompt_logprobs_cpu
3459
        prompt_logprobs_dict: dict[str, LogprobsTensors | None] = {}
3460
3461
3462
3463
3464

        # 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():
3465
3466
3467
3468
            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
3469
3470
3471

            # Get metadata for this request.
            request = self.requests[req_id]
3472
3473
3474
3475
            if request.prompt_token_ids is None:
                # Prompt logprobs is incompatible with prompt embeddings
                continue

3476
3477
            num_prompt_tokens = len(request.prompt_token_ids)
            prompt_token_ids = torch.tensor(request.prompt_token_ids).to(
3478
3479
                self.device, non_blocking=True
            )
3480

3481
3482
3483
3484
3485
3486
            # 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(
3487
3488
                    num_prompt_tokens - 1, num_prompt_logprobs + 1
                )
3489
3490
                in_progress_dict[req_id] = logprobs_tensors

3491
            # Determine number of logits to retrieve.
3492
3493
            start_idx = request.num_computed_tokens
            start_tok = start_idx + 1
3494
            num_remaining_tokens = num_prompt_tokens - start_tok
3495
            if num_tokens <= num_remaining_tokens:
3496
                # This is a chunk, more tokens remain.
3497
3498
3499
                # 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.
3500
3501
3502
3503
3504
                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)
3505
3506
3507
3508
3509
3510
3511
                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
3512
3513
3514
3515
3516

            # 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]
3517
            offset = self.query_start_loc.np[req_idx].item()
3518
            prompt_hidden_states = hidden_states[offset : offset + num_logits]
3519
            logits = self.model.compute_logits(prompt_hidden_states)
3520
3521
3522
3523

            # 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.
3524
            tgt_token_ids = prompt_token_ids[start_tok : start_tok + num_logits]
3525
3526

            # Compute prompt logprobs.
3527
3528
            logprobs = self.sampler.compute_logprobs(logits)
            token_ids, logprobs, ranks = self.sampler.gather_logprobs(
3529
3530
                logprobs, num_prompt_logprobs, tgt_token_ids
            )
3531
3532

            # Transfer GPU->CPU async.
3533
3534
            chunk_slice = slice(start_idx, start_idx + num_logits)
            logprobs_tensors.logprob_token_ids[chunk_slice].copy_(
3535
3536
3537
                token_ids, non_blocking=True
            )
            logprobs_tensors.logprobs[chunk_slice].copy_(logprobs, non_blocking=True)
3538
            logprobs_tensors.selected_token_ranks[chunk_slice].copy_(
3539
3540
                ranks, non_blocking=True
            )
3541
3542
3543
3544
3545

        # 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]
3546
            del in_progress_dict[req_id]
3547
3548

        # Must synchronize the non-blocking GPU->CPU transfers.
3549
        if prompt_logprobs_dict:
3550
            self._sync_device()
3551
3552
3553

        return prompt_logprobs_dict

3554
3555
    def _get_nans_in_logits(
        self,
3556
        logits: torch.Tensor | None,
3557
3558
3559
3560
3561
3562
3563
3564
3565
3566
3567
    ) -> 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])
3568
3569
3570
                    if num_nans_for_index is not None and req_index < logits.shape[0]
                    else 0
                )
3571
3572
3573
3574
            return num_nans_in_logits
        except IndexError:
            return {}

3575
3576
3577
3578
3579
3580
    @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
3581
         - during DP rank dummy run
3582
3583
3584
3585
3586
3587
3588
3589
3590
3591
3592
        """
        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(
3593
                    self.input_ids.gpu,
3594
3595
                    low=0,
                    high=self.model_config.get_vocab_size(),
3596
3597
                    dtype=input_ids.dtype,
                )
3598

3599
            logger.debug_once("Randomizing dummy data for DP Rank")
3600
            input_ids.copy_(rand_input_ids()[: input_ids.size(0)], non_blocking=True)
3601
3602
3603
            yield
            input_ids.fill_(0)

3604
3605
3606
3607
3608
3609
    def _get_mm_dummy_batch(
        self,
        modality: str,
        max_items_per_batch: int,
    ) -> BatchedTensorInputs:
        """Dummy data for profiling and precompiling multimodal models."""
3610
3611
        assert self.mm_budget is not None

3612
3613
        dummy_decoder_data = self.mm_registry.get_decoder_dummy_data(
            model_config=self.model_config,
3614
            seq_len=self.max_model_len,
3615
            mm_counts={modality: 1},
3616
            cache=self.mm_budget.cache,
3617
3618
3619
3620
        )
        dummy_mm_data = dummy_decoder_data.multi_modal_data

        # Result in the maximum GPU consumption of the model
3621
3622
        dummy_mm_item = dummy_mm_data[modality][0]
        dummy_mm_items = [dummy_mm_item] * max_items_per_batch
3623

3624
        model = cast(SupportsMultiModal, self.model)
3625
3626
3627
3628
3629
3630
3631
        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,
3632
                multimodal_cpu_fields=model.multimodal_cpu_fields,
3633
3634
            )
        )
3635

3636
3637
3638
3639
    @torch.inference_mode()
    def _dummy_run(
        self,
        num_tokens: int,
3640
        cudagraph_runtime_mode: CUDAGraphMode | None = None,
3641
3642
        force_attention: bool = False,
        uniform_decode: bool = False,
3643
        allow_microbatching: bool = True,
3644
3645
        skip_eplb: bool = False,
        is_profile: bool = False,
3646
        create_mixed_batch: bool = False,
3647
        remove_lora: bool = True,
3648
        activate_lora: bool = False,
3649
    ) -> tuple[torch.Tensor, torch.Tensor]:
3650
3651
3652
3653
3654
3655
3656
        """
        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.
3657
                - if not set will determine the cudagraph mode based on using
3658
                    the self.cudagraph_dispatcher.
3659
3660
3661
3662
                - CUDAGraphMode.NONE: No cudagraph, for warm up and profile run
                - CUDAGraphMode.PIECEWISE: Piecewise cudagraph.
                - CUDAGraphMode.FULL: Full cudagraph, attention metadata is
                    needed.
3663
            force_attention: If True, always create attention metadata. Used to
3664
3665
3666
3667
                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.
3668
3669
            create_mixed_batch: If True, create a mixed batch with both decode
                (1 token) and prefill (multiple tokens) requests.
3670
            remove_lora: If False, dummy LoRAs are not destroyed after the run
3671
            activate_lora: If False, dummy_run is performed without LoRAs.
3672
        """
3673
3674
3675
3676
        assert (
            cudagraph_runtime_mode is None
            or cudagraph_runtime_mode.valid_runtime_modes()
        )
3677

3678
        # If cudagraph_mode.decode_mode() == FULL and
3679
        # cudagraph_mode.separate_routine(). This means that we are using
3680
3681
3682
3683
3684
3685
3686
3687
3688
3689
3690
        # 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.
3691
        max_query_len = self.uniform_decode_query_len if uniform_decode else num_tokens
3692

3693
3694
3695
3696
3697
        # 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
3698
3699
3700
3701
        if create_mixed_batch:
            assert not uniform_decode
            # Create mixed batch:
            # first half decode tokens, second half one prefill
3702
            num_decode_tokens = min(max_num_reqs - 1, num_tokens // 2)
3703
3704
3705
3706
            num_prefill_tokens = num_tokens - num_decode_tokens
            num_reqs = num_decode_tokens + 1

            # Create decode requests (1 token each) followed by prefill request
3707
            num_scheduled_tokens_list = [1] * num_decode_tokens + [num_prefill_tokens]
3708
3709
3710
            # Note: Overriding max_query_len to be the prefill tokens
            max_query_len = num_prefill_tokens
        elif uniform_decode:
3711
            assert not create_mixed_batch
3712
            num_reqs = min(max_num_reqs, cdiv(num_tokens, max_query_len))
3713
3714
            num_scheduled_tokens_list = [max_query_len] * num_reqs
            if num_tokens % max_query_len != 0:
3715
                num_scheduled_tokens_list[-1] = num_tokens % max_query_len
3716
3717
3718
3719
3720
3721
        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

3722
3723
        assert sum(num_scheduled_tokens_list) == num_tokens
        assert len(num_scheduled_tokens_list) == num_reqs
3724
        num_scheduled_tokens = np.array(num_scheduled_tokens_list, dtype=np.int32)
3725
        total_num_scheduled_tokens = int(num_scheduled_tokens.sum())
3726
        num_sampled_tokens = np.ones(num_reqs, dtype=np.int32)
3727

3728
3729
3730
        # Disable DP padding when running eager
        allow_dp_padding = self.compilation_config.cudagraph_mode != CUDAGraphMode.NONE

3731
3732
        # We currently only microbatch if the number of tokens is
        # over a certain threshold.
3733
        ubatch_slices, num_tokens_across_dp = coordinate_batch_across_dp(
3734
3735
3736
3737
3738
3739
3740
            num_tokens_unpadded=total_num_scheduled_tokens,
            parallel_config=self.vllm_config.parallel_config,
            allow_microbatching=allow_microbatching,
            allow_dp_padding=allow_dp_padding,
            num_tokens_padded=total_num_scheduled_tokens,
            uniform_decode=uniform_decode,
            num_scheduled_tokens_per_request=num_scheduled_tokens,
3741
3742
3743
        )
        num_tokens_after_padding = num_tokens
        if num_tokens_across_dp is not None:
3744
3745
            dp_rank = self.parallel_config.data_parallel_rank
            num_tokens_after_padding = int(num_tokens_across_dp[dp_rank])
3746

3747
3748
3749
3750
3751
3752
3753
3754
3755
3756
3757
3758
3759
3760
3761
3762
3763
3764
3765
3766
3767
3768
3769
3770
3771
        # filter out the valid batch descriptor
        _cg_mode, batch_descriptor = (
            self.cudagraph_dispatcher.dispatch(
                BatchDescriptor(
                    num_tokens=num_tokens_after_padding,
                    uniform_decode=uniform_decode,
                    has_lora=activate_lora and self.lora_config is not None,
                )
            )
            if not is_profile
            else (CUDAGraphMode.NONE, None)
        )
        if cudagraph_runtime_mode is not None:
            # we allow forcing NONE when the dispatcher disagrees to support
            # warm ups for cudagraph capture
            assert (
                cudagraph_runtime_mode == CUDAGraphMode.NONE
                or cudagraph_runtime_mode == _cg_mode
            ), (
                f"Cudagraph runtime mode mismatch at dummy_run. "
                f"Expected {_cg_mode}, but got {cudagraph_runtime_mode}."
            )
        else:
            cudagraph_runtime_mode = _cg_mode

3772
        attn_metadata: PerLayerAttnMetadata | None = None
3773
3774
3775

        # If force_attention is True, we always capture attention. Otherwise,
        # it only happens for cudagraph_runtime_mode=FULL.
3776
        if force_attention or cudagraph_runtime_mode == CUDAGraphMode.FULL:
3777
3778
3779
3780
3781
3782
            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:
3783
                seq_lens = max_query_len  # type: ignore[assignment]
3784
            self.seq_lens.np[:num_reqs] = seq_lens
3785
3786
            self.seq_lens.np[num_reqs:] = 0
            self.seq_lens.copy_to_gpu()
3787

3788
3789
            cum_num_tokens, _ = self._get_cumsum_and_arange(num_scheduled_tokens)
            self.query_start_loc.np[1 : num_reqs + 1] = cum_num_tokens
3790
3791
            self.query_start_loc.copy_to_gpu()

3792
3793
3794
3795
3796
3797
3798
            attn_metadata, _ = self._build_attention_metadata(
                total_num_scheduled_tokens=num_tokens,
                max_num_scheduled_tokens=max_query_len,
                num_reqs=num_reqs,
                ubatch_slices=ubatch_slices,
                for_cudagraph_capture=True,
            )
3799

3800
        with self.maybe_dummy_run_with_lora(
3801
3802
3803
3804
3805
            self.lora_config,
            num_scheduled_tokens,
            num_sampled_tokens,
            activate_lora,
            remove_lora,
3806
        ):
3807
3808
3809
            # Make sure padding doesn't exceed max_num_tokens
            assert num_tokens_after_padding <= self.max_num_tokens
            model_kwargs = self._init_model_kwargs(num_tokens_after_padding)
3810
            if self.supports_mm_inputs and not self.model_config.is_encoder_decoder:
3811
                input_ids = None
3812
                inputs_embeds = self.inputs_embeds.gpu[:num_tokens_after_padding]
3813
                model_kwargs = {
3814
                    **model_kwargs,
3815
3816
                    **self._dummy_mm_kwargs(num_reqs),
                }
3817
3818
            elif self.enable_prompt_embeds:
                input_ids = None
3819
3820
                inputs_embeds = self.inputs_embeds.gpu[:num_tokens_after_padding]
                model_kwargs = self._init_model_kwargs(num_tokens_after_padding)
3821
            else:
3822
                input_ids = self.input_ids.gpu[:num_tokens_after_padding]
3823
                inputs_embeds = None
3824

3825
            if self.uses_mrope:
3826
                positions = self.mrope_positions.gpu[:, :num_tokens_after_padding]
3827
            else:
3828
                positions = self.positions.gpu[:num_tokens_after_padding]
3829
3830
3831
3832
3833
3834
3835
3836
3837

            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,
3838
3839
3840
                            device=self.device,
                        )
                    )
3841
3842

                intermediate_tensors = self.sync_and_slice_intermediate_tensors(
3843
                    num_tokens_after_padding, None, False
3844
                )
3845

3846
            if ubatch_slices is not None:
3847
3848
3849
3850
3851
3852
3853
                # Adjust values to reflect a single ubatch.
                # TODO(sage,lucas): this is cruft that should be addressed in
                #  the padding refactor.
                num_tokens_after_padding = ubatch_slices[0].num_tokens
                if num_tokens_across_dp is not None:
                    num_tokens_across_dp[:] = num_tokens_after_padding

3854
3855
3856
            with (
                self.maybe_randomize_inputs(input_ids),
                set_forward_context(
3857
3858
                    attn_metadata,
                    self.vllm_config,
3859
                    num_tokens=num_tokens_after_padding,
3860
3861
                    num_tokens_across_dp=num_tokens_across_dp,
                    cudagraph_runtime_mode=cudagraph_runtime_mode,
3862
                    batch_descriptor=batch_descriptor,
3863
3864
3865
                    ubatch_slices=ubatch_slices,
                ),
            ):
3866
                outputs = self.model(
3867
3868
3869
3870
                    input_ids=input_ids,
                    positions=positions,
                    intermediate_tensors=intermediate_tensors,
                    inputs_embeds=inputs_embeds,
3871
                    **model_kwargs,
3872
                )
3873

3874
3875
3876
3877
            if self.use_aux_hidden_state_outputs:
                hidden_states, _ = outputs
            else:
                hidden_states = outputs
3878

3879
            if self.speculative_config and self.speculative_config.use_eagle():
3880
                assert isinstance(self.drafter, EagleProposer)
3881
3882
3883
3884
                use_cudagraphs = (
                    cudagraph_runtime_mode == CUDAGraphMode.PIECEWISE
                    and not self.speculative_config.enforce_eager
                )
3885
3886
3887
3888
3889
3890
3891
3892
3893
3894
3895
3896

                # 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,
                )
3897

3898
3899
3900
3901
3902
3903
3904
3905
3906
3907
        # 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)

3908
        logit_indices = np.cumsum(num_scheduled_tokens) - 1
3909
3910
3911
3912
        logit_indices_device = torch.from_numpy(logit_indices).to(
            self.device, non_blocking=True
        )
        return hidden_states, hidden_states[logit_indices_device]
3913
3914
3915
3916
3917
3918

    @torch.inference_mode()
    def _dummy_sampler_run(
        self,
        hidden_states: torch.Tensor,
    ) -> torch.Tensor:
3919
3920
3921
3922
        # 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)
3923

3924
        logits = self.model.compute_logits(hidden_states)
3925
3926
        num_reqs = logits.size(0)

3927
        dummy_tensors = lambda v: torch.full((num_reqs,), v, device=self.device)
3928
3929
3930
3931
3932
3933
3934
3935
3936
3937
3938
3939
3940
3941
3942

        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)],
3943
            spec_token_ids=[[] for _ in range(num_reqs)],
3944
3945
            allowed_token_ids_mask=None,
            bad_words_token_ids={},
3946
            logitsprocs=LogitsProcessors(),
3947
        )
3948
        try:
3949
3950
3951
            sampler_output = self.sampler(
                logits=logits, sampling_metadata=dummy_metadata
            )
3952
        except RuntimeError as e:
3953
            if "out of memory" in str(e):
3954
3955
3956
3957
                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 "
3958
3959
                    "initializing the engine."
                ) from e
3960
3961
            else:
                raise e
3962
        if self.speculative_config:
3963
3964
            draft_token_ids = [[0] for _ in range(num_reqs)]
            dummy_spec_decode_metadata = SpecDecodeMetadata.make_dummy(
3965
3966
                draft_token_ids, self.device
            )
3967
3968
3969
3970
3971
3972

            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
3973
3974
3975
3976
3977
            logits = torch.randn(
                num_tokens + num_reqs,
                logits.shape[-1],
                device=self.device,
                dtype=logits.dtype,
3978
            )
3979
3980
3981
            self.rejection_sampler(
                dummy_spec_decode_metadata,
                draft_probs,
3982
                logits,
3983
3984
                dummy_metadata,
            )
3985
        return sampler_output
3986

3987
    def _dummy_pooler_run_task(
3988
3989
        self,
        hidden_states: torch.Tensor,
3990
3991
        task: PoolingTask,
    ) -> PoolerOutput:
3992
3993
3994
3995
3996
3997
3998
3999
4000
4001
4002
        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

4003
        dummy_prompt_lens = torch.tensor(
4004
4005
            num_scheduled_tokens_list,
            device="cpu",
4006
        )
4007
4008
4009
        dummy_token_ids = torch.zeros(
            (num_reqs, req_num_tokens), dtype=torch.int32, device=self.device
        )
4010

4011
        model = cast(VllmModelForPooling, self.get_model())
4012
        dummy_pooling_params = PoolingParams(task=task)
4013
        dummy_pooling_params.verify(task=task, model_config=self.model_config)
4014
        to_update = model.pooler.get_pooling_updates(task)
4015
4016
        to_update.apply(dummy_pooling_params)

4017
        dummy_metadata = PoolingMetadata(
4018
4019
4020
4021
            prompt_lens=dummy_prompt_lens,
            prompt_token_ids=dummy_token_ids,
            pooling_params=[dummy_pooling_params] * num_reqs,
        )
4022

4023
4024
4025
        dummy_metadata.build_pooling_cursor(
            num_scheduled_tokens_list, device=hidden_states.device
        )
4026

4027
        try:
4028
4029
4030
            return model.pooler(
                hidden_states=hidden_states, pooling_metadata=dummy_metadata
            )
4031
        except RuntimeError as e:
4032
            if "out of memory" in str(e):
4033
                raise RuntimeError(
4034
4035
4036
                    "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 "
4037
4038
                    "initializing the engine."
                ) from e
4039
4040
            else:
                raise e
4041
4042
4043
4044
4045
4046
4047

    @torch.inference_mode()
    def _dummy_pooler_run(
        self,
        hidden_states: torch.Tensor,
    ) -> PoolerOutput:
        # Find the task that has the largest output for subsequent steps
4048
4049
4050
        supported_pooling_tasks = self.get_supported_pooling_tasks()

        if not supported_pooling_tasks:
4051
            if self.scheduler_config.enable_chunked_prefill:
4052
4053
4054
4055
4056
4057
4058
4059
4060
4061
4062
4063
4064
4065
4066
4067
                raise RuntimeError(
                    f"Model {self.model_config.model} does not support "
                    "any pooling tasks with chunked prefill enabled. "
                    "Please add --no-enable-chunked-prefill to your "
                    "config or CLI args. See "
                    "https://docs.vllm.ai/en/latest/models/pooling_models.html "
                    "to learn more."
                )
            else:
                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."
                )

4068
        output_size = dict[PoolingTask, float]()
4069
        for task in supported_pooling_tasks:
4070
4071
            # Run a full batch with each task to ensure none of them OOMs
            output = self._dummy_pooler_run_task(hidden_states, task)
4072
            output_size[task] = sum(o.nbytes for o in output)
4073
4074
4075
4076
            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)
4077

4078
    def profile_run(self) -> None:
4079
        # Profile with multimodal encoder & encoder cache.
4080
        if self.supports_mm_inputs:
4081
4082
            mm_config = self.model_config.multimodal_config
            if mm_config is not None and mm_config.skip_mm_profiling:
4083
                logger.info(
4084
                    "Skipping memory profiling for multimodal encoder and "
4085
4086
                    "encoder cache."
                )
4087
4088
4089
4090
4091
4092
4093
4094
            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.
4095
                    dummy_modality = mm_budget.get_modality_with_max_tokens()
4096
4097
4098
                    max_mm_items_per_batch = mm_budget.max_items_per_batch_by_modality[
                        dummy_modality
                    ]
4099
4100
4101
4102
4103
4104
4105
4106
4107

                    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,
                    )
4108

4109
4110
4111
4112
4113
                    # Create dummy batch of multimodal inputs.
                    batched_dummy_mm_inputs = self._get_mm_dummy_batch(
                        dummy_modality,
                        max_mm_items_per_batch,
                    )
4114

4115
                    # Run multimodal encoder.
4116
                    dummy_encoder_outputs = self.model.embed_multimodal(
4117
4118
                        **batched_dummy_mm_inputs
                    )
4119

4120
4121
4122
4123
                    sanity_check_mm_encoder_outputs(
                        dummy_encoder_outputs,
                        expected_num_items=max_mm_items_per_batch,
                    )
4124

4125
4126
4127
4128
4129
4130
4131
4132
4133
4134
                    # 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
                    # (encode_budget, hidden_size) and scatter encoder
                    # output into it.
                    encoder_output_shape = dummy_encoder_outputs[0].shape
                    if encoder_output_shape[0] < encoder_budget:
                        expanded_outputs = []
                        for output in dummy_encoder_outputs:
                            expanded = output.new_zeros(
4135
4136
                                (encoder_budget, encoder_output_shape[-1])
                            )
4137
4138
4139
4140
4141
4142
                            num_tokens = output.shape[0]
                            expanded[:num_tokens].copy_(output)
                            expanded_outputs.append(expanded)

                        dummy_encoder_outputs = expanded_outputs

4143
                    # Cache the dummy encoder outputs.
4144
                    self.encoder_cache["tmp"] = dict(enumerate(dummy_encoder_outputs))
4145

4146
        # Add `is_profile` here to pre-allocate communication buffers
4147
4148
4149
        hidden_states, last_hidden_states = self._dummy_run(
            self.max_num_tokens, is_profile=True
        )
4150
        if get_pp_group().is_last_rank:
4151
4152
4153
4154
            if self.is_pooling_model:
                output = self._dummy_pooler_run(hidden_states)
            else:
                output = self._dummy_sampler_run(last_hidden_states)
4155
        else:
4156
            output = None
4157
        self._sync_device()
4158
        del hidden_states, output
4159
        self.encoder_cache.clear()
4160
        gc.collect()
4161

4162
    def capture_model(self) -> int:
4163
        if self.compilation_config.cudagraph_mode == CUDAGraphMode.NONE:
4164
            logger.warning(
4165
                "Skipping CUDA graph capture. To turn on CUDA graph capture, "
4166
4167
                "ensure `cudagraph_mode` was not manually set to `NONE`"
            )
4168
            return 0
4169

4170
4171
        compilation_counter.num_gpu_runner_capture_triggers += 1

4172
4173
        start_time = time.perf_counter()

4174
4175
4176
4177
4178
4179
4180
4181
4182
4183
4184
4185
4186
4187
        @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()
4188
                    gc.collect()
4189

4190
4191
4192
        # 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.
4193
        set_cudagraph_capturing_enabled(True)
4194
        with freeze_gc(), graph_capture(device=self.device):
4195
            start_free_gpu_memory = torch.cuda.mem_get_info()[0]
4196
            cudagraph_mode = self.compilation_config.cudagraph_mode
4197
            assert cudagraph_mode is not None
4198
4199
4200
4201
4202
4203
4204
4205
4206

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

4207
4208
            if cudagraph_mode.mixed_mode() != CUDAGraphMode.NONE:
                cudagraph_runtime_mode = cudagraph_mode.mixed_mode()
4209
                # make sure we capture the largest batch size first
4210
4211
4212
                compilation_cases = list(
                    product(reversed(self.cudagraph_batch_sizes), lora_cases)
                )
4213
4214
4215
                self._capture_cudagraphs(
                    compilation_cases,
                    cudagraph_runtime_mode=cudagraph_runtime_mode,
4216
4217
                    uniform_decode=False,
                )
4218

4219
4220
            # Capture full cudagraph for uniform decode batches if we
            # don't already have full mixed prefill-decode cudagraphs.
4221
4222
4223
4224
4225
4226
4227
            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
                )
4228
                decode_cudagraph_batch_sizes = [
4229
4230
                    x
                    for x in self.cudagraph_batch_sizes
4231
                    if max_num_tokens >= x >= self.uniform_decode_query_len
4232
                ]
4233
4234
4235
                compilation_cases_decode = list(
                    product(reversed(decode_cudagraph_batch_sizes), lora_cases)
                )
4236
4237
4238
                self._capture_cudagraphs(
                    compilation_cases=compilation_cases_decode,
                    cudagraph_runtime_mode=CUDAGraphMode.FULL,
4239
4240
                    uniform_decode=True,
                )
4241

4242
4243
4244
            torch.cuda.synchronize()
            end_free_gpu_memory = torch.cuda.mem_get_info()[0]

4245
4246
4247
        # 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
4248
        # we may do lazy capturing in future that still allows capturing
4249
4250
        # after here.
        set_cudagraph_capturing_enabled(False)
4251
4252
4253
4254
4255

        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.
4256
        logger.info_once(
4257
4258
4259
            "Graph capturing finished in %.0f secs, took %.2f GiB",
            elapsed_time,
            cuda_graph_size / (1 << 30),
4260
            scope="local",
4261
        )
4262
        return cuda_graph_size
4263

4264
4265
    def _capture_cudagraphs(
        self,
4266
        compilation_cases: list[tuple[int, bool]],
4267
4268
4269
4270
4271
4272
4273
        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}"
4274
4275
4276
4277
4278
4279
4280
4281

        # 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",
4282
4283
4284
                    cudagraph_runtime_mode.name,
                ),
            )
4285

4286
        # We skip EPLB here since we don't want to record dummy metrics
4287
        for num_tokens, activate_lora in compilation_cases:
4288
            # We currently only capture ubatched graphs when its a FULL
4289
4290
4291
            # 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
4292
4293
4294
4295
            allow_microbatching = (
                self.parallel_config.enable_dbo
                and cudagraph_runtime_mode == CUDAGraphMode.FULL
                and uniform_decode
4296
4297
4298
4299
4300
                and check_ubatch_thresholds(
                    config=self.vllm_config.parallel_config,
                    num_tokens=num_tokens,
                    uniform_decode=uniform_decode,
                )
4301
            )
4302

4303
4304
4305
4306
4307
4308
            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.
4309
4310
4311
4312
4313
4314
4315
4316
4317
                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,
4318
                    activate_lora=activate_lora,
4319
4320
4321
4322
4323
4324
4325
4326
                )
            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,
4327
                activate_lora=activate_lora,
4328
            )
4329
        self.maybe_remove_all_loras(self.lora_config)
4330

4331
4332
4333
4334
    def initialize_attn_backend(self, kv_cache_config: KVCacheConfig) -> None:
        """
        Initialize the attention backends and attention metadata builders.
        """
4335
        assert len(self.attn_groups) == 0, "Attention backends are already initialized"
4336

4337
4338
4339
4340
4341
4342
        class AttentionGroupKey(NamedTuple):
            attn_backend: type[AttentionBackend]
            kv_cache_spec: KVCacheSpec

        def get_attn_backends_for_group(
            kv_cache_group_spec: KVCacheGroupSpec,
4343
        ) -> tuple[dict[AttentionGroupKey, list[str]], set[type[AttentionBackend]]]:
4344
            layer_type = cast(type[Any], AttentionLayerBase)
4345
            layers = get_layers_from_vllm_config(
4346
                self.vllm_config, layer_type, kv_cache_group_spec.layer_names
4347
            )
4348
4349
            attn_backends = {}
            attn_backend_layers = defaultdict(list)
4350
            # Dedupe based on full class name; this is a bit safer than
4351
4352
4353
4354
            # 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.
4355
            for layer_name in kv_cache_group_spec.layer_names:
4356
                attn_backend = layers[layer_name].get_attn_backend()
4357
4358
4359
4360

                if layer_name in self.kv_sharing_fast_prefill_eligible_layers:
                    attn_backend = create_fast_prefill_custom_backend(
                        "FastPrefill",
4361
                        attn_backend,  # type: ignore[arg-type]
4362
4363
                    )

4364
4365
4366
                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):
4367
                    layer_kv_cache_spec = layer_kv_cache_spec.kv_cache_specs[layer_name]
4368
                key = (full_cls_name, layer_kv_cache_spec)
4369
4370
4371
                attn_backends[key] = AttentionGroupKey(
                    attn_backend, layer_kv_cache_spec
                )
4372
                attn_backend_layers[key].append(layer_name)
4373
4374
4375
4376
            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()),
            )
4377
4378

        def create_attn_groups(
4379
            attn_backends_map: dict[AttentionGroupKey, list[str]],
4380
            kv_cache_group_id: int,
4381
4382
        ) -> list[AttentionGroup]:
            attn_groups: list[AttentionGroup] = []
4383
            for (attn_backend, kv_cache_spec), layer_names in attn_backends_map.items():
4384
                attn_group = AttentionGroup(
4385
                    attn_backend,
4386
                    layer_names,
4387
                    kv_cache_spec,
4388
                    kv_cache_group_id,
4389
4390
                )

4391
4392
4393
                attn_groups.append(attn_group)
            return attn_groups

4394
        attention_backend_maps = []
4395
        attention_backend_list = []
4396
        for kv_cache_group_spec in kv_cache_config.kv_cache_groups:
4397
            attn_backends = get_attn_backends_for_group(kv_cache_group_spec)
4398
            attention_backend_maps.append(attn_backends[0])
4399
            attention_backend_list.append(attn_backends[1])
4400
4401

        # Resolve cudagraph_mode before actually initialize metadata_builders
4402
4403
4404
        self._check_and_update_cudagraph_mode(
            attention_backend_list, kv_cache_config.kv_cache_groups
        )
4405

4406
4407
        for i, attn_backend_map in enumerate(attention_backend_maps):
            self.attn_groups.append(create_attn_groups(attn_backend_map, i))
4408

4409
4410
4411
4412
4413
4414
4415
4416
4417
4418
4419
4420
4421
4422
4423
4424
4425
4426
    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
4427
        # Calculate reorder batch threshold (if needed)
4428
4429
        # Note (tdoublep): do this *after* constructing builders,
        # because some of them change the threshold at init time.
4430
4431
        self.calculate_reorder_batch_threshold()

4432
    def _check_and_update_cudagraph_mode(
4433
4434
4435
        self,
        attention_backends: list[set[type[AttentionBackend]]],
        kv_cache_groups: list[KVCacheGroupSpec],
4436
    ) -> None:
4437
        """
4438
        Resolve the cudagraph_mode when there are multiple attention
4439
        groups with potential conflicting CUDA graph support.
4440
4441
4442
        Then initialize the cudagraph_dispatcher based on the resolved
        cudagraph_mode.
        """
4443
        min_cg_support = AttentionCGSupport.ALWAYS
4444
        min_cg_backend_name = None
4445

4446
4447
4448
4449
4450
4451
4452
4453
4454
4455
4456
4457
        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__
4458
4459
        # Flexible resolve the cudagraph mode
        cudagraph_mode = self.compilation_config.cudagraph_mode
4460
        assert cudagraph_mode is not None
4461
        # check cudagraph for mixed batch is supported
4462
4463
4464
4465
4466
4467
        if (
            cudagraph_mode.mixed_mode() == CUDAGraphMode.FULL
            and min_cg_support != AttentionCGSupport.ALWAYS
        ):
            msg = (
                f"CUDAGraphMode.{cudagraph_mode.name} is not supported "
4468
                f"with {min_cg_backend_name} backend (support: "
4469
4470
                f"{min_cg_support})"
            )
4471
4472
            if min_cg_support == AttentionCGSupport.NEVER:
                # if not supported any full cudagraphs, just raise it.
4473
4474
                msg += (
                    "; please try cudagraph_mode=PIECEWISE, and "
4475
                    "make sure compilation mode is VLLM_COMPILE"
4476
                )
4477
4478
4479
4480
4481
                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"
4482
                cudagraph_mode = self.compilation_config.cudagraph_mode = (
4483
                    CUDAGraphMode.FULL_AND_PIECEWISE
4484
                )
4485
4486
            else:
                msg += "; setting cudagraph_mode=FULL_DECODE_ONLY"
4487
                cudagraph_mode = self.compilation_config.cudagraph_mode = (
4488
                    CUDAGraphMode.FULL_DECODE_ONLY
4489
                )
4490
4491
            logger.warning(msg)

4492
        # check that if we are doing decode full-cudagraphs it is supported
4493
4494
4495
4496
4497
4498
        if (
            cudagraph_mode.decode_mode() == CUDAGraphMode.FULL
            and min_cg_support == AttentionCGSupport.NEVER
        ):
            msg = (
                f"CUDAGraphMode.{cudagraph_mode.name} is not supported "
4499
                f"with {min_cg_backend_name} backend (support: "
4500
4501
                f"{min_cg_support})"
            )
4502
            if self.compilation_config.mode == CompilationMode.VLLM_COMPILE and (
4503
4504
4505
4506
4507
                self.compilation_config.splitting_ops_contain_attention()
                or self.compilation_config.use_inductor_graph_partition
            ):
                msg += (
                    "; setting cudagraph_mode=PIECEWISE because "
4508
                    "attention is compiled piecewise"
4509
4510
                )
                cudagraph_mode = self.compilation_config.cudagraph_mode = (
4511
                    CUDAGraphMode.PIECEWISE
4512
                )
4513
            else:
4514
4515
                msg += (
                    "; setting cudagraph_mode=NONE because "
4516
                    "attention is not compiled piecewise"
4517
4518
                )
                cudagraph_mode = self.compilation_config.cudagraph_mode = (
4519
                    CUDAGraphMode.NONE
4520
                )
4521
4522
            logger.warning(msg)

4523
4524
        # check that if we are doing spec-decode + decode full-cudagraphs it is
        # supported
4525
4526
4527
4528
4529
4530
4531
4532
        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 "
4533
                f"{min_cg_backend_name} (support: {min_cg_support})"
4534
            )
4535
4536
            if self.compilation_config.splitting_ops_contain_attention():
                msg += "; setting cudagraph_mode=PIECEWISE"
4537
                cudagraph_mode = self.compilation_config.cudagraph_mode = (
4538
                    CUDAGraphMode.PIECEWISE
4539
                )
4540
4541
            else:
                msg += "; setting cudagraph_mode=NONE"
4542
                cudagraph_mode = self.compilation_config.cudagraph_mode = (
4543
                    CUDAGraphMode.NONE
4544
                )
4545
4546
4547
4548
            logger.warning(msg)

        # double check that we can support full cudagraph if they are requested
        # even after automatic downgrades
4549
4550
4551
4552
4553
4554
        if (
            cudagraph_mode.has_full_cudagraphs()
            and min_cg_support == AttentionCGSupport.NEVER
        ):
            raise ValueError(
                f"CUDAGraphMode.{cudagraph_mode.name} is not "
4555
                f"supported with {min_cg_backend_name} backend ("
4556
4557
                f"support:{min_cg_support}) "
                "; please try cudagraph_mode=PIECEWISE, "
4558
                "and make sure compilation mode is VLLM_COMPILE"
4559
            )
4560

4561
4562
4563
4564
4565
4566
4567
4568
4569
4570
4571
4572
4573
4574
        # 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
            )
4575
4576
4577
4578
            capture_sizes = self.compilation_config.cudagraph_capture_sizes
            self.cudagraph_batch_sizes = (
                capture_sizes if capture_sizes is not None else []
            )
4579

4580
4581
        # Trigger cudagraph dispatching keys initialization after
        # resolved cudagraph mode.
4582
4583
        cudagraph_mode = self.compilation_config.cudagraph_mode
        assert cudagraph_mode is not None
4584
        self.cudagraph_dispatcher.initialize_cudagraph_keys(
4585
            cudagraph_mode, self.uniform_decode_query_len
4586
        )
4587

4588
4589
    def calculate_reorder_batch_threshold(self) -> None:
        """
4590
4591
4592
4593
        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.
4594
        """
4595
4596
        min_none_high = lambda a, b: a if b is None else b if a is None else min(a, b)

4597
        reorder_batch_thresholds: list[int | None] = [
4598
4599
4600
            group.get_metadata_builder().reorder_batch_threshold
            for group in self._attn_group_iterator()
        ]
4601
4602
4603
4604
4605
        # 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
4606
        self.reorder_batch_threshold = reduce(min_none_high, reorder_batch_thresholds)  # type: ignore[assignment]
4607

4608
4609
4610
    @staticmethod
    def select_common_block_size(
        kv_manager_block_size: int, attn_groups: list[AttentionGroup]
4611
4612
    ) -> int:
        """
4613
4614
4615
4616
4617
        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.
4618
4619
4620
4621
4622
4623

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

        Returns:
4624
            The selected block size
4625
4626

        Raises:
4627
            ValueError: If no valid block size found
4628
4629
        """

4630
4631
4632
4633
4634
4635
4636
4637
        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
4638
                for supported_size in backend.get_supported_kernel_block_sizes():
4639
4640
4641
4642
4643
4644
4645
4646
4647
4648
4649
4650
4651
4652
4653
4654
4655
4656
4657
4658
4659
4660
4661
4662
4663
4664
4665
4666
4667
4668
                    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
4669
            for supported_size in backend.get_supported_kernel_block_sizes()
4670
4671
            if isinstance(supported_size, int)
        )
4672

4673
4674
4675
4676
4677
4678
        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}. ")
4679

4680
4681
4682
    def may_reinitialize_input_batch(
        self, kv_cache_config: KVCacheConfig, kernel_block_sizes: list[int]
    ) -> None:
4683
4684
4685
4686
4687
4688
4689
        """
        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.
4690
            kernel_block_sizes: The kernel block sizes for each KV cache group.
4691
4692
4693
4694
        """
        block_sizes = [
            kv_cache_group.kv_cache_spec.block_size
            for kv_cache_group in kv_cache_config.kv_cache_groups
4695
            if not isinstance(kv_cache_group.kv_cache_spec, EncoderOnlyAttentionSpec)
4696
        ]
4697
4698
4699
4700

        if block_sizes != [self.cache_config.block_size] or kernel_block_sizes != [
            self.cache_config.block_size
        ]:
4701
4702
4703
            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
4704
4705
                "for more details."
            )
4706
4707
            self.input_batch = InputBatch(
                max_num_reqs=self.max_num_reqs,
4708
                max_model_len=max(self.max_model_len, self.max_encoder_len),
4709
4710
4711
4712
4713
                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,
4714
                kernel_block_sizes=kernel_block_sizes,
4715
                is_spec_decode=bool(self.vllm_config.speculative_config),
4716
                logitsprocs=self.input_batch.logitsprocs,
4717
                logitsprocs_need_output_token_ids=self.input_batch.logitsprocs_need_output_token_ids,
4718
                is_pooling_model=self.is_pooling_model,
4719
                num_speculative_tokens=self.num_spec_tokens,
4720
4721
            )

4722
    def _allocate_kv_cache_tensors(
4723
4724
        self, kv_cache_config: KVCacheConfig
    ) -> dict[str, torch.Tensor]:
4725
        """
4726
4727
4728
        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.

4729
        Args:
4730
            kv_cache_config: The KV cache config
4731
        Returns:
4732
            dict[str, torch.Tensor]: A map between layer names to their
4733
            corresponding memory buffer for KV cache.
4734
        """
4735
4736
        kv_cache_raw_tensors: dict[str, torch.Tensor] = {}
        for kv_cache_tensor in kv_cache_config.kv_cache_tensors:
4737
4738
4739
            tensor = torch.zeros(
                kv_cache_tensor.size, dtype=torch.int8, device=self.device
            )
4740
4741
4742
4743
4744
            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:
4745
4746
4747
4748
            for layer_name in group.layer_names:
                if layer_name in self.runner_only_attn_layers:
                    continue
                layer_names.add(layer_name)
4749
4750
4751
        assert layer_names == set(kv_cache_raw_tensors.keys()), (
            "Some layers are not correctly initialized"
        )
4752
4753
        return kv_cache_raw_tensors

4754
4755
4756
    def _attn_group_iterator(self) -> Iterator[AttentionGroup]:
        return itertools.chain.from_iterable(self.attn_groups)

4757
    def _kv_cache_spec_attn_group_iterator(self) -> Iterator[AttentionGroup]:
4758
4759
        if not self.kv_cache_config.kv_cache_groups:
            return
4760
4761
        for attn_groups in self.attn_groups:
            yield from attn_groups
4762

4763
4764
4765
4766
4767
4768
4769
4770
4771
4772
4773
4774
4775
4776
4777
    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 = []
4778
        for kv_cache_gid, kv_cache_group in enumerate(kv_cache_config.kv_cache_groups):
4779
4780
4781
4782
4783
4784
            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):
4785
                continue
4786
            elif isinstance(kv_cache_spec, AttentionSpec):
4787
4788
4789
                # This is an attention backend that supports virtual
                # block splitting. Get the supported block sizes from
                # all backends in the group.
4790
                attn_groups = self.attn_groups[kv_cache_gid]
4791
                kv_manager_block_size = kv_cache_group.kv_cache_spec.block_size
4792
                selected_kernel_size = self.select_common_block_size(
4793
4794
4795
                    kv_manager_block_size, attn_groups
                )
                kernel_block_sizes.append(selected_kernel_size)
4796
            elif isinstance(kv_cache_spec, MambaSpec):
4797
4798
                # This is likely Mamba or other non-attention cache,
                # no splitting.
4799
                kernel_block_sizes.append(kv_cache_spec.block_size)
4800
4801
4802
4803
4804
4805
            else:
                raise NotImplementedError(
                    f"unknown kv cache spec {kv_cache_group.kv_cache_spec}"
                )
        return kernel_block_sizes

4806
4807
4808
4809
    def _reshape_kv_cache_tensors(
        self,
        kv_cache_config: KVCacheConfig,
        kv_cache_raw_tensors: dict[str, torch.Tensor],
4810
        kernel_block_sizes: list[int],
4811
    ) -> dict[str, torch.Tensor]:
4812
        """
4813
        Reshape the KV cache tensors to the desired shape and dtype.
4814

4815
        Args:
4816
4817
            kv_cache_config: The KV cache config
            kv_cache_raw_tensors: The KV cache buffer of each layer, with
4818
                correct size but uninitialized shape.
4819
            kernel_block_sizes: The kernel block sizes for each KV cache group.
4820
        Returns:
4821
            Dict[str, torch.Tensor]: A map between layer names to their
4822
4823
            corresponding memory buffer for KV cache.
        """
4824
        kv_caches: dict[str, torch.Tensor] = {}
4825
        has_attn, has_mamba = False, False
4826
4827
        for group in self._kv_cache_spec_attn_group_iterator():
            kv_cache_spec = group.kv_cache_spec
4828
            attn_backend = group.backend
4829
4830
4831
4832
            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]
4833
            for layer_name in group.layer_names:
4834
4835
                if layer_name in self.runner_only_attn_layers:
                    continue
4836
4837
                raw_tensor = kv_cache_raw_tensors[layer_name]
                assert raw_tensor.numel() % kv_cache_spec.page_size_bytes == 0
4838
                num_blocks = raw_tensor.numel() // kv_cache_spec.page_size_bytes
4839
                if isinstance(kv_cache_spec, AttentionSpec):
4840
                    has_attn = True
4841
4842
                    num_blocks_per_kv_block = (
                        kv_cache_spec.block_size // kernel_block_size
4843
4844
4845
                    )
                    kernel_num_blocks = num_blocks * num_blocks_per_kv_block

4846
                    kv_cache_shape = attn_backend.get_kv_cache_shape(
4847
                        kernel_num_blocks,
4848
                        kernel_block_size,
4849
4850
                        kv_cache_spec.num_kv_heads,
                        kv_cache_spec.head_size,
4851
4852
                        cache_dtype_str=self.cache_config.cache_dtype,
                    )
4853
                    dtype = kv_cache_spec.dtype
4854
                    try:
4855
                        kv_cache_stride_order = attn_backend.get_kv_cache_stride_order()
4856
                        assert len(kv_cache_stride_order) == len(kv_cache_shape)
4857
                    except (AttributeError, NotImplementedError):
4858
                        kv_cache_stride_order = tuple(range(len(kv_cache_shape)))
4859
4860
4861
4862
4863
                    # 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.
4864
4865
4866
                    kv_cache_shape = tuple(
                        kv_cache_shape[i] for i in kv_cache_stride_order
                    )
4867
4868
4869
4870
4871
                    # Maintain original KV shape view.
                    inv_order = [
                        kv_cache_stride_order.index(i)
                        for i in range(len(kv_cache_stride_order))
                    ]
4872
4873
4874
4875
4876
4877
                    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
4878
                elif isinstance(kv_cache_spec, MambaSpec):
4879
                    has_mamba = True
Chen Zhang's avatar
Chen Zhang committed
4880
4881
                    raw_tensor = kv_cache_raw_tensors[layer_name]
                    state_tensors = []
4882
                    storage_offset_bytes = 0
4883
                    for shape, dtype in zip(kv_cache_spec.shapes, kv_cache_spec.dtypes):
4884
4885
                        dtype_size = get_dtype_size(dtype)
                        num_element_per_page = (
4886
4887
                            kv_cache_spec.page_size_bytes // dtype_size
                        )
Chen Zhang's avatar
Chen Zhang committed
4888
                        target_shape = (num_blocks, *shape)
4889
4890
                        stride = torch.empty(target_shape).stride()
                        target_stride = (num_element_per_page, *stride[1:])
4891
                        assert storage_offset_bytes % dtype_size == 0
4892
4893
4894
4895
                        tensor = torch.as_strided(
                            raw_tensor.view(dtype),
                            size=target_shape,
                            stride=target_stride,
4896
                            storage_offset=storage_offset_bytes // dtype_size,
4897
                        )
Chen Zhang's avatar
Chen Zhang committed
4898
                        state_tensors.append(tensor)
4899
                        storage_offset_bytes += stride[0] * dtype_size
4900
4901

                    kv_caches[layer_name] = state_tensors
4902
                else:
4903
                    raise NotImplementedError
4904
4905

        if has_attn and has_mamba:
4906
            self._update_hybrid_attention_mamba_layout(kv_caches)
4907

4908
4909
        return kv_caches

4910
    def _update_hybrid_attention_mamba_layout(
4911
4912
        self, kv_caches: dict[str, torch.Tensor]
    ) -> None:
4913
        """
4914
4915
        Update the layout of attention layers from (2, num_blocks, ...) to
        (num_blocks, 2, ...).
4916
4917

        Args:
4918
            kv_caches: The KV cache buffer of each layer.
4919
4920
        """

4921
4922
        for group in self._kv_cache_spec_attn_group_iterator():
            kv_cache_spec = group.kv_cache_spec
4923
            for layer_name in group.layer_names:
4924
                kv_cache = kv_caches[layer_name]
4925
4926
4927
4928
                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 "
4929
                        f"a tensor of shape {kv_cache.shape}"
4930
                    )
4931
                    hidden_size = kv_cache.shape[2:].numel()
4932
4933
4934
4935
                    kv_cache.as_strided_(
                        size=kv_cache.shape,
                        stride=(hidden_size, 2 * hidden_size, *kv_cache.stride()[2:]),
                    )
4936

4937
    def initialize_kv_cache_tensors(
4938
        self, kv_cache_config: KVCacheConfig, kernel_block_sizes: list[int]
4939
    ) -> dict[str, torch.Tensor]:
4940
4941
4942
4943
4944
        """
        Initialize the memory buffer for KV cache.

        Args:
            kv_cache_config: The KV cache config
4945
4946
            kernel_block_sizes: The kernel block sizes for each KV cache group.

4947
        Returns:
4948
            Dict[str, torch.Tensor]: A map between layer names to their
4949
4950
            corresponding memory buffer for KV cache.
        """
4951
4952
4953
4954
4955
4956
4957
4958
4959
4960
4961
4962
4963
4964
4965
4966
4967
4968
4969
4970
4971
4972
4973
4974

        # 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
            )
4975

4976
        # Set up cross-layer KV cache sharing
4977
4978
        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)
4979
4980
            kv_caches[layer_name] = kv_caches[target_layer_name]

4981
4982
4983
4984
4985
4986
4987
4988
4989
        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,
        )
4990
4991
4992
        return kv_caches

    def maybe_add_kv_sharing_layers_to_kv_cache_groups(
4993
4994
        self, kv_cache_config: KVCacheConfig
    ) -> None:
4995
4996
4997
4998
4999
5000
5001
5002
5003
5004
5005
5006
5007
5008
5009
5010
5011
5012
        """
        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.
5013
            attn_layers = get_layers_from_vllm_config(self.vllm_config, Attention)
5014
5015
            for layer_name in reversed(attn_layers):
                if layer_name in self.shared_kv_cache_layers:
5016
                    self.kv_sharing_fast_prefill_eligible_layers.add(layer_name)
5017
5018
                else:
                    break
5019

5020
5021
5022
5023
5024
5025
5026
    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
        """
5027
        kv_cache_config = deepcopy(kv_cache_config)
5028
        self.kv_cache_config = kv_cache_config
5029
        self.may_add_encoder_only_layers_to_kv_cache_config()
5030
        self.maybe_add_kv_sharing_layers_to_kv_cache_groups(kv_cache_config)
5031
        self.initialize_attn_backend(kv_cache_config)
5032
5033
5034
5035
5036
5037
        # 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)
5038
5039
5040
5041

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

5042
        # Reinitialize need to after initialize_attn_backend
5043
5044
5045
5046
        self.may_reinitialize_input_batch(kv_cache_config, kernel_block_sizes)
        kv_caches = self.initialize_kv_cache_tensors(
            kv_cache_config, kernel_block_sizes
        )
5047

5048
5049
5050
5051
5052
5053
        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
5054
        if has_kv_transfer_group():
5055
            kv_transfer_group = get_kv_transfer_group()
5056
5057
5058
5059
5060
5061
5062
            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)
5063
            kv_transfer_group.set_host_xfer_buffer_ops(copy_kv_blocks)
Robert Shaw's avatar
Robert Shaw committed
5064

5065
        if self.dcp_world_size > 1:
5066
5067
            layer_type = cast(type[Any], AttentionLayerBase)
            layers = get_layers_from_vllm_config(self.vllm_config, layer_type)
5068
            for layer in layers.values():
5069
5070
5071
5072
                layer_impl = getattr(layer, "impl", None)
                if layer_impl is None:
                    continue
                assert layer_impl.need_to_return_lse_for_decode, (
5073
5074
                    "DCP requires attention impls to return"
                    " the softmax lse for decode, but the impl "
5075
                    f"{layer_impl.__class__.__name__} "
5076
5077
                    "does not return the softmax lse for decode."
                )
5078

5079
5080
5081
5082
5083
    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
5084
        encoder_only_attn_specs: dict[AttentionSpec, list[str]] = defaultdict(list)
5085
5086
5087
        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:
5088
                attn_spec: AttentionSpec = EncoderOnlyAttentionSpec(
5089
5090
5091
                    block_size=block_size,
                    num_kv_heads=attn_module.num_kv_heads,
                    head_size=attn_module.head_size,
5092
5093
                    dtype=self.kv_cache_dtype,
                )
5094
5095
5096
                encoder_only_attn_specs[attn_spec].append(layer_name)
                self.runner_only_attn_layers.add(layer_name)
        if len(encoder_only_attn_specs) > 0:
5097
5098
5099
            assert len(encoder_only_attn_specs) == 1, (
                "Only support one encoder-only attention spec now"
            )
5100
5101
            spec, layer_names = encoder_only_attn_specs.popitem()
            self.kv_cache_config.kv_cache_groups.append(
5102
5103
                KVCacheGroupSpec(layer_names=layer_names, kv_cache_spec=spec)
            )
5104

5105
    def get_kv_cache_spec(self) -> dict[str, KVCacheSpec]:
5106
        """
5107
        Generates the KVCacheSpec by parsing the kv cache format from each
5108
5109
        Attention module in the static forward context.
        Returns:
5110
            KVCacheSpec: A dictionary mapping layer names to their KV cache
5111
5112
            format. Layers that do not need KV cache are not included.
        """
5113
5114
        if has_ec_transfer() and get_ec_transfer().is_producer:
            return {}
5115
        kv_cache_spec: dict[str, KVCacheSpec] = {}
5116
5117
        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
5118
        for layer_name, attn_module in attn_layers.items():
5119
5120
5121
5122
5123
5124
5125
5126
5127
5128
5129
5130
5131
5132
5133
            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
5134

5135
        return kv_cache_spec
5136

5137
    def _to_list(self, sampled_token_ids: torch.Tensor) -> list[list[int]]:
5138
5139
5140
5141
5142
5143
5144
5145
        # 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.
5146
        pinned = self.sampled_token_ids_pinned_cpu[: sampled_token_ids.shape[0]]
5147
5148
5149
        pinned.copy_(sampled_token_ids, non_blocking=True)
        self.transfer_event.record()
        self.transfer_event.synchronize()
5150
        return pinned.tolist()