gpu_model_runner.py 229 KB
Newer Older
1
# SPDX-License-Identifier: Apache-2.0
2
# SPDX-FileCopyrightText: Copyright contributors to the vLLM project
3

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

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

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

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

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

logger = init_logger(__name__)

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

179

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

        # Event on the copy stream so we can synchronize the non-blocking copy.
195
        self.async_copy_ready_event = torch.Event()
196
197
198
199

        # Keep a reference to the device tensor to avoid it being
        # deallocated until we finish copying it to the host.
        self._sampled_token_ids = sampled_token_ids
200
        self.vocab_size = vocab_size
201
        self._logprobs_tensors = logprobs_tensors
202
203
204
205
206

        # Initiate the copy on a separate stream, but do not synchronize it.
        default_stream = torch.cuda.current_stream()
        with torch.cuda.stream(async_output_copy_stream):
            async_output_copy_stream.wait_stream(default_stream)
207
            self.sampled_token_ids_cpu = self._sampled_token_ids.to(
208
209
                "cpu", non_blocking=True
            )
210
211
212
213
214
            self._logprobs_tensors_cpu = (
                self._logprobs_tensors.to_cpu_nonblocking()
                if self._logprobs_tensors
                else None
            )
215
            self.async_copy_ready_event.record()
216
217
218

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

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

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

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


248
249
250
251
252
253
254
255
256
257
258
class ExecuteModelState(NamedTuple):
    """Ephemeral cached state transferred between execute_model() and
    sample_tokens(), after execute_model() returns None."""

    scheduler_output: "SchedulerOutput"
    logits: torch.Tensor
    spec_decode_metadata: SpecDecodeMetadata | None
    spec_decode_common_attn_metadata: CommonAttentionMetadata | None
    hidden_states: torch.Tensor
    sample_hidden_states: torch.Tensor
    aux_hidden_states: list[torch.Tensor] | None
259
    ec_connector_output: ECConnectorOutput | None
260
261


262
263
264
class GPUModelRunner(
    LoRAModelRunnerMixin, KVConnectorModelRunnerMixin, ECConnectorModelRunnerMixin
):
265
266
    def __init__(
        self,
267
        vllm_config: VllmConfig,
268
        device: torch.device,
269
    ):
270
271
272
        self.vllm_config = vllm_config
        self.model_config = vllm_config.model_config
        self.cache_config = vllm_config.cache_config
273
        self.compilation_config = vllm_config.compilation_config
274
275
276
277
278
279
        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
280

281
        from vllm.model_executor.models.utils import set_cpu_offload_max_bytes
282
283

        set_cpu_offload_max_bytes(int(self.cache_config.cpu_offload_gb * 1024**3))
284

285
286
287
288
        model_config = self.model_config
        cache_config = self.cache_config
        scheduler_config = self.scheduler_config
        parallel_config = self.parallel_config
289
        self.device = device
290
291
        self.pin_memory = is_pin_memory_available()
        self.dtype = self.model_config.dtype
292
293
294
        self.kv_cache_dtype = kv_cache_dtype_str_to_dtype(
            cache_config.cache_dtype, self.model_config
        )
295

296
        self.is_pooling_model = model_config.runner_type == "pooling"
297
        self.enable_prompt_embeds = model_config.enable_prompt_embeds
298
        self.is_multimodal_raw_input_only_model = (
299
300
            model_config.is_multimodal_raw_input_only_model
        )
301
302
        # This will be overridden in load_model()
        self.is_multimodal_pruning_enabled = False
303
        self.max_model_len = model_config.max_model_len
304
305
306

        # Always set to false after the first forward pass
        self.calculate_kv_scales = self.cache_config.calculate_kv_scales
307
        self.dcp_world_size = self.parallel_config.decode_context_parallel_size
308
        self.dcp_rank = 0 if self.dcp_world_size <= 1 else get_dcp_group().rank_in_group
309
        self.max_num_tokens = scheduler_config.max_num_batched_tokens
310
        self.max_num_reqs = scheduler_config.max_num_seqs
311

312
313
314
315
316
        # 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 = (
317
318
319
            self.parallel_config.distributed_executor_backend == "external_launcher"
            and len(get_pp_group().ranks) > 0
        )
320

321
        # Model-related.
322
        self.num_query_heads = model_config.get_num_attention_heads(parallel_config)
323
        self.hidden_size = model_config.get_hidden_size()
324
        self.attention_chunk_size = model_config.attention_chunk_size
325
        # Only relevant for models using ALiBi (e.g, MPT)
326
        self.use_alibi = model_config.uses_alibi
327

328
        self.cascade_attn_enabled = not self.model_config.disable_cascade_attn
329

330
        # Multi-modal data support
331
        self.mm_registry = MULTIMODAL_REGISTRY
332
        self.uses_mrope = model_config.uses_mrope
333
        self.uses_xdrope_dim = model_config.uses_xdrope_dim
334
        self.supports_mm_inputs = self.mm_registry.supports_multimodal_inputs(
335
336
            model_config
        )
337

338
339
340
        if self.model_config.is_encoder_decoder:
            # Maximum length of the encoder input, only for encoder-decoder
            # models.
341
            self.max_encoder_len = scheduler_config.max_num_encoder_input_tokens
342
343
344
        else:
            self.max_encoder_len = 0

345
        # Sampler
346
        self.sampler = Sampler(logprobs_mode=self.model_config.logprobs_mode)
347

348
        self.eplb_state: EplbState | None = None
349
350
351
352
353
354
        """
        State of the expert parallelism load balancer.

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

355
        # Lazy initializations
356
        # self.model: nn.Module  # Set after load_model
357
        # Initialize in initialize_kv_cache
358
        self.kv_caches: list[torch.Tensor] = []
359
360
361
        # Initialize in initialize_kv_cache_tensors
        self.cross_layers_kv_cache: torch.Tensor | None = None
        self.cross_layers_attn_backend: type[AttentionBackend] | None = None
362
363
        # indexes: [kv_cache_group_id][attn_group]
        self.attn_groups: list[list[AttentionGroup]] = []
364
365
        # self.kv_cache_config: KVCacheConfig

366
367
        # mm_hash ->  encoder_output
        self.encoder_cache: dict[str, torch.Tensor] = {}
368

369
        self.use_aux_hidden_state_outputs = False
370
371
372
373
374
        # 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:
375
376
377
            self.drafter: (
                NgramProposer | SuffixDecodingProposer | EagleProposer | MedusaProposer
            )
378
379
            if self.speculative_config.method == "ngram":
                self.drafter = NgramProposer(self.vllm_config)
380
381
            elif self.speculative_config.method == "suffix":
                self.drafter = SuffixDecodingProposer(self.vllm_config)
382
            elif self.speculative_config.use_eagle():
383
                self.drafter = EagleProposer(self.vllm_config, self.device, self)
384
                if self.speculative_config.method == "eagle3":
385
386
387
                    self.use_aux_hidden_state_outputs = (
                        self.drafter.eagle3_use_aux_hidden_state
                    )
388
389
            elif self.speculative_config.method == "medusa":
                self.drafter = MedusaProposer(
390
                    vllm_config=self.vllm_config, device=self.device
391
                )
392
            else:
393
394
395
396
                raise ValueError(
                    "Unknown speculative decoding method: "
                    f"{self.speculative_config.method}"
                )
397
            self.rejection_sampler = RejectionSampler(self.sampler)
398

399
400
401
402
        self.num_spec_tokens = 0
        if self.speculative_config:
            self.num_spec_tokens = self.speculative_config.num_speculative_tokens

403
        # Request states.
404
        self.requests: dict[str, CachedRequestState] = {}
405
406
407
        # NOTE(rob): num_prompt_logprobs only includes reqs
        # that are currently in the prefill phase.
        self.num_prompt_logprobs: dict[str, int] = {}
408
        self.comm_stream = torch.cuda.Stream()
409

410
411
412
413
414
415
416
417
418
        # 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.
419
420
421
422
        logits_processors = model_config.logits_processors
        custom_logitsprocs: Sequence[str | type[LogitsProcessor]] = (
            tuple(logits_processors) if logits_processors is not None else ()
        )
423
424
        self.input_batch = InputBatch(
            max_num_reqs=self.max_num_reqs,
425
426
427
            # 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),
428
429
430
            max_num_batched_tokens=self.max_num_tokens,
            device=self.device,
            pin_memory=self.pin_memory,
431
            vocab_size=self.model_config.get_vocab_size(),
432
            block_sizes=[self.cache_config.block_size],
433
            kernel_block_sizes=[self.cache_config.block_size],
434
            is_spec_decode=bool(self.vllm_config.speculative_config),
435
            logitsprocs=build_logitsprocs(
436
437
438
                self.vllm_config,
                self.device,
                self.pin_memory,
439
                self.is_pooling_model,
440
                custom_logitsprocs,
441
            ),
442
443
444
            # 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),
445
            is_pooling_model=self.is_pooling_model,
446
            cp_kv_cache_interleave_size=self.parallel_config.cp_kv_cache_interleave_size,
447
        )
448

449
        self.use_async_scheduling = self.scheduler_config.async_scheduling
450
451
452
453
454
        # 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.
455
        self.prepare_inputs_event: torch.Event | None = None
456
457
        if self.use_async_scheduling:
            self.async_output_copy_stream = torch.cuda.Stream()
458
            self.prepare_inputs_event = torch.Event()
459

460
        # self.cudagraph_batch_sizes sorts in ascending order.
461
462
463
464
        if (
            self.compilation_config.cudagraph_capture_sizes
            and self.compilation_config.cudagraph_mode != CUDAGraphMode.NONE
        ):
465
466
            self.cudagraph_batch_sizes = sorted(
                self.compilation_config.cudagraph_capture_sizes
467
            )
468

469
        # Cache the device properties.
470
        self._init_device_properties()
471

472
        # Persistent buffers for CUDA graphs.
473
474
475
476
477
        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
        )
478
        self.seq_lens = self._make_buffer(self.max_num_reqs, dtype=torch.int32)
479
        self.encoder_seq_lens = self._make_buffer(self.max_num_reqs, dtype=torch.int32)
480
481
482
483
        if self.dcp_world_size > 1:
            self.dcp_local_seq_lens = self._make_buffer(
                self.max_num_reqs, dtype=torch.int32
            )
484
485
486
        # 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.
487
488
489
490
        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)
491
492
        self.discard_request_mask = self._make_buffer(
            self.max_num_reqs, dtype=torch.bool
493
494
495
496
497
498
499
        )
        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
        )
500

501
502
        # Only relevant for multimodal models
        if self.supports_mm_inputs:
503
            self.is_mm_embed = self._make_buffer(self.max_num_tokens, dtype=torch.bool)
504

505
        # Only relevant for models using M-RoPE (e.g, Qwen2-VL)
506
        if self.uses_mrope:
Roger Wang's avatar
Roger Wang committed
507
508
509
510
            # 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
511
512
513
514
515
516

            # 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
517
            self.mrope_positions = self._make_buffer(
518
519
                (3, self.max_num_tokens + 1), dtype=torch.int64
            )
520

521
522
523
524
525
526
527
        # Only relevant for models using XD-RoPE (e.g, HunYuan-VL)
        if self.uses_xdrope_dim > 0:
            # Similar to mrope but use assigned dimension number for RoPE, 4 as default.
            self.xdrope_positions = self._make_buffer(
                (self.uses_xdrope_dim, self.max_num_tokens + 1), dtype=torch.int64
            )

528
        # None in the first PP rank. The rest are set after load_model.
529
        self.intermediate_tensors: IntermediateTensors | None = None
530

531
        # OPTIMIZATION: Cache the tensors rather than creating them every step.
532
        # Keep in int64 to avoid overflow with long context
533
534
535
536
        self.arange_np = np.arange(
            max(self.max_num_reqs + 1, self.max_model_len, self.max_num_tokens),
            dtype=np.int64,
        )
537

538
539
540
541
542
        # 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] = {}
543
544
545
546
547
        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(
548
549
                self.max_num_tokens, dtype=torch.int32, device=self.device
            )
550

551
        self.uniform_decode_query_len = 1 + self.num_spec_tokens
552
553
554
555

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

556
557
558
559
560
561
562
563
564
        self.mm_budget = (
            MultiModalBudget(
                self.model_config,
                self.scheduler_config,
                self.mm_registry,
            )
            if self.supports_mm_inputs
            else None
        )
565

566
        self.reorder_batch_threshold: int | None = None
567

568
569
570
571
572
        # 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()

573
        # Cached outputs.
574
        self._draft_token_ids: list[list[int]] | torch.Tensor | None = None
575
        self.transfer_event = torch.Event()
576
        self.sampled_token_ids_pinned_cpu = torch.empty(
577
            (self.max_num_reqs, 1),
578
579
            dtype=torch.int64,
            device="cpu",
580
581
            pin_memory=self.pin_memory,
        )
582

583
584
        # Pre-allocated tensor for copying valid sampled token counts to CPU,
        # with dedicated stream for overlapping and event for coordination.
585
        self.valid_sampled_token_count_event: torch.Event | None = None
586
587
        self.valid_sampled_token_count_copy_stream: torch.cuda.Stream | None = None
        if self.use_async_scheduling and self.num_spec_tokens:
588
            self.valid_sampled_token_count_event = torch.Event()
589
590
591
592
593
594
595
596
            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,
        )

597
598
        # Ephemeral state transferred between execute_model() and sample_tokens().
        self.execute_model_state: ExecuteModelState | None = None
599
        self.kv_connector_output: KVConnectorOutput | None = None
600

601
602
603
604
    def reset_mm_cache(self) -> None:
        if self.mm_budget:
            self.mm_budget.reset_cache()

605
606
607
608
    def _get_positions(self, num_tokens: Any):
        if isinstance(num_tokens, int):
            if self.uses_mrope:
                return self.mrope_positions.gpu[:, :num_tokens]
609
610
            if self.uses_xdrope_dim > 0:
                return self.xdrope_positions.gpu[:, :num_tokens]
611
612
613
614
            return self.positions.gpu[:num_tokens]
        else:
            if self.uses_mrope:
                return self.mrope_positions.gpu[:, num_tokens]
615
616
            if self.uses_xdrope_dim > 0:
                return self.xdrope_positions.gpu[:, num_tokens]
617
618
            return self.positions.gpu[num_tokens]

619
    def _make_buffer(
620
        self, *size: int | torch.SymInt, dtype: torch.dtype, numpy: bool = True
621
622
623
624
625
626
627
628
    ) -> CpuGpuBuffer:
        return CpuGpuBuffer(
            *size,
            dtype=dtype,
            device=self.device,
            pin_memory=self.pin_memory,
            with_numpy=numpy,
        )
629

630
631
632
    def _init_model_kwargs(self, num_tokens: int):
        model_kwargs = dict[str, Any]()

633
        if not self.is_pooling_model:
634
635
            return model_kwargs

636
637
        num_reqs = self.input_batch.num_reqs
        pooling_params = self.input_batch.get_pooling_params()
638
639
640

        token_type_id_requests = dict[int, Any]()
        for i, param in enumerate(pooling_params):
641
642
643
644
645
            if (
                param.extra_kwargs is not None
                and (token_types := param.extra_kwargs.get("compressed_token_type_ids"))
                is not None
            ):
646
647
648
649
650
                token_type_id_requests[i] = token_types

        if len(token_type_id_requests) == 0:
            return model_kwargs

651
        seq_lens = self.seq_lens.gpu[:num_reqs]
652
653
654
655
656
657
658
659
        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(
660
661
            device=self.device
        )
662
663
        return model_kwargs

664
    def _may_reorder_batch(self, scheduler_output: "SchedulerOutput") -> None:
665
666
        """
        Update the order of requests in the batch based on the attention
667
        backend's needs. For example, some attention backends (namely MLA) may
668
669
670
671
672
673
        want to separate requests based on if the attention computation will be
        compute-bound or memory-bound.

        Args:
            scheduler_output: The scheduler output.
        """
674
675
676
677
678
679
680
681
        # 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

682
683
684
685
        if self.reorder_batch_threshold is not None:
            reorder_batch_to_split_decodes_and_prefills(
                self.input_batch,
                scheduler_output,
686
687
                decode_threshold=self.reorder_batch_threshold,
            )
688

689
690
    # Note: used for model runner override.
    def _init_device_properties(self) -> None:
691
        """Initialize attributes from torch.cuda.get_device_properties"""
692
693
694
695
696
697
698
        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()

699
    def _update_states(self, scheduler_output: "SchedulerOutput") -> None:
700
701
702
703
704
705
        """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.

706
707
        The SamplingMetadata is updated and copied to the GPU if there is a
        new/resumed/paused/finished request in the batch.
708
709
        """
        # Remove finished requests from the cached states.
710
711
        for req_id in scheduler_output.finished_req_ids:
            self.requests.pop(req_id, None)
712
            self.num_prompt_logprobs.pop(req_id, None)
713
714
715
716
717
718
719
        # 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:
720
            self.input_batch.remove_request(req_id)
721
722

        # Free the cached encoder outputs.
723
724
        for mm_hash in scheduler_output.free_encoder_mm_hashes:
            self.encoder_cache.pop(mm_hash, None)
725

726
727
728
729
730
731
732
733
734
735
736
737
738
        # 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:
739
            self.input_batch.remove_request(req_id)
740

741
        reqs_to_add: list[CachedRequestState] = []
742
        # Add new requests to the cached states.
743
744
745
        for new_req_data in scheduler_output.scheduled_new_reqs:
            req_id = new_req_data.req_id
            sampling_params = new_req_data.sampling_params
746
            pooling_params = new_req_data.pooling_params
747

748
749
750
751
            if (
                sampling_params
                and sampling_params.sampling_type == SamplingType.RANDOM_SEED
            ):
752
753
754
755
756
                generator = torch.Generator(device=self.device)
                generator.manual_seed(sampling_params.seed)
            else:
                generator = None

757
758
            if self.is_pooling_model:
                assert pooling_params is not None
759
760
                task = pooling_params.task
                assert task is not None, "You did not set `task` in the API"
761

762
                model = cast(VllmModelForPooling, self.get_model())
763
                to_update = model.pooler.get_pooling_updates(task)
764
765
                to_update.apply(pooling_params)

766
            req_state = CachedRequestState(
767
                req_id=req_id,
768
                prompt_token_ids=new_req_data.prompt_token_ids,
769
                prompt_embeds=new_req_data.prompt_embeds,
770
                mm_features=new_req_data.mm_features,
771
                sampling_params=sampling_params,
772
                pooling_params=pooling_params,
773
                generator=generator,
774
775
                block_ids=new_req_data.block_ids,
                num_computed_tokens=new_req_data.num_computed_tokens,
776
                output_token_ids=[],
777
                lora_request=new_req_data.lora_request,
778
            )
779
780
            self.requests[req_id] = req_state

781
782
783
784
785
786
787
            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
                )

788
            # Only relevant for models using M-RoPE (e.g, Qwen2-VL)
789
            if self.uses_mrope:
790
                self._init_mrope_positions(req_state)
791

792
793
794
795
            # Only relevant for models using XD-RoPE (e.g, HunYuan-VL)
            if self.uses_xdrope_dim > 0:
                self._init_xdrope_positions(req_state)

796
            reqs_to_add.append(req_state)
797

798
        # Update the states of the running/resumed requests.
799
        is_last_rank = get_pp_group().is_last_rank
800
        req_data = scheduler_output.scheduled_cached_reqs
801
802
803
804
805

        # 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()

806
        for i, req_id in enumerate(req_data.req_ids):
807
            req_state = self.requests[req_id]
808
809
            num_computed_tokens = req_data.num_computed_tokens[i]
            new_block_ids = req_data.new_block_ids[i]
810
            resumed_from_preemption = req_id in req_data.resumed_req_ids
811
            num_output_tokens = req_data.num_output_tokens[i]
812
            req_index = self.input_batch.req_id_to_index.get(req_id)
813

814
815
816
817
818
819
820
821
822
823
824
825
826
827
828
829
830
831
832
833
834
835
836
            # 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)
837

838
            # Update the cached states.
839
            req_state.num_computed_tokens = num_computed_tokens
840
841
842
843
844
845
846
847

            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.
848
849
850
                num_new_tokens = (
                    num_computed_tokens + len(new_token_ids) - req_state.num_tokens
                )
851
852
853
854
                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:
855
                    req_state.output_token_ids.extend(new_token_ids[-num_new_tokens:])
856
857
858
859
860
            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:
861
862
863
864
                    end_idx = (
                        self.input_batch.num_prompt_tokens[req_index]
                        + num_output_tokens
                    )
865
866
                    self.input_batch.num_tokens[req_index] = end_idx
                    self.input_batch.num_tokens_no_spec[req_index] = end_idx
867

868
            # Update the block IDs.
869
            if not resumed_from_preemption:
870
871
                if new_block_ids is not None:
                    # Append the new blocks to the existing block IDs.
872
                    for block_ids, new_ids in zip(req_state.block_ids, new_block_ids):
873
                        block_ids.extend(new_ids)
874
            else:
875
                assert req_index is None
876
                assert new_block_ids is not None
877
878
                # The request is resumed from preemption.
                # Replace the existing block IDs with the new ones.
879
                req_state.block_ids = new_block_ids
880
881
882
883
884

            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.
885
886
887
888
889
890
891

                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:]

892
                reqs_to_add.append(req_state)
893
894
895
                continue

            # Update the persistent batch.
896
            self.input_batch.num_computed_tokens_cpu[req_index] = num_computed_tokens
897
            if new_block_ids is not None:
898
                self.input_batch.block_table.append_row(new_block_ids, req_index)
899
900
901
902
903
904
905

            # 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)
906
                self.input_batch.token_ids_cpu[
907
908
909
                    req_index, start_token_index:end_token_index
                ] = new_token_ids
                self.input_batch.num_tokens_no_spec[req_index] = end_token_index
910
                self.input_batch.num_tokens[req_index] = end_token_index
911

912
            # Add spec_token_ids to token_ids_cpu.
913
            spec_token_ids = scheduler_output.scheduled_spec_decode_tokens.get(
914
                req_id, []
915
            )
916
917
918
919
920
            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:
921
922
923
                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[
924
925
                    req_index, start_index:end_token_index
                ] = spec_token_ids
926
927
                # NOTE(woosuk): `num_tokens` here may include spec tokens.
                self.input_batch.num_tokens[req_index] += num_spec_tokens
928
929
930
931
932
933

            # 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.
934
935
            self.input_batch.spec_token_ids[req_index].clear()
            self.input_batch.spec_token_ids[req_index].extend(spec_token_ids)
936

937
938
939
940
941
942
943
944
945
            # 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)
946
947
        # Add the new or resumed requests to the persistent batch.
        # The smaller empty indices are filled first.
948
949
        for request in reqs_to_add:
            self.input_batch.add_request(request)
950

951
952
953
954
955
956
        # 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()
957

958
    def _update_states_after_model_execute(
959
960
        self, output_token_ids: torch.Tensor
    ) -> None:
961
962
963
964
965
966
967
968
969
970
971
972
        """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.
973
974
975
976
977
978
979
980
981
982
983
984
985
986
987
988
989
990
991
992
        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()
        )
993
994
995
        for i, num_tokens in enumerate(num_accepted_tokens):
            self.input_batch.num_accepted_tokens_cpu[i] = num_tokens

996
    def _init_mrope_positions(self, req_state: CachedRequestState):
997
998
        model = self.get_model()
        assert supports_mrope(model), "M-RoPE support is not implemented."
999
1000
1001
1002
        assert req_state.prompt_token_ids is not None, (
            "M-RoPE requires prompt_token_ids to be available."
        )
        mrope_model = cast(SupportsMRoPE, model)
1003
1004

        req_state.mrope_positions, req_state.mrope_position_delta = (
1005
            mrope_model.get_mrope_input_positions(
1006
                req_state.prompt_token_ids,
1007
                req_state.mm_features,
1008
            )
1009
        )
1010

1011
1012
1013
1014
1015
1016
1017
1018
1019
1020
1021
1022
1023
    def _init_xdrope_positions(self, req_state: CachedRequestState):
        model = self.get_model()
        xdrope_model = cast(SupportsXDRoPE, model)
        assert req_state.prompt_token_ids is not None, (
            "XD-RoPE requires prompt_token_ids to be available."
        )
        assert supports_xdrope(model), "XD-RoPE support is not implemented."

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

1024
    def _extract_mm_kwargs(
1025
        self,
1026
1027
        scheduler_output: "SchedulerOutput",
    ) -> BatchedTensorInputs:
1028
        if not scheduler_output or not self.is_multimodal_raw_input_only_model:
1029
            return {}
1030

1031
1032
        mm_kwargs = list[MultiModalKwargsItem]()
        for req in scheduler_output.scheduled_new_reqs:
1033
1034
1035
            for feature in req.mm_features:
                if feature.data is not None:
                    mm_kwargs.append(feature.data)
1036

1037
        # Input all modalities at once
1038
        model = cast(SupportsMultiModal, self.model)
1039
1040
        mm_kwargs_combined: BatchedTensorInputs = {}
        for _, _, mm_kwargs_group in group_mm_kwargs_by_modality(
1041
1042
1043
1044
            mm_kwargs,
            device=self.device,
            pin_memory=self.pin_memory,
            merge_by_field_config=model.merge_by_field_config,
1045
            multimodal_cpu_fields=model.multimodal_cpu_fields,
1046
1047
        ):
            mm_kwargs_combined.update(mm_kwargs_group)
1048

1049
        return mm_kwargs_combined
1050

1051
    def _dummy_mm_kwargs(self, num_seqs: int) -> BatchedTensorInputs:
1052
        if not self.is_multimodal_raw_input_only_model:
1053
            return {}
1054

1055
1056
1057
1058
1059
        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)
1060

1061
1062
1063
    def _get_cumsum_and_arange(
        self,
        num_tokens: np.ndarray,
1064
        cumsum_dtype: np.dtype | None = None,
1065
1066
1067
1068
1069
1070
1071
1072
1073
1074
1075
1076
1077
1078
1079
1080
    ) -> 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

1081
    def _prepare_input_ids(
1082
1083
1084
1085
        self,
        scheduler_output: "SchedulerOutput",
        total_num_scheduled_tokens: int,
        cu_num_tokens: np.ndarray,
1086
    ) -> None:
1087
        """Prepare the input IDs for the current batch.
1088

1089
1090
1091
1092
1093
1094
1095
        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)
1096
1097
1098
            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)
1099
1100
1101
1102
1103
1104
1105
            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
1106
1107
1108
1109
        sample_flattened_indices: list[int] = []
        spec_flattened_indices: list[int] = []
        prev_common_req_indices: list[int] = []
        prev_draft_token_indices: list[int] = []
1110
1111
        indices_match = True
        max_flattened_index = -1
1112
1113
1114
        total_num_spec_tokens = 0
        scheduled_spec_tokens = scheduler_output.scheduled_spec_decode_tokens

1115
1116
1117
1118
1119
        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.
1120
1121
                draft_len = len(scheduled_spec_tokens.get(req_id, ()))
                total_num_spec_tokens += draft_len
1122
                flattened_index = cu_num_tokens[cur_index].item() - 1
1123
1124
1125
1126
1127
1128
1129
1130
1131
1132
1133
1134
1135
1136
1137
                # 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))
1138
                indices_match &= prev_index == flattened_index
1139
                max_flattened_index = max(max_flattened_index, flattened_index)
1140
1141
1142
        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:
1143
1144
1145
            # 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)
1146
1147
1148
            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)
1149
1150
        if num_commmon_tokens == 0:
            # No requests in common with the previous iteration
1151
            # So input_ids.cpu will have all the input ids.
1152
1153
1154
1155
1156
1157
1158
            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_(
1159
1160
1161
                self.input_batch.prev_sampled_token_ids[:num_commmon_tokens, 0],
                non_blocking=True,
            )
1162
1163
            if self.enable_prompt_embeds:
                self.is_token_ids.gpu[:num_commmon_tokens] = True
1164
            return
1165
        # Upload the index tensors asynchronously so the scatter can be non-blocking.
1166
1167
        sampled_tokens_index_tensor = torch.tensor(
            sample_flattened_indices, dtype=torch.int64, pin_memory=self.pin_memory
1168
        ).to(self.device, non_blocking=True)
1169
        prev_common_req_indices_tensor = torch.tensor(
1170
1171
            prev_common_req_indices, dtype=torch.int64, pin_memory=self.pin_memory
        ).to(self.device, non_blocking=True)
1172
1173
        self.input_ids.gpu.scatter_(
            dim=0,
1174
            index=sampled_tokens_index_tensor,
1175
            src=self.input_batch.prev_sampled_token_ids[
1176
1177
1178
                prev_common_req_indices_tensor, 0
            ],
        )
1179

1180
1181
1182
1183
1184
1185
1186
1187
1188
1189
1190
1191
1192
1193
1194
1195
1196
1197
1198
1199
1200
1201
1202
        # 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],
        )

1203
1204
    def _get_encoder_seq_lens(
        self,
1205
        num_scheduled_tokens: dict[str, int],
1206
1207
        kv_cache_spec: KVCacheSpec,
        num_reqs: int,
1208
    ) -> tuple[torch.Tensor | None, np.ndarray | None]:
1209
        if not isinstance(kv_cache_spec, CrossAttentionSpec):
1210
            return None, None
1211
1212
1213

        # Build encoder_seq_lens array mapping request indices to
        # encoder lengths for inputs scheduled in this batch
1214
        for req_id in num_scheduled_tokens:
1215
            req_index = self.input_batch.req_id_to_index[req_id]
1216
1217
1218
1219
1220
1221
1222
1223
1224
1225
1226
1227
1228
1229
1230
1231
            req_state = self.requests[req_id]
            if req_state.mm_features is None:
                self.encoder_seq_lens.np[req_index] = 0
                continue

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

        self.encoder_seq_lens.copy_to_gpu(num_reqs)
        encoder_seq_lens = self.encoder_seq_lens.gpu[:num_reqs]
        encoder_seq_lens_cpu = self.encoder_seq_lens.np[:num_reqs]
1232

1233
        return encoder_seq_lens, encoder_seq_lens_cpu
1234

1235
    def _prepare_inputs(
1236
1237
1238
        self,
        scheduler_output: "SchedulerOutput",
        num_scheduled_tokens: np.ndarray,
1239
1240
    ) -> tuple[
        torch.Tensor,
1241
        SpecDecodeMetadata | None,
1242
    ]:
1243
1244
        """
        :return: tuple[
1245
            logits_indices, spec_decode_metadata,
1246
1247
        ]
        """
1248
1249
1250
1251
1252
1253
1254
        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.
1255
        self.input_batch.block_table.commit_block_table(num_reqs)
1256
1257
1258

        # Get request indices.
        # E.g., [2, 5, 3] -> [0, 0, 1, 1, 1, 1, 1, 2, 2, 2]
1259
        req_indices = np.repeat(self.arange_np[:num_reqs], num_scheduled_tokens)
1260

1261
1262
        # cu_num_tokens: [2, 5, 3] -> [2, 7, 10]
        # arange: [0, 1, 0, 1, 2, 3, 4, 0, 1, 2]
1263
        cu_num_tokens, arange = self._get_cumsum_and_arange(num_scheduled_tokens)
1264
1265

        # Get positions.
1266
        positions_np = self.positions.np[:total_num_scheduled_tokens]
1267
1268
1269
1270
1271
        np.add(
            self.input_batch.num_computed_tokens_cpu[req_indices],
            arange,
            out=positions_np,
        )
1272

1273
1274
        # Calculate M-RoPE positions.
        # Only relevant for models using M-RoPE (e.g, Qwen2-VL)
1275
        if self.uses_mrope:
1276
1277
            self._calc_mrope_positions(scheduler_output)

1278
1279
1280
1281
1282
        # Calculate XD-RoPE positions.
        # Only relevant for models using XD-RoPE (e.g, HunYuan-VL)
        if self.uses_xdrope_dim > 0:
            self._calc_xdrope_positions(scheduler_output)

1283
1284
1285
1286
        # 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.
1287
1288
1289
        token_indices = (
            positions_np + req_indices * self.input_batch.token_ids_cpu.shape[1]
        )
1290
        token_indices_tensor = torch.from_numpy(token_indices)
1291

1292
1293
1294
        # 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.
1295
1296
1297
1298
1299
1300
        torch.index_select(
            self.input_batch.token_ids_cpu_tensor.flatten(),
            0,
            token_indices_tensor,
            out=self.input_ids.cpu[:total_num_scheduled_tokens],
        )
1301
        if self.enable_prompt_embeds:
1302
            is_token_ids = self.input_batch.is_token_ids_tensor.flatten()
1303
1304
1305
1306
            torch.index_select(
                is_token_ids,
                0,
                token_indices_tensor,
1307
1308
                out=self.is_token_ids.cpu[:total_num_scheduled_tokens],
            )
1309
1310
1311
1312
1313
1314
1315
1316
1317
1318
1319
1320
1321
1322
1323
1324
1325
1326
1327
1328
1329
1330
1331
1332
1333
1334
1335
1336
1337
1338
1339
1340
1341

        # 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:
1342
1343
1344
                    self.inputs_embeds.cpu[
                        output_idx : output_idx + actual_num_sched
                    ].copy_(req_embeds[start_pos:actual_end])
1345
1346

                output_idx += num_sched
1347

1348
1349
        self.input_batch.block_table.compute_slot_mapping(req_indices, positions_np)
        self.input_batch.block_table.commit_slot_mapping(total_num_scheduled_tokens)
1350
1351

        # Prepare the attention metadata.
1352
        self.query_start_loc.np[0] = 0
1353
        self.query_start_loc.np[1 : num_reqs + 1] = cu_num_tokens
1354
1355
        # Note: pad query_start_loc to be non-decreasing, as kernels
        # like FlashAttention requires that
1356
        self.query_start_loc.np[num_reqs + 1 :].fill(cu_num_tokens[-1])
1357
        self.query_start_loc.copy_to_gpu()
1358
        query_start_loc = self.query_start_loc.gpu[: num_reqs + 1]
1359

1360
        self.seq_lens.np[:num_reqs] = (
1361
1362
            self.input_batch.num_computed_tokens_cpu[:num_reqs] + num_scheduled_tokens
        )
1363
        # Fill unused with 0 for full cuda graph mode.
1364
1365
        self.seq_lens.np[num_reqs:].fill(0)
        self.seq_lens.copy_to_gpu()
1366

1367
        num_tokens = [self.requests[r].num_tokens for r in self.input_batch.req_ids]
1368
1369
        num_tokens_np = np.array(num_tokens, dtype=np.int32)

1370
        # Record which requests should not be sampled,
1371
        # so that we could clear the sampled tokens before returning
1372
1373
        self.discard_request_mask.np[:num_reqs] = (
            self.seq_lens.np[:num_reqs] < num_tokens_np
1374
        )
1375
        self.discard_request_mask.copy_to_gpu(num_reqs)
1376

1377
        # Copy the tensors to the GPU.
1378
1379
1380
1381
1382
        self._prepare_input_ids(
            scheduler_output,
            total_num_scheduled_tokens,
            cu_num_tokens,
        )
1383

1384
        if self.uses_mrope:
1385
            # Only relevant for models using M-RoPE (e.g, Qwen2-VL)
1386
1387
            self.mrope_positions.gpu[:, :total_num_scheduled_tokens].copy_(
                self.mrope_positions.cpu[:, :total_num_scheduled_tokens],
1388
1389
                non_blocking=True,
            )
1390
1391
1392
1393
1394
1395
        elif self.uses_xdrope_dim > 0:
            # Only relevant for models using XD-RoPE (e.g, HunYuan-VL)
            self.xdrope_positions.gpu[:, :total_num_scheduled_tokens].copy_(
                self.xdrope_positions.cpu[:, :total_num_scheduled_tokens],
                non_blocking=True,
            )
1396
1397
        else:
            # Common case (1D positions)
1398
            self.positions.copy_to_gpu(total_num_scheduled_tokens)
1399

1400
        use_spec_decode = len(scheduler_output.scheduled_spec_decode_tokens) > 0
1401
1402
1403
1404
1405
1406
1407
        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
1408
            num_draft_tokens = None
1409
            spec_decode_metadata = None
1410
            num_sampled_tokens = np.ones(num_reqs, dtype=np.int32)
1411
1412
1413
1414
1415
        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)
1416
1417
1418
            # 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)
1419
1420
1421
1422
            for (
                req_id,
                draft_token_ids,
            ) in scheduler_output.scheduled_spec_decode_tokens.items():
1423
1424
                req_idx = self.input_batch.req_id_to_index[req_id]
                num_draft_tokens[req_idx] = len(draft_token_ids)
1425
1426
1427
1428
1429
1430
1431
1432
                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
                )
1433
            spec_decode_metadata = self._calc_spec_decode_metadata(
1434
1435
                num_draft_tokens, cu_num_tokens
            )
1436
            logits_indices = spec_decode_metadata.logits_indices
1437
            num_sampled_tokens = num_draft_tokens + 1
1438
            # For DECODE only cuda graph of some attention backends (e.g., GDN).
1439
            self.num_decode_draft_tokens.np[:num_reqs] = num_decode_draft_tokens
1440
1441
            self.num_decode_draft_tokens.np[num_reqs:].fill(-1)
            self.num_decode_draft_tokens.copy_to_gpu()
1442

1443
1444
1445
1446
1447
        # Hot-Swap lora model
        if self.lora_config:
            assert (
                np.sum(num_sampled_tokens)
                <= self.vllm_config.scheduler_config.max_num_batched_tokens
1448
            )
1449
1450
1451
1452
1453
1454
1455
1456
1457
1458
1459
            self.set_active_loras(
                self.input_batch, num_scheduled_tokens, num_sampled_tokens
            )

        return (
            logits_indices,
            spec_decode_metadata,
        )

    def _build_attention_metadata(
        self,
1460
        num_tokens: int,
1461
        num_reqs: int,
1462
1463
1464
        max_query_len: int,
        num_tokens_padded: int | None = None,
        num_reqs_padded: int | None = None,
1465
1466
1467
1468
        ubatch_slices: UBatchSlices | None = None,
        logits_indices: torch.Tensor | None = None,
        use_spec_decode: bool = False,
        for_cudagraph_capture: bool = False,
1469
        num_scheduled_tokens: dict[str, int] | None = None,
1470
1471
1472
1473
1474
        cascade_attn_prefix_lens: list[list[int]] | None = None,
    ) -> tuple[PerLayerAttnMetadata, CommonAttentionMetadata | None]:
        """
        :return: tuple[attn_metadata, spec_decode_common_attn_metadata]
        """
1475
1476
1477
        num_tokens_padded = num_tokens_padded or num_tokens
        num_reqs_padded = num_reqs_padded or num_reqs

1478
        logits_indices_padded = None
1479
        num_logits_indices = None
1480
1481
1482
1483
1484
1485
        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
                )
1486

1487
1488
1489
1490
1491
1492
        # 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,
1493
                self.parallel_config.cp_kv_cache_interleave_size,
1494
            )
1495
1496
            self.dcp_local_seq_lens.cpu[num_reqs:].fill_(0)
            self.dcp_local_seq_lens.copy_to_gpu(num_reqs_padded)
1497

1498
1499
1500
        attn_metadata: PerLayerAttnMetadata = {}
        if ubatch_slices is not None:
            attn_metadata = [dict() for _ in range(len(ubatch_slices))]
1501

1502
1503
1504
1505
1506
1507
1508
1509
        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()

1510
1511
        if use_spec_decode:
            self.num_accepted_tokens.np[:num_reqs] = (
1512
1513
                self.input_batch.num_accepted_tokens_cpu[:num_reqs]
            )
1514
1515
            self.num_accepted_tokens.np[num_reqs:].fill(1)
            self.num_accepted_tokens.copy_to_gpu()
1516

1517
1518
1519
1520
1521
1522
1523
1524
1525
1526
1527
1528
1529
1530
1531
1532
        # Used in the below loop, uses padded shapes
        query_start_loc = self.query_start_loc.gpu[: num_reqs_padded + 1]
        query_start_loc_cpu = self.query_start_loc.cpu[: num_reqs_padded + 1]
        seq_lens = self.seq_lens.gpu[:num_reqs_padded]
        seq_lens_cpu = self.seq_lens.cpu[:num_reqs_padded]
        num_computed_tokens_cpu = self.input_batch.num_computed_tokens_cpu_tensor[
            :num_reqs_padded
        ]

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

        spec_decode_common_attn_metadata = None

1533
1534
        # Prepare the attention metadata for each KV cache group and make layers
        # in the same group share the same metadata.
1535
        for kv_cache_gid, kv_cache_group in enumerate(
1536
1537
            self.kv_cache_config.kv_cache_groups
        ):
1538
1539
            encoder_seq_lens, encoder_seq_lens_cpu = self._get_encoder_seq_lens(
                num_scheduled_tokens or {},
1540
                kv_cache_group.kv_cache_spec,
1541
                num_reqs_padded,
1542
            )
1543

1544
            if isinstance(kv_cache_group.kv_cache_spec, EncoderOnlyAttentionSpec):
1545
1546
1547
                # 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(
1548
                    (num_reqs_padded, 1),
1549
                    dtype=torch.int32,
1550
1551
1552
                    device=self.device,
                )
                slot_mapping = torch.zeros(
1553
                    (num_tokens_padded,),
1554
1555
1556
                    dtype=torch.int64,
                    device=self.device,
                )
1557
            else:
1558
                blk_table = self.input_batch.block_table[kv_cache_gid]
1559
1560
                blk_table_tensor = blk_table.get_device_tensor(num_reqs_padded)
                slot_mapping = blk_table.slot_mapping.gpu[:num_tokens_padded]
1561
1562

                # Fill unused with -1. Needed for reshape_and_cache in full cuda
1563
1564
1565
                # graph mode. `blk_table_tensor` -1 to match mamba PAD_SLOT_ID
                slot_mapping[num_tokens:num_tokens_padded].fill_(-1)
                blk_table_tensor[num_reqs:num_reqs_padded].fill_(-1)
1566

1567
            common_attn_metadata = CommonAttentionMetadata(
1568
1569
1570
1571
1572
                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,
1573
1574
1575
                num_actual_tokens=num_tokens_padded,
                num_reqs=num_reqs_padded,
                max_query_len=max_query_len,
1576
                max_seq_len=max_seq_len,
1577
1578
                block_table_tensor=blk_table_tensor,
                slot_mapping=slot_mapping,
1579
                logits_indices_padded=logits_indices_padded,
1580
                num_logits_indices=num_logits_indices,
1581
                causal=True,
1582
                encoder_seq_lens=encoder_seq_lens,
1583
                encoder_seq_lens_cpu=encoder_seq_lens_cpu,
1584
                dcp_local_seq_lens=dcp_local_seq_lens,
1585
                dcp_local_seq_lens_cpu=dcp_local_seq_lens_cpu,
1586
1587
            )

1588
            if self.speculative_config and spec_decode_common_attn_metadata is None:
1589
                if isinstance(self.drafter, EagleProposer):
1590
                    if self.drafter.attn_layer_names[0] in kv_cache_group.layer_names:
1591
1592
1593
                        spec_decode_common_attn_metadata = common_attn_metadata
                else:
                    spec_decode_common_attn_metadata = common_attn_metadata
1594

1595
1596
1597
1598
1599
1600
            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
                )
1601
                builder = attn_group.get_metadata_builder()
1602

1603
                extra_attn_metadata_args = {}
1604
                if use_spec_decode and isinstance(builder, GDNAttentionMetadataBuilder):
1605
                    extra_attn_metadata_args = dict(
1606
1607
1608
                        num_accepted_tokens=self.num_accepted_tokens.gpu[
                            :num_reqs_padded
                        ],
1609
                        num_decode_draft_tokens_cpu=self.num_decode_draft_tokens.cpu[
1610
                            :num_reqs_padded
1611
                        ],
1612
1613
                    )

1614
1615
                if ubatch_slices is not None:
                    common_attn_metadata_list = split_attn_metadata(
1616
1617
                        ubatch_slices, common_attn_metadata
                    )
1618
                    for ubid, common_attn_metadata in enumerate(
1619
1620
                        common_attn_metadata_list
                    ):
1621
1622
1623
1624
1625
1626
1627
1628
1629
1630
1631
                        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:
1632
1633
1634
1635
                            assert type(attn_metadata) is list
                            attn_metadata[ubid][layer_name] = attn_metadata_i
                else:
                    assert isinstance(attn_metadata, dict)
1636
1637
1638
1639
1640
1641
1642
1643
1644
1645
                    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,
                        )
1646
1647
                    for layer_name in attn_group.layer_names:
                        attn_metadata[layer_name] = attn_metadata_i
1648

1649
1650
1651
1652
1653
1654
1655
1656
1657
1658
        if spec_decode_common_attn_metadata is not None and (
            num_reqs != num_reqs_padded or num_tokens != num_tokens_padded
        ):
            # Currently the drafter still only uses piecewise cudagraphs (and modifies
            # the attention metadata in directly), and therefore does not want to use
            # padded attention metadata.
            spec_decode_common_attn_metadata = (
                spec_decode_common_attn_metadata.unpadded(num_tokens, num_reqs)
            )

1659
        return attn_metadata, spec_decode_common_attn_metadata
1660

1661
1662
1663
    def _compute_cascade_attn_prefix_lens(
        self,
        num_scheduled_tokens: np.ndarray,
1664
        num_computed_tokens: np.ndarray,
1665
1666
1667
1668
1669
1670
1671
        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
        """
1672

1673
1674
1675
1676
1677
1678
1679
1680
1681
1682
1683
1684
1685
1686
        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,
1687
                        num_computed_tokens,
1688
1689
1690
1691
1692
1693
1694
1695
                        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
1696

1697
1698
1699
    def _compute_cascade_attn_prefix_len(
        self,
        num_scheduled_tokens: np.ndarray,
1700
        num_computed_tokens: np.ndarray,
1701
        num_common_prefix_blocks: int,
1702
1703
        kv_cache_spec: KVCacheSpec,
        attn_metadata_builder: AttentionMetadataBuilder,
1704
1705
1706
1707
1708
1709
1710
1711
1712
1713
1714
1715
1716
1717
1718
1719
1720
1721
    ) -> 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.
        """
1722

1723
        common_prefix_len = num_common_prefix_blocks * kv_cache_spec.block_size
1724
1725
1726
1727
1728
1729
1730
1731
1732
1733
1734
1735
1736
1737
1738
1739
1740
1741
1742
1743
1744
1745
1746
1747
1748
1749
1750
1751
1752
1753
1754
1755
1756
1757
1758
1759
1760
        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]
1761
        # Request 3's num_computed_tokens: 3 (i.e., [A, B, C])
1762
1763
1764
1765
1766
        # 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.
1767
        common_prefix_len = min(common_prefix_len, num_computed_tokens.min())
1768
        # common_prefix_len should be a multiple of the block size.
1769
1770
1771
1772
1773
1774
1775
1776
1777
1778
1779
        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
        )
1780
1781
        assert isinstance(kv_cache_spec, AttentionSpec)
        use_cascade = attn_metadata_builder.use_cascade_attention(
1782
1783
1784
            common_prefix_len=common_prefix_len,
            query_lens=num_scheduled_tokens,
            num_query_heads=self.num_query_heads,
1785
            num_kv_heads=kv_cache_spec.num_kv_heads,
1786
            use_alibi=self.use_alibi,
1787
            use_sliding_window=use_sliding_window,
1788
            use_local_attention=use_local_attention,
1789
            num_sms=self.num_sms,
1790
            dcp_world_size=self.dcp_world_size,
1791
1792
1793
        )
        return common_prefix_len if use_cascade else 0

1794
1795
    def _calc_mrope_positions(self, scheduler_output: "SchedulerOutput"):
        mrope_pos_ptr = 0
1796
        for index, req_id in enumerate(self.input_batch.req_ids):
1797
1798
1799
            req = self.requests[req_id]
            assert req.mrope_positions is not None

1800
1801
            num_computed_tokens = self.input_batch.num_computed_tokens_cpu[index]
            num_scheduled_tokens = scheduler_output.num_scheduled_tokens[req_id]
1802
            num_prompt_tokens = length_from_prompt_token_ids_or_embeds(
1803
1804
                req.prompt_token_ids, req.prompt_embeds
            )
1805
1806

            if num_computed_tokens + num_scheduled_tokens > num_prompt_tokens:
1807
1808
                prompt_part_len = max(0, num_prompt_tokens - num_computed_tokens)
                completion_part_len = max(0, num_scheduled_tokens - prompt_part_len)
1809
1810
1811
1812
1813
1814
1815
1816
1817
1818
1819
1820
1821
            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

1822
1823
1824
                self.mrope_positions.cpu[:, dst_start:dst_end] = req.mrope_positions[
                    :, src_start:src_end
                ]
1825
1826
1827
1828
1829
1830
1831
                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

1832
                assert req.mrope_position_delta is not None
1833
                MRotaryEmbedding.get_next_input_positions_tensor(
1834
                    out=self.mrope_positions.np,
1835
1836
1837
1838
1839
                    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,
                )
1840
1841
1842

                mrope_pos_ptr += completion_part_len

1843
1844
1845
1846
1847
1848
1849
1850
1851
1852
1853
1854
1855
1856
1857
1858
1859
1860
1861
1862
1863
1864
1865
1866
1867
1868
1869
1870
1871
1872
1873
1874
1875
1876
1877
1878
1879
1880
1881
1882
1883
1884
1885
1886
1887
1888
1889
    def _calc_xdrope_positions(self, scheduler_output: "SchedulerOutput"):
        xdrope_pos_ptr = 0
        for index, req_id in enumerate(self.input_batch.req_ids):
            req = self.requests[req_id]
            assert req.xdrope_positions is not None

            num_computed_tokens = self.input_batch.num_computed_tokens_cpu[index]
            num_scheduled_tokens = scheduler_output.num_scheduled_tokens[req_id]
            num_prompt_tokens = length_from_prompt_token_ids_or_embeds(
                req.prompt_token_ids, req.prompt_embeds
            )

            if num_computed_tokens + num_scheduled_tokens > num_prompt_tokens:
                prompt_part_len = max(0, num_prompt_tokens - num_computed_tokens)
                completion_part_len = max(0, num_scheduled_tokens - prompt_part_len)
            else:
                prompt_part_len = num_scheduled_tokens
                completion_part_len = 0

            assert num_scheduled_tokens == prompt_part_len + completion_part_len

            if prompt_part_len > 0:
                # prompt's xdrope_positions are pre-computed
                dst_start = xdrope_pos_ptr
                dst_end = xdrope_pos_ptr + prompt_part_len
                src_start = num_computed_tokens
                src_end = num_computed_tokens + prompt_part_len

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

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

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

                xdrope_pos_ptr += completion_part_len

1890
1891
    def _calc_spec_decode_metadata(
        self,
1892
1893
1894
1895
1896
1897
1898
1899
1900
1901
1902
1903
1904
1905
1906
1907
        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
1908
1909
1910
1911

        # 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(
1912
1913
            num_sampled_tokens, cumsum_dtype=np.int32
        )
1914
        # Step 2. [0, 0, 0, 0, 103, 104, 104, 104, 206, 207, 207]
1915
        logits_indices = np.repeat(
1916
1917
            cu_num_scheduled_tokens - num_sampled_tokens, num_sampled_tokens
        )
1918
        # Step 3. [0, 1, 2, 3, 103, 104, 105, 106, 206, 207, 208]
1919
1920
1921
1922
1923
1924
        logits_indices += arange

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

        # Compute the draft logits indices.
1925
1926
1927
        # 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(
1928
1929
            num_draft_tokens, cumsum_dtype=np.int32
        )
1930
1931
        # [0, 0, 0, 5, 5, 9]
        target_logits_indices = np.repeat(
1932
1933
            cu_num_sampled_tokens - num_sampled_tokens, num_draft_tokens
        )
1934
1935
1936
1937
1938
        # [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(
1939
1940
            self.device, non_blocking=True
        )
1941
1942
1943
        cu_num_sampled_tokens = torch.from_numpy(cu_num_sampled_tokens).to(
            self.device, non_blocking=True
        )
1944
1945
1946
        logits_indices = torch.from_numpy(logits_indices).to(
            self.device, non_blocking=True
        )
1947
        target_logits_indices = torch.from_numpy(target_logits_indices).to(
1948
1949
            self.device, non_blocking=True
        )
1950
        bonus_logits_indices = torch.from_numpy(bonus_logits_indices).to(
1951
1952
            self.device, non_blocking=True
        )
1953

1954
1955
        # Compute the draft token ids.
        # draft_token_indices:      [  1,   2,   3, 105, 106, 208]
1956
        draft_token_ids = self.input_ids.gpu[logits_indices]
1957
1958
        draft_token_ids = draft_token_ids[target_logits_indices + 1]

1959
        return SpecDecodeMetadata(
1960
1961
1962
            draft_token_ids=draft_token_ids,
            num_draft_tokens=num_draft_tokens.tolist(),
            cu_num_draft_tokens=cu_num_draft_tokens,
1963
            cu_num_sampled_tokens=cu_num_sampled_tokens,
1964
1965
1966
1967
1968
            target_logits_indices=target_logits_indices,
            bonus_logits_indices=bonus_logits_indices,
            logits_indices=logits_indices,
        )

1969
1970
1971
1972
1973
1974
1975
    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
1976
        self.kv_sharing_fast_prefill_logits_indices[:num_logits].copy_(logits_indices)
1977
1978
1979
1980
1981
        # 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_(
1982
1983
1984
1985
1986
1987
            logits_indices[-1].item()
        )
        if (
            self.compilation_config.cudagraph_mode != CUDAGraphMode.NONE
            and num_logits <= self.cudagraph_batch_sizes[-1]
        ):
1988
1989
1990
1991
1992
            # 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
1993
1994
1995
        logits_indices_padded = self.kv_sharing_fast_prefill_logits_indices[
            :num_logits_padded
        ]
1996
1997
        return logits_indices_padded

1998
1999
2000
2001
2002
2003
2004
2005
    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
2006
                inputs.
2007
2008
2009
2010
2011
2012

        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
        """
2013
2014
        scheduled_encoder_inputs = scheduler_output.scheduled_encoder_inputs
        if not scheduled_encoder_inputs:
2015
            return [], []
2016
        # Batch the multi-modal inputs.
2017
        mm_kwargs = list[MultiModalKwargsItem]()
2018
2019
        # list of tuple (mm_hash, position_info)
        mm_hashes_pos = list[tuple[str, PlaceholderRange]]()
2020
2021
        for req_id, encoder_input_ids in scheduled_encoder_inputs.items():
            req_state = self.requests[req_id]
2022
2023

            for mm_input_id in encoder_input_ids:
2024
                mm_feature = req_state.mm_features[mm_input_id]
2025
2026
                if mm_feature.data is None:
                    continue
2027
2028
2029
                mm_hash = mm_feature.identifier
                mm_kwargs.append(mm_feature.data)
                mm_hashes_pos.append((mm_hash, mm_feature.mm_position))
2030

2031
2032
        return mm_kwargs, mm_hashes_pos

2033
2034
2035
    def _execute_mm_encoder(
        self, scheduler_output: "SchedulerOutput"
    ) -> list[torch.Tensor]:
2036
2037
        # Batch the multi-modal inputs using the helper method.
        mm_kwargs, mm_hashes_pos = self._batch_mm_kwargs_from_scheduler(
2038
2039
            scheduler_output
        )
2040
2041

        if not mm_kwargs:
2042
            return []
2043

2044
2045
2046
2047
2048
2049
2050
        # 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.
2051
        model = cast(SupportsMultiModal, self.model)
2052
        encoder_outputs: list[torch.Tensor] = []
2053
        for modality, num_items, mm_kwargs_group in group_mm_kwargs_by_modality(
2054
2055
2056
2057
            mm_kwargs,
            device=self.device,
            pin_memory=self.pin_memory,
            merge_by_field_config=model.merge_by_field_config,
2058
            multimodal_cpu_fields=model.multimodal_cpu_fields,
2059
        ):
2060
            curr_group_outputs: list[torch.Tensor] = []
2061
2062

            # EVS-related change.
2063
            # (ekhvedchenia): Temporary hack to limit peak memory usage when
2064
            # processing multimodal data. This solves the issue with scheduler
2065
2066
2067
2068
            # 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)
2069
2070
2071
2072
2073
2074
2075
2076
2077
2078
2079
2080
2081
2082
2083
2084
            # 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,
2085
                            multimodal_cpu_fields=model.multimodal_cpu_fields,
2086
                        )
2087
                    )
2088

2089
                    micro_batch_outputs = model.embed_multimodal(
2090
2091
                        **micro_batch_mm_inputs
                    )
2092
2093
2094
2095
2096
2097
2098
2099
2100
2101

                    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.
2102
                curr_group_outputs = model.embed_multimodal(**mm_kwargs_group)  # type: ignore[assignment]
2103

2104
2105
            sanity_check_mm_encoder_outputs(
                curr_group_outputs,
2106
                expected_num_items=num_items,
2107
            )
2108
            encoder_outputs.extend(curr_group_outputs)
2109

2110
2111
2112
        # 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(
2113
2114
2115
                output,
                is_embed=pos_info.is_embed,
            )
2116
2117
            logger.debug("Finish execute for mm hash %s", mm_hash)
            self.maybe_save_ec_to_connector(self.encoder_cache, mm_hash)
2118

2119
2120
        return encoder_outputs

2121
    def _gather_mm_embeddings(
2122
2123
        self,
        scheduler_output: "SchedulerOutput",
2124
        shift_computed_tokens: int = 0,
2125
2126
2127
2128
2129
2130
2131
2132
    ) -> 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
2133
        should_sync_mrope_positions = False
2134
        should_sync_xdrope_positions = False
2135

2136
        for req_id in self.input_batch.req_ids:
2137
2138
            mm_embeds_req: list[torch.Tensor] = []

2139
            num_scheduled_tokens = scheduler_output.num_scheduled_tokens[req_id]
2140
            req_state = self.requests[req_id]
2141
            num_computed_tokens = req_state.num_computed_tokens + shift_computed_tokens
2142

2143
2144
            for mm_feature in req_state.mm_features:
                pos_info = mm_feature.mm_position
2145
2146
                start_pos = pos_info.offset
                num_encoder_tokens = pos_info.length
2147
2148
2149
2150
2151
2152
2153
2154
2155
2156
2157
2158
2159
2160
2161
2162

                # 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,
2163
2164
                    num_encoder_tokens,
                )
2165
                assert start_idx < end_idx
2166

2167
                mm_hash = mm_feature.identifier
2168
                encoder_output = self.encoder_cache.get(mm_hash, None)
2169
                assert encoder_output is not None, f"Encoder cache miss for {mm_hash}."
2170
2171
2172
2173

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

2174
                req_start_pos = req_start_idx + start_pos - num_computed_tokens
2175
2176
2177
                is_mm_embed[req_start_pos + start_idx : req_start_pos + end_idx] = (
                    True if is_embed is None else is_embed
                )
2178

2179
2180
2181
2182
                mm_embeds_item = gather_mm_placeholders(
                    encoder_output[start_idx:end_idx],
                    is_embed=is_embed,
                )
2183
2184
2185
                mm_embeds_req.append(mm_embeds_item)

            if self.is_multimodal_pruning_enabled and self.uses_mrope:
2186
                assert req_state.mrope_positions is not None
2187
2188
2189
2190
2191
2192
2193
                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,
2194
2195
                    )
                )
2196
2197
2198
2199
                req_state.mrope_positions.copy_(new_mrope_positions)
                req_state.mrope_position_delta = new_delta

            mm_embeds.extend(mm_embeds_req)
2200
2201
2202
            req_start_idx += num_scheduled_tokens

        is_mm_embed = self.is_mm_embed.copy_to_gpu(total_num_scheduled_tokens)
2203
2204
2205

        if should_sync_mrope_positions:
            self._calc_mrope_positions(scheduler_output)
2206
            self.mrope_positions.copy_to_gpu(total_num_scheduled_tokens)
2207

2208
2209
2210
2211
        if should_sync_xdrope_positions:
            self._calc_xdrope_positions(scheduler_output)
            self.xdrope_positions.copy_to_gpu(total_num_scheduled_tokens)

2212
        return mm_embeds, is_mm_embed
2213

2214
    def get_model(self) -> nn.Module:
2215
        # get raw model out of the cudagraph wrapper.
2216
        if isinstance(self.model, (CUDAGraphWrapper, UBatchWrapper)):
2217
            return self.model.unwrap()
2218
2219
        return self.model

2220
2221
2222
2223
2224
2225
2226
2227
2228
2229
2230
2231
2232
2233
2234
    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

2235
2236
2237
2238
2239
    def get_supported_pooling_tasks(self) -> list[PoolingTask]:
        model = self.get_model()
        if not is_pooling_model(model):
            return []

2240
2241
        supported_tasks = list(model.pooler.get_supported_tasks())

2242
        if self.scheduler_config.enable_chunked_prefill:
2243
2244
2245
2246
            if "token_embed" in supported_tasks:
                supported_tasks.remove("token_embed")
            if "token_classify" in supported_tasks:
                supported_tasks.remove("token_classify")
2247

2248
2249
            logger.debug_once(
                "Chunked prefill is not supported with "
2250
2251
                "token_embed and token_classify tasks "
                "which using ALL pooling. "
2252
2253
2254
                "Please turn off chunked prefill by "
                "`--no-enable-chunked-prefill` before using it."
            )
2255
2256
2257
2258
2259

        if "score" in supported_tasks:
            num_labels = getattr(self.model_config.hf_config, "num_labels", 0)
            if num_labels != 1:
                supported_tasks.remove("score")
2260
                logger.debug_once("Score API is only enabled for num_labels == 1.")
2261
2262

        return supported_tasks
2263

2264
2265
2266
2267
2268
2269
2270
2271
2272
2273
    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)

2274
    def sync_and_slice_intermediate_tensors(
2275
2276
        self,
        num_tokens: int,
2277
        intermediate_tensors: IntermediateTensors | None,
2278
2279
        sync_self: bool,
    ) -> IntermediateTensors:
2280
2281
2282
        assert self.intermediate_tensors is not None

        tp = self.vllm_config.parallel_config.tensor_parallel_size
2283
        is_rs = is_residual_scattered_for_sp(self.vllm_config, num_tokens)
2284
2285
2286
2287
2288
2289

        # 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():
2290
                is_scattered = k == "residual" and is_rs
2291
                copy_len = num_tokens // tp if is_scattered else num_tokens
2292
                self.intermediate_tensors[k][:copy_len].copy_(
2293
2294
2295
2296
2297
2298
2299
2300
2301
2302
2303
2304
2305
                    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:
2306
2307
2308
2309
2310
2311
2312
        """
        Step for the EPLB (Expert Parallelism Load Balancing) state.
        """
        if not self.parallel_config.enable_eplb:
            return

        assert self.eplb_state is not None
2313
2314
        model = self.get_model()
        assert is_mixture_of_experts(model)
2315
2316
2317
        self.eplb_state.step(
            is_dummy,
            is_profile,
2318
            log_stats=self.parallel_config.eplb_config.log_balancedness,
2319
2320
        )

2321
2322
2323
2324
2325
2326
    def _pool(
        self,
        hidden_states: torch.Tensor,
        num_scheduled_tokens: int,
        num_scheduled_tokens_np: np.ndarray,
    ) -> ModelRunnerOutput:
2327
2328
2329
        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"
        )
2330

2331
        hidden_states = hidden_states[:num_scheduled_tokens]
2332
        pooling_metadata = self.input_batch.get_pooling_metadata()
2333
2334
2335
2336
        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]
2337

2338
2339
2340
2341
2342
2343
2344
2345
2346
2347
        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()
2348

2349
        pooler_output: list[torch.Tensor | None] = []
2350
        for raw_output, seq_len, prompt_len in zip(
2351
2352
            raw_pooler_output, seq_lens_cpu, pooling_metadata.prompt_lens
        ):
2353
            output = raw_output if seq_len == prompt_len else None
2354
            pooler_output.append(output)
2355
2356
2357
2358
2359
2360
2361
2362
2363
2364

        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,
        )

2365
    def _pad_for_sequence_parallelism(self, num_scheduled_tokens: int) -> int:
2366
2367
2368
        # Pad tokens to multiple of tensor_parallel_size when
        # enabled collective fusion for SP
        tp_size = self.vllm_config.parallel_config.tensor_parallel_size
2369
2370
2371
2372
        if (
            self.compilation_config.pass_config.enable_sequence_parallelism
            and tp_size > 1
        ):
2373
2374
2375
            return round_up(num_scheduled_tokens, tp_size)
        return num_scheduled_tokens

2376
    def _preprocess(
2377
2378
        self,
        scheduler_output: "SchedulerOutput",
2379
        num_input_tokens: int,  # Padded
2380
        intermediate_tensors: IntermediateTensors | None = None,
2381
    ) -> tuple[
2382
2383
        torch.Tensor | None,
        torch.Tensor | None,
2384
        torch.Tensor,
2385
        IntermediateTensors | None,
2386
        dict[str, Any],
2387
        ECConnectorOutput | None,
2388
    ]:
2389
        num_scheduled_tokens = scheduler_output.total_num_scheduled_tokens
2390
        is_first_rank = get_pp_group().is_first_rank
2391

2392
2393
        # _prepare_inputs may reorder the batch, so we must gather multi
        # modal outputs after that to ensure the correct order
2394
2395
        ec_connector_output = None

2396
2397
        if (
            self.supports_mm_inputs
2398
            and is_first_rank
2399
2400
            and not self.model_config.is_encoder_decoder
        ):
2401
            # Run the multimodal encoder if any.
2402
2403
2404
2405
2406
2407
            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)
2408

2409
2410
2411
            # 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.
2412
            inputs_embeds_scheduled = self.model.embed_input_ids(
2413
2414
2415
                self.input_ids.gpu[:num_scheduled_tokens],
                multimodal_embeddings=mm_embeds,
                is_multimodal=is_mm_embed,
2416
            )
2417

2418
            # TODO(woosuk): Avoid the copy. Optimize.
2419
            self.inputs_embeds.gpu[:num_scheduled_tokens].copy_(inputs_embeds_scheduled)
2420

2421
            input_ids = None
2422
            inputs_embeds = self.inputs_embeds.gpu[:num_input_tokens]
2423
2424
2425
2426
            model_kwargs = {
                **self._init_model_kwargs(num_scheduled_tokens),
                **self._extract_mm_kwargs(scheduler_output),
            }
2427
        elif self.enable_prompt_embeds and is_first_rank:
2428
2429
2430
2431
2432
2433
2434
2435
2436
2437
2438
2439
            # 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).
2440
2441
2442
            token_ids_idx = (
                self.is_token_ids.gpu[:num_scheduled_tokens]
                .nonzero(as_tuple=False)
2443
                .squeeze(1)
2444
            )
2445
2446
2447
            # Some tokens ids may need to become embeds
            if token_ids_idx.numel() > 0:
                token_ids = self.input_ids.gpu[token_ids_idx]
2448
                tokens_to_embeds = self.model.embed_input_ids(input_ids=token_ids)
2449
2450
2451
2452
2453
                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
2454
        else:
2455
2456
2457
2458
            # 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.
2459
            input_ids = self.input_ids.gpu[:num_input_tokens]
2460
            inputs_embeds = None
2461
            model_kwargs = self._init_model_kwargs(num_input_tokens)
2462

2463
        if self.uses_mrope:
2464
            positions = self.mrope_positions.gpu[:, :num_input_tokens]
2465
2466
        elif self.uses_xdrope_dim > 0:
            positions = self.xdrope_positions.gpu[:, :num_input_tokens]
2467
        else:
2468
            positions = self.positions.gpu[:num_input_tokens]
2469

2470
        if is_first_rank:
2471
2472
            intermediate_tensors = None
        else:
2473
            assert intermediate_tensors is not None
2474
            intermediate_tensors = self.sync_and_slice_intermediate_tensors(
2475
2476
                num_input_tokens, intermediate_tensors, True
            )
2477

2478
2479
2480
2481
        if (
            self.model_config.is_encoder_decoder
            and scheduler_output.scheduled_encoder_inputs
        ):
2482
2483
2484
2485
2486
2487
2488
            # 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})
2489

2490
2491
2492
2493
2494
2495
        return (
            input_ids,
            inputs_embeds,
            positions,
            intermediate_tensors,
            model_kwargs,
2496
            ec_connector_output,
2497
        )
2498

2499
    def _sample(
2500
        self,
2501
2502
        logits: torch.Tensor | None,
        spec_decode_metadata: SpecDecodeMetadata | None,
2503
    ) -> SamplerOutput:
2504
        # Sample the next token and get logprobs if needed.
2505
        sampling_metadata = self.input_batch.sampling_metadata
2506
        if spec_decode_metadata is None:
2507
2508
2509
            # 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()
2510
            return self.sampler(
2511
2512
2513
                logits=logits,
                sampling_metadata=sampling_metadata,
            )
2514

2515
        sampler_output = self.rejection_sampler(
2516
2517
            spec_decode_metadata,
            None,  # draft_probs
2518
            logits,
2519
2520
            sampling_metadata,
        )
2521
        self._update_states_after_model_execute(sampler_output.sampled_token_ids)
2522
2523
2524
        return sampler_output

    def _bookkeeping_sync(
2525
2526
2527
        self,
        scheduler_output: "SchedulerOutput",
        sampler_output: SamplerOutput,
2528
        logits: torch.Tensor | None,
2529
2530
        hidden_states: torch.Tensor,
        num_scheduled_tokens: int,
2531
        spec_decode_metadata: SpecDecodeMetadata | None,
2532
    ) -> tuple[
2533
        dict[str, int],
2534
        LogprobsLists | None,
2535
        list[list[int]],
2536
        dict[str, LogprobsTensors | None],
2537
2538
2539
        list[str],
        dict[str, int],
        list[int],
2540
    ]:
2541
2542
2543
2544
        num_nans_in_logits = {}
        if envs.VLLM_COMPUTE_NANS_IN_LOGITS:
            num_nans_in_logits = self._get_nans_in_logits(logits)

2545
2546
2547
2548
        num_reqs = self.input_batch.num_reqs
        discard_sampled_tokens_req_indices = np.nonzero(
            self.discard_request_mask.np[:num_reqs]
        )[0]
2549
2550
2551
2552
        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)
2553

2554
2555
2556
        # 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()
2557
        req_id_to_index_output_copy = self.input_batch.req_id_to_index.copy()
2558
2559

        num_sampled_tokens = sampler_output.sampled_token_ids.shape[0]
2560
        sampled_token_ids = sampler_output.sampled_token_ids
2561
        logprobs_tensors = sampler_output.logprobs_tensors
2562
        invalid_req_indices = []
2563
        cu_num_tokens: list[int] | None = None
2564
2565
2566
2567
2568
2569
        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)
2570
2571
2572
                # Mask out the sampled tokens that should not be sampled.
                for i in discard_sampled_tokens_req_indices:
                    valid_sampled_token_ids[int(i)].clear()
2573
2574
            else:
                # Includes spec decode tokens.
2575
                valid_sampled_token_ids, cu_num_tokens = RejectionSampler.parse_output(
2576
2577
                    sampled_token_ids,
                    self.input_batch.vocab_size,
2578
2579
                    discard_sampled_tokens_req_indices,
                    return_cu_num_tokens=logprobs_tensors is not None,
2580
                )
2581
        else:
2582
            valid_sampled_token_ids = []
2583
            invalid_req_indices = discard_sampled_tokens_req_indices.tolist()
2584
2585
2586
2587
2588
            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.
2589
2590
2591
2592
            # 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
2593
2594
2595
2596
2597
            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
            }
2598

2599
2600
2601
2602
2603
        # 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.
2604
        req_ids = self.input_batch.req_ids
2605
2606
        for req_idx in range(num_sampled_tokens):
            if self.use_async_scheduling:
2607
                sampled_ids = [-1] if req_idx not in invalid_req_indices_set else None
2608
2609
            else:
                sampled_ids = valid_sampled_token_ids[req_idx]
2610

2611
            num_sampled_ids: int = len(sampled_ids) if sampled_ids else 0
2612

2613
            if not sampled_ids:
2614
2615
2616
                continue

            start_idx = self.input_batch.num_tokens_no_spec[req_idx]
2617
            end_idx = start_idx + num_sampled_ids
2618
2619
2620
2621
            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}"
2622
            )
2623

2624
2625
            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
2626
2627
            self.input_batch.num_tokens_no_spec[req_idx] = end_idx
            self.input_batch.num_tokens[req_idx] = end_idx
2628

2629
            req_id = req_ids[req_idx]
2630
2631
2632
            req_state = self.requests[req_id]
            req_state.output_token_ids.extend(sampled_ids)

2633
        logprobs_lists = (
2634
            logprobs_tensors.tolists(cu_num_tokens)
2635
            if not self.use_async_scheduling and logprobs_tensors is not None
2636
2637
2638
2639
2640
2641
2642
2643
2644
            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,
        )

2645
2646
2647
2648
2649
2650
2651
2652
2653
2654
        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,
        )

2655
2656
2657
2658
2659
2660
2661
2662
2663
2664
2665
2666
2667
2668
2669
    @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()

2670
2671
    def _model_forward(
        self,
2672
2673
2674
2675
        input_ids: torch.Tensor | None = None,
        positions: torch.Tensor | None = None,
        intermediate_tensors: IntermediateTensors | None = None,
        inputs_embeds: torch.Tensor | None = None,
2676
2677
2678
2679
2680
        **model_kwargs: dict[str, Any],
    ) -> Any:
        """Helper method to call the model forward pass.

        This method can be overridden by subclasses for model execution.
2681
        Motivation: We can inspect only this method versus
2682
2683
2684
2685
2686
2687
2688
2689
2690
2691
2692
2693
2694
2695
2696
2697
2698
2699
2700
2701
        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,
        )

2702
2703
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
2735
2736
2737
2738
2739
2740
2741
2742
2743
2744
2745
2746
2747
2748
2749
2750
2751
2752
2753
2754
2755
2756
2757
2758
2759
2760
2761
2762
2763
2764
2765
2766
2767
2768
2769
2770
2771
2772
2773
2774
2775
2776
2777
2778
2779
2780
2781
2782
    def _determine_batch_execution_and_padding(
        self,
        num_tokens: int,
        num_reqs: int,
        num_scheduled_tokens_np: np.ndarray,
        max_num_scheduled_tokens: int,
        use_cascade_attn: bool,
        allow_microbatching: bool = True,
        force_eager: bool = False,
        # For cudagraph capture TODO(lucas): Refactor how we capture cudagraphs (will
        # be improved in model runner v2)
        force_uniform_decode: bool | None = None,
        force_has_lora: bool | None = None,
    ) -> tuple[
        CUDAGraphMode, BatchDescriptor, UBatchSlices | None, torch.Tensor | None
    ]:
        num_tokens_padded = self._pad_for_sequence_parallelism(num_tokens)
        uniform_decode = (
            (
                (max_num_scheduled_tokens == self.uniform_decode_query_len)
                and (num_tokens_padded == max_num_scheduled_tokens * num_reqs)
            )
            if force_uniform_decode is None
            else force_uniform_decode
        )

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

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

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

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

            ubatch_slices, num_tokens_across_dp = coordinate_batch_across_dp(
                num_tokens_unpadded=num_tokens_padded,
                parallel_config=self.parallel_config,
                allow_microbatching=allow_microbatching,
                allow_dp_padding=allow_dp_padding,
                num_tokens_padded=num_tokens_padded,
                uniform_decode=uniform_decode,
                num_scheduled_tokens_per_request=num_scheduled_tokens_np,
            )

            # Extract DP padding if there is any
            if num_tokens_across_dp is not None:
                dp_rank = self.parallel_config.data_parallel_rank
                num_tokens_padded = int(num_tokens_across_dp[dp_rank].item())

                # Re-dispatch with DP padding
                cudagraph_mode, batch_descriptor = dispatch_cudagraph(num_tokens_padded)
                # Assert to make sure the agreed upon token count is correct otherwise
                # num_tokens_across_dp will no-longer be valid
                assert batch_descriptor.num_tokens == num_tokens_padded

        return cudagraph_mode, batch_descriptor, ubatch_slices, num_tokens_across_dp

2783
2784
2785
2786
    @torch.inference_mode()
    def execute_model(
        self,
        scheduler_output: "SchedulerOutput",
2787
        intermediate_tensors: IntermediateTensors | None = None,
2788
2789
2790
2791
2792
2793
    ) -> 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."
            )
2794
2795
2796
2797
2798
2799
2800
2801
2802
2803
2804
2805
2806
2807
2808

        # 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)

2809
        num_scheduled_tokens = scheduler_output.total_num_scheduled_tokens
2810
        with record_function_or_nullcontext("gpu_model_runner: preprocess"):
2811
2812
2813
2814
            with self.synchronize_input_prep():
                # Update persistent batch states.
                self._update_states(scheduler_output)

2815
2816
2817
2818
2819
2820
2821
2822
                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)

2823
                if not num_scheduled_tokens:
2824
2825
2826
2827
2828
2829
2830
2831
2832
2833
2834
                    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.
2835
                        self._dummy_run(1)
2836
2837
2838
2839
                    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(
2840
2841
                        scheduler_output, self.vllm_config
                    )
2842
                if self.cache_config.kv_sharing_fast_prefill:
2843
                    assert not self.num_prompt_logprobs, (
2844
2845
                        "--kv-sharing-fast-prefill produces incorrect "
                        "logprobs for prompt tokens, tokens, please disable "
2846
2847
                        "it when the requests need prompt logprobs"
                    )
2848

2849
2850
2851
2852
2853
                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())
2854
                num_tokens_unpadded = scheduler_output.total_num_scheduled_tokens
2855

2856
2857
2858
                (
                    logits_indices,
                    spec_decode_metadata,
2859
                ) = self._prepare_inputs(
2860
2861
                    scheduler_output,
                    num_scheduled_tokens_np,
2862
2863
2864
2865
                )

                cascade_attn_prefix_lens = None
                # Disable cascade attention when using microbatching (DBO)
2866
                if self.cascade_attn_enabled and not self.parallel_config.enable_dbo:
2867
2868
2869
                    # Pre-compute cascade attention prefix lengths
                    cascade_attn_prefix_lens = self._compute_cascade_attn_prefix_lens(
                        num_scheduled_tokens_np,
2870
                        self.input_batch.num_computed_tokens_cpu[:num_reqs],
2871
2872
2873
                        scheduler_output.num_common_prefix_blocks,
                    )

2874
2875
2876
2877
2878
2879
2880
2881
2882
2883
2884
2885
2886
2887
2888
2889
2890
2891
2892
2893
2894
2895
2896
2897
2898
2899
                (
                    cudagraph_mode,
                    batch_desc,
                    ubatch_slices,
                    num_tokens_across_dp,
                ) = self._determine_batch_execution_and_padding(
                    num_tokens=num_tokens_unpadded,
                    num_reqs=num_reqs,
                    num_scheduled_tokens_np=num_scheduled_tokens_np,
                    max_num_scheduled_tokens=max_num_scheduled_tokens,
                    use_cascade_attn=cascade_attn_prefix_lens is not None,
                )

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

                num_tokens_padded = batch_desc.num_tokens
                num_reqs_padded = (
                    batch_desc.num_reqs if batch_desc.num_reqs is not None else num_reqs
                )
2900
2901

                use_spec_decode = len(scheduler_output.scheduled_spec_decode_tokens) > 0
2902
2903
2904
                pad_attn = cudagraph_mode == CUDAGraphMode.FULL

                (attn_metadata, spec_decode_common_attn_metadata) = (
2905
                    self._build_attention_metadata(
2906
2907
                        num_tokens=num_tokens_unpadded,
                        num_tokens_padded=num_tokens_padded if pad_attn else None,
2908
                        num_reqs=num_reqs,
2909
2910
                        num_reqs_padded=num_reqs_padded if pad_attn else None,
                        max_query_len=max_num_scheduled_tokens,
2911
2912
2913
                        ubatch_slices=ubatch_slices,
                        logits_indices=logits_indices,
                        use_spec_decode=use_spec_decode,
2914
                        num_scheduled_tokens=scheduler_output.num_scheduled_tokens,
2915
2916
2917
                        cascade_attn_prefix_lens=cascade_attn_prefix_lens,
                    )
                )
2918

2919
2920
2921
2922
2923
2924
2925
2926
2927
            (
                input_ids,
                inputs_embeds,
                positions,
                intermediate_tensors,
                model_kwargs,
                ec_connector_output,
            ) = self._preprocess(
                scheduler_output, num_tokens_padded, intermediate_tensors
2928
            )
2929

2930
        # Set cudagraph mode to none if calc_kv_scales is true.
2931
2932
2933
        # KV scales calculation involves dynamic operations that are incompatible
        # with CUDA graph capture.
        if self.calculate_kv_scales:
2934
            cudagraph_mode = CUDAGraphMode.NONE
2935
2936
            # Mark KV scales as calculated after the first forward pass
            self.calculate_kv_scales = False
2937

2938
2939
        # Run the model.
        # Use persistent buffers for CUDA graphs.
2940
2941
        with (
            set_forward_context(
2942
2943
                attn_metadata,
                self.vllm_config,
2944
                num_tokens=num_tokens_padded,
2945
                num_tokens_across_dp=num_tokens_across_dp,
2946
2947
                cudagraph_runtime_mode=cudagraph_mode,
                batch_descriptor=batch_desc,
2948
                ubatch_slices=ubatch_slices,
2949
            ),
2950
            record_function_or_nullcontext("gpu_model_runner: forward"),
2951
2952
            self.maybe_get_kv_connector_output(scheduler_output) as kv_connector_output,
        ):
2953
            model_output = self._model_forward(
2954
2955
2956
2957
2958
2959
2960
                input_ids=input_ids,
                positions=positions,
                intermediate_tensors=intermediate_tensors,
                inputs_embeds=inputs_embeds,
                **model_kwargs,
            )

2961
        with record_function_or_nullcontext("gpu_model_runner: postprocess"):
2962
            if self.use_aux_hidden_state_outputs:
2963
                # True when EAGLE 3 is used.
2964
2965
                hidden_states, aux_hidden_states = model_output
            else:
2966
                # Common case.
2967
2968
2969
                hidden_states = model_output
                aux_hidden_states = None

2970
2971
2972
2973
2974
            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)
2975
                    hidden_states.kv_connector_output = kv_connector_output
2976
                    self.kv_connector_output = kv_connector_output
2977
                    return hidden_states
2978

2979
                if self.is_pooling_model:
2980
                    # Return the pooling output.
2981
2982
2983
                    output = self._pool(
                        hidden_states, num_scheduled_tokens, num_scheduled_tokens_np
                    )
2984
2985
                    output.kv_connector_output = kv_connector_output
                    return output
2986
2987

                sample_hidden_states = hidden_states[logits_indices]
2988
                logits = self.model.compute_logits(sample_hidden_states)
2989
2990
2991
2992
            else:
                # Rare case.
                assert not self.is_pooling_model

2993
                sample_hidden_states = hidden_states[logits_indices]
2994
                if not get_pp_group().is_last_rank:
2995
                    all_gather_tensors = {
2996
                        "residual": not is_residual_scattered_for_sp(
2997
                            self.vllm_config, num_tokens_padded
2998
                        )
2999
                    }
3000
                    get_pp_group().send_tensor_dict(
3001
3002
                        hidden_states.tensors,
                        all_gather_group=get_tp_group(),
3003
3004
                        all_gather_tensors=all_gather_tensors,
                    )
3005
3006
                    logits = None
                else:
3007
                    logits = self.model.compute_logits(sample_hidden_states)
3008

3009
                model_output_broadcast_data: dict[str, Any] = {}
3010
3011
3012
                if logits is not None:
                    model_output_broadcast_data["logits"] = logits.contiguous()

3013
                broadcasted = get_pp_group().broadcast_tensor_dict(
3014
3015
                    model_output_broadcast_data, src=len(get_pp_group().ranks) - 1
                )
3016
3017
                assert broadcasted is not None
                logits = broadcasted["logits"]
3018

3019
3020
3021
3022
3023
3024
3025
3026
        self.execute_model_state = ExecuteModelState(
            scheduler_output,
            logits,
            spec_decode_metadata,
            spec_decode_common_attn_metadata,
            hidden_states,
            sample_hidden_states,
            aux_hidden_states,
3027
            ec_connector_output,
3028
        )
3029
        self.kv_connector_output = kv_connector_output
3030
3031
3032
3033
3034
3035
        return None

    @torch.inference_mode
    def sample_tokens(
        self, grammar_output: "GrammarOutput | None"
    ) -> ModelRunnerOutput | AsyncModelRunnerOutput | IntermediateTensors:
3036
3037
3038
        kv_connector_output = self.kv_connector_output
        self.kv_connector_output = None

3039
3040
        if self.execute_model_state is None:
            # Nothing to do (PP non-final rank case), output isn't used.
3041
            if not kv_connector_output:
3042
                return None  # type: ignore[return-value]
3043
3044
3045
3046
3047
3048
3049
3050
3051

            # 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
3052
3053
3054
3055
3056
3057
3058
3059
3060
3061

        # Unpack ephemeral state.
        (
            scheduler_output,
            logits,
            spec_decode_metadata,
            spec_decode_common_attn_metadata,
            hidden_states,
            sample_hidden_states,
            aux_hidden_states,
3062
            ec_connector_output,
3063
3064
3065
3066
3067
3068
3069
3070
3071
        ) = 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
            )
3072

3073
        with record_function_or_nullcontext("gpu_model_runner: sample"):
3074
3075
            sampler_output = self._sample(logits, spec_decode_metadata)

3076
3077
        self.input_batch.prev_sampled_token_ids = None

3078
        def propose_draft_token_ids(sampled_token_ids):
3079
            assert spec_decode_common_attn_metadata is not None
3080
            with record_function_or_nullcontext("gpu_model_runner: draft"):
3081
3082
3083
3084
3085
3086
3087
3088
3089
3090
3091
                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,
                )

3092
        spec_config = self.speculative_config
3093
        use_padded_batch_for_eagle = (
3094
3095
3096
            spec_config is not None
            and spec_config.use_eagle()
            and not spec_config.disable_padded_drafter_batch
3097
        )
3098
3099
3100
        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
3101
        if (
3102
3103
3104
            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
3105
        ):
3106
            effective_drafter_max_model_len = (
3107
                spec_config.draft_model_config.max_model_len
3108
            )
3109
        input_fits_in_drafter = spec_decode_common_attn_metadata and (
3110
            spec_decode_common_attn_metadata.max_seq_len + self.num_spec_tokens
3111
3112
            <= effective_drafter_max_model_len
        )
3113
        if use_padded_batch_for_eagle:
3114
3115
            assert self.speculative_config is not None
            assert isinstance(self.drafter, EagleProposer)
3116
3117
3118
3119
3120
3121
            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:
3122
                assert spec_decode_common_attn_metadata is not None
3123
3124
3125
3126
3127
3128
                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,
3129
                        self.discard_request_mask.gpu,
3130
3131
3132
3133
3134
                    )
                )
                self._copy_valid_sampled_token_count(
                    next_token_ids, valid_sampled_tokens_count
                )
3135

3136
        with record_function_or_nullcontext("gpu_model_runner: bookkeep"):
3137
3138
3139
3140
3141
3142
3143
3144
            (
                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,
3145
3146
3147
3148
3149
            ) = self._bookkeeping_sync(
                scheduler_output,
                sampler_output,
                logits,
                hidden_states,
3150
                scheduler_output.total_num_scheduled_tokens,
3151
                spec_decode_metadata,
3152
            )
3153

3154
3155
3156
3157
3158
        if (
            self.speculative_config
            and not use_padded_batch_for_eagle
            and input_fits_in_drafter
        ):
3159
3160
3161
            # 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)
3162

3163
        with record_function_or_nullcontext("gpu_model_runner: eplb"):
3164
            self.eplb_step()
3165
3166
3167
3168
3169
3170
3171
3172
3173
        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,
3174
3175
3176
                ec_connector_output=ec_connector_output
                if self.supports_mm_inputs
                else None,
3177
3178
                num_nans_in_logits=num_nans_in_logits,
            )
3179

3180
3181
        if not self.use_async_scheduling:
            return output
3182
3183
3184
3185
3186
3187
3188
3189
3190
        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,
3191
                vocab_size=self.input_batch.vocab_size,
3192
3193
3194
3195
3196
            )
        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
3197
            # any requests with sampling params that require output ids.
3198
3199
3200
3201
            self.input_batch.set_async_sampled_token_ids(
                async_output.sampled_token_ids_cpu,
                async_output.async_copy_ready_event,
            )
3202
3203
3204

        return async_output

3205
    def take_draft_token_ids(self) -> DraftTokenIds | None:
3206
3207
3208
3209
3210
3211
3212
3213
3214
3215
        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)

3216
3217
3218
3219
3220
3221
3222
3223
3224
3225
3226
3227
3228
3229
3230
3231
3232
3233
3234
3235
3236
3237
3238
3239
3240
3241
3242
3243
3244
3245
3246
    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()

3247
3248
3249
    def propose_draft_token_ids(
        self,
        scheduler_output: "SchedulerOutput",
3250
        sampled_token_ids: torch.Tensor | list[list[int]],
3251
3252
3253
        sampling_metadata: SamplingMetadata,
        hidden_states: torch.Tensor,
        sample_hidden_states: torch.Tensor,
3254
3255
        aux_hidden_states: list[torch.Tensor] | None,
        spec_decode_metadata: SpecDecodeMetadata | None,
3256
        common_attn_metadata: CommonAttentionMetadata,
3257
    ) -> list[list[int]] | torch.Tensor:
3258
        num_scheduled_tokens = scheduler_output.total_num_scheduled_tokens
3259
3260
3261
        spec_config = self.speculative_config
        assert spec_config is not None
        if spec_config.method == "ngram":
3262
            assert isinstance(sampled_token_ids, list)
3263
            assert isinstance(self.drafter, NgramProposer)
3264
            draft_token_ids = self.drafter.propose(
3265
3266
                sampled_token_ids,
                self.input_batch.req_ids,
3267
3268
                self.input_batch.num_tokens_no_spec,
                self.input_batch.token_ids_cpu,
3269
3270
                self.input_batch.spec_decode_unsupported_reqs,
            )
3271
        elif spec_config.method == "suffix":
3272
3273
3274
            assert isinstance(sampled_token_ids, list)
            assert isinstance(self.drafter, SuffixDecodingProposer)
            draft_token_ids = self.drafter.propose(self.input_batch, sampled_token_ids)
3275
        elif spec_config.method == "medusa":
3276
            assert isinstance(sampled_token_ids, list)
3277
            assert isinstance(self.drafter, MedusaProposer)
3278

3279
3280
            if sample_hidden_states.shape[0] == len(sampled_token_ids):
                # The input to the target model does not include draft tokens.
3281
3282
3283
3284
                hidden_states = sample_hidden_states
            else:
                indices = []
                offset = 0
3285
3286
3287
                assert spec_decode_metadata is not None, (
                    "No spec decode metadata for medusa"
                )
3288
                for num_draft, tokens in zip(
3289
3290
                    spec_decode_metadata.num_draft_tokens, sampled_token_ids
                ):
3291
                    indices.append(offset + len(tokens) - 1)
3292
                    offset += num_draft + 1
3293
                indices = torch.tensor(indices, device=self.device)
3294
3295
                hidden_states = sample_hidden_states[indices]

3296
            draft_token_ids = self.drafter.propose(
3297
3298
3299
                target_hidden_states=hidden_states,
                sampling_metadata=sampling_metadata,
            )
3300
        elif spec_config.use_eagle():
3301
            assert isinstance(self.drafter, EagleProposer)
3302

3303
            if spec_config.disable_padded_drafter_batch:
3304
3305
3306
                # 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.
3307
3308
                assert isinstance(sampled_token_ids, list), (
                    "sampled_token_ids should be a python list when"
3309
                    "padded-batch is disabled."
3310
                )
3311
                next_token_ids = self.drafter.prepare_next_token_ids_cpu(
3312
3313
3314
3315
3316
                    sampled_token_ids,
                    self.requests,
                    self.input_batch,
                    scheduler_output.num_scheduled_tokens,
                )
3317
3318
3319
3320
3321
            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.
3322
3323
                assert isinstance(sampled_token_ids, torch.Tensor), (
                    "sampled_token_ids should be a torch.Tensor when"
3324
                    "padded-batch is enabled."
3325
3326
                )
                next_token_ids, valid_sampled_tokens_count = (
3327
3328
3329
3330
3331
                    self.drafter.prepare_next_token_ids_padded(
                        common_attn_metadata,
                        sampled_token_ids,
                        self.requests,
                        self.input_batch,
3332
                        self.discard_request_mask.gpu,
3333
                    )
3334
                )
3335
3336
3337
                self._copy_valid_sampled_token_count(
                    next_token_ids, valid_sampled_tokens_count
                )
Jiayi Yao's avatar
Jiayi Yao committed
3338

3339
            if spec_decode_metadata is None:
3340
                token_indices_to_sample = None
3341
                # input_ids can be None for multimodal models.
3342
                target_token_ids = self.input_ids.gpu[:num_scheduled_tokens]
3343
                target_positions = self._get_positions(num_scheduled_tokens)
3344
                if self.use_aux_hidden_state_outputs:
Wentao Ye's avatar
Wentao Ye committed
3345
                    assert aux_hidden_states is not None
3346
                    target_hidden_states = torch.cat(
3347
3348
                        [h[:num_scheduled_tokens] for h in aux_hidden_states], dim=-1
                    )
3349
3350
                else:
                    target_hidden_states = hidden_states[:num_scheduled_tokens]
3351
            else:
3352
                if spec_config.disable_padded_drafter_batch:
3353
                    token_indices_to_sample = None
3354
3355
3356
3357
3358
                    common_attn_metadata, token_indices = self.drafter.prepare_inputs(
                        common_attn_metadata,
                        sampled_token_ids,
                        spec_decode_metadata.num_draft_tokens,
                    )
3359
3360
3361
3362
3363
3364
3365
3366
3367
                    target_token_ids = self.input_ids.gpu[token_indices]
                    target_positions = self._get_positions(token_indices)
                    if self.use_aux_hidden_state_outputs:
                        assert aux_hidden_states is not None
                        target_hidden_states = torch.cat(
                            [h[token_indices] for h in aux_hidden_states], dim=-1
                        )
                    else:
                        target_hidden_states = hidden_states[token_indices]
3368
                else:
3369
                    common_attn_metadata, token_indices_to_sample = (
3370
3371
3372
                        self.drafter.prepare_inputs_padded(
                            common_attn_metadata,
                            spec_decode_metadata,
3373
3374
3375
                            valid_sampled_tokens_count,
                        )
                    )
3376
3377
3378
3379
3380
3381
3382
3383
3384
3385
3386
                    total_num_tokens = common_attn_metadata.num_actual_tokens
                    # When padding the batch, token_indices is just a range
                    target_token_ids = self.input_ids.gpu[:total_num_tokens]
                    target_positions = self._get_positions(total_num_tokens)
                    if self.use_aux_hidden_state_outputs:
                        assert aux_hidden_states is not None
                        target_hidden_states = torch.cat(
                            [h[:total_num_tokens] for h in aux_hidden_states], dim=-1
                        )
                    else:
                        target_hidden_states = hidden_states[:total_num_tokens]
3387

3388
            if self.supports_mm_inputs:
3389
3390
3391
3392
3393
3394
                mm_embed_inputs = self._gather_mm_embeddings(
                    scheduler_output,
                    shift_computed_tokens=1,
                )
            else:
                mm_embed_inputs = None
3395

3396
            draft_token_ids = self.drafter.propose(
3397
3398
3399
3400
                target_token_ids=target_token_ids,
                target_positions=target_positions,
                target_hidden_states=target_hidden_states,
                next_token_ids=next_token_ids,
3401
                last_token_indices=token_indices_to_sample,
3402
                sampling_metadata=sampling_metadata,
3403
                common_attn_metadata=common_attn_metadata,
3404
                mm_embed_inputs=mm_embed_inputs,
3405
            )
3406

3407
        return draft_token_ids
3408

3409
3410
3411
    def update_config(self, overrides: dict[str, Any]) -> None:
        allowed_config_names = {"load_config", "model_config"}
        for config_name, config_overrides in overrides.items():
3412
3413
            assert config_name in allowed_config_names, (
                f"Config `{config_name}` not supported. "
3414
                f"Allowed configs: {allowed_config_names}"
3415
            )
3416
3417
3418
3419
            config = getattr(self, config_name)
            new_config = update_config(config, config_overrides)
            setattr(self, config_name, new_config)

3420
3421
3422
3423
3424
    def load_model(self, eep_scale_up: bool = False) -> None:
        """
        Args:
            eep_scale_up: the model loading is for elastic EP scale up.
        """
3425
3426
3427
3428
3429
        logger.info_once(
            "Starting to load model %s...",
            self.model_config.model,
            scope="global",
        )
3430
3431
3432
3433
3434
        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)
        )
3435

3436
3437
3438
        if self.parallel_config.enable_eplb:
            self.eplb_state = EplbState(self.parallel_config, self.device)
            eplb_models = 0
3439
        with DeviceMemoryProfiler() as m:
3440
            time_before_load = time.perf_counter()
3441
            model_loader = get_model_loader(self.load_config)
3442
            self.model = model_loader.load_model(
3443
3444
                vllm_config=self.vllm_config, model_config=self.model_config
            )
3445
            if self.lora_config:
3446
3447
3448
                self.model = self.load_lora_model(
                    self.model, self.vllm_config, self.device
                )
3449
            if hasattr(self, "drafter"):
3450
                logger.info_once("Loading drafter model...")
3451
                self.drafter.load_model(self.model)
3452
3453
3454
3455
3456
                if (
                    hasattr(self.drafter, "model")
                    and is_mixture_of_experts(self.drafter.model)
                    and self.parallel_config.enable_eplb
                ):
3457
3458
3459
                    spec_config = self.vllm_config.speculative_config
                    assert spec_config is not None
                    assert spec_config.draft_model_config is not None
3460
3461
                    logger.info_once(
                        "EPLB is enabled for drafter model %s.",
3462
                        spec_config.draft_model_config.model,
3463
3464
3465
3466
3467
3468
3469
3470
3471
3472
3473
3474
3475
3476
3477
3478
                    )

                    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,
3479
                        spec_config.draft_model_config,
3480
3481
3482
3483
3484
3485
                        global_expert_load,
                        old_global_expert_indices,
                        rank_mapping,
                    )
                    eplb_models += 1

3486
            if self.use_aux_hidden_state_outputs:
3487
                if not supports_eagle3(self.get_model()):
3488
3489
                    raise RuntimeError(
                        "Model does not support EAGLE3 interface but "
3490
3491
                        "aux_hidden_state_outputs was requested"
                    )
3492
3493
3494
3495
3496
3497
3498
3499
3500
3501
3502
3503
3504

                # 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)
3505
            time_after_load = time.perf_counter()
3506
        self.model_memory_usage = m.consumed_memory
3507
        logger.info_once(
3508
            "Model loading took %.4f GiB memory and %.6f seconds",
3509
3510
            self.model_memory_usage / GiB_bytes,
            time_after_load - time_before_load,
3511
            scope="local",
3512
        )
3513
        prepare_communication_buffer_for_model(self.model)
3514
3515
3516
3517
        if (drafter := getattr(self, "drafter", None)) and (
            drafter_model := getattr(drafter, "model", None)
        ):
            prepare_communication_buffer_for_model(drafter_model)
3518
        mm_config = self.model_config.multimodal_config
3519
        self.is_multimodal_pruning_enabled = (
3520
            supports_multimodal_pruning(self.get_model())
3521
3522
            and mm_config is not None
            and mm_config.is_multimodal_pruning_enabled()
3523
        )
3524

3525
        if is_mixture_of_experts(self.model) and self.parallel_config.enable_eplb:
3526
3527
3528
3529
3530
3531
3532
3533
3534
3535
3536
            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(
3537
                self.model,
3538
                self.model_config,
3539
3540
3541
                global_expert_load,
                old_global_expert_indices,
                rank_mapping,
3542
            )
3543
3544
            if self.eplb_state.is_async:
                self.eplb_state.start_async_loop(rank_mapping=rank_mapping)
3545

3546
        if (
3547
3548
            self.vllm_config.compilation_config.mode
            == CompilationMode.STOCK_TORCH_COMPILE
3549
            and supports_dynamo()
3550
        ):
3551
            backend = self.vllm_config.compilation_config.init_backend(self.vllm_config)
3552
            compilation_counter.stock_torch_compile_count += 1
3553
            self.model.compile(fullgraph=True, backend=backend)
3554
            return
3555
        # for other compilation modes, cudagraph behavior is controlled by
3556
3557
3558
        # CudagraphWraper and CudagraphDispatcher of vllm.

        # wrap the model with full cudagraph wrapper if needed.
3559
3560
3561
        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:
3562
3563
3564
            self.model = CUDAGraphWrapper(
                self.model, self.vllm_config, runtime_mode=CUDAGraphMode.FULL
            )
3565
        elif self.parallel_config.enable_dbo:
3566
            if cudagraph_mode.has_full_cudagraphs():
3567
3568
3569
                self.model = UBatchWrapper(
                    self.model, self.vllm_config, CUDAGraphMode.FULL, self.device
                )
3570
            else:
3571
3572
3573
                self.model = UBatchWrapper(
                    self.model, self.vllm_config, CUDAGraphMode.NONE, self.device
                )
3574

3575
    def _get_eagle3_aux_layers_from_config(self) -> tuple[int, ...] | None:
3576
3577
3578
3579
3580
3581
3582
3583
3584
3585
3586
3587
3588
3589
3590
3591
3592
3593
3594
3595
3596
3597
3598
        """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

3599
    def reload_weights(self) -> None:
3600
        assert getattr(self, "model", None) is not None, (
3601
            "Cannot reload weights before model is loaded."
3602
        )
3603
3604
        model_loader = get_model_loader(self.load_config)
        logger.info("Reloading weights inplace...")
3605
        model_loader.load_weights(self.get_model(), model_config=self.model_config)
3606

3607
3608
3609
3610
3611
    def save_tensorized_model(
        self,
        tensorizer_config: "TensorizerConfig",
    ) -> None:
        TensorizerLoader.save_model(
3612
            self.get_model(),
3613
            tensorizer_config=tensorizer_config,
3614
            model_config=self.model_config,
3615
3616
        )

3617
3618
3619
    def _get_prompt_logprobs_dict(
        self,
        hidden_states: torch.Tensor,
3620
        num_scheduled_tokens: dict[str, int],
3621
    ) -> dict[str, LogprobsTensors | None]:
3622
        num_prompt_logprobs_dict = self.num_prompt_logprobs
3623
3624
3625
        if not num_prompt_logprobs_dict:
            return {}

3626
        in_progress_dict = self.input_batch.in_progress_prompt_logprobs_cpu
3627
        prompt_logprobs_dict: dict[str, LogprobsTensors | None] = {}
3628
3629
3630
3631
3632

        # 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():
3633
3634
3635
3636
            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
3637
3638
3639

            # Get metadata for this request.
            request = self.requests[req_id]
3640
3641
3642
3643
            if request.prompt_token_ids is None:
                # Prompt logprobs is incompatible with prompt embeddings
                continue

3644
3645
            num_prompt_tokens = len(request.prompt_token_ids)
            prompt_token_ids = torch.tensor(request.prompt_token_ids).to(
3646
3647
                self.device, non_blocking=True
            )
3648

3649
3650
3651
3652
3653
3654
            # 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(
3655
3656
                    num_prompt_tokens - 1, num_prompt_logprobs + 1
                )
3657
3658
                in_progress_dict[req_id] = logprobs_tensors

3659
            # Determine number of logits to retrieve.
3660
3661
            start_idx = request.num_computed_tokens
            start_tok = start_idx + 1
3662
            num_remaining_tokens = num_prompt_tokens - start_tok
3663
            if num_tokens <= num_remaining_tokens:
3664
                # This is a chunk, more tokens remain.
3665
3666
3667
                # 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.
3668
3669
3670
3671
3672
                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)
3673
3674
3675
3676
3677
3678
3679
                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
3680
3681
3682
3683
3684

            # 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]
3685
            offset = self.query_start_loc.np[req_idx].item()
3686
            prompt_hidden_states = hidden_states[offset : offset + num_logits]
3687
            logits = self.model.compute_logits(prompt_hidden_states)
3688
3689
3690
3691

            # 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.
3692
            tgt_token_ids = prompt_token_ids[start_tok : start_tok + num_logits]
3693
3694

            # Compute prompt logprobs.
3695
3696
            logprobs = self.sampler.compute_logprobs(logits)
            token_ids, logprobs, ranks = self.sampler.gather_logprobs(
3697
3698
                logprobs, num_prompt_logprobs, tgt_token_ids
            )
3699
3700

            # Transfer GPU->CPU async.
3701
3702
            chunk_slice = slice(start_idx, start_idx + num_logits)
            logprobs_tensors.logprob_token_ids[chunk_slice].copy_(
3703
3704
3705
                token_ids, non_blocking=True
            )
            logprobs_tensors.logprobs[chunk_slice].copy_(logprobs, non_blocking=True)
3706
            logprobs_tensors.selected_token_ranks[chunk_slice].copy_(
3707
3708
                ranks, non_blocking=True
            )
3709
3710
3711
3712
3713

        # 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]
3714
            del in_progress_dict[req_id]
3715
3716

        # Must synchronize the non-blocking GPU->CPU transfers.
3717
        if prompt_logprobs_dict:
3718
            self._sync_device()
3719
3720
3721

        return prompt_logprobs_dict

3722
3723
    def _get_nans_in_logits(
        self,
3724
        logits: torch.Tensor | None,
3725
3726
3727
3728
3729
3730
3731
3732
3733
3734
3735
    ) -> 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])
3736
3737
3738
                    if num_nans_for_index is not None and req_index < logits.shape[0]
                    else 0
                )
3739
3740
3741
3742
            return num_nans_in_logits
        except IndexError:
            return {}

3743
3744
3745
3746
3747
3748
    @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
3749
         - during DP rank dummy run
3750
3751
3752
3753
3754
3755
3756
3757
3758
3759
3760
        """
        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(
3761
                    self.input_ids.gpu,
3762
3763
                    low=0,
                    high=self.model_config.get_vocab_size(),
3764
3765
                    dtype=input_ids.dtype,
                )
3766

3767
            logger.debug_once("Randomizing dummy data for DP Rank")
3768
            input_ids.copy_(rand_input_ids()[: input_ids.size(0)], non_blocking=True)
3769
3770
3771
            yield
            input_ids.fill_(0)

3772
3773
3774
3775
3776
3777
    def _get_mm_dummy_batch(
        self,
        modality: str,
        max_items_per_batch: int,
    ) -> BatchedTensorInputs:
        """Dummy data for profiling and precompiling multimodal models."""
3778
3779
        assert self.mm_budget is not None

3780
3781
        dummy_decoder_data = self.mm_registry.get_decoder_dummy_data(
            model_config=self.model_config,
3782
            seq_len=self.max_model_len,
3783
            mm_counts={modality: 1},
3784
            cache=self.mm_budget.cache,
3785
3786
3787
3788
        )
        dummy_mm_data = dummy_decoder_data.multi_modal_data

        # Result in the maximum GPU consumption of the model
3789
3790
        dummy_mm_item = dummy_mm_data[modality][0]
        dummy_mm_items = [dummy_mm_item] * max_items_per_batch
3791

3792
        model = cast(SupportsMultiModal, self.model)
3793
3794
3795
3796
3797
3798
3799
        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,
3800
                multimodal_cpu_fields=model.multimodal_cpu_fields,
3801
3802
            )
        )
3803

3804
3805
3806
3807
    @torch.inference_mode()
    def _dummy_run(
        self,
        num_tokens: int,
3808
        cudagraph_runtime_mode: CUDAGraphMode | None = None,
3809
3810
        force_attention: bool = False,
        uniform_decode: bool = False,
3811
        allow_microbatching: bool = True,
3812
3813
        skip_eplb: bool = False,
        is_profile: bool = False,
3814
        create_mixed_batch: bool = False,
3815
        remove_lora: bool = True,
3816
        activate_lora: bool = False,
Rémi Delacourt's avatar
Rémi Delacourt committed
3817
        is_graph_capturing: bool = False,
3818
    ) -> tuple[torch.Tensor, torch.Tensor]:
3819
3820
3821
3822
3823
3824
3825
        """
        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.
3826
                - if not set will determine the cudagraph mode based on using
3827
                    the self.cudagraph_dispatcher.
3828
3829
3830
3831
                - CUDAGraphMode.NONE: No cudagraph, for warm up and profile run
                - CUDAGraphMode.PIECEWISE: Piecewise cudagraph.
                - CUDAGraphMode.FULL: Full cudagraph, attention metadata is
                    needed.
3832
            force_attention: If True, always create attention metadata. Used to
3833
3834
3835
3836
                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.
3837
3838
            create_mixed_batch: If True, create a mixed batch with both decode
                (1 token) and prefill (multiple tokens) requests.
3839
            remove_lora: If False, dummy LoRAs are not destroyed after the run
3840
            activate_lora: If False, dummy_run is performed without LoRAs.
3841
        """
3842
3843
3844
3845
        assert (
            cudagraph_runtime_mode is None
            or cudagraph_runtime_mode.valid_runtime_modes()
        )
3846

3847
        # If cudagraph_mode.decode_mode() == FULL and
3848
        # cudagraph_mode.separate_routine(). This means that we are using
3849
3850
3851
3852
3853
3854
3855
3856
3857
3858
3859
        # 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.
3860
        max_query_len = self.uniform_decode_query_len if uniform_decode else num_tokens
3861

3862
3863
3864
3865
3866
        # 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
3867
3868
3869
3870
        if create_mixed_batch:
            assert not uniform_decode
            # Create mixed batch:
            # first half decode tokens, second half one prefill
3871
            num_decode_tokens = min(max_num_reqs - 1, num_tokens // 2)
3872
3873
3874
3875
            num_prefill_tokens = num_tokens - num_decode_tokens
            num_reqs = num_decode_tokens + 1

            # Create decode requests (1 token each) followed by prefill request
3876
            num_scheduled_tokens_list = [1] * num_decode_tokens + [num_prefill_tokens]
3877
3878
3879
            # Note: Overriding max_query_len to be the prefill tokens
            max_query_len = num_prefill_tokens
        elif uniform_decode:
3880
            assert not create_mixed_batch
3881
            num_reqs = min(max_num_reqs, cdiv(num_tokens, max_query_len))
3882
3883
            num_scheduled_tokens_list = [max_query_len] * num_reqs
            if num_tokens % max_query_len != 0:
3884
                num_scheduled_tokens_list[-1] = num_tokens % max_query_len
3885
3886
3887
3888
3889
3890
        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

3891
3892
        assert sum(num_scheduled_tokens_list) == num_tokens
        assert len(num_scheduled_tokens_list) == num_reqs
3893
        num_scheduled_tokens = np.array(num_scheduled_tokens_list, dtype=np.int32)
3894
3895
        num_tokens_unpadded = int(num_scheduled_tokens.sum())

3896
        num_sampled_tokens = np.ones(num_reqs, dtype=np.int32)
3897

3898
3899
3900
3901
3902
3903
3904
3905
3906
3907
3908
3909
3910
3911
3912
3913
3914
3915
3916
        _cudagraph_mode, batch_desc, ubatch_slices, num_tokens_across_dp = (
            self._determine_batch_execution_and_padding(
                num_tokens=num_tokens_unpadded,
                num_reqs=num_reqs,
                num_scheduled_tokens_np=num_scheduled_tokens,
                max_num_scheduled_tokens=max_query_len,
                use_cascade_attn=False,
                allow_microbatching=allow_microbatching,
                force_eager=is_profile
                or (cudagraph_runtime_mode == CUDAGraphMode.NONE),
                # `force_uniform_decode` is used for cudagraph capture; because for
                # capturing mixed prefill-decode batches, we sometimes use
                # num_tokens == num_reqs which looks like a uniform decode batch to the
                # dispatcher; but we actually want to capture a piecewise cudagraph
                force_uniform_decode=uniform_decode,
                # `force_has_lora` is used for cudagraph capture; because LoRA is
                # activated later in the context manager, but we need to know the
                # LoRA state when determining the batch descriptor for capture
                force_has_lora=activate_lora,
3917
3918
            )
        )
3919
3920
3921

        if cudagraph_runtime_mode is None:
            cudagraph_runtime_mode = _cudagraph_mode
3922
        else:
3923
3924
3925
3926
3927
3928
3929
3930
3931
            assert cudagraph_runtime_mode == _cudagraph_mode, (
                f"Cudagraph runtime mode mismatch in dummy_run. "
                f"Expected {_cudagraph_mode}, but got {cudagraph_runtime_mode}."
            )

        num_tokens_padded = batch_desc.num_tokens
        num_reqs_padded = (
            batch_desc.num_reqs if batch_desc.num_reqs is not None else num_reqs
        )
3932

3933
        attn_metadata: PerLayerAttnMetadata | None = None
3934
3935
3936

        # If force_attention is True, we always capture attention. Otherwise,
        # it only happens for cudagraph_runtime_mode=FULL.
3937
        if force_attention or cudagraph_runtime_mode == CUDAGraphMode.FULL:
3938
3939
3940
3941
3942
3943
            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:
3944
                seq_lens = max_query_len  # type: ignore[assignment]
3945
            self.seq_lens.np[:num_reqs] = seq_lens
3946
3947
            self.seq_lens.np[num_reqs:] = 0
            self.seq_lens.copy_to_gpu()
3948

3949
3950
            cum_num_tokens, _ = self._get_cumsum_and_arange(num_scheduled_tokens)
            self.query_start_loc.np[1 : num_reqs + 1] = cum_num_tokens
3951
3952
            self.query_start_loc.copy_to_gpu()

3953
            attn_metadata, _ = self._build_attention_metadata(
3954
3955
3956
                num_tokens=num_tokens_unpadded,
                num_reqs=num_reqs_padded,
                max_query_len=max_query_len,
3957
3958
3959
                ubatch_slices=ubatch_slices,
                for_cudagraph_capture=True,
            )
3960

3961
        with self.maybe_dummy_run_with_lora(
3962
3963
3964
3965
3966
            self.lora_config,
            num_scheduled_tokens,
            num_sampled_tokens,
            activate_lora,
            remove_lora,
3967
        ):
3968
            # Make sure padding doesn't exceed max_num_tokens
3969
3970
            assert num_tokens_padded <= self.max_num_tokens
            model_kwargs = self._init_model_kwargs(num_tokens_padded)
3971
            if self.supports_mm_inputs and not self.model_config.is_encoder_decoder:
3972
                input_ids = None
3973
                inputs_embeds = self.inputs_embeds.gpu[:num_tokens_padded]
3974
                model_kwargs = {
3975
                    **model_kwargs,
3976
3977
                    **self._dummy_mm_kwargs(num_reqs),
                }
3978
3979
            elif self.enable_prompt_embeds:
                input_ids = None
3980
3981
                inputs_embeds = self.inputs_embeds.gpu[:num_tokens_padded]
                model_kwargs = self._init_model_kwargs(num_tokens_padded)
3982
            else:
3983
                input_ids = self.input_ids.gpu[:num_tokens_padded]
3984
                inputs_embeds = None
3985

3986
            if self.uses_mrope:
3987
                positions = self.mrope_positions.gpu[:, :num_tokens_padded]
3988
            elif self.uses_xdrope_dim > 0:
3989
                positions = self.xdrope_positions.gpu[:, :num_tokens_padded]
3990
            else:
3991
                positions = self.positions.gpu[:num_tokens_padded]
3992
3993
3994
3995
3996
3997
3998
3999
4000

            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,
4001
4002
4003
                            device=self.device,
                        )
                    )
4004
4005

                intermediate_tensors = self.sync_and_slice_intermediate_tensors(
4006
                    num_tokens_padded, None, False
4007
                )
4008

4009
            if ubatch_slices is not None:
4010
4011
4012
                # Adjust values to reflect a single ubatch.
                # TODO(sage,lucas): this is cruft that should be addressed in
                #  the padding refactor.
4013
                num_tokens_padded = ubatch_slices[0].num_tokens
4014
                if num_tokens_across_dp is not None:
4015
                    num_tokens_across_dp[:] = num_tokens_padded
4016

4017
4018
4019
            with (
                self.maybe_randomize_inputs(input_ids),
                set_forward_context(
4020
4021
                    attn_metadata,
                    self.vllm_config,
4022
                    num_tokens=num_tokens_padded,
4023
4024
                    num_tokens_across_dp=num_tokens_across_dp,
                    cudagraph_runtime_mode=cudagraph_runtime_mode,
4025
                    batch_descriptor=batch_desc,
4026
4027
4028
                    ubatch_slices=ubatch_slices,
                ),
            ):
4029
                outputs = self.model(
4030
4031
4032
4033
                    input_ids=input_ids,
                    positions=positions,
                    intermediate_tensors=intermediate_tensors,
                    inputs_embeds=inputs_embeds,
4034
                    **model_kwargs,
4035
                )
4036

4037
4038
4039
4040
            if self.use_aux_hidden_state_outputs:
                hidden_states, _ = outputs
            else:
                hidden_states = outputs
4041

4042
            if self.speculative_config and self.speculative_config.use_eagle():
4043
                assert isinstance(self.drafter, EagleProposer)
4044
                use_cudagraphs = (
Rémi Delacourt's avatar
Rémi Delacourt committed
4045
                    cudagraph_runtime_mode.has_mode(CUDAGraphMode.PIECEWISE)
4046
4047
                    and not self.speculative_config.enforce_eager
                )
4048
4049
4050
4051
4052
4053
4054
4055
4056
4057
4058

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

                self.drafter.dummy_run(
                    num_tokens,
                    use_cudagraphs=use_cudagraphs,
Rémi Delacourt's avatar
Rémi Delacourt committed
4059
                    is_graph_capturing=is_graph_capturing,
4060
                )
4061

4062
4063
4064
4065
4066
4067
4068
4069
4070
4071
        # 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)

4072
        logit_indices = np.cumsum(num_scheduled_tokens) - 1
4073
4074
4075
4076
        logit_indices_device = torch.from_numpy(logit_indices).to(
            self.device, non_blocking=True
        )
        return hidden_states, hidden_states[logit_indices_device]
4077
4078
4079
4080
4081
4082

    @torch.inference_mode()
    def _dummy_sampler_run(
        self,
        hidden_states: torch.Tensor,
    ) -> torch.Tensor:
4083
4084
4085
4086
        # 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)
4087

4088
        logits = self.model.compute_logits(hidden_states)
4089
4090
        num_reqs = logits.size(0)

4091
        dummy_tensors = lambda v: torch.full((num_reqs,), v, device=self.device)
4092
4093
4094
4095
4096
4097
4098
4099
4100
4101
4102
4103
4104
4105
4106

        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)],
4107
            spec_token_ids=[[] for _ in range(num_reqs)],
4108
4109
            allowed_token_ids_mask=None,
            bad_words_token_ids={},
4110
            logitsprocs=LogitsProcessors(),
4111
        )
4112
        try:
4113
4114
4115
            sampler_output = self.sampler(
                logits=logits, sampling_metadata=dummy_metadata
            )
4116
        except RuntimeError as e:
4117
            if "out of memory" in str(e):
4118
4119
4120
4121
                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 "
4122
4123
                    "initializing the engine."
                ) from e
4124
4125
            else:
                raise e
4126
        if self.speculative_config:
4127
4128
            draft_token_ids = [[0] for _ in range(num_reqs)]
            dummy_spec_decode_metadata = SpecDecodeMetadata.make_dummy(
4129
4130
                draft_token_ids, self.device
            )
4131
4132
4133
4134
4135
4136

            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
4137
4138
4139
4140
4141
            logits = torch.randn(
                num_tokens + num_reqs,
                logits.shape[-1],
                device=self.device,
                dtype=logits.dtype,
4142
            )
4143
4144
4145
            self.rejection_sampler(
                dummy_spec_decode_metadata,
                draft_probs,
4146
                logits,
4147
4148
                dummy_metadata,
            )
4149
        return sampler_output
4150

4151
    def _dummy_pooler_run_task(
4152
4153
        self,
        hidden_states: torch.Tensor,
4154
4155
        task: PoolingTask,
    ) -> PoolerOutput:
4156
4157
4158
4159
4160
4161
4162
4163
4164
4165
4166
        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

4167
        dummy_prompt_lens = torch.tensor(
4168
4169
            num_scheduled_tokens_list,
            device="cpu",
4170
        )
4171
4172
4173
        dummy_token_ids = torch.zeros(
            (num_reqs, req_num_tokens), dtype=torch.int32, device=self.device
        )
4174

4175
        model = cast(VllmModelForPooling, self.get_model())
4176
        dummy_pooling_params = PoolingParams(task=task)
4177
        dummy_pooling_params.verify(task=task, model_config=self.model_config)
4178
        to_update = model.pooler.get_pooling_updates(task)
4179
4180
        to_update.apply(dummy_pooling_params)

4181
        dummy_metadata = PoolingMetadata(
4182
4183
4184
4185
            prompt_lens=dummy_prompt_lens,
            prompt_token_ids=dummy_token_ids,
            pooling_params=[dummy_pooling_params] * num_reqs,
        )
4186

4187
4188
4189
        dummy_metadata.build_pooling_cursor(
            num_scheduled_tokens_list, device=hidden_states.device
        )
4190

4191
        try:
4192
4193
4194
            return model.pooler(
                hidden_states=hidden_states, pooling_metadata=dummy_metadata
            )
4195
        except RuntimeError as e:
4196
            if "out of memory" in str(e):
4197
                raise RuntimeError(
4198
4199
4200
                    "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 "
4201
4202
                    "initializing the engine."
                ) from e
4203
4204
            else:
                raise e
4205
4206
4207
4208
4209
4210
4211

    @torch.inference_mode()
    def _dummy_pooler_run(
        self,
        hidden_states: torch.Tensor,
    ) -> PoolerOutput:
        # Find the task that has the largest output for subsequent steps
4212
4213
4214
        supported_pooling_tasks = self.get_supported_pooling_tasks()

        if not supported_pooling_tasks:
4215
            if self.scheduler_config.enable_chunked_prefill:
4216
4217
4218
4219
4220
4221
4222
4223
4224
4225
4226
4227
4228
4229
4230
4231
                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."
                )

4232
        output_size = dict[PoolingTask, float]()
4233
        for task in supported_pooling_tasks:
4234
4235
            # Run a full batch with each task to ensure none of them OOMs
            output = self._dummy_pooler_run_task(hidden_states, task)
4236
            output_size[task] = sum(o.nbytes for o in output)
4237
4238
4239
4240
            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)
4241

4242
    def profile_run(self) -> None:
4243
        # Profile with multimodal encoder & encoder cache.
4244
        if self.supports_mm_inputs:
4245
4246
            mm_config = self.model_config.multimodal_config
            if mm_config is not None and mm_config.skip_mm_profiling:
4247
                logger.info(
4248
                    "Skipping memory profiling for multimodal encoder and "
4249
4250
                    "encoder cache."
                )
4251
4252
4253
4254
4255
4256
4257
4258
            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.
4259
                    dummy_modality = mm_budget.get_modality_with_max_tokens()
4260
4261
4262
                    max_mm_items_per_batch = mm_budget.max_items_per_batch_by_modality[
                        dummy_modality
                    ]
4263
4264
4265
4266
4267
4268
4269
4270
4271

                    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,
                    )
4272

4273
4274
4275
4276
4277
                    # Create dummy batch of multimodal inputs.
                    batched_dummy_mm_inputs = self._get_mm_dummy_batch(
                        dummy_modality,
                        max_mm_items_per_batch,
                    )
4278

4279
                    # Run multimodal encoder.
4280
                    dummy_encoder_outputs = self.model.embed_multimodal(
4281
4282
                        **batched_dummy_mm_inputs
                    )
4283

4284
4285
4286
4287
                    sanity_check_mm_encoder_outputs(
                        dummy_encoder_outputs,
                        expected_num_items=max_mm_items_per_batch,
                    )
4288

4289
4290
4291
                    # 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
4292
4293
                    # (max_tokens_for_modality, hidden_size) and scatter
                    # encoder output into it.
4294
                    encoder_output_shape = dummy_encoder_outputs[0].shape
4295
4296
4297
4298
4299
                    max_mm_tokens_per_item = mm_budget.max_tokens_by_modality[
                        dummy_modality
                    ]
                    if encoder_output_shape[0] < max_mm_tokens_per_item:
                        encoder_hidden_size = encoder_output_shape[-1]
4300
4301
4302
                        expanded_outputs = []
                        for output in dummy_encoder_outputs:
                            expanded = output.new_zeros(
4303
                                (max_mm_tokens_per_item, encoder_hidden_size)
4304
                            )
4305
4306
4307
4308
4309
4310
                            num_tokens = output.shape[0]
                            expanded[:num_tokens].copy_(output)
                            expanded_outputs.append(expanded)

                        dummy_encoder_outputs = expanded_outputs

4311
                    # Cache the dummy encoder outputs.
4312
                    self.encoder_cache["tmp"] = dict(enumerate(dummy_encoder_outputs))
4313

4314
        # Add `is_profile` here to pre-allocate communication buffers
4315
4316
4317
        hidden_states, last_hidden_states = self._dummy_run(
            self.max_num_tokens, is_profile=True
        )
4318
        if get_pp_group().is_last_rank:
4319
4320
4321
4322
            if self.is_pooling_model:
                output = self._dummy_pooler_run(hidden_states)
            else:
                output = self._dummy_sampler_run(last_hidden_states)
4323
        else:
4324
            output = None
4325
        self._sync_device()
4326
        del hidden_states, output
4327
        self.encoder_cache.clear()
4328
        gc.collect()
4329

4330
    def capture_model(self) -> int:
4331
        if self.compilation_config.cudagraph_mode == CUDAGraphMode.NONE:
4332
            logger.warning(
4333
                "Skipping CUDA graph capture. To turn on CUDA graph capture, "
4334
4335
                "ensure `cudagraph_mode` was not manually set to `NONE`"
            )
4336
            return 0
4337

4338
4339
        compilation_counter.num_gpu_runner_capture_triggers += 1

4340
4341
        start_time = time.perf_counter()

4342
4343
4344
4345
4346
4347
4348
4349
4350
4351
4352
4353
4354
4355
        @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()
4356
                    gc.collect()
4357

4358
4359
4360
        # 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.
4361
        set_cudagraph_capturing_enabled(True)
4362
        with freeze_gc(), graph_capture(device=self.device):
4363
            start_free_gpu_memory = torch.cuda.mem_get_info()[0]
4364
            cudagraph_mode = self.compilation_config.cudagraph_mode
4365
            assert cudagraph_mode is not None
4366
4367
4368
4369
4370
4371
4372
4373
4374

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

4375
4376
            if cudagraph_mode.mixed_mode() != CUDAGraphMode.NONE:
                cudagraph_runtime_mode = cudagraph_mode.mixed_mode()
4377
                # make sure we capture the largest batch size first
4378
4379
4380
                compilation_cases = list(
                    product(reversed(self.cudagraph_batch_sizes), lora_cases)
                )
4381
4382
4383
                self._capture_cudagraphs(
                    compilation_cases,
                    cudagraph_runtime_mode=cudagraph_runtime_mode,
4384
4385
                    uniform_decode=False,
                )
4386

4387
4388
            # Capture full cudagraph for uniform decode batches if we
            # don't already have full mixed prefill-decode cudagraphs.
4389
4390
4391
4392
4393
4394
4395
            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
                )
4396
                decode_cudagraph_batch_sizes = [
4397
4398
                    x
                    for x in self.cudagraph_batch_sizes
4399
                    if max_num_tokens >= x >= self.uniform_decode_query_len
4400
                ]
4401
4402
4403
                compilation_cases_decode = list(
                    product(reversed(decode_cudagraph_batch_sizes), lora_cases)
                )
4404
4405
4406
                self._capture_cudagraphs(
                    compilation_cases=compilation_cases_decode,
                    cudagraph_runtime_mode=CUDAGraphMode.FULL,
4407
4408
                    uniform_decode=True,
                )
4409

4410
4411
4412
            torch.cuda.synchronize()
            end_free_gpu_memory = torch.cuda.mem_get_info()[0]

4413
4414
4415
        # 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
4416
        # we may do lazy capturing in future that still allows capturing
4417
4418
        # after here.
        set_cudagraph_capturing_enabled(False)
4419
4420
4421
4422
4423

        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.
4424
        logger.info_once(
4425
4426
4427
            "Graph capturing finished in %.0f secs, took %.2f GiB",
            elapsed_time,
            cuda_graph_size / (1 << 30),
4428
            scope="local",
4429
        )
4430
        return cuda_graph_size
4431

4432
4433
    def _capture_cudagraphs(
        self,
4434
        compilation_cases: list[tuple[int, bool]],
4435
4436
4437
4438
4439
4440
4441
        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}"
4442
4443
4444
4445
4446
4447
4448
4449

        # 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",
4450
4451
4452
                    cudagraph_runtime_mode.name,
                ),
            )
4453

4454
        # We skip EPLB here since we don't want to record dummy metrics
4455
        for num_tokens, activate_lora in compilation_cases:
4456
            # We currently only capture ubatched graphs when its a FULL
4457
4458
4459
            # 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
4460
4461
4462
4463
            allow_microbatching = (
                self.parallel_config.enable_dbo
                and cudagraph_runtime_mode == CUDAGraphMode.FULL
                and uniform_decode
4464
4465
4466
4467
4468
                and check_ubatch_thresholds(
                    config=self.vllm_config.parallel_config,
                    num_tokens=num_tokens,
                    uniform_decode=uniform_decode,
                )
4469
            )
4470

4471
4472
4473
4474
4475
4476
            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.
4477
4478
4479
4480
4481
4482
4483
4484
4485
                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,
4486
                    activate_lora=activate_lora,
4487
4488
4489
4490
4491
4492
4493
4494
                )
            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,
4495
                activate_lora=activate_lora,
Rémi Delacourt's avatar
Rémi Delacourt committed
4496
                is_graph_capturing=True,
4497
            )
4498
        self.maybe_remove_all_loras(self.lora_config)
4499

4500
4501
4502
4503
    def initialize_attn_backend(self, kv_cache_config: KVCacheConfig) -> None:
        """
        Initialize the attention backends and attention metadata builders.
        """
4504
        assert len(self.attn_groups) == 0, "Attention backends are already initialized"
4505

4506
4507
4508
4509
4510
4511
        class AttentionGroupKey(NamedTuple):
            attn_backend: type[AttentionBackend]
            kv_cache_spec: KVCacheSpec

        def get_attn_backends_for_group(
            kv_cache_group_spec: KVCacheGroupSpec,
4512
        ) -> tuple[dict[AttentionGroupKey, list[str]], set[type[AttentionBackend]]]:
4513
            layer_type = cast(type[Any], AttentionLayerBase)
4514
            layers = get_layers_from_vllm_config(
4515
                self.vllm_config, layer_type, kv_cache_group_spec.layer_names
4516
            )
4517
4518
            attn_backends = {}
            attn_backend_layers = defaultdict(list)
4519
            # Dedupe based on full class name; this is a bit safer than
4520
4521
4522
4523
            # 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.
4524
            for layer_name in kv_cache_group_spec.layer_names:
4525
                attn_backend = layers[layer_name].get_attn_backend()
4526
4527
4528
4529

                if layer_name in self.kv_sharing_fast_prefill_eligible_layers:
                    attn_backend = create_fast_prefill_custom_backend(
                        "FastPrefill",
4530
                        attn_backend,  # type: ignore[arg-type]
4531
4532
                    )

4533
4534
4535
                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):
4536
                    layer_kv_cache_spec = layer_kv_cache_spec.kv_cache_specs[layer_name]
4537
                key = (full_cls_name, layer_kv_cache_spec)
4538
4539
4540
                attn_backends[key] = AttentionGroupKey(
                    attn_backend, layer_kv_cache_spec
                )
4541
                attn_backend_layers[key].append(layer_name)
4542
4543
4544
4545
            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()),
            )
4546
4547

        def create_attn_groups(
4548
            attn_backends_map: dict[AttentionGroupKey, list[str]],
4549
            kv_cache_group_id: int,
4550
4551
        ) -> list[AttentionGroup]:
            attn_groups: list[AttentionGroup] = []
4552
            for (attn_backend, kv_cache_spec), layer_names in attn_backends_map.items():
4553
                attn_group = AttentionGroup(
4554
                    attn_backend,
4555
                    layer_names,
4556
                    kv_cache_spec,
4557
                    kv_cache_group_id,
4558
4559
                )

4560
4561
4562
                attn_groups.append(attn_group)
            return attn_groups

4563
        attention_backend_maps = []
4564
        attention_backend_list = []
4565
        for kv_cache_group_spec in kv_cache_config.kv_cache_groups:
4566
            attn_backends = get_attn_backends_for_group(kv_cache_group_spec)
4567
            attention_backend_maps.append(attn_backends[0])
4568
            attention_backend_list.append(attn_backends[1])
4569
4570

        # Resolve cudagraph_mode before actually initialize metadata_builders
4571
4572
4573
        self._check_and_update_cudagraph_mode(
            attention_backend_list, kv_cache_config.kv_cache_groups
        )
4574

4575
4576
        for i, attn_backend_map in enumerate(attention_backend_maps):
            self.attn_groups.append(create_attn_groups(attn_backend_map, i))
4577

4578
4579
4580
4581
4582
4583
4584
4585
4586
4587
4588
4589
4590
4591
4592
4593
4594
4595
    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
4596
        # Calculate reorder batch threshold (if needed)
4597
4598
        # Note (tdoublep): do this *after* constructing builders,
        # because some of them change the threshold at init time.
4599
4600
        self.calculate_reorder_batch_threshold()

4601
    def _check_and_update_cudagraph_mode(
4602
4603
4604
        self,
        attention_backends: list[set[type[AttentionBackend]]],
        kv_cache_groups: list[KVCacheGroupSpec],
4605
    ) -> None:
4606
        """
4607
        Resolve the cudagraph_mode when there are multiple attention
4608
        groups with potential conflicting CUDA graph support.
4609
4610
4611
        Then initialize the cudagraph_dispatcher based on the resolved
        cudagraph_mode.
        """
4612
        min_cg_support = AttentionCGSupport.ALWAYS
4613
        min_cg_backend_name = None
4614

4615
4616
4617
4618
4619
4620
4621
4622
4623
4624
4625
4626
        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__
4627
4628
        # Flexible resolve the cudagraph mode
        cudagraph_mode = self.compilation_config.cudagraph_mode
4629
        assert cudagraph_mode is not None
4630
        # check cudagraph for mixed batch is supported
4631
4632
4633
4634
4635
4636
        if (
            cudagraph_mode.mixed_mode() == CUDAGraphMode.FULL
            and min_cg_support != AttentionCGSupport.ALWAYS
        ):
            msg = (
                f"CUDAGraphMode.{cudagraph_mode.name} is not supported "
4637
                f"with {min_cg_backend_name} backend (support: "
4638
4639
                f"{min_cg_support})"
            )
4640
4641
            if min_cg_support == AttentionCGSupport.NEVER:
                # if not supported any full cudagraphs, just raise it.
4642
4643
                msg += (
                    "; please try cudagraph_mode=PIECEWISE, and "
4644
                    "make sure compilation mode is VLLM_COMPILE"
4645
                )
4646
4647
4648
4649
4650
                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"
4651
                cudagraph_mode = self.compilation_config.cudagraph_mode = (
4652
                    CUDAGraphMode.FULL_AND_PIECEWISE
4653
                )
4654
4655
            else:
                msg += "; setting cudagraph_mode=FULL_DECODE_ONLY"
4656
                cudagraph_mode = self.compilation_config.cudagraph_mode = (
4657
                    CUDAGraphMode.FULL_DECODE_ONLY
4658
                )
4659
4660
            logger.warning(msg)

4661
        # check that if we are doing decode full-cudagraphs it is supported
4662
4663
4664
4665
4666
4667
        if (
            cudagraph_mode.decode_mode() == CUDAGraphMode.FULL
            and min_cg_support == AttentionCGSupport.NEVER
        ):
            msg = (
                f"CUDAGraphMode.{cudagraph_mode.name} is not supported "
4668
                f"with {min_cg_backend_name} backend (support: "
4669
4670
                f"{min_cg_support})"
            )
4671
            if self.compilation_config.mode == CompilationMode.VLLM_COMPILE and (
4672
4673
4674
4675
4676
                self.compilation_config.splitting_ops_contain_attention()
                or self.compilation_config.use_inductor_graph_partition
            ):
                msg += (
                    "; setting cudagraph_mode=PIECEWISE because "
4677
                    "attention is compiled piecewise"
4678
4679
                )
                cudagraph_mode = self.compilation_config.cudagraph_mode = (
4680
                    CUDAGraphMode.PIECEWISE
4681
                )
4682
            else:
4683
4684
                msg += (
                    "; setting cudagraph_mode=NONE because "
4685
                    "attention is not compiled piecewise"
4686
4687
                )
                cudagraph_mode = self.compilation_config.cudagraph_mode = (
4688
                    CUDAGraphMode.NONE
4689
                )
4690
4691
            logger.warning(msg)

4692
4693
        # check that if we are doing spec-decode + decode full-cudagraphs it is
        # supported
4694
4695
4696
4697
4698
4699
4700
4701
        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 "
4702
                f"{min_cg_backend_name} (support: {min_cg_support})"
4703
            )
4704
4705
            if self.compilation_config.splitting_ops_contain_attention():
                msg += "; setting cudagraph_mode=PIECEWISE"
4706
                cudagraph_mode = self.compilation_config.cudagraph_mode = (
4707
                    CUDAGraphMode.PIECEWISE
4708
                )
4709
4710
            else:
                msg += "; setting cudagraph_mode=NONE"
4711
                cudagraph_mode = self.compilation_config.cudagraph_mode = (
4712
                    CUDAGraphMode.NONE
4713
                )
4714
4715
4716
4717
            logger.warning(msg)

        # double check that we can support full cudagraph if they are requested
        # even after automatic downgrades
4718
4719
4720
4721
4722
4723
        if (
            cudagraph_mode.has_full_cudagraphs()
            and min_cg_support == AttentionCGSupport.NEVER
        ):
            raise ValueError(
                f"CUDAGraphMode.{cudagraph_mode.name} is not "
4724
                f"supported with {min_cg_backend_name} backend ("
4725
4726
                f"support:{min_cg_support}) "
                "; please try cudagraph_mode=PIECEWISE, "
4727
                "and make sure compilation mode is VLLM_COMPILE"
4728
            )
4729

4730
4731
4732
4733
4734
4735
4736
4737
4738
4739
4740
4741
4742
4743
        # 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
            )
4744
4745
4746
4747
            capture_sizes = self.compilation_config.cudagraph_capture_sizes
            self.cudagraph_batch_sizes = (
                capture_sizes if capture_sizes is not None else []
            )
4748

4749
4750
        # Trigger cudagraph dispatching keys initialization after
        # resolved cudagraph mode.
4751
        self.compilation_config.cudagraph_mode = cudagraph_mode
4752
        self.cudagraph_dispatcher.initialize_cudagraph_keys(
4753
            cudagraph_mode, self.uniform_decode_query_len
4754
        )
4755

4756
4757
    def calculate_reorder_batch_threshold(self) -> None:
        """
4758
4759
4760
4761
        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.
4762
        """
4763
4764
        min_none_high = lambda a, b: a if b is None else b if a is None else min(a, b)

4765
        reorder_batch_thresholds: list[int | None] = [
4766
4767
4768
            group.get_metadata_builder().reorder_batch_threshold
            for group in self._attn_group_iterator()
        ]
4769
4770
4771
4772
4773
        # 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
4774
        self.reorder_batch_threshold = reduce(min_none_high, reorder_batch_thresholds)  # type: ignore[assignment]
4775

4776
4777
4778
    @staticmethod
    def select_common_block_size(
        kv_manager_block_size: int, attn_groups: list[AttentionGroup]
4779
4780
    ) -> int:
        """
4781
4782
4783
4784
4785
        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.
4786
4787
4788
4789
4790
4791

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

        Returns:
4792
            The selected block size
4793
4794

        Raises:
4795
            ValueError: If no valid block size found
4796
4797
        """

4798
4799
4800
4801
4802
4803
4804
4805
        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
4806
                for supported_size in backend.get_supported_kernel_block_sizes():
4807
4808
4809
4810
4811
4812
4813
4814
4815
4816
4817
4818
4819
4820
4821
4822
4823
4824
4825
4826
4827
4828
4829
4830
4831
4832
4833
4834
4835
4836
                    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
4837
            for supported_size in backend.get_supported_kernel_block_sizes()
4838
4839
            if isinstance(supported_size, int)
        )
4840

4841
4842
4843
4844
4845
4846
        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}. ")
4847

4848
4849
4850
    def may_reinitialize_input_batch(
        self, kv_cache_config: KVCacheConfig, kernel_block_sizes: list[int]
    ) -> None:
4851
4852
4853
4854
4855
4856
4857
        """
        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.
4858
            kernel_block_sizes: The kernel block sizes for each KV cache group.
4859
4860
4861
4862
        """
        block_sizes = [
            kv_cache_group.kv_cache_spec.block_size
            for kv_cache_group in kv_cache_config.kv_cache_groups
4863
            if not isinstance(kv_cache_group.kv_cache_spec, EncoderOnlyAttentionSpec)
4864
        ]
4865
4866
4867
4868

        if block_sizes != [self.cache_config.block_size] or kernel_block_sizes != [
            self.cache_config.block_size
        ]:
4869
4870
4871
            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
4872
4873
                "for more details."
            )
4874
4875
            self.input_batch = InputBatch(
                max_num_reqs=self.max_num_reqs,
4876
                max_model_len=max(self.max_model_len, self.max_encoder_len),
4877
4878
4879
4880
4881
                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,
4882
                kernel_block_sizes=kernel_block_sizes,
4883
                is_spec_decode=bool(self.vllm_config.speculative_config),
4884
                logitsprocs=self.input_batch.logitsprocs,
4885
                logitsprocs_need_output_token_ids=self.input_batch.logitsprocs_need_output_token_ids,
4886
                is_pooling_model=self.is_pooling_model,
4887
                num_speculative_tokens=self.num_spec_tokens,
4888
4889
            )

4890
    def _allocate_kv_cache_tensors(
4891
4892
        self, kv_cache_config: KVCacheConfig
    ) -> dict[str, torch.Tensor]:
4893
        """
4894
4895
4896
        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.

4897
        Args:
4898
            kv_cache_config: The KV cache config
4899
        Returns:
4900
            dict[str, torch.Tensor]: A map between layer names to their
4901
            corresponding memory buffer for KV cache.
4902
        """
4903
4904
        kv_cache_raw_tensors: dict[str, torch.Tensor] = {}
        for kv_cache_tensor in kv_cache_config.kv_cache_tensors:
4905
4906
4907
            tensor = torch.zeros(
                kv_cache_tensor.size, dtype=torch.int8, device=self.device
            )
4908
4909
4910
4911
4912
            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:
4913
4914
4915
4916
            for layer_name in group.layer_names:
                if layer_name in self.runner_only_attn_layers:
                    continue
                layer_names.add(layer_name)
4917
4918
4919
        assert layer_names == set(kv_cache_raw_tensors.keys()), (
            "Some layers are not correctly initialized"
        )
4920
4921
        return kv_cache_raw_tensors

4922
4923
4924
    def _attn_group_iterator(self) -> Iterator[AttentionGroup]:
        return itertools.chain.from_iterable(self.attn_groups)

4925
    def _kv_cache_spec_attn_group_iterator(self) -> Iterator[AttentionGroup]:
4926
4927
        if not self.kv_cache_config.kv_cache_groups:
            return
4928
4929
        for attn_groups in self.attn_groups:
            yield from attn_groups
4930

4931
4932
4933
4934
4935
4936
4937
4938
4939
4940
4941
4942
4943
4944
4945
    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 = []
4946
        for kv_cache_gid, kv_cache_group in enumerate(kv_cache_config.kv_cache_groups):
4947
4948
4949
4950
4951
4952
            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):
4953
                continue
4954
            elif isinstance(kv_cache_spec, AttentionSpec):
4955
4956
4957
                # This is an attention backend that supports virtual
                # block splitting. Get the supported block sizes from
                # all backends in the group.
4958
                attn_groups = self.attn_groups[kv_cache_gid]
4959
                kv_manager_block_size = kv_cache_group.kv_cache_spec.block_size
4960
                selected_kernel_size = self.select_common_block_size(
4961
4962
4963
                    kv_manager_block_size, attn_groups
                )
                kernel_block_sizes.append(selected_kernel_size)
4964
            elif isinstance(kv_cache_spec, MambaSpec):
4965
4966
                # This is likely Mamba or other non-attention cache,
                # no splitting.
4967
                kernel_block_sizes.append(kv_cache_spec.block_size)
4968
4969
4970
4971
4972
4973
            else:
                raise NotImplementedError(
                    f"unknown kv cache spec {kv_cache_group.kv_cache_spec}"
                )
        return kernel_block_sizes

4974
4975
4976
4977
    def _reshape_kv_cache_tensors(
        self,
        kv_cache_config: KVCacheConfig,
        kv_cache_raw_tensors: dict[str, torch.Tensor],
4978
        kernel_block_sizes: list[int],
4979
    ) -> dict[str, torch.Tensor]:
4980
        """
4981
        Reshape the KV cache tensors to the desired shape and dtype.
4982

4983
        Args:
4984
4985
            kv_cache_config: The KV cache config
            kv_cache_raw_tensors: The KV cache buffer of each layer, with
4986
                correct size but uninitialized shape.
4987
            kernel_block_sizes: The kernel block sizes for each KV cache group.
4988
        Returns:
4989
            Dict[str, torch.Tensor]: A map between layer names to their
4990
4991
            corresponding memory buffer for KV cache.
        """
4992
        kv_caches: dict[str, torch.Tensor] = {}
4993
        has_attn, has_mamba = False, False
4994
4995
        for group in self._kv_cache_spec_attn_group_iterator():
            kv_cache_spec = group.kv_cache_spec
4996
            attn_backend = group.backend
4997
4998
4999
5000
            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]
5001
            for layer_name in group.layer_names:
5002
5003
                if layer_name in self.runner_only_attn_layers:
                    continue
5004
5005
                raw_tensor = kv_cache_raw_tensors[layer_name]
                assert raw_tensor.numel() % kv_cache_spec.page_size_bytes == 0
5006
                num_blocks = raw_tensor.numel() // kv_cache_spec.page_size_bytes
5007
                if isinstance(kv_cache_spec, AttentionSpec):
5008
                    has_attn = True
5009
5010
                    num_blocks_per_kv_block = (
                        kv_cache_spec.block_size // kernel_block_size
5011
5012
5013
                    )
                    kernel_num_blocks = num_blocks * num_blocks_per_kv_block

5014
                    kv_cache_shape = attn_backend.get_kv_cache_shape(
5015
                        kernel_num_blocks,
5016
                        kernel_block_size,
5017
5018
                        kv_cache_spec.num_kv_heads,
                        kv_cache_spec.head_size,
5019
5020
                        cache_dtype_str=self.cache_config.cache_dtype,
                    )
5021
                    dtype = kv_cache_spec.dtype
5022
                    try:
5023
                        kv_cache_stride_order = attn_backend.get_kv_cache_stride_order()
5024
                        assert len(kv_cache_stride_order) == len(kv_cache_shape)
5025
                    except (AttributeError, NotImplementedError):
5026
                        kv_cache_stride_order = tuple(range(len(kv_cache_shape)))
5027
5028
5029
5030
5031
                    # 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.
5032
5033
5034
                    kv_cache_shape = tuple(
                        kv_cache_shape[i] for i in kv_cache_stride_order
                    )
5035
5036
5037
5038
5039
                    # Maintain original KV shape view.
                    inv_order = [
                        kv_cache_stride_order.index(i)
                        for i in range(len(kv_cache_stride_order))
                    ]
5040
5041
5042
5043
5044
5045
                    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
5046
                elif isinstance(kv_cache_spec, MambaSpec):
5047
                    has_mamba = True
Chen Zhang's avatar
Chen Zhang committed
5048
5049
                    raw_tensor = kv_cache_raw_tensors[layer_name]
                    state_tensors = []
5050
                    storage_offset_bytes = 0
5051
                    for shape, dtype in zip(kv_cache_spec.shapes, kv_cache_spec.dtypes):
5052
5053
                        dtype_size = get_dtype_size(dtype)
                        num_element_per_page = (
5054
5055
                            kv_cache_spec.page_size_bytes // dtype_size
                        )
Chen Zhang's avatar
Chen Zhang committed
5056
                        target_shape = (num_blocks, *shape)
5057
5058
                        stride = torch.empty(target_shape).stride()
                        target_stride = (num_element_per_page, *stride[1:])
5059
                        assert storage_offset_bytes % dtype_size == 0
5060
5061
5062
5063
                        tensor = torch.as_strided(
                            raw_tensor.view(dtype),
                            size=target_shape,
                            stride=target_stride,
5064
                            storage_offset=storage_offset_bytes // dtype_size,
5065
                        )
Chen Zhang's avatar
Chen Zhang committed
5066
                        state_tensors.append(tensor)
5067
                        storage_offset_bytes += stride[0] * dtype_size
5068
5069

                    kv_caches[layer_name] = state_tensors
5070
                else:
5071
                    raise NotImplementedError
5072
5073

        if has_attn and has_mamba:
5074
            self._update_hybrid_attention_mamba_layout(kv_caches)
5075

5076
5077
        return kv_caches

5078
    def _update_hybrid_attention_mamba_layout(
5079
5080
        self, kv_caches: dict[str, torch.Tensor]
    ) -> None:
5081
        """
5082
5083
        Update the layout of attention layers from (2, num_blocks, ...) to
        (num_blocks, 2, ...).
5084
5085

        Args:
5086
            kv_caches: The KV cache buffer of each layer.
5087
5088
        """

5089
5090
        for group in self._kv_cache_spec_attn_group_iterator():
            kv_cache_spec = group.kv_cache_spec
5091
            for layer_name in group.layer_names:
5092
                kv_cache = kv_caches[layer_name]
5093
5094
5095
5096
                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 "
5097
                        f"a tensor of shape {kv_cache.shape}"
5098
                    )
5099
                    hidden_size = kv_cache.shape[2:].numel()
5100
5101
5102
5103
                    kv_cache.as_strided_(
                        size=kv_cache.shape,
                        stride=(hidden_size, 2 * hidden_size, *kv_cache.stride()[2:]),
                    )
5104

5105
    def initialize_kv_cache_tensors(
5106
        self, kv_cache_config: KVCacheConfig, kernel_block_sizes: list[int]
5107
    ) -> dict[str, torch.Tensor]:
5108
5109
5110
5111
5112
        """
        Initialize the memory buffer for KV cache.

        Args:
            kv_cache_config: The KV cache config
5113
5114
            kernel_block_sizes: The kernel block sizes for each KV cache group.

5115
        Returns:
5116
            Dict[str, torch.Tensor]: A map between layer names to their
5117
5118
            corresponding memory buffer for KV cache.
        """
5119
5120
5121
5122
5123
5124
5125
5126
5127
5128
5129
5130
5131
5132
5133
5134
5135
5136
5137
5138
5139
5140
5141
5142

        # 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
            )
5143

5144
        # Set up cross-layer KV cache sharing
5145
5146
        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)
5147
5148
            kv_caches[layer_name] = kv_caches[target_layer_name]

5149
5150
5151
5152
5153
5154
5155
5156
5157
        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,
        )
5158
5159
5160
        return kv_caches

    def maybe_add_kv_sharing_layers_to_kv_cache_groups(
5161
5162
        self, kv_cache_config: KVCacheConfig
    ) -> None:
5163
5164
5165
5166
5167
5168
5169
5170
5171
5172
5173
5174
5175
5176
5177
5178
5179
5180
        """
        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.
5181
            attn_layers = get_layers_from_vllm_config(self.vllm_config, Attention)
5182
5183
            for layer_name in reversed(attn_layers):
                if layer_name in self.shared_kv_cache_layers:
5184
                    self.kv_sharing_fast_prefill_eligible_layers.add(layer_name)
5185
5186
                else:
                    break
5187

5188
5189
5190
5191
5192
5193
5194
    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
        """
5195
        kv_cache_config = deepcopy(kv_cache_config)
5196
        self.kv_cache_config = kv_cache_config
5197
        self.may_add_encoder_only_layers_to_kv_cache_config()
5198
        self.maybe_add_kv_sharing_layers_to_kv_cache_groups(kv_cache_config)
5199
        self.initialize_attn_backend(kv_cache_config)
5200
5201
5202
5203
5204
5205
        # 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)
5206
5207
5208
5209

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

5210
        # Reinitialize need to after initialize_attn_backend
5211
5212
5213
5214
        self.may_reinitialize_input_batch(kv_cache_config, kernel_block_sizes)
        kv_caches = self.initialize_kv_cache_tensors(
            kv_cache_config, kernel_block_sizes
        )
5215

5216
5217
5218
5219
5220
5221
        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
5222
        if has_kv_transfer_group():
5223
            kv_transfer_group = get_kv_transfer_group()
5224
5225
5226
5227
5228
5229
5230
            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)
5231
            kv_transfer_group.set_host_xfer_buffer_ops(copy_kv_blocks)
Robert Shaw's avatar
Robert Shaw committed
5232

5233
        if self.dcp_world_size > 1:
5234
5235
            layer_type = cast(type[Any], AttentionLayerBase)
            layers = get_layers_from_vllm_config(self.vllm_config, layer_type)
5236
            for layer in layers.values():
5237
5238
5239
5240
                layer_impl = getattr(layer, "impl", None)
                if layer_impl is None:
                    continue
                assert layer_impl.need_to_return_lse_for_decode, (
5241
5242
                    "DCP requires attention impls to return"
                    " the softmax lse for decode, but the impl "
5243
                    f"{layer_impl.__class__.__name__} "
5244
5245
                    "does not return the softmax lse for decode."
                )
5246

5247
5248
5249
5250
5251
    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
5252
        encoder_only_attn_specs: dict[AttentionSpec, list[str]] = defaultdict(list)
5253
5254
5255
        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:
5256
                attn_spec: AttentionSpec = EncoderOnlyAttentionSpec(
5257
5258
5259
                    block_size=block_size,
                    num_kv_heads=attn_module.num_kv_heads,
                    head_size=attn_module.head_size,
5260
5261
                    dtype=self.kv_cache_dtype,
                )
5262
5263
5264
                encoder_only_attn_specs[attn_spec].append(layer_name)
                self.runner_only_attn_layers.add(layer_name)
        if len(encoder_only_attn_specs) > 0:
5265
5266
5267
            assert len(encoder_only_attn_specs) == 1, (
                "Only support one encoder-only attention spec now"
            )
5268
5269
            spec, layer_names = encoder_only_attn_specs.popitem()
            self.kv_cache_config.kv_cache_groups.append(
5270
5271
                KVCacheGroupSpec(layer_names=layer_names, kv_cache_spec=spec)
            )
5272

5273
    def get_kv_cache_spec(self) -> dict[str, KVCacheSpec]:
5274
        """
5275
        Generates the KVCacheSpec by parsing the kv cache format from each
5276
5277
        Attention module in the static forward context.
        Returns:
5278
            KVCacheSpec: A dictionary mapping layer names to their KV cache
5279
5280
            format. Layers that do not need KV cache are not included.
        """
5281
5282
        if has_ec_transfer() and get_ec_transfer().is_producer:
            return {}
5283
        kv_cache_spec: dict[str, KVCacheSpec] = {}
5284
5285
        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
5286
        for layer_name, attn_module in attn_layers.items():
5287
5288
5289
5290
5291
5292
5293
5294
5295
5296
5297
5298
5299
5300
5301
            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
5302

5303
        return kv_cache_spec
5304

5305
    def _to_list(self, sampled_token_ids: torch.Tensor) -> list[list[int]]:
5306
5307
5308
5309
5310
5311
5312
5313
        # 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.
5314
        pinned = self.sampled_token_ids_pinned_cpu[: sampled_token_ids.shape[0]]
5315
5316
5317
        pinned.copy_(sampled_token_ids, non_blocking=True)
        self.transfer_event.record()
        self.transfer_event.synchronize()
5318
        return pinned.tolist()