gpu_model_runner.py 249 KB
Newer Older
1
# SPDX-License-Identifier: Apache-2.0
2
# SPDX-FileCopyrightText: Copyright contributors to the vLLM project
3

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

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

21
import vllm.envs as envs
22
from vllm.attention.layer import Attention, MLAAttention
23
from vllm.compilation.counter import compilation_counter
24
from vllm.compilation.cuda_graph import CUDAGraphStat, CUDAGraphWrapper
25
from vllm.compilation.monitor import set_cudagraph_capturing_enabled
26
from vllm.config import (
27
    CompilationMode,
28
29
30
31
32
    CUDAGraphMode,
    VllmConfig,
    get_layers_from_vllm_config,
    update_config,
)
33
from vllm.distributed.ec_transfer import get_ec_transfer, has_ec_transfer
34
from vllm.distributed.eplb.eplb_state import EplbState
35
from vllm.distributed.kv_transfer import get_kv_transfer_group, has_kv_transfer_group
36
from vllm.distributed.kv_transfer.kv_connector.utils import copy_kv_blocks
37
from vllm.distributed.parallel_state import (
38
    get_dcp_group,
39
40
41
42
43
44
    get_pp_group,
    get_tp_group,
    graph_capture,
    is_global_first_rank,
    prepare_communication_buffer_for_model,
)
45
46
47
48
from vllm.forward_context import (
    BatchDescriptor,
    set_forward_context,
)
49
from vllm.logger import init_logger
50
from vllm.lora.layers import LoRAMapping, LoRAMappingType
51
from vllm.model_executor.layers.attention_layer_base import AttentionLayerBase
52
53
54
from vllm.model_executor.layers.fused_moe.routed_experts_capturer import (
    RoutedExpertsCapturer,
)
55
56
57
58
from vllm.model_executor.layers.rotary_embedding import (
    MRotaryEmbedding,
    XDRotaryEmbedding,
)
59
from vllm.model_executor.model_loader import TensorizerLoader, get_model_loader
60
from vllm.model_executor.models.interfaces import (
61
    MultiModalEmbeddings,
62
    SupportsMRoPE,
63
    SupportsMultiModal,
64
    SupportsXDRoPE,
65
66
67
68
69
    is_mixture_of_experts,
    supports_eagle3,
    supports_mrope,
    supports_multimodal_pruning,
    supports_transcription,
70
    supports_xdrope,
71
)
72
from vllm.model_executor.models.interfaces_base import (
73
74
75
76
    VllmModelForPooling,
    is_pooling_model,
    is_text_generation_model,
)
77
from vllm.multimodal import MULTIMODAL_REGISTRY
78
79
80
81
82
from vllm.multimodal.inputs import (
    BatchedTensorInputs,
    MultiModalKwargsItem,
    PlaceholderRange,
)
83
from vllm.multimodal.utils import group_mm_kwargs_by_modality
84
from vllm.pooling_params import PoolingParams
85
from vllm.sampling_params import SamplingType
86
from vllm.sequence import IntermediateTensors
87
from vllm.tasks import GenerationTask, PoolingTask, SupportedTask
88
from vllm.utils import length_from_prompt_token_ids_or_embeds
89
from vllm.utils.jsontree import json_map_leaves
90
from vllm.utils.math_utils import cdiv, round_up
91
from vllm.utils.mem_utils import DeviceMemoryProfiler, format_gib
92
from vllm.utils.nvtx_pytorch_hooks import PytHooks
93
from vllm.utils.platform_utils import is_pin_memory_available
94
95
96
97
from vllm.utils.torch_utils import (
    get_dtype_size,
    kv_cache_dtype_str_to_dtype,
)
98
99
from vllm.v1.attention.backend import (
    AttentionBackend,
100
    AttentionCGSupport,
101
    AttentionMetadata,
102
    AttentionMetadataBuilder,
103
    AttentionType,
104
    CommonAttentionMetadata,
105
106
    MultipleOf,
)
107
from vllm.v1.attention.backends.gdn_attn import GDNAttentionMetadataBuilder
108
from vllm.v1.attention.backends.utils import (
109
    create_fast_prefill_custom_backend,
110
    get_dcp_local_seq_lens,
111
112
    reorder_batch_to_split_decodes_and_prefills,
)
113
from vllm.v1.cudagraph_dispatcher import CudagraphDispatcher
114
115
116
117
118
119
120
121
122
123
124
125
126
127
128
129
130
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,
131
    ECConnectorOutput,
132
    KVConnectorOutput,
133
134
135
136
137
    LogprobsLists,
    LogprobsTensors,
    ModelRunnerOutput,
    PoolerOutput,
    SamplerOutput,
138
    make_empty_encoder_model_runner_output,
139
)
140
from vllm.v1.pool.metadata import PoolingMetadata, PoolingStates
141
from vllm.v1.sample.logits_processor import LogitsProcessors, build_logitsprocs
142
from vllm.v1.sample.logits_processor.interface import LogitsProcessor
143
from vllm.v1.sample.metadata import SamplingMetadata
144
from vllm.v1.sample.rejection_sampler import RejectionSampler
145
from vllm.v1.sample.sampler import Sampler
146
from vllm.v1.spec_decode.draft_model import DraftModelProposer
147
from vllm.v1.spec_decode.eagle import EagleProposer
148
from vllm.v1.spec_decode.medusa import MedusaProposer
149
from vllm.v1.spec_decode.metadata import SpecDecodeMetadata
150
from vllm.v1.spec_decode.ngram_proposer import NgramProposer
151
from vllm.v1.spec_decode.suffix_decoding import SuffixDecodingProposer
152
from vllm.v1.structured_output.utils import apply_grammar_bitmask
153
from vllm.v1.utils import CpuGpuBuffer, record_function_or_nullcontext
154
155
156
157
158
from vllm.v1.worker import mamba_utils
from vllm.v1.worker.cp_utils import (
    check_attention_cp_compatibility,
    get_total_cp_world_size,
)
159
from vllm.v1.worker.dp_utils import coordinate_batch_across_dp
160
from vllm.v1.worker.ec_connector_model_runner_mixin import ECConnectorModelRunnerMixin
161
from vllm.v1.worker.gpu_input_batch import CachedRequestState, InputBatch
162
from vllm.v1.worker.gpu_ubatch_wrapper import UBatchWrapper
163
from vllm.v1.worker.kv_connector_model_runner_mixin import KVConnectorModelRunnerMixin
164
from vllm.v1.worker.lora_model_runner_mixin import LoRAModelRunnerMixin
165
166
167
from vllm.v1.worker.ubatch_utils import (
    UBatchSlices,
    check_ubatch_thresholds,
168
    maybe_create_ubatch_slices,
169
    split_attn_metadata,
170
)
171
from vllm.v1.worker.utils import is_residual_scattered_for_sp
172
from vllm.v1.worker.workspace import lock_workspace
173

174
175
176
177
178
179
180
from .utils import (
    AttentionGroup,
    MultiModalBudget,
    add_kv_sharing_layers_to_kv_cache_groups,
    bind_kv_cache,
    sanity_check_mm_encoder_outputs,
)
181

182
if TYPE_CHECKING:
183
    from vllm.model_executor.model_loader.tensorizer import TensorizerConfig
184
    from vllm.v1.core.sched.output import GrammarOutput, SchedulerOutput
185
186
187

logger = init_logger(__name__)

188
189
AttnMetadataDict: TypeAlias = dict[str, AttentionMetadata]
# list when ubatching is enabled
190
PerLayerAttnMetadata: TypeAlias = list[AttnMetadataDict] | AttnMetadataDict
191

192

193
194
195
196
197
198
# Wrapper for ModelRunnerOutput to support overlapped execution.
class AsyncGPUModelRunnerOutput(AsyncModelRunnerOutput):
    def __init__(
        self,
        model_runner_output: ModelRunnerOutput,
        sampled_token_ids: torch.Tensor,
199
        logprobs_tensors: LogprobsTensors | None,
200
201
        invalid_req_indices: list[int],
        async_output_copy_stream: torch.cuda.Stream,
202
        vocab_size: int,
203
204
205
206
207
    ):
        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.
208
        self.async_copy_ready_event = torch.Event()
209
210
211
212

        # 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
213
        self.vocab_size = vocab_size
214
        self._logprobs_tensors = logprobs_tensors
215
216
217
218
219

        # 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)
220
            self.sampled_token_ids_cpu = self._sampled_token_ids.to(
221
222
                "cpu", non_blocking=True
            )
223
224
225
226
227
            self._logprobs_tensors_cpu = (
                self._logprobs_tensors.to_cpu_nonblocking()
                if self._logprobs_tensors
                else None
            )
228
            self.async_copy_ready_event.record()
229
230
231

    def get_output(self) -> ModelRunnerOutput:
        """Copy the device tensors to the host and return a ModelRunnerOutput.
232

233
234
        This function blocks until the copy is finished.
        """
235
        max_gen_len = self.sampled_token_ids_cpu.shape[-1]
236
        self.async_copy_ready_event.synchronize()
237

238
239
        # Release the device tensors once the copy has completed.
        del self._logprobs_tensors
240
        del self._sampled_token_ids
241
        if max_gen_len == 1:
242
            valid_sampled_token_ids = self.sampled_token_ids_cpu.tolist()
243
244
            for i in self._invalid_req_indices:
                valid_sampled_token_ids[i].clear()
245
246
247
            logprobs_lists = None
            if self._logprobs_tensors_cpu is not None:
                logprobs_lists = self._logprobs_tensors_cpu.tolists()
248
        else:
249
            valid_sampled_token_ids, logprobs_lists = RejectionSampler.parse_output(
250
251
                self.sampled_token_ids_cpu,
                self.vocab_size,
252
                self._invalid_req_indices,
253
                logprobs_tensors=self._logprobs_tensors_cpu,
254
            )
255
256
257

        output = self._model_runner_output
        output.sampled_token_ids = valid_sampled_token_ids
258
        output.logprobs = logprobs_lists
259
260
261
        return output


262
263
264
265
266
267
268
269
270
271
272
273
274
275
276
277
278
279
280
281
282
class AsyncGPUPoolingModelRunnerOutput(AsyncModelRunnerOutput):
    def __init__(
        self,
        model_runner_output: ModelRunnerOutput,
        raw_pooler_output: PoolerOutput,
        finished_mask: list[bool],
        async_output_copy_stream: torch.cuda.Stream,
    ):
        self._model_runner_output = model_runner_output

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

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

        # Initiate the copy on a separate stream, but do not synchronize it.
        default_stream = torch.cuda.current_stream()
        with torch.cuda.stream(async_output_copy_stream):
            async_output_copy_stream.wait_stream(default_stream)
283
            raw_pooler_output_cpu = json_map_leaves(
284
285
286
287
                lambda x: None if x is None else x.to("cpu", non_blocking=True),
                self._raw_pooler_output,
            )
            self.async_copy_ready_event.record()
288
289
290
291
            self._model_runner_output.pooler_output = [
                out if include else None
                for out, include in zip(raw_pooler_output_cpu, finished_mask)
            ]
292
293
294
295
296
297
298
299
300
301
302
303

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

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


304
305
306
307
308
309
310
311
312
313
314
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
315
    ec_connector_output: ECConnectorOutput | None
316
    cudagraph_stats: CUDAGraphStat | None
317
318


319
320
321
class GPUModelRunner(
    LoRAModelRunnerMixin, KVConnectorModelRunnerMixin, ECConnectorModelRunnerMixin
):
322
323
    def __init__(
        self,
324
        vllm_config: VllmConfig,
325
        device: torch.device,
326
    ):
327
328
329
        self.vllm_config = vllm_config
        self.model_config = vllm_config.model_config
        self.cache_config = vllm_config.cache_config
330
        self.compilation_config = vllm_config.compilation_config
331
332
333
334
335
336
        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
337

338
        from vllm.model_executor.models.utils import set_cpu_offload_max_bytes
339
340

        set_cpu_offload_max_bytes(int(self.cache_config.cpu_offload_gb * 1024**3))
341

342
343
344
345
        model_config = self.model_config
        cache_config = self.cache_config
        scheduler_config = self.scheduler_config
        parallel_config = self.parallel_config
346
        self.device = device
347
348
        self.pin_memory = is_pin_memory_available()
        self.dtype = self.model_config.dtype
349

350
351
352
        self.kv_cache_dtype = kv_cache_dtype_str_to_dtype(
            cache_config.cache_dtype, self.model_config
        )
353

354
        self.is_pooling_model = model_config.runner_type == "pooling"
355
        self.enable_prompt_embeds = model_config.enable_prompt_embeds
356
        self.is_multimodal_raw_input_only_model = (
357
358
            model_config.is_multimodal_raw_input_only_model
        )
359
360
        # This will be overridden in load_model()
        self.is_multimodal_pruning_enabled = False
361
        self.max_model_len = model_config.max_model_len
362
363
364

        # Always set to false after the first forward pass
        self.calculate_kv_scales = self.cache_config.calculate_kv_scales
365
        self.dcp_world_size = self.parallel_config.decode_context_parallel_size
366
        self.dcp_rank = 0 if self.dcp_world_size <= 1 else get_dcp_group().rank_in_group
367
        self.max_num_tokens = scheduler_config.max_num_batched_tokens
368
        self.max_num_reqs = scheduler_config.max_num_seqs
369

370
371
372
373
374
        # 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 = (
375
            self.parallel_config.distributed_executor_backend == "external_launcher"
376
            and len(get_pp_group().ranks) > 1
377
        )
378

379
        # Model-related.
380
        self.num_query_heads = model_config.get_num_attention_heads(parallel_config)
381
        self.inputs_embeds_size = model_config.get_inputs_embeds_size()
382
        self.attention_chunk_size = model_config.attention_chunk_size
383
        # Only relevant for models using ALiBi (e.g, MPT)
384
        self.use_alibi = model_config.uses_alibi
385

386
        self.cascade_attn_enabled = not self.model_config.disable_cascade_attn
387
        self.is_mm_prefix_lm = self.model_config.is_mm_prefix_lm
388

389
        # Multi-modal data support
390
        self.mm_registry = MULTIMODAL_REGISTRY
391
        self.uses_mrope = model_config.uses_mrope
392
        self.uses_xdrope_dim = model_config.uses_xdrope_dim
393
        self.supports_mm_inputs = self.mm_registry.supports_multimodal_inputs(
394
            model_config
395
        )
396

397
398
399
        if self.model_config.is_encoder_decoder:
            # Maximum length of the encoder input, only for encoder-decoder
            # models.
400
            self.max_encoder_len = scheduler_config.max_num_encoder_input_tokens
401
402
403
        else:
            self.max_encoder_len = 0

404
405
406
        # Async scheduling
        self.use_async_scheduling = self.scheduler_config.async_scheduling

407
        # Sampler
408
        self.sampler = Sampler(logprobs_mode=self.model_config.logprobs_mode)
409

410
        self.eplb_state: EplbState | None = None
411
412
413
414
415
416
        """
        State of the expert parallelism load balancer.

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

417
        # Lazy initializations
418
        # self.model: nn.Module  # Set after load_model
419
        # Initialize in initialize_kv_cache
420
        self.kv_caches: list[torch.Tensor] = []
421
422
423
        # Initialize in initialize_kv_cache_tensors
        self.cross_layers_kv_cache: torch.Tensor | None = None
        self.cross_layers_attn_backend: type[AttentionBackend] | None = None
424
425
        # indexes: [kv_cache_group_id][attn_group]
        self.attn_groups: list[list[AttentionGroup]] = []
426
427
        # self.kv_cache_config: KVCacheConfig

428
429
        # mm_hash ->  encoder_output
        self.encoder_cache: dict[str, torch.Tensor] = {}
430

431
        self.use_aux_hidden_state_outputs = False
432
433
434
435
436
        # 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:
437
            self.drafter: (
438
439
440
441
442
                NgramProposer
                | SuffixDecodingProposer
                | EagleProposer
                | DraftModelProposer
                | MedusaProposer
443
            )
444
445
            if self.speculative_config.method == "ngram":
                self.drafter = NgramProposer(self.vllm_config)
446
447
448
449
450
451
            elif self.speculative_config.uses_draft_model():
                self.drafter = DraftModelProposer(
                    vllm_config=self.vllm_config,
                    device=self.device,
                    runner=self,
                )
452
453
            elif self.speculative_config.method == "suffix":
                self.drafter = SuffixDecodingProposer(self.vllm_config)
454
            elif self.speculative_config.use_eagle():
455
                self.drafter = EagleProposer(self.vllm_config, self.device, self)
456
                if self.speculative_config.method == "eagle3":
457
458
459
                    self.use_aux_hidden_state_outputs = (
                        self.drafter.eagle3_use_aux_hidden_state
                    )
460
461
            elif self.speculative_config.method == "medusa":
                self.drafter = MedusaProposer(
462
                    vllm_config=self.vllm_config, device=self.device
463
                )
464
            else:
465
466
467
468
                raise ValueError(
                    "Unknown speculative decoding method: "
                    f"{self.speculative_config.method}"
                )
469
            self.rejection_sampler = RejectionSampler(self.sampler)
470

471
472
473
        self.num_spec_tokens = 0
        if self.speculative_config:
            self.num_spec_tokens = self.speculative_config.num_speculative_tokens
474
475
476
477
478
            draft_config = self.speculative_config.draft_model_config
            if draft_config is not None and draft_config.max_model_len is not None:
                self.effective_drafter_max_model_len = draft_config.max_model_len
            else:
                self.effective_drafter_max_model_len = self.max_model_len
479

480
        # Request states.
481
        self.requests: dict[str, CachedRequestState] = {}
482
483
484
        # NOTE(rob): num_prompt_logprobs only includes reqs
        # that are currently in the prefill phase.
        self.num_prompt_logprobs: dict[str, int] = {}
485
        self.comm_stream = torch.cuda.Stream()
486

487
488
489
490
491
492
493
494
495
        # 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.
496
497
498
499
        logits_processors = model_config.logits_processors
        custom_logitsprocs: Sequence[str | type[LogitsProcessor]] = (
            tuple(logits_processors) if logits_processors is not None else ()
        )
500
501
        self.input_batch = InputBatch(
            max_num_reqs=self.max_num_reqs,
502
503
504
            # 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),
505
506
507
            max_num_batched_tokens=self.max_num_tokens,
            device=self.device,
            pin_memory=self.pin_memory,
508
            vocab_size=self.model_config.get_vocab_size(),
509
            block_sizes=[self.cache_config.block_size],
510
            kernel_block_sizes=[self.cache_config.block_size],
511
            is_spec_decode=bool(self.vllm_config.speculative_config),
512
            logitsprocs=build_logitsprocs(
513
514
515
                self.vllm_config,
                self.device,
                self.pin_memory,
516
                self.is_pooling_model,
517
                custom_logitsprocs,
518
            ),
519
520
521
            # 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),
522
            is_pooling_model=self.is_pooling_model,
523
            cp_kv_cache_interleave_size=self.parallel_config.cp_kv_cache_interleave_size,
524
        )
525

526
527
528
529
530
        # 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.
531
        self.prepare_inputs_event: torch.Event | None = None
532
533
        if self.use_async_scheduling:
            self.async_output_copy_stream = torch.cuda.Stream()
534
            self.prepare_inputs_event = torch.Event()
535

536
        # self.cudagraph_batch_sizes sorts in ascending order.
537
538
539
540
        if (
            self.compilation_config.cudagraph_capture_sizes
            and self.compilation_config.cudagraph_mode != CUDAGraphMode.NONE
        ):
541
542
            self.cudagraph_batch_sizes = sorted(
                self.compilation_config.cudagraph_capture_sizes
543
            )
544

545
        # Cache the device properties.
546
        self._init_device_properties()
547

548
        # Persistent buffers for CUDA graphs.
549
550
551
552
553
        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
        )
554
        self.seq_lens = self._make_buffer(self.max_num_reqs, dtype=torch.int32)
555
        self.encoder_seq_lens = self._make_buffer(self.max_num_reqs, dtype=torch.int32)
556
557
558
559
        if self.dcp_world_size > 1:
            self.dcp_local_seq_lens = self._make_buffer(
                self.max_num_reqs, dtype=torch.int32
            )
560
561
562
        # 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.
563
        self.inputs_embeds = self._make_buffer(
564
            self.max_num_tokens, self.inputs_embeds_size, dtype=self.dtype, numpy=False
565
566
        )
        self.is_token_ids = self._make_buffer(self.max_num_tokens, dtype=torch.bool)
567
568
        self.discard_request_mask = self._make_buffer(
            self.max_num_reqs, dtype=torch.bool
569
570
571
572
573
574
575
        )
        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
        )
576

577
578
        # Only relevant for multimodal models
        if self.supports_mm_inputs:
579
580
581
582
583
584
585
            # Double buffer to avoid race condition: previous iteration's async
            # copy may still be reading from CPU while current iteration writes.
            self.is_mm_embed_buffers = [
                self._make_buffer(self.max_num_tokens, dtype=torch.bool),
                self._make_buffer(self.max_num_tokens, dtype=torch.bool),
            ]
            self.is_mm_embed_idx = 0
586

587
        # Only relevant for models using M-RoPE (e.g, Qwen2-VL)
588
        if self.uses_mrope:
Roger Wang's avatar
Roger Wang committed
589
590
591
592
            # 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
593
594
595
596
597
598

            # 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
599
            self.mrope_positions = self._make_buffer(
600
601
                (3, self.max_num_tokens + 1), dtype=torch.int64
            )
602

603
604
605
606
607
608
609
        # 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
            )

610
        # None in the first PP rank. The rest are set after load_model.
611
        self.intermediate_tensors: IntermediateTensors | None = None
612

613
        # OPTIMIZATION: Cache the tensors rather than creating them every step.
614
        # Keep in int64 to avoid overflow with long context
615
616
617
618
        self.arange_np = np.arange(
            max(self.max_num_reqs + 1, self.max_model_len, self.max_num_tokens),
            dtype=np.int64,
        )
619

620
621
622
623
624
        # 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] = {}
625
626
627
628
629
        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(
630
631
                self.max_num_tokens, dtype=torch.int32, device=self.device
            )
632

633
        self.uniform_decode_query_len = 1 + self.num_spec_tokens
634
635
636
637

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

638
        self.mm_budget = (
639
            MultiModalBudget(self.vllm_config, self.mm_registry)
640
641
642
            if self.supports_mm_inputs
            else None
        )
643

644
        self.reorder_batch_threshold: int | None = None
645

646
647
648
649
650
        # 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()

651
        # Cached outputs.
652
        self._draft_token_ids: list[list[int]] | torch.Tensor | None = None
653
        self._draft_token_req_ids: list[str] | None = None
654
        self.transfer_event = torch.Event()
655
        self.sampled_token_ids_pinned_cpu = torch.empty(
656
            (self.max_num_reqs, 1),
657
658
            dtype=torch.int64,
            device="cpu",
659
660
            pin_memory=self.pin_memory,
        )
661

662
663
        # Pre-allocated tensor for copying valid sampled token counts to CPU,
        # with dedicated stream for overlapping and event for coordination.
664
        self.valid_sampled_token_count_event: torch.Event | None = None
665
        self.valid_sampled_token_count_copy_stream: torch.cuda.Stream | None = None
666
667
668
669
670
671
672
673
674
675
676
677
678
679
680
681
682
683
684
685
686
687
688
689
        # We also copy the drafted tokens to the CPU asynchronously,
        # in case we need them for structured outputs.
        self.draft_token_ids_event: torch.Event | None = None
        self.draft_token_ids_copy_stream: torch.cuda.Stream | None = None
        self.valid_sampled_token_count_cpu: torch.Tensor | None = None
        self.draft_token_ids_cpu: torch.Tensor | None = None
        if self.num_spec_tokens:
            self.draft_token_ids_event = torch.Event()
            self.draft_token_ids_copy_stream = torch.cuda.Stream()
            self.draft_token_ids_cpu = torch.empty(
                (self.max_num_reqs, self.num_spec_tokens),
                dtype=torch.int64,
                device="cpu",
                pin_memory=self.pin_memory,
            )
            if self.use_async_scheduling:
                self.valid_sampled_token_count_event = torch.Event()
                self.valid_sampled_token_count_copy_stream = torch.cuda.Stream()
                self.valid_sampled_token_count_cpu = torch.empty(
                    self.max_num_reqs,
                    dtype=torch.int64,
                    device="cpu",
                    pin_memory=self.pin_memory,
                )
690

691
692
        # Ephemeral state transferred between execute_model() and sample_tokens().
        self.execute_model_state: ExecuteModelState | None = None
693
        self.kv_connector_output: KVConnectorOutput | None = None
694
        self.mamba_state_idx: dict[str, int] = {}
695
        self.layerwise_nvtx_hooks_registered = False
696

697
698
699
700
701
702
703
    def update_max_model_len(self, max_model_len: int) -> None:
        self.max_model_len = max_model_len
        if self.speculative_config:
            draft_config = self.speculative_config.draft_model_config
            if draft_config is None or draft_config.max_model_len is None:
                self.effective_drafter_max_model_len = self.max_model_len

704
705
706
707
    def reset_mm_cache(self) -> None:
        if self.mm_budget:
            self.mm_budget.reset_cache()

708
709
710
711
712
713
714
715
716
717
718
719
720
721
722
723
724
725
726
727
728
729
730
731
732
733
734
735
736
737
738
739
740
741
742
743
744
745
746
747
748
749
750
751
    @torch.inference_mode()
    def init_fp8_kv_scales(self) -> None:
        """
        Re-initialize the KV cache and FP8 scales after waking from sleep.
        1. Zero out the KV cache tensors to remove garbage data from re-allocation.
        2. Reset Attention layer scaling factors (_k_scale, _v_scale) to 1.0.
          If these are left at 0.0 (default after wake_up), all KV cache values
          become effectively zero, causing gibberish output.
        """
        if not self.cache_config.cache_dtype.startswith("fp8"):
            return

        kv_caches = getattr(self, "kv_caches", [])
        for cache_tensor in kv_caches:
            if cache_tensor is not None:
                cache_tensor.zero_()

        k_attr_names = ("_k_scale", "k_scale")
        v_attr_names = ("_v_scale", "v_scale")

        attn_layers = self.compilation_config.static_forward_context
        for name, module in attn_layers.items():
            if isinstance(module, (Attention, MLAAttention)):
                # TODO: Generally, scale is 1.0 if user uses on-the-fly fp8
                # kvcache quant. However, to get better accuracy, compression
                # frameworks like llm-compressors allow users to tune the
                # scale. We may need to restore the specific calibrated scales
                # here in the future.
                k_scale_val, v_scale_val = 1.0, 1.0

                # Processing K Scale
                for attr in k_attr_names:
                    if hasattr(module, attr):
                        param = getattr(module, attr)
                        if isinstance(param, torch.Tensor):
                            param.fill_(k_scale_val)

                # Processing V Scale
                for attr in v_attr_names:
                    if hasattr(module, attr):
                        param = getattr(module, attr)
                        if isinstance(param, torch.Tensor):
                            param.fill_(v_scale_val)

752
753
754
755
    def _get_positions(self, num_tokens: Any):
        if isinstance(num_tokens, int):
            if self.uses_mrope:
                return self.mrope_positions.gpu[:, :num_tokens]
756
757
            if self.uses_xdrope_dim > 0:
                return self.xdrope_positions.gpu[:, :num_tokens]
758
759
760
761
            return self.positions.gpu[:num_tokens]
        else:
            if self.uses_mrope:
                return self.mrope_positions.gpu[:, num_tokens]
762
763
            if self.uses_xdrope_dim > 0:
                return self.xdrope_positions.gpu[:, num_tokens]
764
765
            return self.positions.gpu[num_tokens]

766
    def _make_buffer(
767
        self, *size: int | torch.SymInt, dtype: torch.dtype, numpy: bool = True
768
769
770
771
772
773
774
775
    ) -> CpuGpuBuffer:
        return CpuGpuBuffer(
            *size,
            dtype=dtype,
            device=self.device,
            pin_memory=self.pin_memory,
            with_numpy=numpy,
        )
776

777
    def _init_model_kwargs(self):
778
779
        model_kwargs = dict[str, Any]()

780
        if not self.is_pooling_model:
781
782
            return model_kwargs

783
784
        num_reqs = self.input_batch.num_reqs
        pooling_params = self.input_batch.get_pooling_params()
785
786
787

        token_type_id_requests = dict[int, Any]()
        for i, param in enumerate(pooling_params):
788
789
790
791
792
            if (
                param.extra_kwargs is not None
                and (token_types := param.extra_kwargs.get("compressed_token_type_ids"))
                is not None
            ):
793
794
795
796
797
                token_type_id_requests[i] = token_types

        if len(token_type_id_requests) == 0:
            return model_kwargs

798
        seq_lens = self.seq_lens.gpu[:num_reqs]
799
800
801
802
803
804
805
806
        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(
807
808
            device=self.device
        )
809
810
        return model_kwargs

811
    def _may_reorder_batch(self, scheduler_output: "SchedulerOutput") -> None:
812
813
        """
        Update the order of requests in the batch based on the attention
814
        backend's needs. For example, some attention backends (namely MLA) may
815
816
817
818
819
820
        want to separate requests based on if the attention computation will be
        compute-bound or memory-bound.

        Args:
            scheduler_output: The scheduler output.
        """
821
822
823
824
825
826
827
828
        # 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

829
830
831
832
        if self.reorder_batch_threshold is not None:
            reorder_batch_to_split_decodes_and_prefills(
                self.input_batch,
                scheduler_output,
833
834
                decode_threshold=self.reorder_batch_threshold,
            )
835

836
837
    # Note: used for model runner override.
    def _init_device_properties(self) -> None:
838
        """Initialize attributes from torch.cuda.get_device_properties"""
839
840
841
842
843
844
845
        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()

846
    def _update_states(self, scheduler_output: "SchedulerOutput") -> None:
847
848
849
850
851
852
        """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.

853
854
        The SamplingMetadata is updated and copied to the GPU if there is a
        new/resumed/paused/finished request in the batch.
855
856
        """
        # Remove finished requests from the cached states.
857
858
        for req_id in scheduler_output.finished_req_ids:
            self.requests.pop(req_id, None)
859
            self.num_prompt_logprobs.pop(req_id, None)
860
861
862
863
864
865
866
        # 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:
867
            self.input_batch.remove_request(req_id)
868
869

        # Free the cached encoder outputs.
870
871
        for mm_hash in scheduler_output.free_encoder_mm_hashes:
            self.encoder_cache.pop(mm_hash, None)
872

873
874
875
876
877
878
879
        # 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()
880
881
882
883
884
885
886
887
        resumed_req_ids = scheduler_output.scheduled_cached_reqs.resumed_req_ids
        # NOTE(zhuohan): cached_req_ids and resumed_req_ids are usually disjoint,
        # so `(scheduled_req_ids - resumed_req_ids) == scheduled_req_ids` holds
        # apart from the forced-preemption case in reset_prefix_cache. And in
        # that case we include the resumed_req_ids in the unscheduled set so
        # that they get cleared from the persistent batch before being re-scheduled
        # in the normal resumed request path.
        unscheduled_req_ids = cached_req_ids - (scheduled_req_ids - resumed_req_ids)
888
889
890
891
892
        # 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:
893
            self.input_batch.remove_request(req_id)
894

895
        reqs_to_add: list[CachedRequestState] = []
896
        # Add new requests to the cached states.
897
898
899
        for new_req_data in scheduler_output.scheduled_new_reqs:
            req_id = new_req_data.req_id
            sampling_params = new_req_data.sampling_params
900
            pooling_params = new_req_data.pooling_params
901

902
903
904
905
            if (
                sampling_params
                and sampling_params.sampling_type == SamplingType.RANDOM_SEED
            ):
906
907
908
909
910
                generator = torch.Generator(device=self.device)
                generator.manual_seed(sampling_params.seed)
            else:
                generator = None

911
912
            if self.is_pooling_model:
                assert pooling_params is not None
913
914
                task = pooling_params.task
                assert task is not None, "You did not set `task` in the API"
915

916
                model = cast(VllmModelForPooling, self.get_model())
917
                to_update = model.pooler.get_pooling_updates(task)
918
919
                to_update.apply(pooling_params)

920
            req_state = CachedRequestState(
921
                req_id=req_id,
922
                prompt_token_ids=new_req_data.prompt_token_ids,
923
                prompt_embeds=new_req_data.prompt_embeds,
924
                mm_features=new_req_data.mm_features,
925
                sampling_params=sampling_params,
926
                pooling_params=pooling_params,
927
                generator=generator,
928
929
                block_ids=new_req_data.block_ids,
                num_computed_tokens=new_req_data.num_computed_tokens,
930
                output_token_ids=[],
931
                lora_request=new_req_data.lora_request,
932
            )
933
934
            self.requests[req_id] = req_state

935
936
937
938
939
940
941
            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
                )

942
            # Only relevant for models using M-RoPE (e.g, Qwen2-VL)
943
            if self.uses_mrope:
944
                self._init_mrope_positions(req_state)
945

946
947
948
949
            # Only relevant for models using XD-RoPE (e.g, HunYuan-VL)
            if self.uses_xdrope_dim > 0:
                self._init_xdrope_positions(req_state)

950
            reqs_to_add.append(req_state)
951

952
        # Update the states of the running/resumed requests.
953
        is_last_rank = get_pp_group().is_last_rank
954
        req_data = scheduler_output.scheduled_cached_reqs
955
        scheduled_spec_tokens = scheduler_output.scheduled_spec_decode_tokens
956
957
958
959
960

        # 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()

961
        for i, req_id in enumerate(req_data.req_ids):
962
            req_state = self.requests[req_id]
963
964
            num_computed_tokens = req_data.num_computed_tokens[i]
            new_block_ids = req_data.new_block_ids[i]
965
            resumed_from_preemption = req_id in req_data.resumed_req_ids
966
            num_output_tokens = req_data.num_output_tokens[i]
967
            req_index = self.input_batch.req_id_to_index.get(req_id)
968

969
970
971
972
973
974
975
976
977
978
979
980
981
982
            if req_state.prev_num_draft_len and self.use_async_scheduling:
                # prev_num_draft_len is used in async scheduling mode with
                # spec decode. it indicates if need to update num_computed_tokens
                # of the request. for example:
                # fist step: num_computed_tokens = 0, spec_tokens = [],
                # prev_num_draft_len = 0.
                # second step: num_computed_tokens = 100(prompt lenth),
                # spec_tokens = [a,b], prev_num_draft_len = 0.
                # third step: num_computed_tokens = 100 + 2, spec_tokens = [c,d],
                # prev_num_draft_len = 2.
                # num_computed_tokens in first step and second step does't contain
                # the spec tokens length, but in third step it contains the
                # spec tokens length. we only need to update num_computed_tokens
                # when prev_num_draft_len > 0.
983
984
985
986
987
988
989
990
991
                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)
992

993
            # Update the cached states.
994
            req_state.num_computed_tokens = num_computed_tokens
995
996
997
998
999
1000
1001
1002

            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.
1003
1004
1005
                num_new_tokens = (
                    num_computed_tokens + len(new_token_ids) - req_state.num_tokens
                )
1006
1007
1008
1009
                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:
1010
                    req_state.output_token_ids.extend(new_token_ids[-num_new_tokens:])
1011
1012
1013
1014
1015
            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:
1016
1017
1018
1019
                    end_idx = (
                        self.input_batch.num_prompt_tokens[req_index]
                        + num_output_tokens
                    )
1020
                    self.input_batch.num_tokens_no_spec[req_index] = end_idx
1021

1022
            # Update the block IDs.
1023
            if not resumed_from_preemption:
1024
1025
                if new_block_ids is not None:
                    # Append the new blocks to the existing block IDs.
1026
                    for block_ids, new_ids in zip(req_state.block_ids, new_block_ids):
1027
                        block_ids.extend(new_ids)
1028
            else:
1029
                assert req_index is None
1030
                assert new_block_ids is not None
1031
1032
                # The request is resumed from preemption.
                # Replace the existing block IDs with the new ones.
1033
                req_state.block_ids = new_block_ids
1034
1035
1036
1037
1038

            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.
1039
1040
1041
1042
1043
1044
1045

                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:]

1046
                reqs_to_add.append(req_state)
1047
1048
1049
                continue

            # Update the persistent batch.
1050
            self.input_batch.num_computed_tokens_cpu[req_index] = num_computed_tokens
1051
            if new_block_ids is not None:
1052
                self.input_batch.block_table.append_row(new_block_ids, req_index)
1053
1054
1055
1056
1057
1058
1059

            # 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)
1060
                self.input_batch.token_ids_cpu[
1061
1062
1063
                    req_index, start_token_index:end_token_index
                ] = new_token_ids
                self.input_batch.num_tokens_no_spec[req_index] = end_token_index
1064

1065
            # Add spec_token_ids to token_ids_cpu.
1066
            self.input_batch.update_req_spec_token_ids(req_state, scheduled_spec_tokens)
1067

1068
1069
        # Add the new or resumed requests to the persistent batch.
        # The smaller empty indices are filled first.
1070
1071
        for request in reqs_to_add:
            self.input_batch.add_request(request)
1072
            self.input_batch.update_req_spec_token_ids(request, scheduled_spec_tokens)
1073

1074
1075
1076
1077
1078
1079
        # 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()
1080

1081
    def _update_states_after_model_execute(
1082
        self, output_token_ids: torch.Tensor, scheduler_output: "SchedulerOutput"
1083
    ) -> None:
1084
1085
1086
1087
1088
1089
1090
1091
        """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.
        """
1092
        if not self.speculative_config or not self.model_config.is_hybrid:
1093
1094
1095
            return

        # Find the number of accepted tokens for each sequence.
1096
1097
1098
1099
1100
1101
1102
1103
1104
1105
1106
1107
1108
1109
1110
1111
1112
1113
1114
1115
        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()
        )
1116
1117
        for i, num_tokens in enumerate(num_accepted_tokens):
            self.input_batch.num_accepted_tokens_cpu[i] = num_tokens
1118
1119
1120
1121
1122
1123
1124
1125
1126
1127
        if self.cache_config.mamba_cache_mode == "align":
            mamba_utils.postprocess_mamba(
                scheduler_output,
                self.kv_cache_config,
                self.input_batch,
                self.requests,
                self.mamba_state_idx,
                self.compilation_config.static_forward_context,
                self.model.get_mamba_state_copy_func(),
            )
1128

1129
    def _init_mrope_positions(self, req_state: CachedRequestState):
1130
1131
        model = self.get_model()
        assert supports_mrope(model), "M-RoPE support is not implemented."
1132
1133
1134
1135
        assert req_state.prompt_token_ids is not None, (
            "M-RoPE requires prompt_token_ids to be available."
        )
        mrope_model = cast(SupportsMRoPE, model)
1136
1137

        req_state.mrope_positions, req_state.mrope_position_delta = (
1138
            mrope_model.get_mrope_input_positions(
1139
                req_state.prompt_token_ids,
1140
                req_state.mm_features,
1141
            )
1142
        )
1143

1144
1145
1146
1147
1148
1149
1150
1151
1152
1153
1154
1155
1156
    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,
        )

1157
    def _extract_mm_kwargs(
1158
        self,
1159
1160
        scheduler_output: "SchedulerOutput",
    ) -> BatchedTensorInputs:
1161
        if not scheduler_output or not self.is_multimodal_raw_input_only_model:
1162
            return {}
1163

1164
1165
        mm_kwargs = list[MultiModalKwargsItem]()
        for req in scheduler_output.scheduled_new_reqs:
1166
1167
1168
            for feature in req.mm_features:
                if feature.data is not None:
                    mm_kwargs.append(feature.data)
1169

1170
1171
1172
        # Input all modalities at once
        mm_kwargs_combined: BatchedTensorInputs = {}
        for _, _, mm_kwargs_group in group_mm_kwargs_by_modality(
1173
1174
1175
            mm_kwargs,
            device=self.device,
            pin_memory=self.pin_memory,
1176
1177
        ):
            mm_kwargs_combined.update(mm_kwargs_group)
1178

1179
        return mm_kwargs_combined
1180

1181
    def _dummy_mm_kwargs(self, num_seqs: int) -> BatchedTensorInputs:
1182
        if not self.is_multimodal_raw_input_only_model:
1183
            return {}
1184

1185
1186
1187
1188
1189
        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)
1190

1191
1192
1193
    def _get_cumsum_and_arange(
        self,
        num_tokens: np.ndarray,
1194
        cumsum_dtype: np.dtype | None = None,
1195
1196
1197
1198
1199
1200
1201
1202
1203
1204
1205
1206
1207
1208
1209
1210
    ) -> 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

1211
    def _prepare_input_ids(
1212
1213
1214
1215
        self,
        scheduler_output: "SchedulerOutput",
        total_num_scheduled_tokens: int,
        cu_num_tokens: np.ndarray,
1216
    ) -> None:
1217
        """Prepare the input IDs for the current batch.
1218

1219
1220
1221
1222
1223
1224
1225
        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)
1226
1227
1228
            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)
1229
1230
1231
1232
1233
1234
1235
            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
1236
1237
1238
1239
        sample_flattened_indices: list[int] = []
        spec_flattened_indices: list[int] = []
        prev_common_req_indices: list[int] = []
        prev_draft_token_indices: list[int] = []
1240
1241
        indices_match = True
        max_flattened_index = -1
1242
1243
1244
        total_num_spec_tokens = 0
        scheduled_spec_tokens = scheduler_output.scheduled_spec_decode_tokens

1245
1246
1247
1248
1249
        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.
1250
1251
                draft_len = len(scheduled_spec_tokens.get(req_id, ()))
                total_num_spec_tokens += draft_len
1252
                flattened_index = cu_num_tokens[cur_index].item() - 1
1253
1254
1255
1256
1257
1258
1259
1260
1261
1262
1263
1264
1265
1266
1267
                # 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))
1268
                indices_match &= prev_index == flattened_index
1269
                max_flattened_index = max(max_flattened_index, flattened_index)
1270
1271
1272
        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:
1273
1274
1275
            # 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)
1276
1277
1278
            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)
1279
1280
        if num_commmon_tokens == 0:
            # No requests in common with the previous iteration
1281
            # So input_ids.cpu will have all the input ids.
1282
1283
1284
1285
1286
1287
1288
            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_(
1289
1290
1291
                self.input_batch.prev_sampled_token_ids[:num_commmon_tokens, 0],
                non_blocking=True,
            )
1292
1293
            if self.enable_prompt_embeds:
                self.is_token_ids.gpu[:num_commmon_tokens] = True
1294
            return
1295
        # Upload the index tensors asynchronously so the scatter can be non-blocking.
1296
1297
        sampled_tokens_index_tensor = torch.tensor(
            sample_flattened_indices, dtype=torch.int64, pin_memory=self.pin_memory
1298
        ).to(self.device, non_blocking=True)
1299
        prev_common_req_indices_tensor = torch.tensor(
1300
1301
            prev_common_req_indices, dtype=torch.int64, pin_memory=self.pin_memory
        ).to(self.device, non_blocking=True)
1302
1303
        self.input_ids.gpu.scatter_(
            dim=0,
1304
            index=sampled_tokens_index_tensor,
1305
            src=self.input_batch.prev_sampled_token_ids[
1306
1307
1308
                prev_common_req_indices_tensor, 0
            ],
        )
1309

1310
1311
1312
1313
1314
1315
1316
1317
1318
1319
1320
1321
1322
1323
1324
1325
1326
1327
1328
1329
1330
1331
        # Scatter the draft tokens after the sampled tokens are scattered.
        if self._draft_token_ids is None or not spec_flattened_indices:
            return

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

        # because input_ids dtype is torch.int32,
        # so convert draft_token_ids to torch.int32 here.
        draft_token_ids = self._draft_token_ids.to(dtype=torch.int32)

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

1332
1333
    def _get_encoder_seq_lens(
        self,
1334
        num_scheduled_tokens: dict[str, int],
1335
1336
        kv_cache_spec: KVCacheSpec,
        num_reqs: int,
1337
    ) -> tuple[torch.Tensor | None, np.ndarray | None]:
1338
        if not isinstance(kv_cache_spec, CrossAttentionSpec):
1339
            return None, None
1340

1341
1342
        # Zero out buffer for padding requests that are not actually scheduled (CGs)
        self.encoder_seq_lens.np[:num_reqs] = 0
1343
1344
        # Build encoder_seq_lens array mapping request indices to
        # encoder lengths for inputs scheduled in this batch
1345
        for req_id in num_scheduled_tokens:
1346
            req_index = self.input_batch.req_id_to_index[req_id]
1347
1348
1349
1350
1351
1352
1353
1354
1355
1356
1357
1358
1359
1360
1361
1362
            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]
1363

1364
        return encoder_seq_lens, encoder_seq_lens_cpu
1365

1366
    def _prepare_inputs(
1367
1368
1369
        self,
        scheduler_output: "SchedulerOutput",
        num_scheduled_tokens: np.ndarray,
1370
1371
    ) -> tuple[
        torch.Tensor,
1372
        SpecDecodeMetadata | None,
1373
    ]:
1374
1375
        """
        :return: tuple[
1376
            logits_indices, spec_decode_metadata,
1377
1378
        ]
        """
1379
1380
1381
1382
1383
1384
1385
        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.
1386
        self.input_batch.block_table.commit_block_table(num_reqs)
1387
1388
1389

        # Get request indices.
        # E.g., [2, 5, 3] -> [0, 0, 1, 1, 1, 1, 1, 2, 2, 2]
1390
        req_indices = np.repeat(self.arange_np[:num_reqs], num_scheduled_tokens)
1391

1392
1393
        # cu_num_tokens: [2, 5, 3] -> [2, 7, 10]
        # arange: [0, 1, 0, 1, 2, 3, 4, 0, 1, 2]
1394
        cu_num_tokens, arange = self._get_cumsum_and_arange(num_scheduled_tokens)
1395
1396

        # Get positions.
1397
        positions_np = self.positions.np[:total_num_scheduled_tokens]
1398
1399
1400
1401
1402
        np.add(
            self.input_batch.num_computed_tokens_cpu[req_indices],
            arange,
            out=positions_np,
        )
1403

1404
1405
        # Calculate M-RoPE positions.
        # Only relevant for models using M-RoPE (e.g, Qwen2-VL)
1406
        if self.uses_mrope:
1407
1408
            self._calc_mrope_positions(scheduler_output)

1409
1410
1411
1412
1413
        # 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)

1414
1415
1416
1417
        # 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.
1418
1419
1420
        token_indices = (
            positions_np + req_indices * self.input_batch.token_ids_cpu.shape[1]
        )
1421
        token_indices_tensor = torch.from_numpy(token_indices)
1422

1423
1424
1425
        # 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.
1426
1427
1428
1429
1430
1431
        torch.index_select(
            self.input_batch.token_ids_cpu_tensor.flatten(),
            0,
            token_indices_tensor,
            out=self.input_ids.cpu[:total_num_scheduled_tokens],
        )
1432
        if self.enable_prompt_embeds:
1433
            is_token_ids = self.input_batch.is_token_ids_tensor.flatten()
1434
1435
1436
1437
            torch.index_select(
                is_token_ids,
                0,
                token_indices_tensor,
1438
1439
                out=self.is_token_ids.cpu[:total_num_scheduled_tokens],
            )
1440
1441
1442
1443
1444
1445
1446
1447
1448
1449
1450
1451
1452
1453
1454
1455
1456
1457
1458
1459
1460
1461
1462
1463
1464
1465
1466
1467
1468
1469
1470
1471
1472

        # 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:
1473
1474
1475
                    self.inputs_embeds.cpu[
                        output_idx : output_idx + actual_num_sched
                    ].copy_(req_embeds[start_pos:actual_end])
1476
1477

                output_idx += num_sched
1478

1479
1480
        self.input_batch.block_table.compute_slot_mapping(req_indices, positions_np)
        self.input_batch.block_table.commit_slot_mapping(total_num_scheduled_tokens)
1481
1482

        # Prepare the attention metadata.
1483
        self.query_start_loc.np[0] = 0
1484
        self.query_start_loc.np[1 : num_reqs + 1] = cu_num_tokens
1485
1486
        # Note: pad query_start_loc to be non-decreasing, as kernels
        # like FlashAttention requires that
1487
        self.query_start_loc.np[num_reqs + 1 :].fill(cu_num_tokens[-1])
1488
        self.query_start_loc.copy_to_gpu()
1489
        query_start_loc = self.query_start_loc.gpu[: num_reqs + 1]
1490

1491
        self.seq_lens.np[:num_reqs] = (
1492
1493
            self.input_batch.num_computed_tokens_cpu[:num_reqs] + num_scheduled_tokens
        )
1494
        # Fill unused with 0 for full cuda graph mode.
1495
1496
        self.seq_lens.np[num_reqs:].fill(0)
        self.seq_lens.copy_to_gpu()
1497

1498
        num_tokens = [self.requests[r].num_tokens for r in self.input_batch.req_ids]
1499
1500
        num_tokens_np = np.array(num_tokens, dtype=np.int32)

1501
        # Record which requests should not be sampled,
1502
        # so that we could clear the sampled tokens before returning
1503
1504
        self.discard_request_mask.np[:num_reqs] = (
            self.seq_lens.np[:num_reqs] < num_tokens_np
1505
        )
1506
        self.discard_request_mask.copy_to_gpu(num_reqs)
1507

1508
        # Copy the tensors to the GPU.
1509
1510
1511
1512
1513
        self._prepare_input_ids(
            scheduler_output,
            total_num_scheduled_tokens,
            cu_num_tokens,
        )
1514

1515
        if self.uses_mrope:
1516
            # Only relevant for models using M-RoPE (e.g, Qwen2-VL)
1517
1518
            self.mrope_positions.gpu[:, :total_num_scheduled_tokens].copy_(
                self.mrope_positions.cpu[:, :total_num_scheduled_tokens],
1519
1520
                non_blocking=True,
            )
1521
1522
1523
1524
1525
1526
        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,
            )
1527
1528
        else:
            # Common case (1D positions)
1529
            self.positions.copy_to_gpu(total_num_scheduled_tokens)
1530

1531
        use_spec_decode = len(scheduler_output.scheduled_spec_decode_tokens) > 0
1532
1533
1534
1535
1536
1537
1538
1539
        if not use_spec_decode:
            # NOTE(woosuk): Due to chunked prefills, the batch may contain
            # partial requests. While we should not sample any token
            # from these partial requests, we do so for simplicity.
            # We will ignore the sampled tokens from the partial requests.
            # TODO: Support prompt logprobs.
            logits_indices = query_start_loc[1:] - 1
            spec_decode_metadata = None
1540
            num_sampled_tokens = np.ones(num_reqs, dtype=np.int32)
1541
1542
1543
1544
1545
        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)
1546
1547
1548
            # 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)
1549
1550
1551
1552
            for (
                req_id,
                draft_token_ids,
            ) in scheduler_output.scheduled_spec_decode_tokens.items():
1553
1554
                req_idx = self.input_batch.req_id_to_index[req_id]
                num_draft_tokens[req_idx] = len(draft_token_ids)
1555
1556
1557
1558
1559
                if (
                    self.input_batch.num_computed_tokens_cpu[req_idx]
                    >= self.input_batch.num_prompt_tokens[req_idx]
                ):
                    num_decode_draft_tokens[req_idx] = len(draft_token_ids)
1560
            spec_decode_metadata = self._calc_spec_decode_metadata(
1561
1562
                num_draft_tokens, cu_num_tokens
            )
1563
            logits_indices = spec_decode_metadata.logits_indices
1564
            num_sampled_tokens = num_draft_tokens + 1
1565
            # For DECODE only cuda graph of some attention backends (e.g., GDN).
1566
            self.num_decode_draft_tokens.np[:num_reqs] = num_decode_draft_tokens
1567
1568
            self.num_decode_draft_tokens.np[num_reqs:].fill(-1)
            self.num_decode_draft_tokens.copy_to_gpu()
1569

1570
1571
1572
1573
1574
        # Hot-Swap lora model
        if self.lora_config:
            assert (
                np.sum(num_sampled_tokens)
                <= self.vllm_config.scheduler_config.max_num_batched_tokens
1575
            )
1576
1577
1578
1579
1580
1581
1582
1583
1584
1585
1586
            self.set_active_loras(
                self.input_batch, num_scheduled_tokens, num_sampled_tokens
            )

        return (
            logits_indices,
            spec_decode_metadata,
        )

    def _build_attention_metadata(
        self,
1587
        num_tokens: int,
1588
        num_reqs: int,
1589
1590
1591
        max_query_len: int,
        num_tokens_padded: int | None = None,
        num_reqs_padded: int | None = None,
1592
1593
1594
1595
        ubatch_slices: UBatchSlices | None = None,
        logits_indices: torch.Tensor | None = None,
        use_spec_decode: bool = False,
        for_cudagraph_capture: bool = False,
1596
        num_scheduled_tokens: dict[str, int] | None = None,
1597
1598
1599
1600
1601
        cascade_attn_prefix_lens: list[list[int]] | None = None,
    ) -> tuple[PerLayerAttnMetadata, CommonAttentionMetadata | None]:
        """
        :return: tuple[attn_metadata, spec_decode_common_attn_metadata]
        """
1602
1603
1604
1605
        # Attention metadata is not needed for attention free models
        if len(self.kv_cache_config.kv_cache_groups) == 0:
            return {}, None

1606
1607
        num_tokens_padded = num_tokens_padded or num_tokens
        num_reqs_padded = num_reqs_padded or num_reqs
1608
        assert num_reqs_padded is not None and num_tokens_padded is not None
1609

1610
1611
1612
        attn_metadata: PerLayerAttnMetadata = {}
        if ubatch_slices is not None:
            attn_metadata = [dict() for _ in range(len(ubatch_slices))]
1613

1614
1615
1616
1617
1618
1619
1620
1621
        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()

1622
1623
        if use_spec_decode:
            self.num_accepted_tokens.np[:num_reqs] = (
1624
1625
                self.input_batch.num_accepted_tokens_cpu[:num_reqs]
            )
1626
1627
            self.num_accepted_tokens.np[num_reqs:].fill(1)
            self.num_accepted_tokens.copy_to_gpu()
1628

1629
        kv_cache_groups = self.kv_cache_config.kv_cache_groups
1630

1631
1632
1633
1634
        def _get_block_table_and_slot_mapping(kv_cache_gid: int):
            assert num_reqs_padded is not None and num_tokens_padded is not None
            kv_cache_spec = kv_cache_groups[kv_cache_gid].kv_cache_spec
            if isinstance(kv_cache_spec, EncoderOnlyAttentionSpec):
1635
                blk_table_tensor = torch.zeros(
1636
                    (num_reqs_padded, 1),
1637
                    dtype=torch.int32,
1638
1639
1640
                    device=self.device,
                )
                slot_mapping = torch.zeros(
1641
                    (num_tokens_padded,),
1642
1643
1644
                    dtype=torch.int64,
                    device=self.device,
                )
1645
            else:
1646
                blk_table = self.input_batch.block_table[kv_cache_gid]
1647
1648
                blk_table_tensor = blk_table.get_device_tensor(num_reqs_padded)
                slot_mapping = blk_table.slot_mapping.gpu[:num_tokens_padded]
1649

1650
1651
1652
1653
1654
1655
1656
1657
            # Fill unused with -1. Needed for reshape_and_cache in full cuda
            # graph mode. `blk_table_tensor` -1 to match mamba PAD_SLOT_ID
            slot_mapping[num_tokens:num_tokens_padded].fill_(-1)
            blk_table_tensor[num_reqs:num_reqs_padded].fill_(-1)

            return blk_table_tensor, slot_mapping

        block_table_gid_0, slot_mapping_gid_0 = _get_block_table_and_slot_mapping(0)
1658
1659
        if self.model_config.enable_return_routed_experts:
            self.slot_mapping = slot_mapping_gid_0[:num_tokens].cpu().numpy()
1660
1661
1662
1663
1664
1665
1666
1667
1668
1669
1670
1671
1672
1673
1674
1675
1676
1677
1678
1679
1680
1681
1682
1683
1684
1685
1686
1687
1688
1689
1690
1691
1692
1693
1694
1695
1696
1697
        cm_base = CommonAttentionMetadata(
            query_start_loc=self.query_start_loc.gpu[: num_reqs_padded + 1],
            query_start_loc_cpu=self.query_start_loc.cpu[: num_reqs_padded + 1],
            seq_lens=self.seq_lens.gpu[:num_reqs_padded],
            _seq_lens_cpu=self.seq_lens.cpu[:num_reqs_padded],
            _num_computed_tokens_cpu=self.input_batch.num_computed_tokens_cpu_tensor[
                :num_reqs_padded
            ],
            num_reqs=num_reqs_padded,
            num_actual_tokens=num_tokens_padded,
            max_query_len=max_query_len,
            max_seq_len=max_seq_len,
            block_table_tensor=block_table_gid_0,
            slot_mapping=slot_mapping_gid_0,
            causal=True,
        )

        if self.dcp_world_size > 1:
            self.dcp_local_seq_lens.cpu[:num_reqs] = get_dcp_local_seq_lens(
                self.seq_lens.cpu[:num_reqs],
                self.dcp_world_size,
                self.dcp_rank,
                self.parallel_config.cp_kv_cache_interleave_size,
            )
            self.dcp_local_seq_lens.cpu[num_reqs:].fill_(0)
            self.dcp_local_seq_lens.copy_to_gpu(num_reqs_padded)

            cm_base.dcp_local_seq_lens = self.dcp_local_seq_lens.gpu[:num_reqs_padded]
            cm_base.dcp_local_seq_lens_cpu = self.dcp_local_seq_lens.cpu[
                :num_reqs_padded
            ]

        if logits_indices is not None and self.cache_config.kv_sharing_fast_prefill:
            cm_base.num_logits_indices = logits_indices.size(0)
            cm_base.logits_indices_padded = self._prepare_kv_sharing_fast_prefill(
                logits_indices
            )

1698
1699
1700
1701
1702
1703
1704
1705
1706
        # Cache attention metadata builds across hybrid KV-cache groups
        # The only thing that changes between different hybrid KV-cache groups when the
        # same metadata builder and KVCacheSpec is the same is the block table, so we
        # can cache the attention metadata builds and just update the block table using
        # `builder.update_block_table` if the builder supports it.
        cached_attn_metadata: dict[
            tuple[KVCacheSpec, type[AttentionMetadataBuilder]], AttentionMetadata
        ] = {}

1707
1708
1709
1710
1711
1712
1713
        def _build_attn_group_metadata(
            kv_cache_gid: int,
            attn_gid: int,
            common_attn_metadata: CommonAttentionMetadata,
            ubid: int | None = None,
        ) -> None:
            attn_group = self.attn_groups[kv_cache_gid][attn_gid]
1714
            builder = attn_group.get_metadata_builder(ubid or 0)
1715
1716
1717
1718
            kv_cache_spec = kv_cache_groups[kv_cache_gid].kv_cache_spec
            if isinstance(kv_cache_spec, UniformTypeKVCacheSpecs):
                kv_cache_spec = kv_cache_spec.kv_cache_specs[attn_group.layer_names[0]]
            cache_key = (kv_cache_spec, type(builder))
1719

1720
1721
1722
1723
1724
1725
1726
1727
1728
1729
1730
1731
1732
1733
1734
1735
1736
1737
1738
1739
            cascade_attn_prefix_len = (
                cascade_attn_prefix_lens[kv_cache_gid][attn_gid]
                if cascade_attn_prefix_lens
                else 0
            )

            extra_attn_metadata_args = {}
            if use_spec_decode and isinstance(builder, GDNAttentionMetadataBuilder):
                assert ubid is None, "UBatching not supported with GDN yet"
                extra_attn_metadata_args = dict(
                    num_accepted_tokens=self.num_accepted_tokens.gpu[:num_reqs_padded],
                    num_decode_draft_tokens_cpu=self.num_decode_draft_tokens.cpu[
                        :num_reqs_padded
                    ],
                )

            if for_cudagraph_capture:
                attn_metadata_i = builder.build_for_cudagraph_capture(
                    common_attn_metadata
                )
1740
1741
1742
1743
1744
1745
1746
1747
1748
            elif (
                cache_key in cached_attn_metadata
                and builder.supports_update_block_table
            ):
                attn_metadata_i = builder.update_block_table(
                    cached_attn_metadata[cache_key],
                    common_attn_metadata.block_table_tensor,
                    common_attn_metadata.slot_mapping,
                )
1749
1750
1751
1752
1753
1754
            else:
                attn_metadata_i = builder.build(
                    common_prefix_len=cascade_attn_prefix_len,
                    common_attn_metadata=common_attn_metadata,
                    **extra_attn_metadata_args,
                )
1755
1756
                if builder.supports_update_block_table:
                    cached_attn_metadata[cache_key] = attn_metadata_i
1757
1758
1759
1760
1761
1762
1763
1764
1765
1766
1767
1768
1769
1770
1771
1772
1773
1774
1775
1776
1777
1778
1779

            if ubid is None:
                assert isinstance(attn_metadata, dict)
                attn_metadata_dict = attn_metadata
            else:
                assert isinstance(attn_metadata, list)
                attn_metadata_dict = attn_metadata[ubid]

            for layer_name in attn_group.layer_names:
                attn_metadata_dict[layer_name] = attn_metadata_i

        # Prepare the attention metadata for each KV cache group and make layers
        # in the same group share the same metadata.
        spec_decode_common_attn_metadata = None
        for kv_cache_gid, kv_cache_group in enumerate(kv_cache_groups):
            cm = copy(cm_base)  # shallow copy

            # Basically only the encoder seq_lens, block_table and slot_mapping change
            # for each kv_cache_group.
            cm.encoder_seq_lens, cm.encoder_seq_lens_cpu = self._get_encoder_seq_lens(
                num_scheduled_tokens or {},
                kv_cache_group.kv_cache_spec,
                num_reqs_padded,
1780
            )
1781
1782
1783
1784
            if kv_cache_gid > 0:
                cm.block_table_tensor, cm.slot_mapping = (
                    _get_block_table_and_slot_mapping(kv_cache_gid)
                )
1785

1786
            if self.speculative_config and spec_decode_common_attn_metadata is None:
1787
                if isinstance(self.drafter, EagleProposer):
1788
                    if self.drafter.attn_layer_names[0] in kv_cache_group.layer_names:
1789
                        spec_decode_common_attn_metadata = cm
1790
                else:
1791
                    spec_decode_common_attn_metadata = cm
1792

1793
            for attn_gid in range(len(self.attn_groups[kv_cache_gid])):
1794
                if ubatch_slices is not None:
1795
1796
1797
                    for ubid, _cm in enumerate(split_attn_metadata(ubatch_slices, cm)):
                        _build_attn_group_metadata(kv_cache_gid, attn_gid, _cm, ubid)

1798
                else:
1799
                    _build_attn_group_metadata(kv_cache_gid, attn_gid, cm)
1800

1801
1802
1803
1804
1805
1806
1807
1808
1809
1810
1811
1812
1813
1814
1815
1816
1817
1818
1819
1820
        if self.is_mm_prefix_lm:
            req_doc_ranges = {}
            for req_id in self.input_batch.req_ids:
                image_doc_ranges = []
                req_state = self.requests[req_id]
                for mm_feature in req_state.mm_features:
                    pos_info = mm_feature.mm_position
                    img_doc_range = pos_info.extract_embeds_range()
                    image_doc_ranges.extend(img_doc_range)
                req_idx = self.input_batch.req_id_to_index[req_id]
                req_doc_ranges[req_idx] = image_doc_ranges

            if isinstance(attn_metadata, list):
                for ub_metadata in attn_metadata:
                    for _metadata in ub_metadata.values():
                        _metadata.mm_prefix_range = req_doc_ranges  # type: ignore[attr-defined]
            else:
                for _metadata in attn_metadata.values():
                    _metadata.mm_prefix_range = req_doc_ranges  # type: ignore[attr-defined]

1821
1822
1823
1824
1825
1826
1827
1828
1829
1830
        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)
            )

1831
        return attn_metadata, spec_decode_common_attn_metadata
1832

1833
1834
1835
    def _compute_cascade_attn_prefix_lens(
        self,
        num_scheduled_tokens: np.ndarray,
1836
        num_computed_tokens: np.ndarray,
1837
1838
1839
1840
1841
1842
1843
        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
        """
1844

1845
1846
1847
1848
1849
1850
1851
1852
1853
1854
1855
1856
1857
1858
        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,
1859
                        num_computed_tokens,
1860
1861
1862
1863
1864
1865
1866
1867
                        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
1868

1869
1870
1871
    def _compute_cascade_attn_prefix_len(
        self,
        num_scheduled_tokens: np.ndarray,
1872
        num_computed_tokens: np.ndarray,
1873
        num_common_prefix_blocks: int,
1874
1875
        kv_cache_spec: KVCacheSpec,
        attn_metadata_builder: AttentionMetadataBuilder,
1876
1877
1878
1879
1880
1881
1882
1883
1884
1885
1886
1887
1888
1889
1890
1891
1892
1893
    ) -> 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.
        """
1894

1895
        common_prefix_len = num_common_prefix_blocks * kv_cache_spec.block_size
1896
1897
1898
1899
1900
1901
1902
1903
1904
1905
1906
1907
1908
1909
1910
1911
1912
1913
1914
1915
1916
1917
1918
1919
1920
1921
1922
1923
1924
1925
1926
1927
1928
1929
1930
1931
1932
        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]
1933
        # Request 3's num_computed_tokens: 3 (i.e., [A, B, C])
1934
1935
1936
1937
1938
        # 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.
1939
        common_prefix_len = min(common_prefix_len, num_computed_tokens.min())
1940
        # common_prefix_len should be a multiple of the block size.
1941
1942
1943
1944
1945
1946
1947
1948
1949
1950
1951
        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
        )
1952
1953
        assert isinstance(kv_cache_spec, AttentionSpec)
        use_cascade = attn_metadata_builder.use_cascade_attention(
1954
1955
1956
            common_prefix_len=common_prefix_len,
            query_lens=num_scheduled_tokens,
            num_query_heads=self.num_query_heads,
1957
            num_kv_heads=kv_cache_spec.num_kv_heads,
1958
            use_alibi=self.use_alibi,
1959
            use_sliding_window=use_sliding_window,
1960
            use_local_attention=use_local_attention,
1961
            num_sms=self.num_sms,
1962
            dcp_world_size=self.dcp_world_size,
1963
1964
1965
        )
        return common_prefix_len if use_cascade else 0

1966
1967
    def _calc_mrope_positions(self, scheduler_output: "SchedulerOutput"):
        mrope_pos_ptr = 0
1968
        for index, req_id in enumerate(self.input_batch.req_ids):
1969
1970
1971
            req = self.requests[req_id]
            assert req.mrope_positions is not None

1972
1973
            num_computed_tokens = self.input_batch.num_computed_tokens_cpu[index]
            num_scheduled_tokens = scheduler_output.num_scheduled_tokens[req_id]
1974
            num_prompt_tokens = length_from_prompt_token_ids_or_embeds(
1975
1976
                req.prompt_token_ids, req.prompt_embeds
            )
1977
1978

            if num_computed_tokens + num_scheduled_tokens > num_prompt_tokens:
1979
1980
                prompt_part_len = max(0, num_prompt_tokens - num_computed_tokens)
                completion_part_len = max(0, num_scheduled_tokens - prompt_part_len)
1981
1982
1983
1984
1985
1986
1987
1988
1989
1990
1991
1992
1993
            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

1994
1995
1996
                self.mrope_positions.cpu[:, dst_start:dst_end] = req.mrope_positions[
                    :, src_start:src_end
                ]
1997
1998
1999
2000
2001
2002
2003
                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

2004
                assert req.mrope_position_delta is not None
2005
                MRotaryEmbedding.get_next_input_positions_tensor(
2006
                    out=self.mrope_positions.np,
2007
2008
2009
2010
2011
                    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,
                )
2012
2013
2014

                mrope_pos_ptr += completion_part_len

2015
2016
2017
2018
2019
2020
2021
2022
2023
2024
2025
2026
2027
2028
2029
2030
2031
2032
2033
2034
2035
2036
2037
2038
2039
2040
2041
2042
2043
2044
2045
2046
2047
2048
2049
2050
2051
2052
2053
2054
2055
2056
2057
2058
2059
2060
2061
    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

2062
2063
    def _calc_spec_decode_metadata(
        self,
2064
2065
2066
2067
2068
2069
2070
2071
2072
2073
2074
2075
2076
2077
2078
2079
        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
2080
2081
2082
2083

        # 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(
2084
2085
            num_sampled_tokens, cumsum_dtype=np.int32
        )
2086
        # Step 2. [0, 0, 0, 0, 103, 104, 104, 104, 206, 207, 207]
2087
        logits_indices = np.repeat(
2088
2089
            cu_num_scheduled_tokens - num_sampled_tokens, num_sampled_tokens
        )
2090
        # Step 3. [0, 1, 2, 3, 103, 104, 105, 106, 206, 207, 208]
2091
2092
2093
2094
2095
2096
        logits_indices += arange

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

        # Compute the draft logits indices.
2097
2098
2099
        # 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(
2100
2101
            num_draft_tokens, cumsum_dtype=np.int32
        )
2102
2103
        # [0, 0, 0, 5, 5, 9]
        target_logits_indices = np.repeat(
2104
2105
            cu_num_sampled_tokens - num_sampled_tokens, num_draft_tokens
        )
2106
2107
2108
2109
2110
        # [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(
2111
2112
            self.device, non_blocking=True
        )
2113
2114
2115
        cu_num_sampled_tokens = torch.from_numpy(cu_num_sampled_tokens).to(
            self.device, non_blocking=True
        )
2116
2117
2118
        logits_indices = torch.from_numpy(logits_indices).to(
            self.device, non_blocking=True
        )
2119
        target_logits_indices = torch.from_numpy(target_logits_indices).to(
2120
2121
            self.device, non_blocking=True
        )
2122
        bonus_logits_indices = torch.from_numpy(bonus_logits_indices).to(
2123
2124
            self.device, non_blocking=True
        )
2125

2126
2127
        # Compute the draft token ids.
        # draft_token_indices:      [  1,   2,   3, 105, 106, 208]
2128
        draft_token_ids = self.input_ids.gpu[logits_indices]
2129
2130
        draft_token_ids = draft_token_ids[target_logits_indices + 1]

2131
        return SpecDecodeMetadata(
2132
2133
2134
            draft_token_ids=draft_token_ids,
            num_draft_tokens=num_draft_tokens.tolist(),
            cu_num_draft_tokens=cu_num_draft_tokens,
2135
            cu_num_sampled_tokens=cu_num_sampled_tokens,
2136
2137
2138
2139
2140
            target_logits_indices=target_logits_indices,
            bonus_logits_indices=bonus_logits_indices,
            logits_indices=logits_indices,
        )

2141
2142
2143
2144
2145
2146
2147
    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
2148
        self.kv_sharing_fast_prefill_logits_indices[:num_logits].copy_(logits_indices)
2149
2150
2151
2152
2153
        # 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_(
2154
2155
            logits_indices[-1].item()
        )
2156
2157
2158
2159
2160
        # Dispatch for the decoder portion of the model.
        _, batch_desc = self.cudagraph_dispatcher.dispatch(
            num_logits, disable_full=True
        )
        num_logits_padded = batch_desc.num_tokens
2161
2162
2163
        logits_indices_padded = self.kv_sharing_fast_prefill_logits_indices[
            :num_logits_padded
        ]
2164
2165
        return logits_indices_padded

2166
    def _batch_mm_inputs_from_scheduler(
2167
2168
        self,
        scheduler_output: "SchedulerOutput",
2169
2170
2171
2172
2173
    ) -> tuple[
        list[str],
        list[MultiModalKwargsItem],
        list[tuple[str, PlaceholderRange]],
    ]:
2174
        """Batch multimodal inputs from scheduled encoder inputs.
2175
2176
2177

        Args:
            scheduler_output: The scheduler output containing scheduled encoder
2178
                inputs.
2179
2180

        Returns:
2181
            A tuple of (mm_hashes, mm_kwargs, mm_lora_refs) where:
2182
2183
            - mm_hashes: List of multimodal hashes for each item
            - mm_kwargs: List of multimodal kwargs for each item
2184
            - mm_lora_refs: List of (req_id, placeholder_range) for each item
2185
        """
2186
2187
        scheduled_encoder_inputs = scheduler_output.scheduled_encoder_inputs
        if not scheduled_encoder_inputs:
2188
            return [], [], []
2189
2190

        mm_hashes = list[str]()
2191
        mm_kwargs = list[MultiModalKwargsItem]()
2192
2193
2194
        # Multimodal LoRA reference info to map each multimodal item
        # back to its request & position
        mm_lora_refs = list[tuple[str, PlaceholderRange]]()
2195
2196
        for req_id, encoder_input_ids in scheduled_encoder_inputs.items():
            req_state = self.requests[req_id]
2197
2198

            for mm_input_id in encoder_input_ids:
2199
                mm_feature = req_state.mm_features[mm_input_id]
2200
2201
                if mm_feature.data is None:
                    continue
2202
2203

                mm_hashes.append(mm_feature.identifier)
2204
                mm_kwargs.append(mm_feature.data)
2205
                mm_lora_refs.append((req_id, mm_feature.mm_position))
2206

2207
        return mm_hashes, mm_kwargs, mm_lora_refs
2208

2209
2210
2211
    def _execute_mm_encoder(
        self, scheduler_output: "SchedulerOutput"
    ) -> list[torch.Tensor]:
2212
2213
2214
        mm_hashes, mm_kwargs, mm_lora_refs = self._batch_mm_inputs_from_scheduler(
            scheduler_output
        )
2215
2216

        if not mm_kwargs:
2217
            return []
2218

2219
2220
2221
2222
2223
2224
2225
        # 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.
2226
        model = cast(SupportsMultiModal, self.model)
2227
2228
2229
2230
2231
2232
2233
2234
2235
2236
2237
2238
2239
2240
2241
2242
2243
2244
2245
2246
2247
2248
2249
2250
2251
2252
2253
2254
2255
2256
2257
2258
2259
2260
2261
2262
2263
2264
2265
2266
2267
2268
2269
2270
2271
2272
2273
2274
2275
2276
2277
2278
2279
2280
2281
2282
2283

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

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

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

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

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

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

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

                self.lora_manager.set_active_adapters(
                    lora_requests,
                    connector_mapping,
                )

2284
        encoder_outputs: list[torch.Tensor] = []
2285
        for modality, num_items, mm_kwargs_group in group_mm_kwargs_by_modality(
2286
2287
2288
            mm_kwargs,
            device=self.device,
            pin_memory=self.pin_memory,
2289
        ):
2290
            curr_group_outputs: MultiModalEmbeddings
2291
2292

            # EVS-related change.
2293
            # (ekhvedchenia): Temporary hack to limit peak memory usage when
2294
            # processing multimodal data. This solves the issue with scheduler
2295
2296
2297
2298
            # 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)
2299
2300
2301
2302
2303
2304
2305
            # 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
            ):
2306
                curr_group_outputs_lst = list[torch.Tensor]()
2307
2308
2309
2310
2311
2312
2313
2314
2315
                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,
                        )
2316
                    )
2317

2318
                    micro_batch_outputs = model.embed_multimodal(
2319
2320
                        **micro_batch_mm_inputs
                    )
2321

2322
2323
2324
                    curr_group_outputs_lst.extend(micro_batch_outputs)

                curr_group_outputs = curr_group_outputs_lst
2325
2326
2327
2328
2329
2330
2331
2332
            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.
2333
                curr_group_outputs = model.embed_multimodal(**mm_kwargs_group)
2334

2335
2336
            sanity_check_mm_encoder_outputs(
                curr_group_outputs,
2337
                expected_num_items=num_items,
2338
            )
2339
            encoder_outputs.extend(curr_group_outputs)
2340

2341
        # Cache the encoder outputs by mm_hash
2342
        for mm_hash, output in zip(mm_hashes, encoder_outputs):
2343
            self.encoder_cache[mm_hash] = output
2344
2345
            logger.debug("Finish execute for mm hash %s", mm_hash)
            self.maybe_save_ec_to_connector(self.encoder_cache, mm_hash)
2346

2347
2348
        return encoder_outputs

2349
    def _gather_mm_embeddings(
2350
2351
        self,
        scheduler_output: "SchedulerOutput",
2352
        shift_computed_tokens: int = 0,
2353
2354
2355
    ) -> tuple[list[torch.Tensor], torch.Tensor]:
        total_num_scheduled_tokens = scheduler_output.total_num_scheduled_tokens

2356
2357
2358
2359
2360
        # Swap to the other buffer to avoid race condition with previous
        # iteration's async copy that may still be reading from CPU.
        self.is_mm_embed_idx = 1 - self.is_mm_embed_idx
        is_mm_embed_buf = self.is_mm_embed_buffers[self.is_mm_embed_idx]

2361
        mm_embeds = list[torch.Tensor]()
2362
        is_mm_embed = is_mm_embed_buf.cpu
2363
2364
2365
        is_mm_embed[:total_num_scheduled_tokens] = False

        req_start_idx = 0
2366
        should_sync_mrope_positions = False
2367
        should_sync_xdrope_positions = False
2368

2369
        for req_id in self.input_batch.req_ids:
2370
2371
            mm_embeds_req: list[torch.Tensor] = []

2372
            num_scheduled_tokens = scheduler_output.num_scheduled_tokens[req_id]
2373
            req_state = self.requests[req_id]
2374
            num_computed_tokens = req_state.num_computed_tokens + shift_computed_tokens
2375

2376
2377
            for mm_feature in req_state.mm_features:
                pos_info = mm_feature.mm_position
2378
2379
                start_pos = pos_info.offset
                num_encoder_tokens = pos_info.length
2380
2381
2382
2383
2384
2385
2386
2387
2388
2389
2390
2391
2392
2393
2394
2395

                # 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,
2396
2397
                    num_encoder_tokens,
                )
2398
                assert start_idx < end_idx
2399
2400
2401
2402
2403
2404
2405
                curr_embeds_start, curr_embeds_end = (
                    pos_info.get_embeds_indices_in_range(start_idx, end_idx)
                )
                # If there are no embeddings in the current range, we skip
                # gathering the embeddings.
                if curr_embeds_start == curr_embeds_end:
                    continue
2406

2407
                mm_hash = mm_feature.identifier
2408
                encoder_output = self.encoder_cache.get(mm_hash, None)
2409
                assert encoder_output is not None, f"Encoder cache miss for {mm_hash}."
2410
2411
2412

                if (is_embed := pos_info.is_embed) is not None:
                    is_embed = is_embed[start_idx:end_idx]
2413
2414
2415
                    mm_embeds_item = encoder_output[curr_embeds_start:curr_embeds_end]
                else:
                    mm_embeds_item = encoder_output[start_idx:end_idx]
2416

2417
                req_start_pos = req_start_idx + start_pos - num_computed_tokens
2418
2419
2420
                is_mm_embed[req_start_pos + start_idx : req_start_pos + end_idx] = (
                    True if is_embed is None else is_embed
                )
2421
2422
2423
                mm_embeds_req.append(mm_embeds_item)

            if self.is_multimodal_pruning_enabled and self.uses_mrope:
2424
                assert req_state.mrope_positions is not None
2425
2426
2427
2428
2429
2430
2431
                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,
2432
2433
                    )
                )
2434
2435
2436
2437
                req_state.mrope_positions.copy_(new_mrope_positions)
                req_state.mrope_position_delta = new_delta

            mm_embeds.extend(mm_embeds_req)
2438
2439
            req_start_idx += num_scheduled_tokens

2440
        is_mm_embed = is_mm_embed_buf.copy_to_gpu(total_num_scheduled_tokens)
2441
2442
2443

        if should_sync_mrope_positions:
            self._calc_mrope_positions(scheduler_output)
2444
            self.mrope_positions.copy_to_gpu(total_num_scheduled_tokens)
2445

2446
2447
2448
2449
        if should_sync_xdrope_positions:
            self._calc_xdrope_positions(scheduler_output)
            self.xdrope_positions.copy_to_gpu(total_num_scheduled_tokens)

2450
        return mm_embeds, is_mm_embed
2451

2452
    def get_model(self) -> nn.Module:
2453
        # get raw model out of the cudagraph wrapper.
2454
        if isinstance(self.model, (CUDAGraphWrapper, UBatchWrapper)):
2455
            return self.model.unwrap()
2456
2457
        return self.model

2458
2459
2460
2461
2462
2463
2464
2465
2466
2467
2468
2469
2470
2471
2472
    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

2473
2474
2475
2476
2477
    def get_supported_pooling_tasks(self) -> list[PoolingTask]:
        model = self.get_model()
        if not is_pooling_model(model):
            return []

2478
2479
        supported_tasks = list(model.pooler.get_supported_tasks())

2480
2481
2482
2483
        if "score" in supported_tasks:
            num_labels = getattr(self.model_config.hf_config, "num_labels", 0)
            if num_labels != 1:
                supported_tasks.remove("score")
2484
                logger.debug_once("Score API is only enabled for num_labels == 1.")
2485
2486

        return supported_tasks
2487

2488
2489
2490
2491
2492
2493
2494
2495
2496
2497
    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)

2498
    def sync_and_slice_intermediate_tensors(
2499
2500
        self,
        num_tokens: int,
2501
        intermediate_tensors: IntermediateTensors | None,
2502
2503
        sync_self: bool,
    ) -> IntermediateTensors:
2504
2505
2506
        assert self.intermediate_tensors is not None

        tp = self.vllm_config.parallel_config.tensor_parallel_size
2507
        is_rs = is_residual_scattered_for_sp(self.vllm_config, num_tokens)
2508
2509
2510
2511
2512
2513

        # 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():
2514
                is_scattered = k == "residual" and is_rs
2515
                copy_len = num_tokens // tp if is_scattered else num_tokens
2516
                self.intermediate_tensors[k][:copy_len].copy_(
2517
2518
2519
2520
2521
2522
2523
2524
2525
2526
2527
2528
2529
                    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:
2530
2531
2532
2533
2534
2535
2536
        """
        Step for the EPLB (Expert Parallelism Load Balancing) state.
        """
        if not self.parallel_config.enable_eplb:
            return

        assert self.eplb_state is not None
2537
2538
        model = self.get_model()
        assert is_mixture_of_experts(model)
2539
2540
2541
        self.eplb_state.step(
            is_dummy,
            is_profile,
2542
            log_stats=self.parallel_config.eplb_config.log_balancedness,
2543
2544
        )

2545
2546
2547
2548
2549
    def _pool(
        self,
        hidden_states: torch.Tensor,
        num_scheduled_tokens: int,
        num_scheduled_tokens_np: np.ndarray,
2550
2551
2552
2553
        kv_connector_output: KVConnectorOutput | None,
    ) -> ModelRunnerOutput | AsyncModelRunnerOutput:
        num_reqs = self.input_batch.num_reqs
        assert num_reqs == len(self.input_batch.pooling_params), (
2554
2555
            "Either all or none of the requests in a batch must be pooling request"
        )
2556

2557
        hidden_states = hidden_states[:num_scheduled_tokens]
2558
        seq_lens_cpu = self.seq_lens.cpu[:num_reqs]
2559

2560
        pooling_metadata = self.input_batch.get_pooling_metadata()
2561
        pooling_metadata.build_pooling_cursor(
2562
            num_scheduled_tokens_np, seq_lens_cpu, device=hidden_states.device
2563
        )
2564

2565
2566
        model = cast(VllmModelForPooling, self.model)
        raw_pooler_output: PoolerOutput = model.pooler(
2567
            hidden_states=hidden_states, pooling_metadata=pooling_metadata
2568
        )
2569
2570
2571
2572
2573
2574
2575
2576
2577
2578
2579
2580
2581
2582
2583
2584
2585
2586
2587
2588
2589
2590
2591
2592

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

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

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

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

2593
        raw_pooler_output = json_map_leaves(
2594
            lambda x: None if x is None else x.to("cpu", non_blocking=True),
2595
2596
            raw_pooler_output,
        )
2597
2598
2599
2600
        model_runner_output.pooler_output = [
            out if include else None
            for out, include in zip(raw_pooler_output, finished_mask)
        ]
2601
2602
        self._sync_device()

2603
        return model_runner_output
2604

2605
    def _pad_for_sequence_parallelism(self, num_scheduled_tokens: int) -> int:
2606
2607
2608
        # Pad tokens to multiple of tensor_parallel_size when
        # enabled collective fusion for SP
        tp_size = self.vllm_config.parallel_config.tensor_parallel_size
2609
        if self.compilation_config.pass_config.enable_sp and tp_size > 1:
2610
2611
2612
            return round_up(num_scheduled_tokens, tp_size)
        return num_scheduled_tokens

Patrick von Platen's avatar
Patrick von Platen committed
2613
2614
2615
2616
2617
2618
2619
2620
2621
2622
2623
    def _prepare_mm_inputs(
        self, num_tokens: int
    ) -> tuple[torch.Tensor | None, torch.Tensor]:
        if self.model.requires_raw_input_tokens:
            input_ids = self.input_ids.gpu[:num_tokens]
        else:
            input_ids = None

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

2624
    def _preprocess(
2625
2626
        self,
        scheduler_output: "SchedulerOutput",
2627
        num_input_tokens: int,  # Padded
2628
        intermediate_tensors: IntermediateTensors | None = None,
2629
    ) -> tuple[
2630
2631
        torch.Tensor | None,
        torch.Tensor | None,
2632
        torch.Tensor,
2633
        IntermediateTensors | None,
2634
        dict[str, Any],
2635
        ECConnectorOutput | None,
2636
    ]:
2637
        num_scheduled_tokens = scheduler_output.total_num_scheduled_tokens
2638
        is_first_rank = get_pp_group().is_first_rank
2639
        is_encoder_decoder = self.model_config.is_encoder_decoder
2640

2641
2642
        # _prepare_inputs may reorder the batch, so we must gather multi
        # modal outputs after that to ensure the correct order
2643
2644
        ec_connector_output = None

2645
        if self.supports_mm_inputs and is_first_rank and not is_encoder_decoder:
2646
            # Run the multimodal encoder if any.
2647
2648
2649
2650
2651
2652
            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)
2653

2654
2655
2656
            # 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.
2657
            inputs_embeds_scheduled = self.model.embed_input_ids(
2658
2659
2660
                self.input_ids.gpu[:num_scheduled_tokens],
                multimodal_embeddings=mm_embeds,
                is_multimodal=is_mm_embed,
2661
            )
2662

2663
            # TODO(woosuk): Avoid the copy. Optimize.
2664
            self.inputs_embeds.gpu[:num_scheduled_tokens].copy_(inputs_embeds_scheduled)
2665

Patrick von Platen's avatar
Patrick von Platen committed
2666
            input_ids, inputs_embeds = self._prepare_mm_inputs(num_input_tokens)
2667
            model_kwargs = {
2668
                **self._init_model_kwargs(),
2669
2670
                **self._extract_mm_kwargs(scheduler_output),
            }
2671
        elif self.enable_prompt_embeds and is_first_rank:
2672
2673
2674
2675
2676
2677
2678
2679
2680
2681
2682
2683
            # 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).
2684
2685
2686
            token_ids_idx = (
                self.is_token_ids.gpu[:num_scheduled_tokens]
                .nonzero(as_tuple=False)
2687
                .squeeze(1)
2688
            )
2689
2690
2691
            # Some tokens ids may need to become embeds
            if token_ids_idx.numel() > 0:
                token_ids = self.input_ids.gpu[token_ids_idx]
2692
                tokens_to_embeds = self.model.embed_input_ids(input_ids=token_ids)
2693
2694
2695
                self.inputs_embeds.gpu[token_ids_idx] = tokens_to_embeds

            inputs_embeds = self.inputs_embeds.gpu[:num_input_tokens]
2696
            model_kwargs = self._init_model_kwargs()
2697
            input_ids = None
2698
        else:
2699
2700
2701
2702
            # 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.
2703
            input_ids = self.input_ids.gpu[:num_input_tokens]
2704
            inputs_embeds = None
2705
            model_kwargs = self._init_model_kwargs()
2706

2707
        if self.uses_mrope:
2708
            positions = self.mrope_positions.gpu[:, :num_input_tokens]
2709
2710
        elif self.uses_xdrope_dim > 0:
            positions = self.xdrope_positions.gpu[:, :num_input_tokens]
2711
        else:
2712
            positions = self.positions.gpu[:num_input_tokens]
2713

2714
        if is_first_rank:
2715
2716
            intermediate_tensors = None
        else:
2717
            assert intermediate_tensors is not None
2718
            intermediate_tensors = self.sync_and_slice_intermediate_tensors(
2719
2720
                num_input_tokens, intermediate_tensors, True
            )
2721

2722
        if is_encoder_decoder and scheduler_output.scheduled_encoder_inputs:
2723
2724
2725
2726
2727
2728
2729
            # 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})
2730

2731
2732
2733
2734
2735
2736
        return (
            input_ids,
            inputs_embeds,
            positions,
            intermediate_tensors,
            model_kwargs,
2737
            ec_connector_output,
2738
        )
2739

2740
    def _sample(
2741
        self,
2742
2743
        logits: torch.Tensor | None,
        spec_decode_metadata: SpecDecodeMetadata | None,
2744
    ) -> SamplerOutput:
2745
        # Sample the next token and get logprobs if needed.
2746
        sampling_metadata = self.input_batch.sampling_metadata
2747
2748
2749
        # 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()
2750
        if spec_decode_metadata is None:
2751
            return self.sampler(
2752
2753
2754
                logits=logits,
                sampling_metadata=sampling_metadata,
            )
2755

2756
2757
2758
2759
2760
2761
        # Update spec_token_ids with real draft tokens from pre step only when
        # output_token_ids is needed (penalties or bad_words are in use).
        if self.use_async_scheduling and self._draft_token_req_ids is not None:
            draft_token_ids_cpu, _ = self._get_draft_token_ids_cpu()
            self.input_batch.update_async_spec_token_ids(draft_token_ids_cpu)

2762
        sampler_output = self.rejection_sampler(
2763
2764
            spec_decode_metadata,
            None,  # draft_probs
2765
            logits,
2766
2767
            sampling_metadata,
        )
2768
2769
2770
        return sampler_output

    def _bookkeeping_sync(
2771
2772
2773
        self,
        scheduler_output: "SchedulerOutput",
        sampler_output: SamplerOutput,
2774
        logits: torch.Tensor | None,
2775
2776
        hidden_states: torch.Tensor,
        num_scheduled_tokens: int,
2777
        spec_decode_metadata: SpecDecodeMetadata | None,
2778
    ) -> tuple[
2779
        dict[str, int],
2780
        LogprobsLists | None,
2781
        list[list[int]],
2782
        dict[str, LogprobsTensors | None],
2783
2784
2785
        list[str],
        dict[str, int],
        list[int],
2786
    ]:
2787
2788
2789
2790
        num_nans_in_logits = {}
        if envs.VLLM_COMPUTE_NANS_IN_LOGITS:
            num_nans_in_logits = self._get_nans_in_logits(logits)

2791
2792
2793
2794
        num_reqs = self.input_batch.num_reqs
        discard_sampled_tokens_req_indices = np.nonzero(
            self.discard_request_mask.np[:num_reqs]
        )[0]
2795
2796
2797
2798
        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)
2799

2800
2801
2802
        # 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()
2803
        req_id_to_index_output_copy = self.input_batch.req_id_to_index.copy()
2804
2805

        num_sampled_tokens = sampler_output.sampled_token_ids.shape[0]
2806
        sampled_token_ids = sampler_output.sampled_token_ids
2807
        logprobs_tensors = sampler_output.logprobs_tensors
2808
        invalid_req_indices = []
2809
        logprobs_lists = None
2810
2811
2812
2813
2814
2815
        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)
2816
2817
2818
                # 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()
2819
2820
2821

                if logprobs_tensors is not None:
                    logprobs_lists = logprobs_tensors.tolists()
2822
2823
            else:
                # Includes spec decode tokens.
2824
                valid_sampled_token_ids, logprobs_lists = RejectionSampler.parse_output(
2825
2826
                    sampled_token_ids,
                    self.input_batch.vocab_size,
2827
                    discard_sampled_tokens_req_indices,
2828
                    logprobs_tensors=logprobs_tensors,
2829
                )
2830
        else:
2831
            valid_sampled_token_ids = []
2832
            invalid_req_indices = discard_sampled_tokens_req_indices.tolist()
2833
2834
2835
2836
2837
            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.
2838
2839
2840
2841
            # 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
2842
2843
2844
2845
2846
            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
            }
2847

2848
2849
2850
2851
2852
        # 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.
2853
        req_ids = self.input_batch.req_ids
2854
2855
        for req_idx in range(num_sampled_tokens):
            if self.use_async_scheduling:
2856
                sampled_ids = [-1] if req_idx not in invalid_req_indices_set else None
2857
2858
            else:
                sampled_ids = valid_sampled_token_ids[req_idx]
2859

2860
            num_sampled_ids: int = len(sampled_ids) if sampled_ids else 0
2861

2862
            if not sampled_ids:
2863
2864
2865
                continue

            start_idx = self.input_batch.num_tokens_no_spec[req_idx]
2866
            end_idx = start_idx + num_sampled_ids
2867
2868
2869
2870
            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}"
2871
            )
2872

2873
2874
            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
2875
            self.input_batch.num_tokens_no_spec[req_idx] = end_idx
2876

2877
            req_id = req_ids[req_idx]
2878
2879
2880
            req_state = self.requests[req_id]
            req_state.output_token_ids.extend(sampled_ids)

2881
2882
2883
2884
2885
2886
        # Compute prompt logprobs if needed.
        prompt_logprobs_dict = self._get_prompt_logprobs_dict(
            hidden_states[:num_scheduled_tokens],
            scheduler_output.num_scheduled_tokens,
        )

2887
2888
2889
2890
2891
2892
2893
2894
2895
2896
        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,
        )

2897
2898
2899
2900
2901
2902
2903
2904
2905
2906
2907
2908
2909
2910
2911
    @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()

2912
2913
    def _model_forward(
        self,
2914
2915
2916
2917
        input_ids: torch.Tensor | None = None,
        positions: torch.Tensor | None = None,
        intermediate_tensors: IntermediateTensors | None = None,
        inputs_embeds: torch.Tensor | None = None,
2918
2919
2920
2921
2922
        **model_kwargs: dict[str, Any],
    ) -> Any:
        """Helper method to call the model forward pass.

        This method can be overridden by subclasses for model execution.
2923
        Motivation: We can inspect only this method versus
2924
2925
2926
2927
2928
2929
2930
2931
2932
2933
2934
2935
2936
2937
2938
2939
2940
2941
2942
2943
        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,
        )

2944
2945
2946
2947
2948
2949
2950
2951
2952
2953
2954
2955
2956
2957
2958
2959
2960
2961
2962
2963
2964
    @staticmethod
    def _is_uniform_decode(
        max_num_scheduled_tokens: int,
        uniform_decode_query_len: int,
        num_tokens: int,
        num_reqs: int,
        force_uniform_decode: bool | None = None,
    ) -> bool:
        """
        Checks if it's a decode batch with same amount scheduled tokens
        across all requests.
        """
        return (
            (
                (max_num_scheduled_tokens == uniform_decode_query_len)
                and (num_tokens == max_num_scheduled_tokens * num_reqs)
            )
            if force_uniform_decode is None
            else force_uniform_decode
        )

2965
2966
2967
2968
2969
2970
2971
2972
2973
2974
2975
2976
2977
    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,
2978
        num_encoder_reqs: int = 0,
2979
    ) -> tuple[
2980
2981
        CUDAGraphMode,
        BatchDescriptor,
2982
        bool,
2983
2984
        torch.Tensor | None,
        CUDAGraphStat | None,
2985
    ]:
2986
2987
2988
2989
2990
2991
        uniform_decode = self._is_uniform_decode(
            max_num_scheduled_tokens=max_num_scheduled_tokens,
            uniform_decode_query_len=self.uniform_decode_query_len,
            num_tokens=num_tokens,
            num_reqs=num_reqs,
            force_uniform_decode=force_uniform_decode,
2992
        )
2993
2994
2995
2996
2997
        # Encoder-decoder models only support CG for decoder_step > 0 (no enc_output
        # is present). Also, chunked-prefill is disabled, so batch are uniform.
        has_encoder_output = (
            self.model_config.is_encoder_decoder and num_encoder_reqs > 0
        )
2998
2999
3000
3001
3002
3003
3004

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

3005
        num_tokens_padded = self._pad_for_sequence_parallelism(num_tokens)
3006
        dispatch_cudagraph = (
3007
            lambda num_tokens, disable_full: self.cudagraph_dispatcher.dispatch(
3008
3009
3010
                num_tokens=num_tokens,
                has_lora=has_lora,
                uniform_decode=uniform_decode,
3011
                disable_full=disable_full,
3012
3013
3014
3015
3016
            )
            if not force_eager
            else (CUDAGraphMode.NONE, BatchDescriptor(num_tokens_padded))
        )

3017
        cudagraph_mode, batch_descriptor = dispatch_cudagraph(
3018
            num_tokens_padded, use_cascade_attn or has_encoder_output
3019
        )
3020
        num_tokens_padded = batch_descriptor.num_tokens
3021
3022
3023
3024
3025
3026
3027
3028
3029
        if self.compilation_config.pass_config.enable_sp:
            assert (
                batch_descriptor.num_tokens
                % self.vllm_config.parallel_config.tensor_parallel_size
                == 0
            ), (
                "Sequence parallelism requires num_tokens to be "
                "a multiple of tensor parallel size"
            )
3030
3031
3032

        # Extra coordination when running data-parallel since we need to coordinate
        # across ranks
3033
        should_ubatch, num_tokens_across_dp = False, None
3034
3035
3036
3037
3038
3039
3040
3041
3042
        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
            )

3043
3044
3045
3046
3047
3048
3049
3050
3051
3052
3053
            should_ubatch, num_tokens_across_dp, synced_cudagraph_mode = (
                coordinate_batch_across_dp(
                    num_tokens_unpadded=num_tokens,
                    parallel_config=self.parallel_config,
                    allow_microbatching=allow_microbatching,
                    allow_dp_padding=allow_dp_padding,
                    num_tokens_padded=num_tokens_padded,
                    uniform_decode=uniform_decode,
                    num_scheduled_tokens_per_request=num_scheduled_tokens_np,
                    cudagraph_mode=cudagraph_mode.value,
                )
3054
3055
            )

3056
            # Extract DP-synced values
3057
3058
3059
            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())
3060
3061
3062
3063
3064
                # Re-dispatch with DP padding so we have the correct batch_descriptor
                cudagraph_mode, batch_descriptor = dispatch_cudagraph(
                    num_tokens_padded,
                    disable_full=synced_cudagraph_mode <= CUDAGraphMode.PIECEWISE.value,
                )
3065
3066
3067
3068
                # 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

3069
3070
3071
3072
3073
3074
3075
3076
3077
3078
3079
3080
        cudagraph_stats = None
        if self.vllm_config.observability_config.cudagraph_metrics:
            cudagraph_stats = CUDAGraphStat(
                num_unpadded_tokens=num_tokens,
                num_padded_tokens=batch_descriptor.num_tokens,
                num_paddings=batch_descriptor.num_tokens - num_tokens,
                runtime_mode=str(cudagraph_mode),
            )

        return (
            cudagraph_mode,
            batch_descriptor,
3081
            should_ubatch,
3082
3083
3084
            num_tokens_across_dp,
            cudagraph_stats,
        )
3085

3086
3087
3088
3089
3090
3091
3092
3093
3094
3095
3096
3097
3098
3099
3100
3101
3102
3103
3104
3105
3106
3107
3108
3109
3110
3111
3112
3113
3114
3115
3116
3117
3118
3119
3120
3121
    def _register_layerwise_nvtx_hooks(self) -> None:
        """
        Register layerwise NVTX hooks if --enable-layerwise-nvtx-tracing is enabled
        to trace detailed information of each layer or module in the model.
        """

        if (
            self.vllm_config.observability_config.enable_layerwise_nvtx_tracing
            and not self.layerwise_nvtx_hooks_registered
        ):
            if self.compilation_config.cudagraph_mode != CUDAGraphMode.NONE:
                logger.debug_once(
                    "layerwise NVTX tracing is not supported when CUDA graph is "
                    "turned off; you may observe part or all of the model "
                    "missing NVTX markers"
                )

            # In STOCK_TORCH_COMPILE mode, after registering hooks here,
            # the __call__ function of nn.module will be recompiled with
            # fullgraph=True. Since nvtx.range_push/pop are not traceable
            # by torch dynamo, we can't register hook functions here
            # because hook functions will also be traced by torch dynamo.
            if (
                self.vllm_config.compilation_config.mode
                == CompilationMode.STOCK_TORCH_COMPILE
            ):
                logger.debug_once(
                    "layerwise NVTX tracing is not supported when "
                    "CompilationMode is STOCK_TORCH_COMPILE, skipping "
                    "function hooks registration"
                )
            else:
                pyt_hooks = PytHooks()
                pyt_hooks.register_hooks(self.model, self.model.__class__.__name__)
                self.layerwise_nvtx_hooks_registered = True

3122
3123
3124
3125
    @torch.inference_mode()
    def execute_model(
        self,
        scheduler_output: "SchedulerOutput",
3126
        intermediate_tensors: IntermediateTensors | None = None,
3127
    ) -> ModelRunnerOutput | AsyncModelRunnerOutput | IntermediateTensors | None:
3128
3129
3130
3131
3132
        if self.execute_model_state is not None:
            raise RuntimeError(
                "State error: sample_tokens() must be called "
                "after execute_model() returns None."
            )
3133

3134
3135
3136
3137
3138
3139
3140
        if self.vllm_config.model_config.enable_return_routed_experts:
            capturer = RoutedExpertsCapturer.get_instance()
            if capturer is not None:
                capturer.clear_buffer()  # noqa
            else:
                logger.error("RoutedExpertsCapturer not initialized.")

3141
3142
3143
3144
3145
        if scheduler_output.preempted_req_ids and has_kv_transfer_group():
            get_kv_transfer_group().handle_preemptions(
                scheduler_output.preempted_req_ids
            )

3146
        num_scheduled_tokens = scheduler_output.total_num_scheduled_tokens
3147
3148
3149
3150
3151
3152
        with (
            record_function_or_nullcontext("gpu_model_runner: preprocess"),
            self.synchronize_input_prep(),
        ):
            # Update persistent batch states.
            self._update_states(scheduler_output)
3153

3154
3155
            if has_ec_transfer() and get_ec_transfer().is_producer:
                with self.maybe_get_ec_connector_output(
3156
                    scheduler_output,
3157
3158
3159
3160
3161
3162
3163
3164
3165
3166
3167
3168
3169
3170
3171
3172
3173
3174
3175
3176
3177
3178
3179
3180
3181
3182
3183
3184
                    encoder_cache=self.encoder_cache,
                ) as ec_connector_output:
                    self._execute_mm_encoder(scheduler_output)
                    return make_empty_encoder_model_runner_output(scheduler_output)

            if not num_scheduled_tokens:
                if (
                    self.parallel_config.distributed_executor_backend
                    == "external_launcher"
                    and self.parallel_config.data_parallel_size > 1
                ):
                    # this is a corner case when both external launcher
                    # and DP are enabled, num_scheduled_tokens could be
                    # 0, and has_unfinished_requests in the outer loop
                    # returns True. before returning early here we call
                    # dummy run to ensure coordinate_batch_across_dp
                    # is called into to avoid out of sync issues.
                    self._dummy_run(1)
                if not has_kv_transfer_group():
                    # Return empty ModelRunnerOutput if no work to do.
                    return EMPTY_MODEL_RUNNER_OUTPUT
                return self.kv_connector_no_forward(scheduler_output, self.vllm_config)

            if self.cache_config.kv_sharing_fast_prefill:
                assert not self.num_prompt_logprobs, (
                    "--kv-sharing-fast-prefill produces incorrect "
                    "logprobs for prompt tokens, tokens, please disable "
                    "it when the requests need prompt logprobs"
3185
3186
                )

3187
3188
3189
3190
3191
3192
            num_reqs = self.input_batch.num_reqs
            req_ids = self.input_batch.req_ids
            tokens = [scheduler_output.num_scheduled_tokens[i] for i in req_ids]
            num_scheduled_tokens_np = np.array(tokens, dtype=np.int32)
            max_num_scheduled_tokens = int(num_scheduled_tokens_np.max())
            num_tokens_unpadded = scheduler_output.total_num_scheduled_tokens
3193

3194
3195
3196
3197
            logits_indices, spec_decode_metadata = self._prepare_inputs(
                scheduler_output,
                num_scheduled_tokens_np,
            )
3198

3199
3200
3201
3202
3203
            cascade_attn_prefix_lens = None
            # Disable cascade attention when using microbatching (DBO)
            if self.cascade_attn_enabled and not self.parallel_config.use_ubatching:
                # Pre-compute cascade attention prefix lengths
                cascade_attn_prefix_lens = self._compute_cascade_attn_prefix_lens(
3204
                    num_scheduled_tokens_np,
3205
3206
                    self.input_batch.num_computed_tokens_cpu[:num_reqs],
                    scheduler_output.num_common_prefix_blocks,
3207
3208
                )

3209
3210
3211
3212
3213
3214
3215
3216
3217
3218
3219
3220
3221
3222
            (
                cudagraph_mode,
                batch_desc,
                should_ubatch,
                num_tokens_across_dp,
                cudagraph_stats,
            ) = self._determine_batch_execution_and_padding(
                num_tokens=num_tokens_unpadded,
                num_reqs=num_reqs,
                num_scheduled_tokens_np=num_scheduled_tokens_np,
                max_num_scheduled_tokens=max_num_scheduled_tokens,
                use_cascade_attn=cascade_attn_prefix_lens is not None,
                num_encoder_reqs=len(scheduler_output.scheduled_encoder_inputs),
            )
3223

3224
3225
3226
3227
3228
3229
3230
3231
3232
3233
3234
3235
3236
3237
3238
3239
3240
3241
3242
3243
3244
3245
3246
3247
3248
3249
3250
3251
3252
            logger.debug(
                "Running batch with cudagraph_mode: %s, batch_descriptor: %s, "
                "should_ubatch: %s, num_tokens_across_dp: %s",
                cudagraph_mode,
                batch_desc,
                should_ubatch,
                num_tokens_across_dp,
            )

            num_tokens_padded = batch_desc.num_tokens
            num_reqs_padded = (
                batch_desc.num_reqs if batch_desc.num_reqs is not None else num_reqs
            )
            ubatch_slices, ubatch_slices_padded = maybe_create_ubatch_slices(
                should_ubatch,
                num_scheduled_tokens_np,
                num_tokens_padded,
                num_reqs_padded,
                self.parallel_config.num_ubatches,
            )

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

            pad_attn = cudagraph_mode == CUDAGraphMode.FULL

3253
3254
3255
3256
3257
3258
3259
3260
3261
3262
3263
3264
            if self.cache_config.mamba_cache_mode == "align":
                mamba_utils.preprocess_mamba(
                    scheduler_output,
                    self.kv_cache_config,
                    self.cache_config,
                    self.mamba_state_idx,
                    self.input_batch,
                    self.requests,
                    self.compilation_config.static_forward_context,
                    self.model.get_mamba_state_copy_func(),
                )

3265
3266
3267
3268
3269
3270
3271
3272
3273
3274
3275
3276
3277
3278
3279
            use_spec_decode = len(scheduler_output.scheduled_spec_decode_tokens) > 0
            ubatch_slices_attn = ubatch_slices_padded if pad_attn else ubatch_slices

            attn_metadata, spec_decode_common_attn_metadata = (
                self._build_attention_metadata(
                    num_tokens=num_tokens_unpadded,
                    num_tokens_padded=num_tokens_padded if pad_attn else None,
                    num_reqs=num_reqs,
                    num_reqs_padded=num_reqs_padded if pad_attn else None,
                    max_query_len=max_num_scheduled_tokens,
                    ubatch_slices=ubatch_slices_attn,
                    logits_indices=logits_indices,
                    use_spec_decode=use_spec_decode,
                    num_scheduled_tokens=scheduler_output.num_scheduled_tokens,
                    cascade_attn_prefix_lens=cascade_attn_prefix_lens,
3280
                )
3281
            )
3282

3283
3284
3285
3286
3287
3288
3289
3290
3291
            (
                input_ids,
                inputs_embeds,
                positions,
                intermediate_tensors,
                model_kwargs,
                ec_connector_output,
            ) = self._preprocess(
                scheduler_output, num_tokens_padded, intermediate_tensors
3292
            )
3293

3294
        # Set cudagraph mode to none if calc_kv_scales is true.
3295
3296
3297
        # KV scales calculation involves dynamic operations that are incompatible
        # with CUDA graph capture.
        if self.calculate_kv_scales:
3298
            cudagraph_mode = CUDAGraphMode.NONE
3299
3300
            # Mark KV scales as calculated after the first forward pass
            self.calculate_kv_scales = False
3301

3302
3303
3304
3305
3306
3307
3308
        # Encoder-decoder models can only compile the pure decode steps where no
        # encoder inputs are present. Use eager for the first pass.
        num_encoder_reqs = len(scheduler_output.scheduled_encoder_inputs)
        has_encoder_input = (
            self.model_config.is_encoder_decoder and num_encoder_reqs > 0
        )

3309
3310
        # Run the model.
        # Use persistent buffers for CUDA graphs.
3311
3312
        with (
            set_forward_context(
3313
3314
                attn_metadata,
                self.vllm_config,
3315
                num_tokens=num_tokens_padded,
3316
                num_tokens_across_dp=num_tokens_across_dp,
3317
3318
                cudagraph_runtime_mode=cudagraph_mode,
                batch_descriptor=batch_desc,
3319
                ubatch_slices=ubatch_slices_padded,
3320
                skip_compiled=has_encoder_input,
3321
            ),
3322
            record_function_or_nullcontext("gpu_model_runner: forward"),
3323
3324
            self.maybe_get_kv_connector_output(scheduler_output) as kv_connector_output,
        ):
3325
            model_output = self._model_forward(
3326
3327
3328
3329
3330
3331
3332
                input_ids=input_ids,
                positions=positions,
                intermediate_tensors=intermediate_tensors,
                inputs_embeds=inputs_embeds,
                **model_kwargs,
            )

3333
        with record_function_or_nullcontext("gpu_model_runner: postprocess"):
3334
            if self.use_aux_hidden_state_outputs:
3335
                # True when EAGLE 3 is used.
3336
3337
                hidden_states, aux_hidden_states = model_output
            else:
3338
                # Common case.
3339
3340
3341
                hidden_states = model_output
                aux_hidden_states = None

3342
3343
3344
3345
3346
            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)
3347
                    hidden_states.kv_connector_output = kv_connector_output
3348
                    self.kv_connector_output = kv_connector_output
3349
                    return hidden_states
3350

3351
                if self.is_pooling_model:
3352
                    # Return the pooling output.
3353
3354
3355
3356
3357
                    return self._pool(
                        hidden_states,
                        num_scheduled_tokens,
                        num_scheduled_tokens_np,
                        kv_connector_output,
3358
                    )
3359
3360

                sample_hidden_states = hidden_states[logits_indices]
3361
                logits = self.model.compute_logits(sample_hidden_states)
3362
3363
3364
3365
            else:
                # Rare case.
                assert not self.is_pooling_model

3366
                sample_hidden_states = hidden_states[logits_indices]
3367
                if not get_pp_group().is_last_rank:
3368
                    all_gather_tensors = {
3369
                        "residual": not is_residual_scattered_for_sp(
3370
                            self.vllm_config, num_tokens_padded
3371
                        )
3372
                    }
3373
                    get_pp_group().send_tensor_dict(
3374
3375
                        hidden_states.tensors,
                        all_gather_group=get_tp_group(),
3376
3377
                        all_gather_tensors=all_gather_tensors,
                    )
3378
3379
                    logits = None
                else:
3380
                    logits = self.model.compute_logits(sample_hidden_states)
3381

3382
                model_output_broadcast_data: dict[str, Any] = {}
3383
3384
3385
                if logits is not None:
                    model_output_broadcast_data["logits"] = logits.contiguous()

3386
                broadcasted = get_pp_group().broadcast_tensor_dict(
3387
3388
                    model_output_broadcast_data, src=len(get_pp_group().ranks) - 1
                )
3389
3390
                assert broadcasted is not None
                logits = broadcasted["logits"]
3391

3392
3393
3394
3395
3396
3397
3398
3399
        self.execute_model_state = ExecuteModelState(
            scheduler_output,
            logits,
            spec_decode_metadata,
            spec_decode_common_attn_metadata,
            hidden_states,
            sample_hidden_states,
            aux_hidden_states,
3400
            ec_connector_output,
3401
            cudagraph_stats,
3402
        )
3403
        self.kv_connector_output = kv_connector_output
3404
3405
3406
3407
3408
3409
        return None

    @torch.inference_mode
    def sample_tokens(
        self, grammar_output: "GrammarOutput | None"
    ) -> ModelRunnerOutput | AsyncModelRunnerOutput | IntermediateTensors:
3410
3411
3412
        kv_connector_output = self.kv_connector_output
        self.kv_connector_output = None

3413
3414
        if self.execute_model_state is None:
            # Nothing to do (PP non-final rank case), output isn't used.
3415
            if not kv_connector_output:
3416
                return None  # type: ignore[return-value]
3417
3418
3419
3420
3421
3422
3423
3424
3425

            # 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
3426
3427
3428
3429
3430
3431
3432
3433
3434
3435

        # Unpack ephemeral state.
        (
            scheduler_output,
            logits,
            spec_decode_metadata,
            spec_decode_common_attn_metadata,
            hidden_states,
            sample_hidden_states,
            aux_hidden_states,
3436
            ec_connector_output,
3437
            cudagraph_stats,
3438
3439
3440
3441
3442
3443
3444
3445
3446
        ) = 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
            )
3447

3448
        with record_function_or_nullcontext("gpu_model_runner: sample"):
3449
3450
            sampler_output = self._sample(logits, spec_decode_metadata)

3451
3452
3453
3454
        self._update_states_after_model_execute(
            sampler_output.sampled_token_ids, scheduler_output
        )

3455
3456
        self._draft_token_ids = None
        self._draft_token_req_ids = None
3457
3458
        self.input_batch.prev_sampled_token_ids = None

3459
        def propose_draft_token_ids(sampled_token_ids):
3460
            assert spec_decode_common_attn_metadata is not None
3461
            with record_function_or_nullcontext("gpu_model_runner: draft"):
3462
3463
3464
3465
3466
3467
3468
3469
3470
3471
                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,
                )
3472
                self._copy_draft_token_ids_to_cpu(scheduler_output)
3473

3474
        spec_config = self.speculative_config
3475
3476
3477
3478
3479
        propose_drafts_after_bookkeeping = False
        if spec_config is not None:
            input_fits_in_drafter = spec_decode_common_attn_metadata is not None and (
                spec_decode_common_attn_metadata.max_seq_len + self.num_spec_tokens
                <= self.effective_drafter_max_model_len
3480
            )
3481
3482
3483
3484
3485
            use_gpu_toks = (
                spec_config.use_eagle() or spec_config.uses_draft_model()
            ) and not spec_config.disable_padded_drafter_batch
            if use_gpu_toks:
                # EAGLE/DraftModel speculative decoding can use the GPU sampled tokens
3486
                # as inputs, and does not need to wait for bookkeeping to finish.
3487
                assert isinstance(self.drafter, EagleProposer | DraftModelProposer)
3488
3489
3490
3491
3492
3493
3494
3495
3496
3497
3498
3499
3500
                sampled_token_ids = sampler_output.sampled_token_ids
                if input_fits_in_drafter:
                    propose_draft_token_ids(sampled_token_ids)
                elif self.valid_sampled_token_count_event is not None:
                    assert spec_decode_common_attn_metadata is not None
                    next_token_ids, valid_sampled_tokens_count = (
                        self.drafter.prepare_next_token_ids_padded(
                            spec_decode_common_attn_metadata,
                            sampled_token_ids,
                            self.requests,
                            self.input_batch,
                            self.discard_request_mask.gpu,
                        )
3501
                    )
3502
3503
3504
3505
3506
3507
3508
3509
3510
3511
3512
                    self._copy_valid_sampled_token_count(
                        next_token_ids, valid_sampled_tokens_count
                    )
                    # Since we couldn't run the drafter,
                    # just use zeros for the draft tokens.
                    self._draft_token_ids = torch.zeros(
                        1, device=self.device, dtype=torch.int32
                    ).expand(len(self.input_batch.req_ids), self.num_spec_tokens)
                    self._copy_draft_token_ids_to_cpu(scheduler_output, zeros_only=True)
            else:
                propose_drafts_after_bookkeeping = input_fits_in_drafter
3513

3514
        with record_function_or_nullcontext("gpu_model_runner: bookkeep"):
3515
3516
3517
3518
3519
3520
3521
3522
            (
                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,
3523
3524
3525
3526
3527
            ) = self._bookkeeping_sync(
                scheduler_output,
                sampler_output,
                logits,
                hidden_states,
3528
                scheduler_output.total_num_scheduled_tokens,
3529
                spec_decode_metadata,
3530
            )
3531

3532
        if propose_drafts_after_bookkeeping:
3533
3534
3535
            # 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)
3536

3537
        with record_function_or_nullcontext("gpu_model_runner: eplb"):
3538
            self.eplb_step()
3539

3540
        with record_function_or_nullcontext("gpu_model_runner: ModelRunnerOutput"):
3541
3542
3543
3544
3545
3546
3547
            if self.model_config.enable_return_routed_experts:
                capturer = RoutedExpertsCapturer.get_instance()
                if capturer is not None:
                    capturer.save_captured_experts(indices=self.slot_mapping)  # noqa
                else:
                    logger.error("RoutedExpertsCapturer not initialized.")

3548
3549
3550
3551
3552
3553
3554
            output = ModelRunnerOutput(
                req_ids=req_ids_output_copy,
                req_id_to_index=req_id_to_index_output_copy,
                sampled_token_ids=valid_sampled_token_ids,
                logprobs=logprobs_lists,
                prompt_logprobs_dict=prompt_logprobs_dict,
                kv_connector_output=kv_connector_output,
3555
3556
3557
                ec_connector_output=ec_connector_output
                if self.supports_mm_inputs
                else None,
3558
                num_nans_in_logits=num_nans_in_logits,
3559
                cudagraph_stats=cudagraph_stats,
3560
            )
3561

3562
3563
        if not self.use_async_scheduling:
            return output
3564

3565
3566
3567
3568
3569
3570
3571
3572
3573
        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,
3574
                vocab_size=self.input_batch.vocab_size,
3575
3576
3577
3578
3579
            )
        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
3580
            # any requests with sampling params that require output ids.
3581
3582
3583
3584
            self.input_batch.set_async_sampled_token_ids(
                async_output.sampled_token_ids_cpu,
                async_output.async_copy_ready_event,
            )
3585
3586
3587

        return async_output

3588
    def take_draft_token_ids(self) -> DraftTokenIds | None:
3589
        if not self.num_spec_tokens or not self._draft_token_req_ids:
3590
            return None
3591
        draft_token_ids, req_ids = self._get_draft_token_ids_cpu()
3592
        return DraftTokenIds(req_ids, draft_token_ids)
3593

3594
3595
3596
    def _copy_draft_token_ids_to_cpu(
        self, scheduler_output: "SchedulerOutput", zeros_only: bool = False
    ) -> None:
3597
3598
3599
3600
3601
3602
        # Check if we need to copy draft tokens to CPU. In async scheduling,
        # we only copy when needed for structured output, penalties or bad_words.
        if self.use_async_scheduling and not (
            scheduler_output.has_structured_output_requests
            or self.input_batch.sampling_metadata.output_token_ids
        ):
3603
3604
3605
            return
        # We must also set the corresponding request ids.
        self._draft_token_req_ids = self.input_batch.req_ids.copy()
3606

3607
3608
3609
3610
3611
3612
3613
3614
3615
3616
3617
3618
3619
3620
3621
3622
3623
3624
3625
3626
        draft_token_ids: torch.Tensor = self._draft_token_ids
        if not torch.is_tensor(draft_token_ids):
            return
        assert self.draft_token_ids_event is not None
        assert self.draft_token_ids_copy_stream is not None
        assert self.draft_token_ids_cpu is not None
        default_stream = torch.cuda.current_stream()
        num_reqs = draft_token_ids.shape[0]
        with torch.cuda.stream(self.draft_token_ids_copy_stream):
            if not zeros_only:
                # Trigger async copy of draft token ids to cpu.
                self.draft_token_ids_copy_stream.wait_stream(default_stream)
                self.draft_token_ids_cpu[:num_reqs].copy_(
                    draft_token_ids, non_blocking=True
                )
            else:
                # No copy needed, just zero-out cpu tensor.
                self.draft_token_ids_cpu[:num_reqs] = 0
            self.draft_token_ids_event.record()

3627
    def _get_draft_token_ids_cpu(self) -> tuple[list[list[int]], list[str]]:
3628
        if isinstance(self._draft_token_ids, list):
3629
3630
3631
3632
            return self._draft_token_ids, self.input_batch.req_ids
        req_ids = self._draft_token_req_ids
        if req_ids is None:
            return [], []
3633
3634
3635
        assert self.draft_token_ids_event is not None
        assert self.draft_token_ids_cpu is not None
        self.draft_token_ids_event.synchronize()
3636
        return self.draft_token_ids_cpu[: len(req_ids)].tolist(), req_ids
3637

3638
3639
3640
3641
3642
3643
3644
3645
3646
3647
3648
3649
3650
    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
3651
            assert counts_cpu is not None
3652
3653
3654
3655
3656
3657
3658
3659
            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
3660
3661
        sampled_count_event = self.valid_sampled_token_count_event
        if sampled_count_event is None or prev_sampled_token_ids is None:
3662
3663
3664
            return []

        counts_cpu = self.valid_sampled_token_count_cpu
3665
3666
        assert counts_cpu is not None
        sampled_count_event.synchronize()
3667
3668
        return counts_cpu[: prev_sampled_token_ids.shape[0]].tolist()

3669
3670
3671
    def propose_draft_token_ids(
        self,
        scheduler_output: "SchedulerOutput",
3672
        sampled_token_ids: torch.Tensor | list[list[int]],
3673
3674
3675
        sampling_metadata: SamplingMetadata,
        hidden_states: torch.Tensor,
        sample_hidden_states: torch.Tensor,
3676
3677
        aux_hidden_states: list[torch.Tensor] | None,
        spec_decode_metadata: SpecDecodeMetadata | None,
3678
        common_attn_metadata: CommonAttentionMetadata,
3679
    ) -> list[list[int]] | torch.Tensor:
3680
        num_scheduled_tokens = scheduler_output.total_num_scheduled_tokens
3681
3682
3683
        spec_config = self.speculative_config
        assert spec_config is not None
        if spec_config.method == "ngram":
3684
            assert isinstance(sampled_token_ids, list)
3685
            assert isinstance(self.drafter, NgramProposer)
3686
            draft_token_ids = self.drafter.propose(
3687
                sampled_token_ids,
3688
3689
                self.input_batch.num_tokens_no_spec,
                self.input_batch.token_ids_cpu,
3690
            )
3691
        elif spec_config.method == "suffix":
3692
3693
3694
            assert isinstance(sampled_token_ids, list)
            assert isinstance(self.drafter, SuffixDecodingProposer)
            draft_token_ids = self.drafter.propose(self.input_batch, sampled_token_ids)
3695
        elif spec_config.method == "medusa":
3696
            assert isinstance(sampled_token_ids, list)
3697
            assert isinstance(self.drafter, MedusaProposer)
3698

3699
3700
            if sample_hidden_states.shape[0] == len(sampled_token_ids):
                # The input to the target model does not include draft tokens.
3701
3702
3703
3704
                hidden_states = sample_hidden_states
            else:
                indices = []
                offset = 0
3705
3706
3707
                assert spec_decode_metadata is not None, (
                    "No spec decode metadata for medusa"
                )
3708
                for num_draft, tokens in zip(
3709
3710
                    spec_decode_metadata.num_draft_tokens, sampled_token_ids
                ):
3711
                    indices.append(offset + len(tokens) - 1)
3712
                    offset += num_draft + 1
3713
                indices = torch.tensor(indices, device=self.device)
3714
3715
                hidden_states = sample_hidden_states[indices]

3716
            draft_token_ids = self.drafter.propose(
3717
3718
3719
                target_hidden_states=hidden_states,
                sampling_metadata=sampling_metadata,
            )
3720
3721
        elif spec_config.use_eagle() or spec_config.uses_draft_model():
            assert isinstance(self.drafter, EagleProposer | DraftModelProposer)
3722

3723
            if spec_config.disable_padded_drafter_batch:
3724
3725
3726
                # 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.
3727
3728
                assert isinstance(sampled_token_ids, list), (
                    "sampled_token_ids should be a python list when"
3729
                    "padded-batch is disabled."
3730
                )
3731
                next_token_ids = self.drafter.prepare_next_token_ids_cpu(
3732
3733
3734
3735
3736
                    sampled_token_ids,
                    self.requests,
                    self.input_batch,
                    scheduler_output.num_scheduled_tokens,
                )
3737
3738
3739
3740
3741
            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.
3742
3743
                assert isinstance(sampled_token_ids, torch.Tensor), (
                    "sampled_token_ids should be a torch.Tensor when"
3744
                    "padded-batch is enabled."
3745
3746
                )
                next_token_ids, valid_sampled_tokens_count = (
3747
3748
3749
3750
3751
                    self.drafter.prepare_next_token_ids_padded(
                        common_attn_metadata,
                        sampled_token_ids,
                        self.requests,
                        self.input_batch,
3752
                        self.discard_request_mask.gpu,
3753
                    )
3754
                )
3755
3756
3757
                self._copy_valid_sampled_token_count(
                    next_token_ids, valid_sampled_tokens_count
                )
Jiayi Yao's avatar
Jiayi Yao committed
3758

3759
            num_rejected_tokens_gpu = None
3760
            if spec_decode_metadata is None:
3761
                token_indices_to_sample = None
3762
                # input_ids can be None for multimodal models.
3763
                target_token_ids = self.input_ids.gpu[:num_scheduled_tokens]
3764
                target_positions = self._get_positions(num_scheduled_tokens)
3765
                if self.use_aux_hidden_state_outputs:
Wentao Ye's avatar
Wentao Ye committed
3766
                    assert aux_hidden_states is not None
3767
                    target_hidden_states = torch.cat(
3768
3769
                        [h[:num_scheduled_tokens] for h in aux_hidden_states], dim=-1
                    )
3770
3771
                else:
                    target_hidden_states = hidden_states[:num_scheduled_tokens]
3772
            else:
3773
                if spec_config.disable_padded_drafter_batch:
3774
                    token_indices_to_sample = None
3775
3776
3777
3778
3779
                    common_attn_metadata, token_indices = self.drafter.prepare_inputs(
                        common_attn_metadata,
                        sampled_token_ids,
                        spec_decode_metadata.num_draft_tokens,
                    )
3780
3781
3782
3783
3784
3785
3786
3787
3788
                    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]
3789
                else:
3790
3791
3792
3793
3794
3795
3796
3797
                    (
                        common_attn_metadata,
                        token_indices_to_sample,
                        num_rejected_tokens_gpu,
                    ) = self.drafter.prepare_inputs_padded(
                        common_attn_metadata,
                        spec_decode_metadata,
                        valid_sampled_tokens_count,
3798
                    )
3799
3800
3801
3802
3803
3804
3805
3806
3807
3808
3809
                    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]
3810

3811
            if self.supports_mm_inputs:
3812
3813
3814
3815
3816
3817
                mm_embed_inputs = self._gather_mm_embeddings(
                    scheduler_output,
                    shift_computed_tokens=1,
                )
            else:
                mm_embed_inputs = None
3818

3819
            draft_token_ids = self.drafter.propose(
3820
3821
3822
3823
                target_token_ids=target_token_ids,
                target_positions=target_positions,
                target_hidden_states=target_hidden_states,
                next_token_ids=next_token_ids,
3824
                last_token_indices=token_indices_to_sample,
3825
                sampling_metadata=sampling_metadata,
3826
                common_attn_metadata=common_attn_metadata,
3827
                mm_embed_inputs=mm_embed_inputs,
3828
                num_rejected_tokens_gpu=num_rejected_tokens_gpu,
3829
            )
3830

3831
        return draft_token_ids
3832

3833
3834
3835
    def update_config(self, overrides: dict[str, Any]) -> None:
        allowed_config_names = {"load_config", "model_config"}
        for config_name, config_overrides in overrides.items():
3836
3837
            assert config_name in allowed_config_names, (
                f"Config `{config_name}` not supported. "
3838
                f"Allowed configs: {allowed_config_names}"
3839
            )
3840
3841
3842
3843
            config = getattr(self, config_name)
            new_config = update_config(config, config_overrides)
            setattr(self, config_name, new_config)

3844
3845
3846
3847
3848
    def load_model(self, eep_scale_up: bool = False) -> None:
        """
        Args:
            eep_scale_up: the model loading is for elastic EP scale up.
        """
3849
3850
3851
3852
3853
        logger.info_once(
            "Starting to load model %s...",
            self.model_config.model,
            scope="global",
        )
3854
3855
3856
3857
3858
        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)
        )
3859

3860
3861
3862
3863
        if self.parallel_config.enable_eplb:
            self.eplb_state = EplbState(self.parallel_config, self.device)
            eplb_models = 0

3864
3865
3866
3867
3868
3869
3870
3871
3872
3873
        try:
            with DeviceMemoryProfiler() as m:
                time_before_load = time.perf_counter()
                model_loader = get_model_loader(self.load_config)
                self.model = model_loader.load_model(
                    vllm_config=self.vllm_config, model_config=self.model_config
                )
                if self.lora_config:
                    self.model = self.load_lora_model(
                        self.model, self.vllm_config, self.device
3874
                    )
3875
3876
3877
3878
3879
3880
3881
3882
3883
3884
3885
3886
3887
3888
3889
                if hasattr(self, "drafter"):
                    logger.info_once("Loading drafter model...")
                    self.drafter.load_model(self.model)
                    if (
                        hasattr(self.drafter, "model")
                        and is_mixture_of_experts(self.drafter.model)
                        and self.parallel_config.enable_eplb
                    ):
                        spec_config = self.vllm_config.speculative_config
                        assert spec_config is not None
                        assert spec_config.draft_model_config is not None
                        logger.info_once(
                            "EPLB is enabled for drafter model %s.",
                            spec_config.draft_model_config.model,
                        )
3890

3891
3892
3893
3894
3895
3896
3897
3898
3899
3900
3901
3902
3903
3904
3905
3906
3907
3908
3909
3910
3911
3912
                        global_expert_load = (
                            global_expert_loads[eplb_models]
                            if global_expert_loads
                            else None
                        )
                        old_global_expert_indices = (
                            old_global_expert_indices_per_model[eplb_models]
                            if old_global_expert_indices_per_model
                            else None
                        )
                        if self.eplb_state is None:
                            self.eplb_state = EplbState(
                                self.parallel_config, self.device
                            )
                        self.eplb_state.add_model(
                            self.drafter.model,
                            spec_config.draft_model_config,
                            global_expert_load,
                            old_global_expert_indices,
                            rank_mapping,
                        )
                        eplb_models += 1
3913

3914
3915
3916
3917
3918
3919
                if self.use_aux_hidden_state_outputs:
                    if not supports_eagle3(self.get_model()):
                        raise RuntimeError(
                            "Model does not support EAGLE3 interface but "
                            "aux_hidden_state_outputs was requested"
                        )
3920

3921
3922
3923
3924
3925
3926
3927
3928
3929
3930
3931
3932
3933
3934
3935
3936
3937
3938
3939
3940
3941
3942
3943
3944
3945
                    # Try to get auxiliary layers from speculative config,
                    # otherwise use model's default layers
                    aux_layers = self._get_eagle3_aux_layers_from_config()
                    if aux_layers:
                        logger.info(
                            "Using auxiliary layers from speculative config: %s",
                            aux_layers,
                        )
                    else:
                        aux_layers = self.model.get_eagle3_aux_hidden_state_layers()

                    self.model.set_aux_hidden_state_layers(aux_layers)
                time_after_load = time.perf_counter()
            self.model_memory_usage = m.consumed_memory
        except torch.cuda.OutOfMemoryError as e:
            msg = (
                "Failed to load model - not enough GPU memory. "
                "Try lowering --gpu-memory-utilization to free memory for weights, "
                "increasing --tensor-parallel-size, or using --quantization. "
                "See https://docs.vllm.ai/en/latest/configuration/conserving_memory/ "
                "for more tips."
            )
            combined_msg = f"{msg} (original error: {e})"
            logger.error(combined_msg)
            raise e
3946
        logger.info_once(
3947
3948
            "Model loading took %s GiB memory and %.6f seconds",
            format_gib(self.model_memory_usage),
3949
            time_after_load - time_before_load,
3950
            scope="local",
3951
        )
3952
        prepare_communication_buffer_for_model(self.model)
3953
3954
3955
3956
        if (drafter := getattr(self, "drafter", None)) and (
            drafter_model := getattr(drafter, "model", None)
        ):
            prepare_communication_buffer_for_model(drafter_model)
3957
        mm_config = self.model_config.multimodal_config
3958
        self.is_multimodal_pruning_enabled = (
3959
            supports_multimodal_pruning(self.get_model())
3960
3961
            and mm_config is not None
            and mm_config.is_multimodal_pruning_enabled()
3962
        )
3963

3964
        if is_mixture_of_experts(self.model) and self.parallel_config.enable_eplb:
3965
3966
3967
3968
3969
3970
3971
3972
3973
3974
3975
            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(
3976
                self.model,
3977
                self.model_config,
3978
3979
3980
                global_expert_load,
                old_global_expert_indices,
                rank_mapping,
3981
            )
3982
3983
            if self.eplb_state.is_async:
                self.eplb_state.start_async_loop(rank_mapping=rank_mapping)
3984

3985
        if (
3986
3987
            self.vllm_config.compilation_config.mode
            == CompilationMode.STOCK_TORCH_COMPILE
3988
        ):
3989
            backend = self.vllm_config.compilation_config.init_backend(self.vllm_config)
3990
            compilation_counter.stock_torch_compile_count += 1
3991
            self.model.compile(fullgraph=True, backend=backend)
3992
            return
3993
        # for other compilation modes, cudagraph behavior is controlled by
3994
3995
3996
        # CudagraphWraper and CudagraphDispatcher of vllm.

        # wrap the model with full cudagraph wrapper if needed.
3997
3998
        cudagraph_mode = self.compilation_config.cudagraph_mode
        assert cudagraph_mode is not None
3999
4000
4001
4002
        if (
            cudagraph_mode.has_full_cudagraphs()
            and not self.parallel_config.use_ubatching
        ):
4003
4004
4005
            self.model = CUDAGraphWrapper(
                self.model, self.vllm_config, runtime_mode=CUDAGraphMode.FULL
            )
4006
        elif self.parallel_config.use_ubatching:
4007
            if cudagraph_mode.has_full_cudagraphs():
4008
4009
4010
                self.model = UBatchWrapper(
                    self.model, self.vllm_config, CUDAGraphMode.FULL, self.device
                )
4011
            else:
4012
4013
4014
                self.model = UBatchWrapper(
                    self.model, self.vllm_config, CUDAGraphMode.NONE, self.device
                )
4015

4016
    def _get_eagle3_aux_layers_from_config(self) -> tuple[int, ...] | None:
4017
4018
4019
4020
4021
4022
4023
4024
4025
4026
4027
4028
4029
4030
4031
4032
4033
4034
4035
4036
4037
4038
4039
        """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

4040
    def reload_weights(self) -> None:
4041
        assert getattr(self, "model", None) is not None, (
4042
            "Cannot reload weights before model is loaded."
4043
        )
4044
4045
        model_loader = get_model_loader(self.load_config)
        logger.info("Reloading weights inplace...")
4046
        model_loader.load_weights(self.get_model(), model_config=self.model_config)
4047

4048
4049
4050
4051
4052
    def save_tensorized_model(
        self,
        tensorizer_config: "TensorizerConfig",
    ) -> None:
        TensorizerLoader.save_model(
4053
            self.get_model(),
4054
            tensorizer_config=tensorizer_config,
4055
            model_config=self.model_config,
4056
4057
        )

4058
4059
4060
    def _get_prompt_logprobs_dict(
        self,
        hidden_states: torch.Tensor,
4061
        num_scheduled_tokens: dict[str, int],
4062
    ) -> dict[str, LogprobsTensors | None]:
4063
        num_prompt_logprobs_dict = self.num_prompt_logprobs
4064
4065
4066
        if not num_prompt_logprobs_dict:
            return {}

4067
        in_progress_dict = self.input_batch.in_progress_prompt_logprobs_cpu
4068
        prompt_logprobs_dict: dict[str, LogprobsTensors | None] = {}
4069
4070
4071
4072
4073

        # 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():
4074
4075
4076
4077
            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
4078
4079
4080

            # Get metadata for this request.
            request = self.requests[req_id]
4081
4082
4083
4084
            if request.prompt_token_ids is None:
                # Prompt logprobs is incompatible with prompt embeddings
                continue

4085
4086
            num_prompt_tokens = len(request.prompt_token_ids)
            prompt_token_ids = torch.tensor(request.prompt_token_ids).to(
4087
4088
                self.device, non_blocking=True
            )
4089

4090
4091
4092
4093
4094
4095
            # 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(
4096
4097
                    num_prompt_tokens - 1, num_prompt_logprobs + 1
                )
4098
4099
                in_progress_dict[req_id] = logprobs_tensors

4100
            # Determine number of logits to retrieve.
4101
4102
            start_idx = request.num_computed_tokens
            start_tok = start_idx + 1
4103
            num_remaining_tokens = num_prompt_tokens - start_tok
4104
            if num_tokens <= num_remaining_tokens:
4105
                # This is a chunk, more tokens remain.
4106
4107
4108
                # 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.
4109
4110
4111
4112
4113
                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)
4114
4115
4116
4117
4118
4119
4120
                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
4121
4122
4123
4124
4125

            # 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]
4126
            offset = self.query_start_loc.np[req_idx].item()
4127
            prompt_hidden_states = hidden_states[offset : offset + num_logits]
4128
            logits = self.model.compute_logits(prompt_hidden_states)
4129
4130
4131
4132

            # 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.
4133
            tgt_token_ids = prompt_token_ids[start_tok : start_tok + num_logits]
4134
4135

            # Compute prompt logprobs.
4136
4137
            logprobs = self.sampler.compute_logprobs(logits)
            token_ids, logprobs, ranks = self.sampler.gather_logprobs(
4138
4139
                logprobs, num_prompt_logprobs, tgt_token_ids
            )
4140
4141

            # Transfer GPU->CPU async.
4142
4143
            chunk_slice = slice(start_idx, start_idx + num_logits)
            logprobs_tensors.logprob_token_ids[chunk_slice].copy_(
4144
4145
4146
                token_ids, non_blocking=True
            )
            logprobs_tensors.logprobs[chunk_slice].copy_(logprobs, non_blocking=True)
4147
            logprobs_tensors.selected_token_ranks[chunk_slice].copy_(
4148
4149
                ranks, non_blocking=True
            )
4150
4151
4152
4153
4154

        # 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]
4155
            del in_progress_dict[req_id]
4156
4157

        # Must synchronize the non-blocking GPU->CPU transfers.
4158
        if prompt_logprobs_dict:
4159
            self._sync_device()
4160
4161
4162

        return prompt_logprobs_dict

4163
4164
    def _get_nans_in_logits(
        self,
4165
        logits: torch.Tensor | None,
4166
4167
4168
4169
4170
4171
4172
4173
4174
4175
4176
    ) -> 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])
4177
4178
4179
                    if num_nans_for_index is not None and req_index < logits.shape[0]
                    else 0
                )
4180
4181
4182
4183
            return num_nans_in_logits
        except IndexError:
            return {}

4184
    @contextmanager
4185
4186
4187
    def maybe_randomize_inputs(
        self, input_ids: torch.Tensor | None, inputs_embeds: torch.Tensor | None
    ):
4188
4189
4190
4191
        """
        Randomize input_ids if VLLM_RANDOMIZE_DP_DUMMY_INPUTS is set.
        This is to help balance expert-selection
         - during profile_run
4192
         - during DP rank dummy run
4193
        """
4194

4195
4196
4197
4198
        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
4199
        elif input_ids is not None:
4200
4201
4202
4203

            @functools.cache
            def rand_input_ids() -> torch.Tensor:
                return torch.randint_like(
4204
                    self.input_ids.gpu,
4205
4206
                    low=0,
                    high=self.model_config.get_vocab_size(),
4207
                )
4208

4209
            logger.debug_once("Randomizing dummy input_ids for DP Rank")
4210
            input_ids.copy_(rand_input_ids()[: input_ids.size(0)], non_blocking=True)
4211
4212
            yield
            input_ids.fill_(0)
4213
4214
4215
4216
4217
4218
4219
4220
4221
4222
4223
4224
4225
4226
4227
        else:

            @functools.cache
            def rand_inputs_embeds() -> torch.Tensor:
                return torch.randn_like(
                    self.inputs_embeds.gpu,
                )

            assert inputs_embeds is not None
            logger.debug_once("Randomizing dummy inputs_embeds for DP Rank")
            inputs_embeds.copy_(
                rand_inputs_embeds()[: inputs_embeds.size(0)], non_blocking=True
            )
            yield
            inputs_embeds.fill_(0)
4228

4229
4230
4231
4232
4233
4234
    def _get_mm_dummy_batch(
        self,
        modality: str,
        max_items_per_batch: int,
    ) -> BatchedTensorInputs:
        """Dummy data for profiling and precompiling multimodal models."""
4235
4236
        assert self.mm_budget is not None

4237
4238
4239
        # Don't use `max_items_per_batch` here to avoid redundant computation
        dummy_mm_inputs = self.mm_registry.get_dummy_mm_inputs(
            self.model_config,
4240
            mm_counts={modality: 1},
4241
            cache=self.mm_budget.cache,
4242
        )
4243
4244
4245
4246
4247
        dummy_mm_item = dummy_mm_inputs["mm_kwargs"][modality][0]

        # We use the cache so that the item is saved to the cache,
        # but not read from the cache
        assert dummy_mm_item is not None, "Item should not already be cached"
4248

4249
        dummy_mm_items = [dummy_mm_item] * max_items_per_batch
4250

4251
4252
4253
4254
4255
4256
4257
4258
        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,
            )
        )
4259

4260
4261
4262
4263
    @torch.inference_mode()
    def _dummy_run(
        self,
        num_tokens: int,
4264
        cudagraph_runtime_mode: CUDAGraphMode | None = None,
4265
4266
        force_attention: bool = False,
        uniform_decode: bool = False,
4267
        allow_microbatching: bool = True,
4268
4269
        skip_eplb: bool = False,
        is_profile: bool = False,
4270
        create_mixed_batch: bool = False,
4271
        remove_lora: bool = True,
4272
        activate_lora: bool = False,
Rémi Delacourt's avatar
Rémi Delacourt committed
4273
        is_graph_capturing: bool = False,
4274
    ) -> tuple[torch.Tensor, torch.Tensor]:
4275
4276
4277
4278
4279
4280
4281
        """
        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.
4282
                - if not set will determine the cudagraph mode based on using
4283
                    the self.cudagraph_dispatcher.
4284
4285
4286
4287
                - CUDAGraphMode.NONE: No cudagraph, for warm up and profile run
                - CUDAGraphMode.PIECEWISE: Piecewise cudagraph.
                - CUDAGraphMode.FULL: Full cudagraph, attention metadata is
                    needed.
4288
            force_attention: If True, always create attention metadata. Used to
4289
4290
4291
4292
                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.
4293
4294
            create_mixed_batch: If True, create a mixed batch with both decode
                (1 token) and prefill (multiple tokens) requests.
4295
            remove_lora: If False, dummy LoRAs are not destroyed after the run
4296
            activate_lora: If False, dummy_run is performed without LoRAs.
4297
        """
4298
4299
        mm_config = self.vllm_config.model_config.multimodal_config
        if mm_config and mm_config.mm_encoder_only:
4300
4301
4302
4303
            # The current dummy run only covers LM execution, so we can skip it.
            # mm encoder dummy run may need to add in the future.
            return torch.tensor([]), torch.tensor([])

4304
4305
4306
4307
        assert (
            cudagraph_runtime_mode is None
            or cudagraph_runtime_mode.valid_runtime_modes()
        )
4308

4309
        # If cudagraph_mode.decode_mode() == FULL and
4310
        # cudagraph_mode.separate_routine(). This means that we are using
4311
4312
4313
4314
4315
4316
4317
4318
4319
4320
4321
        # 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.
4322
        max_query_len = self.uniform_decode_query_len if uniform_decode else num_tokens
4323

4324
4325
4326
4327
4328
        # 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
4329
4330
4331
4332
        if create_mixed_batch:
            assert not uniform_decode
            # Create mixed batch:
            # first half decode tokens, second half one prefill
4333
            num_decode_tokens = min(max_num_reqs - 1, num_tokens // 2)
4334
4335
4336
4337
            num_prefill_tokens = num_tokens - num_decode_tokens
            num_reqs = num_decode_tokens + 1

            # Create decode requests (1 token each) followed by prefill request
4338
            num_scheduled_tokens_list = [1] * num_decode_tokens + [num_prefill_tokens]
4339
4340
4341
            # Note: Overriding max_query_len to be the prefill tokens
            max_query_len = num_prefill_tokens
        elif uniform_decode:
4342
            assert not create_mixed_batch
4343
            num_reqs = min(max_num_reqs, cdiv(num_tokens, max_query_len))
4344
4345
            num_scheduled_tokens_list = [max_query_len] * num_reqs
            if num_tokens % max_query_len != 0:
4346
                num_scheduled_tokens_list[-1] = num_tokens % max_query_len
4347
4348
4349
4350
4351
4352
        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

4353
4354
        assert sum(num_scheduled_tokens_list) == num_tokens
        assert len(num_scheduled_tokens_list) == num_reqs
4355
        num_scheduled_tokens = np.array(num_scheduled_tokens_list, dtype=np.int32)
4356
4357
        num_tokens_unpadded = int(num_scheduled_tokens.sum())

4358
        num_sampled_tokens = np.ones(num_reqs, dtype=np.int32)
4359

4360
        _cudagraph_mode, batch_desc, should_ubatch, num_tokens_across_dp, _ = (
4361
4362
4363
4364
4365
4366
4367
4368
4369
4370
4371
4372
4373
4374
4375
4376
4377
4378
            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,
4379
4380
            )
        )
4381
4382
4383

        if cudagraph_runtime_mode is None:
            cudagraph_runtime_mode = _cudagraph_mode
4384
        else:
4385
4386
4387
4388
4389
4390
4391
4392
4393
            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
        )
4394
        ubatch_slices, ubatch_slices_padded = maybe_create_ubatch_slices(
4395
4396
4397
4398
4399
4400
4401
4402
4403
4404
            should_ubatch,
            num_scheduled_tokens,
            num_tokens_padded,
            num_reqs_padded,
            self.vllm_config.parallel_config.num_ubatches,
        )
        logger.debug(
            "ubatch_slices: %s, ubatch_slices_padded: %s",
            ubatch_slices,
            ubatch_slices_padded,
4405
        )
4406

4407
        attn_metadata: PerLayerAttnMetadata | None = None
4408
4409
4410

        # If force_attention is True, we always capture attention. Otherwise,
        # it only happens for cudagraph_runtime_mode=FULL.
4411
        if force_attention or cudagraph_runtime_mode == CUDAGraphMode.FULL:
4412
4413
4414
4415
4416
4417
            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:
4418
                seq_lens = max_query_len  # type: ignore[assignment]
4419
            self.seq_lens.np[:num_reqs] = seq_lens
4420
4421
            self.seq_lens.np[num_reqs:] = 0
            self.seq_lens.copy_to_gpu()
4422

4423
4424
            cum_num_tokens, _ = self._get_cumsum_and_arange(num_scheduled_tokens)
            self.query_start_loc.np[1 : num_reqs + 1] = cum_num_tokens
4425
4426
            self.query_start_loc.copy_to_gpu()

4427
            pad_attn = cudagraph_runtime_mode == CUDAGraphMode.FULL
4428
            attn_metadata, _ = self._build_attention_metadata(
4429
4430
4431
                num_tokens=num_tokens_unpadded,
                num_reqs=num_reqs_padded,
                max_query_len=max_query_len,
4432
                ubatch_slices=ubatch_slices_padded if pad_attn else ubatch_slices,
4433
                for_cudagraph_capture=is_graph_capturing,
4434
            )
4435

4436
        with self.maybe_dummy_run_with_lora(
4437
4438
4439
4440
4441
            self.lora_config,
            num_scheduled_tokens,
            num_sampled_tokens,
            activate_lora,
            remove_lora,
4442
        ):
4443
            # Make sure padding doesn't exceed max_num_tokens
4444
            assert num_tokens_padded <= self.max_num_tokens
4445
            model_kwargs = self._init_model_kwargs()
4446
            if self.supports_mm_inputs and not self.model_config.is_encoder_decoder:
Patrick von Platen's avatar
Patrick von Platen committed
4447
4448
                input_ids, inputs_embeds = self._prepare_mm_inputs(num_tokens_padded)

4449
                model_kwargs = {
4450
                    **model_kwargs,
4451
4452
                    **self._dummy_mm_kwargs(num_reqs),
                }
4453
4454
            elif self.enable_prompt_embeds:
                input_ids = None
4455
                inputs_embeds = self.inputs_embeds.gpu[:num_tokens_padded]
4456
                model_kwargs = self._init_model_kwargs()
4457
            else:
4458
                input_ids = self.input_ids.gpu[:num_tokens_padded]
4459
                inputs_embeds = None
4460

4461
            if self.uses_mrope:
4462
                positions = self.mrope_positions.gpu[:, :num_tokens_padded]
4463
            elif self.uses_xdrope_dim > 0:
4464
                positions = self.xdrope_positions.gpu[:, :num_tokens_padded]
4465
            else:
4466
                positions = self.positions.gpu[:num_tokens_padded]
4467
4468
4469
4470
4471
4472
4473
4474
4475

            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,
4476
4477
4478
                            device=self.device,
                        )
                    )
4479
4480

                intermediate_tensors = self.sync_and_slice_intermediate_tensors(
4481
                    num_tokens_padded, None, False
4482
                )
4483

4484
            if ubatch_slices_padded is not None:
4485
4486
4487
                # Adjust values to reflect a single ubatch.
                # TODO(sage,lucas): this is cruft that should be addressed in
                #  the padding refactor.
4488
                num_tokens_padded = ubatch_slices_padded[0].num_tokens
4489
                if num_tokens_across_dp is not None:
4490
                    num_tokens_across_dp[:] = num_tokens_padded
4491

4492
            with (
4493
                self.maybe_randomize_inputs(input_ids, inputs_embeds),
4494
                set_forward_context(
4495
4496
                    attn_metadata,
                    self.vllm_config,
4497
                    num_tokens=num_tokens_padded,
4498
4499
                    num_tokens_across_dp=num_tokens_across_dp,
                    cudagraph_runtime_mode=cudagraph_runtime_mode,
4500
                    batch_descriptor=batch_desc,
4501
                    ubatch_slices=ubatch_slices_padded,
4502
4503
                ),
            ):
4504
                outputs = self.model(
4505
4506
4507
4508
                    input_ids=input_ids,
                    positions=positions,
                    intermediate_tensors=intermediate_tensors,
                    inputs_embeds=inputs_embeds,
4509
                    **model_kwargs,
4510
                )
4511

4512
4513
4514
4515
            if self.use_aux_hidden_state_outputs:
                hidden_states, _ = outputs
            else:
                hidden_states = outputs
4516

4517
4518
4519
4520
4521
4522
            if self.speculative_config and (
                self.speculative_config.use_eagle()
                or self.speculative_config.uses_draft_model()
            ):
                assert isinstance(self.drafter, EagleProposer | DraftModelProposer)
                assert self.speculative_config is not None
4523
4524
4525
                # Eagle currently only supports PIECEWISE cudagraphs.
                # Therefore only use cudagraphs if the main model uses PIECEWISE
                # NOTE(lucas): this is a hack, need to clean up.
4526
                use_cudagraphs = (
4527
4528
4529
4530
4531
4532
4533
4534
4535
                    (
                        is_graph_capturing
                        and cudagraph_runtime_mode == CUDAGraphMode.PIECEWISE
                    )
                    or (
                        not is_graph_capturing
                        and cudagraph_runtime_mode != CUDAGraphMode.NONE
                    )
                ) and not self.speculative_config.enforce_eager
4536
4537
4538
4539
4540
4541
4542
4543
4544
4545
4546

                # 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
4547
                    is_graph_capturing=is_graph_capturing,
4548
                )
4549

4550
4551
4552
4553
4554
4555
4556
4557
4558
4559
4560
        # We register layerwise NVTX hooks here after the first dynamo tracing is
        # done to avoid nvtx operations in hook functions being traced by
        # torch dynamo and causing graph breaks.
        # Note that for DYNAMO_ONCE and VLLM_COMPILE mode,
        # compiled model's dynamo tracing is only done once and the compiled model's
        # __call__ function is replaced by calling the compiled function.
        # So it's safe to register hooks here. Hooks will be registered to
        # both compiled and uncompiled models but they will never
        # be called on the compiled model execution path.
        self._register_layerwise_nvtx_hooks()

4561
4562
4563
4564
4565
4566
4567
4568
4569
4570
        # 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)

4571
        logit_indices = np.cumsum(num_scheduled_tokens) - 1
4572
4573
4574
4575
        logit_indices_device = torch.from_numpy(logit_indices).to(
            self.device, non_blocking=True
        )
        return hidden_states, hidden_states[logit_indices_device]
4576
4577
4578
4579
4580
4581

    @torch.inference_mode()
    def _dummy_sampler_run(
        self,
        hidden_states: torch.Tensor,
    ) -> torch.Tensor:
4582
4583
4584
        # The dummy hidden states may contain special values,
        # like `inf` or `nan`.
        # To avoid breaking the sampler, we use a random tensor here instead.
4585

4586
4587
        mm_config = self.vllm_config.model_config.multimodal_config
        if mm_config and mm_config.mm_encoder_only:
4588
4589
4590
            # MM Encoder only model no need to run sampler.
            return torch.tensor([])

4591
        hidden_states = torch.rand_like(hidden_states)
4592

4593
        logits = self.model.compute_logits(hidden_states)
4594
4595
        num_reqs = logits.size(0)

4596
        dummy_tensors = lambda v: torch.full((num_reqs,), v, device=self.device)
4597
4598
4599
4600
4601
4602
4603
4604
4605
4606
4607
4608
4609
4610
4611

        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)],
4612
            spec_token_ids=[[] for _ in range(num_reqs)],
4613
4614
            allowed_token_ids_mask=None,
            bad_words_token_ids={},
4615
            logitsprocs=LogitsProcessors(),
4616
        )
4617
        try:
4618
4619
4620
            sampler_output = self.sampler(
                logits=logits, sampling_metadata=dummy_metadata
            )
4621
        except RuntimeError as e:
4622
            if "out of memory" in str(e):
4623
4624
4625
4626
                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 "
4627
4628
                    "initializing the engine."
                ) from e
4629
4630
            else:
                raise e
4631
        if self.speculative_config:
4632
4633
            draft_token_ids = [[0] for _ in range(num_reqs)]
            dummy_spec_decode_metadata = SpecDecodeMetadata.make_dummy(
4634
4635
                draft_token_ids, self.device
            )
4636
4637
4638
4639
4640
4641

            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
4642
4643
4644
4645
4646
            logits = torch.randn(
                num_tokens + num_reqs,
                logits.shape[-1],
                device=self.device,
                dtype=logits.dtype,
4647
            )
4648
4649
4650
            self.rejection_sampler(
                dummy_spec_decode_metadata,
                draft_probs,
4651
                logits,
4652
4653
                dummy_metadata,
            )
4654
        return sampler_output
4655

4656
    def _dummy_pooler_run_task(
4657
4658
        self,
        hidden_states: torch.Tensor,
4659
4660
        task: PoolingTask,
    ) -> PoolerOutput:
4661
4662
4663
4664
        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
4665
4666
4667
4668
        num_scheduled_tokens_np = np.full(num_reqs, min_tokens_per_req)
        num_scheduled_tokens_np[-1] += num_tokens % num_reqs
        assert np.sum(num_scheduled_tokens_np) == num_tokens
        assert len(num_scheduled_tokens_np) == num_reqs
4669
4670
4671

        req_num_tokens = num_tokens // num_reqs

4672
        dummy_prompt_lens = torch.from_numpy(num_scheduled_tokens_np)
4673
4674
4675
        dummy_token_ids = torch.zeros(
            (num_reqs, req_num_tokens), dtype=torch.int32, device=self.device
        )
4676

4677
        model = cast(VllmModelForPooling, self.get_model())
4678
        dummy_pooling_params = PoolingParams(task=task)
4679
        dummy_pooling_params.verify(task=task, model_config=self.model_config)
4680
        to_update = model.pooler.get_pooling_updates(task)
4681
4682
        to_update.apply(dummy_pooling_params)

4683
        dummy_metadata = PoolingMetadata(
4684
4685
4686
            prompt_lens=dummy_prompt_lens,
            prompt_token_ids=dummy_token_ids,
            pooling_params=[dummy_pooling_params] * num_reqs,
4687
            pooling_states=[PoolingStates() for i in range(num_reqs)],
4688
        )
4689

4690
        dummy_metadata.build_pooling_cursor(
4691
            num_scheduled_tokens_np,
4692
4693
            seq_lens_cpu=dummy_prompt_lens,
            device=hidden_states.device,
4694
        )
4695

4696
        try:
4697
4698
4699
            return model.pooler(
                hidden_states=hidden_states, pooling_metadata=dummy_metadata
            )
4700
        except RuntimeError as e:
4701
            if "out of memory" in str(e):
4702
                raise RuntimeError(
4703
4704
4705
                    "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 "
4706
4707
                    "initializing the engine."
                ) from e
4708
4709
            else:
                raise e
4710
4711
4712
4713
4714
4715

    @torch.inference_mode()
    def _dummy_pooler_run(
        self,
        hidden_states: torch.Tensor,
    ) -> PoolerOutput:
4716
4717
        mm_config = self.vllm_config.model_config.multimodal_config
        if mm_config and mm_config.mm_encoder_only:
4718
4719
4720
            # MM Encoder only model not need to run pooler.
            return torch.tensor([])

4721
        # Find the task that has the largest output for subsequent steps
4722
4723
4724
        supported_pooling_tasks = self.get_supported_pooling_tasks()

        if not supported_pooling_tasks:
4725
4726
4727
4728
4729
4730
            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."
            )
4731

4732
        output_size = dict[PoolingTask, float]()
4733
        for task in supported_pooling_tasks:
4734
4735
            # Run a full batch with each task to ensure none of them OOMs
            output = self._dummy_pooler_run_task(hidden_states, task)
4736
            output_size[task] = sum(o.nbytes for o in output if o is not None)
4737
4738
4739
4740
            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)
4741

4742
    def profile_run(self) -> None:
4743
        # Profile with multimodal encoder & encoder cache.
4744
        if self.supports_mm_inputs:
4745
4746
            mm_config = self.model_config.multimodal_config
            if mm_config is not None and mm_config.skip_mm_profiling:
4747
                logger.info(
4748
                    "Skipping memory profiling for multimodal encoder and "
4749
4750
                    "encoder cache."
                )
4751
4752
4753
4754
4755
4756
4757
4758
            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.
4759
                    dummy_modality = mm_budget.get_modality_with_max_tokens()
4760
4761
4762
                    max_mm_items_per_batch = mm_budget.max_items_per_batch_by_modality[
                        dummy_modality
                    ]
4763
4764
4765
4766
4767
4768
4769
4770
4771

                    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,
                    )
4772

4773
4774
4775
4776
4777
                    # Create dummy batch of multimodal inputs.
                    batched_dummy_mm_inputs = self._get_mm_dummy_batch(
                        dummy_modality,
                        max_mm_items_per_batch,
                    )
4778

4779
                    # Run multimodal encoder.
4780
                    dummy_encoder_outputs = self.model.embed_multimodal(
4781
4782
                        **batched_dummy_mm_inputs
                    )
4783

4784
4785
4786
4787
                    sanity_check_mm_encoder_outputs(
                        dummy_encoder_outputs,
                        expected_num_items=max_mm_items_per_batch,
                    )
4788
4789
                    for i, output in enumerate(dummy_encoder_outputs):
                        self.encoder_cache[f"tmp_{i}"] = output
4790

4791
        # Add `is_profile` here to pre-allocate communication buffers
4792
4793
4794
        hidden_states, last_hidden_states = self._dummy_run(
            self.max_num_tokens, is_profile=True
        )
4795
        if get_pp_group().is_last_rank:
4796
4797
4798
4799
            if self.is_pooling_model:
                output = self._dummy_pooler_run(hidden_states)
            else:
                output = self._dummy_sampler_run(last_hidden_states)
4800
        else:
4801
            output = None
4802
        self._sync_device()
4803
        del hidden_states, output
4804
        self.encoder_cache.clear()
4805
        gc.collect()
4806

4807
    def capture_model(self) -> int:
4808
        if self.compilation_config.cudagraph_mode == CUDAGraphMode.NONE:
4809
            logger.warning(
4810
                "Skipping CUDA graph capture. To turn on CUDA graph capture, "
4811
4812
                "ensure `cudagraph_mode` was not manually set to `NONE`"
            )
4813
            return 0
4814

4815
4816
        compilation_counter.num_gpu_runner_capture_triggers += 1

4817
4818
        start_time = time.perf_counter()

4819
4820
4821
4822
4823
4824
4825
4826
4827
4828
4829
4830
4831
4832
        @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()
4833
                    gc.collect()
4834

4835
4836
4837
        # 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.
4838
        set_cudagraph_capturing_enabled(True)
4839
        with freeze_gc(), graph_capture(device=self.device):
4840
            start_free_gpu_memory = torch.cuda.mem_get_info()[0]
4841

4842
4843
4844
4845
            for (
                runtime_mode,
                batch_descs,
            ) in self.cudagraph_dispatcher.get_capture_descs():
4846
                self._capture_cudagraphs(
4847
4848
                    batch_descriptors=batch_descs,
                    cudagraph_runtime_mode=runtime_mode,
4849
                )
4850

4851
4852
4853
            torch.cuda.synchronize()
            end_free_gpu_memory = torch.cuda.mem_get_info()[0]

4854
4855
4856
        # 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
4857
        # we may do lazy capturing in future that still allows capturing
4858
4859
        # after here.
        set_cudagraph_capturing_enabled(False)
4860

4861
4862
4863
4864
        # Lock workspace to prevent resizing during execution.
        # Max workspace sizes should have been captured during warmup/profiling.
        lock_workspace()

4865
4866
4867
4868
        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.
4869
        logger.info_once(
4870
4871
4872
            "Graph capturing finished in %.0f secs, took %.2f GiB",
            elapsed_time,
            cuda_graph_size / (1 << 30),
4873
            scope="local",
4874
        )
4875
        return cuda_graph_size
4876

4877
4878
    def _capture_cudagraphs(
        self,
4879
        batch_descriptors: list[BatchDescriptor],
4880
4881
4882
4883
4884
4885
        cudagraph_runtime_mode: CUDAGraphMode,
    ):
        assert (
            cudagraph_runtime_mode != CUDAGraphMode.NONE
            and cudagraph_runtime_mode.valid_runtime_modes()
        ), f"Invalid cudagraph runtime mode: {cudagraph_runtime_mode}"
4886

4887
4888
4889
4890
4891
4892
4893
4894
4895
4896
4897
4898
4899
4900
        if not batch_descriptors:
            return

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

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

4901
4902
        # Only rank 0 should print progress bar during capture
        if is_global_first_rank():
4903
4904
            batch_descriptors = tqdm(
                batch_descriptors,
4905
4906
4907
                disable=not self.load_config.use_tqdm_on_load,
                desc="Capturing CUDA graphs ({}, {})".format(
                    "decode" if uniform_decode else "mixed prefill-decode",
4908
4909
4910
                    cudagraph_runtime_mode.name,
                ),
            )
4911

4912
        # We skip EPLB here since we don't want to record dummy metrics
4913
4914
4915
4916
        for batch_desc in batch_descriptors:
            num_tokens = batch_desc.num_tokens
            activate_lora = batch_desc.has_lora

4917
            # We currently only capture ubatched graphs when its a FULL
4918
4919
4920
            # 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
4921
            allow_microbatching = (
4922
                self.parallel_config.use_ubatching
4923
4924
                and cudagraph_runtime_mode == CUDAGraphMode.FULL
                and uniform_decode
4925
4926
4927
4928
4929
                and check_ubatch_thresholds(
                    config=self.vllm_config.parallel_config,
                    num_tokens=num_tokens,
                    uniform_decode=uniform_decode,
                )
4930
            )
4931

4932
4933
            for _ in range(self.compilation_config.cudagraph_num_of_warmups):
                # Use CUDAGraphRuntimeStyle.NONE (default) for warmup.
4934
                # But be careful, warm up with `NONE` is orthogonal to
4935
4936
4937
                # 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.
4938
                dummy_run(
4939
4940
4941
                    num_tokens,
                    cudagraph_runtime_mode=CUDAGraphMode.NONE,
                    allow_microbatching=allow_microbatching,
4942
                    activate_lora=activate_lora,
4943
                )
4944
4945
4946

            # Capture run
            dummy_run(
4947
4948
4949
                num_tokens,
                cudagraph_runtime_mode=cudagraph_runtime_mode,
                allow_microbatching=allow_microbatching,
4950
                activate_lora=activate_lora,
Rémi Delacourt's avatar
Rémi Delacourt committed
4951
                is_graph_capturing=True,
4952
            )
4953
        self.maybe_remove_all_loras(self.lora_config)
4954

4955
4956
4957
4958
    def initialize_attn_backend(self, kv_cache_config: KVCacheConfig) -> None:
        """
        Initialize the attention backends and attention metadata builders.
        """
4959
        assert len(self.attn_groups) == 0, "Attention backends are already initialized"
4960

4961
4962
4963
4964
4965
4966
        class AttentionGroupKey(NamedTuple):
            attn_backend: type[AttentionBackend]
            kv_cache_spec: KVCacheSpec

        def get_attn_backends_for_group(
            kv_cache_group_spec: KVCacheGroupSpec,
4967
        ) -> tuple[dict[AttentionGroupKey, list[str]], set[type[AttentionBackend]]]:
4968
            layer_type = cast(type[Any], AttentionLayerBase)
4969
            layers = get_layers_from_vllm_config(
4970
                self.vllm_config, layer_type, kv_cache_group_spec.layer_names
4971
            )
4972
4973
            attn_backends = {}
            attn_backend_layers = defaultdict(list)
4974
            # Dedupe based on full class name; this is a bit safer than
4975
4976
4977
4978
            # 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.
4979
            for layer_name in kv_cache_group_spec.layer_names:
4980
                attn_backend = layers[layer_name].get_attn_backend()
4981
4982
4983
4984

                if layer_name in self.kv_sharing_fast_prefill_eligible_layers:
                    attn_backend = create_fast_prefill_custom_backend(
                        "FastPrefill",
4985
                        attn_backend,  # type: ignore[arg-type]
4986
4987
                    )

4988
4989
4990
                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):
4991
                    layer_kv_cache_spec = layer_kv_cache_spec.kv_cache_specs[layer_name]
4992
                key = (full_cls_name, layer_kv_cache_spec)
4993
4994
4995
                attn_backends[key] = AttentionGroupKey(
                    attn_backend, layer_kv_cache_spec
                )
4996
                attn_backend_layers[key].append(layer_name)
4997
4998
4999
5000
            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()),
            )
5001
5002

        def create_attn_groups(
5003
            attn_backends_map: dict[AttentionGroupKey, list[str]],
5004
            kv_cache_group_id: int,
5005
5006
        ) -> list[AttentionGroup]:
            attn_groups: list[AttentionGroup] = []
5007
            for (attn_backend, kv_cache_spec), layer_names in attn_backends_map.items():
5008
                attn_group = AttentionGroup(
5009
                    attn_backend,
5010
                    layer_names,
5011
                    kv_cache_spec,
5012
                    kv_cache_group_id,
5013
5014
                )

5015
5016
5017
                attn_groups.append(attn_group)
            return attn_groups

5018
        attention_backend_maps = []
5019
        attention_backend_list = []
5020
        for kv_cache_group_spec in kv_cache_config.kv_cache_groups:
5021
            attn_backends = get_attn_backends_for_group(kv_cache_group_spec)
5022
            attention_backend_maps.append(attn_backends[0])
5023
            attention_backend_list.append(attn_backends[1])
5024
5025

        # Resolve cudagraph_mode before actually initialize metadata_builders
5026
5027
5028
        self._check_and_update_cudagraph_mode(
            attention_backend_list, kv_cache_config.kv_cache_groups
        )
5029

5030
5031
5032
        # Check if attention backend supports PCP&DCP and related features.
        check_attention_cp_compatibility(self.vllm_config)

5033
5034
        for i, attn_backend_map in enumerate(attention_backend_maps):
            self.attn_groups.append(create_attn_groups(attn_backend_map, i))
5035

5036
5037
5038
5039
5040
5041
5042
5043
5044
5045
5046
5047
5048
5049
5050
    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
5051
5052
                    if not self.parallel_config.use_ubatching
                    else self.parallel_config.num_ubatches,
5053
                )
co63oc's avatar
co63oc committed
5054
        # Calculate reorder batch threshold (if needed)
5055
5056
        # Note (tdoublep): do this *after* constructing builders,
        # because some of them change the threshold at init time.
5057
5058
        self.calculate_reorder_batch_threshold()

5059
    def _check_and_update_cudagraph_mode(
5060
5061
5062
        self,
        attention_backends: list[set[type[AttentionBackend]]],
        kv_cache_groups: list[KVCacheGroupSpec],
5063
    ) -> None:
5064
        """
5065
        Resolve the cudagraph_mode when there are multiple attention
5066
        groups with potential conflicting CUDA graph support.
5067
5068
5069
        Then initialize the cudagraph_dispatcher based on the resolved
        cudagraph_mode.
        """
5070
        min_cg_support = AttentionCGSupport.ALWAYS
5071
        min_cg_backend_name = None
5072

5073
5074
5075
5076
5077
5078
5079
5080
5081
5082
5083
5084
        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__
5085
5086
        # Flexible resolve the cudagraph mode
        cudagraph_mode = self.compilation_config.cudagraph_mode
5087
        assert cudagraph_mode is not None
5088
        # check cudagraph for mixed batch is supported
5089
5090
5091
5092
5093
5094
        if (
            cudagraph_mode.mixed_mode() == CUDAGraphMode.FULL
            and min_cg_support != AttentionCGSupport.ALWAYS
        ):
            msg = (
                f"CUDAGraphMode.{cudagraph_mode.name} is not supported "
5095
                f"with {min_cg_backend_name} backend (support: "
5096
5097
                f"{min_cg_support})"
            )
5098
5099
            if min_cg_support == AttentionCGSupport.NEVER:
                # if not supported any full cudagraphs, just raise it.
5100
5101
                msg += (
                    "; please try cudagraph_mode=PIECEWISE, and "
5102
                    "make sure compilation mode is VLLM_COMPILE"
5103
                )
5104
5105
5106
5107
5108
                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"
5109
                cudagraph_mode = self.compilation_config.cudagraph_mode = (
5110
                    CUDAGraphMode.FULL_AND_PIECEWISE
5111
                )
5112
5113
            else:
                msg += "; setting cudagraph_mode=FULL_DECODE_ONLY"
5114
                cudagraph_mode = self.compilation_config.cudagraph_mode = (
5115
                    CUDAGraphMode.FULL_DECODE_ONLY
5116
                )
5117
5118
            logger.warning(msg)

5119
        # check that if we are doing decode full-cudagraphs it is supported
5120
5121
5122
5123
5124
5125
        if (
            cudagraph_mode.decode_mode() == CUDAGraphMode.FULL
            and min_cg_support == AttentionCGSupport.NEVER
        ):
            msg = (
                f"CUDAGraphMode.{cudagraph_mode.name} is not supported "
5126
                f"with {min_cg_backend_name} backend (support: "
5127
5128
                f"{min_cg_support})"
            )
5129
            if self.compilation_config.mode == CompilationMode.VLLM_COMPILE and (
5130
5131
5132
5133
5134
                self.compilation_config.splitting_ops_contain_attention()
                or self.compilation_config.use_inductor_graph_partition
            ):
                msg += (
                    "; setting cudagraph_mode=PIECEWISE because "
5135
                    "attention is compiled piecewise"
5136
5137
                )
                cudagraph_mode = self.compilation_config.cudagraph_mode = (
5138
                    CUDAGraphMode.PIECEWISE
5139
                )
5140
            else:
5141
5142
                msg += (
                    "; setting cudagraph_mode=NONE because "
5143
                    "attention is not compiled piecewise"
5144
5145
                )
                cudagraph_mode = self.compilation_config.cudagraph_mode = (
5146
                    CUDAGraphMode.NONE
5147
                )
5148
5149
            logger.warning(msg)

5150
5151
        # check that if we are doing spec-decode + decode full-cudagraphs it is
        # supported
5152
5153
5154
5155
5156
5157
5158
5159
        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 "
5160
                f"{min_cg_backend_name} (support: {min_cg_support})"
5161
            )
5162
5163
            if self.compilation_config.splitting_ops_contain_attention():
                msg += "; setting cudagraph_mode=PIECEWISE"
5164
                cudagraph_mode = self.compilation_config.cudagraph_mode = (
5165
                    CUDAGraphMode.PIECEWISE
5166
                )
5167
5168
            else:
                msg += "; setting cudagraph_mode=NONE"
5169
                cudagraph_mode = self.compilation_config.cudagraph_mode = (
5170
                    CUDAGraphMode.NONE
5171
                )
5172
5173
5174
5175
            logger.warning(msg)

        # double check that we can support full cudagraph if they are requested
        # even after automatic downgrades
5176
5177
5178
5179
5180
5181
        if (
            cudagraph_mode.has_full_cudagraphs()
            and min_cg_support == AttentionCGSupport.NEVER
        ):
            raise ValueError(
                f"CUDAGraphMode.{cudagraph_mode.name} is not "
5182
                f"supported with {min_cg_backend_name} backend ("
5183
5184
                f"support:{min_cg_support}) "
                "; please try cudagraph_mode=PIECEWISE, "
5185
                "and make sure compilation mode is VLLM_COMPILE"
5186
            )
5187

5188
5189
5190
5191
        # 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
5192
        # Will be removed in the near future when we have separate cudagraph capture
5193
5194
5195
5196
5197
5198
5199
5200
5201
        # 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
            )
5202
5203
5204
5205
            capture_sizes = self.compilation_config.cudagraph_capture_sizes
            self.cudagraph_batch_sizes = (
                capture_sizes if capture_sizes is not None else []
            )
5206

5207
5208
        # Trigger cudagraph dispatching keys initialization after
        # resolved cudagraph mode.
5209
        self.compilation_config.cudagraph_mode = cudagraph_mode
5210
        self.cudagraph_dispatcher.initialize_cudagraph_keys(
5211
            cudagraph_mode, self.uniform_decode_query_len
5212
        )
5213

5214
5215
5216
5217
5218
        # Initialize eagle's cudagraph dispatcher if using eagle spec decode.
        if self.speculative_config and self.speculative_config.use_eagle():
            assert isinstance(self.drafter, EagleProposer)
            self.drafter.initialize_cudagraph_keys(cudagraph_mode)

5219
5220
    def calculate_reorder_batch_threshold(self) -> None:
        """
5221
5222
5223
5224
        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.
5225
        """
5226
5227
        min_none_high = lambda a, b: a if b is None else b if a is None else min(a, b)

5228
        reorder_batch_thresholds: list[int | None] = [
5229
5230
5231
            group.get_metadata_builder().reorder_batch_threshold
            for group in self._attn_group_iterator()
        ]
5232
5233
5234
5235
5236
        # 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
5237
        self.reorder_batch_threshold = reduce(min_none_high, reorder_batch_thresholds)  # type: ignore[assignment]
5238

5239
5240
5241
    @staticmethod
    def select_common_block_size(
        kv_manager_block_size: int, attn_groups: list[AttentionGroup]
5242
5243
    ) -> int:
        """
5244
5245
5246
5247
5248
        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.
5249
5250
5251
5252
5253
5254

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

        Returns:
5255
            The selected block size
5256
5257

        Raises:
5258
            ValueError: If no valid block size found
5259
5260
        """

5261
5262
5263
5264
5265
5266
5267
5268
        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
5269
                for supported_size in backend.get_supported_kernel_block_sizes():
5270
5271
5272
5273
5274
5275
5276
5277
5278
5279
5280
5281
5282
5283
5284
5285
5286
5287
5288
5289
5290
5291
5292
5293
5294
5295
5296
5297
5298
5299
                    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
5300
            for supported_size in backend.get_supported_kernel_block_sizes()
5301
5302
            if isinstance(supported_size, int)
        )
5303

5304
5305
5306
5307
5308
5309
        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}. ")
5310

5311
5312
5313
    def may_reinitialize_input_batch(
        self, kv_cache_config: KVCacheConfig, kernel_block_sizes: list[int]
    ) -> None:
5314
5315
5316
5317
5318
5319
5320
        """
        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.
5321
            kernel_block_sizes: The kernel block sizes for each KV cache group.
5322
5323
5324
5325
        """
        block_sizes = [
            kv_cache_group.kv_cache_spec.block_size
            for kv_cache_group in kv_cache_config.kv_cache_groups
5326
            if not isinstance(kv_cache_group.kv_cache_spec, EncoderOnlyAttentionSpec)
5327
        ]
5328
5329
5330
5331
5332
5333
5334
5335
5336
5337
5338
5339
5340
5341
5342
5343
5344
5345
        max_num_blocks = []
        max_model_len = max(self.max_model_len, self.max_encoder_len)
        for i, kv_cache_group in enumerate(kv_cache_config.kv_cache_groups):
            if isinstance(kv_cache_group.kv_cache_spec, EncoderOnlyAttentionSpec):
                continue
            max_num_blocks_per_req = cdiv(
                max_model_len, block_sizes[i] * get_total_cp_world_size()
            )
            if isinstance(kv_cache_group.kv_cache_spec, MambaSpec):
                mamba_blocks_per_req = (
                    max_num_blocks_per_req
                    if self.cache_config.enable_prefix_caching
                    else 1
                ) + kv_cache_group.kv_cache_spec.num_speculative_blocks
                max_num_blocks_per_req = max(
                    max_num_blocks_per_req, mamba_blocks_per_req
                )
            max_num_blocks.append(max_num_blocks_per_req)
5346
5347
5348
5349

        if block_sizes != [self.cache_config.block_size] or kernel_block_sizes != [
            self.cache_config.block_size
        ]:
5350
5351
5352
            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
5353
5354
                "for more details."
            )
5355
5356
            self.input_batch = InputBatch(
                max_num_reqs=self.max_num_reqs,
5357
                max_model_len=max_model_len,
5358
5359
5360
5361
5362
                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,
5363
                kernel_block_sizes=kernel_block_sizes,
5364
                max_num_blocks_per_req=max_num_blocks,
5365
                is_spec_decode=bool(self.vllm_config.speculative_config),
5366
                logitsprocs=self.input_batch.logitsprocs,
5367
                logitsprocs_need_output_token_ids=self.input_batch.logitsprocs_need_output_token_ids,
5368
                is_pooling_model=self.is_pooling_model,
5369
5370
            )

5371
    def _allocate_kv_cache_tensors(
5372
5373
        self, kv_cache_config: KVCacheConfig
    ) -> dict[str, torch.Tensor]:
5374
        """
5375
5376
5377
        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.

5378
        Args:
5379
            kv_cache_config: The KV cache config
5380
        Returns:
5381
            dict[str, torch.Tensor]: A map between layer names to their
5382
            corresponding memory buffer for KV cache.
5383
        """
5384
5385
        kv_cache_raw_tensors: dict[str, torch.Tensor] = {}
        for kv_cache_tensor in kv_cache_config.kv_cache_tensors:
5386
5387
5388
            tensor = torch.zeros(
                kv_cache_tensor.size, dtype=torch.int8, device=self.device
            )
5389
5390
5391
5392
5393
            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:
5394
5395
5396
5397
            for layer_name in group.layer_names:
                if layer_name in self.runner_only_attn_layers:
                    continue
                layer_names.add(layer_name)
5398
5399
5400
        assert layer_names == set(kv_cache_raw_tensors.keys()), (
            "Some layers are not correctly initialized"
        )
5401
5402
        return kv_cache_raw_tensors

5403
5404
5405
    def _attn_group_iterator(self) -> Iterator[AttentionGroup]:
        return itertools.chain.from_iterable(self.attn_groups)

5406
    def _kv_cache_spec_attn_group_iterator(self) -> Iterator[AttentionGroup]:
5407
5408
        if not self.kv_cache_config.kv_cache_groups:
            return
5409
5410
        for attn_groups in self.attn_groups:
            yield from attn_groups
5411

5412
5413
5414
5415
5416
5417
5418
5419
5420
5421
5422
5423
5424
5425
5426
    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 = []
5427
        for kv_cache_gid, kv_cache_group in enumerate(kv_cache_config.kv_cache_groups):
5428
5429
5430
5431
5432
5433
            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):
5434
                continue
5435
            elif isinstance(kv_cache_spec, AttentionSpec):
5436
5437
5438
                # This is an attention backend that supports virtual
                # block splitting. Get the supported block sizes from
                # all backends in the group.
5439
                attn_groups = self.attn_groups[kv_cache_gid]
5440
                kv_manager_block_size = kv_cache_group.kv_cache_spec.block_size
5441
                selected_kernel_size = self.select_common_block_size(
5442
5443
5444
                    kv_manager_block_size, attn_groups
                )
                kernel_block_sizes.append(selected_kernel_size)
5445
            elif isinstance(kv_cache_spec, MambaSpec):
5446
5447
                # This is likely Mamba or other non-attention cache,
                # no splitting.
5448
                kernel_block_sizes.append(kv_cache_spec.block_size)
5449
5450
5451
5452
5453
5454
            else:
                raise NotImplementedError(
                    f"unknown kv cache spec {kv_cache_group.kv_cache_spec}"
                )
        return kernel_block_sizes

5455
5456
5457
5458
    def _reshape_kv_cache_tensors(
        self,
        kv_cache_config: KVCacheConfig,
        kv_cache_raw_tensors: dict[str, torch.Tensor],
5459
        kernel_block_sizes: list[int],
5460
    ) -> dict[str, torch.Tensor]:
5461
        """
5462
        Reshape the KV cache tensors to the desired shape and dtype.
5463

5464
        Args:
5465
5466
            kv_cache_config: The KV cache config
            kv_cache_raw_tensors: The KV cache buffer of each layer, with
5467
                correct size but uninitialized shape.
5468
            kernel_block_sizes: The kernel block sizes for each KV cache group.
5469
        Returns:
5470
            Dict[str, torch.Tensor]: A map between layer names to their
5471
5472
            corresponding memory buffer for KV cache.
        """
5473
        kv_caches: dict[str, torch.Tensor] = {}
5474
        has_attn, has_mamba = False, False
5475
5476
        for group in self._kv_cache_spec_attn_group_iterator():
            kv_cache_spec = group.kv_cache_spec
5477
            attn_backend = group.backend
5478
5479
5480
5481
            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]
5482
            for layer_name in group.layer_names:
5483
5484
                if layer_name in self.runner_only_attn_layers:
                    continue
5485
5486
                raw_tensor = kv_cache_raw_tensors[layer_name]
                assert raw_tensor.numel() % kv_cache_spec.page_size_bytes == 0
5487
                num_blocks = raw_tensor.numel() // kv_cache_spec.page_size_bytes
5488
                if isinstance(kv_cache_spec, AttentionSpec):
5489
                    has_attn = True
5490
5491
                    num_blocks_per_kv_block = (
                        kv_cache_spec.block_size // kernel_block_size
5492
5493
5494
                    )
                    kernel_num_blocks = num_blocks * num_blocks_per_kv_block

5495
                    kv_cache_shape = attn_backend.get_kv_cache_shape(
5496
                        kernel_num_blocks,
5497
                        kernel_block_size,
5498
5499
                        kv_cache_spec.num_kv_heads,
                        kv_cache_spec.head_size,
5500
5501
                        cache_dtype_str=self.cache_config.cache_dtype,
                    )
5502
                    dtype = kv_cache_spec.dtype
5503
                    try:
5504
                        kv_cache_stride_order = attn_backend.get_kv_cache_stride_order()
5505
                        assert len(kv_cache_stride_order) == len(kv_cache_shape)
5506
                    except (AttributeError, NotImplementedError):
5507
                        kv_cache_stride_order = tuple(range(len(kv_cache_shape)))
5508
5509
5510
5511
5512
                    # 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.
5513
5514
5515
                    kv_cache_shape = tuple(
                        kv_cache_shape[i] for i in kv_cache_stride_order
                    )
5516
5517
5518
5519
5520
                    # Maintain original KV shape view.
                    inv_order = [
                        kv_cache_stride_order.index(i)
                        for i in range(len(kv_cache_stride_order))
                    ]
5521
5522
5523
5524
5525
5526
                    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
5527
                elif isinstance(kv_cache_spec, MambaSpec):
5528
                    has_mamba = True
Chen Zhang's avatar
Chen Zhang committed
5529
5530
                    raw_tensor = kv_cache_raw_tensors[layer_name]
                    state_tensors = []
5531
                    storage_offset_bytes = 0
5532
                    for shape, dtype in zip(kv_cache_spec.shapes, kv_cache_spec.dtypes):
5533
5534
                        dtype_size = get_dtype_size(dtype)
                        num_element_per_page = (
5535
5536
                            kv_cache_spec.page_size_bytes // dtype_size
                        )
Chen Zhang's avatar
Chen Zhang committed
5537
                        target_shape = (num_blocks, *shape)
5538
5539
                        stride = torch.empty(target_shape).stride()
                        target_stride = (num_element_per_page, *stride[1:])
5540
                        assert storage_offset_bytes % dtype_size == 0
5541
5542
5543
5544
                        tensor = torch.as_strided(
                            raw_tensor.view(dtype),
                            size=target_shape,
                            stride=target_stride,
5545
                            storage_offset=storage_offset_bytes // dtype_size,
5546
                        )
Chen Zhang's avatar
Chen Zhang committed
5547
                        state_tensors.append(tensor)
5548
                        storage_offset_bytes += stride[0] * dtype_size
5549
5550

                    kv_caches[layer_name] = state_tensors
5551
                else:
5552
                    raise NotImplementedError
5553
5554

        if has_attn and has_mamba:
5555
            self._update_hybrid_attention_mamba_layout(kv_caches)
5556

5557
5558
        return kv_caches

5559
    def _update_hybrid_attention_mamba_layout(
5560
5561
        self, kv_caches: dict[str, torch.Tensor]
    ) -> None:
5562
        """
5563
5564
        Update the layout of attention layers from (2, num_blocks, ...) to
        (num_blocks, 2, ...).
5565
5566

        Args:
5567
            kv_caches: The KV cache buffer of each layer.
5568
5569
        """

5570
5571
        for group in self._kv_cache_spec_attn_group_iterator():
            kv_cache_spec = group.kv_cache_spec
5572
            for layer_name in group.layer_names:
5573
                kv_cache = kv_caches[layer_name]
5574
5575
5576
5577
                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 "
5578
                        f"a tensor of shape {kv_cache.shape}"
5579
                    )
5580
                    hidden_size = kv_cache.shape[2:].numel()
5581
5582
5583
5584
                    kv_cache.as_strided_(
                        size=kv_cache.shape,
                        stride=(hidden_size, 2 * hidden_size, *kv_cache.stride()[2:]),
                    )
5585

5586
    def initialize_kv_cache_tensors(
5587
        self, kv_cache_config: KVCacheConfig, kernel_block_sizes: list[int]
5588
    ) -> dict[str, torch.Tensor]:
5589
5590
5591
5592
5593
        """
        Initialize the memory buffer for KV cache.

        Args:
            kv_cache_config: The KV cache config
5594
5595
            kernel_block_sizes: The kernel block sizes for each KV cache group.

5596
        Returns:
5597
            Dict[str, torch.Tensor]: A map between layer names to their
5598
5599
            corresponding memory buffer for KV cache.
        """
5600
5601
5602
5603
5604
5605
5606
5607
5608
5609
5610
5611
5612
5613
5614
5615
5616
5617
5618
5619
5620
5621
5622
5623

        # 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
            )
5624

5625
        # Set up cross-layer KV cache sharing
5626
5627
        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)
5628
5629
            kv_caches[layer_name] = kv_caches[target_layer_name]

5630
5631
5632
5633
5634
5635
5636
5637
5638
        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,
        )
5639
5640
5641
        return kv_caches

    def maybe_add_kv_sharing_layers_to_kv_cache_groups(
5642
5643
        self, kv_cache_config: KVCacheConfig
    ) -> None:
5644
5645
5646
5647
5648
5649
5650
5651
5652
5653
5654
5655
5656
5657
5658
5659
5660
5661
        """
        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.
5662
            attn_layers = get_layers_from_vllm_config(self.vllm_config, Attention)
5663
5664
            for layer_name in reversed(attn_layers):
                if layer_name in self.shared_kv_cache_layers:
5665
                    self.kv_sharing_fast_prefill_eligible_layers.add(layer_name)
5666
5667
                else:
                    break
5668

5669
5670
5671
5672
5673
5674
5675
    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
        """
5676
        kv_cache_config = deepcopy(kv_cache_config)
5677
        self.kv_cache_config = kv_cache_config
5678
        self.may_add_encoder_only_layers_to_kv_cache_config()
5679
        self.maybe_add_kv_sharing_layers_to_kv_cache_groups(kv_cache_config)
5680
        self.initialize_attn_backend(kv_cache_config)
5681
5682
5683
5684
5685
5686
        # 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)
5687
5688
5689
5690

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

5691
        # Reinitialize need to after initialize_attn_backend
5692
5693
5694
5695
        self.may_reinitialize_input_batch(kv_cache_config, kernel_block_sizes)
        kv_caches = self.initialize_kv_cache_tensors(
            kv_cache_config, kernel_block_sizes
        )
5696

5697
5698
5699
5700
5701
        if self.speculative_config and (
            self.speculative_config.use_eagle()
            or self.speculative_config.uses_draft_model()
        ):
            assert isinstance(self.drafter, EagleProposer | DraftModelProposer)
5702
5703
5704
5705
            # 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
5706
        if has_kv_transfer_group():
5707
            kv_transfer_group = get_kv_transfer_group()
5708
5709
5710
5711
5712
5713
5714
            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)
5715
            kv_transfer_group.set_host_xfer_buffer_ops(copy_kv_blocks)
Robert Shaw's avatar
Robert Shaw committed
5716

5717
5718
5719
5720
5721
5722
5723
5724
5725
5726
5727
5728
5729
5730
5731
5732
5733
        if self.model_config.enable_return_routed_experts:
            self.init_routed_experts_capturer()

    def init_routed_experts_capturer(self):
        logger.info(
            "Initializing routed experts capturer, enable_return_routed_experts: %s",
            self.model_config.enable_return_routed_experts,
        )
        routed_experts_capturer = RoutedExpertsCapturer.create()
        block_size = self.cache_config.block_size
        self.max_num_kv_tokens = (
            self.kv_cache_config.num_blocks // len(self.kv_cache_config.kv_cache_groups)
            + 1
        ) * block_size
        routed_experts_capturer.init_buffer(
            max_num_batched_tokens=self.scheduler_config.max_num_batched_tokens,
            max_num_kv_tokens=self.max_num_kv_tokens,
5734
            vllm_config=self.vllm_config,
5735
5736
        )

5737
5738
5739
5740
5741
    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
5742
        encoder_only_attn_specs: dict[AttentionSpec, list[str]] = defaultdict(list)
5743
5744
5745
        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:
5746
                attn_spec: AttentionSpec = EncoderOnlyAttentionSpec(
5747
5748
5749
                    block_size=block_size,
                    num_kv_heads=attn_module.num_kv_heads,
                    head_size=attn_module.head_size,
5750
5751
                    dtype=self.kv_cache_dtype,
                )
5752
5753
5754
                encoder_only_attn_specs[attn_spec].append(layer_name)
                self.runner_only_attn_layers.add(layer_name)
        if len(encoder_only_attn_specs) > 0:
5755
5756
5757
            assert len(encoder_only_attn_specs) == 1, (
                "Only support one encoder-only attention spec now"
            )
5758
5759
            spec, layer_names = encoder_only_attn_specs.popitem()
            self.kv_cache_config.kv_cache_groups.append(
5760
5761
                KVCacheGroupSpec(layer_names=layer_names, kv_cache_spec=spec)
            )
5762

5763
    def get_kv_cache_spec(self) -> dict[str, KVCacheSpec]:
5764
        """
5765
        Generates the KVCacheSpec by parsing the kv cache format from each
5766
5767
        Attention module in the static forward context.
        Returns:
5768
            KVCacheSpec: A dictionary mapping layer names to their KV cache
5769
5770
            format. Layers that do not need KV cache are not included.
        """
5771
5772
        if has_ec_transfer() and get_ec_transfer().is_producer:
            return {}
5773
        kv_cache_spec: dict[str, KVCacheSpec] = {}
5774
5775
        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
5776
        for layer_name, attn_module in attn_layers.items():
5777
5778
5779
5780
5781
5782
5783
5784
5785
5786
5787
5788
5789
5790
5791
            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
5792

5793
        return kv_cache_spec
5794

5795
    def _to_list(self, sampled_token_ids: torch.Tensor) -> list[list[int]]:
5796
5797
5798
5799
5800
5801
5802
5803
        # 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.
5804
        pinned = self.sampled_token_ids_pinned_cpu[: sampled_token_ids.shape[0]]
5805
5806
5807
        pinned.copy_(sampled_token_ids, non_blocking=True)
        self.transfer_event.record()
        self.transfer_event.synchronize()
5808
        return pinned.tolist()