gpu_model_runner.py 247 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 itertools import product
14
from typing import TYPE_CHECKING, Any, NamedTuple, TypeAlias, cast
15
16
17
18
19

import numpy as np
import torch
import torch.distributed
import torch.nn as nn
20
from tqdm import tqdm
21

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

172
173
174
175
176
177
178
from .utils import (
    AttentionGroup,
    MultiModalBudget,
    add_kv_sharing_layers_to_kv_cache_groups,
    bind_kv_cache,
    sanity_check_mm_encoder_outputs,
)
179

180
if TYPE_CHECKING:
181
    from vllm.model_executor.model_loader.tensorizer import TensorizerConfig
182
    from vllm.v1.core.sched.output import GrammarOutput, SchedulerOutput
183
184
185

logger = init_logger(__name__)

186
187
AttnMetadataDict: TypeAlias = dict[str, AttentionMetadata]
# list when ubatching is enabled
188
PerLayerAttnMetadata: TypeAlias = list[AttnMetadataDict] | AttnMetadataDict
189

190

191
192
193
194
195
196
# Wrapper for ModelRunnerOutput to support overlapped execution.
class AsyncGPUModelRunnerOutput(AsyncModelRunnerOutput):
    def __init__(
        self,
        model_runner_output: ModelRunnerOutput,
        sampled_token_ids: torch.Tensor,
197
        logprobs_tensors: LogprobsTensors | None,
198
199
        invalid_req_indices: list[int],
        async_output_copy_stream: torch.cuda.Stream,
200
        vocab_size: int,
201
202
203
204
205
    ):
        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.
206
        self.async_copy_ready_event = torch.Event()
207
208
209
210

        # 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
211
        self.vocab_size = vocab_size
212
        self._logprobs_tensors = logprobs_tensors
213
214
215
216
217

        # 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)
218
            self.sampled_token_ids_cpu = self._sampled_token_ids.to(
219
220
                "cpu", non_blocking=True
            )
221
222
223
224
225
            self._logprobs_tensors_cpu = (
                self._logprobs_tensors.to_cpu_nonblocking()
                if self._logprobs_tensors
                else None
            )
226
            self.async_copy_ready_event.record()
227
228
229

    def get_output(self) -> ModelRunnerOutput:
        """Copy the device tensors to the host and return a ModelRunnerOutput.
230

231
232
        This function blocks until the copy is finished.
        """
233
        max_gen_len = self.sampled_token_ids_cpu.shape[-1]
234
        self.async_copy_ready_event.synchronize()
235

236
237
        # Release the device tensors once the copy has completed.
        del self._logprobs_tensors
238
        del self._sampled_token_ids
239
        if max_gen_len == 1:
240
            valid_sampled_token_ids = self.sampled_token_ids_cpu.tolist()
241
242
            for i in self._invalid_req_indices:
                valid_sampled_token_ids[i].clear()
243
244
245
            logprobs_lists = None
            if self._logprobs_tensors_cpu is not None:
                logprobs_lists = self._logprobs_tensors_cpu.tolists()
246
        else:
247
            valid_sampled_token_ids, logprobs_lists = RejectionSampler.parse_output(
248
249
                self.sampled_token_ids_cpu,
                self.vocab_size,
250
                self._invalid_req_indices,
251
                logprobs_tensors=self._logprobs_tensors_cpu,
252
            )
253
254
255

        output = self._model_runner_output
        output.sampled_token_ids = valid_sampled_token_ids
256
        output.logprobs = logprobs_lists
257
258
259
        return output


260
261
262
263
264
265
266
267
268
269
270
271
272
273
274
275
276
277
278
279
280
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)
281
            raw_pooler_output_cpu = json_map_leaves(
282
283
284
285
                lambda x: None if x is None else x.to("cpu", non_blocking=True),
                self._raw_pooler_output,
            )
            self.async_copy_ready_event.record()
286
287
288
289
            self._model_runner_output.pooler_output = [
                out if include else None
                for out, include in zip(raw_pooler_output_cpu, finished_mask)
            ]
290
291
292
293
294
295
296
297
298
299
300
301

    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


302
303
304
305
306
307
308
309
310
311
312
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
313
    ec_connector_output: ECConnectorOutput | None
314
    cudagraph_stats: CUDAGraphStat | None
315
316


317
318
319
class GPUModelRunner(
    LoRAModelRunnerMixin, KVConnectorModelRunnerMixin, ECConnectorModelRunnerMixin
):
320
321
    def __init__(
        self,
322
        vllm_config: VllmConfig,
323
        device: torch.device,
324
    ):
325
326
327
        self.vllm_config = vllm_config
        self.model_config = vllm_config.model_config
        self.cache_config = vllm_config.cache_config
328
        self.compilation_config = vllm_config.compilation_config
329
330
331
332
333
334
        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
335

336
        from vllm.model_executor.models.utils import set_cpu_offload_max_bytes
337
338

        set_cpu_offload_max_bytes(int(self.cache_config.cpu_offload_gb * 1024**3))
339

340
341
342
343
        model_config = self.model_config
        cache_config = self.cache_config
        scheduler_config = self.scheduler_config
        parallel_config = self.parallel_config
344
        self.device = device
345
346
        self.pin_memory = is_pin_memory_available()
        self.dtype = self.model_config.dtype
347

348
349
350
        self.kv_cache_dtype = kv_cache_dtype_str_to_dtype(
            cache_config.cache_dtype, self.model_config
        )
351

352
        self.is_pooling_model = model_config.runner_type == "pooling"
353
        self.enable_prompt_embeds = model_config.enable_prompt_embeds
354
        self.is_multimodal_raw_input_only_model = (
355
356
            model_config.is_multimodal_raw_input_only_model
        )
357
358
        # This will be overridden in load_model()
        self.is_multimodal_pruning_enabled = False
359
        self.max_model_len = model_config.max_model_len
360
361
362

        # Always set to false after the first forward pass
        self.calculate_kv_scales = self.cache_config.calculate_kv_scales
363
        self.dcp_world_size = self.parallel_config.decode_context_parallel_size
364
        self.dcp_rank = 0 if self.dcp_world_size <= 1 else get_dcp_group().rank_in_group
365
        self.max_num_tokens = scheduler_config.max_num_batched_tokens
366
        self.max_num_reqs = scheduler_config.max_num_seqs
367

368
369
370
371
372
        # 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 = (
373
            self.parallel_config.distributed_executor_backend == "external_launcher"
374
            and len(get_pp_group().ranks) > 1
375
        )
376

377
        # Model-related.
378
        self.num_query_heads = model_config.get_num_attention_heads(parallel_config)
379
        self.inputs_embeds_size = model_config.get_inputs_embeds_size()
380
        self.attention_chunk_size = model_config.attention_chunk_size
381
        # Only relevant for models using ALiBi (e.g, MPT)
382
        self.use_alibi = model_config.uses_alibi
383

384
        self.cascade_attn_enabled = not self.model_config.disable_cascade_attn
385
        self.is_mm_prefix_lm = self.model_config.is_mm_prefix_lm
386

387
        # Multi-modal data support
388
        self.mm_registry = MULTIMODAL_REGISTRY
389
        self.uses_mrope = model_config.uses_mrope
390
        self.uses_xdrope_dim = model_config.uses_xdrope_dim
391
        self.supports_mm_inputs = self.mm_registry.supports_multimodal_inputs(
392
            model_config
393
        )
394

395
396
397
        if self.model_config.is_encoder_decoder:
            # Maximum length of the encoder input, only for encoder-decoder
            # models.
398
            self.max_encoder_len = scheduler_config.max_num_encoder_input_tokens
399
400
401
        else:
            self.max_encoder_len = 0

402
403
404
        # Async scheduling
        self.use_async_scheduling = self.scheduler_config.async_scheduling

405
        # Sampler
406
        self.sampler = Sampler(logprobs_mode=self.model_config.logprobs_mode)
407

408
        self.eplb_state: EplbState | None = None
409
410
411
412
413
414
        """
        State of the expert parallelism load balancer.

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

415
        # Lazy initializations
416
        # self.model: nn.Module  # Set after load_model
417
        # Initialize in initialize_kv_cache
418
        self.kv_caches: list[torch.Tensor] = []
419
420
421
        # Initialize in initialize_kv_cache_tensors
        self.cross_layers_kv_cache: torch.Tensor | None = None
        self.cross_layers_attn_backend: type[AttentionBackend] | None = None
422
423
        # indexes: [kv_cache_group_id][attn_group]
        self.attn_groups: list[list[AttentionGroup]] = []
424
425
        # self.kv_cache_config: KVCacheConfig

426
427
        # mm_hash ->  encoder_output
        self.encoder_cache: dict[str, torch.Tensor] = {}
428

429
        self.use_aux_hidden_state_outputs = False
430
431
432
433
434
        # 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:
435
436
437
            self.drafter: (
                NgramProposer | SuffixDecodingProposer | EagleProposer | MedusaProposer
            )
438
439
            if self.speculative_config.method == "ngram":
                self.drafter = NgramProposer(self.vllm_config)
440
441
            elif self.speculative_config.method == "suffix":
                self.drafter = SuffixDecodingProposer(self.vllm_config)
442
            elif self.speculative_config.use_eagle():
443
                self.drafter = EagleProposer(self.vllm_config, self.device, self)
444
                if self.speculative_config.method == "eagle3":
445
446
447
                    self.use_aux_hidden_state_outputs = (
                        self.drafter.eagle3_use_aux_hidden_state
                    )
448
449
            elif self.speculative_config.method == "medusa":
                self.drafter = MedusaProposer(
450
                    vllm_config=self.vllm_config, device=self.device
451
                )
452
            else:
453
454
455
456
                raise ValueError(
                    "Unknown speculative decoding method: "
                    f"{self.speculative_config.method}"
                )
457
            self.rejection_sampler = RejectionSampler(self.sampler)
458

459
460
461
        self.num_spec_tokens = 0
        if self.speculative_config:
            self.num_spec_tokens = self.speculative_config.num_speculative_tokens
462
463
464
465
466
            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
467

468
        # Request states.
469
        self.requests: dict[str, CachedRequestState] = {}
470
471
472
        # NOTE(rob): num_prompt_logprobs only includes reqs
        # that are currently in the prefill phase.
        self.num_prompt_logprobs: dict[str, int] = {}
473
        self.comm_stream = torch.cuda.Stream()
474

475
476
477
478
479
480
481
482
483
        # 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.
484
485
486
487
        logits_processors = model_config.logits_processors
        custom_logitsprocs: Sequence[str | type[LogitsProcessor]] = (
            tuple(logits_processors) if logits_processors is not None else ()
        )
488
489
        self.input_batch = InputBatch(
            max_num_reqs=self.max_num_reqs,
490
491
492
            # 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),
493
494
495
            max_num_batched_tokens=self.max_num_tokens,
            device=self.device,
            pin_memory=self.pin_memory,
496
            vocab_size=self.model_config.get_vocab_size(),
497
            block_sizes=[self.cache_config.block_size],
498
            kernel_block_sizes=[self.cache_config.block_size],
499
            is_spec_decode=bool(self.vllm_config.speculative_config),
500
            logitsprocs=build_logitsprocs(
501
502
503
                self.vllm_config,
                self.device,
                self.pin_memory,
504
                self.is_pooling_model,
505
                custom_logitsprocs,
506
            ),
507
508
509
            # 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),
510
            is_pooling_model=self.is_pooling_model,
511
            cp_kv_cache_interleave_size=self.parallel_config.cp_kv_cache_interleave_size,
512
        )
513

514
515
516
517
518
        # 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.
519
        self.prepare_inputs_event: torch.Event | None = None
520
521
        if self.use_async_scheduling:
            self.async_output_copy_stream = torch.cuda.Stream()
522
            self.prepare_inputs_event = torch.Event()
523

524
        # self.cudagraph_batch_sizes sorts in ascending order.
525
526
527
528
        if (
            self.compilation_config.cudagraph_capture_sizes
            and self.compilation_config.cudagraph_mode != CUDAGraphMode.NONE
        ):
529
530
            self.cudagraph_batch_sizes = sorted(
                self.compilation_config.cudagraph_capture_sizes
531
            )
532

533
        # Cache the device properties.
534
        self._init_device_properties()
535

536
        # Persistent buffers for CUDA graphs.
537
538
539
540
541
        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
        )
542
        self.seq_lens = self._make_buffer(self.max_num_reqs, dtype=torch.int32)
543
        self.encoder_seq_lens = self._make_buffer(self.max_num_reqs, dtype=torch.int32)
544
545
546
547
        if self.dcp_world_size > 1:
            self.dcp_local_seq_lens = self._make_buffer(
                self.max_num_reqs, dtype=torch.int32
            )
548
549
550
        # 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.
551
        self.inputs_embeds = self._make_buffer(
552
            self.max_num_tokens, self.inputs_embeds_size, dtype=self.dtype, numpy=False
553
554
        )
        self.is_token_ids = self._make_buffer(self.max_num_tokens, dtype=torch.bool)
555
556
        self.discard_request_mask = self._make_buffer(
            self.max_num_reqs, dtype=torch.bool
557
558
559
560
561
562
563
        )
        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
        )
564

565
566
        # Only relevant for multimodal models
        if self.supports_mm_inputs:
567
568
569
570
571
572
573
            # 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
574

575
        # Only relevant for models using M-RoPE (e.g, Qwen2-VL)
576
        if self.uses_mrope:
Roger Wang's avatar
Roger Wang committed
577
578
579
580
            # 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
581
582
583
584
585
586

            # 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
587
            self.mrope_positions = self._make_buffer(
588
589
                (3, self.max_num_tokens + 1), dtype=torch.int64
            )
590

591
592
593
594
595
596
597
        # 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
            )

598
        # None in the first PP rank. The rest are set after load_model.
599
        self.intermediate_tensors: IntermediateTensors | None = None
600

601
        # OPTIMIZATION: Cache the tensors rather than creating them every step.
602
        # Keep in int64 to avoid overflow with long context
603
604
605
606
        self.arange_np = np.arange(
            max(self.max_num_reqs + 1, self.max_model_len, self.max_num_tokens),
            dtype=np.int64,
        )
607

608
609
610
611
612
        # 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] = {}
613
614
615
616
617
        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(
618
619
                self.max_num_tokens, dtype=torch.int32, device=self.device
            )
620

621
        self.uniform_decode_query_len = 1 + self.num_spec_tokens
622
623
624
625

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

626
627
        self.mm_budget = (
            MultiModalBudget(
628
                self.model_config,
629
630
631
632
633
634
                self.scheduler_config,
                self.mm_registry,
            )
            if self.supports_mm_inputs
            else None
        )
635

636
        self.reorder_batch_threshold: int | None = None
637

638
639
640
641
642
        # 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()

643
        # Cached outputs.
644
        self._draft_token_ids: list[list[int]] | torch.Tensor | None = None
645
        self._draft_token_req_ids: list[str] | None = None
646
        self.transfer_event = torch.Event()
647
        self.sampled_token_ids_pinned_cpu = torch.empty(
648
            (self.max_num_reqs, 1),
649
650
            dtype=torch.int64,
            device="cpu",
651
652
            pin_memory=self.pin_memory,
        )
653

654
655
        # Pre-allocated tensor for copying valid sampled token counts to CPU,
        # with dedicated stream for overlapping and event for coordination.
656
        self.valid_sampled_token_count_event: torch.Event | None = None
657
        self.valid_sampled_token_count_copy_stream: torch.cuda.Stream | None = None
658
659
660
661
662
663
664
665
666
667
668
669
670
671
672
673
674
675
676
677
678
679
680
681
        # 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,
                )
682

683
684
        # Ephemeral state transferred between execute_model() and sample_tokens().
        self.execute_model_state: ExecuteModelState | None = None
685
        self.kv_connector_output: KVConnectorOutput | None = None
686
        self.layerwise_nvtx_hooks_registered = False
687

688
689
690
691
692
693
694
    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

695
696
697
698
    def reset_mm_cache(self) -> None:
        if self.mm_budget:
            self.mm_budget.reset_cache()

699
700
701
702
703
704
705
706
707
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
    @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)

743
744
745
746
    def _get_positions(self, num_tokens: Any):
        if isinstance(num_tokens, int):
            if self.uses_mrope:
                return self.mrope_positions.gpu[:, :num_tokens]
747
748
            if self.uses_xdrope_dim > 0:
                return self.xdrope_positions.gpu[:, :num_tokens]
749
750
751
752
            return self.positions.gpu[:num_tokens]
        else:
            if self.uses_mrope:
                return self.mrope_positions.gpu[:, num_tokens]
753
754
            if self.uses_xdrope_dim > 0:
                return self.xdrope_positions.gpu[:, num_tokens]
755
756
            return self.positions.gpu[num_tokens]

757
    def _make_buffer(
758
        self, *size: int | torch.SymInt, dtype: torch.dtype, numpy: bool = True
759
760
761
762
763
764
765
766
    ) -> CpuGpuBuffer:
        return CpuGpuBuffer(
            *size,
            dtype=dtype,
            device=self.device,
            pin_memory=self.pin_memory,
            with_numpy=numpy,
        )
767

768
    def _init_model_kwargs(self):
769
770
        model_kwargs = dict[str, Any]()

771
        if not self.is_pooling_model:
772
773
            return model_kwargs

774
775
        num_reqs = self.input_batch.num_reqs
        pooling_params = self.input_batch.get_pooling_params()
776
777
778

        token_type_id_requests = dict[int, Any]()
        for i, param in enumerate(pooling_params):
779
780
781
782
783
            if (
                param.extra_kwargs is not None
                and (token_types := param.extra_kwargs.get("compressed_token_type_ids"))
                is not None
            ):
784
785
786
787
788
                token_type_id_requests[i] = token_types

        if len(token_type_id_requests) == 0:
            return model_kwargs

789
        seq_lens = self.seq_lens.gpu[:num_reqs]
790
791
792
793
794
795
796
797
        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(
798
799
            device=self.device
        )
800
801
        return model_kwargs

802
    def _may_reorder_batch(self, scheduler_output: "SchedulerOutput") -> None:
803
804
        """
        Update the order of requests in the batch based on the attention
805
        backend's needs. For example, some attention backends (namely MLA) may
806
807
808
809
810
811
        want to separate requests based on if the attention computation will be
        compute-bound or memory-bound.

        Args:
            scheduler_output: The scheduler output.
        """
812
813
814
815
816
817
818
819
        # 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

820
821
822
823
        if self.reorder_batch_threshold is not None:
            reorder_batch_to_split_decodes_and_prefills(
                self.input_batch,
                scheduler_output,
824
825
                decode_threshold=self.reorder_batch_threshold,
            )
826

827
828
    # Note: used for model runner override.
    def _init_device_properties(self) -> None:
829
        """Initialize attributes from torch.cuda.get_device_properties"""
830
831
832
833
834
835
836
        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()

837
    def _update_states(self, scheduler_output: "SchedulerOutput") -> None:
838
839
840
841
842
843
        """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.

844
845
        The SamplingMetadata is updated and copied to the GPU if there is a
        new/resumed/paused/finished request in the batch.
846
847
        """
        # Remove finished requests from the cached states.
848
849
        for req_id in scheduler_output.finished_req_ids:
            self.requests.pop(req_id, None)
850
            self.num_prompt_logprobs.pop(req_id, None)
851
852
853
854
855
856
857
        # 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:
858
            self.input_batch.remove_request(req_id)
859
860

        # Free the cached encoder outputs.
861
862
        for mm_hash in scheduler_output.free_encoder_mm_hashes:
            self.encoder_cache.pop(mm_hash, None)
863

864
865
866
867
868
869
870
        # 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()
871
872
873
874
875
876
877
878
        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)
879
880
881
882
883
        # 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:
884
            self.input_batch.remove_request(req_id)
885

886
        reqs_to_add: list[CachedRequestState] = []
887
        # Add new requests to the cached states.
888
889
890
        for new_req_data in scheduler_output.scheduled_new_reqs:
            req_id = new_req_data.req_id
            sampling_params = new_req_data.sampling_params
891
            pooling_params = new_req_data.pooling_params
892

893
894
895
896
            if (
                sampling_params
                and sampling_params.sampling_type == SamplingType.RANDOM_SEED
            ):
897
898
899
900
901
                generator = torch.Generator(device=self.device)
                generator.manual_seed(sampling_params.seed)
            else:
                generator = None

902
903
            if self.is_pooling_model:
                assert pooling_params is not None
904
905
                task = pooling_params.task
                assert task is not None, "You did not set `task` in the API"
906

907
                model = cast(VllmModelForPooling, self.get_model())
908
                to_update = model.pooler.get_pooling_updates(task)
909
910
                to_update.apply(pooling_params)

911
            req_state = CachedRequestState(
912
                req_id=req_id,
913
                prompt_token_ids=new_req_data.prompt_token_ids,
914
                prompt_embeds=new_req_data.prompt_embeds,
915
                mm_features=new_req_data.mm_features,
916
                sampling_params=sampling_params,
917
                pooling_params=pooling_params,
918
                generator=generator,
919
920
                block_ids=new_req_data.block_ids,
                num_computed_tokens=new_req_data.num_computed_tokens,
921
                output_token_ids=[],
922
                lora_request=new_req_data.lora_request,
923
            )
924
925
            self.requests[req_id] = req_state

926
927
928
929
930
931
932
            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
                )

933
            # Only relevant for models using M-RoPE (e.g, Qwen2-VL)
934
            if self.uses_mrope:
935
                self._init_mrope_positions(req_state)
936

937
938
939
940
            # Only relevant for models using XD-RoPE (e.g, HunYuan-VL)
            if self.uses_xdrope_dim > 0:
                self._init_xdrope_positions(req_state)

941
            reqs_to_add.append(req_state)
942

943
        # Update the states of the running/resumed requests.
944
        is_last_rank = get_pp_group().is_last_rank
945
        req_data = scheduler_output.scheduled_cached_reqs
946
        scheduled_spec_tokens = scheduler_output.scheduled_spec_decode_tokens
947
948
949
950
951

        # 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()

952
        for i, req_id in enumerate(req_data.req_ids):
953
            req_state = self.requests[req_id]
954
955
            num_computed_tokens = req_data.num_computed_tokens[i]
            new_block_ids = req_data.new_block_ids[i]
956
            resumed_from_preemption = req_id in req_data.resumed_req_ids
957
            num_output_tokens = req_data.num_output_tokens[i]
958
            req_index = self.input_batch.req_id_to_index.get(req_id)
959

960
961
962
963
964
965
966
967
968
969
970
971
972
973
            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.
974
975
976
977
978
979
980
981
982
                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)
983

984
            # Update the cached states.
985
            req_state.num_computed_tokens = num_computed_tokens
986
987
988
989
990
991
992
993

            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.
994
995
996
                num_new_tokens = (
                    num_computed_tokens + len(new_token_ids) - req_state.num_tokens
                )
997
998
999
1000
                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:
1001
                    req_state.output_token_ids.extend(new_token_ids[-num_new_tokens:])
1002
1003
1004
1005
1006
            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:
1007
1008
1009
1010
                    end_idx = (
                        self.input_batch.num_prompt_tokens[req_index]
                        + num_output_tokens
                    )
1011
                    self.input_batch.num_tokens_no_spec[req_index] = end_idx
1012

1013
            # Update the block IDs.
1014
            if not resumed_from_preemption:
1015
1016
                if new_block_ids is not None:
                    # Append the new blocks to the existing block IDs.
1017
                    for block_ids, new_ids in zip(req_state.block_ids, new_block_ids):
1018
                        block_ids.extend(new_ids)
1019
            else:
1020
                assert req_index is None
1021
                assert new_block_ids is not None
1022
1023
                # The request is resumed from preemption.
                # Replace the existing block IDs with the new ones.
1024
                req_state.block_ids = new_block_ids
1025
1026
1027
1028
1029

            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.
1030
1031
1032
1033
1034
1035
1036

                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:]

1037
                reqs_to_add.append(req_state)
1038
1039
1040
                continue

            # Update the persistent batch.
1041
            self.input_batch.num_computed_tokens_cpu[req_index] = num_computed_tokens
1042
            if new_block_ids is not None:
1043
                self.input_batch.block_table.append_row(new_block_ids, req_index)
1044
1045
1046
1047
1048
1049
1050

            # 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)
1051
                self.input_batch.token_ids_cpu[
1052
1053
1054
                    req_index, start_token_index:end_token_index
                ] = new_token_ids
                self.input_batch.num_tokens_no_spec[req_index] = end_token_index
1055

1056
            # Add spec_token_ids to token_ids_cpu.
1057
            self.input_batch.update_req_spec_token_ids(req_state, scheduled_spec_tokens)
1058

1059
1060
        # Add the new or resumed requests to the persistent batch.
        # The smaller empty indices are filled first.
1061
1062
        for request in reqs_to_add:
            self.input_batch.add_request(request)
1063
            self.input_batch.update_req_spec_token_ids(request, scheduled_spec_tokens)
1064

1065
1066
1067
1068
1069
1070
        # 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()
1071

1072
    def _update_states_after_model_execute(
1073
1074
        self, output_token_ids: torch.Tensor
    ) -> None:
1075
1076
1077
1078
1079
1080
1081
1082
        """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.
        """
1083
        if not self.speculative_config or not self.model_config.is_hybrid:
1084
1085
1086
            return

        # Find the number of accepted tokens for each sequence.
1087
1088
1089
1090
1091
1092
1093
1094
1095
1096
1097
1098
1099
1100
1101
1102
1103
1104
1105
1106
        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()
        )
1107
1108
1109
        for i, num_tokens in enumerate(num_accepted_tokens):
            self.input_batch.num_accepted_tokens_cpu[i] = num_tokens

1110
    def _init_mrope_positions(self, req_state: CachedRequestState):
1111
1112
        model = self.get_model()
        assert supports_mrope(model), "M-RoPE support is not implemented."
1113
1114
1115
1116
        assert req_state.prompt_token_ids is not None, (
            "M-RoPE requires prompt_token_ids to be available."
        )
        mrope_model = cast(SupportsMRoPE, model)
1117
1118

        req_state.mrope_positions, req_state.mrope_position_delta = (
1119
            mrope_model.get_mrope_input_positions(
1120
                req_state.prompt_token_ids,
1121
                req_state.mm_features,
1122
            )
1123
        )
1124

1125
1126
1127
1128
1129
1130
1131
1132
1133
1134
1135
1136
1137
    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,
        )

1138
    def _extract_mm_kwargs(
1139
        self,
1140
1141
        scheduler_output: "SchedulerOutput",
    ) -> BatchedTensorInputs:
1142
        if not scheduler_output or not self.is_multimodal_raw_input_only_model:
1143
            return {}
1144

1145
1146
        mm_kwargs = list[MultiModalKwargsItem]()
        for req in scheduler_output.scheduled_new_reqs:
1147
1148
1149
            for feature in req.mm_features:
                if feature.data is not None:
                    mm_kwargs.append(feature.data)
1150

1151
1152
1153
        # Input all modalities at once
        mm_kwargs_combined: BatchedTensorInputs = {}
        for _, _, mm_kwargs_group in group_mm_kwargs_by_modality(
1154
1155
1156
            mm_kwargs,
            device=self.device,
            pin_memory=self.pin_memory,
1157
1158
        ):
            mm_kwargs_combined.update(mm_kwargs_group)
1159

1160
        return mm_kwargs_combined
1161

1162
    def _dummy_mm_kwargs(self, num_seqs: int) -> BatchedTensorInputs:
1163
        if not self.is_multimodal_raw_input_only_model:
1164
            return {}
1165

1166
1167
1168
1169
1170
        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)
1171

1172
1173
1174
    def _get_cumsum_and_arange(
        self,
        num_tokens: np.ndarray,
1175
        cumsum_dtype: np.dtype | None = None,
1176
1177
1178
1179
1180
1181
1182
1183
1184
1185
1186
1187
1188
1189
1190
1191
    ) -> 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

1192
    def _prepare_input_ids(
1193
1194
1195
1196
        self,
        scheduler_output: "SchedulerOutput",
        total_num_scheduled_tokens: int,
        cu_num_tokens: np.ndarray,
1197
    ) -> None:
1198
        """Prepare the input IDs for the current batch.
1199

1200
1201
1202
1203
1204
1205
1206
        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)
1207
1208
1209
            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)
1210
1211
1212
1213
1214
1215
1216
            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
1217
1218
1219
1220
        sample_flattened_indices: list[int] = []
        spec_flattened_indices: list[int] = []
        prev_common_req_indices: list[int] = []
        prev_draft_token_indices: list[int] = []
1221
1222
        indices_match = True
        max_flattened_index = -1
1223
1224
1225
        total_num_spec_tokens = 0
        scheduled_spec_tokens = scheduler_output.scheduled_spec_decode_tokens

1226
1227
1228
1229
1230
        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.
1231
1232
                draft_len = len(scheduled_spec_tokens.get(req_id, ()))
                total_num_spec_tokens += draft_len
1233
                flattened_index = cu_num_tokens[cur_index].item() - 1
1234
1235
1236
1237
1238
1239
1240
1241
1242
1243
1244
1245
1246
1247
1248
                # 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))
1249
                indices_match &= prev_index == flattened_index
1250
                max_flattened_index = max(max_flattened_index, flattened_index)
1251
1252
1253
        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:
1254
1255
1256
            # 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)
1257
1258
1259
            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)
1260
1261
        if num_commmon_tokens == 0:
            # No requests in common with the previous iteration
1262
            # So input_ids.cpu will have all the input ids.
1263
1264
1265
1266
1267
1268
1269
            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_(
1270
1271
1272
                self.input_batch.prev_sampled_token_ids[:num_commmon_tokens, 0],
                non_blocking=True,
            )
1273
1274
            if self.enable_prompt_embeds:
                self.is_token_ids.gpu[:num_commmon_tokens] = True
1275
            return
1276
        # Upload the index tensors asynchronously so the scatter can be non-blocking.
1277
1278
        sampled_tokens_index_tensor = torch.tensor(
            sample_flattened_indices, dtype=torch.int64, pin_memory=self.pin_memory
1279
        ).to(self.device, non_blocking=True)
1280
        prev_common_req_indices_tensor = torch.tensor(
1281
1282
            prev_common_req_indices, dtype=torch.int64, pin_memory=self.pin_memory
        ).to(self.device, non_blocking=True)
1283
1284
        self.input_ids.gpu.scatter_(
            dim=0,
1285
            index=sampled_tokens_index_tensor,
1286
            src=self.input_batch.prev_sampled_token_ids[
1287
1288
1289
                prev_common_req_indices_tensor, 0
            ],
        )
1290

1291
1292
1293
1294
1295
1296
1297
1298
1299
1300
1301
1302
1303
1304
1305
1306
1307
1308
1309
1310
1311
1312
        # 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],
        )

1313
1314
    def _get_encoder_seq_lens(
        self,
1315
        num_scheduled_tokens: dict[str, int],
1316
1317
        kv_cache_spec: KVCacheSpec,
        num_reqs: int,
1318
    ) -> tuple[torch.Tensor | None, np.ndarray | None]:
1319
        if not isinstance(kv_cache_spec, CrossAttentionSpec):
1320
            return None, None
1321

1322
1323
        # Zero out buffer for padding requests that are not actually scheduled (CGs)
        self.encoder_seq_lens.np[:num_reqs] = 0
1324
1325
        # Build encoder_seq_lens array mapping request indices to
        # encoder lengths for inputs scheduled in this batch
1326
        for req_id in num_scheduled_tokens:
1327
            req_index = self.input_batch.req_id_to_index[req_id]
1328
1329
1330
1331
1332
1333
1334
1335
1336
1337
1338
1339
1340
1341
1342
1343
            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]
1344

1345
        return encoder_seq_lens, encoder_seq_lens_cpu
1346

1347
    def _prepare_inputs(
1348
1349
1350
        self,
        scheduler_output: "SchedulerOutput",
        num_scheduled_tokens: np.ndarray,
1351
1352
    ) -> tuple[
        torch.Tensor,
1353
        SpecDecodeMetadata | None,
1354
    ]:
1355
1356
        """
        :return: tuple[
1357
            logits_indices, spec_decode_metadata,
1358
1359
        ]
        """
1360
1361
1362
1363
1364
1365
1366
        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.
1367
        self.input_batch.block_table.commit_block_table(num_reqs)
1368
1369
1370

        # Get request indices.
        # E.g., [2, 5, 3] -> [0, 0, 1, 1, 1, 1, 1, 2, 2, 2]
1371
        req_indices = np.repeat(self.arange_np[:num_reqs], num_scheduled_tokens)
1372

1373
1374
        # cu_num_tokens: [2, 5, 3] -> [2, 7, 10]
        # arange: [0, 1, 0, 1, 2, 3, 4, 0, 1, 2]
1375
        cu_num_tokens, arange = self._get_cumsum_and_arange(num_scheduled_tokens)
1376
1377

        # Get positions.
1378
        positions_np = self.positions.np[:total_num_scheduled_tokens]
1379
1380
1381
1382
1383
        np.add(
            self.input_batch.num_computed_tokens_cpu[req_indices],
            arange,
            out=positions_np,
        )
1384

1385
1386
        # Calculate M-RoPE positions.
        # Only relevant for models using M-RoPE (e.g, Qwen2-VL)
1387
        if self.uses_mrope:
1388
1389
            self._calc_mrope_positions(scheduler_output)

1390
1391
1392
1393
1394
        # 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)

1395
1396
1397
1398
        # 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.
1399
1400
1401
        token_indices = (
            positions_np + req_indices * self.input_batch.token_ids_cpu.shape[1]
        )
1402
        token_indices_tensor = torch.from_numpy(token_indices)
1403

1404
1405
1406
        # 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.
1407
1408
1409
1410
1411
1412
        torch.index_select(
            self.input_batch.token_ids_cpu_tensor.flatten(),
            0,
            token_indices_tensor,
            out=self.input_ids.cpu[:total_num_scheduled_tokens],
        )
1413
        if self.enable_prompt_embeds:
1414
            is_token_ids = self.input_batch.is_token_ids_tensor.flatten()
1415
1416
1417
1418
            torch.index_select(
                is_token_ids,
                0,
                token_indices_tensor,
1419
1420
                out=self.is_token_ids.cpu[:total_num_scheduled_tokens],
            )
1421
1422
1423
1424
1425
1426
1427
1428
1429
1430
1431
1432
1433
1434
1435
1436
1437
1438
1439
1440
1441
1442
1443
1444
1445
1446
1447
1448
1449
1450
1451
1452
1453

        # 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:
1454
1455
1456
                    self.inputs_embeds.cpu[
                        output_idx : output_idx + actual_num_sched
                    ].copy_(req_embeds[start_pos:actual_end])
1457
1458

                output_idx += num_sched
1459

1460
1461
        self.input_batch.block_table.compute_slot_mapping(req_indices, positions_np)
        self.input_batch.block_table.commit_slot_mapping(total_num_scheduled_tokens)
1462
1463

        # Prepare the attention metadata.
1464
        self.query_start_loc.np[0] = 0
1465
        self.query_start_loc.np[1 : num_reqs + 1] = cu_num_tokens
1466
1467
        # Note: pad query_start_loc to be non-decreasing, as kernels
        # like FlashAttention requires that
1468
        self.query_start_loc.np[num_reqs + 1 :].fill(cu_num_tokens[-1])
1469
        self.query_start_loc.copy_to_gpu()
1470
        query_start_loc = self.query_start_loc.gpu[: num_reqs + 1]
1471

1472
        self.seq_lens.np[:num_reqs] = (
1473
1474
            self.input_batch.num_computed_tokens_cpu[:num_reqs] + num_scheduled_tokens
        )
1475
        # Fill unused with 0 for full cuda graph mode.
1476
1477
        self.seq_lens.np[num_reqs:].fill(0)
        self.seq_lens.copy_to_gpu()
1478

1479
        num_tokens = [self.requests[r].num_tokens for r in self.input_batch.req_ids]
1480
1481
        num_tokens_np = np.array(num_tokens, dtype=np.int32)

1482
        # Record which requests should not be sampled,
1483
        # so that we could clear the sampled tokens before returning
1484
1485
        self.discard_request_mask.np[:num_reqs] = (
            self.seq_lens.np[:num_reqs] < num_tokens_np
1486
        )
1487
        self.discard_request_mask.copy_to_gpu(num_reqs)
1488

1489
        # Copy the tensors to the GPU.
1490
1491
1492
1493
1494
        self._prepare_input_ids(
            scheduler_output,
            total_num_scheduled_tokens,
            cu_num_tokens,
        )
1495

1496
        if self.uses_mrope:
1497
            # Only relevant for models using M-RoPE (e.g, Qwen2-VL)
1498
1499
            self.mrope_positions.gpu[:, :total_num_scheduled_tokens].copy_(
                self.mrope_positions.cpu[:, :total_num_scheduled_tokens],
1500
1501
                non_blocking=True,
            )
1502
1503
1504
1505
1506
1507
        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,
            )
1508
1509
        else:
            # Common case (1D positions)
1510
            self.positions.copy_to_gpu(total_num_scheduled_tokens)
1511

1512
        use_spec_decode = len(scheduler_output.scheduled_spec_decode_tokens) > 0
1513
1514
1515
1516
1517
1518
1519
1520
        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
1521
            num_sampled_tokens = np.ones(num_reqs, dtype=np.int32)
1522
1523
1524
1525
1526
        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)
1527
1528
1529
            # 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)
1530
1531
1532
1533
            for (
                req_id,
                draft_token_ids,
            ) in scheduler_output.scheduled_spec_decode_tokens.items():
1534
1535
                req_idx = self.input_batch.req_id_to_index[req_id]
                num_draft_tokens[req_idx] = len(draft_token_ids)
1536
1537
1538
1539
1540
                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)
1541
            spec_decode_metadata = self._calc_spec_decode_metadata(
1542
1543
                num_draft_tokens, cu_num_tokens
            )
1544
            logits_indices = spec_decode_metadata.logits_indices
1545
            num_sampled_tokens = num_draft_tokens + 1
1546
            # For DECODE only cuda graph of some attention backends (e.g., GDN).
1547
            self.num_decode_draft_tokens.np[:num_reqs] = num_decode_draft_tokens
1548
1549
            self.num_decode_draft_tokens.np[num_reqs:].fill(-1)
            self.num_decode_draft_tokens.copy_to_gpu()
1550

1551
1552
1553
1554
1555
        # Hot-Swap lora model
        if self.lora_config:
            assert (
                np.sum(num_sampled_tokens)
                <= self.vllm_config.scheduler_config.max_num_batched_tokens
1556
            )
1557
1558
1559
1560
1561
1562
1563
1564
1565
1566
1567
            self.set_active_loras(
                self.input_batch, num_scheduled_tokens, num_sampled_tokens
            )

        return (
            logits_indices,
            spec_decode_metadata,
        )

    def _build_attention_metadata(
        self,
1568
        num_tokens: int,
1569
        num_reqs: int,
1570
1571
1572
        max_query_len: int,
        num_tokens_padded: int | None = None,
        num_reqs_padded: int | None = None,
1573
1574
1575
1576
        ubatch_slices: UBatchSlices | None = None,
        logits_indices: torch.Tensor | None = None,
        use_spec_decode: bool = False,
        for_cudagraph_capture: bool = False,
1577
        num_scheduled_tokens: dict[str, int] | None = None,
1578
1579
1580
1581
1582
        cascade_attn_prefix_lens: list[list[int]] | None = None,
    ) -> tuple[PerLayerAttnMetadata, CommonAttentionMetadata | None]:
        """
        :return: tuple[attn_metadata, spec_decode_common_attn_metadata]
        """
1583
1584
1585
1586
        # Attention metadata is not needed for attention free models
        if len(self.kv_cache_config.kv_cache_groups) == 0:
            return {}, None

1587
1588
        num_tokens_padded = num_tokens_padded or num_tokens
        num_reqs_padded = num_reqs_padded or num_reqs
1589
        assert num_reqs_padded is not None and num_tokens_padded is not None
1590

1591
1592
1593
        attn_metadata: PerLayerAttnMetadata = {}
        if ubatch_slices is not None:
            attn_metadata = [dict() for _ in range(len(ubatch_slices))]
1594

1595
1596
1597
1598
1599
1600
1601
1602
        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()

1603
1604
        if use_spec_decode:
            self.num_accepted_tokens.np[:num_reqs] = (
1605
1606
                self.input_batch.num_accepted_tokens_cpu[:num_reqs]
            )
1607
1608
            self.num_accepted_tokens.np[num_reqs:].fill(1)
            self.num_accepted_tokens.copy_to_gpu()
1609

1610
        kv_cache_groups = self.kv_cache_config.kv_cache_groups
1611

1612
1613
1614
1615
        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):
1616
                blk_table_tensor = torch.zeros(
1617
                    (num_reqs_padded, 1),
1618
                    dtype=torch.int32,
1619
1620
1621
                    device=self.device,
                )
                slot_mapping = torch.zeros(
1622
                    (num_tokens_padded,),
1623
1624
1625
                    dtype=torch.int64,
                    device=self.device,
                )
1626
            else:
1627
                blk_table = self.input_batch.block_table[kv_cache_gid]
1628
1629
                blk_table_tensor = blk_table.get_device_tensor(num_reqs_padded)
                slot_mapping = blk_table.slot_mapping.gpu[:num_tokens_padded]
1630

1631
1632
1633
1634
1635
1636
1637
1638
            # 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)
1639
1640
        if self.model_config.enable_return_routed_experts:
            self.slot_mapping = slot_mapping_gid_0[:num_tokens].cpu().numpy()
1641
1642
1643
1644
1645
1646
1647
1648
1649
1650
1651
1652
1653
1654
1655
1656
1657
1658
1659
1660
1661
1662
1663
1664
1665
1666
1667
1668
1669
1670
1671
1672
1673
1674
1675
1676
1677
1678
        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
            )

1679
1680
1681
1682
1683
1684
1685
1686
1687
        # 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
        ] = {}

1688
1689
1690
1691
1692
1693
1694
        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]
1695
            builder = attn_group.get_metadata_builder(ubid or 0)
1696
1697
1698
1699
            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))
1700

1701
1702
1703
1704
1705
1706
1707
1708
1709
1710
1711
1712
1713
1714
1715
1716
1717
1718
1719
1720
            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
                )
1721
1722
1723
1724
1725
1726
1727
1728
1729
            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,
                )
1730
1731
1732
1733
1734
1735
            else:
                attn_metadata_i = builder.build(
                    common_prefix_len=cascade_attn_prefix_len,
                    common_attn_metadata=common_attn_metadata,
                    **extra_attn_metadata_args,
                )
1736
1737
                if builder.supports_update_block_table:
                    cached_attn_metadata[cache_key] = attn_metadata_i
1738
1739
1740
1741
1742
1743
1744
1745
1746
1747
1748
1749
1750
1751
1752
1753
1754
1755
1756
1757
1758
1759
1760

            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,
1761
            )
1762
1763
1764
1765
            if kv_cache_gid > 0:
                cm.block_table_tensor, cm.slot_mapping = (
                    _get_block_table_and_slot_mapping(kv_cache_gid)
                )
1766

1767
            if self.speculative_config and spec_decode_common_attn_metadata is None:
1768
                if isinstance(self.drafter, EagleProposer):
1769
                    if self.drafter.attn_layer_names[0] in kv_cache_group.layer_names:
1770
                        spec_decode_common_attn_metadata = cm
1771
                else:
1772
                    spec_decode_common_attn_metadata = cm
1773

1774
            for attn_gid in range(len(self.attn_groups[kv_cache_gid])):
1775
                if ubatch_slices is not None:
1776
1777
1778
                    for ubid, _cm in enumerate(split_attn_metadata(ubatch_slices, cm)):
                        _build_attn_group_metadata(kv_cache_gid, attn_gid, _cm, ubid)

1779
                else:
1780
                    _build_attn_group_metadata(kv_cache_gid, attn_gid, cm)
1781

1782
1783
1784
1785
1786
1787
1788
1789
1790
1791
1792
1793
1794
1795
1796
1797
1798
1799
1800
1801
        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]

1802
1803
1804
1805
1806
1807
1808
1809
1810
1811
        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)
            )

1812
        return attn_metadata, spec_decode_common_attn_metadata
1813

1814
1815
1816
    def _compute_cascade_attn_prefix_lens(
        self,
        num_scheduled_tokens: np.ndarray,
1817
        num_computed_tokens: np.ndarray,
1818
1819
1820
1821
1822
1823
1824
        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
        """
1825

1826
1827
1828
1829
1830
1831
1832
1833
1834
1835
1836
1837
1838
1839
        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,
1840
                        num_computed_tokens,
1841
1842
1843
1844
1845
1846
1847
1848
                        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
1849

1850
1851
1852
    def _compute_cascade_attn_prefix_len(
        self,
        num_scheduled_tokens: np.ndarray,
1853
        num_computed_tokens: np.ndarray,
1854
        num_common_prefix_blocks: int,
1855
1856
        kv_cache_spec: KVCacheSpec,
        attn_metadata_builder: AttentionMetadataBuilder,
1857
1858
1859
1860
1861
1862
1863
1864
1865
1866
1867
1868
1869
1870
1871
1872
1873
1874
    ) -> 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.
        """
1875

1876
        common_prefix_len = num_common_prefix_blocks * kv_cache_spec.block_size
1877
1878
1879
1880
1881
1882
1883
1884
1885
1886
1887
1888
1889
1890
1891
1892
1893
1894
1895
1896
1897
1898
1899
1900
1901
1902
1903
1904
1905
1906
1907
1908
1909
1910
1911
1912
1913
        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]
1914
        # Request 3's num_computed_tokens: 3 (i.e., [A, B, C])
1915
1916
1917
1918
1919
        # 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.
1920
        common_prefix_len = min(common_prefix_len, num_computed_tokens.min())
1921
        # common_prefix_len should be a multiple of the block size.
1922
1923
1924
1925
1926
1927
1928
1929
1930
1931
1932
        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
        )
1933
1934
        assert isinstance(kv_cache_spec, AttentionSpec)
        use_cascade = attn_metadata_builder.use_cascade_attention(
1935
1936
1937
            common_prefix_len=common_prefix_len,
            query_lens=num_scheduled_tokens,
            num_query_heads=self.num_query_heads,
1938
            num_kv_heads=kv_cache_spec.num_kv_heads,
1939
            use_alibi=self.use_alibi,
1940
            use_sliding_window=use_sliding_window,
1941
            use_local_attention=use_local_attention,
1942
            num_sms=self.num_sms,
1943
            dcp_world_size=self.dcp_world_size,
1944
1945
1946
        )
        return common_prefix_len if use_cascade else 0

1947
1948
    def _calc_mrope_positions(self, scheduler_output: "SchedulerOutput"):
        mrope_pos_ptr = 0
1949
        for index, req_id in enumerate(self.input_batch.req_ids):
1950
1951
1952
            req = self.requests[req_id]
            assert req.mrope_positions is not None

1953
1954
            num_computed_tokens = self.input_batch.num_computed_tokens_cpu[index]
            num_scheduled_tokens = scheduler_output.num_scheduled_tokens[req_id]
1955
            num_prompt_tokens = length_from_prompt_token_ids_or_embeds(
1956
1957
                req.prompt_token_ids, req.prompt_embeds
            )
1958
1959

            if num_computed_tokens + num_scheduled_tokens > num_prompt_tokens:
1960
1961
                prompt_part_len = max(0, num_prompt_tokens - num_computed_tokens)
                completion_part_len = max(0, num_scheduled_tokens - prompt_part_len)
1962
1963
1964
1965
1966
1967
1968
1969
1970
1971
1972
1973
1974
            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

1975
1976
1977
                self.mrope_positions.cpu[:, dst_start:dst_end] = req.mrope_positions[
                    :, src_start:src_end
                ]
1978
1979
1980
1981
1982
1983
1984
                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

1985
                assert req.mrope_position_delta is not None
1986
                MRotaryEmbedding.get_next_input_positions_tensor(
1987
                    out=self.mrope_positions.np,
1988
1989
1990
1991
1992
                    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,
                )
1993
1994
1995

                mrope_pos_ptr += completion_part_len

1996
1997
1998
1999
2000
2001
2002
2003
2004
2005
2006
2007
2008
2009
2010
2011
2012
2013
2014
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
    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

2043
2044
    def _calc_spec_decode_metadata(
        self,
2045
2046
2047
2048
2049
2050
2051
2052
2053
2054
2055
2056
2057
2058
2059
2060
        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
2061
2062
2063
2064

        # 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(
2065
2066
            num_sampled_tokens, cumsum_dtype=np.int32
        )
2067
        # Step 2. [0, 0, 0, 0, 103, 104, 104, 104, 206, 207, 207]
2068
        logits_indices = np.repeat(
2069
2070
            cu_num_scheduled_tokens - num_sampled_tokens, num_sampled_tokens
        )
2071
        # Step 3. [0, 1, 2, 3, 103, 104, 105, 106, 206, 207, 208]
2072
2073
2074
2075
2076
2077
        logits_indices += arange

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

        # Compute the draft logits indices.
2078
2079
2080
        # 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(
2081
2082
            num_draft_tokens, cumsum_dtype=np.int32
        )
2083
2084
        # [0, 0, 0, 5, 5, 9]
        target_logits_indices = np.repeat(
2085
2086
            cu_num_sampled_tokens - num_sampled_tokens, num_draft_tokens
        )
2087
2088
2089
2090
2091
        # [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(
2092
2093
            self.device, non_blocking=True
        )
2094
2095
2096
        cu_num_sampled_tokens = torch.from_numpy(cu_num_sampled_tokens).to(
            self.device, non_blocking=True
        )
2097
2098
2099
        logits_indices = torch.from_numpy(logits_indices).to(
            self.device, non_blocking=True
        )
2100
        target_logits_indices = torch.from_numpy(target_logits_indices).to(
2101
2102
            self.device, non_blocking=True
        )
2103
        bonus_logits_indices = torch.from_numpy(bonus_logits_indices).to(
2104
2105
            self.device, non_blocking=True
        )
2106

2107
2108
        # Compute the draft token ids.
        # draft_token_indices:      [  1,   2,   3, 105, 106, 208]
2109
        draft_token_ids = self.input_ids.gpu[logits_indices]
2110
2111
        draft_token_ids = draft_token_ids[target_logits_indices + 1]

2112
        return SpecDecodeMetadata(
2113
2114
2115
            draft_token_ids=draft_token_ids,
            num_draft_tokens=num_draft_tokens.tolist(),
            cu_num_draft_tokens=cu_num_draft_tokens,
2116
            cu_num_sampled_tokens=cu_num_sampled_tokens,
2117
2118
2119
2120
2121
            target_logits_indices=target_logits_indices,
            bonus_logits_indices=bonus_logits_indices,
            logits_indices=logits_indices,
        )

2122
2123
2124
2125
2126
2127
2128
    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
2129
        self.kv_sharing_fast_prefill_logits_indices[:num_logits].copy_(logits_indices)
2130
2131
2132
2133
2134
        # 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_(
2135
2136
2137
2138
2139
2140
            logits_indices[-1].item()
        )
        if (
            self.compilation_config.cudagraph_mode != CUDAGraphMode.NONE
            and num_logits <= self.cudagraph_batch_sizes[-1]
        ):
2141
2142
2143
2144
2145
            # Use piecewise CUDA graphs.
            # Add padding to the batch size.
            num_logits_padded = self.vllm_config.pad_for_cudagraph(num_logits)
        else:
            num_logits_padded = num_logits
2146
2147
2148
        logits_indices_padded = self.kv_sharing_fast_prefill_logits_indices[
            :num_logits_padded
        ]
2149
2150
        return logits_indices_padded

2151
    def _batch_mm_inputs_from_scheduler(
2152
2153
        self,
        scheduler_output: "SchedulerOutput",
2154
2155
2156
2157
2158
    ) -> tuple[
        list[str],
        list[MultiModalKwargsItem],
        list[tuple[str, PlaceholderRange]],
    ]:
2159
        """Batch multimodal inputs from scheduled encoder inputs.
2160
2161
2162

        Args:
            scheduler_output: The scheduler output containing scheduled encoder
2163
                inputs.
2164
2165

        Returns:
2166
            A tuple of (mm_hashes, mm_kwargs, mm_lora_refs) where:
2167
2168
            - mm_hashes: List of multimodal hashes for each item
            - mm_kwargs: List of multimodal kwargs for each item
2169
            - mm_lora_refs: List of (req_id, placeholder_range) for each item
2170
        """
2171
2172
        scheduled_encoder_inputs = scheduler_output.scheduled_encoder_inputs
        if not scheduled_encoder_inputs:
2173
            return [], [], []
2174
2175

        mm_hashes = list[str]()
2176
        mm_kwargs = list[MultiModalKwargsItem]()
2177
2178
2179
        # Multimodal LoRA reference info to map each multimodal item
        # back to its request & position
        mm_lora_refs = list[tuple[str, PlaceholderRange]]()
2180
2181
        for req_id, encoder_input_ids in scheduled_encoder_inputs.items():
            req_state = self.requests[req_id]
2182
2183

            for mm_input_id in encoder_input_ids:
2184
                mm_feature = req_state.mm_features[mm_input_id]
2185
2186
                if mm_feature.data is None:
                    continue
2187
2188

                mm_hashes.append(mm_feature.identifier)
2189
                mm_kwargs.append(mm_feature.data)
2190
                mm_lora_refs.append((req_id, mm_feature.mm_position))
2191

2192
        return mm_hashes, mm_kwargs, mm_lora_refs
2193

2194
2195
2196
    def _execute_mm_encoder(
        self, scheduler_output: "SchedulerOutput"
    ) -> list[torch.Tensor]:
2197
2198
2199
        mm_hashes, mm_kwargs, mm_lora_refs = self._batch_mm_inputs_from_scheduler(
            scheduler_output
        )
2200
2201

        if not mm_kwargs:
2202
            return []
2203

2204
2205
2206
2207
2208
2209
2210
        # 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.
2211
        model = cast(SupportsMultiModal, self.model)
2212
2213
2214
2215
2216
2217
2218
2219
2220
2221
2222
2223
2224
2225
2226
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

        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,
                )

2269
        encoder_outputs: list[torch.Tensor] = []
2270
        for modality, num_items, mm_kwargs_group in group_mm_kwargs_by_modality(
2271
2272
2273
            mm_kwargs,
            device=self.device,
            pin_memory=self.pin_memory,
2274
        ):
2275
            curr_group_outputs: MultiModalEmbeddings
2276
2277

            # EVS-related change.
2278
            # (ekhvedchenia): Temporary hack to limit peak memory usage when
2279
            # processing multimodal data. This solves the issue with scheduler
2280
2281
2282
2283
            # 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)
2284
2285
2286
2287
2288
2289
2290
            # 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
            ):
2291
                curr_group_outputs_lst = list[torch.Tensor]()
2292
2293
2294
2295
2296
2297
2298
2299
2300
                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,
                        )
2301
                    )
2302

2303
                    micro_batch_outputs = model.embed_multimodal(
2304
2305
                        **micro_batch_mm_inputs
                    )
2306

2307
2308
2309
                    curr_group_outputs_lst.extend(micro_batch_outputs)

                curr_group_outputs = curr_group_outputs_lst
2310
2311
2312
2313
2314
2315
2316
2317
            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.
2318
                curr_group_outputs = model.embed_multimodal(**mm_kwargs_group)
2319

2320
2321
            sanity_check_mm_encoder_outputs(
                curr_group_outputs,
2322
                expected_num_items=num_items,
2323
            )
2324
            encoder_outputs.extend(curr_group_outputs)
2325

2326
        # Cache the encoder outputs by mm_hash
2327
        for mm_hash, output in zip(mm_hashes, encoder_outputs):
2328
            self.encoder_cache[mm_hash] = output
2329
2330
            logger.debug("Finish execute for mm hash %s", mm_hash)
            self.maybe_save_ec_to_connector(self.encoder_cache, mm_hash)
2331

2332
2333
        return encoder_outputs

2334
    def _gather_mm_embeddings(
2335
2336
        self,
        scheduler_output: "SchedulerOutput",
2337
        shift_computed_tokens: int = 0,
2338
2339
2340
    ) -> tuple[list[torch.Tensor], torch.Tensor]:
        total_num_scheduled_tokens = scheduler_output.total_num_scheduled_tokens

2341
2342
2343
2344
2345
        # 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]

2346
        mm_embeds = list[torch.Tensor]()
2347
        is_mm_embed = is_mm_embed_buf.cpu
2348
2349
2350
        is_mm_embed[:total_num_scheduled_tokens] = False

        req_start_idx = 0
2351
        should_sync_mrope_positions = False
2352
        should_sync_xdrope_positions = False
2353

2354
        for req_id in self.input_batch.req_ids:
2355
2356
            mm_embeds_req: list[torch.Tensor] = []

2357
            num_scheduled_tokens = scheduler_output.num_scheduled_tokens[req_id]
2358
            req_state = self.requests[req_id]
2359
            num_computed_tokens = req_state.num_computed_tokens + shift_computed_tokens
2360

2361
2362
            for mm_feature in req_state.mm_features:
                pos_info = mm_feature.mm_position
2363
2364
                start_pos = pos_info.offset
                num_encoder_tokens = pos_info.length
2365
2366
2367
2368
2369
2370
2371
2372
2373
2374
2375
2376
2377
2378
2379
2380

                # 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,
2381
2382
                    num_encoder_tokens,
                )
2383
                assert start_idx < end_idx
2384
2385
2386
2387
2388
2389
2390
                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
2391

2392
                mm_hash = mm_feature.identifier
2393
                encoder_output = self.encoder_cache.get(mm_hash, None)
2394
                assert encoder_output is not None, f"Encoder cache miss for {mm_hash}."
2395
2396
2397

                if (is_embed := pos_info.is_embed) is not None:
                    is_embed = is_embed[start_idx:end_idx]
2398
2399
2400
                    mm_embeds_item = encoder_output[curr_embeds_start:curr_embeds_end]
                else:
                    mm_embeds_item = encoder_output[start_idx:end_idx]
2401

2402
                req_start_pos = req_start_idx + start_pos - num_computed_tokens
2403
2404
2405
                is_mm_embed[req_start_pos + start_idx : req_start_pos + end_idx] = (
                    True if is_embed is None else is_embed
                )
2406
2407
2408
                mm_embeds_req.append(mm_embeds_item)

            if self.is_multimodal_pruning_enabled and self.uses_mrope:
2409
                assert req_state.mrope_positions is not None
2410
2411
2412
2413
2414
2415
2416
                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,
2417
2418
                    )
                )
2419
2420
2421
2422
                req_state.mrope_positions.copy_(new_mrope_positions)
                req_state.mrope_position_delta = new_delta

            mm_embeds.extend(mm_embeds_req)
2423
2424
            req_start_idx += num_scheduled_tokens

2425
        is_mm_embed = is_mm_embed_buf.copy_to_gpu(total_num_scheduled_tokens)
2426
2427
2428

        if should_sync_mrope_positions:
            self._calc_mrope_positions(scheduler_output)
2429
            self.mrope_positions.copy_to_gpu(total_num_scheduled_tokens)
2430

2431
2432
2433
2434
        if should_sync_xdrope_positions:
            self._calc_xdrope_positions(scheduler_output)
            self.xdrope_positions.copy_to_gpu(total_num_scheduled_tokens)

2435
        return mm_embeds, is_mm_embed
2436

2437
    def get_model(self) -> nn.Module:
2438
        # get raw model out of the cudagraph wrapper.
2439
        if isinstance(self.model, (CUDAGraphWrapper, UBatchWrapper)):
2440
            return self.model.unwrap()
2441
2442
        return self.model

2443
2444
2445
2446
2447
2448
2449
2450
2451
2452
2453
2454
2455
2456
2457
    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

2458
2459
2460
2461
2462
    def get_supported_pooling_tasks(self) -> list[PoolingTask]:
        model = self.get_model()
        if not is_pooling_model(model):
            return []

2463
2464
        supported_tasks = list(model.pooler.get_supported_tasks())

2465
2466
2467
2468
        if "score" in supported_tasks:
            num_labels = getattr(self.model_config.hf_config, "num_labels", 0)
            if num_labels != 1:
                supported_tasks.remove("score")
2469
                logger.debug_once("Score API is only enabled for num_labels == 1.")
2470
2471

        return supported_tasks
2472

2473
2474
2475
2476
2477
2478
2479
2480
2481
2482
    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)

2483
    def sync_and_slice_intermediate_tensors(
2484
2485
        self,
        num_tokens: int,
2486
        intermediate_tensors: IntermediateTensors | None,
2487
2488
        sync_self: bool,
    ) -> IntermediateTensors:
2489
2490
2491
        assert self.intermediate_tensors is not None

        tp = self.vllm_config.parallel_config.tensor_parallel_size
2492
        is_rs = is_residual_scattered_for_sp(self.vllm_config, num_tokens)
2493
2494
2495
2496
2497
2498

        # 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():
2499
                is_scattered = k == "residual" and is_rs
2500
                copy_len = num_tokens // tp if is_scattered else num_tokens
2501
                self.intermediate_tensors[k][:copy_len].copy_(
2502
2503
2504
2505
2506
2507
2508
2509
2510
2511
2512
2513
2514
                    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:
2515
2516
2517
2518
2519
2520
2521
        """
        Step for the EPLB (Expert Parallelism Load Balancing) state.
        """
        if not self.parallel_config.enable_eplb:
            return

        assert self.eplb_state is not None
2522
2523
        model = self.get_model()
        assert is_mixture_of_experts(model)
2524
2525
2526
        self.eplb_state.step(
            is_dummy,
            is_profile,
2527
            log_stats=self.parallel_config.eplb_config.log_balancedness,
2528
2529
        )

2530
2531
2532
2533
2534
    def _pool(
        self,
        hidden_states: torch.Tensor,
        num_scheduled_tokens: int,
        num_scheduled_tokens_np: np.ndarray,
2535
2536
2537
2538
        kv_connector_output: KVConnectorOutput | None,
    ) -> ModelRunnerOutput | AsyncModelRunnerOutput:
        num_reqs = self.input_batch.num_reqs
        assert num_reqs == len(self.input_batch.pooling_params), (
2539
2540
            "Either all or none of the requests in a batch must be pooling request"
        )
2541

2542
        hidden_states = hidden_states[:num_scheduled_tokens]
2543
        seq_lens_cpu = self.seq_lens.cpu[:num_reqs]
2544

2545
        pooling_metadata = self.input_batch.get_pooling_metadata()
2546
        pooling_metadata.build_pooling_cursor(
2547
            num_scheduled_tokens_np, seq_lens_cpu, device=hidden_states.device
2548
        )
2549

2550
2551
        model = cast(VllmModelForPooling, self.model)
        raw_pooler_output: PoolerOutput = model.pooler(
2552
            hidden_states=hidden_states, pooling_metadata=pooling_metadata
2553
        )
2554
2555
2556
2557
2558
2559
2560
2561
2562
2563
2564
2565
2566
2567
2568
2569
2570
2571
2572
2573
2574
2575
2576
2577

        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,
            )

2578
        raw_pooler_output = json_map_leaves(
2579
            lambda x: None if x is None else x.to("cpu", non_blocking=True),
2580
2581
            raw_pooler_output,
        )
2582
2583
2584
2585
        model_runner_output.pooler_output = [
            out if include else None
            for out, include in zip(raw_pooler_output, finished_mask)
        ]
2586
2587
        self._sync_device()

2588
        return model_runner_output
2589

2590
    def _pad_for_sequence_parallelism(self, num_scheduled_tokens: int) -> int:
2591
2592
2593
        # Pad tokens to multiple of tensor_parallel_size when
        # enabled collective fusion for SP
        tp_size = self.vllm_config.parallel_config.tensor_parallel_size
2594
        if self.compilation_config.pass_config.enable_sp and tp_size > 1:
2595
2596
2597
            return round_up(num_scheduled_tokens, tp_size)
        return num_scheduled_tokens

Patrick von Platen's avatar
Patrick von Platen committed
2598
2599
2600
2601
2602
2603
2604
2605
2606
2607
2608
    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

2609
    def _preprocess(
2610
2611
        self,
        scheduler_output: "SchedulerOutput",
2612
        num_input_tokens: int,  # Padded
2613
        intermediate_tensors: IntermediateTensors | None = None,
2614
    ) -> tuple[
2615
2616
        torch.Tensor | None,
        torch.Tensor | None,
2617
        torch.Tensor,
2618
        IntermediateTensors | None,
2619
        dict[str, Any],
2620
        ECConnectorOutput | None,
2621
    ]:
2622
        num_scheduled_tokens = scheduler_output.total_num_scheduled_tokens
2623
        is_first_rank = get_pp_group().is_first_rank
2624
        is_encoder_decoder = self.model_config.is_encoder_decoder
2625

2626
2627
        # _prepare_inputs may reorder the batch, so we must gather multi
        # modal outputs after that to ensure the correct order
2628
2629
        ec_connector_output = None

2630
        if self.supports_mm_inputs and is_first_rank and not is_encoder_decoder:
2631
            # Run the multimodal encoder if any.
2632
2633
2634
2635
2636
2637
            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)
2638

2639
2640
2641
            # 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.
2642
            inputs_embeds_scheduled = self.model.embed_input_ids(
2643
2644
2645
                self.input_ids.gpu[:num_scheduled_tokens],
                multimodal_embeddings=mm_embeds,
                is_multimodal=is_mm_embed,
2646
            )
2647

2648
            # TODO(woosuk): Avoid the copy. Optimize.
2649
            self.inputs_embeds.gpu[:num_scheduled_tokens].copy_(inputs_embeds_scheduled)
2650

Patrick von Platen's avatar
Patrick von Platen committed
2651
            input_ids, inputs_embeds = self._prepare_mm_inputs(num_input_tokens)
2652
            model_kwargs = {
2653
                **self._init_model_kwargs(),
2654
2655
                **self._extract_mm_kwargs(scheduler_output),
            }
2656
        elif self.enable_prompt_embeds and is_first_rank:
2657
2658
2659
2660
2661
2662
2663
2664
2665
2666
2667
2668
            # 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).
2669
2670
2671
            token_ids_idx = (
                self.is_token_ids.gpu[:num_scheduled_tokens]
                .nonzero(as_tuple=False)
2672
                .squeeze(1)
2673
            )
2674
2675
2676
            # Some tokens ids may need to become embeds
            if token_ids_idx.numel() > 0:
                token_ids = self.input_ids.gpu[token_ids_idx]
2677
                tokens_to_embeds = self.model.embed_input_ids(input_ids=token_ids)
2678
2679
2680
                self.inputs_embeds.gpu[token_ids_idx] = tokens_to_embeds

            inputs_embeds = self.inputs_embeds.gpu[:num_input_tokens]
2681
            model_kwargs = self._init_model_kwargs()
2682
            input_ids = None
2683
        else:
2684
2685
2686
2687
            # 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.
2688
            input_ids = self.input_ids.gpu[:num_input_tokens]
2689
            inputs_embeds = None
2690
            model_kwargs = self._init_model_kwargs()
2691

2692
        if self.uses_mrope:
2693
            positions = self.mrope_positions.gpu[:, :num_input_tokens]
2694
2695
        elif self.uses_xdrope_dim > 0:
            positions = self.xdrope_positions.gpu[:, :num_input_tokens]
2696
        else:
2697
            positions = self.positions.gpu[:num_input_tokens]
2698

2699
        if is_first_rank:
2700
2701
            intermediate_tensors = None
        else:
2702
            assert intermediate_tensors is not None
2703
            intermediate_tensors = self.sync_and_slice_intermediate_tensors(
2704
2705
                num_input_tokens, intermediate_tensors, True
            )
2706

2707
        if is_encoder_decoder and scheduler_output.scheduled_encoder_inputs:
2708
2709
2710
2711
2712
2713
2714
            # 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})
2715

2716
2717
2718
2719
2720
2721
        return (
            input_ids,
            inputs_embeds,
            positions,
            intermediate_tensors,
            model_kwargs,
2722
            ec_connector_output,
2723
        )
2724

2725
    def _sample(
2726
        self,
2727
2728
        logits: torch.Tensor | None,
        spec_decode_metadata: SpecDecodeMetadata | None,
2729
    ) -> SamplerOutput:
2730
        # Sample the next token and get logprobs if needed.
2731
        sampling_metadata = self.input_batch.sampling_metadata
2732
2733
2734
        # 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()
2735
        if spec_decode_metadata is None:
2736
            return self.sampler(
2737
2738
2739
                logits=logits,
                sampling_metadata=sampling_metadata,
            )
2740

2741
2742
2743
2744
2745
2746
        # 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)

2747
        sampler_output = self.rejection_sampler(
2748
2749
            spec_decode_metadata,
            None,  # draft_probs
2750
            logits,
2751
2752
            sampling_metadata,
        )
2753
        self._update_states_after_model_execute(sampler_output.sampled_token_ids)
2754
2755
2756
        return sampler_output

    def _bookkeeping_sync(
2757
2758
2759
        self,
        scheduler_output: "SchedulerOutput",
        sampler_output: SamplerOutput,
2760
        logits: torch.Tensor | None,
2761
2762
        hidden_states: torch.Tensor,
        num_scheduled_tokens: int,
2763
        spec_decode_metadata: SpecDecodeMetadata | None,
2764
    ) -> tuple[
2765
        dict[str, int],
2766
        LogprobsLists | None,
2767
        list[list[int]],
2768
        dict[str, LogprobsTensors | None],
2769
2770
2771
        list[str],
        dict[str, int],
        list[int],
2772
    ]:
2773
2774
2775
2776
        num_nans_in_logits = {}
        if envs.VLLM_COMPUTE_NANS_IN_LOGITS:
            num_nans_in_logits = self._get_nans_in_logits(logits)

2777
2778
2779
2780
        num_reqs = self.input_batch.num_reqs
        discard_sampled_tokens_req_indices = np.nonzero(
            self.discard_request_mask.np[:num_reqs]
        )[0]
2781
2782
2783
2784
        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)
2785

2786
2787
2788
        # 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()
2789
        req_id_to_index_output_copy = self.input_batch.req_id_to_index.copy()
2790
2791

        num_sampled_tokens = sampler_output.sampled_token_ids.shape[0]
2792
        sampled_token_ids = sampler_output.sampled_token_ids
2793
        logprobs_tensors = sampler_output.logprobs_tensors
2794
        invalid_req_indices = []
2795
        logprobs_lists = None
2796
2797
2798
2799
2800
2801
        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)
2802
2803
2804
                # 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()
2805
2806
2807

                if logprobs_tensors is not None:
                    logprobs_lists = logprobs_tensors.tolists()
2808
2809
            else:
                # Includes spec decode tokens.
2810
                valid_sampled_token_ids, logprobs_lists = RejectionSampler.parse_output(
2811
2812
                    sampled_token_ids,
                    self.input_batch.vocab_size,
2813
                    discard_sampled_tokens_req_indices,
2814
                    logprobs_tensors=logprobs_tensors,
2815
                )
2816
        else:
2817
            valid_sampled_token_ids = []
2818
            invalid_req_indices = discard_sampled_tokens_req_indices.tolist()
2819
2820
2821
2822
2823
            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.
2824
2825
2826
2827
            # 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
2828
2829
2830
2831
2832
            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
            }
2833

2834
2835
2836
2837
2838
        # 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.
2839
        req_ids = self.input_batch.req_ids
2840
2841
        for req_idx in range(num_sampled_tokens):
            if self.use_async_scheduling:
2842
                sampled_ids = [-1] if req_idx not in invalid_req_indices_set else None
2843
2844
            else:
                sampled_ids = valid_sampled_token_ids[req_idx]
2845

2846
            num_sampled_ids: int = len(sampled_ids) if sampled_ids else 0
2847

2848
            if not sampled_ids:
2849
2850
2851
                continue

            start_idx = self.input_batch.num_tokens_no_spec[req_idx]
2852
            end_idx = start_idx + num_sampled_ids
2853
2854
2855
2856
            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}"
2857
            )
2858

2859
2860
            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
2861
            self.input_batch.num_tokens_no_spec[req_idx] = end_idx
2862

2863
            req_id = req_ids[req_idx]
2864
2865
2866
            req_state = self.requests[req_id]
            req_state.output_token_ids.extend(sampled_ids)

2867
2868
2869
2870
2871
2872
        # Compute prompt logprobs if needed.
        prompt_logprobs_dict = self._get_prompt_logprobs_dict(
            hidden_states[:num_scheduled_tokens],
            scheduler_output.num_scheduled_tokens,
        )

2873
2874
2875
2876
2877
2878
2879
2880
2881
2882
        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,
        )

2883
2884
2885
2886
2887
2888
2889
2890
2891
2892
2893
2894
2895
2896
2897
    @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()

2898
2899
    def _model_forward(
        self,
2900
2901
2902
2903
        input_ids: torch.Tensor | None = None,
        positions: torch.Tensor | None = None,
        intermediate_tensors: IntermediateTensors | None = None,
        inputs_embeds: torch.Tensor | None = None,
2904
2905
2906
2907
2908
        **model_kwargs: dict[str, Any],
    ) -> Any:
        """Helper method to call the model forward pass.

        This method can be overridden by subclasses for model execution.
2909
        Motivation: We can inspect only this method versus
2910
2911
2912
2913
2914
2915
2916
2917
2918
2919
2920
2921
2922
2923
2924
2925
2926
2927
2928
2929
        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,
        )

2930
2931
2932
2933
2934
2935
2936
2937
2938
2939
2940
2941
2942
2943
2944
2945
2946
2947
2948
2949
2950
    @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
        )

2951
2952
2953
2954
2955
2956
2957
2958
2959
2960
2961
2962
2963
    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,
2964
        num_encoder_reqs: int = 0,
2965
    ) -> tuple[
2966
2967
        CUDAGraphMode,
        BatchDescriptor,
2968
        bool,
2969
2970
        torch.Tensor | None,
        CUDAGraphStat | None,
2971
    ]:
2972
2973
2974
2975
2976
2977
        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,
2978
        )
2979
2980
2981
2982
2983
        # 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
        )
2984
2985
2986
2987
2988
2989
2990

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

2991
        num_tokens_padded = self._pad_for_sequence_parallelism(num_tokens)
2992
        dispatch_cudagraph = (
2993
            lambda num_tokens, disable_full: self.cudagraph_dispatcher.dispatch(
2994
2995
2996
                num_tokens=num_tokens,
                has_lora=has_lora,
                uniform_decode=uniform_decode,
2997
                disable_full=disable_full,
2998
2999
3000
3001
3002
            )
            if not force_eager
            else (CUDAGraphMode.NONE, BatchDescriptor(num_tokens_padded))
        )

3003
        cudagraph_mode, batch_descriptor = dispatch_cudagraph(
3004
            num_tokens_padded, use_cascade_attn or has_encoder_output
3005
        )
3006
        num_tokens_padded = batch_descriptor.num_tokens
3007
3008
3009
3010
3011
3012
3013
3014
3015
        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"
            )
3016
3017
3018

        # Extra coordination when running data-parallel since we need to coordinate
        # across ranks
3019
        should_ubatch, num_tokens_across_dp = False, None
3020
3021
3022
3023
3024
3025
3026
3027
3028
        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
            )

3029
3030
3031
3032
3033
3034
3035
3036
3037
3038
3039
            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,
                )
3040
3041
            )

3042
            # Extract DP-synced values
3043
3044
3045
            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())
3046
3047
3048
3049
3050
                # 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,
                )
3051
3052
3053
3054
                # 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

3055
3056
3057
3058
3059
3060
3061
3062
3063
3064
3065
3066
        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,
3067
            should_ubatch,
3068
3069
3070
            num_tokens_across_dp,
            cudagraph_stats,
        )
3071

3072
3073
3074
3075
3076
3077
3078
3079
3080
3081
3082
3083
3084
3085
3086
3087
3088
3089
3090
3091
3092
3093
3094
3095
3096
3097
3098
3099
3100
3101
3102
3103
3104
3105
3106
3107
    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

3108
3109
3110
3111
    @torch.inference_mode()
    def execute_model(
        self,
        scheduler_output: "SchedulerOutput",
3112
        intermediate_tensors: IntermediateTensors | None = None,
3113
    ) -> ModelRunnerOutput | AsyncModelRunnerOutput | IntermediateTensors | None:
3114
3115
3116
3117
3118
        if self.execute_model_state is not None:
            raise RuntimeError(
                "State error: sample_tokens() must be called "
                "after execute_model() returns None."
            )
3119

3120
3121
3122
3123
3124
3125
3126
        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.")

3127
3128
3129
3130
3131
        if scheduler_output.preempted_req_ids and has_kv_transfer_group():
            get_kv_transfer_group().handle_preemptions(
                scheduler_output.preempted_req_ids
            )

3132
        num_scheduled_tokens = scheduler_output.total_num_scheduled_tokens
3133
3134
3135
3136
3137
3138
        with (
            record_function_or_nullcontext("gpu_model_runner: preprocess"),
            self.synchronize_input_prep(),
        ):
            # Update persistent batch states.
            self._update_states(scheduler_output)
3139

3140
3141
            if has_ec_transfer() and get_ec_transfer().is_producer:
                with self.maybe_get_ec_connector_output(
3142
                    scheduler_output,
3143
3144
3145
3146
3147
3148
3149
3150
3151
3152
3153
3154
3155
3156
3157
3158
3159
3160
3161
3162
3163
3164
3165
3166
3167
3168
3169
3170
                    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"
3171
3172
                )

3173
3174
3175
3176
3177
3178
            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
3179

3180
3181
3182
3183
            logits_indices, spec_decode_metadata = self._prepare_inputs(
                scheduler_output,
                num_scheduled_tokens_np,
            )
3184

3185
3186
3187
3188
3189
            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(
3190
                    num_scheduled_tokens_np,
3191
3192
                    self.input_batch.num_computed_tokens_cpu[:num_reqs],
                    scheduler_output.num_common_prefix_blocks,
3193
3194
                )

3195
3196
3197
3198
3199
3200
3201
3202
3203
3204
3205
3206
3207
3208
            (
                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),
            )
3209

3210
3211
3212
3213
3214
3215
3216
3217
3218
3219
3220
3221
3222
3223
3224
3225
3226
3227
3228
3229
3230
3231
3232
3233
3234
3235
3236
3237
3238
3239
3240
3241
3242
3243
3244
3245
3246
3247
3248
3249
3250
3251
3252
3253
            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

            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,
3254
                )
3255
            )
3256

3257
3258
3259
3260
3261
3262
3263
3264
3265
            (
                input_ids,
                inputs_embeds,
                positions,
                intermediate_tensors,
                model_kwargs,
                ec_connector_output,
            ) = self._preprocess(
                scheduler_output, num_tokens_padded, intermediate_tensors
3266
            )
3267

3268
        # Set cudagraph mode to none if calc_kv_scales is true.
3269
3270
3271
        # KV scales calculation involves dynamic operations that are incompatible
        # with CUDA graph capture.
        if self.calculate_kv_scales:
3272
            cudagraph_mode = CUDAGraphMode.NONE
3273
3274
            # Mark KV scales as calculated after the first forward pass
            self.calculate_kv_scales = False
3275

3276
3277
        # Run the model.
        # Use persistent buffers for CUDA graphs.
3278
3279
        with (
            set_forward_context(
3280
3281
                attn_metadata,
                self.vllm_config,
3282
                num_tokens=num_tokens_padded,
3283
                num_tokens_across_dp=num_tokens_across_dp,
3284
3285
                cudagraph_runtime_mode=cudagraph_mode,
                batch_descriptor=batch_desc,
3286
                ubatch_slices=ubatch_slices_padded,
3287
            ),
3288
            record_function_or_nullcontext("gpu_model_runner: forward"),
3289
3290
            self.maybe_get_kv_connector_output(scheduler_output) as kv_connector_output,
        ):
3291
            model_output = self._model_forward(
3292
3293
3294
3295
3296
3297
3298
                input_ids=input_ids,
                positions=positions,
                intermediate_tensors=intermediate_tensors,
                inputs_embeds=inputs_embeds,
                **model_kwargs,
            )

3299
        with record_function_or_nullcontext("gpu_model_runner: postprocess"):
3300
            if self.use_aux_hidden_state_outputs:
3301
                # True when EAGLE 3 is used.
3302
3303
                hidden_states, aux_hidden_states = model_output
            else:
3304
                # Common case.
3305
3306
3307
                hidden_states = model_output
                aux_hidden_states = None

3308
3309
3310
3311
3312
            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)
3313
                    hidden_states.kv_connector_output = kv_connector_output
3314
                    self.kv_connector_output = kv_connector_output
3315
                    return hidden_states
3316

3317
                if self.is_pooling_model:
3318
                    # Return the pooling output.
3319
3320
3321
3322
3323
                    return self._pool(
                        hidden_states,
                        num_scheduled_tokens,
                        num_scheduled_tokens_np,
                        kv_connector_output,
3324
                    )
3325
3326

                sample_hidden_states = hidden_states[logits_indices]
3327
                logits = self.model.compute_logits(sample_hidden_states)
3328
3329
3330
3331
            else:
                # Rare case.
                assert not self.is_pooling_model

3332
                sample_hidden_states = hidden_states[logits_indices]
3333
                if not get_pp_group().is_last_rank:
3334
                    all_gather_tensors = {
3335
                        "residual": not is_residual_scattered_for_sp(
3336
                            self.vllm_config, num_tokens_padded
3337
                        )
3338
                    }
3339
                    get_pp_group().send_tensor_dict(
3340
3341
                        hidden_states.tensors,
                        all_gather_group=get_tp_group(),
3342
3343
                        all_gather_tensors=all_gather_tensors,
                    )
3344
3345
                    logits = None
                else:
3346
                    logits = self.model.compute_logits(sample_hidden_states)
3347

3348
                model_output_broadcast_data: dict[str, Any] = {}
3349
3350
3351
                if logits is not None:
                    model_output_broadcast_data["logits"] = logits.contiguous()

3352
                broadcasted = get_pp_group().broadcast_tensor_dict(
3353
3354
                    model_output_broadcast_data, src=len(get_pp_group().ranks) - 1
                )
3355
3356
                assert broadcasted is not None
                logits = broadcasted["logits"]
3357

3358
3359
3360
3361
3362
3363
3364
3365
        self.execute_model_state = ExecuteModelState(
            scheduler_output,
            logits,
            spec_decode_metadata,
            spec_decode_common_attn_metadata,
            hidden_states,
            sample_hidden_states,
            aux_hidden_states,
3366
            ec_connector_output,
3367
            cudagraph_stats,
3368
        )
3369
        self.kv_connector_output = kv_connector_output
3370
3371
3372
3373
3374
3375
        return None

    @torch.inference_mode
    def sample_tokens(
        self, grammar_output: "GrammarOutput | None"
    ) -> ModelRunnerOutput | AsyncModelRunnerOutput | IntermediateTensors:
3376
3377
3378
        kv_connector_output = self.kv_connector_output
        self.kv_connector_output = None

3379
3380
        if self.execute_model_state is None:
            # Nothing to do (PP non-final rank case), output isn't used.
3381
            if not kv_connector_output:
3382
                return None  # type: ignore[return-value]
3383
3384
3385
3386
3387
3388
3389
3390
3391

            # 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
3392
3393
3394
3395
3396
3397
3398
3399
3400
3401

        # Unpack ephemeral state.
        (
            scheduler_output,
            logits,
            spec_decode_metadata,
            spec_decode_common_attn_metadata,
            hidden_states,
            sample_hidden_states,
            aux_hidden_states,
3402
            ec_connector_output,
3403
            cudagraph_stats,
3404
3405
3406
3407
3408
3409
3410
3411
3412
        ) = 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
            )
3413

3414
        with record_function_or_nullcontext("gpu_model_runner: sample"):
3415
3416
            sampler_output = self._sample(logits, spec_decode_metadata)

3417
3418
        self._draft_token_ids = None
        self._draft_token_req_ids = None
3419
3420
        self.input_batch.prev_sampled_token_ids = None

3421
        def propose_draft_token_ids(sampled_token_ids):
3422
            assert spec_decode_common_attn_metadata is not None
3423
            with record_function_or_nullcontext("gpu_model_runner: draft"):
3424
3425
3426
3427
3428
3429
3430
3431
3432
3433
                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,
                )
3434
                self._copy_draft_token_ids_to_cpu(scheduler_output)
3435

3436
        spec_config = self.speculative_config
3437
3438
3439
3440
3441
        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
3442
            )
3443
            if spec_config.use_eagle() and not spec_config.disable_padded_drafter_batch:
3444
3445
                # EAGLE speculative decoding can use the GPU sampled tokens
                # as inputs, and does not need to wait for bookkeeping to finish.
3446
3447
3448
3449
3450
3451
3452
3453
3454
3455
3456
3457
3458
3459
                assert isinstance(self.drafter, EagleProposer)
                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,
                        )
3460
                    )
3461
3462
3463
3464
3465
3466
3467
3468
3469
3470
3471
                    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
3472

3473
        with record_function_or_nullcontext("gpu_model_runner: bookkeep"):
3474
3475
3476
3477
3478
3479
3480
3481
            (
                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,
3482
3483
3484
3485
3486
            ) = self._bookkeeping_sync(
                scheduler_output,
                sampler_output,
                logits,
                hidden_states,
3487
                scheduler_output.total_num_scheduled_tokens,
3488
                spec_decode_metadata,
3489
            )
3490

3491
        if propose_drafts_after_bookkeeping:
3492
3493
3494
            # 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)
3495

3496
        with record_function_or_nullcontext("gpu_model_runner: eplb"):
3497
            self.eplb_step()
3498

3499
        with record_function_or_nullcontext("gpu_model_runner: ModelRunnerOutput"):
3500
3501
3502
3503
3504
3505
3506
            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.")

3507
3508
3509
3510
3511
3512
3513
            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,
3514
3515
3516
                ec_connector_output=ec_connector_output
                if self.supports_mm_inputs
                else None,
3517
                num_nans_in_logits=num_nans_in_logits,
3518
                cudagraph_stats=cudagraph_stats,
3519
            )
3520

3521
3522
        if not self.use_async_scheduling:
            return output
3523

3524
3525
3526
3527
3528
3529
3530
3531
3532
        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,
3533
                vocab_size=self.input_batch.vocab_size,
3534
3535
3536
3537
3538
            )
        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
3539
            # any requests with sampling params that require output ids.
3540
3541
3542
3543
            self.input_batch.set_async_sampled_token_ids(
                async_output.sampled_token_ids_cpu,
                async_output.async_copy_ready_event,
            )
3544
3545
3546

        return async_output

3547
    def take_draft_token_ids(self) -> DraftTokenIds | None:
3548
        if not self.num_spec_tokens or not self._draft_token_req_ids:
3549
            return None
3550
        draft_token_ids, req_ids = self._get_draft_token_ids_cpu()
3551
        return DraftTokenIds(req_ids, draft_token_ids)
3552

3553
3554
3555
    def _copy_draft_token_ids_to_cpu(
        self, scheduler_output: "SchedulerOutput", zeros_only: bool = False
    ) -> None:
3556
3557
3558
3559
3560
3561
        # 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
        ):
3562
3563
3564
            return
        # We must also set the corresponding request ids.
        self._draft_token_req_ids = self.input_batch.req_ids.copy()
3565

3566
3567
3568
3569
3570
3571
3572
3573
3574
3575
3576
3577
3578
3579
3580
3581
3582
3583
3584
3585
        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()

3586
    def _get_draft_token_ids_cpu(self) -> tuple[list[list[int]], list[str]]:
3587
        if isinstance(self._draft_token_ids, list):
3588
3589
3590
3591
            return self._draft_token_ids, self.input_batch.req_ids
        req_ids = self._draft_token_req_ids
        if req_ids is None:
            return [], []
3592
3593
3594
        assert self.draft_token_ids_event is not None
        assert self.draft_token_ids_cpu is not None
        self.draft_token_ids_event.synchronize()
3595
        return self.draft_token_ids_cpu[: len(req_ids)].tolist(), req_ids
3596

3597
3598
3599
3600
3601
3602
3603
3604
3605
3606
3607
3608
3609
    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
3610
            assert counts_cpu is not None
3611
3612
3613
3614
3615
3616
3617
3618
            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
3619
3620
        sampled_count_event = self.valid_sampled_token_count_event
        if sampled_count_event is None or prev_sampled_token_ids is None:
3621
3622
3623
            return []

        counts_cpu = self.valid_sampled_token_count_cpu
3624
3625
        assert counts_cpu is not None
        sampled_count_event.synchronize()
3626
3627
        return counts_cpu[: prev_sampled_token_ids.shape[0]].tolist()

3628
3629
3630
    def propose_draft_token_ids(
        self,
        scheduler_output: "SchedulerOutput",
3631
        sampled_token_ids: torch.Tensor | list[list[int]],
3632
3633
3634
        sampling_metadata: SamplingMetadata,
        hidden_states: torch.Tensor,
        sample_hidden_states: torch.Tensor,
3635
3636
        aux_hidden_states: list[torch.Tensor] | None,
        spec_decode_metadata: SpecDecodeMetadata | None,
3637
        common_attn_metadata: CommonAttentionMetadata,
3638
    ) -> list[list[int]] | torch.Tensor:
3639
        num_scheduled_tokens = scheduler_output.total_num_scheduled_tokens
3640
3641
3642
        spec_config = self.speculative_config
        assert spec_config is not None
        if spec_config.method == "ngram":
3643
            assert isinstance(sampled_token_ids, list)
3644
            assert isinstance(self.drafter, NgramProposer)
3645
            draft_token_ids = self.drafter.propose(
3646
                sampled_token_ids,
3647
3648
                self.input_batch.num_tokens_no_spec,
                self.input_batch.token_ids_cpu,
3649
            )
3650
        elif spec_config.method == "suffix":
3651
3652
3653
            assert isinstance(sampled_token_ids, list)
            assert isinstance(self.drafter, SuffixDecodingProposer)
            draft_token_ids = self.drafter.propose(self.input_batch, sampled_token_ids)
3654
        elif spec_config.method == "medusa":
3655
            assert isinstance(sampled_token_ids, list)
3656
            assert isinstance(self.drafter, MedusaProposer)
3657

3658
3659
            if sample_hidden_states.shape[0] == len(sampled_token_ids):
                # The input to the target model does not include draft tokens.
3660
3661
3662
3663
                hidden_states = sample_hidden_states
            else:
                indices = []
                offset = 0
3664
3665
3666
                assert spec_decode_metadata is not None, (
                    "No spec decode metadata for medusa"
                )
3667
                for num_draft, tokens in zip(
3668
3669
                    spec_decode_metadata.num_draft_tokens, sampled_token_ids
                ):
3670
                    indices.append(offset + len(tokens) - 1)
3671
                    offset += num_draft + 1
3672
                indices = torch.tensor(indices, device=self.device)
3673
3674
                hidden_states = sample_hidden_states[indices]

3675
            draft_token_ids = self.drafter.propose(
3676
3677
3678
                target_hidden_states=hidden_states,
                sampling_metadata=sampling_metadata,
            )
3679
        elif spec_config.use_eagle():
3680
            assert isinstance(self.drafter, EagleProposer)
3681

3682
            if spec_config.disable_padded_drafter_batch:
3683
3684
3685
                # 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.
3686
3687
                assert isinstance(sampled_token_ids, list), (
                    "sampled_token_ids should be a python list when"
3688
                    "padded-batch is disabled."
3689
                )
3690
                next_token_ids = self.drafter.prepare_next_token_ids_cpu(
3691
3692
3693
3694
3695
                    sampled_token_ids,
                    self.requests,
                    self.input_batch,
                    scheduler_output.num_scheduled_tokens,
                )
3696
3697
3698
3699
3700
            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.
3701
3702
                assert isinstance(sampled_token_ids, torch.Tensor), (
                    "sampled_token_ids should be a torch.Tensor when"
3703
                    "padded-batch is enabled."
3704
3705
                )
                next_token_ids, valid_sampled_tokens_count = (
3706
3707
3708
3709
3710
                    self.drafter.prepare_next_token_ids_padded(
                        common_attn_metadata,
                        sampled_token_ids,
                        self.requests,
                        self.input_batch,
3711
                        self.discard_request_mask.gpu,
3712
                    )
3713
                )
3714
3715
3716
                self._copy_valid_sampled_token_count(
                    next_token_ids, valid_sampled_tokens_count
                )
Jiayi Yao's avatar
Jiayi Yao committed
3717

3718
            num_rejected_tokens_gpu = None
3719
            if spec_decode_metadata is None:
3720
                token_indices_to_sample = None
3721
                # input_ids can be None for multimodal models.
3722
                target_token_ids = self.input_ids.gpu[:num_scheduled_tokens]
3723
                target_positions = self._get_positions(num_scheduled_tokens)
3724
                if self.use_aux_hidden_state_outputs:
Wentao Ye's avatar
Wentao Ye committed
3725
                    assert aux_hidden_states is not None
3726
                    target_hidden_states = torch.cat(
3727
3728
                        [h[:num_scheduled_tokens] for h in aux_hidden_states], dim=-1
                    )
3729
3730
                else:
                    target_hidden_states = hidden_states[:num_scheduled_tokens]
3731
            else:
3732
                if spec_config.disable_padded_drafter_batch:
3733
                    token_indices_to_sample = None
3734
3735
3736
3737
3738
                    common_attn_metadata, token_indices = self.drafter.prepare_inputs(
                        common_attn_metadata,
                        sampled_token_ids,
                        spec_decode_metadata.num_draft_tokens,
                    )
3739
3740
3741
3742
3743
3744
3745
3746
3747
                    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]
3748
                else:
3749
3750
3751
3752
3753
3754
3755
3756
                    (
                        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,
3757
                    )
3758
3759
3760
3761
3762
3763
3764
3765
3766
3767
3768
                    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]
3769

3770
            if self.supports_mm_inputs:
3771
3772
3773
3774
3775
3776
                mm_embed_inputs = self._gather_mm_embeddings(
                    scheduler_output,
                    shift_computed_tokens=1,
                )
            else:
                mm_embed_inputs = None
3777

3778
            draft_token_ids = self.drafter.propose(
3779
3780
3781
3782
                target_token_ids=target_token_ids,
                target_positions=target_positions,
                target_hidden_states=target_hidden_states,
                next_token_ids=next_token_ids,
3783
                last_token_indices=token_indices_to_sample,
3784
                sampling_metadata=sampling_metadata,
3785
                common_attn_metadata=common_attn_metadata,
3786
                mm_embed_inputs=mm_embed_inputs,
3787
                num_rejected_tokens_gpu=num_rejected_tokens_gpu,
3788
            )
3789

3790
        return draft_token_ids
3791

3792
3793
3794
    def update_config(self, overrides: dict[str, Any]) -> None:
        allowed_config_names = {"load_config", "model_config"}
        for config_name, config_overrides in overrides.items():
3795
3796
            assert config_name in allowed_config_names, (
                f"Config `{config_name}` not supported. "
3797
                f"Allowed configs: {allowed_config_names}"
3798
            )
3799
3800
3801
3802
            config = getattr(self, config_name)
            new_config = update_config(config, config_overrides)
            setattr(self, config_name, new_config)

3803
3804
3805
3806
3807
    def load_model(self, eep_scale_up: bool = False) -> None:
        """
        Args:
            eep_scale_up: the model loading is for elastic EP scale up.
        """
3808
3809
3810
3811
3812
        logger.info_once(
            "Starting to load model %s...",
            self.model_config.model,
            scope="global",
        )
3813
3814
3815
3816
3817
        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)
        )
3818

3819
3820
3821
3822
        if self.parallel_config.enable_eplb:
            self.eplb_state = EplbState(self.parallel_config, self.device)
            eplb_models = 0

3823
3824
3825
3826
3827
3828
3829
3830
3831
3832
        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
3833
                    )
3834
3835
3836
3837
3838
3839
3840
3841
3842
3843
3844
3845
3846
3847
3848
                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,
                        )
3849

3850
3851
3852
3853
3854
3855
3856
3857
3858
3859
3860
3861
3862
3863
3864
3865
3866
3867
3868
3869
3870
3871
                        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
3872

3873
3874
3875
3876
3877
3878
                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"
                        )
3879

3880
3881
3882
3883
3884
3885
3886
3887
3888
3889
3890
3891
3892
3893
3894
3895
3896
3897
3898
3899
3900
3901
3902
3903
3904
                    # 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
3905
        logger.info_once(
3906
3907
            "Model loading took %s GiB memory and %.6f seconds",
            format_gib(self.model_memory_usage),
3908
            time_after_load - time_before_load,
3909
            scope="local",
3910
        )
3911
        prepare_communication_buffer_for_model(self.model)
3912
3913
3914
3915
        if (drafter := getattr(self, "drafter", None)) and (
            drafter_model := getattr(drafter, "model", None)
        ):
            prepare_communication_buffer_for_model(drafter_model)
3916
        mm_config = self.model_config.multimodal_config
3917
        self.is_multimodal_pruning_enabled = (
3918
            supports_multimodal_pruning(self.get_model())
3919
3920
            and mm_config is not None
            and mm_config.is_multimodal_pruning_enabled()
3921
        )
3922

3923
        if is_mixture_of_experts(self.model) and self.parallel_config.enable_eplb:
3924
3925
3926
3927
3928
3929
3930
3931
3932
3933
3934
            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(
3935
                self.model,
3936
                self.model_config,
3937
3938
3939
                global_expert_load,
                old_global_expert_indices,
                rank_mapping,
3940
            )
3941
3942
            if self.eplb_state.is_async:
                self.eplb_state.start_async_loop(rank_mapping=rank_mapping)
3943

3944
        if (
3945
3946
            self.vllm_config.compilation_config.mode
            == CompilationMode.STOCK_TORCH_COMPILE
3947
            and supports_dynamo()
3948
        ):
3949
            backend = self.vllm_config.compilation_config.init_backend(self.vllm_config)
3950
            compilation_counter.stock_torch_compile_count += 1
3951
            self.model.compile(fullgraph=True, backend=backend)
3952
            return
3953
        # for other compilation modes, cudagraph behavior is controlled by
3954
3955
3956
        # CudagraphWraper and CudagraphDispatcher of vllm.

        # wrap the model with full cudagraph wrapper if needed.
3957
3958
        cudagraph_mode = self.compilation_config.cudagraph_mode
        assert cudagraph_mode is not None
3959
3960
3961
3962
        if (
            cudagraph_mode.has_full_cudagraphs()
            and not self.parallel_config.use_ubatching
        ):
3963
3964
3965
            self.model = CUDAGraphWrapper(
                self.model, self.vllm_config, runtime_mode=CUDAGraphMode.FULL
            )
3966
        elif self.parallel_config.use_ubatching:
3967
            if cudagraph_mode.has_full_cudagraphs():
3968
3969
3970
                self.model = UBatchWrapper(
                    self.model, self.vllm_config, CUDAGraphMode.FULL, self.device
                )
3971
            else:
3972
3973
3974
                self.model = UBatchWrapper(
                    self.model, self.vllm_config, CUDAGraphMode.NONE, self.device
                )
3975

3976
    def _get_eagle3_aux_layers_from_config(self) -> tuple[int, ...] | None:
3977
3978
3979
3980
3981
3982
3983
3984
3985
3986
3987
3988
3989
3990
3991
3992
3993
3994
3995
3996
3997
3998
3999
        """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

4000
    def reload_weights(self) -> None:
4001
        assert getattr(self, "model", None) is not None, (
4002
            "Cannot reload weights before model is loaded."
4003
        )
4004
4005
        model_loader = get_model_loader(self.load_config)
        logger.info("Reloading weights inplace...")
4006
        model_loader.load_weights(self.get_model(), model_config=self.model_config)
4007

4008
4009
4010
4011
4012
    def save_tensorized_model(
        self,
        tensorizer_config: "TensorizerConfig",
    ) -> None:
        TensorizerLoader.save_model(
4013
            self.get_model(),
4014
            tensorizer_config=tensorizer_config,
4015
            model_config=self.model_config,
4016
4017
        )

4018
4019
4020
    def _get_prompt_logprobs_dict(
        self,
        hidden_states: torch.Tensor,
4021
        num_scheduled_tokens: dict[str, int],
4022
    ) -> dict[str, LogprobsTensors | None]:
4023
        num_prompt_logprobs_dict = self.num_prompt_logprobs
4024
4025
4026
        if not num_prompt_logprobs_dict:
            return {}

4027
        in_progress_dict = self.input_batch.in_progress_prompt_logprobs_cpu
4028
        prompt_logprobs_dict: dict[str, LogprobsTensors | None] = {}
4029
4030
4031
4032
4033

        # 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():
4034
4035
4036
4037
            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
4038
4039
4040

            # Get metadata for this request.
            request = self.requests[req_id]
4041
4042
4043
4044
            if request.prompt_token_ids is None:
                # Prompt logprobs is incompatible with prompt embeddings
                continue

4045
4046
            num_prompt_tokens = len(request.prompt_token_ids)
            prompt_token_ids = torch.tensor(request.prompt_token_ids).to(
4047
4048
                self.device, non_blocking=True
            )
4049

4050
4051
4052
4053
4054
4055
            # 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(
4056
4057
                    num_prompt_tokens - 1, num_prompt_logprobs + 1
                )
4058
4059
                in_progress_dict[req_id] = logprobs_tensors

4060
            # Determine number of logits to retrieve.
4061
4062
            start_idx = request.num_computed_tokens
            start_tok = start_idx + 1
4063
            num_remaining_tokens = num_prompt_tokens - start_tok
4064
            if num_tokens <= num_remaining_tokens:
4065
                # This is a chunk, more tokens remain.
4066
4067
4068
                # 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.
4069
4070
4071
4072
4073
                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)
4074
4075
4076
4077
4078
4079
4080
                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
4081
4082
4083
4084
4085

            # 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]
4086
            offset = self.query_start_loc.np[req_idx].item()
4087
            prompt_hidden_states = hidden_states[offset : offset + num_logits]
4088
            logits = self.model.compute_logits(prompt_hidden_states)
4089
4090
4091
4092

            # 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.
4093
            tgt_token_ids = prompt_token_ids[start_tok : start_tok + num_logits]
4094
4095

            # Compute prompt logprobs.
4096
4097
            logprobs = self.sampler.compute_logprobs(logits)
            token_ids, logprobs, ranks = self.sampler.gather_logprobs(
4098
4099
                logprobs, num_prompt_logprobs, tgt_token_ids
            )
4100
4101

            # Transfer GPU->CPU async.
4102
4103
            chunk_slice = slice(start_idx, start_idx + num_logits)
            logprobs_tensors.logprob_token_ids[chunk_slice].copy_(
4104
4105
4106
                token_ids, non_blocking=True
            )
            logprobs_tensors.logprobs[chunk_slice].copy_(logprobs, non_blocking=True)
4107
            logprobs_tensors.selected_token_ranks[chunk_slice].copy_(
4108
4109
                ranks, non_blocking=True
            )
4110
4111
4112
4113
4114

        # 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]
4115
            del in_progress_dict[req_id]
4116
4117

        # Must synchronize the non-blocking GPU->CPU transfers.
4118
        if prompt_logprobs_dict:
4119
            self._sync_device()
4120
4121
4122

        return prompt_logprobs_dict

4123
4124
    def _get_nans_in_logits(
        self,
4125
        logits: torch.Tensor | None,
4126
4127
4128
4129
4130
4131
4132
4133
4134
4135
4136
    ) -> 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])
4137
4138
4139
                    if num_nans_for_index is not None and req_index < logits.shape[0]
                    else 0
                )
4140
4141
4142
4143
            return num_nans_in_logits
        except IndexError:
            return {}

4144
    @contextmanager
4145
4146
4147
    def maybe_randomize_inputs(
        self, input_ids: torch.Tensor | None, inputs_embeds: torch.Tensor | None
    ):
4148
4149
4150
4151
        """
        Randomize input_ids if VLLM_RANDOMIZE_DP_DUMMY_INPUTS is set.
        This is to help balance expert-selection
         - during profile_run
4152
         - during DP rank dummy run
4153
        """
4154

4155
4156
4157
4158
        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
4159
        elif input_ids is not None:
4160
4161
4162
4163

            @functools.cache
            def rand_input_ids() -> torch.Tensor:
                return torch.randint_like(
4164
                    self.input_ids.gpu,
4165
4166
                    low=0,
                    high=self.model_config.get_vocab_size(),
4167
                )
4168

4169
            logger.debug_once("Randomizing dummy input_ids for DP Rank")
4170
            input_ids.copy_(rand_input_ids()[: input_ids.size(0)], non_blocking=True)
4171
4172
            yield
            input_ids.fill_(0)
4173
4174
4175
4176
4177
4178
4179
4180
4181
4182
4183
4184
4185
4186
4187
        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)
4188

4189
4190
4191
4192
4193
4194
    def _get_mm_dummy_batch(
        self,
        modality: str,
        max_items_per_batch: int,
    ) -> BatchedTensorInputs:
        """Dummy data for profiling and precompiling multimodal models."""
4195
4196
        assert self.mm_budget is not None

4197
        dummy_decoder_data = self.mm_registry.get_decoder_dummy_data(
4198
            model_config=self.model_config,
4199
            seq_len=self.max_model_len,
4200
            mm_counts={modality: 1},
4201
            cache=self.mm_budget.cache,
4202
4203
4204
4205
        )
        dummy_mm_data = dummy_decoder_data.multi_modal_data

        # Result in the maximum GPU consumption of the model
4206
4207
        dummy_mm_item = dummy_mm_data[modality][0]
        dummy_mm_items = [dummy_mm_item] * max_items_per_batch
4208

4209
4210
4211
4212
4213
4214
4215
4216
        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,
            )
        )
4217

4218
4219
4220
4221
    @torch.inference_mode()
    def _dummy_run(
        self,
        num_tokens: int,
4222
        cudagraph_runtime_mode: CUDAGraphMode | None = None,
4223
4224
        force_attention: bool = False,
        uniform_decode: bool = False,
4225
        allow_microbatching: bool = True,
4226
4227
        skip_eplb: bool = False,
        is_profile: bool = False,
4228
        create_mixed_batch: bool = False,
4229
        remove_lora: bool = True,
4230
        activate_lora: bool = False,
Rémi Delacourt's avatar
Rémi Delacourt committed
4231
        is_graph_capturing: bool = False,
4232
    ) -> tuple[torch.Tensor, torch.Tensor]:
4233
4234
4235
4236
4237
4238
4239
        """
        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.
4240
                - if not set will determine the cudagraph mode based on using
4241
                    the self.cudagraph_dispatcher.
4242
4243
4244
4245
                - CUDAGraphMode.NONE: No cudagraph, for warm up and profile run
                - CUDAGraphMode.PIECEWISE: Piecewise cudagraph.
                - CUDAGraphMode.FULL: Full cudagraph, attention metadata is
                    needed.
4246
            force_attention: If True, always create attention metadata. Used to
4247
4248
4249
4250
                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.
4251
4252
            create_mixed_batch: If True, create a mixed batch with both decode
                (1 token) and prefill (multiple tokens) requests.
4253
            remove_lora: If False, dummy LoRAs are not destroyed after the run
4254
            activate_lora: If False, dummy_run is performed without LoRAs.
4255
        """
4256
4257
4258
4259
4260
        if supports_mm_encoder_only(self.model):
            # 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([])

4261
4262
4263
4264
        assert (
            cudagraph_runtime_mode is None
            or cudagraph_runtime_mode.valid_runtime_modes()
        )
4265

4266
        # If cudagraph_mode.decode_mode() == FULL and
4267
        # cudagraph_mode.separate_routine(). This means that we are using
4268
4269
4270
4271
4272
4273
4274
4275
4276
4277
4278
        # 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.
4279
        max_query_len = self.uniform_decode_query_len if uniform_decode else num_tokens
4280

4281
4282
4283
4284
4285
        # 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
4286
4287
4288
4289
        if create_mixed_batch:
            assert not uniform_decode
            # Create mixed batch:
            # first half decode tokens, second half one prefill
4290
            num_decode_tokens = min(max_num_reqs - 1, num_tokens // 2)
4291
4292
4293
4294
            num_prefill_tokens = num_tokens - num_decode_tokens
            num_reqs = num_decode_tokens + 1

            # Create decode requests (1 token each) followed by prefill request
4295
            num_scheduled_tokens_list = [1] * num_decode_tokens + [num_prefill_tokens]
4296
4297
4298
            # Note: Overriding max_query_len to be the prefill tokens
            max_query_len = num_prefill_tokens
        elif uniform_decode:
4299
            assert not create_mixed_batch
4300
            num_reqs = min(max_num_reqs, cdiv(num_tokens, max_query_len))
4301
4302
            num_scheduled_tokens_list = [max_query_len] * num_reqs
            if num_tokens % max_query_len != 0:
4303
                num_scheduled_tokens_list[-1] = num_tokens % max_query_len
4304
4305
4306
4307
4308
4309
        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

4310
4311
        assert sum(num_scheduled_tokens_list) == num_tokens
        assert len(num_scheduled_tokens_list) == num_reqs
4312
        num_scheduled_tokens = np.array(num_scheduled_tokens_list, dtype=np.int32)
4313
4314
        num_tokens_unpadded = int(num_scheduled_tokens.sum())

4315
        num_sampled_tokens = np.ones(num_reqs, dtype=np.int32)
4316

4317
        _cudagraph_mode, batch_desc, should_ubatch, num_tokens_across_dp, _ = (
4318
4319
4320
4321
4322
4323
4324
4325
4326
4327
4328
4329
4330
4331
4332
4333
4334
4335
            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,
4336
4337
            )
        )
4338
4339
4340

        if cudagraph_runtime_mode is None:
            cudagraph_runtime_mode = _cudagraph_mode
4341
        else:
4342
4343
4344
4345
4346
4347
4348
4349
4350
            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
        )
4351
        ubatch_slices, ubatch_slices_padded = maybe_create_ubatch_slices(
4352
4353
4354
4355
4356
4357
4358
4359
4360
4361
            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,
4362
        )
4363

4364
        attn_metadata: PerLayerAttnMetadata | None = None
4365
4366
4367

        # If force_attention is True, we always capture attention. Otherwise,
        # it only happens for cudagraph_runtime_mode=FULL.
4368
        if force_attention or cudagraph_runtime_mode == CUDAGraphMode.FULL:
4369
4370
4371
4372
4373
4374
            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:
4375
                seq_lens = max_query_len  # type: ignore[assignment]
4376
            self.seq_lens.np[:num_reqs] = seq_lens
4377
4378
            self.seq_lens.np[num_reqs:] = 0
            self.seq_lens.copy_to_gpu()
4379

4380
4381
            cum_num_tokens, _ = self._get_cumsum_and_arange(num_scheduled_tokens)
            self.query_start_loc.np[1 : num_reqs + 1] = cum_num_tokens
4382
4383
            self.query_start_loc.copy_to_gpu()

4384
            pad_attn = cudagraph_runtime_mode == CUDAGraphMode.FULL
4385
            attn_metadata, _ = self._build_attention_metadata(
4386
4387
4388
                num_tokens=num_tokens_unpadded,
                num_reqs=num_reqs_padded,
                max_query_len=max_query_len,
4389
                ubatch_slices=ubatch_slices_padded if pad_attn else ubatch_slices,
4390
                for_cudagraph_capture=is_graph_capturing,
4391
            )
4392

4393
        with self.maybe_dummy_run_with_lora(
4394
4395
4396
4397
4398
            self.lora_config,
            num_scheduled_tokens,
            num_sampled_tokens,
            activate_lora,
            remove_lora,
4399
        ):
4400
            # Make sure padding doesn't exceed max_num_tokens
4401
            assert num_tokens_padded <= self.max_num_tokens
4402
            model_kwargs = self._init_model_kwargs()
4403
            if self.supports_mm_inputs and not self.model_config.is_encoder_decoder:
Patrick von Platen's avatar
Patrick von Platen committed
4404
4405
                input_ids, inputs_embeds = self._prepare_mm_inputs(num_tokens_padded)

4406
                model_kwargs = {
4407
                    **model_kwargs,
4408
4409
                    **self._dummy_mm_kwargs(num_reqs),
                }
4410
4411
            elif self.enable_prompt_embeds:
                input_ids = None
4412
                inputs_embeds = self.inputs_embeds.gpu[:num_tokens_padded]
4413
                model_kwargs = self._init_model_kwargs()
4414
            else:
4415
                input_ids = self.input_ids.gpu[:num_tokens_padded]
4416
                inputs_embeds = None
4417

4418
            if self.uses_mrope:
4419
                positions = self.mrope_positions.gpu[:, :num_tokens_padded]
4420
            elif self.uses_xdrope_dim > 0:
4421
                positions = self.xdrope_positions.gpu[:, :num_tokens_padded]
4422
            else:
4423
                positions = self.positions.gpu[:num_tokens_padded]
4424
4425
4426
4427
4428
4429
4430
4431
4432

            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,
4433
4434
4435
                            device=self.device,
                        )
                    )
4436
4437

                intermediate_tensors = self.sync_and_slice_intermediate_tensors(
4438
                    num_tokens_padded, None, False
4439
                )
4440

4441
            if ubatch_slices_padded is not None:
4442
4443
4444
                # Adjust values to reflect a single ubatch.
                # TODO(sage,lucas): this is cruft that should be addressed in
                #  the padding refactor.
4445
                num_tokens_padded = ubatch_slices_padded[0].num_tokens
4446
                if num_tokens_across_dp is not None:
4447
                    num_tokens_across_dp[:] = num_tokens_padded
4448

4449
            with (
4450
                self.maybe_randomize_inputs(input_ids, inputs_embeds),
4451
                set_forward_context(
4452
4453
                    attn_metadata,
                    self.vllm_config,
4454
                    num_tokens=num_tokens_padded,
4455
4456
                    num_tokens_across_dp=num_tokens_across_dp,
                    cudagraph_runtime_mode=cudagraph_runtime_mode,
4457
                    batch_descriptor=batch_desc,
4458
                    ubatch_slices=ubatch_slices_padded,
4459
4460
                ),
            ):
4461
                outputs = self.model(
4462
4463
4464
4465
                    input_ids=input_ids,
                    positions=positions,
                    intermediate_tensors=intermediate_tensors,
                    inputs_embeds=inputs_embeds,
4466
                    **model_kwargs,
4467
                )
4468

4469
4470
4471
4472
            if self.use_aux_hidden_state_outputs:
                hidden_states, _ = outputs
            else:
                hidden_states = outputs
4473

4474
            if self.speculative_config and self.speculative_config.use_eagle():
4475
                assert isinstance(self.drafter, EagleProposer)
4476
4477
4478
                # 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.
4479
                use_cudagraphs = (
4480
4481
4482
4483
4484
4485
4486
4487
4488
                    (
                        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
4489
4490
4491
4492
4493
4494
4495
4496
4497
4498
4499

                # 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
4500
                    is_graph_capturing=is_graph_capturing,
4501
                )
4502

4503
4504
4505
4506
4507
4508
4509
4510
4511
4512
4513
        # 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()

4514
4515
4516
4517
4518
4519
4520
4521
4522
4523
        # 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)

4524
        logit_indices = np.cumsum(num_scheduled_tokens) - 1
4525
4526
4527
4528
        logit_indices_device = torch.from_numpy(logit_indices).to(
            self.device, non_blocking=True
        )
        return hidden_states, hidden_states[logit_indices_device]
4529
4530
4531
4532
4533
4534

    @torch.inference_mode()
    def _dummy_sampler_run(
        self,
        hidden_states: torch.Tensor,
    ) -> torch.Tensor:
4535
4536
4537
        # The dummy hidden states may contain special values,
        # like `inf` or `nan`.
        # To avoid breaking the sampler, we use a random tensor here instead.
4538
4539
4540
4541
4542

        if supports_mm_encoder_only(self.model):
            # MM Encoder only model no need to run sampler.
            return torch.tensor([])

4543
        hidden_states = torch.rand_like(hidden_states)
4544

4545
        logits = self.model.compute_logits(hidden_states)
4546
4547
        num_reqs = logits.size(0)

4548
        dummy_tensors = lambda v: torch.full((num_reqs,), v, device=self.device)
4549
4550
4551
4552
4553
4554
4555
4556
4557
4558
4559
4560
4561
4562
4563

        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)],
4564
            spec_token_ids=[[] for _ in range(num_reqs)],
4565
4566
            allowed_token_ids_mask=None,
            bad_words_token_ids={},
4567
            logitsprocs=LogitsProcessors(),
4568
        )
4569
        try:
4570
4571
4572
            sampler_output = self.sampler(
                logits=logits, sampling_metadata=dummy_metadata
            )
4573
        except RuntimeError as e:
4574
            if "out of memory" in str(e):
4575
4576
4577
4578
                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 "
4579
4580
                    "initializing the engine."
                ) from e
4581
4582
            else:
                raise e
4583
        if self.speculative_config:
4584
4585
            draft_token_ids = [[0] for _ in range(num_reqs)]
            dummy_spec_decode_metadata = SpecDecodeMetadata.make_dummy(
4586
4587
                draft_token_ids, self.device
            )
4588
4589
4590
4591
4592
4593

            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
4594
4595
4596
4597
4598
            logits = torch.randn(
                num_tokens + num_reqs,
                logits.shape[-1],
                device=self.device,
                dtype=logits.dtype,
4599
            )
4600
4601
4602
            self.rejection_sampler(
                dummy_spec_decode_metadata,
                draft_probs,
4603
                logits,
4604
4605
                dummy_metadata,
            )
4606
        return sampler_output
4607

4608
    def _dummy_pooler_run_task(
4609
4610
        self,
        hidden_states: torch.Tensor,
4611
4612
        task: PoolingTask,
    ) -> PoolerOutput:
4613
4614
4615
4616
        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
4617
4618
4619
4620
        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
4621
4622
4623

        req_num_tokens = num_tokens // num_reqs

4624
        dummy_prompt_lens = torch.from_numpy(num_scheduled_tokens_np)
4625
4626
4627
        dummy_token_ids = torch.zeros(
            (num_reqs, req_num_tokens), dtype=torch.int32, device=self.device
        )
4628

4629
        model = cast(VllmModelForPooling, self.get_model())
4630
        dummy_pooling_params = PoolingParams(task=task)
4631
        dummy_pooling_params.verify(task=task, model_config=self.model_config)
4632
        to_update = model.pooler.get_pooling_updates(task)
4633
4634
        to_update.apply(dummy_pooling_params)

4635
        dummy_metadata = PoolingMetadata(
4636
4637
4638
            prompt_lens=dummy_prompt_lens,
            prompt_token_ids=dummy_token_ids,
            pooling_params=[dummy_pooling_params] * num_reqs,
4639
            pooling_states=[PoolingStates() for i in range(num_reqs)],
4640
        )
4641

4642
        dummy_metadata.build_pooling_cursor(
4643
            num_scheduled_tokens_np,
4644
4645
            seq_lens_cpu=dummy_prompt_lens,
            device=hidden_states.device,
4646
        )
4647

4648
        try:
4649
4650
4651
            return model.pooler(
                hidden_states=hidden_states, pooling_metadata=dummy_metadata
            )
4652
        except RuntimeError as e:
4653
            if "out of memory" in str(e):
4654
                raise RuntimeError(
4655
4656
4657
                    "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 "
4658
4659
                    "initializing the engine."
                ) from e
4660
4661
            else:
                raise e
4662
4663
4664
4665
4666
4667

    @torch.inference_mode()
    def _dummy_pooler_run(
        self,
        hidden_states: torch.Tensor,
    ) -> PoolerOutput:
4668
4669
4670
4671
        if supports_mm_encoder_only(self.model):
            # MM Encoder only model not need to run pooler.
            return torch.tensor([])

4672
        # Find the task that has the largest output for subsequent steps
4673
4674
4675
        supported_pooling_tasks = self.get_supported_pooling_tasks()

        if not supported_pooling_tasks:
4676
4677
4678
4679
4680
4681
            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."
            )
4682

4683
        output_size = dict[PoolingTask, float]()
4684
        for task in supported_pooling_tasks:
4685
4686
            # Run a full batch with each task to ensure none of them OOMs
            output = self._dummy_pooler_run_task(hidden_states, task)
4687
            output_size[task] = sum(o.nbytes for o in output if o is not None)
4688
4689
4690
4691
            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)
4692

4693
    def profile_run(self) -> None:
4694
        # Profile with multimodal encoder & encoder cache.
4695
        if self.supports_mm_inputs:
4696
4697
            mm_config = self.model_config.multimodal_config
            if mm_config is not None and mm_config.skip_mm_profiling:
4698
                logger.info(
4699
                    "Skipping memory profiling for multimodal encoder and "
4700
4701
                    "encoder cache."
                )
4702
4703
4704
4705
4706
4707
4708
4709
            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.
4710
                    dummy_modality = mm_budget.get_modality_with_max_tokens()
4711
4712
4713
                    max_mm_items_per_batch = mm_budget.max_items_per_batch_by_modality[
                        dummy_modality
                    ]
4714
4715
4716
4717
4718
4719
4720
4721
4722

                    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,
                    )
4723

4724
4725
4726
4727
4728
                    # Create dummy batch of multimodal inputs.
                    batched_dummy_mm_inputs = self._get_mm_dummy_batch(
                        dummy_modality,
                        max_mm_items_per_batch,
                    )
4729

4730
                    # Run multimodal encoder.
4731
                    dummy_encoder_outputs = self.model.embed_multimodal(
4732
4733
                        **batched_dummy_mm_inputs
                    )
4734

4735
4736
4737
4738
                    sanity_check_mm_encoder_outputs(
                        dummy_encoder_outputs,
                        expected_num_items=max_mm_items_per_batch,
                    )
4739
4740
                    for i, output in enumerate(dummy_encoder_outputs):
                        self.encoder_cache[f"tmp_{i}"] = output
4741

4742
        # Add `is_profile` here to pre-allocate communication buffers
4743
4744
4745
        hidden_states, last_hidden_states = self._dummy_run(
            self.max_num_tokens, is_profile=True
        )
4746
        if get_pp_group().is_last_rank:
4747
4748
4749
4750
            if self.is_pooling_model:
                output = self._dummy_pooler_run(hidden_states)
            else:
                output = self._dummy_sampler_run(last_hidden_states)
4751
        else:
4752
            output = None
4753
        self._sync_device()
4754
        del hidden_states, output
4755
        self.encoder_cache.clear()
4756
        gc.collect()
4757

4758
    def capture_model(self) -> int:
4759
        if self.compilation_config.cudagraph_mode == CUDAGraphMode.NONE:
4760
            logger.warning(
4761
                "Skipping CUDA graph capture. To turn on CUDA graph capture, "
4762
4763
                "ensure `cudagraph_mode` was not manually set to `NONE`"
            )
4764
            return 0
4765

4766
4767
        compilation_counter.num_gpu_runner_capture_triggers += 1

4768
4769
        start_time = time.perf_counter()

4770
4771
4772
4773
4774
4775
4776
4777
4778
4779
4780
4781
4782
4783
        @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()
4784
                    gc.collect()
4785

4786
4787
4788
        # 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.
4789
        set_cudagraph_capturing_enabled(True)
4790
        with freeze_gc(), graph_capture(device=self.device):
4791
            start_free_gpu_memory = torch.cuda.mem_get_info()[0]
4792
            cudagraph_mode = self.compilation_config.cudagraph_mode
4793
            assert cudagraph_mode is not None
4794
4795
4796
4797
4798
4799
4800
4801
4802

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

4803
4804
            if cudagraph_mode.mixed_mode() != CUDAGraphMode.NONE:
                cudagraph_runtime_mode = cudagraph_mode.mixed_mode()
4805
                # make sure we capture the largest batch size first
4806
4807
4808
                compilation_cases = list(
                    product(reversed(self.cudagraph_batch_sizes), lora_cases)
                )
4809
4810
4811
                self._capture_cudagraphs(
                    compilation_cases,
                    cudagraph_runtime_mode=cudagraph_runtime_mode,
4812
4813
                    uniform_decode=False,
                )
4814

4815
4816
            # Capture full cudagraph for uniform decode batches if we
            # don't already have full mixed prefill-decode cudagraphs.
4817
4818
4819
4820
4821
4822
4823
            if (
                cudagraph_mode.decode_mode() == CUDAGraphMode.FULL
                and cudagraph_mode.separate_routine()
            ):
                max_num_tokens = (
                    self.scheduler_config.max_num_seqs * self.uniform_decode_query_len
                )
4824
                decode_cudagraph_batch_sizes = [
4825
4826
                    x
                    for x in self.cudagraph_batch_sizes
4827
                    if max_num_tokens >= x >= self.uniform_decode_query_len
4828
                ]
4829
4830
4831
                compilation_cases_decode = list(
                    product(reversed(decode_cudagraph_batch_sizes), lora_cases)
                )
4832
4833
4834
                self._capture_cudagraphs(
                    compilation_cases=compilation_cases_decode,
                    cudagraph_runtime_mode=CUDAGraphMode.FULL,
4835
4836
                    uniform_decode=True,
                )
4837

4838
4839
4840
            torch.cuda.synchronize()
            end_free_gpu_memory = torch.cuda.mem_get_info()[0]

4841
4842
4843
        # 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
4844
        # we may do lazy capturing in future that still allows capturing
4845
4846
        # after here.
        set_cudagraph_capturing_enabled(False)
4847

4848
4849
4850
4851
        # Lock workspace to prevent resizing during execution.
        # Max workspace sizes should have been captured during warmup/profiling.
        lock_workspace()

4852
4853
4854
4855
        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.
4856
        logger.info_once(
4857
4858
4859
            "Graph capturing finished in %.0f secs, took %.2f GiB",
            elapsed_time,
            cuda_graph_size / (1 << 30),
4860
            scope="local",
4861
        )
4862
        return cuda_graph_size
4863

4864
4865
    def _capture_cudagraphs(
        self,
4866
        compilation_cases: list[tuple[int, bool]],
4867
4868
4869
4870
4871
4872
4873
        cudagraph_runtime_mode: CUDAGraphMode,
        uniform_decode: bool,
    ):
        assert (
            cudagraph_runtime_mode != CUDAGraphMode.NONE
            and cudagraph_runtime_mode.valid_runtime_modes()
        ), f"Invalid cudagraph runtime mode: {cudagraph_runtime_mode}"
4874
4875
4876
4877
4878
4879
4880
4881

        # Only rank 0 should print progress bar during capture
        if is_global_first_rank():
            compilation_cases = tqdm(
                compilation_cases,
                disable=not self.load_config.use_tqdm_on_load,
                desc="Capturing CUDA graphs ({}, {})".format(
                    "decode" if uniform_decode else "mixed prefill-decode",
4882
4883
4884
                    cudagraph_runtime_mode.name,
                ),
            )
4885

4886
        # We skip EPLB here since we don't want to record dummy metrics
4887
        for num_tokens, activate_lora in compilation_cases:
4888
            # We currently only capture ubatched graphs when its a FULL
4889
4890
4891
            # 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
4892
            allow_microbatching = (
4893
                self.parallel_config.use_ubatching
4894
4895
                and cudagraph_runtime_mode == CUDAGraphMode.FULL
                and uniform_decode
4896
4897
4898
4899
4900
                and check_ubatch_thresholds(
                    config=self.vllm_config.parallel_config,
                    num_tokens=num_tokens,
                    uniform_decode=uniform_decode,
                )
4901
            )
4902

4903
4904
4905
4906
4907
4908
            for _ in range(self.compilation_config.cudagraph_num_of_warmups):
                # Use CUDAGraphRuntimeStyle.NONE (default) for warmup.
                # But be careful, warm up with `NONE`is orthogonal to
                # if we want to warm up attention or not. This is
                # different from the case where `FULL` implies capture
                # attention while `PIECEWISE` implies no attention.
4909
4910
4911
4912
4913
4914
4915
4916
4917
                force_attention = cudagraph_runtime_mode == CUDAGraphMode.FULL
                self._dummy_run(
                    num_tokens,
                    cudagraph_runtime_mode=CUDAGraphMode.NONE,
                    force_attention=force_attention,
                    uniform_decode=uniform_decode,
                    allow_microbatching=allow_microbatching,
                    skip_eplb=True,
                    remove_lora=False,
4918
                    activate_lora=activate_lora,
4919
4920
4921
4922
4923
4924
4925
4926
                )
            self._dummy_run(
                num_tokens,
                cudagraph_runtime_mode=cudagraph_runtime_mode,
                uniform_decode=uniform_decode,
                allow_microbatching=allow_microbatching,
                skip_eplb=True,
                remove_lora=False,
4927
                activate_lora=activate_lora,
Rémi Delacourt's avatar
Rémi Delacourt committed
4928
                is_graph_capturing=True,
4929
            )
4930
        self.maybe_remove_all_loras(self.lora_config)
4931

4932
4933
4934
4935
    def initialize_attn_backend(self, kv_cache_config: KVCacheConfig) -> None:
        """
        Initialize the attention backends and attention metadata builders.
        """
4936
        assert len(self.attn_groups) == 0, "Attention backends are already initialized"
4937

4938
4939
4940
4941
4942
4943
        class AttentionGroupKey(NamedTuple):
            attn_backend: type[AttentionBackend]
            kv_cache_spec: KVCacheSpec

        def get_attn_backends_for_group(
            kv_cache_group_spec: KVCacheGroupSpec,
4944
        ) -> tuple[dict[AttentionGroupKey, list[str]], set[type[AttentionBackend]]]:
4945
            layer_type = cast(type[Any], AttentionLayerBase)
4946
            layers = get_layers_from_vllm_config(
4947
                self.vllm_config, layer_type, kv_cache_group_spec.layer_names
4948
            )
4949
4950
            attn_backends = {}
            attn_backend_layers = defaultdict(list)
4951
            # Dedupe based on full class name; this is a bit safer than
4952
4953
4954
4955
            # 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.
4956
            for layer_name in kv_cache_group_spec.layer_names:
4957
                attn_backend = layers[layer_name].get_attn_backend()
4958
4959
4960
4961

                if layer_name in self.kv_sharing_fast_prefill_eligible_layers:
                    attn_backend = create_fast_prefill_custom_backend(
                        "FastPrefill",
4962
                        attn_backend,  # type: ignore[arg-type]
4963
4964
                    )

4965
4966
4967
                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):
4968
                    layer_kv_cache_spec = layer_kv_cache_spec.kv_cache_specs[layer_name]
4969
                key = (full_cls_name, layer_kv_cache_spec)
4970
4971
4972
                attn_backends[key] = AttentionGroupKey(
                    attn_backend, layer_kv_cache_spec
                )
4973
                attn_backend_layers[key].append(layer_name)
4974
4975
4976
4977
            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()),
            )
4978
4979

        def create_attn_groups(
4980
            attn_backends_map: dict[AttentionGroupKey, list[str]],
4981
            kv_cache_group_id: int,
4982
4983
        ) -> list[AttentionGroup]:
            attn_groups: list[AttentionGroup] = []
4984
            for (attn_backend, kv_cache_spec), layer_names in attn_backends_map.items():
4985
                attn_group = AttentionGroup(
4986
                    attn_backend,
4987
                    layer_names,
4988
                    kv_cache_spec,
4989
                    kv_cache_group_id,
4990
4991
                )

4992
4993
4994
                attn_groups.append(attn_group)
            return attn_groups

4995
        attention_backend_maps = []
4996
        attention_backend_list = []
4997
        for kv_cache_group_spec in kv_cache_config.kv_cache_groups:
4998
            attn_backends = get_attn_backends_for_group(kv_cache_group_spec)
4999
            attention_backend_maps.append(attn_backends[0])
5000
            attention_backend_list.append(attn_backends[1])
5001
5002

        # Resolve cudagraph_mode before actually initialize metadata_builders
5003
5004
5005
        self._check_and_update_cudagraph_mode(
            attention_backend_list, kv_cache_config.kv_cache_groups
        )
5006

5007
5008
5009
        # Check if attention backend supports PCP&DCP and related features.
        check_attention_cp_compatibility(self.vllm_config)

5010
5011
        for i, attn_backend_map in enumerate(attention_backend_maps):
            self.attn_groups.append(create_attn_groups(attn_backend_map, i))
5012

5013
5014
5015
5016
5017
5018
5019
5020
5021
5022
5023
5024
5025
5026
5027
    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
5028
5029
                    if not self.parallel_config.use_ubatching
                    else self.parallel_config.num_ubatches,
5030
                )
co63oc's avatar
co63oc committed
5031
        # Calculate reorder batch threshold (if needed)
5032
5033
        # Note (tdoublep): do this *after* constructing builders,
        # because some of them change the threshold at init time.
5034
5035
        self.calculate_reorder_batch_threshold()

5036
    def _check_and_update_cudagraph_mode(
5037
5038
5039
        self,
        attention_backends: list[set[type[AttentionBackend]]],
        kv_cache_groups: list[KVCacheGroupSpec],
5040
    ) -> None:
5041
        """
5042
        Resolve the cudagraph_mode when there are multiple attention
5043
        groups with potential conflicting CUDA graph support.
5044
5045
5046
        Then initialize the cudagraph_dispatcher based on the resolved
        cudagraph_mode.
        """
5047
        min_cg_support = AttentionCGSupport.ALWAYS
5048
        min_cg_backend_name = None
5049

5050
5051
5052
5053
5054
5055
5056
5057
5058
5059
5060
5061
        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__
5062
5063
        # Flexible resolve the cudagraph mode
        cudagraph_mode = self.compilation_config.cudagraph_mode
5064
        assert cudagraph_mode is not None
5065
        # check cudagraph for mixed batch is supported
5066
5067
5068
5069
5070
5071
        if (
            cudagraph_mode.mixed_mode() == CUDAGraphMode.FULL
            and min_cg_support != AttentionCGSupport.ALWAYS
        ):
            msg = (
                f"CUDAGraphMode.{cudagraph_mode.name} is not supported "
5072
                f"with {min_cg_backend_name} backend (support: "
5073
5074
                f"{min_cg_support})"
            )
5075
5076
            if min_cg_support == AttentionCGSupport.NEVER:
                # if not supported any full cudagraphs, just raise it.
5077
5078
                msg += (
                    "; please try cudagraph_mode=PIECEWISE, and "
5079
                    "make sure compilation mode is VLLM_COMPILE"
5080
                )
5081
5082
5083
5084
5085
                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"
5086
                cudagraph_mode = self.compilation_config.cudagraph_mode = (
5087
                    CUDAGraphMode.FULL_AND_PIECEWISE
5088
                )
5089
5090
            else:
                msg += "; setting cudagraph_mode=FULL_DECODE_ONLY"
5091
                cudagraph_mode = self.compilation_config.cudagraph_mode = (
5092
                    CUDAGraphMode.FULL_DECODE_ONLY
5093
                )
5094
5095
            logger.warning(msg)

5096
        # check that if we are doing decode full-cudagraphs it is supported
5097
5098
5099
5100
5101
5102
        if (
            cudagraph_mode.decode_mode() == CUDAGraphMode.FULL
            and min_cg_support == AttentionCGSupport.NEVER
        ):
            msg = (
                f"CUDAGraphMode.{cudagraph_mode.name} is not supported "
5103
                f"with {min_cg_backend_name} backend (support: "
5104
5105
                f"{min_cg_support})"
            )
5106
            if self.compilation_config.mode == CompilationMode.VLLM_COMPILE and (
5107
5108
5109
5110
5111
                self.compilation_config.splitting_ops_contain_attention()
                or self.compilation_config.use_inductor_graph_partition
            ):
                msg += (
                    "; setting cudagraph_mode=PIECEWISE because "
5112
                    "attention is compiled piecewise"
5113
5114
                )
                cudagraph_mode = self.compilation_config.cudagraph_mode = (
5115
                    CUDAGraphMode.PIECEWISE
5116
                )
5117
            else:
5118
5119
                msg += (
                    "; setting cudagraph_mode=NONE because "
5120
                    "attention is not compiled piecewise"
5121
5122
                )
                cudagraph_mode = self.compilation_config.cudagraph_mode = (
5123
                    CUDAGraphMode.NONE
5124
                )
5125
5126
            logger.warning(msg)

5127
5128
        # check that if we are doing spec-decode + decode full-cudagraphs it is
        # supported
5129
5130
5131
5132
5133
5134
5135
5136
        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 "
5137
                f"{min_cg_backend_name} (support: {min_cg_support})"
5138
            )
5139
5140
            if self.compilation_config.splitting_ops_contain_attention():
                msg += "; setting cudagraph_mode=PIECEWISE"
5141
                cudagraph_mode = self.compilation_config.cudagraph_mode = (
5142
                    CUDAGraphMode.PIECEWISE
5143
                )
5144
5145
            else:
                msg += "; setting cudagraph_mode=NONE"
5146
                cudagraph_mode = self.compilation_config.cudagraph_mode = (
5147
                    CUDAGraphMode.NONE
5148
                )
5149
5150
5151
5152
            logger.warning(msg)

        # double check that we can support full cudagraph if they are requested
        # even after automatic downgrades
5153
5154
5155
5156
5157
5158
        if (
            cudagraph_mode.has_full_cudagraphs()
            and min_cg_support == AttentionCGSupport.NEVER
        ):
            raise ValueError(
                f"CUDAGraphMode.{cudagraph_mode.name} is not "
5159
                f"supported with {min_cg_backend_name} backend ("
5160
5161
                f"support:{min_cg_support}) "
                "; please try cudagraph_mode=PIECEWISE, "
5162
                "and make sure compilation mode is VLLM_COMPILE"
5163
            )
5164

5165
5166
5167
5168
        # 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
5169
        # Will be removed in the near future when we have separate cudagraph capture
5170
5171
5172
5173
5174
5175
5176
5177
5178
        # 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
            )
5179
5180
5181
5182
            capture_sizes = self.compilation_config.cudagraph_capture_sizes
            self.cudagraph_batch_sizes = (
                capture_sizes if capture_sizes is not None else []
            )
5183

5184
5185
        # Trigger cudagraph dispatching keys initialization after
        # resolved cudagraph mode.
5186
        self.compilation_config.cudagraph_mode = cudagraph_mode
5187
        self.cudagraph_dispatcher.initialize_cudagraph_keys(
5188
            cudagraph_mode, self.uniform_decode_query_len
5189
        )
5190

5191
5192
    def calculate_reorder_batch_threshold(self) -> None:
        """
5193
5194
5195
5196
        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.
5197
        """
5198
5199
        min_none_high = lambda a, b: a if b is None else b if a is None else min(a, b)

5200
        reorder_batch_thresholds: list[int | None] = [
5201
5202
5203
            group.get_metadata_builder().reorder_batch_threshold
            for group in self._attn_group_iterator()
        ]
5204
5205
5206
5207
5208
        # 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
5209
        self.reorder_batch_threshold = reduce(min_none_high, reorder_batch_thresholds)  # type: ignore[assignment]
5210

5211
5212
5213
    @staticmethod
    def select_common_block_size(
        kv_manager_block_size: int, attn_groups: list[AttentionGroup]
5214
5215
    ) -> int:
        """
5216
5217
5218
5219
5220
        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.
5221
5222
5223
5224
5225
5226

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

        Returns:
5227
            The selected block size
5228
5229

        Raises:
5230
            ValueError: If no valid block size found
5231
5232
        """

5233
5234
5235
5236
5237
5238
5239
5240
        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
5241
                for supported_size in backend.get_supported_kernel_block_sizes():
5242
5243
5244
5245
5246
5247
5248
5249
5250
5251
5252
5253
5254
5255
5256
5257
5258
5259
5260
5261
5262
5263
5264
5265
5266
5267
5268
5269
5270
5271
                    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
5272
            for supported_size in backend.get_supported_kernel_block_sizes()
5273
5274
            if isinstance(supported_size, int)
        )
5275

5276
5277
5278
5279
5280
5281
        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}. ")
5282

5283
5284
5285
    def may_reinitialize_input_batch(
        self, kv_cache_config: KVCacheConfig, kernel_block_sizes: list[int]
    ) -> None:
5286
5287
5288
5289
5290
5291
5292
        """
        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.
5293
            kernel_block_sizes: The kernel block sizes for each KV cache group.
5294
5295
5296
5297
        """
        block_sizes = [
            kv_cache_group.kv_cache_spec.block_size
            for kv_cache_group in kv_cache_config.kv_cache_groups
5298
            if not isinstance(kv_cache_group.kv_cache_spec, EncoderOnlyAttentionSpec)
5299
        ]
5300
5301
5302
5303

        if block_sizes != [self.cache_config.block_size] or kernel_block_sizes != [
            self.cache_config.block_size
        ]:
5304
5305
5306
            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
5307
5308
                "for more details."
            )
5309
5310
            self.input_batch = InputBatch(
                max_num_reqs=self.max_num_reqs,
5311
                max_model_len=max(self.max_model_len, self.max_encoder_len),
5312
5313
5314
5315
5316
                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,
5317
                kernel_block_sizes=kernel_block_sizes,
5318
                is_spec_decode=bool(self.vllm_config.speculative_config),
5319
                logitsprocs=self.input_batch.logitsprocs,
5320
                logitsprocs_need_output_token_ids=self.input_batch.logitsprocs_need_output_token_ids,
5321
                is_pooling_model=self.is_pooling_model,
5322
                num_speculative_tokens=self.num_spec_tokens,
5323
5324
            )

5325
    def _allocate_kv_cache_tensors(
5326
5327
        self, kv_cache_config: KVCacheConfig
    ) -> dict[str, torch.Tensor]:
5328
        """
5329
5330
5331
        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.

5332
        Args:
5333
            kv_cache_config: The KV cache config
5334
        Returns:
5335
            dict[str, torch.Tensor]: A map between layer names to their
5336
            corresponding memory buffer for KV cache.
5337
        """
5338
5339
        kv_cache_raw_tensors: dict[str, torch.Tensor] = {}
        for kv_cache_tensor in kv_cache_config.kv_cache_tensors:
5340
5341
5342
            tensor = torch.zeros(
                kv_cache_tensor.size, dtype=torch.int8, device=self.device
            )
5343
5344
5345
5346
5347
            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:
5348
5349
5350
5351
            for layer_name in group.layer_names:
                if layer_name in self.runner_only_attn_layers:
                    continue
                layer_names.add(layer_name)
5352
5353
5354
        assert layer_names == set(kv_cache_raw_tensors.keys()), (
            "Some layers are not correctly initialized"
        )
5355
5356
        return kv_cache_raw_tensors

5357
5358
5359
    def _attn_group_iterator(self) -> Iterator[AttentionGroup]:
        return itertools.chain.from_iterable(self.attn_groups)

5360
    def _kv_cache_spec_attn_group_iterator(self) -> Iterator[AttentionGroup]:
5361
5362
        if not self.kv_cache_config.kv_cache_groups:
            return
5363
5364
        for attn_groups in self.attn_groups:
            yield from attn_groups
5365

5366
5367
5368
5369
5370
5371
5372
5373
5374
5375
5376
5377
5378
5379
5380
    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 = []
5381
        for kv_cache_gid, kv_cache_group in enumerate(kv_cache_config.kv_cache_groups):
5382
5383
5384
5385
5386
5387
            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):
5388
                continue
5389
            elif isinstance(kv_cache_spec, AttentionSpec):
5390
5391
5392
                # This is an attention backend that supports virtual
                # block splitting. Get the supported block sizes from
                # all backends in the group.
5393
                attn_groups = self.attn_groups[kv_cache_gid]
5394
                kv_manager_block_size = kv_cache_group.kv_cache_spec.block_size
5395
                selected_kernel_size = self.select_common_block_size(
5396
5397
5398
                    kv_manager_block_size, attn_groups
                )
                kernel_block_sizes.append(selected_kernel_size)
5399
            elif isinstance(kv_cache_spec, MambaSpec):
5400
5401
                # This is likely Mamba or other non-attention cache,
                # no splitting.
5402
                kernel_block_sizes.append(kv_cache_spec.block_size)
5403
5404
5405
5406
5407
5408
            else:
                raise NotImplementedError(
                    f"unknown kv cache spec {kv_cache_group.kv_cache_spec}"
                )
        return kernel_block_sizes

5409
5410
5411
5412
    def _reshape_kv_cache_tensors(
        self,
        kv_cache_config: KVCacheConfig,
        kv_cache_raw_tensors: dict[str, torch.Tensor],
5413
        kernel_block_sizes: list[int],
5414
    ) -> dict[str, torch.Tensor]:
5415
        """
5416
        Reshape the KV cache tensors to the desired shape and dtype.
5417

5418
        Args:
5419
5420
            kv_cache_config: The KV cache config
            kv_cache_raw_tensors: The KV cache buffer of each layer, with
5421
                correct size but uninitialized shape.
5422
            kernel_block_sizes: The kernel block sizes for each KV cache group.
5423
        Returns:
5424
            Dict[str, torch.Tensor]: A map between layer names to their
5425
5426
            corresponding memory buffer for KV cache.
        """
5427
        kv_caches: dict[str, torch.Tensor] = {}
5428
        has_attn, has_mamba = False, False
5429
5430
        for group in self._kv_cache_spec_attn_group_iterator():
            kv_cache_spec = group.kv_cache_spec
5431
            attn_backend = group.backend
5432
5433
5434
5435
            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]
5436
            for layer_name in group.layer_names:
5437
5438
                if layer_name in self.runner_only_attn_layers:
                    continue
5439
5440
                raw_tensor = kv_cache_raw_tensors[layer_name]
                assert raw_tensor.numel() % kv_cache_spec.page_size_bytes == 0
5441
                num_blocks = raw_tensor.numel() // kv_cache_spec.page_size_bytes
5442
                if isinstance(kv_cache_spec, AttentionSpec):
5443
                    has_attn = True
5444
5445
                    num_blocks_per_kv_block = (
                        kv_cache_spec.block_size // kernel_block_size
5446
5447
5448
                    )
                    kernel_num_blocks = num_blocks * num_blocks_per_kv_block

5449
                    kv_cache_shape = attn_backend.get_kv_cache_shape(
5450
                        kernel_num_blocks,
5451
                        kernel_block_size,
5452
5453
                        kv_cache_spec.num_kv_heads,
                        kv_cache_spec.head_size,
5454
5455
                        cache_dtype_str=self.cache_config.cache_dtype,
                    )
5456
                    dtype = kv_cache_spec.dtype
5457
                    try:
5458
                        kv_cache_stride_order = attn_backend.get_kv_cache_stride_order()
5459
                        assert len(kv_cache_stride_order) == len(kv_cache_shape)
5460
                    except (AttributeError, NotImplementedError):
5461
                        kv_cache_stride_order = tuple(range(len(kv_cache_shape)))
5462
5463
5464
5465
5466
                    # 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.
5467
5468
5469
                    kv_cache_shape = tuple(
                        kv_cache_shape[i] for i in kv_cache_stride_order
                    )
5470
5471
5472
5473
5474
                    # Maintain original KV shape view.
                    inv_order = [
                        kv_cache_stride_order.index(i)
                        for i in range(len(kv_cache_stride_order))
                    ]
5475
5476
5477
5478
5479
5480
                    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
5481
                elif isinstance(kv_cache_spec, MambaSpec):
5482
                    has_mamba = True
Chen Zhang's avatar
Chen Zhang committed
5483
5484
                    raw_tensor = kv_cache_raw_tensors[layer_name]
                    state_tensors = []
5485
                    storage_offset_bytes = 0
5486
                    for shape, dtype in zip(kv_cache_spec.shapes, kv_cache_spec.dtypes):
5487
5488
                        dtype_size = get_dtype_size(dtype)
                        num_element_per_page = (
5489
5490
                            kv_cache_spec.page_size_bytes // dtype_size
                        )
Chen Zhang's avatar
Chen Zhang committed
5491
                        target_shape = (num_blocks, *shape)
5492
5493
                        stride = torch.empty(target_shape).stride()
                        target_stride = (num_element_per_page, *stride[1:])
5494
                        assert storage_offset_bytes % dtype_size == 0
5495
5496
5497
5498
                        tensor = torch.as_strided(
                            raw_tensor.view(dtype),
                            size=target_shape,
                            stride=target_stride,
5499
                            storage_offset=storage_offset_bytes // dtype_size,
5500
                        )
Chen Zhang's avatar
Chen Zhang committed
5501
                        state_tensors.append(tensor)
5502
                        storage_offset_bytes += stride[0] * dtype_size
5503
5504

                    kv_caches[layer_name] = state_tensors
5505
                else:
5506
                    raise NotImplementedError
5507
5508

        if has_attn and has_mamba:
5509
            self._update_hybrid_attention_mamba_layout(kv_caches)
5510

5511
5512
        return kv_caches

5513
    def _update_hybrid_attention_mamba_layout(
5514
5515
        self, kv_caches: dict[str, torch.Tensor]
    ) -> None:
5516
        """
5517
5518
        Update the layout of attention layers from (2, num_blocks, ...) to
        (num_blocks, 2, ...).
5519
5520

        Args:
5521
            kv_caches: The KV cache buffer of each layer.
5522
5523
        """

5524
5525
        for group in self._kv_cache_spec_attn_group_iterator():
            kv_cache_spec = group.kv_cache_spec
5526
            for layer_name in group.layer_names:
5527
                kv_cache = kv_caches[layer_name]
5528
5529
5530
5531
                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 "
5532
                        f"a tensor of shape {kv_cache.shape}"
5533
                    )
5534
                    hidden_size = kv_cache.shape[2:].numel()
5535
5536
5537
5538
                    kv_cache.as_strided_(
                        size=kv_cache.shape,
                        stride=(hidden_size, 2 * hidden_size, *kv_cache.stride()[2:]),
                    )
5539

5540
    def initialize_kv_cache_tensors(
5541
        self, kv_cache_config: KVCacheConfig, kernel_block_sizes: list[int]
5542
    ) -> dict[str, torch.Tensor]:
5543
5544
5545
5546
5547
        """
        Initialize the memory buffer for KV cache.

        Args:
            kv_cache_config: The KV cache config
5548
5549
            kernel_block_sizes: The kernel block sizes for each KV cache group.

5550
        Returns:
5551
            Dict[str, torch.Tensor]: A map between layer names to their
5552
5553
            corresponding memory buffer for KV cache.
        """
5554
5555
5556
5557
5558
5559
5560
5561
5562
5563
5564
5565
5566
5567
5568
5569
5570
5571
5572
5573
5574
5575
5576
5577

        # 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
            )
5578

5579
        # Set up cross-layer KV cache sharing
5580
5581
        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)
5582
5583
            kv_caches[layer_name] = kv_caches[target_layer_name]

5584
5585
5586
5587
5588
5589
5590
5591
5592
        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,
        )
5593
5594
5595
        return kv_caches

    def maybe_add_kv_sharing_layers_to_kv_cache_groups(
5596
5597
        self, kv_cache_config: KVCacheConfig
    ) -> None:
5598
5599
5600
5601
5602
5603
5604
5605
5606
5607
5608
5609
5610
5611
5612
5613
5614
5615
        """
        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.
5616
            attn_layers = get_layers_from_vllm_config(self.vllm_config, Attention)
5617
5618
            for layer_name in reversed(attn_layers):
                if layer_name in self.shared_kv_cache_layers:
5619
                    self.kv_sharing_fast_prefill_eligible_layers.add(layer_name)
5620
5621
                else:
                    break
5622

5623
5624
5625
5626
5627
5628
5629
    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
        """
5630
        kv_cache_config = deepcopy(kv_cache_config)
5631
        self.kv_cache_config = kv_cache_config
5632
        self.may_add_encoder_only_layers_to_kv_cache_config()
5633
        self.maybe_add_kv_sharing_layers_to_kv_cache_groups(kv_cache_config)
5634
        self.initialize_attn_backend(kv_cache_config)
5635
5636
5637
5638
5639
5640
        # 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)
5641
5642
5643
5644

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

5645
        # Reinitialize need to after initialize_attn_backend
5646
5647
5648
5649
        self.may_reinitialize_input_batch(kv_cache_config, kernel_block_sizes)
        kv_caches = self.initialize_kv_cache_tensors(
            kv_cache_config, kernel_block_sizes
        )
5650

5651
5652
5653
5654
5655
5656
        if self.speculative_config and self.speculative_config.use_eagle():
            assert isinstance(self.drafter, EagleProposer)
            # validate all draft model layers belong to the same kv cache
            # group
            self.drafter.validate_same_kv_cache_group(kv_cache_config)

Robert Shaw's avatar
Robert Shaw committed
5657
        if has_kv_transfer_group():
5658
            kv_transfer_group = get_kv_transfer_group()
5659
5660
5661
5662
5663
5664
5665
            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)
5666
            kv_transfer_group.set_host_xfer_buffer_ops(copy_kv_blocks)
Robert Shaw's avatar
Robert Shaw committed
5667

5668
5669
5670
5671
5672
5673
5674
5675
5676
5677
5678
5679
5680
5681
5682
5683
5684
5685
5686
5687
5688
5689
        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,
            model_config=self.model_config,
            instance_id=self.vllm_config.instance_id,
        )

5690
5691
5692
5693
5694
    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
5695
        encoder_only_attn_specs: dict[AttentionSpec, list[str]] = defaultdict(list)
5696
5697
5698
        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:
5699
                attn_spec: AttentionSpec = EncoderOnlyAttentionSpec(
5700
5701
5702
                    block_size=block_size,
                    num_kv_heads=attn_module.num_kv_heads,
                    head_size=attn_module.head_size,
5703
5704
                    dtype=self.kv_cache_dtype,
                )
5705
5706
5707
                encoder_only_attn_specs[attn_spec].append(layer_name)
                self.runner_only_attn_layers.add(layer_name)
        if len(encoder_only_attn_specs) > 0:
5708
5709
5710
            assert len(encoder_only_attn_specs) == 1, (
                "Only support one encoder-only attention spec now"
            )
5711
5712
            spec, layer_names = encoder_only_attn_specs.popitem()
            self.kv_cache_config.kv_cache_groups.append(
5713
5714
                KVCacheGroupSpec(layer_names=layer_names, kv_cache_spec=spec)
            )
5715

5716
    def get_kv_cache_spec(self) -> dict[str, KVCacheSpec]:
5717
        """
5718
        Generates the KVCacheSpec by parsing the kv cache format from each
5719
5720
        Attention module in the static forward context.
        Returns:
5721
            KVCacheSpec: A dictionary mapping layer names to their KV cache
5722
5723
            format. Layers that do not need KV cache are not included.
        """
5724
5725
        if has_ec_transfer() and get_ec_transfer().is_producer:
            return {}
5726
        kv_cache_spec: dict[str, KVCacheSpec] = {}
5727
5728
        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
5729
        for layer_name, attn_module in attn_layers.items():
5730
5731
5732
5733
5734
5735
5736
5737
5738
5739
5740
5741
5742
5743
5744
            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
5745

5746
        return kv_cache_spec
5747

5748
    def _to_list(self, sampled_token_ids: torch.Tensor) -> list[list[int]]:
5749
5750
5751
5752
5753
5754
5755
5756
        # 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.
5757
        pinned = self.sampled_token_ids_pinned_cpu[: sampled_token_ids.shape[0]]
5758
5759
5760
        pinned.copy_(sampled_token_ids, non_blocking=True)
        self.transfer_event.record()
        self.transfer_event.synchronize()
5761
        return pinned.tolist()