gpu_model_runner.py 265 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 threading
8
import time
9
from collections import defaultdict
10
from collections.abc import Iterator, Sequence
11
from contextlib import contextmanager
12
from copy import copy, deepcopy
13
from dataclasses import dataclass
14
from functools import reduce
王敏's avatar
王敏 committed
15
from typing import TYPE_CHECKING, Any, NamedTuple, TypeAlias, cast, Optional
16
17
18
19
20

import numpy as np
import torch
import torch.distributed
import torch.nn as nn
21
from tqdm import tqdm
22

23
import vllm.envs as envs
24
from vllm.attention.layer import Attention, MLAAttention
25
from vllm.compilation.counter import compilation_counter
26
from vllm.compilation.cuda_graph import CUDAGraphStat, CUDAGraphWrapper
27
from vllm.compilation.monitor import set_cudagraph_capturing_enabled
28
from vllm.config import (
29
    CompilationMode,
30
31
32
33
34
    CUDAGraphMode,
    VllmConfig,
    get_layers_from_vllm_config,
    update_config,
)
35
from vllm.distributed.ec_transfer import get_ec_transfer, has_ec_transfer
36
from vllm.distributed.eplb.eplb_state import EplbState
37
from vllm.distributed.kv_transfer import get_kv_transfer_group, has_kv_transfer_group
38
from vllm.distributed.kv_transfer.kv_connector.utils import copy_kv_blocks
39
from vllm.distributed.parallel_state import (
40
    get_dcp_group,
41
42
43
44
45
46
    get_pp_group,
    get_tp_group,
    graph_capture,
    is_global_first_rank,
    prepare_communication_buffer_for_model,
)
47
48
49
50
from vllm.forward_context import (
    BatchDescriptor,
    set_forward_context,
)
51
from vllm.logger import init_logger
52
from vllm.lora.layers import LoRAMapping, LoRAMappingType
53
from vllm.model_executor.layers.attention_layer_base import AttentionLayerBase
54
55
56
from vllm.model_executor.layers.fused_moe.routed_experts_capturer import (
    RoutedExpertsCapturer,
)
57
58
59
60
from vllm.model_executor.layers.rotary_embedding import (
    MRotaryEmbedding,
    XDRotaryEmbedding,
)
61
from vllm.model_executor.model_loader import TensorizerLoader, get_model_loader
62
from vllm.model_executor.models.interfaces import (
63
    MultiModalEmbeddings,
64
    SupportsMRoPE,
65
    SupportsMultiModal,
66
    SupportsXDRoPE,
67
68
69
70
71
    is_mixture_of_experts,
    supports_eagle3,
    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
from vllm.utils.torch_utils import (
    get_dtype_size,
    kv_cache_dtype_str_to_dtype,
)
100
101
from vllm.v1.attention.backend import (
    AttentionBackend,
102
    AttentionCGSupport,
103
    AttentionMetadata,
104
    AttentionMetadataBuilder,
105
    AttentionType,
106
    CommonAttentionMetadata,
107
108
    MultipleOf,
)
109
from vllm.v1.attention.backends.gdn_attn import GDNAttentionMetadataBuilder
110
from vllm.v1.attention.backends.utils import (
111
    create_fast_prefill_custom_backend,
112
    get_dcp_local_seq_lens,
113
114
    reorder_batch_to_split_decodes_and_prefills,
)
115
from vllm.v1.core.sched.output import NewRequestData
116
from vllm.v1.cudagraph_dispatcher import CudagraphDispatcher
117
118
119
120
121
122
123
124
125
126
127
128
129
130
131
132
133
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,
134
    ECConnectorOutput,
135
    KVConnectorOutput,
136
137
138
139
140
    LogprobsLists,
    LogprobsTensors,
    ModelRunnerOutput,
    PoolerOutput,
    SamplerOutput,
141
    make_empty_encoder_model_runner_output,
142
)
143
from vllm.v1.pool.metadata import PoolingMetadata, PoolingStates
144
from vllm.v1.sample.logits_processor import LogitsProcessors, build_logitsprocs
145
from vllm.v1.sample.logits_processor.interface import LogitsProcessor
146
from vllm.v1.sample.metadata import SamplingMetadata
147
from vllm.v1.sample.rejection_sampler import RejectionSampler
王敏's avatar
王敏 committed
148
from vllm.v1.sample.rejection_sampler_opt import OptRejectionSampler
149
from vllm.v1.sample.sampler import Sampler
150
from vllm.v1.spec_decode.draft_model import DraftModelProposer
151
from vllm.v1.spec_decode.eagle import EagleProposer
152
from vllm.v1.spec_decode.medusa import MedusaProposer
153
from vllm.v1.spec_decode.metadata import SpecDecodeMetadata
154
from vllm.v1.spec_decode.ngram_proposer import NgramProposer
155
from vllm.v1.spec_decode.suffix_decoding import SuffixDecodingProposer
156
from vllm.v1.structured_output.utils import apply_grammar_bitmask
157
from vllm.v1.utils import CpuGpuBuffer, record_function_or_nullcontext
158
159
160
161
162
from vllm.v1.worker import mamba_utils
from vllm.v1.worker.cp_utils import (
    check_attention_cp_compatibility,
    get_total_cp_world_size,
)
163
from vllm.v1.worker.dp_utils import coordinate_batch_across_dp
164
from vllm.v1.worker.ec_connector_model_runner_mixin import ECConnectorModelRunnerMixin
165
from vllm.v1.worker.gpu_input_batch import CachedRequestState, InputBatch
166
from vllm.v1.worker.gpu_ubatch_wrapper import UBatchWrapper
167
from vllm.v1.worker.kv_connector_model_runner_mixin import KVConnectorModelRunnerMixin
168
from vllm.v1.worker.lora_model_runner_mixin import LoRAModelRunnerMixin
169
170
171
from vllm.v1.worker.ubatch_utils import (
    UBatchSlices,
    check_ubatch_thresholds,
172
    maybe_create_ubatch_slices,
173
    split_attn_metadata,
174
)
175
from vllm.v1.worker.utils import is_residual_scattered_for_sp
176
from vllm.v1.worker.workspace import lock_workspace
177

178
179
180
181
182
183
184
from .utils import (
    AttentionGroup,
    MultiModalBudget,
    add_kv_sharing_layers_to_kv_cache_groups,
    bind_kv_cache,
    sanity_check_mm_encoder_outputs,
)
王敏's avatar
王敏 committed
185
from vllm.v1.spec_decode.utils import DraftProbs
186

187
if TYPE_CHECKING:
188
    from vllm.model_executor.model_loader.tensorizer import TensorizerConfig
189
    from vllm.v1.core.sched.output import GrammarOutput, SchedulerOutput
190
191
192

logger = init_logger(__name__)

193
194
AttnMetadataDict: TypeAlias = dict[str, AttentionMetadata]
# list when ubatching is enabled
195
PerLayerAttnMetadata: TypeAlias = list[AttnMetadataDict] | AttnMetadataDict
196

197

198
199
200
201
202
203
# Wrapper for ModelRunnerOutput to support overlapped execution.
class AsyncGPUModelRunnerOutput(AsyncModelRunnerOutput):
    def __init__(
        self,
        model_runner_output: ModelRunnerOutput,
        sampled_token_ids: torch.Tensor,
204
        logprobs_tensors: LogprobsTensors | None,
205
206
        invalid_req_indices: list[int],
        async_output_copy_stream: torch.cuda.Stream,
207
        vocab_size: int,
208
209
210
211
212
    ):
        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.
213
        self.async_copy_ready_event = torch.Event()
214
215
216
217

        # 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
218
        self.vocab_size = vocab_size
219
        self._logprobs_tensors = logprobs_tensors
220
221
222
223
224

        # 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)
225
            self.sampled_token_ids_cpu = self._sampled_token_ids.to(
226
227
                "cpu", non_blocking=True
            )
228
229
230
231
232
            self._logprobs_tensors_cpu = (
                self._logprobs_tensors.to_cpu_nonblocking()
                if self._logprobs_tensors
                else None
            )
233
            self.async_copy_ready_event.record()
234
235
236

    def get_output(self) -> ModelRunnerOutput:
        """Copy the device tensors to the host and return a ModelRunnerOutput.
237

238
239
        This function blocks until the copy is finished.
        """
240
        max_gen_len = self.sampled_token_ids_cpu.shape[-1]
241
        self.async_copy_ready_event.synchronize()
242

243
244
        # Release the device tensors once the copy has completed.
        del self._logprobs_tensors
245
        del self._sampled_token_ids
246
        if max_gen_len == 1:
247
            valid_sampled_token_ids = self.sampled_token_ids_cpu.tolist()
248
249
            for i in self._invalid_req_indices:
                valid_sampled_token_ids[i].clear()
250
251
252
            logprobs_lists = None
            if self._logprobs_tensors_cpu is not None:
                logprobs_lists = self._logprobs_tensors_cpu.tolists()
253
        else:
254
            valid_sampled_token_ids, logprobs_lists = RejectionSampler.parse_output(
255
256
                self.sampled_token_ids_cpu,
                self.vocab_size,
257
                self._invalid_req_indices,
258
                logprobs_tensors=self._logprobs_tensors_cpu,
259
            )
260
261
262

        output = self._model_runner_output
        output.sampled_token_ids = valid_sampled_token_ids
263
        output.logprobs = logprobs_lists
264
265
266
        return output


267
268
269
270
271
272
273
274
275
276
277
278
279
280
281
282
283
284
285
286
287
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)
288
            raw_pooler_output_cpu = json_map_leaves(
289
290
291
292
                lambda x: None if x is None else x.to("cpu", non_blocking=True),
                self._raw_pooler_output,
            )
            self.async_copy_ready_event.record()
293
294
295
296
            self._model_runner_output.pooler_output = [
                out if include else None
                for out, include in zip(raw_pooler_output_cpu, finished_mask)
            ]
297
298
299
300
301
302
303
304
305
306
307
308

    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


309
310
311
class ExecuteModelState(NamedTuple):
    """Ephemeral cached state transferred between execute_model() and
    sample_tokens(), after execute_model() returns None."""
312

313
314
315
316
317
318
319
    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
320
    ec_connector_output: ECConnectorOutput | None
321
    cudagraph_stats: CUDAGraphStat | None
322
    slot_mappings: dict[str, torch.Tensor] | list[dict[str, torch.Tensor]] | None
323
324


325
326
327
class GPUModelRunner(
    LoRAModelRunnerMixin, KVConnectorModelRunnerMixin, ECConnectorModelRunnerMixin
):
328
329
    def __init__(
        self,
330
        vllm_config: VllmConfig,
331
        device: torch.device,
332
    ):
333
334
335
        self.vllm_config = vllm_config
        self.model_config = vllm_config.model_config
        self.cache_config = vllm_config.cache_config
336
        self.compilation_config = vllm_config.compilation_config
337
338
339
340
341
342
        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
343

344
        from vllm.model_executor.models.utils import set_cpu_offload_max_bytes
345
346

        set_cpu_offload_max_bytes(int(self.cache_config.cpu_offload_gb * 1024**3))
347

348
349
350
351
        model_config = self.model_config
        cache_config = self.cache_config
        scheduler_config = self.scheduler_config
        parallel_config = self.parallel_config
352
        self.device = device
353
354
        self.pin_memory = is_pin_memory_available()
        self.dtype = self.model_config.dtype
355

356
357
358
        self.kv_cache_dtype = kv_cache_dtype_str_to_dtype(
            cache_config.cache_dtype, self.model_config
        )
359

360
        self.is_pooling_model = model_config.runner_type == "pooling"
361
        self.enable_prompt_embeds = model_config.enable_prompt_embeds
362
        self.is_multimodal_raw_input_only_model = (
363
364
            model_config.is_multimodal_raw_input_only_model
        )
365
366
        # This will be overridden in load_model()
        self.is_multimodal_pruning_enabled = False
367
        self.max_model_len = model_config.max_model_len
368
369
370

        # Always set to false after the first forward pass
        self.calculate_kv_scales = self.cache_config.calculate_kv_scales
371
        self.dcp_world_size = self.parallel_config.decode_context_parallel_size
372
        self.dcp_rank = 0 if self.dcp_world_size <= 1 else get_dcp_group().rank_in_group
373
        self.max_num_tokens = scheduler_config.max_num_batched_tokens
374
        self.max_num_reqs = scheduler_config.max_num_seqs
375

376
377
378
379
380
        # 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 = (
381
            self.parallel_config.distributed_executor_backend == "external_launcher"
382
            and len(get_pp_group().ranks) > 1
383
        )
384

385
        # Model-related.
386
        self.num_query_heads = model_config.get_num_attention_heads(parallel_config)
387
        self.inputs_embeds_size = model_config.get_inputs_embeds_size()
388
        self.attention_chunk_size = model_config.attention_chunk_size
389
        # Only relevant for models using ALiBi (e.g, MPT)
390
        self.use_alibi = model_config.uses_alibi
391

392
        self.cascade_attn_enabled = not self.model_config.disable_cascade_attn
393
        self.is_mm_prefix_lm = self.model_config.is_mm_prefix_lm
394

395
        # Multi-modal data support
396
        self.mm_registry = MULTIMODAL_REGISTRY
397
        self.uses_mrope = model_config.uses_mrope
398
        self.uses_xdrope_dim = model_config.uses_xdrope_dim
399
        self.supports_mm_inputs = self.mm_registry.supports_multimodal_inputs(
400
401
            model_config
        )
402

403
404
405
        if self.model_config.is_encoder_decoder:
            # Maximum length of the encoder input, only for encoder-decoder
            # models.
406
            self.max_encoder_len = scheduler_config.max_num_encoder_input_tokens
407
408
409
        else:
            self.max_encoder_len = 0

410
411
412
        # Async scheduling
        self.use_async_scheduling = self.scheduler_config.async_scheduling

413
        # Sampler
414
        self.sampler = Sampler(logprobs_mode=self.model_config.logprobs_mode)
415

416
        self.eplb_state: EplbState | None = None
417
418
419
420
421
422
        """
        State of the expert parallelism load balancer.

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

423
        # Lazy initializations
424
        # self.model: nn.Module  # Set after load_model
425
        # Initialize in initialize_kv_cache
426
        self.kv_caches: list[torch.Tensor] = []
427
428
429
        # Initialize in initialize_kv_cache_tensors
        self.cross_layers_kv_cache: torch.Tensor | None = None
        self.cross_layers_attn_backend: type[AttentionBackend] | None = None
430
431
        # indexes: [kv_cache_group_id][attn_group]
        self.attn_groups: list[list[AttentionGroup]] = []
432
433
        # self.kv_cache_config: KVCacheConfig

434
435
        # mm_hash ->  encoder_output
        self.encoder_cache: dict[str, torch.Tensor] = {}
436

437
        self.use_aux_hidden_state_outputs = False
438
439
440
441
442
        # 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:
443
            self.drafter: (
444
445
446
447
448
                NgramProposer
                | SuffixDecodingProposer
                | EagleProposer
                | DraftModelProposer
                | MedusaProposer
449
            )
450
451
            if self.speculative_config.method == "ngram":
                self.drafter = NgramProposer(self.vllm_config)
452
453
454
455
456
457
            elif self.speculative_config.uses_draft_model():
                self.drafter = DraftModelProposer(
                    vllm_config=self.vllm_config,
                    device=self.device,
                    runner=self,
                )
458
459
            elif self.speculative_config.method == "suffix":
                self.drafter = SuffixDecodingProposer(self.vllm_config)
460
            elif self.speculative_config.use_eagle():
461
                self.drafter = EagleProposer(self.vllm_config, self.device, self)
462
                if self.speculative_config.method == "eagle3":
463
464
465
                    self.use_aux_hidden_state_outputs = (
                        self.drafter.eagle3_use_aux_hidden_state
                    )
466
467
            elif self.speculative_config.method == "medusa":
                self.drafter = MedusaProposer(
468
                    vllm_config=self.vllm_config, device=self.device
469
                )
470
            else:
471
472
473
474
                raise ValueError(
                    "Unknown speculative decoding method: "
                    f"{self.speculative_config.method}"
                )
王敏's avatar
王敏 committed
475
476
477
478
479
            
            if not envs.VLLM_REJECT_SAMPLE_OPT:
                self.rejection_sampler = RejectionSampler(self.sampler)
            else:
                self.rejection_sampler = OptRejectionSampler(self.sampler)
480

481
482
483
        self.num_spec_tokens = 0
        if self.speculative_config:
            self.num_spec_tokens = self.speculative_config.num_speculative_tokens
484
485
486
487
488
            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
489

490
        # Request states.
491
        self.requests: dict[str, CachedRequestState] = {}
492
493
494
        # NOTE(rob): num_prompt_logprobs only includes reqs
        # that are currently in the prefill phase.
        self.num_prompt_logprobs: dict[str, int] = {}
495
        self.comm_stream = torch.cuda.Stream()
496

497
498
499
500
501
502
503
504
505
        # 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.
506
507
508
509
        logits_processors = model_config.logits_processors
        custom_logitsprocs: Sequence[str | type[LogitsProcessor]] = (
            tuple(logits_processors) if logits_processors is not None else ()
        )
510
511
        self.input_batch = InputBatch(
            max_num_reqs=self.max_num_reqs,
512
513
514
            # 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),
515
516
517
            max_num_batched_tokens=self.max_num_tokens,
            device=self.device,
            pin_memory=self.pin_memory,
518
            vocab_size=self.model_config.get_vocab_size(),
519
            block_sizes=[self.cache_config.block_size],
520
            kernel_block_sizes=[self.cache_config.block_size],
521
            is_spec_decode=bool(self.vllm_config.speculative_config),
522
            logitsprocs=build_logitsprocs(
523
524
525
                self.vllm_config,
                self.device,
                self.pin_memory,
526
                self.is_pooling_model,
527
                custom_logitsprocs,
528
            ),
529
530
531
            # 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),
532
            is_pooling_model=self.is_pooling_model,
533
            cp_kv_cache_interleave_size=self.parallel_config.cp_kv_cache_interleave_size,
534
        )
535

536
537
538
539
540
        # 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.
541
        self.prepare_inputs_event: torch.Event | None = None
542
543
        if self.use_async_scheduling:
            self.async_output_copy_stream = torch.cuda.Stream()
544
            self.prepare_inputs_event = torch.Event()
545

546
        # self.cudagraph_batch_sizes sorts in ascending order.
547
548
549
550
        if (
            self.compilation_config.cudagraph_capture_sizes
            and self.compilation_config.cudagraph_mode != CUDAGraphMode.NONE
        ):
551
552
            self.cudagraph_batch_sizes = sorted(
                self.compilation_config.cudagraph_capture_sizes
553
            )
554

555
        # Cache the device properties.
556
        self._init_device_properties()
557

558
559
560
561
        # Encoder timing registry for observability
        self.encoder_timing_registry: dict[str, EncoderTimingStats] = {}
        self._encoder_timing_lock = threading.Lock()

562
        # Persistent buffers for CUDA graphs.
563
564
565
566
567
        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
        )
568
        self.seq_lens = self._make_buffer(self.max_num_reqs, dtype=torch.int32)
569
        self.encoder_seq_lens = self._make_buffer(self.max_num_reqs, dtype=torch.int32)
570
571
572
573
        if self.dcp_world_size > 1:
            self.dcp_local_seq_lens = self._make_buffer(
                self.max_num_reqs, dtype=torch.int32
            )
574
575
576
        # 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.
577
        self.inputs_embeds = self._make_buffer(
578
            self.max_num_tokens, self.inputs_embeds_size, dtype=self.dtype, numpy=False
579
580
        )
        self.is_token_ids = self._make_buffer(self.max_num_tokens, dtype=torch.bool)
581
582
        self.discard_request_mask = self._make_buffer(
            self.max_num_reqs, dtype=torch.bool
583
584
585
586
587
588
589
        )
        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
        )
590

591
592
        # Only relevant for multimodal models
        if self.supports_mm_inputs:
593
594
595
596
597
598
599
            # 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
600
601

        # Only relevant for models using M-RoPE (e.g, Qwen2-VL)
602
        if self.uses_mrope:
Roger Wang's avatar
Roger Wang committed
603
604
605
606
            # 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
607
608
609
610
611
612

            # 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
613
            self.mrope_positions = self._make_buffer(
614
615
                (3, self.max_num_tokens + 1), dtype=torch.int64
            )
616

617
618
619
620
621
622
        # 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
            )
623

624
        # None in the first PP rank. The rest are set after load_model.
625
        self.intermediate_tensors: IntermediateTensors | None = None
626

627
        # OPTIMIZATION: Cache the tensors rather than creating them every step.
628
        # Keep in int64 to avoid overflow with long context
629
630
631
632
        self.arange_np = np.arange(
            max(self.max_num_reqs + 1, self.max_model_len, self.max_num_tokens),
            dtype=np.int64,
        )
633

634
635
636
637
638
        # 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] = {}
639
640
641
642
643
        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(
644
645
                self.max_num_tokens, dtype=torch.int32, device=self.device
            )
646

647
        self.uniform_decode_query_len = 1 + self.num_spec_tokens
648
649
650
651

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

652
        self.mm_budget = (
653
            MultiModalBudget(self.vllm_config, self.mm_registry)
654
655
656
            if self.supports_mm_inputs
            else None
        )
657

658
        self.reorder_batch_threshold: int | None = None
659

660
661
662
663
664
        # 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()

665
        # Cached outputs.
666
        self._draft_token_ids: list[list[int]] | torch.Tensor | None = None
667
        self._draft_token_req_ids: list[str] | None = None
668
        self.transfer_event = torch.Event()
669
        self.sampled_token_ids_pinned_cpu = torch.empty(
670
            (self.max_num_reqs, 1),
671
672
            dtype=torch.int64,
            device="cpu",
673
674
            pin_memory=self.pin_memory,
        )
675

676
677
        # Pre-allocated tensor for copying valid sampled token counts to CPU,
        # with dedicated stream for overlapping and event for coordination.
678
        self.valid_sampled_token_count_event: torch.Event | None = None
679
        self.valid_sampled_token_count_copy_stream: torch.cuda.Stream | None = None
680
681
682
683
684
685
686
687
688
689
690
691
692
693
694
695
696
697
698
699
700
701
702
703
        # 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,
                )
704

705
706
        # Ephemeral state transferred between execute_model() and sample_tokens().
        self.execute_model_state: ExecuteModelState | None = None
707
        self.kv_connector_output: KVConnectorOutput | None = None
708
        self.mamba_state_idx: dict[str, int] = {}
709
        self.layerwise_nvtx_hooks_registered = False
710

王敏's avatar
王敏 committed
711
712
        self.draft_probs : Optional[DraftProbs] = None

713
714
715
716
717
718
719
    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

720
721
722
723
    def reset_mm_cache(self) -> None:
        if self.mm_budget:
            self.mm_budget.reset_cache()

724
725
726
727
728
729
730
731
732
733
734
735
736
737
738
739
740
741
742
743
744
745
746
747
748
749
750
751
752
753
754
755
756
757
758
759
760
761
762
763
764
765
766
767
    @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)

768
769
770
771
    def _get_positions(self, num_tokens: Any):
        if isinstance(num_tokens, int):
            if self.uses_mrope:
                return self.mrope_positions.gpu[:, :num_tokens]
772
773
            if self.uses_xdrope_dim > 0:
                return self.xdrope_positions.gpu[:, :num_tokens]
774
775
776
777
            return self.positions.gpu[:num_tokens]
        else:
            if self.uses_mrope:
                return self.mrope_positions.gpu[:, num_tokens]
778
779
            if self.uses_xdrope_dim > 0:
                return self.xdrope_positions.gpu[:, num_tokens]
780
781
            return self.positions.gpu[num_tokens]

782
    def _make_buffer(
783
        self, *size: int | torch.SymInt, dtype: torch.dtype, numpy: bool = True
784
785
786
787
788
789
790
791
    ) -> CpuGpuBuffer:
        return CpuGpuBuffer(
            *size,
            dtype=dtype,
            device=self.device,
            pin_memory=self.pin_memory,
            with_numpy=numpy,
        )
792

793
    def _init_model_kwargs(self):
794
795
        model_kwargs = dict[str, Any]()

796
        if not self.is_pooling_model:
797
798
            return model_kwargs

799
800
        num_reqs = self.input_batch.num_reqs
        pooling_params = self.input_batch.get_pooling_params()
801
802
803

        token_type_id_requests = dict[int, Any]()
        for i, param in enumerate(pooling_params):
804
805
806
807
808
            if (
                param.extra_kwargs is not None
                and (token_types := param.extra_kwargs.get("compressed_token_type_ids"))
                is not None
            ):
809
810
811
812
813
                token_type_id_requests[i] = token_types

        if len(token_type_id_requests) == 0:
            return model_kwargs

814
        seq_lens = self.seq_lens.gpu[:num_reqs]
815
816
817
818
819
820
821
822
        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(
823
824
            device=self.device
        )
825
        return model_kwargs
826

827
    def _may_reorder_batch(self, scheduler_output: "SchedulerOutput") -> None:
828
829
        """
        Update the order of requests in the batch based on the attention
830
        backend's needs. For example, some attention backends (namely MLA) may
831
832
833
834
835
836
        want to separate requests based on if the attention computation will be
        compute-bound or memory-bound.

        Args:
            scheduler_output: The scheduler output.
        """
837
838
839
840
841
842
843
844
        # 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

845
846
847
848
        if self.reorder_batch_threshold is not None:
            reorder_batch_to_split_decodes_and_prefills(
                self.input_batch,
                scheduler_output,
849
850
                decode_threshold=self.reorder_batch_threshold,
            )
851

852
853
    # Note: used for model runner override.
    def _init_device_properties(self) -> None:
854
        """Initialize attributes from torch.cuda.get_device_properties"""
855
856
857
858
859
860
861
        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()

862
    def _update_states(self, scheduler_output: "SchedulerOutput") -> None:
863
864
865
866
867
868
        """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.

869
870
        The SamplingMetadata is updated and copied to the GPU if there is a
        new/resumed/paused/finished request in the batch.
871
872
        """
        # Remove finished requests from the cached states.
873
874
        for req_id in scheduler_output.finished_req_ids:
            self.requests.pop(req_id, None)
875
            self.num_prompt_logprobs.pop(req_id, None)
876
877
878
879
880
881
882
        # 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:
883
            self.input_batch.remove_request(req_id)
884

王敏's avatar
王敏 committed
885
886
887
888
        # prune draft probs of finished requests
        if envs.VLLM_REJECT_SAMPLE_OPT and self.draft_probs is not None and len(scheduler_output.finished_req_ids) > 0:
            self.draft_probs.prune(list(scheduler_output.finished_req_ids))

889
        # Free the cached encoder outputs.
890
891
        for mm_hash in scheduler_output.free_encoder_mm_hashes:
            self.encoder_cache.pop(mm_hash, None)
892

893
894
895
896
897
898
899
        # 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()
900
901
902
903
904
905
906
907
        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)
908
909
910
911
912
        # 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:
913
            self.input_batch.remove_request(req_id)
914

915
        reqs_to_add: list[CachedRequestState] = []
916
        # Add new requests to the cached states.
917
918
        for new_req_data in scheduler_output.scheduled_new_reqs:
            req_id = new_req_data.req_id
919
920
921
922
923
924
            if req_id in self.requests:
                # For streaming case only.
                req_state = self._update_streaming_request(req_id, new_req_data)
                reqs_to_add.append(req_state)
                continue

925
            sampling_params = new_req_data.sampling_params
926
            pooling_params = new_req_data.pooling_params
927

928
929
930
931
            if (
                sampling_params
                and sampling_params.sampling_type == SamplingType.RANDOM_SEED
            ):
932
933
934
935
936
                generator = torch.Generator(device=self.device)
                generator.manual_seed(sampling_params.seed)
            else:
                generator = None

937
938
            if self.is_pooling_model:
                assert pooling_params is not None
939
940
                task = pooling_params.task
                assert task is not None, "You did not set `task` in the API"
941

942
                model = cast(VllmModelForPooling, self.get_model())
943
                to_update = model.pooler.get_pooling_updates(task)
944
945
                to_update.apply(pooling_params)

946
            req_state = CachedRequestState(
947
                req_id=req_id,
948
                prompt_token_ids=new_req_data.prompt_token_ids,
949
                prompt_embeds=new_req_data.prompt_embeds,
950
                mm_features=new_req_data.mm_features,
951
                sampling_params=sampling_params,
952
                pooling_params=pooling_params,
953
                generator=generator,
954
955
                block_ids=new_req_data.block_ids,
                num_computed_tokens=new_req_data.num_computed_tokens,
956
                output_token_ids=[],
957
                lora_request=new_req_data.lora_request,
958
            )
959
            self.requests[req_id] = req_state
960

961
962
963
964
965
966
967
            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
                )

968
            # Only relevant for models using M-RoPE (e.g, Qwen2-VL)
969
            if self.uses_mrope:
970
                self._init_mrope_positions(req_state)
971

972
973
974
975
            # Only relevant for models using XD-RoPE (e.g, HunYuan-VL)
            if self.uses_xdrope_dim > 0:
                self._init_xdrope_positions(req_state)

976
            reqs_to_add.append(req_state)
977

978
        # Update the states of the running/resumed requests.
979
        is_last_rank = get_pp_group().is_last_rank
980
        req_data = scheduler_output.scheduled_cached_reqs
981
        scheduled_spec_tokens = scheduler_output.scheduled_spec_decode_tokens
982
983
984
985
986

        # 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()

987
        for i, req_id in enumerate(req_data.req_ids):
988
            req_state = self.requests[req_id]
989
990
            num_computed_tokens = req_data.num_computed_tokens[i]
            new_block_ids = req_data.new_block_ids[i]
991
            resumed_from_preemption = req_id in req_data.resumed_req_ids
992
            num_output_tokens = req_data.num_output_tokens[i]
993
            req_index = self.input_batch.req_id_to_index.get(req_id)
994

995
996
997
998
999
1000
1001
1002
1003
1004
1005
1006
1007
1008
            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.
1009
1010
1011
1012
1013
1014
1015
1016
1017
                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)
1018

1019
            # Update the cached states.
1020
            req_state.num_computed_tokens = num_computed_tokens
1021
1022
1023
1024
1025
1026
1027
1028

            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.
1029
1030
1031
                num_new_tokens = (
                    num_computed_tokens + len(new_token_ids) - req_state.num_tokens
                )
1032
1033
1034
1035
                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:
1036
                    req_state.output_token_ids.extend(new_token_ids[-num_new_tokens:])
1037
1038
1039
1040
1041
            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:
1042
1043
1044
1045
                    end_idx = (
                        self.input_batch.num_prompt_tokens[req_index]
                        + num_output_tokens
                    )
1046
                    self.input_batch.num_tokens_no_spec[req_index] = end_idx
1047

1048
            # Update the block IDs.
1049
            if not resumed_from_preemption:
1050
1051
                if new_block_ids is not None:
                    # Append the new blocks to the existing block IDs.
1052
                    for block_ids, new_ids in zip(req_state.block_ids, new_block_ids):
1053
                        block_ids.extend(new_ids)
1054
            else:
1055
                assert req_index is None
1056
                assert new_block_ids is not None
1057
1058
                # The request is resumed from preemption.
                # Replace the existing block IDs with the new ones.
1059
                req_state.block_ids = new_block_ids
1060
1061
1062
1063
1064

            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.
1065
1066
1067
1068
1069
1070
1071

                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:]

1072
                reqs_to_add.append(req_state)
1073
1074
1075
                continue

            # Update the persistent batch.
1076
            self.input_batch.num_computed_tokens_cpu[req_index] = num_computed_tokens
1077
            if new_block_ids is not None:
1078
                self.input_batch.block_table.append_row(new_block_ids, req_index)
1079
1080
1081
1082
1083
1084

            # 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
zhuwenwen's avatar
zhuwenwen committed
1085
                end_token_index = num_computed_tokens + len(new_token_ids)
1086
                self.input_batch.token_ids_cpu[
1087
1088
1089
                    req_index, start_token_index:end_token_index
                ] = new_token_ids
                self.input_batch.num_tokens_no_spec[req_index] = end_token_index
1090

1091
            # Add spec_token_ids to token_ids_cpu.
1092
            self.input_batch.update_req_spec_token_ids(req_state, scheduled_spec_tokens)
1093

1094
1095
        # Add the new or resumed requests to the persistent batch.
        # The smaller empty indices are filled first.
1096
1097
        for request in reqs_to_add:
            self.input_batch.add_request(request)
1098
            self.input_batch.update_req_spec_token_ids(request, scheduled_spec_tokens)
1099

1100
1101
1102
1103
1104
1105
        # 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()
1106

1107
    def _update_states_after_model_execute(
1108
        self, output_token_ids: torch.Tensor, scheduler_output: "SchedulerOutput"
1109
    ) -> None:
1110
1111
1112
1113
1114
1115
1116
1117
        """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.
        """
1118
        if not self.speculative_config or not self.model_config.is_hybrid:
1119
1120
1121
            return

        # Find the number of accepted tokens for each sequence.
1122
1123
1124
1125
1126
1127
1128
1129
1130
1131
1132
1133
1134
1135
1136
1137
1138
1139
1140
1141
        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()
        )
1142
1143
        for i, num_tokens in enumerate(num_accepted_tokens):
            self.input_batch.num_accepted_tokens_cpu[i] = num_tokens
1144
1145
1146
1147
1148
1149
1150
1151
1152
1153
        if self.cache_config.mamba_cache_mode == "align":
            mamba_utils.postprocess_mamba(
                scheduler_output,
                self.kv_cache_config,
                self.input_batch,
                self.requests,
                self.mamba_state_idx,
                self.compilation_config.static_forward_context,
                self.model.get_mamba_state_copy_func(),
            )
1154

1155
1156
1157
1158
1159
1160
1161
1162
1163
1164
1165
1166
1167
1168
1169
1170
1171
1172
1173
1174
1175
1176
1177
1178
1179
1180
1181
1182
1183
1184
1185
1186
1187
    def _update_streaming_request(
        self, req_id: str, new_req_data: NewRequestData
    ) -> CachedRequestState:
        """Updates streaming session request from `scheduled_new_reqs`.

        Removes the request from InputBatch (if present), updates the cached
        state, and prepares it for re-addition to the batch.

        NOTE: prompt_token_ids includes intermediate output tokens - tokens
        previously generated but now are input context (part of the prompt).
        """
        self.input_batch.remove_request(req_id)
        req_state = self.requests[req_id]

        req_state.prompt_token_ids = new_req_data.prompt_token_ids
        req_state.mm_features = new_req_data.mm_features
        req_state.prompt_embeds = new_req_data.prompt_embeds
        req_state.sampling_params = new_req_data.sampling_params
        req_state.pooling_params = new_req_data.pooling_params
        req_state.block_ids = new_req_data.block_ids
        req_state.num_computed_tokens = new_req_data.num_computed_tokens
        req_state.num_prompt_tokens = length_from_prompt_token_ids_or_embeds(
            req_state.prompt_token_ids, req_state.prompt_embeds
        )

        # Clear `output_token_ids` as previous output tokens are now part of
        # `prompt_token_ids`.
        req_state.output_token_ids.clear()

        if self.uses_mrope:
            self._init_mrope_positions(req_state)

        return req_state
1188

1189
    def _init_mrope_positions(self, req_state: CachedRequestState):
1190
1191
        model = self.get_model()
        assert supports_mrope(model), "M-RoPE support is not implemented."
1192
1193
1194
1195
        assert req_state.prompt_token_ids is not None, (
            "M-RoPE requires prompt_token_ids to be available."
        )
        mrope_model = cast(SupportsMRoPE, model)
1196
1197

        req_state.mrope_positions, req_state.mrope_position_delta = (
1198
            mrope_model.get_mrope_input_positions(
1199
                req_state.prompt_token_ids,
1200
                req_state.mm_features,
1201
            )
1202
        )
1203

1204
1205
1206
1207
1208
1209
1210
1211
1212
1213
1214
1215
    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,
        )
1216

1217
    def _extract_mm_kwargs(
1218
        self,
1219
1220
        scheduler_output: "SchedulerOutput",
    ) -> BatchedTensorInputs:
1221
        if not scheduler_output or not self.is_multimodal_raw_input_only_model:
1222
            return {}
1223

1224
1225
        mm_kwargs = list[MultiModalKwargsItem]()
        for req in scheduler_output.scheduled_new_reqs:
1226
1227
1228
            for feature in req.mm_features:
                if feature.data is not None:
                    mm_kwargs.append(feature.data)
1229

1230
1231
1232
        # Input all modalities at once
        mm_kwargs_combined: BatchedTensorInputs = {}
        for _, _, mm_kwargs_group in group_mm_kwargs_by_modality(
1233
1234
1235
            mm_kwargs,
            device=self.device,
            pin_memory=self.pin_memory,
1236
1237
        ):
            mm_kwargs_combined.update(mm_kwargs_group)
1238

1239
        return mm_kwargs_combined
1240
1241

    def _dummy_mm_kwargs(self, num_seqs: int) -> BatchedTensorInputs:
1242
        if not self.is_multimodal_raw_input_only_model:
1243
            return {}
1244

1245
1246
        mm_budget = self.mm_budget
        assert mm_budget is not None
1247

1248
1249
        dummy_modality = mm_budget.get_modality_with_max_tokens()
        return self._get_mm_dummy_batch(dummy_modality, num_seqs)
1250

1251
1252
1253
    def _get_cumsum_and_arange(
        self,
        num_tokens: np.ndarray,
1254
        cumsum_dtype: np.dtype | None = None,
1255
1256
1257
1258
1259
1260
1261
1262
1263
1264
1265
1266
1267
1268
1269
1270
    ) -> 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

1271
    def _prepare_input_ids(
1272
1273
1274
1275
        self,
        scheduler_output: "SchedulerOutput",
        total_num_scheduled_tokens: int,
        cu_num_tokens: np.ndarray,
1276
    ) -> None:
1277
        """Prepare the input IDs for the current batch.
1278

1279
1280
1281
1282
1283
1284
1285
        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)
1286
1287
1288
            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)
1289
1290
1291
1292
1293
1294
1295
            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
1296
1297
1298
1299
        sample_flattened_indices: list[int] = []
        spec_flattened_indices: list[int] = []
        prev_common_req_indices: list[int] = []
        prev_draft_token_indices: list[int] = []
1300
1301
        indices_match = True
        max_flattened_index = -1
1302
1303
1304
        total_num_spec_tokens = 0
        scheduled_spec_tokens = scheduler_output.scheduled_spec_decode_tokens

1305
1306
1307
1308
1309
        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.
1310
1311
                draft_len = len(scheduled_spec_tokens.get(req_id, ()))
                total_num_spec_tokens += draft_len
1312
                flattened_index = cu_num_tokens[cur_index].item() - 1
1313
1314
1315
1316
1317
1318
1319
1320
1321
1322
1323
1324
1325
1326
1327
                # 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))
1328
                indices_match &= prev_index == flattened_index
1329
                max_flattened_index = max(max_flattened_index, flattened_index)
1330
1331
1332
        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:
1333
1334
1335
            # 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)
1336
1337
1338
            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)
1339
1340
        if num_commmon_tokens == 0:
            # No requests in common with the previous iteration
1341
            # So input_ids.cpu will have all the input ids.
1342
1343
1344
1345
1346
1347
1348
            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_(
1349
1350
1351
                self.input_batch.prev_sampled_token_ids[:num_commmon_tokens, 0],
                non_blocking=True,
            )
1352
1353
            if self.enable_prompt_embeds:
                self.is_token_ids.gpu[:num_commmon_tokens] = True
1354
            return
1355
        # Upload the index tensors asynchronously so the scatter can be non-blocking.
1356
1357
        sampled_tokens_index_tensor = torch.tensor(
            sample_flattened_indices, dtype=torch.int64, pin_memory=self.pin_memory
1358
        ).to(self.device, non_blocking=True)
1359
        prev_common_req_indices_tensor = torch.tensor(
1360
1361
            prev_common_req_indices, dtype=torch.int64, pin_memory=self.pin_memory
        ).to(self.device, non_blocking=True)
1362
1363
        self.input_ids.gpu.scatter_(
            dim=0,
1364
            index=sampled_tokens_index_tensor,
1365
            src=self.input_batch.prev_sampled_token_ids[
1366
1367
1368
                prev_common_req_indices_tensor, 0
            ],
        )
1369

1370
1371
1372
1373
1374
1375
1376
1377
1378
1379
1380
1381
1382
1383
1384
1385
1386
1387
1388
1389
1390
        # 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],
        )
1391

1392
1393
    def _get_encoder_seq_lens(
        self,
1394
        num_scheduled_tokens: dict[str, int],
1395
1396
        kv_cache_spec: KVCacheSpec,
        num_reqs: int,
1397
        for_cudagraph_capture: bool = False,
1398
    ) -> tuple[torch.Tensor | None, np.ndarray | None]:
1399
        if not isinstance(kv_cache_spec, CrossAttentionSpec):
1400
            return None, None
1401

1402
1403
        # Zero out buffer for padding requests that are not actually scheduled (CGs)
        self.encoder_seq_lens.np[:num_reqs] = 0
1404

1405
1406
        # Build encoder_seq_lens array mapping request indices to
        # encoder lengths for inputs scheduled in this batch
1407
        for req_id in num_scheduled_tokens:
1408
            req_index = self.input_batch.req_id_to_index[req_id]
1409
1410
1411
1412
1413
1414
1415
1416
1417
1418
1419
1420
            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
1421
1422
1423
1424
1425
1426
1427
1428
1429
        if for_cudagraph_capture:
            # During CUDA graph capture, we need to use realistic encoder lengths
            # so that max_seqlen_k is captured with the correct value.
            max_encoder_len = getattr(
                self.model_config.hf_config,
                "max_source_positions",
                self.max_encoder_len,
            )
            self.encoder_seq_lens.np[:num_reqs] = max_encoder_len
1430

1431
1432
1433
        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]
1434

1435
        return encoder_seq_lens, encoder_seq_lens_cpu
1436

1437
    def _prepare_inputs(
1438
1439
1440
        self,
        scheduler_output: "SchedulerOutput",
        num_scheduled_tokens: np.ndarray,
1441
1442
    ) -> tuple[
        torch.Tensor,
1443
        SpecDecodeMetadata | None,
1444
    ]:
1445
1446
        """
        :return: tuple[
1447
            logits_indices, spec_decode_metadata,
1448
1449
        ]
        """
1450
1451
1452
1453
1454
1455
1456
        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.
1457
        self.input_batch.block_table.commit_block_table(num_reqs)
1458
1459
1460

        # Get request indices.
        # E.g., [2, 5, 3] -> [0, 0, 1, 1, 1, 1, 1, 2, 2, 2]
1461
        req_indices = np.repeat(self.arange_np[:num_reqs], num_scheduled_tokens)
1462

1463
1464
        # cu_num_tokens: [2, 5, 3] -> [2, 7, 10]
        # arange: [0, 1, 0, 1, 2, 3, 4, 0, 1, 2]
1465
        cu_num_tokens, arange = self._get_cumsum_and_arange(num_scheduled_tokens)
1466
1467

        # Get positions.
1468
        positions_np = self.positions.np[:total_num_scheduled_tokens]
1469
1470
1471
1472
1473
        np.add(
            self.input_batch.num_computed_tokens_cpu[req_indices],
            arange,
            out=positions_np,
        )
1474

1475
1476
        # Calculate M-RoPE positions.
        # Only relevant for models using M-RoPE (e.g, Qwen2-VL)
1477
        if self.uses_mrope:
1478
1479
            self._calc_mrope_positions(scheduler_output)

1480
1481
1482
1483
1484
        # 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)

1485
1486
1487
1488
        # 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.
1489
1490
1491
        token_indices = (
            positions_np + req_indices * self.input_batch.token_ids_cpu.shape[1]
        )
1492
        token_indices_tensor = torch.from_numpy(token_indices)
1493

1494
1495
1496
        # 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.
1497
1498
1499
1500
1501
1502
        torch.index_select(
            self.input_batch.token_ids_cpu_tensor.flatten(),
            0,
            token_indices_tensor,
            out=self.input_ids.cpu[:total_num_scheduled_tokens],
        )
1503
        if self.enable_prompt_embeds:
1504
            is_token_ids = self.input_batch.is_token_ids_tensor.flatten()
1505
1506
1507
1508
            torch.index_select(
                is_token_ids,
                0,
                token_indices_tensor,
1509
1510
                out=self.is_token_ids.cpu[:total_num_scheduled_tokens],
            )
1511
1512
1513
1514
1515
1516
1517
1518
1519
1520
1521
1522
1523
1524
1525
1526
1527
1528
1529
1530
1531
1532
1533
1534
1535
1536
1537
1538
1539
1540
1541
1542
1543

        # 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:
1544
1545
1546
                    self.inputs_embeds.cpu[
                        output_idx : output_idx + actual_num_sched
                    ].copy_(req_embeds[start_pos:actual_end])
1547
1548

                output_idx += num_sched
1549

1550
1551
        self.input_batch.block_table.compute_slot_mapping(req_indices, positions_np)
        self.input_batch.block_table.commit_slot_mapping(total_num_scheduled_tokens)
1552
1553

        # Prepare the attention metadata.
1554
        self.query_start_loc.np[0] = 0
1555
        self.query_start_loc.np[1 : num_reqs + 1] = cu_num_tokens
1556
1557
        # Note: pad query_start_loc to be non-decreasing, as kernels
        # like FlashAttention requires that
1558
        self.query_start_loc.np[num_reqs + 1 :].fill(cu_num_tokens[-1])
1559
        self.query_start_loc.copy_to_gpu()
1560
        query_start_loc = self.query_start_loc.gpu[: num_reqs + 1]
1561

1562
        self.seq_lens.np[:num_reqs] = (
1563
1564
            self.input_batch.num_computed_tokens_cpu[:num_reqs] + num_scheduled_tokens
        )
1565
        # Fill unused with 0 for full cuda graph mode.
1566
1567
        self.seq_lens.np[num_reqs:].fill(0)
        self.seq_lens.copy_to_gpu()
1568

1569
        num_tokens = [self.requests[r].num_tokens for r in self.input_batch.req_ids]
1570
1571
        num_tokens_np = np.array(num_tokens, dtype=np.int32)

1572
        # Record which requests should not be sampled,
1573
        # so that we could clear the sampled tokens before returning
1574
1575
        self.discard_request_mask.np[:num_reqs] = (
            self.seq_lens.np[:num_reqs] < num_tokens_np
1576
        )
1577
        self.discard_request_mask.copy_to_gpu(num_reqs)
1578

1579
        # Copy the tensors to the GPU.
1580
1581
1582
1583
1584
        self._prepare_input_ids(
            scheduler_output,
            total_num_scheduled_tokens,
            cu_num_tokens,
        )
1585

1586
        if self.uses_mrope:
1587
            # Only relevant for models using M-RoPE (e.g, Qwen2-VL)
1588
1589
            self.mrope_positions.gpu[:, :total_num_scheduled_tokens].copy_(
                self.mrope_positions.cpu[:, :total_num_scheduled_tokens],
1590
1591
                non_blocking=True,
            )
1592
1593
1594
1595
1596
1597
        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,
            )
1598
1599
        else:
            # Common case (1D positions)
1600
            self.positions.copy_to_gpu(total_num_scheduled_tokens)
1601

1602
        use_spec_decode = len(scheduler_output.scheduled_spec_decode_tokens) > 0
1603
1604
1605
1606
1607
1608
1609
1610
        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
1611
            num_sampled_tokens = np.ones(num_reqs, dtype=np.int32)
1612
1613
1614
1615
1616
        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)
1617
1618
1619
            # 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)
1620
1621
1622
1623
            for (
                req_id,
                draft_token_ids,
            ) in scheduler_output.scheduled_spec_decode_tokens.items():
1624
1625
                req_idx = self.input_batch.req_id_to_index[req_id]
                num_draft_tokens[req_idx] = len(draft_token_ids)
1626
1627
1628
1629
1630
                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)
王敏's avatar
王敏 committed
1631
1632
1633
1634
1635

            spec_decode_ids = None
            if envs.VLLM_REJECT_SAMPLE_OPT:
                spec_decode_ids = scheduler_output.scheduled_spec_decode_tokens.keys()

1636
            spec_decode_metadata = self._calc_spec_decode_metadata(
王敏's avatar
王敏 committed
1637
                num_draft_tokens, cu_num_tokens, spec_decode_ids
1638
            )
1639
            logits_indices = spec_decode_metadata.logits_indices
1640
            num_sampled_tokens = num_draft_tokens + 1
1641
            # For DECODE only cuda graph of some attention backends (e.g., GDN).
1642
            self.num_decode_draft_tokens.np[:num_reqs] = num_decode_draft_tokens
1643
1644
            self.num_decode_draft_tokens.np[num_reqs:].fill(-1)
            self.num_decode_draft_tokens.copy_to_gpu()
1645

1646
1647
1648
1649
1650
        # Hot-Swap lora model
        if self.lora_config:
            assert (
                np.sum(num_sampled_tokens)
                <= self.vllm_config.scheduler_config.max_num_batched_tokens
1651
            )
1652
1653
1654
1655
1656
1657
1658
1659
1660
1661
1662
            self.set_active_loras(
                self.input_batch, num_scheduled_tokens, num_sampled_tokens
            )

        return (
            logits_indices,
            spec_decode_metadata,
        )

    def _build_attention_metadata(
        self,
1663
        num_tokens: int,
1664
        num_reqs: int,
1665
1666
1667
        max_query_len: int,
        num_tokens_padded: int | None = None,
        num_reqs_padded: int | None = None,
1668
1669
1670
1671
        ubatch_slices: UBatchSlices | None = None,
        logits_indices: torch.Tensor | None = None,
        use_spec_decode: bool = False,
        for_cudagraph_capture: bool = False,
1672
        num_scheduled_tokens: dict[str, int] | None = None,
1673
        cascade_attn_prefix_lens: list[list[int]] | None = None,
1674
        slot_mappings: dict[int, torch.Tensor] | None = None,
1675
1676
1677
1678
    ) -> tuple[PerLayerAttnMetadata, CommonAttentionMetadata | None]:
        """
        :return: tuple[attn_metadata, spec_decode_common_attn_metadata]
        """
1679
1680
1681
1682
        # Attention metadata is not needed for attention free models
        if len(self.kv_cache_config.kv_cache_groups) == 0:
            return {}, None

1683
1684
        num_tokens_padded = num_tokens_padded or num_tokens
        num_reqs_padded = num_reqs_padded or num_reqs
1685
        assert num_reqs_padded is not None and num_tokens_padded is not None
1686

1687
1688
1689
        attn_metadata: PerLayerAttnMetadata = {}
        if ubatch_slices is not None:
            attn_metadata = [dict() for _ in range(len(ubatch_slices))]
1690

1691
1692
1693
1694
1695
1696
1697
1698
        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()

1699
1700
        if use_spec_decode:
            self.num_accepted_tokens.np[:num_reqs] = (
1701
1702
                self.input_batch.num_accepted_tokens_cpu[:num_reqs]
            )
1703
1704
            self.num_accepted_tokens.np[num_reqs:].fill(1)
            self.num_accepted_tokens.copy_to_gpu()
1705

1706
        kv_cache_groups = self.kv_cache_config.kv_cache_groups
1707

1708
        def _get_block_table(kv_cache_gid: int):
1709
1710
1711
            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):
1712
                blk_table_tensor = torch.zeros(
1713
                    (num_reqs_padded, 1),
1714
                    dtype=torch.int32,
1715
1716
                    device=self.device,
                )
1717
            else:
1718
                blk_table = self.input_batch.block_table[kv_cache_gid]
1719
                blk_table_tensor = blk_table.get_device_tensor(num_reqs_padded)
1720

1721
1722
1723
            # Fill unused with -1. Needed for reshape_and_cache in full cuda
            # graph mode. `blk_table_tensor` -1 to match mamba PAD_SLOT_ID
            blk_table_tensor[num_reqs:num_reqs_padded].fill_(-1)
1724
            return blk_table_tensor
1725

1726
1727
1728
        assert slot_mappings is not None
        block_table_gid_0 = _get_block_table(0)
        slot_mapping_gid_0 = slot_mappings[0]
1729

1730
1731
        if self.model_config.enable_return_routed_experts:
            self.slot_mapping = slot_mapping_gid_0[:num_tokens].cpu().numpy()
1732
1733
1734
1735
1736
1737
1738
1739
1740
1741
1742
1743
1744
1745
1746
1747
1748
1749
1750
1751
1752
1753
1754
1755
1756
1757
1758
1759
1760
1761
1762
1763
1764
1765
1766
1767
1768
1769
        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
            )

1770
1771
1772
1773
1774
1775
1776
1777
1778
        # 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
        ] = {}

1779
1780
1781
1782
1783
1784
1785
        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]
1786
            builder = attn_group.get_metadata_builder(ubid or 0)
1787
1788
1789
1790
            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))
1791

1792
1793
1794
1795
            cascade_attn_prefix_len = (
                cascade_attn_prefix_lens[kv_cache_gid][attn_gid]
                if cascade_attn_prefix_lens
                else 0
1796
1797
            )

1798
1799
1800
1801
1802
1803
1804
1805
1806
1807
1808
1809
1810
1811
            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
                )
1812
1813
1814
1815
1816
1817
1818
1819
1820
            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,
                )
1821
1822
1823
1824
1825
1826
            else:
                attn_metadata_i = builder.build(
                    common_prefix_len=cascade_attn_prefix_len,
                    common_attn_metadata=common_attn_metadata,
                    **extra_attn_metadata_args,
                )
1827
1828
                if builder.supports_update_block_table:
                    cached_attn_metadata[cache_key] = attn_metadata_i
1829
1830
1831
1832
1833
1834
1835
1836
1837
1838
1839
1840
1841
1842
1843
1844
1845
1846
1847
1848
1849
1850
1851

            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,
1852
                for_cudagraph_capture=for_cudagraph_capture,
1853
            )
1854
            if kv_cache_gid > 0:
1855
1856
                cm.block_table_tensor = _get_block_table(kv_cache_gid)
                cm.slot_mapping = slot_mappings[kv_cache_gid]
1857

1858
            if self.speculative_config and spec_decode_common_attn_metadata is None:
1859
                if isinstance(self.drafter, EagleProposer):
1860
                    if self.drafter.attn_layer_names[0] in kv_cache_group.layer_names:
1861
                        spec_decode_common_attn_metadata = cm
1862
                else:
1863
                    spec_decode_common_attn_metadata = cm
1864

1865
            for attn_gid in range(len(self.attn_groups[kv_cache_gid])):
1866
                if ubatch_slices is not None:
1867
1868
1869
                    for ubid, _cm in enumerate(split_attn_metadata(ubatch_slices, cm)):
                        _build_attn_group_metadata(kv_cache_gid, attn_gid, _cm, ubid)

1870
                else:
1871
                    _build_attn_group_metadata(kv_cache_gid, attn_gid, cm)
1872

1873
1874
1875
1876
1877
1878
1879
1880
1881
1882
1883
1884
1885
1886
1887
1888
1889
1890
1891
        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]
1892

1893
1894
1895
1896
1897
1898
1899
1900
1901
1902
        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)
            )

1903
        return attn_metadata, spec_decode_common_attn_metadata
1904

1905
1906
1907
    def _compute_cascade_attn_prefix_lens(
        self,
        num_scheduled_tokens: np.ndarray,
1908
        num_computed_tokens: np.ndarray,
1909
1910
1911
1912
1913
1914
1915
        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
        """
1916

1917
1918
1919
1920
1921
1922
1923
1924
1925
1926
1927
1928
1929
1930
        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,
1931
                        num_computed_tokens,
1932
1933
1934
1935
1936
1937
                        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
1938

1939
        return cascade_attn_prefix_lens if use_cascade_attn else None
1940

1941
1942
1943
    def _compute_cascade_attn_prefix_len(
        self,
        num_scheduled_tokens: np.ndarray,
1944
        num_computed_tokens: np.ndarray,
1945
        num_common_prefix_blocks: int,
1946
1947
        kv_cache_spec: KVCacheSpec,
        attn_metadata_builder: AttentionMetadataBuilder,
1948
1949
1950
1951
1952
1953
1954
1955
1956
1957
1958
1959
1960
1961
1962
1963
1964
1965
    ) -> 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.
        """
1966

1967
        common_prefix_len = num_common_prefix_blocks * kv_cache_spec.block_size
1968
1969
1970
1971
1972
1973
1974
1975
1976
1977
1978
1979
1980
1981
1982
1983
1984
1985
1986
1987
1988
1989
1990
1991
1992
1993
1994
1995
1996
1997
1998
1999
2000
2001
2002
2003
2004
        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]
2005
        # Request 3's num_computed_tokens: 3 (i.e., [A, B, C])
2006
2007
2008
2009
2010
        # 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.
2011
        common_prefix_len = min(common_prefix_len, num_computed_tokens.min())
2012
        # common_prefix_len should be a multiple of the block size.
2013
2014
2015
2016
2017
2018
2019
2020
2021
2022
2023
        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
        )
2024
2025
        assert isinstance(kv_cache_spec, AttentionSpec)
        use_cascade = attn_metadata_builder.use_cascade_attention(
2026
2027
2028
            common_prefix_len=common_prefix_len,
            query_lens=num_scheduled_tokens,
            num_query_heads=self.num_query_heads,
2029
            num_kv_heads=kv_cache_spec.num_kv_heads,
2030
            use_alibi=self.use_alibi,
2031
            use_sliding_window=use_sliding_window,
2032
            use_local_attention=use_local_attention,
2033
            num_sms=self.num_sms,
2034
            dcp_world_size=self.dcp_world_size,
2035
2036
2037
        )
        return common_prefix_len if use_cascade else 0

2038
2039
    def _calc_mrope_positions(self, scheduler_output: "SchedulerOutput"):
        mrope_pos_ptr = 0
2040
        for index, req_id in enumerate(self.input_batch.req_ids):
2041
2042
2043
            req = self.requests[req_id]
            assert req.mrope_positions is not None

2044
2045
            num_computed_tokens = self.input_batch.num_computed_tokens_cpu[index]
            num_scheduled_tokens = scheduler_output.num_scheduled_tokens[req_id]
2046
            num_prompt_tokens = length_from_prompt_token_ids_or_embeds(
2047
2048
                req.prompt_token_ids, req.prompt_embeds
            )
2049
2050

            if num_computed_tokens + num_scheduled_tokens > num_prompt_tokens:
2051
2052
                prompt_part_len = max(0, num_prompt_tokens - num_computed_tokens)
                completion_part_len = max(0, num_scheduled_tokens - prompt_part_len)
2053
2054
2055
2056
2057
2058
2059
2060
2061
2062
2063
2064
2065
            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

2066
2067
2068
                self.mrope_positions.cpu[:, dst_start:dst_end] = req.mrope_positions[
                    :, src_start:src_end
                ]
2069
2070
2071
2072
2073
2074
2075
                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

2076
                assert req.mrope_position_delta is not None
2077
                MRotaryEmbedding.get_next_input_positions_tensor(
2078
                    out=self.mrope_positions.np,
2079
2080
2081
2082
2083
                    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,
                )
2084
2085
2086

                mrope_pos_ptr += completion_part_len

2087
2088
2089
2090
2091
    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
2092

2093
2094
2095
2096
2097
            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
            )
2098

2099
2100
2101
2102
2103
2104
2105
2106
2107
2108
2109
2110
2111
2112
2113
2114
2115
2116
2117
2118
2119
2120
2121
2122
2123
2124
2125
2126
2127
2128
2129
2130
2131
2132
2133
            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

2134
2135
    def _calc_spec_decode_metadata(
        self,
2136
2137
        num_draft_tokens: np.ndarray,
        cu_num_scheduled_tokens: np.ndarray,
王敏's avatar
王敏 committed
2138
        spec_decode_ids: Optional[list[str]] = None
2139
2140
2141
2142
2143
2144
2145
2146
2147
2148
2149
2150
2151
2152
    ) -> 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
2153
2154
2155
2156

        # 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(
2157
2158
            num_sampled_tokens, cumsum_dtype=np.int32
        )
2159
        # Step 2. [0, 0, 0, 0, 103, 104, 104, 104, 206, 207, 207]
2160
        logits_indices = np.repeat(
2161
2162
            cu_num_scheduled_tokens - num_sampled_tokens, num_sampled_tokens
        )
2163
        # Step 3. [0, 1, 2, 3, 103, 104, 105, 106, 206, 207, 208]
2164
2165
2166
2167
2168
2169
        logits_indices += arange

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

        # Compute the draft logits indices.
2170
2171
2172
        # 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(
2173
2174
            num_draft_tokens, cumsum_dtype=np.int32
        )
2175
2176
        # [0, 0, 0, 5, 5, 9]
        target_logits_indices = np.repeat(
2177
2178
            cu_num_sampled_tokens - num_sampled_tokens, num_draft_tokens
        )
2179
2180
2181
        # [0, 1, 2, 5, 6, 9]
        target_logits_indices += arange

2182
2183
        # TODO: Optimize the CPU -> GPU copy.
        cu_num_draft_tokens = torch.from_numpy(cu_num_draft_tokens).to(
2184
2185
            self.device, non_blocking=True
        )
2186
2187
2188
        cu_num_sampled_tokens = torch.from_numpy(cu_num_sampled_tokens).to(
            self.device, non_blocking=True
        )
2189
2190
2191
        logits_indices = torch.from_numpy(logits_indices).to(
            self.device, non_blocking=True
        )
2192
        target_logits_indices = torch.from_numpy(target_logits_indices).to(
2193
2194
            self.device, non_blocking=True
        )
2195
        bonus_logits_indices = torch.from_numpy(bonus_logits_indices).to(
2196
2197
            self.device, non_blocking=True
        )
2198

2199
2200
        # Compute the draft token ids.
        # draft_token_indices:      [  1,   2,   3, 105, 106, 208]
2201
        draft_token_ids = self.input_ids.gpu[logits_indices]
2202
2203
        draft_token_ids = draft_token_ids[target_logits_indices + 1]

2204
        return SpecDecodeMetadata(
2205
2206
2207
            draft_token_ids=draft_token_ids,
            num_draft_tokens=num_draft_tokens.tolist(),
            cu_num_draft_tokens=cu_num_draft_tokens,
2208
            cu_num_sampled_tokens=cu_num_sampled_tokens,
2209
2210
2211
            target_logits_indices=target_logits_indices,
            bonus_logits_indices=bonus_logits_indices,
            logits_indices=logits_indices,
王敏's avatar
王敏 committed
2212
            spec_decode_ids=spec_decode_ids,
2213
2214
        )

2215
2216
2217
2218
2219
2220
2221
    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
2222
        self.kv_sharing_fast_prefill_logits_indices[:num_logits].copy_(logits_indices)
2223
2224
2225
2226
2227
        # 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_(
2228
2229
            logits_indices[-1].item()
        )
2230
2231
2232
2233
2234
        # Dispatch for the decoder portion of the model.
        _, batch_desc = self.cudagraph_dispatcher.dispatch(
            num_logits, disable_full=True
        )
        num_logits_padded = batch_desc.num_tokens
2235
2236
2237
        logits_indices_padded = self.kv_sharing_fast_prefill_logits_indices[
            :num_logits_padded
        ]
2238
2239
        return logits_indices_padded

2240
    def _batch_mm_inputs_from_scheduler(
2241
2242
        self,
        scheduler_output: "SchedulerOutput",
2243
2244
2245
2246
2247
    ) -> tuple[
        list[str],
        list[MultiModalKwargsItem],
        list[tuple[str, PlaceholderRange]],
    ]:
2248
        """Batch multimodal inputs from scheduled encoder inputs.
2249
2250
2251

        Args:
            scheduler_output: The scheduler output containing scheduled encoder
2252
                inputs.
2253
2254

        Returns:
2255
            A tuple of (mm_hashes, mm_kwargs, mm_lora_refs) where:
2256
2257
            - mm_hashes: List of multimodal hashes for each item
            - mm_kwargs: List of multimodal kwargs for each item
2258
            - mm_lora_refs: List of (req_id, placeholder_range) for each item
2259
        """
2260
2261
        scheduled_encoder_inputs = scheduler_output.scheduled_encoder_inputs
        if not scheduled_encoder_inputs:
2262
            return [], [], []
2263
2264

        mm_hashes = list[str]()
2265
        mm_kwargs = list[MultiModalKwargsItem]()
2266
2267
2268
        # Multimodal LoRA reference info to map each multimodal item
        # back to its request & position
        mm_lora_refs = list[tuple[str, PlaceholderRange]]()
2269
2270
        for req_id, encoder_input_ids in scheduled_encoder_inputs.items():
            req_state = self.requests[req_id]
2271
2272

            for mm_input_id in encoder_input_ids:
2273
                mm_feature = req_state.mm_features[mm_input_id]
2274
2275
                if mm_feature.data is None:
                    continue
2276
2277

                mm_hashes.append(mm_feature.identifier)
2278
                mm_kwargs.append(mm_feature.data)
2279
                mm_lora_refs.append((req_id, mm_feature.mm_position))
2280

2281
        return mm_hashes, mm_kwargs, mm_lora_refs
2282

2283
2284
2285
    def _execute_mm_encoder(
        self, scheduler_output: "SchedulerOutput"
    ) -> list[torch.Tensor]:
2286
        mm_hashes, mm_kwargs, mm_lora_refs = self._batch_mm_inputs_from_scheduler(
2287
2288
            scheduler_output
        )
2289
2290

        if not mm_kwargs:
2291
            return []
2292

2293
2294
2295
2296
2297
2298
        should_time = bool(
            self.observability_config
            and self.observability_config.enable_mm_processor_stats
            and scheduler_output.scheduled_encoder_inputs
        )

2299
2300
2301
2302
2303
2304
2305
        # 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.
2306
        model = cast(SupportsMultiModal, self.model)
2307
2308
2309
2310
2311
2312
2313
2314
2315
2316
2317
2318
2319
2320
2321
2322
2323
2324
2325
2326
2327
2328
2329
2330
2331
2332
2333
2334
2335
2336
2337
2338
2339
2340
2341
2342
2343
2344
2345
2346
2347
2348
2349
2350
2351
2352
2353
2354
2355
2356
2357
2358
2359
2360
2361
2362
2363

        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,
                )

2364
        encoder_outputs: list[torch.Tensor] = []
2365
2366
        # Track the current index in mm_kwargs/mm_lora_refs to map groups to request IDs
        current_item_idx = 0
2367
        for modality, num_items, mm_kwargs_group in group_mm_kwargs_by_modality(
2368
2369
2370
            mm_kwargs,
            device=self.device,
            pin_memory=self.pin_memory,
2371
        ):
2372
            curr_group_outputs: MultiModalEmbeddings
2373
2374

            # EVS-related change.
2375
            # (ekhvedchenia): Temporary hack to limit peak memory usage when
2376
            # processing multimodal data. This solves the issue with scheduler
2377
2378
2379
2380
            # 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)
2381
2382
2383
2384
2385
2386
2387
            # 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
            ):
2388
                curr_group_outputs_lst = list[torch.Tensor]()
2389
2390
2391
2392
2393
2394
2395
2396
2397
2398
2399
                for video_idx in range(num_items):
                    video_mm_kwargs_item = mm_kwargs[current_item_idx + video_idx]
                    with self.timed_encoder_operation(
                        should_time, mm_lora_refs, current_item_idx + video_idx, 1
                    ):
                        _, _, micro_batch_mm_inputs = next(
                            group_mm_kwargs_by_modality(
                                [video_mm_kwargs_item],
                                device=self.device,
                                pin_memory=self.pin_memory,
                            )
2400
                        )
2401

2402
2403
2404
                        micro_batch_outputs = model.embed_multimodal(
                            **micro_batch_mm_inputs
                        )
2405

2406
                        curr_group_outputs_lst.extend(micro_batch_outputs)
2407
2408

                curr_group_outputs = curr_group_outputs_lst
2409
2410
2411
2412
2413
2414
2415
2416
            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.
2417
2418
2419
2420
2421

                with self.timed_encoder_operation(
                    should_time, mm_lora_refs, current_item_idx, num_items
                ):
                    curr_group_outputs = model.embed_multimodal(**mm_kwargs_group)
2422

2423
2424
            sanity_check_mm_encoder_outputs(
                curr_group_outputs,
2425
                expected_num_items=num_items,
2426
            )
2427
            encoder_outputs.extend(curr_group_outputs)
2428

2429
2430
            current_item_idx += num_items

2431
        # Cache the encoder outputs by mm_hash
2432
        for mm_hash, output in zip(mm_hashes, encoder_outputs):
2433
            self.encoder_cache[mm_hash] = output
2434
2435
            logger.debug("Finish execute for mm hash %s", mm_hash)
            self.maybe_save_ec_to_connector(self.encoder_cache, mm_hash)
2436

2437
        return encoder_outputs
2438
2439

    def _gather_mm_embeddings(
2440
2441
        self,
        scheduler_output: "SchedulerOutput",
2442
        shift_computed_tokens: int = 0,
2443
2444
2445
    ) -> tuple[list[torch.Tensor], torch.Tensor]:
        total_num_scheduled_tokens = scheduler_output.total_num_scheduled_tokens

2446
2447
2448
2449
2450
        # 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]

2451
        mm_embeds = list[torch.Tensor]()
2452
        is_mm_embed = is_mm_embed_buf.cpu
2453
2454
2455
        is_mm_embed[:total_num_scheduled_tokens] = False

        req_start_idx = 0
2456
        should_sync_mrope_positions = False
2457
        should_sync_xdrope_positions = False
2458

2459
        for req_id in self.input_batch.req_ids:
2460
2461
            mm_embeds_req: list[torch.Tensor] = []

2462
            num_scheduled_tokens = scheduler_output.num_scheduled_tokens[req_id]
2463
            req_state = self.requests[req_id]
2464
            num_computed_tokens = req_state.num_computed_tokens + shift_computed_tokens
2465

2466
2467
            for mm_feature in req_state.mm_features:
                pos_info = mm_feature.mm_position
2468
2469
                start_pos = pos_info.offset
                num_encoder_tokens = pos_info.length
2470
2471
2472
2473
2474
2475
2476
2477
2478
2479
2480
2481
2482
2483
2484
2485

                # 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,
2486
2487
                    num_encoder_tokens,
                )
2488
                assert start_idx < end_idx
2489
2490
2491
2492
2493
2494
2495
                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
2496

2497
                mm_hash = mm_feature.identifier
2498
                encoder_output = self.encoder_cache.get(mm_hash, None)
2499
                assert encoder_output is not None, f"Encoder cache miss for {mm_hash}."
2500
2501
2502

                if (is_embed := pos_info.is_embed) is not None:
                    is_embed = is_embed[start_idx:end_idx]
2503
2504
2505
                    mm_embeds_item = encoder_output[curr_embeds_start:curr_embeds_end]
                else:
                    mm_embeds_item = encoder_output[start_idx:end_idx]
2506

2507
                req_start_pos = req_start_idx + start_pos - num_computed_tokens
2508
2509
2510
2511
2512
2513
2514
2515
2516
                # OR mask for overlapping mm_features (use_audio_in_video)
                if is_embed is None:
                    is_mm_embed[req_start_pos + start_idx : req_start_pos + end_idx] = (
                        True
                    )
                else:
                    is_mm_embed[
                        req_start_pos + start_idx : req_start_pos + end_idx
                    ] |= is_embed
2517
2518
2519
                mm_embeds_req.append(mm_embeds_item)

            if self.is_multimodal_pruning_enabled and self.uses_mrope:
2520
                assert req_state.mrope_positions is not None
2521
2522
2523
2524
2525
2526
2527
                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,
2528
2529
                    )
                )
2530
2531
2532
2533
                req_state.mrope_positions.copy_(new_mrope_positions)
                req_state.mrope_position_delta = new_delta

            mm_embeds.extend(mm_embeds_req)
2534
2535
            req_start_idx += num_scheduled_tokens

2536
        is_mm_embed = is_mm_embed_buf.copy_to_gpu(total_num_scheduled_tokens)
2537
2538
2539

        if should_sync_mrope_positions:
            self._calc_mrope_positions(scheduler_output)
2540
            self.mrope_positions.copy_to_gpu(total_num_scheduled_tokens)
2541

2542
2543
2544
        if should_sync_xdrope_positions:
            self._calc_xdrope_positions(scheduler_output)
            self.xdrope_positions.copy_to_gpu(total_num_scheduled_tokens)
2545

2546
        return mm_embeds, is_mm_embed
2547

2548
    def get_model(self) -> nn.Module:
2549
        # get raw model out of the cudagraph wrapper.
2550
        if isinstance(self.model, (CUDAGraphWrapper, UBatchWrapper)):
2551
            return self.model.unwrap()
2552
2553
        return self.model

2554
2555
2556
2557
2558
2559
2560
2561
2562
2563
2564
2565
2566
2567
2568
    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

2569
2570
2571
2572
2573
    def get_supported_pooling_tasks(self) -> list[PoolingTask]:
        model = self.get_model()
        if not is_pooling_model(model):
            return []

2574
2575
        supported_tasks = list(model.pooler.get_supported_tasks())

2576
2577
2578
2579
        if "score" in supported_tasks:
            num_labels = getattr(self.model_config.hf_config, "num_labels", 0)
            if num_labels != 1:
                supported_tasks.remove("score")
2580
                logger.debug_once("Score API is only enabled for num_labels == 1.")
2581
2582

        return supported_tasks
2583

2584
2585
2586
2587
2588
2589
2590
2591
2592
    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)
2593

2594
    def sync_and_slice_intermediate_tensors(
2595
2596
        self,
        num_tokens: int,
2597
        intermediate_tensors: IntermediateTensors | None,
2598
2599
        sync_self: bool,
    ) -> IntermediateTensors:
2600
2601
2602
        assert self.intermediate_tensors is not None

        tp = self.vllm_config.parallel_config.tensor_parallel_size
2603
        is_rs = is_residual_scattered_for_sp(self.vllm_config, num_tokens)
2604
2605
2606
2607
2608
2609

        # 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():
2610
                is_scattered = k == "residual" and is_rs
2611
                copy_len = num_tokens // tp if is_scattered else num_tokens
2612
                self.intermediate_tensors[k][:copy_len].copy_(
2613
2614
2615
2616
2617
2618
2619
2620
2621
2622
2623
2624
2625
                    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:
2626
2627
2628
2629
2630
2631
2632
        """
        Step for the EPLB (Expert Parallelism Load Balancing) state.
        """
        if not self.parallel_config.enable_eplb:
            return

        assert self.eplb_state is not None
2633
2634
        model = self.get_model()
        assert is_mixture_of_experts(model)
2635
2636
2637
        self.eplb_state.step(
            is_dummy,
            is_profile,
2638
            log_stats=self.parallel_config.eplb_config.log_balancedness,
2639
2640
        )

2641
2642
2643
2644
2645
    def _pool(
        self,
        hidden_states: torch.Tensor,
        num_scheduled_tokens: int,
        num_scheduled_tokens_np: np.ndarray,
2646
2647
2648
2649
        kv_connector_output: KVConnectorOutput | None,
    ) -> ModelRunnerOutput | AsyncModelRunnerOutput:
        num_reqs = self.input_batch.num_reqs
        assert num_reqs == len(self.input_batch.pooling_params), (
2650
2651
            "Either all or none of the requests in a batch must be pooling request"
        )
2652

2653
        hidden_states = hidden_states[:num_scheduled_tokens]
2654
        seq_lens_cpu = self.seq_lens.cpu[:num_reqs]
2655

2656
        pooling_metadata = self.input_batch.get_pooling_metadata()
2657
        pooling_metadata.build_pooling_cursor(
2658
            num_scheduled_tokens_np, seq_lens_cpu, device=hidden_states.device
2659
        )
2660

2661
2662
        model = cast(VllmModelForPooling, self.model)
        raw_pooler_output: PoolerOutput = model.pooler(
2663
            hidden_states=hidden_states, pooling_metadata=pooling_metadata
2664
        )
2665
2666
2667
2668
2669
2670
2671
2672
2673
2674
2675
2676
2677
2678
2679
2680
2681
2682
2683
2684
2685
2686
2687
2688

        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,
            )

2689
        raw_pooler_output = json_map_leaves(
2690
            lambda x: None if x is None else x.to("cpu", non_blocking=True),
2691
2692
            raw_pooler_output,
        )
2693
2694
2695
2696
        model_runner_output.pooler_output = [
            out if include else None
            for out, include in zip(raw_pooler_output, finished_mask)
        ]
2697
        self._sync_device()
2698

2699
        return model_runner_output
2700

2701
    def _pad_for_sequence_parallelism(self, num_scheduled_tokens: int) -> int:
2702
2703
2704
        # Pad tokens to multiple of tensor_parallel_size when
        # enabled collective fusion for SP
        tp_size = self.vllm_config.parallel_config.tensor_parallel_size
2705
        if self.compilation_config.pass_config.enable_sp and tp_size > 1:
2706
2707
2708
            return round_up(num_scheduled_tokens, tp_size)
        return num_scheduled_tokens

Patrick von Platen's avatar
Patrick von Platen committed
2709
2710
2711
2712
2713
2714
2715
2716
2717
2718
2719
    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

2720
    def _preprocess(
2721
2722
        self,
        scheduler_output: "SchedulerOutput",
2723
        num_input_tokens: int,  # Padded
2724
        intermediate_tensors: IntermediateTensors | None = None,
2725
    ) -> tuple[
2726
2727
        torch.Tensor | None,
        torch.Tensor | None,
2728
        torch.Tensor,
2729
        IntermediateTensors | None,
2730
        dict[str, Any],
2731
        ECConnectorOutput | None,
2732
    ]:
2733
        num_scheduled_tokens = scheduler_output.total_num_scheduled_tokens
2734
        is_first_rank = get_pp_group().is_first_rank
2735
        is_encoder_decoder = self.model_config.is_encoder_decoder
2736

2737
2738
        # _prepare_inputs may reorder the batch, so we must gather multi
        # modal outputs after that to ensure the correct order
2739
2740
        ec_connector_output = None

2741
        if self.supports_mm_inputs and is_first_rank and not is_encoder_decoder:
2742
            # Run the multimodal encoder if any.
2743
2744
2745
2746
2747
2748
            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)
2749

2750
2751
2752
            # 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.
2753
            inputs_embeds_scheduled = self.model.embed_input_ids(
2754
2755
2756
                self.input_ids.gpu[:num_scheduled_tokens],
                multimodal_embeddings=mm_embeds,
                is_multimodal=is_mm_embed,
2757
            )
2758

2759
            # TODO(woosuk): Avoid the copy. Optimize.
2760
            self.inputs_embeds.gpu[:num_scheduled_tokens].copy_(inputs_embeds_scheduled)
2761

Patrick von Platen's avatar
Patrick von Platen committed
2762
            input_ids, inputs_embeds = self._prepare_mm_inputs(num_input_tokens)
2763
            model_kwargs = {
2764
                **self._init_model_kwargs(),
2765
2766
                **self._extract_mm_kwargs(scheduler_output),
            }
2767
        elif self.enable_prompt_embeds and is_first_rank:
2768
2769
2770
2771
2772
2773
2774
2775
2776
2777
2778
2779
            # 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).
2780
2781
2782
            token_ids_idx = (
                self.is_token_ids.gpu[:num_scheduled_tokens]
                .nonzero(as_tuple=False)
2783
                .squeeze(1)
2784
            )
2785
2786
2787
            # Some tokens ids may need to become embeds
            if token_ids_idx.numel() > 0:
                token_ids = self.input_ids.gpu[token_ids_idx]
2788
                tokens_to_embeds = self.model.embed_input_ids(input_ids=token_ids)
2789
2790
2791
                self.inputs_embeds.gpu[token_ids_idx] = tokens_to_embeds

            inputs_embeds = self.inputs_embeds.gpu[:num_input_tokens]
2792
            model_kwargs = self._init_model_kwargs()
2793
            input_ids = None
2794
        else:
2795
2796
2797
2798
            # 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.
2799
            input_ids = self.input_ids.gpu[:num_input_tokens]
2800
            inputs_embeds = None
2801
            model_kwargs = self._init_model_kwargs()
2802

2803
        if self.uses_mrope:
2804
            positions = self.mrope_positions.gpu[:, :num_input_tokens]
2805
2806
        elif self.uses_xdrope_dim > 0:
            positions = self.xdrope_positions.gpu[:, :num_input_tokens]
2807
        else:
2808
            positions = self.positions.gpu[:num_input_tokens]
2809

2810
        if is_first_rank:
2811
2812
            intermediate_tensors = None
        else:
2813
            assert intermediate_tensors is not None
2814
            intermediate_tensors = self.sync_and_slice_intermediate_tensors(
2815
2816
                num_input_tokens, intermediate_tensors, True
            )
2817

2818
        if is_encoder_decoder and scheduler_output.scheduled_encoder_inputs:
2819
2820
2821
2822
2823
2824
2825
            # 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})
2826

2827
2828
2829
2830
2831
2832
        return (
            input_ids,
            inputs_embeds,
            positions,
            intermediate_tensors,
            model_kwargs,
2833
            ec_connector_output,
2834
        )
2835

2836
    def _sample(
2837
        self,
2838
2839
        logits: torch.Tensor | None,
        spec_decode_metadata: SpecDecodeMetadata | None,
2840
    ) -> SamplerOutput:
2841
        # Sample the next token and get logprobs if needed.
2842
        sampling_metadata = self.input_batch.sampling_metadata
2843
2844
2845
        # 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()
2846
        if spec_decode_metadata is None:
2847
            return self.sampler(
2848
2849
2850
                logits=logits,
                sampling_metadata=sampling_metadata,
            )
2851

2852
2853
2854
2855
2856
2857
        # 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)

2858
        sampler_output = self.rejection_sampler(
2859
            spec_decode_metadata,
王敏's avatar
王敏 committed
2860
2861
            None if self.draft_probs is None else \
                self.draft_probs.get_probs(spec_decode_metadata.spec_decode_ids),  # draft_probs
2862
            logits,
2863
2864
            sampling_metadata,
        )
2865
2866
2867
        return sampler_output

    def _bookkeeping_sync(
2868
2869
2870
        self,
        scheduler_output: "SchedulerOutput",
        sampler_output: SamplerOutput,
2871
        logits: torch.Tensor | None,
2872
2873
        hidden_states: torch.Tensor,
        num_scheduled_tokens: int,
2874
        spec_decode_metadata: SpecDecodeMetadata | None,
2875
    ) -> tuple[
2876
        dict[str, int],
2877
        LogprobsLists | None,
2878
        list[list[int]],
2879
        dict[str, LogprobsTensors | None],
2880
2881
2882
        list[str],
        dict[str, int],
        list[int],
2883
    ]:
2884
2885
2886
2887
        num_nans_in_logits = {}
        if envs.VLLM_COMPUTE_NANS_IN_LOGITS:
            num_nans_in_logits = self._get_nans_in_logits(logits)

2888
2889
2890
2891
        num_reqs = self.input_batch.num_reqs
        discard_sampled_tokens_req_indices = np.nonzero(
            self.discard_request_mask.np[:num_reqs]
        )[0]
2892
2893
2894
2895
        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)
2896

2897
2898
2899
        # 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()
2900
        req_id_to_index_output_copy = self.input_batch.req_id_to_index.copy()
2901

2902
        num_sampled_tokens = sampler_output.sampled_token_ids.shape[0]
2903
        sampled_token_ids = sampler_output.sampled_token_ids
2904
        logprobs_tensors = sampler_output.logprobs_tensors
2905
        invalid_req_indices = []
2906
        logprobs_lists = None
2907
2908
2909
2910
2911
2912
        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)
2913
2914
2915
                # 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()
2916
2917
2918

                if logprobs_tensors is not None:
                    logprobs_lists = logprobs_tensors.tolists()
2919
2920
            else:
                # Includes spec decode tokens.
2921
                valid_sampled_token_ids, logprobs_lists = RejectionSampler.parse_output(
2922
2923
                    sampled_token_ids,
                    self.input_batch.vocab_size,
2924
                    discard_sampled_tokens_req_indices,
2925
                    logprobs_tensors=logprobs_tensors,
2926
                )
2927
        else:
2928
            valid_sampled_token_ids = []
2929
            invalid_req_indices = discard_sampled_tokens_req_indices.tolist()
2930
2931
2932
2933
2934
            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.
2935
2936
2937
2938
            # 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
2939
2940
2941
2942
2943
            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
            }
2944

2945
2946
2947
2948
2949
        # 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.
2950
        req_ids = self.input_batch.req_ids
2951
2952
        for req_idx in range(num_sampled_tokens):
            if self.use_async_scheduling:
2953
                sampled_ids = [-1] if req_idx not in invalid_req_indices_set else None
2954
2955
            else:
                sampled_ids = valid_sampled_token_ids[req_idx]
2956

2957
            num_sampled_ids: int = len(sampled_ids) if sampled_ids else 0
2958

2959
2960
2961
2962
            if not sampled_ids:
                continue

            start_idx = self.input_batch.num_tokens_no_spec[req_idx]
2963
            end_idx = start_idx + num_sampled_ids
2964
2965
2966
            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: "
2967
                f"{self.max_model_len}"
2968
            )
2969

2970
            self.input_batch.token_ids_cpu[req_idx, start_idx:end_idx] = sampled_ids
2971
            self.input_batch.is_token_ids[req_idx, start_idx:end_idx] = True
2972
            self.input_batch.num_tokens_no_spec[req_idx] = end_idx
2973

2974
            req_id = req_ids[req_idx]
2975
2976
2977
            req_state = self.requests[req_id]
            req_state.output_token_ids.extend(sampled_ids)

2978
2979
2980
2981
2982
2983
        # Compute prompt logprobs if needed.
        prompt_logprobs_dict = self._get_prompt_logprobs_dict(
            hidden_states[:num_scheduled_tokens],
            scheduler_output.num_scheduled_tokens,
        )

2984
2985
2986
2987
2988
2989
2990
2991
2992
2993
        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,
        )

2994
2995
2996
2997
2998
2999
3000
3001
3002
3003
3004
3005
3006
3007
3008
    @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()

3009
3010
    def _model_forward(
        self,
3011
3012
3013
3014
        input_ids: torch.Tensor | None = None,
        positions: torch.Tensor | None = None,
        intermediate_tensors: IntermediateTensors | None = None,
        inputs_embeds: torch.Tensor | None = None,
3015
3016
3017
3018
3019
        **model_kwargs: dict[str, Any],
    ) -> Any:
        """Helper method to call the model forward pass.

        This method can be overridden by subclasses for model execution.
3020
        Motivation: We can inspect only this method versus
3021
3022
3023
3024
3025
3026
3027
3028
3029
3030
3031
3032
3033
3034
3035
3036
3037
3038
3039
3040
        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,
        )

3041
3042
3043
3044
3045
3046
3047
3048
3049
3050
3051
3052
3053
3054
3055
3056
3057
3058
3059
3060
3061
    @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
        )

3062
3063
3064
3065
3066
3067
3068
3069
3070
3071
3072
3073
3074
    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,
3075
        num_encoder_reqs: int = 0,
3076
    ) -> tuple[
3077
3078
        CUDAGraphMode,
        BatchDescriptor,
3079
        bool,
3080
3081
        torch.Tensor | None,
        CUDAGraphStat | None,
3082
    ]:
3083
3084
3085
3086
3087
3088
        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,
3089
        )
3090
3091
3092
3093
3094
        # 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
        )
3095
3096
3097
3098
3099
3100
3101

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

3102
        num_tokens_padded = self._pad_for_sequence_parallelism(num_tokens)
3103
        dispatch_cudagraph = (
3104
            lambda num_tokens, disable_full: self.cudagraph_dispatcher.dispatch(
3105
3106
3107
                num_tokens=num_tokens,
                has_lora=has_lora,
                uniform_decode=uniform_decode,
3108
                disable_full=disable_full,
3109
3110
3111
3112
3113
            )
            if not force_eager
            else (CUDAGraphMode.NONE, BatchDescriptor(num_tokens_padded))
        )

3114
        cudagraph_mode, batch_descriptor = dispatch_cudagraph(
3115
            num_tokens_padded, use_cascade_attn or has_encoder_output
3116
        )
3117
        num_tokens_padded = batch_descriptor.num_tokens
3118
3119
3120
3121
3122
3123
3124
3125
3126
        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"
            )
3127
3128
3129

        # Extra coordination when running data-parallel since we need to coordinate
        # across ranks
3130
        should_ubatch, num_tokens_across_dp = False, None
3131
3132
3133
3134
3135
3136
3137
3138
3139
        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
            )

3140
3141
3142
3143
3144
3145
3146
3147
3148
3149
3150
            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,
                )
3151
3152
            )

3153
            # Extract DP-synced values
3154
3155
3156
            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())
3157
3158
3159
3160
3161
                # 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,
                )
3162
3163
3164
3165
                # 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

3166
3167
3168
3169
3170
3171
3172
3173
3174
3175
3176
3177
        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,
3178
            should_ubatch,
3179
3180
3181
            num_tokens_across_dp,
            cudagraph_stats,
        )
3182

3183
3184
3185
3186
3187
3188
3189
3190
3191
3192
3193
3194
3195
3196
3197
3198
3199
3200
3201
3202
3203
3204
3205
3206
3207
3208
3209
3210
3211
3212
3213
3214
3215
3216
3217
    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
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
3254
3255
3256
3257
3258
3259
3260
3261
3262
3263
3264
3265
3266
3267
3268
3269
3270
3271
3272
3273
3274
3275
3276
3277
3278
3279
3280
3281
3282
3283
3284
3285
3286
3287
3288
3289
3290
3291
3292
    def _get_slot_mappings(
        self,
        num_tokens_padded: int,
        num_reqs_padded: int,
        num_tokens_unpadded: int,
        ubatch_slices: "UBatchSlices | None" = None,
    ) -> tuple[
        dict[int, torch.Tensor] | None,
        dict[str, torch.Tensor] | list[dict[str, torch.Tensor]] | None,
    ]:
        """
        Build slot mappings in both formats needed by the system.

        Args:
            num_tokens_padded: Total number of tokens (padded)
            num_reqs_padded: Total number of requests (padded)
            num_tokens_unpadded: Actual number of tokens (unpadded)
            ubatch_slices: Optional ubatch slicing info for DBO

        Returns:
            A tuple of:
            - slot_mappings_by_gid: dict[int, torch.Tensor] for attention metadata
            - slot_mappings_by_layer: dict[str, torch.Tensor] or list for ForwardContext
        """
        if not (
            hasattr(self, "kv_cache_config")
            and self.kv_cache_config is not None
            and len(self.kv_cache_config.kv_cache_groups) > 0
        ):
            return None, None

        def _get_slot_mapping(kv_cache_gid: int):
            assert num_reqs_padded is not None and num_tokens_padded is not None
            kv_cache_spec = self.kv_cache_config.kv_cache_groups[
                kv_cache_gid
            ].kv_cache_spec
            if isinstance(kv_cache_spec, EncoderOnlyAttentionSpec):
                slot_mapping = torch.zeros(
                    (num_tokens_padded,),
                    dtype=torch.int64,
                    device=self.device,
                )
            else:
                blk_table = self.input_batch.block_table[kv_cache_gid]
                slot_mapping = blk_table.slot_mapping.gpu[:num_tokens_padded]

            # 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_unpadded:num_tokens_padded].fill_(-1)

            return slot_mapping

        slot_mappings_by_gid = {
            gid: _get_slot_mapping(gid)
            for gid, _ in enumerate(self.kv_cache_config.kv_cache_groups)
        }

        slot_mappings_by_layer: dict[str, torch.Tensor] = {}
        for gid, kv_cache_group in enumerate(self.kv_cache_config.kv_cache_groups):
            slot_mapping = slot_mappings_by_gid[gid]
            for layer_name in kv_cache_group.layer_names:
                slot_mappings_by_layer[layer_name] = slot_mapping

        if ubatch_slices is not None:
            result: list[dict[str, torch.Tensor]] = []
            for ubatch in ubatch_slices:
                sliced_mappings: dict[str, torch.Tensor] = {}
                for layer_name, slot_mapping in slot_mappings_by_layer.items():
                    sliced_mappings[layer_name] = slot_mapping[ubatch.token_slice]
                result.append(sliced_mappings)
            return slot_mappings_by_gid, result

        return slot_mappings_by_gid, slot_mappings_by_layer

3293
3294
3295
3296
    @torch.inference_mode()
    def execute_model(
        self,
        scheduler_output: "SchedulerOutput",
3297
        intermediate_tensors: IntermediateTensors | None = None,
3298
    ) -> ModelRunnerOutput | AsyncModelRunnerOutput | IntermediateTensors | None:
3299
3300
3301
3302
3303
        if self.execute_model_state is not None:
            raise RuntimeError(
                "State error: sample_tokens() must be called "
                "after execute_model() returns None."
            )
3304

3305
3306
3307
3308
3309
3310
        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.")
3311

3312
3313
3314
3315
        if scheduler_output.preempted_req_ids and has_kv_transfer_group():
            get_kv_transfer_group().handle_preemptions(
                scheduler_output.preempted_req_ids
            )
3316

3317
        num_scheduled_tokens = scheduler_output.total_num_scheduled_tokens
3318
3319
3320
3321
3322
3323
        with (
            record_function_or_nullcontext("gpu_model_runner: preprocess"),
            self.synchronize_input_prep(),
        ):
            # Update persistent batch states.
            self._update_states(scheduler_output)
3324

3325
3326
            if has_ec_transfer() and get_ec_transfer().is_producer:
                with self.maybe_get_ec_connector_output(
3327
                    scheduler_output,
3328
3329
3330
3331
3332
3333
3334
3335
3336
3337
3338
3339
3340
3341
3342
3343
3344
3345
3346
3347
3348
3349
3350
3351
3352
3353
3354
3355
                    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"
3356
                )
3357

3358
3359
3360
3361
3362
3363
            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
3364

3365
3366
3367
3368
            logits_indices, spec_decode_metadata = self._prepare_inputs(
                scheduler_output,
                num_scheduled_tokens_np,
            )
3369

3370
3371
3372
3373
3374
            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(
3375
                    num_scheduled_tokens_np,
3376
3377
                    self.input_batch.num_computed_tokens_cpu[:num_reqs],
                    scheduler_output.num_common_prefix_blocks,
3378
                )
3379

3380
3381
3382
3383
3384
3385
3386
3387
3388
3389
3390
3391
3392
3393
            (
                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),
            )
3394

3395
3396
3397
3398
3399
3400
3401
3402
3403
3404
3405
3406
3407
3408
3409
3410
3411
3412
3413
3414
3415
3416
3417
3418
3419
3420
3421
            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,
            )

3422
3423
3424
3425
3426
3427
3428
3429
3430
3431
3432
            # True if any attention backend handles KV cache update separately
            # from forward() (i.e., forward_includes_kv_cache_update=False). When true,
            # slot_mappings must use padded dimensions to match the key/value tensors.
            has_separate_kv_update = not all(
                all(
                    g.backend.forward_includes_kv_cache_update
                    for g in self.attn_groups[id]
                )
                for id, spec in enumerate(self.kv_cache_config.kv_cache_groups)
                if not isinstance(spec.kv_cache_spec, EncoderOnlyAttentionSpec)
            )
3433
3434
            pad_attn = cudagraph_mode == CUDAGraphMode.FULL

3435
3436
3437
3438
3439
3440
3441
3442
3443
3444
3445
3446
            if self.cache_config.mamba_cache_mode == "align":
                mamba_utils.preprocess_mamba(
                    scheduler_output,
                    self.kv_cache_config,
                    self.cache_config,
                    self.mamba_state_idx,
                    self.input_batch,
                    self.requests,
                    self.compilation_config.static_forward_context,
                    self.model.get_mamba_state_copy_func(),
                )

3447
3448
3449
            use_spec_decode = len(scheduler_output.scheduled_spec_decode_tokens) > 0
            ubatch_slices_attn = ubatch_slices_padded if pad_attn else ubatch_slices

3450
3451
3452
3453
3454
3455
3456
3457
3458
3459
3460
            slot_mappings_by_group, slot_mappings = self._get_slot_mappings(
                num_tokens_padded=num_tokens_padded
                if pad_attn or has_separate_kv_update
                else num_tokens_unpadded,
                num_reqs_padded=(
                    num_reqs_padded if pad_attn or has_separate_kv_update else num_reqs
                ),
                num_tokens_unpadded=num_tokens_unpadded,
                ubatch_slices=ubatch_slices_padded,
            )

3461
3462
3463
3464
3465
3466
3467
3468
3469
3470
3471
3472
            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,
3473
                    slot_mappings=slot_mappings_by_group,
3474
                )
3475
            )
3476
3477
3478
3479
3480
3481
3482

            (
                input_ids,
                inputs_embeds,
                positions,
                intermediate_tensors,
                model_kwargs,
3483
3484
3485
                ec_connector_output,
            ) = self._preprocess(
                scheduler_output, num_tokens_padded, intermediate_tensors
3486
            )
3487

3488
        # Set cudagraph mode to none if calc_kv_scales is true.
3489
3490
3491
        # KV scales calculation involves dynamic operations that are incompatible
        # with CUDA graph capture.
        if self.calculate_kv_scales:
3492
            cudagraph_mode = CUDAGraphMode.NONE
3493
3494
            # Mark KV scales as calculated after the first forward pass
            self.calculate_kv_scales = False
3495

3496
3497
3498
3499
3500
3501
3502
        # Encoder-decoder models can only compile the pure decode steps where no
        # encoder inputs are present. Use eager for the first pass.
        num_encoder_reqs = len(scheduler_output.scheduled_encoder_inputs)
        has_encoder_input = (
            self.model_config.is_encoder_decoder and num_encoder_reqs > 0
        )

3503
3504
        # Run the model.
        # Use persistent buffers for CUDA graphs.
3505
3506
        with (
            set_forward_context(
3507
3508
                attn_metadata,
                self.vllm_config,
3509
                num_tokens=num_tokens_padded,
3510
                num_tokens_across_dp=num_tokens_across_dp,
3511
3512
                cudagraph_runtime_mode=cudagraph_mode,
                batch_descriptor=batch_desc,
3513
                ubatch_slices=ubatch_slices_padded,
3514
                slot_mapping=slot_mappings,
3515
                skip_compiled=has_encoder_input,
3516
            ),
3517
            record_function_or_nullcontext("gpu_model_runner: forward"),
3518
3519
            self.maybe_get_kv_connector_output(scheduler_output) as kv_connector_output,
        ):
3520
            model_output = self._model_forward(
3521
3522
3523
3524
3525
3526
3527
                input_ids=input_ids,
                positions=positions,
                intermediate_tensors=intermediate_tensors,
                inputs_embeds=inputs_embeds,
                **model_kwargs,
            )

3528
        with record_function_or_nullcontext("gpu_model_runner: postprocess"):
3529
            if self.use_aux_hidden_state_outputs:
3530
                # True when EAGLE 3 is used.
3531
3532
                hidden_states, aux_hidden_states = model_output
            else:
3533
                # Common case.
3534
3535
3536
                hidden_states = model_output
                aux_hidden_states = None

3537
3538
3539
3540
3541
            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)
3542
                    hidden_states.kv_connector_output = kv_connector_output
3543
                    self.kv_connector_output = kv_connector_output
3544
                    return hidden_states
3545

3546
                if self.is_pooling_model:
3547
                    # Return the pooling output.
3548
3549
3550
3551
3552
                    return self._pool(
                        hidden_states,
                        num_scheduled_tokens,
                        num_scheduled_tokens_np,
                        kv_connector_output,
3553
                    )
3554
3555

                sample_hidden_states = hidden_states[logits_indices]
3556
                logits = self.model.compute_logits(sample_hidden_states)
3557
3558
3559
3560
            else:
                # Rare case.
                assert not self.is_pooling_model

3561
                sample_hidden_states = hidden_states[logits_indices]
3562
                if not get_pp_group().is_last_rank:
3563
                    all_gather_tensors = {
3564
                        "residual": not is_residual_scattered_for_sp(
3565
                            self.vllm_config, num_tokens_padded
3566
                        )
3567
                    }
3568
                    get_pp_group().send_tensor_dict(
3569
3570
                        hidden_states.tensors,
                        all_gather_group=get_tp_group(),
3571
3572
                        all_gather_tensors=all_gather_tensors,
                    )
3573
3574
                    logits = None
                else:
3575
                    logits = self.model.compute_logits(sample_hidden_states)
3576

3577
                model_output_broadcast_data: dict[str, Any] = {}
3578
3579
3580
                if logits is not None:
                    model_output_broadcast_data["logits"] = logits.contiguous()

3581
                broadcasted = get_pp_group().broadcast_tensor_dict(
3582
3583
                    model_output_broadcast_data, src=len(get_pp_group().ranks) - 1
                )
3584
3585
                assert broadcasted is not None
                logits = broadcasted["logits"]
3586

3587
3588
3589
3590
3591
3592
3593
3594
        self.execute_model_state = ExecuteModelState(
            scheduler_output,
            logits,
            spec_decode_metadata,
            spec_decode_common_attn_metadata,
            hidden_states,
            sample_hidden_states,
            aux_hidden_states,
3595
            ec_connector_output,
3596
            cudagraph_stats,
3597
            slot_mappings,
3598
        )
3599
        self.kv_connector_output = kv_connector_output
3600
3601
3602
3603
3604
3605
        return None

    @torch.inference_mode
    def sample_tokens(
        self, grammar_output: "GrammarOutput | None"
    ) -> ModelRunnerOutput | AsyncModelRunnerOutput | IntermediateTensors:
3606
3607
3608
        kv_connector_output = self.kv_connector_output
        self.kv_connector_output = None

3609
3610
        if self.execute_model_state is None:
            # Nothing to do (PP non-final rank case), output isn't used.
3611
            if not kv_connector_output:
3612
                return None  # type: ignore[return-value]
3613
3614
3615
3616
3617
3618
3619
3620
3621

            # 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
3622

3623
3624
3625
3626
3627
3628
3629
3630
3631
        # Unpack ephemeral state.
        (
            scheduler_output,
            logits,
            spec_decode_metadata,
            spec_decode_common_attn_metadata,
            hidden_states,
            sample_hidden_states,
            aux_hidden_states,
3632
            ec_connector_output,
3633
            cudagraph_stats,
3634
            slot_mappings,
3635
3636
3637
3638
3639
3640
3641
3642
3643
        ) = 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
            )
3644

3645
        with record_function_or_nullcontext("gpu_model_runner: sample"):
3646
3647
            sampler_output = self._sample(logits, spec_decode_metadata)

3648
3649
3650
3651
        self._update_states_after_model_execute(
            sampler_output.sampled_token_ids, scheduler_output
        )

3652
3653
        self._draft_token_ids = None
        self._draft_token_req_ids = None
3654
3655
        self.input_batch.prev_sampled_token_ids = None

3656
3657
        def propose_draft_token_ids(sampled_token_ids):
            assert spec_decode_common_attn_metadata is not None
3658
            with record_function_or_nullcontext("gpu_model_runner: draft"):
3659
3660
3661
3662
3663
3664
3665
3666
3667
                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,
3668
                    slot_mappings,
3669
                )
3670
                self._copy_draft_token_ids_to_cpu(scheduler_output)
3671

3672
        spec_config = self.speculative_config
3673
3674
3675
3676
3677
        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
3678
            )
3679
3680
3681
3682
3683
            use_gpu_toks = (
                spec_config.use_eagle() or spec_config.uses_draft_model()
            ) and not spec_config.disable_padded_drafter_batch
            if use_gpu_toks:
                # EAGLE/DraftModel speculative decoding can use the GPU sampled tokens
3684
                # as inputs, and does not need to wait for bookkeeping to finish.
3685
                assert isinstance(self.drafter, EagleProposer | DraftModelProposer)
3686
3687
3688
3689
3690
3691
3692
3693
3694
3695
3696
3697
3698
                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,
                        )
3699
                    )
3700
3701
3702
3703
3704
3705
3706
3707
3708
3709
3710
                    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
3711

3712
        with record_function_or_nullcontext("gpu_model_runner: bookkeep"):
3713
3714
3715
3716
3717
3718
3719
3720
            (
                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,
3721
3722
3723
3724
3725
            ) = self._bookkeeping_sync(
                scheduler_output,
                sampler_output,
                logits,
                hidden_states,
3726
                scheduler_output.total_num_scheduled_tokens,
3727
                spec_decode_metadata,
3728
            )
3729

3730
        if propose_drafts_after_bookkeeping:
3731
3732
3733
            # 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)
3734

3735
        with record_function_or_nullcontext("gpu_model_runner: eplb"):
3736
            self.eplb_step()
3737

3738
        with record_function_or_nullcontext("gpu_model_runner: ModelRunnerOutput"):
3739
3740
3741
3742
3743
3744
3745
            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.")

3746
3747
3748
3749
3750
3751
3752
            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,
3753
3754
3755
                ec_connector_output=ec_connector_output
                if self.supports_mm_inputs
                else None,
3756
                num_nans_in_logits=num_nans_in_logits,
3757
                cudagraph_stats=cudagraph_stats,
3758
            )
3759

3760
3761
        if not self.use_async_scheduling:
            return output
3762

3763
3764
3765
3766
3767
3768
3769
3770
3771
        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,
3772
                vocab_size=self.input_batch.vocab_size,
3773
3774
3775
3776
3777
            )
        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
3778
            # any requests with sampling params that require output ids.
3779
3780
3781
3782
            self.input_batch.set_async_sampled_token_ids(
                async_output.sampled_token_ids_cpu,
                async_output.async_copy_ready_event,
            )
3783

3784
        return async_output
3785

3786
    def take_draft_token_ids(self) -> DraftTokenIds | None:
3787
        if not self.num_spec_tokens or not self._draft_token_req_ids:
3788
            return None
3789
        draft_token_ids, req_ids = self._get_draft_token_ids_cpu()
3790
3791
        return DraftTokenIds(req_ids, draft_token_ids)

3792
3793
3794
    def _copy_draft_token_ids_to_cpu(
        self, scheduler_output: "SchedulerOutput", zeros_only: bool = False
    ) -> None:
3795
3796
3797
3798
3799
3800
        # 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
        ):
3801
3802
3803
            return
        # We must also set the corresponding request ids.
        self._draft_token_req_ids = self.input_batch.req_ids.copy()
3804

3805
3806
3807
3808
3809
3810
3811
3812
3813
3814
3815
3816
3817
3818
3819
3820
3821
3822
3823
3824
        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()

3825
    def _get_draft_token_ids_cpu(self) -> tuple[list[list[int]], list[str]]:
3826
        if isinstance(self._draft_token_ids, list):
3827
3828
3829
3830
            return self._draft_token_ids, self.input_batch.req_ids
        req_ids = self._draft_token_req_ids
        if req_ids is None:
            return [], []
3831
3832
3833
        assert self.draft_token_ids_event is not None
        assert self.draft_token_ids_cpu is not None
        self.draft_token_ids_event.synchronize()
3834
        return self.draft_token_ids_cpu[: len(req_ids)].tolist(), req_ids
3835

3836
3837
3838
3839
3840
3841
3842
3843
3844
3845
3846
3847
3848
    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
3849
            assert counts_cpu is not None
3850
3851
3852
3853
3854
3855
3856
3857
            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
3858
3859
        sampled_count_event = self.valid_sampled_token_count_event
        if sampled_count_event is None or prev_sampled_token_ids is None:
3860
3861
3862
            return []

        counts_cpu = self.valid_sampled_token_count_cpu
3863
3864
        assert counts_cpu is not None
        sampled_count_event.synchronize()
3865
3866
        return counts_cpu[: prev_sampled_token_ids.shape[0]].tolist()

3867
3868
3869
    def propose_draft_token_ids(
        self,
        scheduler_output: "SchedulerOutput",
3870
        sampled_token_ids: torch.Tensor | list[list[int]],
3871
3872
3873
        sampling_metadata: SamplingMetadata,
        hidden_states: torch.Tensor,
        sample_hidden_states: torch.Tensor,
3874
3875
        aux_hidden_states: list[torch.Tensor] | None,
        spec_decode_metadata: SpecDecodeMetadata | None,
3876
        common_attn_metadata: CommonAttentionMetadata,
3877
        slot_mappings: dict[str, torch.Tensor] | list[dict[str, torch.Tensor]] | None,
3878
    ) -> list[list[int]] | torch.Tensor:
3879
        num_scheduled_tokens = scheduler_output.total_num_scheduled_tokens
3880
3881
3882
        spec_config = self.speculative_config
        assert spec_config is not None
        if spec_config.method == "ngram":
3883
            assert isinstance(sampled_token_ids, list)
3884
            assert isinstance(self.drafter, NgramProposer)
3885
            draft_token_ids = self.drafter.propose(
3886
                sampled_token_ids,
3887
3888
                self.input_batch.num_tokens_no_spec,
                self.input_batch.token_ids_cpu,
3889
                slot_mappings=slot_mappings,
3890
            )
3891
        elif spec_config.method == "suffix":
3892
3893
            assert isinstance(sampled_token_ids, list)
            assert isinstance(self.drafter, SuffixDecodingProposer)
3894
3895
3896
            draft_token_ids = self.drafter.propose(
                self.input_batch, sampled_token_ids, slot_mappings=slot_mappings
            )
3897
        elif spec_config.method == "medusa":
3898
            assert isinstance(sampled_token_ids, list)
3899
            assert isinstance(self.drafter, MedusaProposer)
3900

3901
3902
            if sample_hidden_states.shape[0] == len(sampled_token_ids):
                # The input to the target model does not include draft tokens.
3903
3904
3905
3906
                hidden_states = sample_hidden_states
            else:
                indices = []
                offset = 0
3907
3908
3909
                assert spec_decode_metadata is not None, (
                    "No spec decode metadata for medusa"
                )
3910
                for num_draft, tokens in zip(
3911
3912
                    spec_decode_metadata.num_draft_tokens, sampled_token_ids
                ):
3913
3914
                    indices.append(offset + len(tokens) - 1)
                    offset += num_draft + 1
3915
                indices = torch.tensor(indices, device=self.device)
3916
3917
                hidden_states = sample_hidden_states[indices]

3918
            draft_token_ids = self.drafter.propose(
3919
3920
                target_hidden_states=hidden_states,
                sampling_metadata=sampling_metadata,
3921
                slot_mappings=slot_mappings,
3922
            )
3923
3924
        elif spec_config.use_eagle() or spec_config.uses_draft_model():
            assert isinstance(self.drafter, EagleProposer | DraftModelProposer)
3925

3926
            if spec_config.disable_padded_drafter_batch:
3927
3928
3929
                # 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.
3930
3931
                assert isinstance(sampled_token_ids, list), (
                    "sampled_token_ids should be a python list when"
3932
                    "padded-batch is disabled."
3933
                )
3934
                next_token_ids = self.drafter.prepare_next_token_ids_cpu(
3935
3936
3937
3938
3939
                    sampled_token_ids,
                    self.requests,
                    self.input_batch,
                    scheduler_output.num_scheduled_tokens,
                )
3940
3941
3942
3943
3944
            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.
3945
3946
                assert isinstance(sampled_token_ids, torch.Tensor), (
                    "sampled_token_ids should be a torch.Tensor when"
3947
                    "padded-batch is enabled."
3948
3949
                )
                next_token_ids, valid_sampled_tokens_count = (
3950
3951
3952
3953
3954
                    self.drafter.prepare_next_token_ids_padded(
                        common_attn_metadata,
                        sampled_token_ids,
                        self.requests,
                        self.input_batch,
3955
                        self.discard_request_mask.gpu,
3956
                    )
3957
                )
3958
3959
3960
                self._copy_valid_sampled_token_count(
                    next_token_ids, valid_sampled_tokens_count
                )
Jiayi Yao's avatar
Jiayi Yao committed
3961

3962
            num_rejected_tokens_gpu = None
3963
            if spec_decode_metadata is None:
3964
                token_indices_to_sample = None
3965
                # input_ids can be None for multimodal models.
3966
                target_token_ids = self.input_ids.gpu[:num_scheduled_tokens]
3967
                target_positions = self._get_positions(num_scheduled_tokens)
3968
                if self.use_aux_hidden_state_outputs:
Wentao Ye's avatar
Wentao Ye committed
3969
                    assert aux_hidden_states is not None
3970
                    target_hidden_states = torch.cat(
3971
3972
                        [h[:num_scheduled_tokens] for h in aux_hidden_states], dim=-1
                    )
3973
3974
                else:
                    target_hidden_states = hidden_states[:num_scheduled_tokens]
3975
            else:
3976
                if spec_config.disable_padded_drafter_batch:
3977
                    token_indices_to_sample = None
3978
3979
3980
3981
3982
                    common_attn_metadata, token_indices = self.drafter.prepare_inputs(
                        common_attn_metadata,
                        sampled_token_ids,
                        spec_decode_metadata.num_draft_tokens,
                    )
3983
3984
3985
3986
3987
3988
3989
3990
3991
                    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]
3992
                else:
3993
3994
3995
3996
3997
3998
3999
4000
                    (
                        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,
4001
                    )
4002
4003
4004
4005
4006
4007
4008
4009
4010
4011
4012
                    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]
4013

4014
            if self.supports_mm_inputs:
4015
4016
4017
4018
4019
4020
                mm_embed_inputs = self._gather_mm_embeddings(
                    scheduler_output,
                    shift_computed_tokens=1,
                )
            else:
                mm_embed_inputs = None
4021

王敏's avatar
王敏 committed
4022
            draft_result = self.drafter.propose(
4023
4024
4025
4026
                target_token_ids=target_token_ids,
                target_positions=target_positions,
                target_hidden_states=target_hidden_states,
                next_token_ids=next_token_ids,
4027
                last_token_indices=token_indices_to_sample,
4028
                sampling_metadata=sampling_metadata,
4029
                common_attn_metadata=common_attn_metadata,
4030
                mm_embed_inputs=mm_embed_inputs,
4031
                num_rejected_tokens_gpu=num_rejected_tokens_gpu,
4032
                slot_mappings=slot_mappings,
4033
            )
4034

王敏's avatar
王敏 committed
4035
4036
4037
4038
4039
4040
4041
4042
4043
4044
4045
4046
4047
            if not envs.VLLM_REJECT_SAMPLE_OPT:
                draft_token_ids = draft_result
            else:
                draft_token_ids, draft_probs = draft_result

            if envs.VLLM_REJECT_SAMPLE_OPT:
                draft_req_ids = list(scheduler_output.num_scheduled_tokens.keys())
                if self.draft_probs is None:
                    self.draft_probs = DraftProbs(
                        draft_probs, draft_req_ids)
                else:
                    self.draft_probs.update(draft_probs, draft_req_ids)

4048
        return draft_token_ids
4049

4050
4051
4052
    def update_config(self, overrides: dict[str, Any]) -> None:
        allowed_config_names = {"load_config", "model_config"}
        for config_name, config_overrides in overrides.items():
4053
4054
            assert config_name in allowed_config_names, (
                f"Config `{config_name}` not supported. "
4055
                f"Allowed configs: {allowed_config_names}"
4056
            )
4057
4058
4059
4060
            config = getattr(self, config_name)
            new_config = update_config(config, config_overrides)
            setattr(self, config_name, new_config)

4061
4062
4063
4064
4065
    def load_model(self, eep_scale_up: bool = False) -> None:
        """
        Args:
            eep_scale_up: the model loading is for elastic EP scale up.
        """
4066
4067
4068
4069
4070
        logger.info_once(
            "Starting to load model %s...",
            self.model_config.model,
            scope="global",
        )
4071
4072
4073
4074
4075
        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)
        )
4076

4077
4078
4079
4080
        if self.parallel_config.enable_eplb:
            self.eplb_state = EplbState(self.parallel_config, self.device)
            eplb_models = 0

4081
4082
4083
4084
4085
4086
        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
4087
                )
4088
4089
4090
                if self.lora_config:
                    self.model = self.load_lora_model(
                        self.model, self.vllm_config, self.device
4091
                    )
4092
4093
4094
4095
4096
4097
4098
4099
4100
4101
4102
4103
4104
4105
4106
                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,
                        )
4107

4108
4109
4110
4111
4112
4113
4114
4115
4116
4117
4118
4119
4120
4121
4122
4123
4124
4125
4126
4127
4128
4129
                        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
4130

4131
4132
4133
4134
4135
4136
                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"
                        )
4137

4138
4139
4140
4141
4142
4143
4144
4145
4146
4147
                    # 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()
4148

4149
4150
4151
4152
4153
4154
4155
4156
4157
4158
4159
4160
4161
4162
                    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
4163
        logger.info_once(
4164
4165
            "Model loading took %s GiB memory and %.6f seconds",
            format_gib(self.model_memory_usage),
4166
            time_after_load - time_before_load,
4167
            scope="local",
4168
        )
4169
        prepare_communication_buffer_for_model(self.model)
4170
4171
4172
4173
        if (drafter := getattr(self, "drafter", None)) and (
            drafter_model := getattr(drafter, "model", None)
        ):
            prepare_communication_buffer_for_model(drafter_model)
4174
        mm_config = self.model_config.multimodal_config
4175
        self.is_multimodal_pruning_enabled = (
4176
            supports_multimodal_pruning(self.get_model())
4177
4178
            and mm_config is not None
            and mm_config.is_multimodal_pruning_enabled()
4179
        )
4180

4181
        if is_mixture_of_experts(self.model) and self.parallel_config.enable_eplb:
4182
4183
4184
4185
4186
4187
4188
4189
4190
4191
4192
            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(
4193
                self.model,
4194
                self.model_config,
4195
4196
4197
                global_expert_load,
                old_global_expert_indices,
                rank_mapping,
4198
            )
4199
4200
            if self.eplb_state.is_async:
                self.eplb_state.start_async_loop(rank_mapping=rank_mapping)
4201

4202
        if (
4203
4204
            self.vllm_config.compilation_config.mode
            == CompilationMode.STOCK_TORCH_COMPILE
4205
        ):
4206
            backend = self.vllm_config.compilation_config.init_backend(self.vllm_config)
4207
            compilation_counter.stock_torch_compile_count += 1
4208
            self.model.compile(fullgraph=True, backend=backend)
4209
            return
4210
        # for other compilation modes, cudagraph behavior is controlled by
4211
4212
4213
        # CudagraphWraper and CudagraphDispatcher of vllm.

        # wrap the model with full cudagraph wrapper if needed.
4214
4215
        cudagraph_mode = self.compilation_config.cudagraph_mode
        assert cudagraph_mode is not None
4216
4217
4218
4219
        if (
            cudagraph_mode.has_full_cudagraphs()
            and not self.parallel_config.use_ubatching
        ):
4220
4221
4222
            self.model = CUDAGraphWrapper(
                self.model, self.vllm_config, runtime_mode=CUDAGraphMode.FULL
            )
4223
        elif self.parallel_config.use_ubatching:
4224
            if cudagraph_mode.has_full_cudagraphs():
4225
4226
4227
                self.model = UBatchWrapper(
                    self.model, self.vllm_config, CUDAGraphMode.FULL, self.device
                )
4228
            else:
4229
4230
4231
                self.model = UBatchWrapper(
                    self.model, self.vllm_config, CUDAGraphMode.NONE, self.device
                )
4232

4233
    def _get_eagle3_aux_layers_from_config(self) -> tuple[int, ...] | None:
4234
4235
4236
4237
4238
4239
4240
4241
4242
4243
4244
4245
4246
4247
4248
4249
4250
4251
4252
4253
4254
4255
        """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
4256

4257
    def reload_weights(self) -> None:
4258
        assert getattr(self, "model", None) is not None, (
4259
            "Cannot reload weights before model is loaded."
4260
        )
4261
4262
        model_loader = get_model_loader(self.load_config)
        logger.info("Reloading weights inplace...")
4263
        model_loader.load_weights(self.get_model(), model_config=self.model_config)
4264

4265
4266
4267
4268
4269
    def save_tensorized_model(
        self,
        tensorizer_config: "TensorizerConfig",
    ) -> None:
        TensorizerLoader.save_model(
4270
            self.get_model(),
4271
            tensorizer_config=tensorizer_config,
4272
            model_config=self.model_config,
4273
4274
        )

4275
4276
4277
    def _get_prompt_logprobs_dict(
        self,
        hidden_states: torch.Tensor,
4278
        num_scheduled_tokens: dict[str, int],
4279
    ) -> dict[str, LogprobsTensors | None]:
4280
        num_prompt_logprobs_dict = self.num_prompt_logprobs
4281
4282
4283
        if not num_prompt_logprobs_dict:
            return {}

4284
        in_progress_dict = self.input_batch.in_progress_prompt_logprobs_cpu
4285
        prompt_logprobs_dict: dict[str, LogprobsTensors | None] = {}
4286
4287
4288
4289
4290

        # 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():
4291
4292
4293
4294
            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
4295
4296
4297

            # Get metadata for this request.
            request = self.requests[req_id]
4298
4299
4300
4301
            if request.prompt_token_ids is None:
                # Prompt logprobs is incompatible with prompt embeddings
                continue

4302
4303
            num_prompt_tokens = len(request.prompt_token_ids)
            prompt_token_ids = torch.tensor(request.prompt_token_ids).to(
4304
4305
                self.device, non_blocking=True
            )
4306

4307
4308
4309
4310
4311
4312
            # 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(
4313
4314
                    num_prompt_tokens - 1, num_prompt_logprobs + 1
                )
4315
4316
                in_progress_dict[req_id] = logprobs_tensors

4317
            # Determine number of logits to retrieve.
4318
4319
            start_idx = request.num_computed_tokens
            start_tok = start_idx + 1
4320
            num_remaining_tokens = num_prompt_tokens - start_tok
4321
            if num_tokens <= num_remaining_tokens:
4322
                # This is a chunk, more tokens remain.
4323
4324
4325
                # 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.
4326
4327
4328
4329
4330
                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)
4331
4332
4333
4334
4335
4336
4337
                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
4338
4339
4340
4341
4342

            # 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]
4343
            offset = self.query_start_loc.np[req_idx].item()
4344
            prompt_hidden_states = hidden_states[offset : offset + num_logits]
4345
            logits = self.model.compute_logits(prompt_hidden_states)
4346
4347
4348
4349

            # 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.
4350
            tgt_token_ids = prompt_token_ids[start_tok : start_tok + num_logits]
4351
4352

            # Compute prompt logprobs.
4353
4354
            logprobs = self.sampler.compute_logprobs(logits)
            token_ids, logprobs, ranks = self.sampler.gather_logprobs(
4355
4356
                logprobs, num_prompt_logprobs, tgt_token_ids
            )
4357
4358

            # Transfer GPU->CPU async.
4359
4360
            chunk_slice = slice(start_idx, start_idx + num_logits)
            logprobs_tensors.logprob_token_ids[chunk_slice].copy_(
4361
4362
4363
                token_ids, non_blocking=True
            )
            logprobs_tensors.logprobs[chunk_slice].copy_(logprobs, non_blocking=True)
4364
            logprobs_tensors.selected_token_ranks[chunk_slice].copy_(
4365
4366
                ranks, non_blocking=True
            )
4367
4368
4369
4370
4371

        # 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]
4372
            del in_progress_dict[req_id]
4373
4374

        # Must synchronize the non-blocking GPU->CPU transfers.
4375
        if prompt_logprobs_dict:
4376
            self._sync_device()
4377
4378
4379

        return prompt_logprobs_dict

4380
4381
    def _get_nans_in_logits(
        self,
4382
        logits: torch.Tensor | None,
4383
4384
4385
4386
4387
4388
4389
4390
4391
4392
4393
    ) -> 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])
4394
4395
4396
                    if num_nans_for_index is not None and req_index < logits.shape[0]
                    else 0
                )
4397
4398
4399
4400
            return num_nans_in_logits
        except IndexError:
            return {}

4401
    @contextmanager
4402
4403
4404
    def maybe_randomize_inputs(
        self, input_ids: torch.Tensor | None, inputs_embeds: torch.Tensor | None
    ):
4405
4406
4407
4408
        """
        Randomize input_ids if VLLM_RANDOMIZE_DP_DUMMY_INPUTS is set.
        This is to help balance expert-selection
         - during profile_run
4409
         - during DP rank dummy run
4410
        """
4411

4412
4413
4414
4415
        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
4416
        elif input_ids is not None:
4417
4418
4419
4420

            @functools.cache
            def rand_input_ids() -> torch.Tensor:
                return torch.randint_like(
4421
                    self.input_ids.gpu,
4422
4423
                    low=0,
                    high=self.model_config.get_vocab_size(),
4424
                )
4425

4426
            logger.debug_once("Randomizing dummy input_ids for DP Rank")
4427
            input_ids.copy_(rand_input_ids()[: input_ids.size(0)], non_blocking=True)
4428
4429
            yield
            input_ids.fill_(0)
4430
4431
4432
4433
4434
4435
4436
4437
4438
4439
4440
4441
4442
4443
4444
        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)
4445

4446
4447
4448
4449
4450
4451
    def _get_mm_dummy_batch(
        self,
        modality: str,
        max_items_per_batch: int,
    ) -> BatchedTensorInputs:
        """Dummy data for profiling and precompiling multimodal models."""
4452
4453
        assert self.mm_budget is not None

4454
4455
4456
        # Don't use `max_items_per_batch` here to avoid redundant computation
        dummy_mm_inputs = self.mm_registry.get_dummy_mm_inputs(
            self.model_config,
4457
            mm_counts={modality: 1},
4458
            cache=self.mm_budget.cache,
4459
        )
4460
4461
4462
4463
4464
        dummy_mm_item = dummy_mm_inputs["mm_kwargs"][modality][0]

        # We use the cache so that the item is saved to the cache,
        # but not read from the cache
        assert dummy_mm_item is not None, "Item should not already be cached"
4465

4466
        dummy_mm_items = [dummy_mm_item] * max_items_per_batch
4467

4468
4469
4470
4471
4472
4473
4474
4475
        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,
            )
        )
4476

4477
4478
4479
4480
    @torch.inference_mode()
    def _dummy_run(
        self,
        num_tokens: int,
4481
        cudagraph_runtime_mode: CUDAGraphMode | None = None,
4482
4483
        force_attention: bool = False,
        uniform_decode: bool = False,
4484
        allow_microbatching: bool = True,
4485
4486
        skip_eplb: bool = False,
        is_profile: bool = False,
4487
        create_mixed_batch: bool = False,
4488
        remove_lora: bool = True,
4489
        activate_lora: bool = False,
Rémi Delacourt's avatar
Rémi Delacourt committed
4490
        is_graph_capturing: bool = False,
4491
    ) -> tuple[torch.Tensor, torch.Tensor]:
4492
4493
4494
4495
4496
4497
4498
        """
        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.
4499
                - if not set will determine the cudagraph mode based on using
4500
                    the self.cudagraph_dispatcher.
4501
4502
4503
4504
                - CUDAGraphMode.NONE: No cudagraph, for warm up and profile run
                - CUDAGraphMode.PIECEWISE: Piecewise cudagraph.
                - CUDAGraphMode.FULL: Full cudagraph, attention metadata is
                    needed.
4505
            force_attention: If True, always create attention metadata. Used to
4506
4507
4508
4509
                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.
4510
4511
            create_mixed_batch: If True, create a mixed batch with both decode
                (1 token) and prefill (multiple tokens) requests.
4512
            remove_lora: If False, dummy LoRAs are not destroyed after the run
4513
            activate_lora: If False, dummy_run is performed without LoRAs.
4514
        """
4515
4516
        mm_config = self.vllm_config.model_config.multimodal_config
        if mm_config and mm_config.mm_encoder_only:
4517
4518
4519
4520
            # 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([])

4521
4522
4523
4524
        assert (
            cudagraph_runtime_mode is None
            or cudagraph_runtime_mode.valid_runtime_modes()
        )
4525

4526
        # If cudagraph_mode.decode_mode() == FULL and
4527
        # cudagraph_mode.separate_routine(). This means that we are using
4528
4529
4530
4531
4532
4533
4534
4535
4536
4537
4538
        # 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.
4539
        max_query_len = self.uniform_decode_query_len if uniform_decode else num_tokens
4540

4541
4542
4543
4544
4545
        # 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
4546
4547
4548
4549
        if create_mixed_batch:
            assert not uniform_decode
            # Create mixed batch:
            # first half decode tokens, second half one prefill
4550
            num_decode_tokens = min(max_num_reqs - 1, num_tokens // 2)
4551
4552
4553
4554
            num_prefill_tokens = num_tokens - num_decode_tokens
            num_reqs = num_decode_tokens + 1

            # Create decode requests (1 token each) followed by prefill request
4555
            num_scheduled_tokens_list = [1] * num_decode_tokens + [num_prefill_tokens]
4556
4557
4558
            # Note: Overriding max_query_len to be the prefill tokens
            max_query_len = num_prefill_tokens
        elif uniform_decode:
4559
            assert not create_mixed_batch
4560
            num_reqs = min(max_num_reqs, cdiv(num_tokens, max_query_len))
4561
4562
            num_scheduled_tokens_list = [max_query_len] * num_reqs
            if num_tokens % max_query_len != 0:
4563
                num_scheduled_tokens_list[-1] = num_tokens % max_query_len
4564
4565
4566
4567
4568
4569
        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

4570
4571
        assert sum(num_scheduled_tokens_list) == num_tokens
        assert len(num_scheduled_tokens_list) == num_reqs
4572
        num_scheduled_tokens = np.array(num_scheduled_tokens_list, dtype=np.int32)
4573
4574
        num_tokens_unpadded = int(num_scheduled_tokens.sum())

4575
        num_sampled_tokens = np.ones(num_reqs, dtype=np.int32)
4576

4577
        _cudagraph_mode, batch_desc, should_ubatch, num_tokens_across_dp, _ = (
4578
4579
4580
4581
4582
4583
4584
4585
4586
4587
4588
4589
4590
4591
4592
4593
4594
4595
            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,
4596
            )
4597
        )
4598
4599
4600

        if cudagraph_runtime_mode is None:
            cudagraph_runtime_mode = _cudagraph_mode
4601
        else:
4602
4603
4604
4605
            assert cudagraph_runtime_mode == _cudagraph_mode, (
                f"Cudagraph runtime mode mismatch in dummy_run. "
                f"Expected {_cudagraph_mode}, but got {cudagraph_runtime_mode}."
            )
4606

4607
4608
4609
4610
        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
        )
4611
        ubatch_slices, ubatch_slices_padded = maybe_create_ubatch_slices(
4612
4613
4614
4615
4616
4617
4618
4619
4620
4621
            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,
4622
        )
4623

4624
        attn_metadata: PerLayerAttnMetadata | None = None
4625

4626
4627
4628
4629
4630
4631
4632
        slot_mappings_by_group, slot_mappings = self._get_slot_mappings(
            num_tokens_padded=num_tokens,
            num_reqs_padded=num_reqs_padded,
            num_tokens_unpadded=num_tokens_unpadded,
            ubatch_slices=ubatch_slices_padded,
        )

4633
4634
        # If force_attention is True, we always capture attention. Otherwise,
        # it only happens for cudagraph_runtime_mode=FULL.
4635
        if force_attention or cudagraph_runtime_mode == CUDAGraphMode.FULL:
4636
4637
4638
4639
4640
4641
            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:
4642
4643
4644
4645
4646
                if not envs.VLLM_USE_PIECEWISE:
                    seq_lens = max_query_len
                else:
                    # Make sure max_model_len is used at the graph capture time.
                    seq_lens = self.max_model_len
4647
            self.seq_lens.np[:num_reqs] = seq_lens
4648
4649
            self.seq_lens.np[num_reqs:] = 0
            self.seq_lens.copy_to_gpu()
4650

4651
4652
            cum_num_tokens, _ = self._get_cumsum_and_arange(num_scheduled_tokens)
            self.query_start_loc.np[1 : num_reqs + 1] = cum_num_tokens
4653
4654
            self.query_start_loc.copy_to_gpu()

4655
            pad_attn = cudagraph_runtime_mode == CUDAGraphMode.FULL
4656
            attn_metadata, _ = self._build_attention_metadata(
4657
4658
4659
                num_tokens=num_tokens_unpadded,
                num_reqs=num_reqs_padded,
                max_query_len=max_query_len,
4660
                ubatch_slices=ubatch_slices_padded if pad_attn else ubatch_slices,
4661
                for_cudagraph_capture=is_graph_capturing,
4662
                slot_mappings=slot_mappings_by_group,
4663
            )
4664

4665
        with self.maybe_dummy_run_with_lora(
4666
4667
4668
4669
4670
            self.lora_config,
            num_scheduled_tokens,
            num_sampled_tokens,
            activate_lora,
            remove_lora,
4671
        ):
4672
            # Make sure padding doesn't exceed max_num_tokens
4673
            assert num_tokens_padded <= self.max_num_tokens
4674
            model_kwargs = self._init_model_kwargs()
4675
            if self.supports_mm_inputs and not self.model_config.is_encoder_decoder:
Patrick von Platen's avatar
Patrick von Platen committed
4676
4677
                input_ids, inputs_embeds = self._prepare_mm_inputs(num_tokens_padded)

4678
                model_kwargs = {
4679
                    **model_kwargs,
4680
4681
                    **self._dummy_mm_kwargs(num_reqs),
                }
4682
4683
            elif self.enable_prompt_embeds:
                input_ids = None
4684
                inputs_embeds = self.inputs_embeds.gpu[:num_tokens_padded]
4685
                model_kwargs = self._init_model_kwargs()
4686
            else:
王敏's avatar
王敏 committed
4687
4688
4689
                self.input_ids.gpu[:num_tokens_padded] = torch.randint(0, self.model_config.get_vocab_size(),
                                                                        (num_tokens_padded,),
                                                                        dtype=torch.int32)
4690
                input_ids = self.input_ids.gpu[:num_tokens_padded]
4691
                inputs_embeds = None
4692

4693
            if self.uses_mrope:
4694
                positions = self.mrope_positions.gpu[:, :num_tokens_padded]
4695
            elif self.uses_xdrope_dim > 0:
4696
                positions = self.xdrope_positions.gpu[:, :num_tokens_padded]
4697
            else:
4698
                positions = self.positions.gpu[:num_tokens_padded]
4699
4700
4701
4702
4703
4704
4705
4706
4707

            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,
4708
4709
4710
                            device=self.device,
                        )
                    )
4711
4712

                intermediate_tensors = self.sync_and_slice_intermediate_tensors(
4713
                    num_tokens_padded, None, False
4714
                )
4715

4716
            if ubatch_slices_padded is not None:
4717
4718
4719
                # Adjust values to reflect a single ubatch.
                # TODO(sage,lucas): this is cruft that should be addressed in
                #  the padding refactor.
4720
                num_tokens_padded = ubatch_slices_padded[0].num_tokens
4721
                if num_tokens_across_dp is not None:
4722
                    num_tokens_across_dp[:] = num_tokens_padded
4723

4724
            with (
4725
                self.maybe_randomize_inputs(input_ids, inputs_embeds),
4726
                set_forward_context(
4727
4728
                    attn_metadata,
                    self.vllm_config,
4729
                    num_tokens=num_tokens_padded,
4730
4731
                    num_tokens_across_dp=num_tokens_across_dp,
                    cudagraph_runtime_mode=cudagraph_runtime_mode,
4732
                    batch_descriptor=batch_desc,
4733
                    ubatch_slices=ubatch_slices_padded,
4734
                    slot_mapping=slot_mappings,
4735
4736
                ),
            ):
4737
                outputs = self.model(
4738
4739
4740
4741
                    input_ids=input_ids,
                    positions=positions,
                    intermediate_tensors=intermediate_tensors,
                    inputs_embeds=inputs_embeds,
4742
                    **model_kwargs,
4743
                )
4744

4745
4746
4747
4748
            if self.use_aux_hidden_state_outputs:
                hidden_states, _ = outputs
            else:
                hidden_states = outputs
4749

4750
4751
4752
4753
4754
4755
            if self.speculative_config and (
                self.speculative_config.use_eagle()
                or self.speculative_config.uses_draft_model()
            ):
                assert isinstance(self.drafter, EagleProposer | DraftModelProposer)
                assert self.speculative_config is not None
4756
4757
4758
                # 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.
4759
                use_cudagraphs = (
4760
4761
4762
4763
4764
4765
4766
4767
4768
                    (
                        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
4769
4770
4771
4772
4773
4774
4775
4776
4777
4778
4779

                # 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
4780
                    is_graph_capturing=is_graph_capturing,
4781
                    slot_mappings=slot_mappings,
4782
                )
4783

4784
4785
4786
4787
4788
4789
4790
4791
4792
4793
4794
        # 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()

4795
4796
4797
4798
4799
4800
4801
4802
4803
4804
        # 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)

4805
        logit_indices = np.cumsum(num_scheduled_tokens) - 1
4806
4807
4808
4809
        logit_indices_device = torch.from_numpy(logit_indices).to(
            self.device, non_blocking=True
        )
        return hidden_states, hidden_states[logit_indices_device]
4810
4811
4812
4813
4814
4815

    @torch.inference_mode()
    def _dummy_sampler_run(
        self,
        hidden_states: torch.Tensor,
    ) -> torch.Tensor:
4816
4817
4818
        # The dummy hidden states may contain special values,
        # like `inf` or `nan`.
        # To avoid breaking the sampler, we use a random tensor here instead.
4819

4820
4821
        mm_config = self.vllm_config.model_config.multimodal_config
        if mm_config and mm_config.mm_encoder_only:
4822
4823
4824
            # MM Encoder only model no need to run sampler.
            return torch.tensor([])

4825
        hidden_states = torch.rand_like(hidden_states)
4826

4827
        logits = self.model.compute_logits(hidden_states)
4828
4829
        num_reqs = logits.size(0)

4830
        dummy_tensors = lambda v: torch.full((num_reqs,), v, device=self.device)
4831
4832
4833
4834
4835
4836
4837
4838
4839
4840
4841
4842
4843
4844
4845

        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)],
4846
            spec_token_ids=[[] for _ in range(num_reqs)],
4847
4848
            allowed_token_ids_mask=None,
            bad_words_token_ids={},
4849
            logitsprocs=LogitsProcessors(),
4850
        )
4851
        try:
4852
4853
4854
            sampler_output = self.sampler(
                logits=logits, sampling_metadata=dummy_metadata
            )
4855
        except RuntimeError as e:
4856
            if "out of memory" in str(e):
4857
4858
4859
4860
                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 "
4861
4862
                    "initializing the engine."
                ) from e
4863
4864
            else:
                raise e
4865
        if self.speculative_config:
4866
4867
            draft_token_ids = [[0] for _ in range(num_reqs)]
            dummy_spec_decode_metadata = SpecDecodeMetadata.make_dummy(
4868
4869
                draft_token_ids, self.device
            )
4870
4871

            num_tokens = sum(len(ids) for ids in draft_token_ids)
4872
4873
4874
            # draft_probs = torch.randn(
            #     num_tokens, logits.shape[-1], device=self.device,
            #     dtype=logits.dtype)
王敏's avatar
王敏 committed
4875
4876
4877
4878
4879
4880
4881
4882
            
            if not envs.VLLM_REJECT_SAMPLE_OPT:
                draft_probs = None
            else:
                draft_probs = torch.randn(
                    num_reqs, self.speculative_config.num_speculative_tokens, logits.shape[-1], device=self.device,
                    dtype=logits.dtype)

4883
4884
4885
4886
4887
            logits = torch.randn(
                num_tokens + num_reqs,
                logits.shape[-1],
                device=self.device,
                dtype=logits.dtype,
4888
            )
4889
4890
4891
            self.rejection_sampler(
                dummy_spec_decode_metadata,
                draft_probs,
4892
                logits,
4893
4894
                dummy_metadata,
            )
4895
        return sampler_output
4896

4897
    def _dummy_pooler_run_task(
4898
4899
        self,
        hidden_states: torch.Tensor,
4900
4901
        task: PoolingTask,
    ) -> PoolerOutput:
4902
4903
4904
4905
        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
4906
4907
4908
4909
        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
4910
4911
4912

        req_num_tokens = num_tokens // num_reqs

4913
        dummy_prompt_lens = torch.from_numpy(num_scheduled_tokens_np)
4914
4915
4916
        dummy_token_ids = torch.zeros(
            (num_reqs, req_num_tokens), dtype=torch.int32, device=self.device
        )
4917

4918
        model = cast(VllmModelForPooling, self.get_model())
4919
        dummy_pooling_params = PoolingParams(task=task)
4920
        dummy_pooling_params.verify(task=task, model_config=self.model_config)
4921
        to_update = model.pooler.get_pooling_updates(task)
4922
4923
        to_update.apply(dummy_pooling_params)

4924
        dummy_metadata = PoolingMetadata(
4925
4926
4927
            prompt_lens=dummy_prompt_lens,
            prompt_token_ids=dummy_token_ids,
            pooling_params=[dummy_pooling_params] * num_reqs,
4928
            pooling_states=[PoolingStates() for i in range(num_reqs)],
4929
        )
4930

4931
        dummy_metadata.build_pooling_cursor(
4932
            num_scheduled_tokens_np,
4933
4934
            seq_lens_cpu=dummy_prompt_lens,
            device=hidden_states.device,
4935
        )
4936

4937
        try:
4938
4939
4940
            return model.pooler(
                hidden_states=hidden_states, pooling_metadata=dummy_metadata
            )
4941
        except RuntimeError as e:
4942
            if "out of memory" in str(e):
4943
                raise RuntimeError(
4944
4945
4946
                    "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 "
4947
4948
                    "initializing the engine."
                ) from e
4949
4950
            else:
                raise e
4951
4952
4953
4954
4955
4956

    @torch.inference_mode()
    def _dummy_pooler_run(
        self,
        hidden_states: torch.Tensor,
    ) -> PoolerOutput:
4957
4958
        mm_config = self.vllm_config.model_config.multimodal_config
        if mm_config and mm_config.mm_encoder_only:
4959
4960
4961
            # MM Encoder only model not need to run pooler.
            return torch.tensor([])

4962
        # Find the task that has the largest output for subsequent steps
4963
4964
4965
        supported_pooling_tasks = self.get_supported_pooling_tasks()

        if not supported_pooling_tasks:
4966
4967
4968
4969
4970
4971
            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."
            )
4972

4973
        output_size = dict[PoolingTask, float]()
4974
        for task in supported_pooling_tasks:
4975
4976
            # Run a full batch with each task to ensure none of them OOMs
            output = self._dummy_pooler_run_task(hidden_states, task)
4977
            output_size[task] = sum(o.nbytes for o in output if o is not None)
4978
4979
4980
4981
            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)
4982

4983
    def profile_run(self) -> None:
4984
        # Profile with multimodal encoder & encoder cache.
4985
        if self.supports_mm_inputs:
4986
4987
            mm_config = self.model_config.multimodal_config
            if mm_config is not None and mm_config.skip_mm_profiling:
4988
                logger.info(
4989
                    "Skipping memory profiling for multimodal encoder and "
4990
4991
                    "encoder cache."
                )
4992
4993
4994
4995
4996
4997
4998
4999
            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.
5000
                    dummy_modality = mm_budget.get_modality_with_max_tokens()
5001
5002
5003
                    max_mm_items_per_batch = mm_budget.max_items_per_batch_by_modality[
                        dummy_modality
                    ]
5004
5005
5006
5007
5008
5009
5010
5011
5012

                    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,
                    )
5013

5014
5015
5016
5017
5018
                    # Create dummy batch of multimodal inputs.
                    batched_dummy_mm_inputs = self._get_mm_dummy_batch(
                        dummy_modality,
                        max_mm_items_per_batch,
                    )
5019

5020
                    # Run multimodal encoder.
5021
                    dummy_encoder_outputs = self.model.embed_multimodal(
5022
5023
                        **batched_dummy_mm_inputs
                    )
5024

5025
5026
5027
5028
                    sanity_check_mm_encoder_outputs(
                        dummy_encoder_outputs,
                        expected_num_items=max_mm_items_per_batch,
                    )
5029
5030
                    for i, output in enumerate(dummy_encoder_outputs):
                        self.encoder_cache[f"tmp_{i}"] = output
5031

5032
        # Add `is_profile` here to pre-allocate communication buffers
5033
5034
5035
        hidden_states, last_hidden_states = self._dummy_run(
            self.max_num_tokens, is_profile=True
        )
5036
        if get_pp_group().is_last_rank:
5037
5038
5039
5040
            if self.is_pooling_model:
                output = self._dummy_pooler_run(hidden_states)
            else:
                output = self._dummy_sampler_run(last_hidden_states)
5041
        else:
5042
            output = None
5043
        self._sync_device()
5044
        del hidden_states, output
5045
        self.encoder_cache.clear()
5046
        gc.collect()
5047

5048
    def capture_model(self) -> int:
5049
        if self.compilation_config.cudagraph_mode == CUDAGraphMode.NONE:
5050
            logger.warning(
5051
                "Skipping CUDA graph capture. To turn on CUDA graph capture, "
5052
5053
                "ensure `cudagraph_mode` was not manually set to `NONE`"
            )
5054
            return 0
5055

5056
5057
        compilation_counter.num_gpu_runner_capture_triggers += 1

5058
5059
        start_time = time.perf_counter()

5060
5061
5062
5063
5064
5065
5066
5067
5068
5069
5070
5071
5072
5073
        @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()
5074
                    gc.collect()
5075

5076
5077
5078
        # 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.
5079
        set_cudagraph_capturing_enabled(True)
5080
        with freeze_gc(), graph_capture(device=self.device):
5081
            start_free_gpu_memory = torch.cuda.mem_get_info()[0]
5082

5083
5084
5085
5086
            for (
                runtime_mode,
                batch_descs,
            ) in self.cudagraph_dispatcher.get_capture_descs():
5087
                self._capture_cudagraphs(
5088
5089
                    batch_descriptors=batch_descs,
                    cudagraph_runtime_mode=runtime_mode,
5090
                )
5091

5092
5093
            torch.cuda.synchronize()
            end_free_gpu_memory = torch.cuda.mem_get_info()[0]
5094
5095
5096
5097

        # 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
5098
        # we may do lazy capturing in future that still allows capturing
5099
5100
        # after here.
        set_cudagraph_capturing_enabled(False)
5101

5102
5103
5104
5105
        # Lock workspace to prevent resizing during execution.
        # Max workspace sizes should have been captured during warmup/profiling.
        lock_workspace()

5106
5107
5108
5109
        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.
5110
        logger.info_once(
5111
5112
5113
            "Graph capturing finished in %.0f secs, took %.2f GiB",
            elapsed_time,
            cuda_graph_size / (1 << 30),
5114
            scope="local",
5115
        )
5116
        return cuda_graph_size
5117

5118
5119
    def _capture_cudagraphs(
        self,
5120
        batch_descriptors: list[BatchDescriptor],
5121
5122
5123
5124
5125
5126
        cudagraph_runtime_mode: CUDAGraphMode,
    ):
        assert (
            cudagraph_runtime_mode != CUDAGraphMode.NONE
            and cudagraph_runtime_mode.valid_runtime_modes()
        ), f"Invalid cudagraph runtime mode: {cudagraph_runtime_mode}"
5127

5128
5129
5130
5131
5132
5133
5134
5135
5136
5137
5138
5139
5140
5141
        if not batch_descriptors:
            return

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

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

5142
5143
        # Only rank 0 should print progress bar during capture
        if is_global_first_rank():
5144
5145
            batch_descriptors = tqdm(
                batch_descriptors,
5146
5147
5148
                disable=not self.load_config.use_tqdm_on_load,
                desc="Capturing CUDA graphs ({}, {})".format(
                    "decode" if uniform_decode else "mixed prefill-decode",
5149
5150
5151
                    cudagraph_runtime_mode.name,
                ),
            )
5152

5153
        # We skip EPLB here since we don't want to record dummy metrics
5154
5155
5156
5157
        for batch_desc in batch_descriptors:
            num_tokens = batch_desc.num_tokens
            activate_lora = batch_desc.has_lora

5158
            # We currently only capture ubatched graphs when its a FULL
5159
5160
5161
            # 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
5162
            allow_microbatching = (
5163
                self.parallel_config.use_ubatching
5164
5165
                and cudagraph_runtime_mode == CUDAGraphMode.FULL
                and uniform_decode
5166
5167
5168
5169
5170
                and check_ubatch_thresholds(
                    config=self.vllm_config.parallel_config,
                    num_tokens=num_tokens,
                    uniform_decode=uniform_decode,
                )
5171
            )
5172

5173
5174
            for _ in range(self.compilation_config.cudagraph_num_of_warmups):
                # Use CUDAGraphRuntimeStyle.NONE (default) for warmup.
5175
                # But be careful, warm up with `NONE` is orthogonal to
5176
5177
5178
                # 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.
5179
                dummy_run(
5180
5181
5182
                    num_tokens,
                    cudagraph_runtime_mode=CUDAGraphMode.NONE,
                    allow_microbatching=allow_microbatching,
5183
                    activate_lora=activate_lora,
5184
                )
5185
5186
5187

            # Capture run
            dummy_run(
5188
5189
5190
                num_tokens,
                cudagraph_runtime_mode=cudagraph_runtime_mode,
                allow_microbatching=allow_microbatching,
5191
                activate_lora=activate_lora,
Rémi Delacourt's avatar
Rémi Delacourt committed
5192
                is_graph_capturing=True,
5193
            )
5194
        self.maybe_remove_all_loras(self.lora_config)
5195

5196
5197
5198
5199
    def initialize_attn_backend(self, kv_cache_config: KVCacheConfig) -> None:
        """
        Initialize the attention backends and attention metadata builders.
        """
5200
        assert len(self.attn_groups) == 0, "Attention backends are already initialized"
5201

5202
5203
5204
5205
5206
5207
        class AttentionGroupKey(NamedTuple):
            attn_backend: type[AttentionBackend]
            kv_cache_spec: KVCacheSpec

        def get_attn_backends_for_group(
            kv_cache_group_spec: KVCacheGroupSpec,
5208
        ) -> tuple[dict[AttentionGroupKey, list[str]], set[type[AttentionBackend]]]:
5209
            layer_type = cast(type[Any], AttentionLayerBase)
5210
            layers = get_layers_from_vllm_config(
5211
                self.vllm_config, layer_type, kv_cache_group_spec.layer_names
5212
            )
5213
5214
            attn_backends = {}
            attn_backend_layers = defaultdict(list)
5215
            # Dedupe based on full class name; this is a bit safer than
5216
5217
5218
5219
            # 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.
5220
            for layer_name in kv_cache_group_spec.layer_names:
5221
                attn_backend = layers[layer_name].get_attn_backend()
5222
5223
5224
5225

                if layer_name in self.kv_sharing_fast_prefill_eligible_layers:
                    attn_backend = create_fast_prefill_custom_backend(
                        "FastPrefill",
5226
                        attn_backend,  # type: ignore[arg-type]
5227
5228
                    )

5229
5230
5231
                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):
5232
                    layer_kv_cache_spec = layer_kv_cache_spec.kv_cache_specs[layer_name]
5233
                key = (full_cls_name, layer_kv_cache_spec)
5234
5235
5236
                attn_backends[key] = AttentionGroupKey(
                    attn_backend, layer_kv_cache_spec
                )
5237
                attn_backend_layers[key].append(layer_name)
5238
5239
5240
5241
            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()),
            )
5242
5243

        def create_attn_groups(
5244
            attn_backends_map: dict[AttentionGroupKey, list[str]],
5245
            kv_cache_group_id: int,
5246
5247
        ) -> list[AttentionGroup]:
            attn_groups: list[AttentionGroup] = []
5248
            for (attn_backend, kv_cache_spec), layer_names in attn_backends_map.items():
5249
                attn_group = AttentionGroup(
5250
                    attn_backend,
5251
                    layer_names,
5252
                    kv_cache_spec,
5253
                    kv_cache_group_id,
5254
                )
5255

5256
5257
5258
                attn_groups.append(attn_group)
            return attn_groups

5259
        attention_backend_maps = []
5260
        attention_backend_list = []
5261
        for kv_cache_group_spec in kv_cache_config.kv_cache_groups:
5262
            attn_backends = get_attn_backends_for_group(kv_cache_group_spec)
5263
            attention_backend_maps.append(attn_backends[0])
5264
            attention_backend_list.append(attn_backends[1])
5265
5266

        # Resolve cudagraph_mode before actually initialize metadata_builders
5267
5268
5269
        self._check_and_update_cudagraph_mode(
            attention_backend_list, kv_cache_config.kv_cache_groups
        )
5270

5271
5272
5273
        # Check if attention backend supports PCP&DCP and related features.
        check_attention_cp_compatibility(self.vllm_config)

5274
5275
        for i, attn_backend_map in enumerate(attention_backend_maps):
            self.attn_groups.append(create_attn_groups(attn_backend_map, i))
5276

5277
5278
5279
5280
5281
5282
5283
5284
5285
5286
5287
5288
5289
5290
5291
    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
5292
5293
                    if not self.parallel_config.use_ubatching
                    else self.parallel_config.num_ubatches,
5294
                )
co63oc's avatar
co63oc committed
5295
        # Calculate reorder batch threshold (if needed)
5296
5297
        # Note (tdoublep): do this *after* constructing builders,
        # because some of them change the threshold at init time.
5298
5299
        self.calculate_reorder_batch_threshold()

5300
    def _check_and_update_cudagraph_mode(
5301
5302
5303
        self,
        attention_backends: list[set[type[AttentionBackend]]],
        kv_cache_groups: list[KVCacheGroupSpec],
5304
    ) -> None:
5305
        """
5306
        Resolve the cudagraph_mode when there are multiple attention
5307
        groups with potential conflicting CUDA graph support.
5308
5309
5310
        Then initialize the cudagraph_dispatcher based on the resolved
        cudagraph_mode.
        """
5311
        min_cg_support = AttentionCGSupport.ALWAYS
5312
        min_cg_backend_name = None
5313

5314
5315
5316
5317
5318
        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()
5319

5320
5321
5322
5323
5324
5325
                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__
5326
5327
        # Flexible resolve the cudagraph mode
        cudagraph_mode = self.compilation_config.cudagraph_mode
5328
        assert cudagraph_mode is not None
5329
        # check cudagraph for mixed batch is supported
5330
5331
5332
5333
5334
5335
        if (
            cudagraph_mode.mixed_mode() == CUDAGraphMode.FULL
            and min_cg_support != AttentionCGSupport.ALWAYS
        ):
            msg = (
                f"CUDAGraphMode.{cudagraph_mode.name} is not supported "
5336
                f"with {min_cg_backend_name} backend (support: "
5337
5338
                f"{min_cg_support})"
            )
5339
5340
            if min_cg_support == AttentionCGSupport.NEVER:
                # if not supported any full cudagraphs, just raise it.
5341
5342
                msg += (
                    "; please try cudagraph_mode=PIECEWISE, and "
5343
                    "make sure compilation mode is VLLM_COMPILE"
5344
                )
5345
5346
5347
5348
5349
                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"
5350
                cudagraph_mode = self.compilation_config.cudagraph_mode = (
5351
                    CUDAGraphMode.FULL_AND_PIECEWISE
5352
                )
5353
5354
            else:
                msg += "; setting cudagraph_mode=FULL_DECODE_ONLY"
5355
                cudagraph_mode = self.compilation_config.cudagraph_mode = (
5356
                    CUDAGraphMode.FULL_DECODE_ONLY
5357
                )
5358
5359
            logger.warning(msg)

5360
        # check that if we are doing decode full-cudagraphs it is supported
5361
5362
5363
5364
        if not envs.VLLM_USE_PIECEWISE:
            if (
                cudagraph_mode.decode_mode() == CUDAGraphMode.FULL
                and min_cg_support == AttentionCGSupport.NEVER
5365
            ):
5366
5367
5368
5369
                msg = (
                    f"CUDAGraphMode.{cudagraph_mode.name} is not supported "
                    f"with {min_cg_backend_name} backend (support: "
                    f"{min_cg_support})"
5370
                )
5371
5372
5373
5374
5375
5376
5377
5378
5379
5380
5381
5382
5383
5384
5385
5386
5387
5388
5389
5390
                if self.compilation_config.mode == CompilationMode.VLLM_COMPILE and (
                    self.compilation_config.splitting_ops_contain_attention()
                    or self.compilation_config.use_inductor_graph_partition
                ):
                    msg += (
                        "; setting cudagraph_mode=PIECEWISE because "
                        "attention is compiled piecewise"
                    )
                    cudagraph_mode = self.compilation_config.cudagraph_mode = (
                        CUDAGraphMode.PIECEWISE
                    )
                else:
                    msg += (
                        "; setting cudagraph_mode=NONE because "
                        "attention is not compiled piecewise"
                    )
                    cudagraph_mode = self.compilation_config.cudagraph_mode = (
                        CUDAGraphMode.NONE
                    )
                logger.warning(msg)
5391

5392
5393
        # check that if we are doing spec-decode + decode full-cudagraphs it is
        # supported
5394
5395
5396
5397
5398
5399
5400
5401
        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 "
5402
                f"{min_cg_backend_name} (support: {min_cg_support})"
5403
            )
5404
5405
            if self.compilation_config.splitting_ops_contain_attention():
                msg += "; setting cudagraph_mode=PIECEWISE"
5406
                cudagraph_mode = self.compilation_config.cudagraph_mode = (
5407
                    CUDAGraphMode.PIECEWISE
5408
                )
5409
5410
            else:
                msg += "; setting cudagraph_mode=NONE"
5411
                cudagraph_mode = self.compilation_config.cudagraph_mode = (
5412
                    CUDAGraphMode.NONE
5413
                )
5414
5415
5416
5417
            logger.warning(msg)

        # double check that we can support full cudagraph if they are requested
        # even after automatic downgrades
5418
5419
5420
5421
5422
5423
        if (
            cudagraph_mode.has_full_cudagraphs()
            and min_cg_support == AttentionCGSupport.NEVER
        ):
            raise ValueError(
                f"CUDAGraphMode.{cudagraph_mode.name} is not "
5424
                f"supported with {min_cg_backend_name} backend ("
5425
5426
                f"support:{min_cg_support}) "
                "; please try cudagraph_mode=PIECEWISE, "
5427
                "and make sure compilation mode is VLLM_COMPILE"
5428
            )
5429

5430
5431
5432
5433
        # 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
5434
        # Will be removed in the near future when we have separate cudagraph capture
5435
5436
5437
5438
5439
5440
5441
5442
5443
        # 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
            )
5444
5445
5446
5447
            capture_sizes = self.compilation_config.cudagraph_capture_sizes
            self.cudagraph_batch_sizes = (
                capture_sizes if capture_sizes is not None else []
            )
5448

5449
5450
        # Trigger cudagraph dispatching keys initialization after
        # resolved cudagraph mode.
5451
        self.compilation_config.cudagraph_mode = cudagraph_mode
5452
        self.cudagraph_dispatcher.initialize_cudagraph_keys(
5453
            cudagraph_mode, self.uniform_decode_query_len
5454
        )
5455

5456
5457
5458
5459
5460
        # Initialize eagle's cudagraph dispatcher if using eagle spec decode.
        if self.speculative_config and self.speculative_config.use_eagle():
            assert isinstance(self.drafter, EagleProposer)
            self.drafter.initialize_cudagraph_keys(cudagraph_mode)

5461
5462
    def calculate_reorder_batch_threshold(self) -> None:
        """
5463
5464
5465
5466
        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.
5467
        """
5468
5469
        min_none_high = lambda a, b: a if b is None else b if a is None else min(a, b)

5470
        reorder_batch_thresholds: list[int | None] = [
5471
5472
5473
            group.get_metadata_builder().reorder_batch_threshold
            for group in self._attn_group_iterator()
        ]
5474
5475
5476
5477
5478
        # 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
5479
        self.reorder_batch_threshold = reduce(min_none_high, reorder_batch_thresholds)  # type: ignore[assignment]
5480

5481
5482
5483
    @staticmethod
    def select_common_block_size(
        kv_manager_block_size: int, attn_groups: list[AttentionGroup]
5484
5485
    ) -> int:
        """
5486
5487
5488
5489
5490
        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.
5491

5492
5493
5494
5495
5496
        Args:
            kv_manager_block_size: Block size of KV cache
            attn_groups: List of attention groups

        Returns:
5497
            The selected block size
5498
5499

        Raises:
5500
            ValueError: If no valid block size found
5501
5502
        """

5503
5504
5505
5506
5507
5508
5509
5510
        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
5511
                for supported_size in backend.get_supported_kernel_block_sizes():
5512
5513
5514
5515
5516
5517
5518
5519
5520
5521
5522
5523
5524
5525
5526
5527
5528
5529
5530
5531
5532
5533
5534
5535
5536
5537
5538
5539
5540
5541
                    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
5542
            for supported_size in backend.get_supported_kernel_block_sizes()
5543
5544
            if isinstance(supported_size, int)
        )
5545

5546
5547
5548
5549
5550
5551
        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}. ")
5552

5553
5554
5555
    def may_reinitialize_input_batch(
        self, kv_cache_config: KVCacheConfig, kernel_block_sizes: list[int]
    ) -> None:
5556
5557
5558
5559
5560
5561
5562
        """
        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.
5563
            kernel_block_sizes: The kernel block sizes for each KV cache group.
5564
5565
5566
5567
        """
        block_sizes = [
            kv_cache_group.kv_cache_spec.block_size
            for kv_cache_group in kv_cache_config.kv_cache_groups
5568
            if not isinstance(kv_cache_group.kv_cache_spec, EncoderOnlyAttentionSpec)
5569
        ]
5570
5571
5572
5573
5574
5575
5576
5577
5578
5579
5580
5581
5582
5583
5584
5585
5586
5587
        max_num_blocks = []
        max_model_len = max(self.max_model_len, self.max_encoder_len)
        for i, kv_cache_group in enumerate(kv_cache_config.kv_cache_groups):
            if isinstance(kv_cache_group.kv_cache_spec, EncoderOnlyAttentionSpec):
                continue
            max_num_blocks_per_req = cdiv(
                max_model_len, block_sizes[i] * get_total_cp_world_size()
            )
            if isinstance(kv_cache_group.kv_cache_spec, MambaSpec):
                mamba_blocks_per_req = (
                    max_num_blocks_per_req
                    if self.cache_config.enable_prefix_caching
                    else 1
                ) + kv_cache_group.kv_cache_spec.num_speculative_blocks
                max_num_blocks_per_req = max(
                    max_num_blocks_per_req, mamba_blocks_per_req
                )
            max_num_blocks.append(max_num_blocks_per_req)
5588
5589
5590
5591

        if block_sizes != [self.cache_config.block_size] or kernel_block_sizes != [
            self.cache_config.block_size
        ]:
5592
5593
5594
            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
5595
5596
                "for more details."
            )
5597
5598
            self.input_batch = InputBatch(
                max_num_reqs=self.max_num_reqs,
5599
                max_model_len=max_model_len,
5600
5601
5602
5603
5604
                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,
5605
                kernel_block_sizes=kernel_block_sizes,
5606
                max_num_blocks_per_req=max_num_blocks,
5607
                is_spec_decode=bool(self.vllm_config.speculative_config),
5608
                logitsprocs=self.input_batch.logitsprocs,
5609
                logitsprocs_need_output_token_ids=self.input_batch.logitsprocs_need_output_token_ids,
5610
                is_pooling_model=self.is_pooling_model,
5611
5612
            )

5613
    def _allocate_kv_cache_tensors(
5614
5615
        self, kv_cache_config: KVCacheConfig
    ) -> dict[str, torch.Tensor]:
5616
        """
5617
5618
5619
        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.

5620
        Args:
5621
            kv_cache_config: The KV cache config
5622
        Returns:
5623
            dict[str, torch.Tensor]: A map between layer names to their
5624
            corresponding memory buffer for KV cache.
5625
        """
5626
5627
        kv_cache_raw_tensors: dict[str, torch.Tensor] = {}
        for kv_cache_tensor in kv_cache_config.kv_cache_tensors:
5628
5629
5630
            tensor = torch.zeros(
                kv_cache_tensor.size, dtype=torch.int8, device=self.device
            )
5631
5632
5633
5634
5635
            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:
5636
5637
5638
5639
            for layer_name in group.layer_names:
                if layer_name in self.runner_only_attn_layers:
                    continue
                layer_names.add(layer_name)
5640
5641
5642
        assert layer_names == set(kv_cache_raw_tensors.keys()), (
            "Some layers are not correctly initialized"
        )
5643
5644
        return kv_cache_raw_tensors

5645
5646
5647
    def _attn_group_iterator(self) -> Iterator[AttentionGroup]:
        return itertools.chain.from_iterable(self.attn_groups)

5648
    def _kv_cache_spec_attn_group_iterator(self) -> Iterator[AttentionGroup]:
5649
5650
        if not self.kv_cache_config.kv_cache_groups:
            return
5651
5652
        for attn_groups in self.attn_groups:
            yield from attn_groups
5653

5654
5655
5656
5657
5658
5659
5660
5661
5662
5663
5664
5665
5666
5667
5668
    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 = []
5669
        for kv_cache_gid, kv_cache_group in enumerate(kv_cache_config.kv_cache_groups):
5670
5671
5672
5673
5674
5675
            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):
5676
                continue
5677
            elif isinstance(kv_cache_spec, AttentionSpec):
5678
5679
5680
                # This is an attention backend that supports virtual
                # block splitting. Get the supported block sizes from
                # all backends in the group.
5681
                attn_groups = self.attn_groups[kv_cache_gid]
5682
                kv_manager_block_size = kv_cache_group.kv_cache_spec.block_size
5683
                selected_kernel_size = self.select_common_block_size(
5684
5685
5686
                    kv_manager_block_size, attn_groups
                )
                kernel_block_sizes.append(selected_kernel_size)
5687
            elif isinstance(kv_cache_spec, MambaSpec):
5688
5689
                # This is likely Mamba or other non-attention cache,
                # no splitting.
5690
                kernel_block_sizes.append(kv_cache_spec.block_size)
5691
5692
5693
5694
5695
5696
            else:
                raise NotImplementedError(
                    f"unknown kv cache spec {kv_cache_group.kv_cache_spec}"
                )
        return kernel_block_sizes

5697
5698
5699
5700
    def _reshape_kv_cache_tensors(
        self,
        kv_cache_config: KVCacheConfig,
        kv_cache_raw_tensors: dict[str, torch.Tensor],
5701
        kernel_block_sizes: list[int],
5702
    ) -> dict[str, torch.Tensor]:
5703
        """
5704
        Reshape the KV cache tensors to the desired shape and dtype.
5705

5706
        Args:
5707
5708
            kv_cache_config: The KV cache config
            kv_cache_raw_tensors: The KV cache buffer of each layer, with
5709
                correct size but uninitialized shape.
5710
            kernel_block_sizes: The kernel block sizes for each KV cache group.
5711
        Returns:
5712
            Dict[str, torch.Tensor]: A map between layer names to their
5713
5714
            corresponding memory buffer for KV cache.
        """
5715
        kv_caches: dict[str, torch.Tensor] = {}
5716
        has_attn, has_mamba = False, False
5717
5718
        for group in self._kv_cache_spec_attn_group_iterator():
            kv_cache_spec = group.kv_cache_spec
5719
            attn_backend = group.backend
5720
5721
5722
5723
            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]
5724
            for layer_name in group.layer_names:
5725
5726
                if layer_name in self.runner_only_attn_layers:
                    continue
5727
5728
                raw_tensor = kv_cache_raw_tensors[layer_name]
                assert raw_tensor.numel() % kv_cache_spec.page_size_bytes == 0
5729
                num_blocks = raw_tensor.numel() // kv_cache_spec.page_size_bytes
5730
                if isinstance(kv_cache_spec, AttentionSpec):
5731
                    has_attn = True
5732
5733
                    num_blocks_per_kv_block = (
                        kv_cache_spec.block_size // kernel_block_size
5734
5735
5736
                    )
                    kernel_num_blocks = num_blocks * num_blocks_per_kv_block

5737
                    if envs.VLLM_USE_FLASH_ATTN_PA and not self.vllm_config.model_config.use_mla:
5738
                        key_cache_shape, value_cache_shape = attn_backend.get_kv_cache_shape(
5739
5740
                            kernel_num_blocks,
                            kernel_block_size,
5741
5742
                            kv_cache_spec.num_kv_heads,
                            kv_cache_spec.head_size,
5743
5744
                            cache_dtype_str=self.cache_config.cache_dtype,
                        )
5745
5746
5747
                        dtype = kv_cache_spec.dtype
                        try:
                            key_stride_order, value_stride_order = attn_backend.get_kv_cache_stride_order()
5748
5749
                            assert len(key_stride_order) == len(key_stride_order)
                            assert len(value_stride_order) == len(value_cache_shape)
5750
                        except (AttributeError, NotImplementedError):
5751
5752
                            key_stride_order = tuple(range(len(key_cache_shape)))
                            value_stride_order = tuple(range(len(value_cache_shape)))
5753
5754
5755
5756
5757
                        # 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.
5758
5759
5760
5761
                        key_cache_shape = tuple(
                            key_cache_shape[i] for i in key_stride_order)
                        value_cache_shape = tuple(
                            value_cache_shape[i] for i in value_stride_order)
5762
5763
5764
5765
5766
5767
5768
5769
5770
5771
5772
5773
5774
5775
5776
5777
5778
5779
5780
5781
5782
5783
5784
5785
5786
5787
5788
                        # Maintain original KV shape view.
                        inv_key_order = [
                            key_stride_order.index(i)
                            for i in range(len(key_stride_order))
                        ]
                        inv_value_order = [
                            value_stride_order.index(i)
                            for i in range(len(value_stride_order))
                        ]
                        
                        raw_tensor = kv_cache_raw_tensors[layer_name].view(dtype)
                        total_elements = raw_tensor.numel()
                        key_elements = (key_cache_shape[0] * key_cache_shape[1] * 
                                        key_cache_shape[2] * key_cache_shape[3])
                        value_elements = (value_cache_shape[0] * value_cache_shape[1] *
                                        value_cache_shape[2] * value_cache_shape[3])

                        assert total_elements == key_elements + value_elements

                        key_cache = raw_tensor[:key_elements].view(key_cache_shape).permute(
                            *inv_key_order)
                        value_cache = raw_tensor[key_elements:].view(value_cache_shape).permute(
                            *inv_value_order)
                        kv_caches[layer_name] = (key_cache, value_cache)

                    else:
                        kv_cache_shape = attn_backend.get_kv_cache_shape(
5789
5790
                            kernel_num_blocks,
                            kernel_block_size,
5791
5792
                            kv_cache_spec.num_kv_heads,
                            kv_cache_spec.head_size,
5793
5794
                            cache_dtype_str=self.cache_config.cache_dtype,
                        )
5795
5796
                        dtype = kv_cache_spec.dtype
                        try:
5797
5798
                            kv_cache_stride_order = attn_backend.get_kv_cache_stride_order()
                            assert len(kv_cache_stride_order) == len(kv_cache_shape)
5799
                        except (AttributeError, NotImplementedError):
5800
                            kv_cache_stride_order = tuple(range(len(kv_cache_shape)))
5801
5802
5803
5804
5805
                        # 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.
5806
5807
5808
                        kv_cache_shape = tuple(
                            kv_cache_shape[i] for i in kv_cache_stride_order
                        )
5809
5810
5811
5812
5813
                        # Maintain original KV shape view.
                        inv_order = [
                            kv_cache_stride_order.index(i)
                            for i in range(len(kv_cache_stride_order))
                        ]
5814
5815
5816
5817
5818
5819
                        kv_caches[layer_name] = (
                            kv_cache_raw_tensors[layer_name]
                            .view(dtype)
                            .view(kv_cache_shape)
                            .permute(*inv_order)
                        )
5820

Chen Zhang's avatar
Chen Zhang committed
5821
                elif isinstance(kv_cache_spec, MambaSpec):
5822
                    has_mamba = True
Chen Zhang's avatar
Chen Zhang committed
5823
5824
                    raw_tensor = kv_cache_raw_tensors[layer_name]
                    state_tensors = []
5825
                    storage_offset_bytes = 0
5826
                    for shape, dtype in zip(kv_cache_spec.shapes, kv_cache_spec.dtypes):
5827
5828
                        dtype_size = get_dtype_size(dtype)
                        num_element_per_page = (
5829
5830
                            kv_cache_spec.page_size_bytes // dtype_size
                        )
Chen Zhang's avatar
Chen Zhang committed
5831
                        target_shape = (num_blocks, *shape)
5832
5833
                        stride = torch.empty(target_shape).stride()
                        target_stride = (num_element_per_page, *stride[1:])
5834
                        assert storage_offset_bytes % dtype_size == 0
5835
5836
5837
5838
                        tensor = torch.as_strided(
                            raw_tensor.view(dtype),
                            size=target_shape,
                            stride=target_stride,
5839
                            storage_offset=storage_offset_bytes // dtype_size,
5840
                        )
Chen Zhang's avatar
Chen Zhang committed
5841
                        state_tensors.append(tensor)
5842
                        storage_offset_bytes += stride[0] * dtype_size
5843
5844

                    kv_caches[layer_name] = state_tensors
5845
                else:
5846
                    raise NotImplementedError
5847
5848

        if has_attn and has_mamba:
5849
            self._update_hybrid_attention_mamba_layout(kv_caches)
5850

5851
5852
        return kv_caches

5853
    def _update_hybrid_attention_mamba_layout(
5854
5855
        self, kv_caches: dict[str, torch.Tensor]
    ) -> None:
5856
        """
5857
5858
        Update the layout of attention layers from (2, num_blocks, ...) to
        (num_blocks, 2, ...).
5859
5860

        Args:
5861
            kv_caches: The KV cache buffer of each layer.
5862
5863
        """

5864
5865
        for group in self._kv_cache_spec_attn_group_iterator():
            kv_cache_spec = group.kv_cache_spec
5866
            for layer_name in group.layer_names:
5867
                kv_cache = kv_caches[layer_name]
5868
5869
5870
5871
                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 "
5872
                        f"a tensor of shape {kv_cache.shape}"
5873
                    )
5874
                    hidden_size = kv_cache.shape[2:].numel()
5875
5876
5877
5878
                    kv_cache.as_strided_(
                        size=kv_cache.shape,
                        stride=(hidden_size, 2 * hidden_size, *kv_cache.stride()[2:]),
                    )
5879

5880
    def initialize_kv_cache_tensors(
5881
        self, kv_cache_config: KVCacheConfig, kernel_block_sizes: list[int]
5882
    ) -> dict[str, torch.Tensor]:
5883
5884
5885
5886
5887
        """
        Initialize the memory buffer for KV cache.

        Args:
            kv_cache_config: The KV cache config
5888
5889
            kernel_block_sizes: The kernel block sizes for each KV cache group.

5890
        Returns:
5891
            Dict[str, torch.Tensor]: A map between layer names to their
5892
5893
            corresponding memory buffer for KV cache.
        """
5894
5895
5896
5897
5898
5899
5900
5901
5902
5903
5904
5905
5906
5907
5908
5909
5910
5911
5912
5913
5914
5915
5916
5917

        # 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
            )
5918

5919
        # Set up cross-layer KV cache sharing
5920
5921
        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)
5922
5923
            kv_caches[layer_name] = kv_caches[target_layer_name]

5924
5925
5926
5927
5928
5929
5930
5931
5932
        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,
        )
5933
5934
5935
        return kv_caches

    def maybe_add_kv_sharing_layers_to_kv_cache_groups(
5936
5937
        self, kv_cache_config: KVCacheConfig
    ) -> None:
5938
5939
5940
5941
5942
5943
5944
5945
5946
5947
5948
5949
5950
5951
5952
5953
5954
5955
        """
        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.
5956
            attn_layers = get_layers_from_vllm_config(self.vllm_config, Attention)
5957
5958
            for layer_name in reversed(attn_layers):
                if layer_name in self.shared_kv_cache_layers:
5959
                    self.kv_sharing_fast_prefill_eligible_layers.add(layer_name)
5960
5961
                else:
                    break
5962

5963
5964
5965
5966
5967
5968
5969
    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
        """
5970
        kv_cache_config = deepcopy(kv_cache_config)
5971
        self.kv_cache_config = kv_cache_config
5972
        self.may_add_encoder_only_layers_to_kv_cache_config()
5973
        self.maybe_add_kv_sharing_layers_to_kv_cache_groups(kv_cache_config)
5974
        self.initialize_attn_backend(kv_cache_config)
5975
5976
5977
5978
5979
5980
        # 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)
5981
5982
5983
5984

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

5985
        # Reinitialize need to after initialize_attn_backend
5986
5987
5988
5989
        self.may_reinitialize_input_batch(kv_cache_config, kernel_block_sizes)
        kv_caches = self.initialize_kv_cache_tensors(
            kv_cache_config, kernel_block_sizes
        )
5990

5991
5992
5993
5994
5995
        if self.speculative_config and (
            self.speculative_config.use_eagle()
            or self.speculative_config.uses_draft_model()
        ):
            assert isinstance(self.drafter, EagleProposer | DraftModelProposer)
5996
5997
            # validate all draft model layers belong to the same kv cache
            # group
zhuwenwen's avatar
zhuwenwen committed
5998
            self.drafter.validate_same_kv_cache_group(kv_cache_config)
5999

Robert Shaw's avatar
Robert Shaw committed
6000
        if has_kv_transfer_group():
6001
            kv_transfer_group = get_kv_transfer_group()
6002
6003
6004
6005
6006
6007
6008
            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)
6009
            kv_transfer_group.set_host_xfer_buffer_ops(copy_kv_blocks)
Robert Shaw's avatar
Robert Shaw committed
6010

6011
6012
6013
6014
6015
6016
6017
6018
6019
6020
6021
6022
6023
6024
6025
6026
6027
        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,
6028
            vllm_config=self.vllm_config,
6029
6030
        )

6031
6032
6033
6034
6035
    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
6036
        encoder_only_attn_specs: dict[AttentionSpec, list[str]] = defaultdict(list)
6037
6038
6039
        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:
6040
                attn_spec: AttentionSpec = EncoderOnlyAttentionSpec(
6041
6042
6043
                    block_size=block_size,
                    num_kv_heads=attn_module.num_kv_heads,
                    head_size=attn_module.head_size,
6044
6045
                    dtype=self.kv_cache_dtype,
                )
6046
6047
6048
                encoder_only_attn_specs[attn_spec].append(layer_name)
                self.runner_only_attn_layers.add(layer_name)
        if len(encoder_only_attn_specs) > 0:
6049
6050
6051
            assert len(encoder_only_attn_specs) == 1, (
                "Only support one encoder-only attention spec now"
            )
6052
6053
            spec, layer_names = encoder_only_attn_specs.popitem()
            self.kv_cache_config.kv_cache_groups.append(
6054
6055
                KVCacheGroupSpec(layer_names=layer_names, kv_cache_spec=spec)
            )
6056

6057
    def get_kv_cache_spec(self) -> dict[str, KVCacheSpec]:
6058
        """
6059
        Generates the KVCacheSpec by parsing the kv cache format from each
6060
6061
        Attention module in the static forward context.
        Returns:
6062
            KVCacheSpec: A dictionary mapping layer names to their KV cache
6063
6064
            format. Layers that do not need KV cache are not included.
        """
6065
6066
        if has_ec_transfer() and get_ec_transfer().is_producer:
            return {}
6067
        kv_cache_spec: dict[str, KVCacheSpec] = {}
6068
6069
        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
6070
        for layer_name, attn_module in attn_layers.items():
6071
6072
6073
            if isinstance(attn_module, Attention) and (
                kv_tgt_layer := attn_module.kv_sharing_target_layer_name
            ):
6074
6075
6076
6077
6078
6079
6080
6081
6082
                # 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
6083
6084
6085
            # 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
6086

6087
        return kv_cache_spec
6088

6089
6090
6091
6092
6093
6094
6095
6096
6097
    def _to_list(self, sampled_token_ids: torch.Tensor) -> list[list[int]]:
        # 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.
6098
        pinned = self.sampled_token_ids_pinned_cpu[: sampled_token_ids.shape[0]]
6099
6100
6101
6102
        pinned.copy_(sampled_token_ids, non_blocking=True)
        self.transfer_event.record()
        self.transfer_event.synchronize()
        return pinned.tolist()
6103
6104
6105
6106
6107
6108
6109
6110
6111
6112
6113
6114
6115
6116
6117
6118
6119
6120
6121
6122
6123
6124
6125
6126
6127
6128
6129
6130
6131
6132
6133
6134
6135
6136
6137
6138
6139
6140
6141
6142
6143
6144
6145
6146
6147
6148
6149
6150
6151
6152
6153
6154
6155
6156
6157
6158
6159
6160
6161
6162
6163
6164
6165
6166
6167
6168
6169
6170
6171
6172
6173
6174
6175
6176
6177
6178

    def get_encoder_timing_stats(self) -> dict[str, dict[str, float | int]]:
        """
        Get encoder timing stats for all requests and clear the registry.

        Returns:
            Dictionary mapping request_id to stats dict.
        """
        with self._encoder_timing_lock:
            stats = {
                req_id: stats_obj.to_dict()
                for req_id, stats_obj in self.encoder_timing_registry.items()
            }
            self.encoder_timing_registry.clear()
            return stats

    @contextmanager
    def timed_encoder_operation(
        self,
        should_time: bool,
        group_lora_refs: list[tuple[str, Any]],
        current_item_idx: int,
        num_items: int,
    ):
        """
        Context manager to time encoder forward operations.

        Args:
            should_time: Whether timing is enabled
            group_lora_refs: Full list of (request_id, pos_info) tuples
            current_item_idx: Starting index for this group
            num_items: Number of items in this group
        """
        if not should_time:
            yield
            return

        group_refs = group_lora_refs[current_item_idx : current_item_idx + num_items]
        group_request_ids = {req_id for req_id, _ in group_refs}

        torch.cuda.synchronize()
        start_time = time.perf_counter()

        try:
            yield
        finally:
            torch.cuda.synchronize()
            elapsed = time.perf_counter() - start_time

            per_request_time = elapsed / max(len(group_request_ids), 1)

            with self._encoder_timing_lock:
                for req_id in group_request_ids:
                    if req_id not in self.encoder_timing_registry:
                        self.encoder_timing_registry[req_id] = EncoderTimingStats()

                    stats = self.encoder_timing_registry[req_id]
                    stats.encoder_forward_time += per_request_time
                    stats.num_encoder_calls += 1


@dataclass
class EncoderTimingStats:
    """Per-request timing statistics for encoder forward pass."""

    encoder_forward_time: float = 0.0
    """Time spent in vision encoder forward pass (seconds)."""

    num_encoder_calls: int = 0
    """Number of times encoder was called for this request."""

    def to_dict(self) -> dict[str, float | int]:
        return {
            "encoder_forward_time": self.encoder_forward_time,
            "num_encoder_calls": self.num_encoder_calls,
        }