gpu_model_runner.py 283 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
    get_pp_group,
    get_tp_group,
    graph_capture,
    is_global_first_rank,
王敏's avatar
王敏 committed
45
46
    get_tensor_model_parallel_rank,
    get_tensor_model_parallel_world_size,
47
48
    prepare_communication_buffer_for_model,
)
王敏's avatar
王敏 committed
49
50
51
from vllm.distributed import (
    tensor_model_parallel_all_gather
)
52
53
54
55
from vllm.forward_context import (
    BatchDescriptor,
    set_forward_context,
)
56
from vllm.logger import init_logger
57
from vllm.lora.layers import LoRAMapping, LoRAMappingType
58
from vllm.model_executor.layers.attention_layer_base import AttentionLayerBase
59
60
61
from vllm.model_executor.layers.fused_moe.routed_experts_capturer import (
    RoutedExpertsCapturer,
)
62
63
64
65
from vllm.model_executor.layers.rotary_embedding import (
    MRotaryEmbedding,
    XDRotaryEmbedding,
)
66
from vllm.model_executor.model_loader import TensorizerLoader, get_model_loader
67
from vllm.model_executor.models.interfaces import (
68
    MultiModalEmbeddings,
69
    SupportsMRoPE,
70
    SupportsMultiModal,
71
    SupportsXDRoPE,
72
73
74
75
76
    is_mixture_of_experts,
    supports_eagle3,
    supports_mrope,
    supports_multimodal_pruning,
    supports_transcription,
77
    supports_xdrope,
78
)
79
from vllm.model_executor.models.interfaces_base import (
80
81
82
83
    VllmModelForPooling,
    is_pooling_model,
    is_text_generation_model,
)
84
from vllm.multimodal import MULTIMODAL_REGISTRY
85
86
87
88
89
from vllm.multimodal.inputs import (
    BatchedTensorInputs,
    MultiModalKwargsItem,
    PlaceholderRange,
)
90
from vllm.multimodal.utils import group_mm_kwargs_by_modality
91
from vllm.pooling_params import PoolingParams
92
from vllm.sampling_params import SamplingType
93
from vllm.sequence import IntermediateTensors
94
from vllm.tasks import GenerationTask, PoolingTask, SupportedTask
95
from vllm.utils import length_from_prompt_token_ids_or_embeds
96
from vllm.utils.jsontree import json_map_leaves
97
from vllm.utils.math_utils import cdiv, round_up
98
from vllm.utils.mem_utils import DeviceMemoryProfiler, format_gib
99
from vllm.utils.nvtx_pytorch_hooks import PytHooks
100
from vllm.utils.platform_utils import is_pin_memory_available
101
102
103
104
from vllm.utils.torch_utils import (
    get_dtype_size,
    kv_cache_dtype_str_to_dtype,
)
105
106
from vllm.v1.attention.backend import (
    AttentionBackend,
107
    AttentionCGSupport,
108
    AttentionMetadata,
109
    AttentionMetadataBuilder,
110
    AttentionType,
111
    CommonAttentionMetadata,
王敏's avatar
王敏 committed
112
    CpCommonAttentionMetadata,
113
114
    MultipleOf,
)
115
from vllm.v1.attention.backends.gdn_attn import GDNAttentionMetadataBuilder
116
from vllm.v1.attention.backends.utils import (
117
    create_fast_prefill_custom_backend,
118
    get_dcp_local_seq_lens,
119
120
    reorder_batch_to_split_decodes_and_prefills,
)
121
from vllm.v1.core.sched.output import NewRequestData
122
from vllm.v1.cudagraph_dispatcher import CudagraphDispatcher
123
124
125
126
127
128
129
130
131
132
133
134
135
136
137
138
139
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,
140
    ECConnectorOutput,
141
    KVConnectorOutput,
142
143
144
145
146
    LogprobsLists,
    LogprobsTensors,
    ModelRunnerOutput,
    PoolerOutput,
    SamplerOutput,
147
    make_empty_encoder_model_runner_output,
148
)
149
from vllm.v1.pool.metadata import PoolingMetadata, PoolingStates
150
from vllm.v1.sample.logits_processor import LogitsProcessors, build_logitsprocs
151
from vllm.v1.sample.logits_processor.interface import LogitsProcessor
152
from vllm.v1.sample.metadata import SamplingMetadata
153
from vllm.v1.sample.rejection_sampler import RejectionSampler
王敏's avatar
王敏 committed
154
from vllm.v1.sample.rejection_sampler_opt import OptRejectionSampler
155
from vllm.v1.sample.sampler import Sampler
156
from vllm.v1.spec_decode.draft_model import DraftModelProposer
157
from vllm.v1.spec_decode.eagle import EagleProposer
158
from vllm.v1.spec_decode.medusa import MedusaProposer
159
from vllm.v1.spec_decode.metadata import SpecDecodeMetadata
160
from vllm.v1.spec_decode.ngram_proposer import NgramProposer
161
from vllm.v1.spec_decode.suffix_decoding import SuffixDecodingProposer
162
from vllm.v1.structured_output.utils import apply_grammar_bitmask
163
from vllm.v1.utils import CpuGpuBuffer, record_function_or_nullcontext
164
165
166
167
168
from vllm.v1.worker import mamba_utils
from vllm.v1.worker.cp_utils import (
    check_attention_cp_compatibility,
    get_total_cp_world_size,
)
169
from vllm.v1.worker.dp_utils import coordinate_batch_across_dp
170
from vllm.v1.worker.ec_connector_model_runner_mixin import ECConnectorModelRunnerMixin
171
from vllm.v1.worker.gpu_input_batch import CachedRequestState, InputBatch
172
from vllm.v1.worker.gpu_ubatch_wrapper import UBatchWrapper
173
from vllm.v1.worker.kv_connector_model_runner_mixin import KVConnectorModelRunnerMixin
174
from vllm.v1.worker.lora_model_runner_mixin import LoRAModelRunnerMixin
175
176
177
from vllm.v1.worker.ubatch_utils import (
    UBatchSlices,
    check_ubatch_thresholds,
178
    maybe_create_ubatch_slices,
179
    split_attn_metadata,
180
)
181
from vllm.v1.worker.utils import is_residual_scattered_for_sp
182
from vllm.v1.worker.workspace import lock_workspace
183

184
185
186
187
188
189
190
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
191
from vllm.v1.spec_decode.utils import DraftProbs
192
from vllm.utils.torch_utils import async_tensor_h2d
193

194
if TYPE_CHECKING:
195
    from vllm.model_executor.model_loader.tensorizer import TensorizerConfig
196
    from vllm.v1.core.sched.output import GrammarOutput, SchedulerOutput
197
198
199

logger = init_logger(__name__)

200
201
AttnMetadataDict: TypeAlias = dict[str, AttentionMetadata]
# list when ubatching is enabled
202
PerLayerAttnMetadata: TypeAlias = list[AttnMetadataDict] | AttnMetadataDict
203

204

205
206
207
208
209
210
# Wrapper for ModelRunnerOutput to support overlapped execution.
class AsyncGPUModelRunnerOutput(AsyncModelRunnerOutput):
    def __init__(
        self,
        model_runner_output: ModelRunnerOutput,
        sampled_token_ids: torch.Tensor,
211
        logprobs_tensors: LogprobsTensors | None,
212
213
        invalid_req_indices: list[int],
        async_output_copy_stream: torch.cuda.Stream,
214
        vocab_size: int,
215
216
217
218
219
    ):
        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.
220
        self.async_copy_ready_event = torch.Event()
221
222
223
224

        # 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
225
        self.vocab_size = vocab_size
226
        self._logprobs_tensors = logprobs_tensors
227
228
229
230
231

        # 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)
232
            self.sampled_token_ids_cpu = self._sampled_token_ids.to(
233
234
                "cpu", non_blocking=True
            )
235
236
237
238
239
            self._logprobs_tensors_cpu = (
                self._logprobs_tensors.to_cpu_nonblocking()
                if self._logprobs_tensors
                else None
            )
240
            self.async_copy_ready_event.record()
241
242
243

    def get_output(self) -> ModelRunnerOutput:
        """Copy the device tensors to the host and return a ModelRunnerOutput.
244

245
246
        This function blocks until the copy is finished.
        """
247
        max_gen_len = self.sampled_token_ids_cpu.shape[-1]
248
        self.async_copy_ready_event.synchronize()
249

250
251
        # Release the device tensors once the copy has completed.
        del self._logprobs_tensors
252
        del self._sampled_token_ids
253
        if max_gen_len == 1:
254
            valid_sampled_token_ids = self.sampled_token_ids_cpu.tolist()
255
256
            for i in self._invalid_req_indices:
                valid_sampled_token_ids[i].clear()
257
258
259
            logprobs_lists = None
            if self._logprobs_tensors_cpu is not None:
                logprobs_lists = self._logprobs_tensors_cpu.tolists()
260
        else:
261
            valid_sampled_token_ids, logprobs_lists = RejectionSampler.parse_output(
262
263
                self.sampled_token_ids_cpu,
                self.vocab_size,
264
                self._invalid_req_indices,
265
                logprobs_tensors=self._logprobs_tensors_cpu,
266
            )
267
268
269

        output = self._model_runner_output
        output.sampled_token_ids = valid_sampled_token_ids
270
        output.logprobs = logprobs_lists
271
272
        return output

273
274
275
    def get_output_async(self) -> ModelRunnerOutput:
        return self._model_runner_output

276

277
278
279
280
281
282
283
284
285
286
287
288
289
290
291
292
293
294
295
296
297
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)
298
            raw_pooler_output_cpu = json_map_leaves(
299
300
301
302
                lambda x: None if x is None else x.to("cpu", non_blocking=True),
                self._raw_pooler_output,
            )
            self.async_copy_ready_event.record()
303
304
305
306
            self._model_runner_output.pooler_output = [
                out if include else None
                for out, include in zip(raw_pooler_output_cpu, finished_mask)
            ]
307
308
309
310
311
312
313
314
315
316
317
318

    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


319
320
321
class ExecuteModelState(NamedTuple):
    """Ephemeral cached state transferred between execute_model() and
    sample_tokens(), after execute_model() returns None."""
322

323
324
325
326
327
328
329
    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
330
    ec_connector_output: ECConnectorOutput | None
331
    cudagraph_stats: CUDAGraphStat | None
332
    slot_mappings: dict[str, torch.Tensor] | list[dict[str, torch.Tensor]] | None
333
334


335
336
337
class GPUModelRunner(
    LoRAModelRunnerMixin, KVConnectorModelRunnerMixin, ECConnectorModelRunnerMixin
):
338
339
    def __init__(
        self,
340
        vllm_config: VllmConfig,
341
        device: torch.device,
342
    ):
343
344
345
        self.vllm_config = vllm_config
        self.model_config = vllm_config.model_config
        self.cache_config = vllm_config.cache_config
346
        self.compilation_config = vllm_config.compilation_config
347
348
349
350
351
352
        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
353

354
        from vllm.model_executor.models.utils import set_cpu_offload_max_bytes
355
356

        set_cpu_offload_max_bytes(int(self.cache_config.cpu_offload_gb * 1024**3))
357

358
359
360
361
        model_config = self.model_config
        cache_config = self.cache_config
        scheduler_config = self.scheduler_config
        parallel_config = self.parallel_config
362
        self.device = device
363
364
        self.pin_memory = is_pin_memory_available()
        self.dtype = self.model_config.dtype
365

366
367
368
        self.kv_cache_dtype = kv_cache_dtype_str_to_dtype(
            cache_config.cache_dtype, self.model_config
        )
369

370
        self.is_pooling_model = model_config.runner_type == "pooling"
371
        self.enable_prompt_embeds = model_config.enable_prompt_embeds
372
        self.is_multimodal_raw_input_only_model = (
373
374
            model_config.is_multimodal_raw_input_only_model
        )
375
376
        # This will be overridden in load_model()
        self.is_multimodal_pruning_enabled = False
377
        self.max_model_len = model_config.max_model_len
378
379
380

        # Always set to false after the first forward pass
        self.calculate_kv_scales = self.cache_config.calculate_kv_scales
王敏's avatar
王敏 committed
381
        self.tp_size = self.parallel_config.tensor_parallel_size
382
        self.dcp_world_size = self.parallel_config.decode_context_parallel_size
383
        self.dcp_rank = 0 if self.dcp_world_size <= 1 else get_dcp_group().rank_in_group
384
        self.max_num_tokens = scheduler_config.max_num_batched_tokens
王敏's avatar
王敏 committed
385
        #self.max_num_reqs = scheduler_config.max_num_seqs
386
387
388

        self.enable_lightly_cp = self.parallel_config.enable_lightly_cp
        self.enable_lightly_cplb = self.enable_lightly_cp and self.parallel_config.enable_lightly_cplb
王敏's avatar
王敏 committed
389
390
        self.max_num_reqs = (
            scheduler_config.max_num_seqs
391
            if not self.enable_lightly_cplb
王敏's avatar
王敏 committed
392
393
            else scheduler_config.max_num_seqs * 2
        )
394
        self.lightly_cp_threshould = envs.VLLM_LIGHTLY_CP_THRESHOULD
395

396
397
398
399
400
        # 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 = (
401
            self.parallel_config.distributed_executor_backend == "external_launcher"
402
            and len(get_pp_group().ranks) > 1
403
        )
404

405
        # Model-related.
406
        self.num_query_heads = model_config.get_num_attention_heads(parallel_config)
407
        self.inputs_embeds_size = model_config.get_inputs_embeds_size()
408
        self.attention_chunk_size = model_config.attention_chunk_size
409
        # Only relevant for models using ALiBi (e.g, MPT)
410
        self.use_alibi = model_config.uses_alibi
411

412
        self.cascade_attn_enabled = not self.model_config.disable_cascade_attn
413
        self.is_mm_prefix_lm = self.model_config.is_mm_prefix_lm
414

415
        # Multi-modal data support
416
        self.mm_registry = MULTIMODAL_REGISTRY
417
        self.uses_mrope = model_config.uses_mrope
guanyu1's avatar
guanyu1 committed
418
        self.use_1d_mrope = self.uses_mrope and envs.VLLM_1D_MROPE
419
        self.uses_xdrope_dim = model_config.uses_xdrope_dim
420
        self.supports_mm_inputs = self.mm_registry.supports_multimodal_inputs(
421
422
            model_config
        )
423

424
425
426
        if self.model_config.is_encoder_decoder:
            # Maximum length of the encoder input, only for encoder-decoder
            # models.
427
            self.max_encoder_len = scheduler_config.max_num_encoder_input_tokens
428
429
430
        else:
            self.max_encoder_len = 0

431
432
433
        # Async scheduling
        self.use_async_scheduling = self.scheduler_config.async_scheduling

434
        # Sampler
435
        self.sampler = Sampler(logprobs_mode=self.model_config.logprobs_mode)
436

437
        self.eplb_state: EplbState | None = None
438
439
440
441
442
443
        """
        State of the expert parallelism load balancer.

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

444
        # Lazy initializations
445
        # self.model: nn.Module  # Set after load_model
446
        # Initialize in initialize_kv_cache
447
        self.kv_caches: list[torch.Tensor] = []
448
449
450
        # Initialize in initialize_kv_cache_tensors
        self.cross_layers_kv_cache: torch.Tensor | None = None
        self.cross_layers_attn_backend: type[AttentionBackend] | None = None
451
452
        # indexes: [kv_cache_group_id][attn_group]
        self.attn_groups: list[list[AttentionGroup]] = []
453
454
        # self.kv_cache_config: KVCacheConfig

455
456
        # mm_hash ->  encoder_output
        self.encoder_cache: dict[str, torch.Tensor] = {}
457

458
        self.use_aux_hidden_state_outputs = False
459
460
461
462
463
        # 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:
464
            self.drafter: (
465
466
467
468
469
                NgramProposer
                | SuffixDecodingProposer
                | EagleProposer
                | DraftModelProposer
                | MedusaProposer
470
            )
471
472
            if self.speculative_config.method == "ngram":
                self.drafter = NgramProposer(self.vllm_config)
473
474
475
476
477
478
            elif self.speculative_config.uses_draft_model():
                self.drafter = DraftModelProposer(
                    vllm_config=self.vllm_config,
                    device=self.device,
                    runner=self,
                )
479
480
            elif self.speculative_config.method == "suffix":
                self.drafter = SuffixDecodingProposer(self.vllm_config)
481
            elif self.speculative_config.use_eagle():
482
                self.drafter = EagleProposer(self.vllm_config, self.device, self)
483
                if self.speculative_config.method == "eagle3":
484
485
486
                    self.use_aux_hidden_state_outputs = (
                        self.drafter.eagle3_use_aux_hidden_state
                    )
487
488
            elif self.speculative_config.method == "medusa":
                self.drafter = MedusaProposer(
489
                    vllm_config=self.vllm_config, device=self.device
490
                )
491
            else:
492
493
494
495
                raise ValueError(
                    "Unknown speculative decoding method: "
                    f"{self.speculative_config.method}"
                )
王敏's avatar
王敏 committed
496
497
498
499
500
            
            if not envs.VLLM_REJECT_SAMPLE_OPT:
                self.rejection_sampler = RejectionSampler(self.sampler)
            else:
                self.rejection_sampler = OptRejectionSampler(self.sampler)
501

502
503
504
        self.num_spec_tokens = 0
        if self.speculative_config:
            self.num_spec_tokens = self.speculative_config.num_speculative_tokens
505
506
507
508
509
            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
510

511
        # Request states.
512
        self.requests: dict[str, CachedRequestState] = {}
513
514
515
        # NOTE(rob): num_prompt_logprobs only includes reqs
        # that are currently in the prefill phase.
        self.num_prompt_logprobs: dict[str, int] = {}
516
        self.comm_stream = torch.cuda.Stream()
517

518
519
520
521
522
523
524
525
526
        # 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.
527
528
529
530
        logits_processors = model_config.logits_processors
        custom_logitsprocs: Sequence[str | type[LogitsProcessor]] = (
            tuple(logits_processors) if logits_processors is not None else ()
        )
531
532
        self.input_batch = InputBatch(
            max_num_reqs=self.max_num_reqs,
533
534
535
            # 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),
536
537
538
            max_num_batched_tokens=self.max_num_tokens,
            device=self.device,
            pin_memory=self.pin_memory,
539
            vocab_size=self.model_config.get_vocab_size(),
540
            block_sizes=[self.cache_config.block_size],
541
            kernel_block_sizes=[self.cache_config.block_size],
542
            is_spec_decode=bool(self.vllm_config.speculative_config),
543
            logitsprocs=build_logitsprocs(
544
545
546
                self.vllm_config,
                self.device,
                self.pin_memory,
547
                self.is_pooling_model,
548
                custom_logitsprocs,
549
            ),
550
551
552
            # 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),
553
            is_pooling_model=self.is_pooling_model,
554
            cp_kv_cache_interleave_size=self.parallel_config.cp_kv_cache_interleave_size,
555
        )
556

557
558
559
560
561
        # 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.
562
        self.prepare_inputs_event: torch.Event | None = None
563
564
        if self.use_async_scheduling:
            self.async_output_copy_stream = torch.cuda.Stream()
565
            self.prepare_inputs_event = torch.Event()
566

567
        # self.cudagraph_batch_sizes sorts in ascending order.
568
569
570
571
        if (
            self.compilation_config.cudagraph_capture_sizes
            and self.compilation_config.cudagraph_mode != CUDAGraphMode.NONE
        ):
572
573
            self.cudagraph_batch_sizes = sorted(
                self.compilation_config.cudagraph_capture_sizes
574
            )
575

576
        # Cache the device properties.
577
        self._init_device_properties()
578

579
580
581
582
        # Encoder timing registry for observability
        self.encoder_timing_registry: dict[str, EncoderTimingStats] = {}
        self._encoder_timing_lock = threading.Lock()

583
        # Persistent buffers for CUDA graphs.
584
585
586
587
588
        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
        )
589
        self.seq_lens = self._make_buffer(self.max_num_reqs, dtype=torch.int32)
590
        self.encoder_seq_lens = self._make_buffer(self.max_num_reqs, dtype=torch.int32)
591
592
593
594
        if self.dcp_world_size > 1:
            self.dcp_local_seq_lens = self._make_buffer(
                self.max_num_reqs, dtype=torch.int32
            )
595
596
597
        # 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.
598
        self.inputs_embeds = self._make_buffer(
599
            self.max_num_tokens, self.inputs_embeds_size, dtype=self.dtype, numpy=False
600
601
        )
        self.is_token_ids = self._make_buffer(self.max_num_tokens, dtype=torch.bool)
602
603
        self.discard_request_mask = self._make_buffer(
            self.max_num_reqs, dtype=torch.bool
604
605
606
607
608
609
610
        )
        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
        )
611

612
613
        # Only relevant for multimodal models
        if self.supports_mm_inputs:
614
615
616
617
618
619
620
            # 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
621
622

        # Only relevant for models using M-RoPE (e.g, Qwen2-VL)
623
        if self.uses_mrope:
Roger Wang's avatar
Roger Wang committed
624
625
626
627
            # 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
628
629
630
631
632
633

            # 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
guanyu1's avatar
guanyu1 committed
634
635
636
637
638
639
640
641
            if self.use_1d_mrope:
                self.mrope_positions = self._make_buffer(
                    3 * (self.max_num_tokens + 1), dtype=torch.int64
                )
            else:
                self.mrope_positions = self._make_buffer(
                    (3, self.max_num_tokens + 1), dtype=torch.int64
                )
642

643
644
645
646
647
648
        # 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
            )
649

650
        # None in the first PP rank. The rest are set after load_model.
651
        self.intermediate_tensors: IntermediateTensors | None = None
652

653
        # OPTIMIZATION: Cache the tensors rather than creating them every step.
654
        # Keep in int64 to avoid overflow with long context
655
656
657
658
        self.arange_np = np.arange(
            max(self.max_num_reqs + 1, self.max_model_len, self.max_num_tokens),
            dtype=np.int64,
        )
659

660
661
662
663
664
        # 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] = {}
665
666
667
668
669
        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(
670
671
                self.max_num_tokens, dtype=torch.int32, device=self.device
            )
672

673
        self.uniform_decode_query_len = 1 + self.num_spec_tokens
674
675
676
677

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

678
        self.mm_budget = (
679
            MultiModalBudget(self.vllm_config, self.mm_registry)
680
681
682
            if self.supports_mm_inputs
            else None
        )
683

684
        self.reorder_batch_threshold: int | None = None
685

686
687
688
689
690
        # 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()

691
        # Cached outputs.
692
        self._draft_token_ids: list[list[int]] | torch.Tensor | None = None
693
        self._draft_token_req_ids: list[str] | None = None
694
        self.transfer_event = torch.Event()
695
        self.sampled_token_ids_pinned_cpu = torch.empty(
696
            (self.max_num_reqs, 1),
697
698
            dtype=torch.int64,
            device="cpu",
699
700
            pin_memory=self.pin_memory,
        )
701

702
703
        # Pre-allocated tensor for copying valid sampled token counts to CPU,
        # with dedicated stream for overlapping and event for coordination.
704
        self.valid_sampled_token_count_event: torch.Event | None = None
705
        self.valid_sampled_token_count_copy_stream: torch.cuda.Stream | None = None
706
707
708
709
710
711
712
713
714
715
716
717
718
719
720
721
722
723
724
725
726
727
728
729
        # 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,
                )
730

731
732
        # Ephemeral state transferred between execute_model() and sample_tokens().
        self.execute_model_state: ExecuteModelState | None = None
733
        self.kv_connector_output: KVConnectorOutput | None = None
734
        self.mamba_state_idx: dict[str, int] = {}
735
        self.layerwise_nvtx_hooks_registered = False
736

王敏's avatar
王敏 committed
737
738
        self.draft_probs : Optional[DraftProbs] = None

739
740
741
742
743
744
745
    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

746
747
748
749
    def reset_mm_cache(self) -> None:
        if self.mm_budget:
            self.mm_budget.reset_cache()

750
751
752
753
754
755
756
757
758
759
760
761
762
763
764
765
766
767
768
769
770
771
772
773
774
775
776
777
778
779
780
781
782
783
784
785
786
787
788
789
790
791
792
793
    @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)

794
795
796
    def _get_positions(self, num_tokens: Any):
        if isinstance(num_tokens, int):
            if self.uses_mrope:
guanyu1's avatar
guanyu1 committed
797
798
799
800
                if self.use_1d_mrope:
                    return self.mrope_positions.gpu[: 3 * num_tokens].view(
                        num_tokens, 3
                    ).T
801
                return self.mrope_positions.gpu[:, :num_tokens]
802
803
            if self.uses_xdrope_dim > 0:
                return self.xdrope_positions.gpu[:, :num_tokens]
804
805
806
            return self.positions.gpu[:num_tokens]
        else:
            if self.uses_mrope:
guanyu1's avatar
guanyu1 committed
807
808
                if self.use_1d_mrope:
                    return self.mrope_positions.gpu.view(-1, 3)[num_tokens].T
809
                return self.mrope_positions.gpu[:, num_tokens]
810
811
            if self.uses_xdrope_dim > 0:
                return self.xdrope_positions.gpu[:, num_tokens]
812
813
            return self.positions.gpu[num_tokens]

guanyu1's avatar
guanyu1 committed
814

815
    def _make_buffer(
816
        self, *size: int | torch.SymInt, dtype: torch.dtype, numpy: bool = True
817
818
819
820
821
822
823
824
    ) -> CpuGpuBuffer:
        return CpuGpuBuffer(
            *size,
            dtype=dtype,
            device=self.device,
            pin_memory=self.pin_memory,
            with_numpy=numpy,
        )
guanyu1's avatar
guanyu1 committed
825

guanyu1's avatar
guanyu1 committed
826
827
828
    def _copy_mrope_positions_to_gpu(self, num_tokens: int) -> None:
        if not self.uses_mrope:
            return
guanyu1's avatar
guanyu1 committed
829
830
831
832
833
834
835
        if self.use_1d_mrope:
            num_values = 3 * num_tokens
            self.mrope_positions.gpu[:num_values].copy_(
                self.mrope_positions.cpu[:num_values],
                non_blocking=True,
            )
            return
836
837
838
839
        # self.mrope_positions.gpu[:, :num_tokens].copy_(
        #     self.mrope_positions.cpu[:, :num_tokens],
        #     non_blocking=True,
        # )
guanyu1's avatar
guanyu1 committed
840
        self.mrope_positions.gpu[:, :num_tokens].copy_(
841
            self.mrope_positions.cpu[:, :num_tokens].contiguous().pin_memory(),
guanyu1's avatar
guanyu1 committed
842
843
844
845
846
847
            non_blocking=True,
        )

    def _copy_xdrope_positions_to_gpu(self, num_tokens: int) -> None:
        if self.uses_xdrope_dim <= 0:
            return
guanyu1's avatar
guanyu1 committed
848

guanyu1's avatar
guanyu1 committed
849
850
851
852
853
        self.xdrope_positions.gpu[:, :num_tokens].copy_(
            self.xdrope_positions.cpu[:, :num_tokens],
            non_blocking=True,
        )

854

855
    def _init_model_kwargs(self):
856
857
        model_kwargs = dict[str, Any]()

858
        if not self.is_pooling_model:
859
860
            return model_kwargs

861
862
        num_reqs = self.input_batch.num_reqs
        pooling_params = self.input_batch.get_pooling_params()
863
864
865

        token_type_id_requests = dict[int, Any]()
        for i, param in enumerate(pooling_params):
866
867
868
869
870
            if (
                param.extra_kwargs is not None
                and (token_types := param.extra_kwargs.get("compressed_token_type_ids"))
                is not None
            ):
871
872
873
874
875
                token_type_id_requests[i] = token_types

        if len(token_type_id_requests) == 0:
            return model_kwargs

876
        seq_lens = self.seq_lens.gpu[:num_reqs]
877
878
879
880
881
882
883
884
        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(
885
886
            device=self.device
        )
887
        return model_kwargs
888

889
    def _may_reorder_batch(self, scheduler_output: "SchedulerOutput") -> None:
890
891
        """
        Update the order of requests in the batch based on the attention
892
        backend's needs. For example, some attention backends (namely MLA) may
893
894
895
896
897
898
        want to separate requests based on if the attention computation will be
        compute-bound or memory-bound.

        Args:
            scheduler_output: The scheduler output.
        """
899
900
901
902
903
904
905
906
        # 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

907
908
909
910
        if self.reorder_batch_threshold is not None:
            reorder_batch_to_split_decodes_and_prefills(
                self.input_batch,
                scheduler_output,
911
912
                decode_threshold=self.reorder_batch_threshold,
            )
913

914
915
    # Note: used for model runner override.
    def _init_device_properties(self) -> None:
916
        """Initialize attributes from torch.cuda.get_device_properties"""
917
918
919
920
921
922
923
        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()

924
    def _update_states(self, scheduler_output: "SchedulerOutput") -> None:
925
926
927
928
929
930
        """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.

931
932
        The SamplingMetadata is updated and copied to the GPU if there is a
        new/resumed/paused/finished request in the batch.
933
934
        """
        # Remove finished requests from the cached states.
935
936
        for req_id in scheduler_output.finished_req_ids:
            self.requests.pop(req_id, None)
937
            self.num_prompt_logprobs.pop(req_id, None)
938
939
940
941
942
943
944
        # 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:
945
            self.input_batch.remove_request(req_id)
946

王敏's avatar
王敏 committed
947
948
949
950
        # 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))

951
        # Free the cached encoder outputs.
952
953
        for mm_hash in scheduler_output.free_encoder_mm_hashes:
            self.encoder_cache.pop(mm_hash, None)
954

955
956
957
958
959
960
961
        # 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()
962
963
964
965
966
967
968
969
        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)
970
971
972
973
974
        # 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:
975
            self.input_batch.remove_request(req_id)
976

977
        reqs_to_add: list[CachedRequestState] = []
978
        # Add new requests to the cached states.
979
980
        for new_req_data in scheduler_output.scheduled_new_reqs:
            req_id = new_req_data.req_id
981
982
983
984
985
986
            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

987
            sampling_params = new_req_data.sampling_params
988
            pooling_params = new_req_data.pooling_params
989

990
991
992
993
            if (
                sampling_params
                and sampling_params.sampling_type == SamplingType.RANDOM_SEED
            ):
994
995
996
997
998
                generator = torch.Generator(device=self.device)
                generator.manual_seed(sampling_params.seed)
            else:
                generator = None

999
1000
            if self.is_pooling_model:
                assert pooling_params is not None
1001
1002
                task = pooling_params.task
                assert task is not None, "You did not set `task` in the API"
1003

1004
                model = cast(VllmModelForPooling, self.get_model())
1005
                to_update = model.pooler.get_pooling_updates(task)
1006
1007
                to_update.apply(pooling_params)

1008
            req_state = CachedRequestState(
1009
                req_id=req_id,
1010
                prompt_token_ids=new_req_data.prompt_token_ids,
1011
                prompt_embeds=new_req_data.prompt_embeds,
1012
                mm_features=new_req_data.mm_features,
1013
                sampling_params=sampling_params,
1014
                pooling_params=pooling_params,
1015
                generator=generator,
1016
1017
                block_ids=new_req_data.block_ids,
                num_computed_tokens=new_req_data.num_computed_tokens,
1018
                output_token_ids=[],
1019
                lora_request=new_req_data.lora_request,
1020
            )
1021
            self.requests[req_id] = req_state
1022

1023
1024
1025
1026
1027
1028
1029
            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
                )

1030
            # Only relevant for models using M-RoPE (e.g, Qwen2-VL)
1031
            if self.uses_mrope:
1032
                self._init_mrope_positions(req_state)
1033

1034
1035
1036
1037
            # Only relevant for models using XD-RoPE (e.g, HunYuan-VL)
            if self.uses_xdrope_dim > 0:
                self._init_xdrope_positions(req_state)

1038
            reqs_to_add.append(req_state)
1039

1040
        # Update the states of the running/resumed requests.
1041
        is_last_rank = get_pp_group().is_last_rank
1042
        req_data = scheduler_output.scheduled_cached_reqs
1043
        scheduled_spec_tokens = scheduler_output.scheduled_spec_decode_tokens
1044
1045
1046
1047
1048

        # 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()

1049
        for i, req_id in enumerate(req_data.req_ids):
1050
            req_state = self.requests[req_id]
1051
1052
            num_computed_tokens = req_data.num_computed_tokens[i]
            new_block_ids = req_data.new_block_ids[i]
1053
            resumed_from_preemption = req_id in req_data.resumed_req_ids
1054
            num_output_tokens = req_data.num_output_tokens[i]
1055
            req_index = self.input_batch.req_id_to_index.get(req_id)
1056

1057
1058
1059
1060
1061
1062
1063
1064
1065
1066
1067
1068
1069
1070
            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.
1071
1072
1073
1074
1075
1076
1077
1078
1079
                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)
1080

1081
            # Update the cached states.
1082
            req_state.num_computed_tokens = num_computed_tokens
1083
1084
1085
1086
1087
1088
1089
1090

            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.
1091
1092
1093
                num_new_tokens = (
                    num_computed_tokens + len(new_token_ids) - req_state.num_tokens
                )
1094
1095
1096
1097
                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:
1098
                    req_state.output_token_ids.extend(new_token_ids[-num_new_tokens:])
1099
1100
1101
1102
1103
            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:
1104
1105
1106
1107
                    end_idx = (
                        self.input_batch.num_prompt_tokens[req_index]
                        + num_output_tokens
                    )
1108
                    self.input_batch.num_tokens_no_spec[req_index] = end_idx
1109

1110
            # Update the block IDs.
1111
            if not resumed_from_preemption:
1112
1113
                if new_block_ids is not None:
                    # Append the new blocks to the existing block IDs.
1114
                    for block_ids, new_ids in zip(req_state.block_ids, new_block_ids):
1115
                        block_ids.extend(new_ids)
1116
            else:
1117
                assert req_index is None
1118
                assert new_block_ids is not None
1119
1120
                # The request is resumed from preemption.
                # Replace the existing block IDs with the new ones.
1121
                req_state.block_ids = new_block_ids
1122
1123
1124
1125
1126

            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.
1127
1128
1129
1130
1131
1132
1133

                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:]

1134
                reqs_to_add.append(req_state)
1135
1136
1137
                continue

            # Update the persistent batch.
1138
            self.input_batch.num_computed_tokens_cpu[req_index] = num_computed_tokens
1139
            if new_block_ids is not None:
1140
                self.input_batch.block_table.append_row(new_block_ids, req_index)
1141
1142
1143
1144
1145
1146

            # 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
1147
                end_token_index = num_computed_tokens + len(new_token_ids)
1148
                self.input_batch.token_ids_cpu[
1149
1150
1151
                    req_index, start_token_index:end_token_index
                ] = new_token_ids
                self.input_batch.num_tokens_no_spec[req_index] = end_token_index
1152

1153
            # Add spec_token_ids to token_ids_cpu.
1154
            self.input_batch.update_req_spec_token_ids(req_state, scheduled_spec_tokens)
1155

1156
1157
        # Add the new or resumed requests to the persistent batch.
        # The smaller empty indices are filled first.
1158
1159
        for request in reqs_to_add:
            self.input_batch.add_request(request)
1160
            self.input_batch.update_req_spec_token_ids(request, scheduled_spec_tokens)
1161

1162
1163
1164
1165
1166
        # 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.
1167
1168
1169
1170
1171
1172
1173
1174
1175
1176
1177
        repeat_counts = None
        if envs.VLLM_REJECT_SAMPLE_OPT and \
                scheduler_output.scheduled_spec_decode_tokens:
            repeat_counts = [1] * self.input_batch.num_reqs
            for req_id, draft_token_ids in (
                    scheduler_output.scheduled_spec_decode_tokens.items()):
                req_idx = self.input_batch.req_id_to_index.get(req_id)
                if req_idx is not None:
                    repeat_counts[req_idx] += len(draft_token_ids)
            repeat_counts = torch.tensor(repeat_counts, dtype=torch.int32, device="cpu")
        self.input_batch.refresh_metadata(repeat_counts)
1178

1179
    def _update_states_after_model_execute(
1180
        self, output_token_ids: torch.Tensor, scheduler_output: "SchedulerOutput"
1181
    ) -> None:
1182
1183
1184
1185
1186
1187
1188
1189
        """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.
        """
1190
        if not self.speculative_config or not self.model_config.is_hybrid:
1191
1192
1193
            return

        # Find the number of accepted tokens for each sequence.
1194
1195
1196
1197
1198
1199
1200
1201
1202
1203
1204
1205
1206
1207
1208
1209
1210
1211
1212
1213
        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()
        )
1214
1215
        for i, num_tokens in enumerate(num_accepted_tokens):
            self.input_batch.num_accepted_tokens_cpu[i] = num_tokens
1216
1217
1218
1219
1220
1221
1222
1223
1224
1225
        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(),
            )
1226

1227
1228
1229
1230
1231
1232
1233
1234
1235
1236
1237
1238
1239
1240
1241
1242
1243
1244
1245
1246
1247
1248
1249
1250
1251
1252
1253
1254
1255
1256
1257
1258
1259
    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
1260

1261
    def _init_mrope_positions(self, req_state: CachedRequestState):
1262
1263
        model = self.get_model()
        assert supports_mrope(model), "M-RoPE support is not implemented."
1264
1265
1266
1267
        assert req_state.prompt_token_ids is not None, (
            "M-RoPE requires prompt_token_ids to be available."
        )
        mrope_model = cast(SupportsMRoPE, model)
1268
1269

        req_state.mrope_positions, req_state.mrope_position_delta = (
1270
            mrope_model.get_mrope_input_positions(
1271
                req_state.prompt_token_ids,
1272
                req_state.mm_features,
1273
            )
1274
        )
1275

1276
1277
1278
1279
1280
1281
1282
1283
1284
1285
1286
1287
    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,
        )
1288

1289
    def _extract_mm_kwargs(
1290
        self,
1291
1292
        scheduler_output: "SchedulerOutput",
    ) -> BatchedTensorInputs:
1293
        if not scheduler_output or not self.is_multimodal_raw_input_only_model:
1294
            return {}
1295

1296
1297
        mm_kwargs = list[MultiModalKwargsItem]()
        for req in scheduler_output.scheduled_new_reqs:
1298
1299
1300
            for feature in req.mm_features:
                if feature.data is not None:
                    mm_kwargs.append(feature.data)
1301

1302
1303
1304
        # Input all modalities at once
        mm_kwargs_combined: BatchedTensorInputs = {}
        for _, _, mm_kwargs_group in group_mm_kwargs_by_modality(
1305
1306
1307
            mm_kwargs,
            device=self.device,
            pin_memory=self.pin_memory,
1308
1309
        ):
            mm_kwargs_combined.update(mm_kwargs_group)
1310

1311
        return mm_kwargs_combined
1312
1313

    def _dummy_mm_kwargs(self, num_seqs: int) -> BatchedTensorInputs:
1314
        if not self.is_multimodal_raw_input_only_model:
1315
            return {}
1316

1317
1318
        mm_budget = self.mm_budget
        assert mm_budget is not None
1319

1320
1321
        dummy_modality = mm_budget.get_modality_with_max_tokens()
        return self._get_mm_dummy_batch(dummy_modality, num_seqs)
1322

1323
1324
1325
    def _get_cumsum_and_arange(
        self,
        num_tokens: np.ndarray,
1326
        cumsum_dtype: np.dtype | None = None,
1327
1328
1329
1330
1331
1332
1333
1334
1335
1336
1337
1338
1339
1340
1341
1342
    ) -> 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

1343
    def _prepare_input_ids(
1344
1345
1346
1347
        self,
        scheduler_output: "SchedulerOutput",
        total_num_scheduled_tokens: int,
        cu_num_tokens: np.ndarray,
1348
    ) -> None:
1349
        """Prepare the input IDs for the current batch.
1350

1351
1352
1353
1354
1355
1356
1357
        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)
1358
1359
1360
            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)
1361
1362
1363
1364
1365
1366
1367
            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
1368
1369
1370
1371
        sample_flattened_indices: list[int] = []
        spec_flattened_indices: list[int] = []
        prev_common_req_indices: list[int] = []
        prev_draft_token_indices: list[int] = []
1372
1373
        indices_match = True
        max_flattened_index = -1
1374
1375
1376
        total_num_spec_tokens = 0
        scheduled_spec_tokens = scheduler_output.scheduled_spec_decode_tokens

1377
1378
1379
1380
1381
        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.
1382
1383
                draft_len = len(scheduled_spec_tokens.get(req_id, ()))
                total_num_spec_tokens += draft_len
1384
                flattened_index = cu_num_tokens[cur_index].item() - 1
1385
1386
1387
1388
1389
1390
1391
1392
1393
1394
1395
1396
1397
1398
1399
                # 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))
1400
                indices_match &= prev_index == flattened_index
1401
                max_flattened_index = max(max_flattened_index, flattened_index)
1402
1403
1404
        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:
1405
1406
1407
            # 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)
1408
1409
1410
            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)
1411
1412
        if num_commmon_tokens == 0:
            # No requests in common with the previous iteration
1413
            # So input_ids.cpu will have all the input ids.
1414
1415
1416
1417
1418
1419
1420
            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_(
1421
1422
1423
                self.input_batch.prev_sampled_token_ids[:num_commmon_tokens, 0],
                non_blocking=True,
            )
1424
1425
            if self.enable_prompt_embeds:
                self.is_token_ids.gpu[:num_commmon_tokens] = True
1426
            return
1427
        # Upload the index tensors asynchronously so the scatter can be non-blocking.
1428
1429
        sampled_tokens_index_tensor = torch.tensor(
            sample_flattened_indices, dtype=torch.int64, pin_memory=self.pin_memory
1430
        ).to(self.device, non_blocking=True)
1431
        prev_common_req_indices_tensor = torch.tensor(
1432
1433
            prev_common_req_indices, dtype=torch.int64, pin_memory=self.pin_memory
        ).to(self.device, non_blocking=True)
1434
1435
        self.input_ids.gpu.scatter_(
            dim=0,
1436
            index=sampled_tokens_index_tensor,
1437
            src=self.input_batch.prev_sampled_token_ids[
1438
1439
1440
                prev_common_req_indices_tensor, 0
            ],
        )
1441

1442
1443
1444
1445
1446
1447
1448
1449
1450
1451
1452
1453
1454
1455
1456
1457
1458
1459
1460
1461
1462
        # 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],
        )
1463

1464
1465
    def _get_encoder_seq_lens(
        self,
1466
        num_scheduled_tokens: dict[str, int],
1467
1468
        kv_cache_spec: KVCacheSpec,
        num_reqs: int,
1469
        for_cudagraph_capture: bool = False,
1470
    ) -> tuple[torch.Tensor | None, np.ndarray | None]:
1471
        if not isinstance(kv_cache_spec, CrossAttentionSpec):
1472
            return None, None
1473

1474
1475
        # Zero out buffer for padding requests that are not actually scheduled (CGs)
        self.encoder_seq_lens.np[:num_reqs] = 0
1476

1477
1478
        # Build encoder_seq_lens array mapping request indices to
        # encoder lengths for inputs scheduled in this batch
1479
        for req_id in num_scheduled_tokens:
1480
            req_index = self.input_batch.req_id_to_index[req_id]
1481
1482
1483
1484
1485
1486
1487
1488
1489
1490
1491
1492
            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
1493
1494
1495
1496
1497
1498
1499
1500
1501
        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
1502

1503
1504
1505
        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]
1506

1507
        return encoder_seq_lens, encoder_seq_lens_cpu
1508

王敏's avatar
王敏 committed
1509
1510
1511
1512
1513
1514
1515
1516
1517
1518
1519
1520
1521
1522
1523
1524
1525
1526
1527
1528
1529
1530
1531
    def _distribute_tokens_to_cp_ranks(
        self,
        total_q_len: int,
        q_lens_cpu: np.ndarray,
        kv_lens_cpu: np.ndarray,
        tp_rank: int,
        tp_size: int,
        req_ids: list[str],
    ):
        tokens_per_rank = (total_q_len + tp_size - 1) // tp_size

        start_token = tp_rank * tokens_per_rank
        end_token = min((tp_rank + 1) * tokens_per_rank, total_q_len)

        q_lens = []
        seq_count = 0
        seq_indexes = []
        kv_lens = []
        local_req_ids = []

        local_scatter_indexes_tensor = None
        gather_indexes_tensor = None

1532
        if self.enable_lightly_cp:
王敏's avatar
王敏 committed
1533
1534
1535
1536
1537
1538
1539
1540
1541
1542
1543
1544
1545
1546
1547
1548
1549
1550
1551
1552
1553
1554
1555
1556
1557
1558
1559
1560
1561
1562
1563
1564
1565
1566
1567
1568
1569
1570
1571
1572
1573
1574
1575
1576
1577
1578
1579
1580
1581
1582
1583
1584
1585
1586
1587
1588
1589
1590
1591
1592
1593
1594
1595
1596
1597
1598
1599
1600
1601
1602
1603
1604
1605
1606
1607
1608
1609
1610
1611
1612
1613
1614
1615
1616
1617
1618
1619
1620
1621
1622
1623
1624
1625
1626
1627
1628
1629
1630
1631
1632
1633
1634
1635
1636
1637
1638
1639
1640
1641
1642
1643
1644
1645
1646
1647
1648
1649
1650
1651
1652
1653
1654
            rank_tokens = 0
            rank_pad_tokens = 0
            accu_q_start = 0
            scatter_indexes: list[int] = []
            num_requests = len(q_lens_cpu)
            for i in range(num_requests):
                req_q_len = q_lens_cpu[i]
                req_pad_q_len = round_up(q_lens_cpu[i], 2 * tp_size)
                kv_len = kv_lens_cpu[i]

                chunk_q_len = req_pad_q_len // (2 * tp_size)

                q_1_start = tp_rank * chunk_q_len
                q_1_end = (tp_rank + 1) * chunk_q_len
                q_2_start = req_pad_q_len - (tp_rank + 1) * chunk_q_len
                q_2_end = req_pad_q_len - tp_rank * chunk_q_len

                q_len_1 = (
                    chunk_q_len
                    if q_1_end <= req_q_len
                    else max(0, req_q_len - q_1_start)
                )
                q_len_2 = (
                    chunk_q_len
                    if q_2_end <= req_q_len
                    else max(0, req_q_len - q_2_start)
                )

                kv_len_1 = kv_len - req_q_len + min(req_q_len, q_1_end)
                kv_len_2 = kv_len - req_q_len + min(req_q_len, q_2_end)

                scatter_index1 = range(
                    accu_q_start + q_1_start, accu_q_start + q_1_start + q_len_1
                )

                scatter_index2 = range(
                    accu_q_start + q_2_start, accu_q_start + q_2_start + q_len_2
                )
                accu_q_start += req_q_len

                if q_len_1 > 0:
                    q_lens.append(q_len_1)
                    kv_lens.append(kv_len_1)
                    seq_indexes.append(i)
                    local_req_ids.append(req_ids[i])
                    scatter_indexes.extend(scatter_index1)
                    seq_count += 1
                    rank_tokens += q_len_1

                if q_len_2 > 0:
                    q_lens.append(q_len_2)
                    kv_lens.append(kv_len_2)
                    seq_indexes.append(i)
                    local_req_ids.append(req_ids[i])
                    scatter_indexes.extend(scatter_index2)
                    seq_count += 1
                    rank_tokens += q_len_2

                rank_pad_tokens += chunk_q_len * 2

            if len(scatter_indexes) < rank_pad_tokens:
                scatter_indexes.extend([-1] * (rank_pad_tokens - len(scatter_indexes)))

            local_scatter_indexes_tensor = torch.tensor(
                scatter_indexes, dtype=torch.int64, device=self.device
            )
            global_scatter_indexes_tensor = tensor_model_parallel_all_gather(
                local_scatter_indexes_tensor.contiguous(), dim=0
            )
            non_neg_mask = global_scatter_indexes_tensor != -1
            non_neg_values = global_scatter_indexes_tensor[non_neg_mask]
            non_neg_positions = torch.where(non_neg_mask)[0]
            sorted_indices = torch.argsort(non_neg_values)
            gather_indexes_tensor = non_neg_positions[sorted_indices]
            if isinstance(rank_tokens, torch.Tensor):
                rank_tokens = rank_tokens.item()
        else:
            current_seq = 0
            current_pos = 0
            rank_tokens = min(tokens_per_rank, end_token - start_token)
            while start_token < end_token and current_seq < len(q_lens_cpu):
                q_len = q_lens_cpu[current_seq]
                q_start = current_pos
                q_end = current_pos + q_len
                kv_len = kv_lens_cpu[current_seq]

                # Find overlap between this sequence and rank's token range
                overlap_start = max(start_token, q_start)
                overlap_end = min(end_token, q_end)

                if overlap_start < overlap_end:
                    # This sequence contributes tokens to this rank
                    token_count = overlap_end - overlap_start
                    q_lens.append(token_count)
                    start_token = overlap_end
                    seq_count += 1
                    seq_indexes.append(current_seq)
                    local_req_ids.append(req_ids[current_seq])

                    if q_end <= end_token:
                        kv_lens.append(kv_len)
                    else:
                        kv_lens.append(kv_len - (q_end - end_token))

                current_pos = q_end
                current_seq += 1

        return (
            rank_tokens,
            np.array(q_lens, dtype=np.int32),
            seq_count,
            np.array(kv_lens, dtype=np.int32),
            np.array(local_req_ids, dtype=str),
            local_scatter_indexes_tensor,
            gather_indexes_tensor,
            seq_indexes,
        )

    def _prepare_cp_metadata(
        self,
        num_reqs_padded,
        max_query_len,
王敏's avatar
王敏 committed
1655
        max_seq_len,
王敏's avatar
王敏 committed
1656
1657
1658
        num_tokens,
        block_table_gid_0,
        slot_mapping_gid_0,
王敏's avatar
王敏 committed
1659
        num_computed_tokens_cpu
王敏's avatar
王敏 committed
1660
1661
1662
1663
1664
1665
1666
1667
1668
1669
    ):
        tp_size = self.vllm_config.parallel_config.tensor_parallel_size
        tp_rank = get_tensor_model_parallel_rank()

        cp_common_metadata = CpCommonAttentionMetadata(
            query_start_loc=self.query_start_loc.gpu[: num_reqs_padded + 1].clone(),
            query_start_loc_cpu=self.query_start_loc.cpu[: num_reqs_padded + 1].clone(),
            seq_lens=self.seq_lens.gpu[:num_reqs_padded].clone(),
            _seq_lens_cpu=self.seq_lens.cpu[:num_reqs_padded].clone(),
            max_query_len=max_query_len,
王敏's avatar
王敏 committed
1670
            max_seq_len=max_seq_len,
王敏's avatar
王敏 committed
1671
1672
1673
            num_reqs=num_reqs_padded,
            req_ids=self.input_batch.req_ids,
            num_actual_tokens=num_tokens,
王敏's avatar
王敏 committed
1674
1675
1676
1677
            num_kv_actual_tokens=num_tokens,
            block_table_tensor=block_table_gid_0,
            slot_mapping=slot_mapping_gid_0,
            _num_computed_tokens_cpu=num_computed_tokens_cpu
王敏's avatar
王敏 committed
1678
1679
1680
1681
1682
1683
1684
1685
1686
1687
1688
1689
1690
1691
1692
1693
1694
1695
1696
1697
1698
1699
1700
1701
1702
1703
1704
1705
1706
1707
1708
1709
1710
1711
1712
1713
1714
1715
1716
1717
1718
1719
1720
1721
1722
1723
1724
1725
1726
1727
1728
1729
1730
1731
1732
1733
1734
1735
1736
1737
1738
1739
1740
1741
1742
        )

        query_start_loc_cpu = self.query_start_loc.cpu[: num_reqs_padded + 1]
        q_lens_cpu = query_start_loc_cpu[1:] - query_start_loc_cpu[:-1]
        kv_lens_cpu = self.seq_lens.cpu[:num_reqs_padded]
        total_q_len = num_tokens
        total_kv_len = num_tokens

        (
            total_q_len,
            q_lens_cpu,
            seq_count,
            kv_lens_cpu,
            local_req_ids,
            scatter_indexes_tensor,
            gather_indexes_tensor,
            seq_indexes_list,
        ) = self._distribute_tokens_to_cp_ranks(
            total_q_len,
            q_lens_cpu,
            kv_lens_cpu,
            tp_rank,
            tp_size,
            self.input_batch.req_ids,
        )

        num_reqs = seq_count

        cu_num_tokens = np.cumsum(q_lens_cpu)
        self.query_start_loc.np[0] = 0
        self.query_start_loc.np[1 : num_reqs + 1] = cu_num_tokens
        self.query_start_loc.np[num_reqs + 1 :].fill(cu_num_tokens[-1])
        self.query_start_loc.copy_to_gpu()
        q_acc_lens = self.query_start_loc.gpu[: num_reqs + 1]
        q_acc_lens_cpu = self.query_start_loc.cpu[: num_reqs + 1]
        max_q_len = max(q_acc_lens_cpu)

        self.seq_lens.np[:num_reqs] = kv_lens_cpu
        self.seq_lens.np[num_reqs:].fill(0)
        self.seq_lens.copy_to_gpu()
        kv_lens = self.seq_lens.gpu[:num_reqs]
        kv_lens_cpu = self.seq_lens.cpu[:num_reqs]
        max_kv_len = max(kv_lens_cpu)

        num_computed_tokens_cpu = kv_lens_cpu - q_acc_lens_cpu[1:]
        blk_table_tensor = block_table_gid_0[seq_indexes_list]

        cm_base = CommonAttentionMetadata(
            query_start_loc=q_acc_lens,
            query_start_loc_cpu=q_acc_lens_cpu,
            seq_lens=kv_lens,
            _seq_lens_cpu=kv_lens_cpu,
            _num_computed_tokens_cpu=num_computed_tokens_cpu,
            num_reqs=num_reqs,
            num_actual_tokens=total_q_len,
            max_query_len=max_q_len,
            max_seq_len=max_kv_len,
            block_table_tensor=blk_table_tensor,
            slot_mapping=slot_mapping_gid_0,
            causal=True,
            num_kv_actual_tokens=total_kv_len,
            seq_indexes_list=seq_indexes_list,
            cp_common_metadata=cp_common_metadata,
            scatter_indexes_tensor=scatter_indexes_tensor,
            gather_indexes_tensor=gather_indexes_tensor,
1743
            enable_lightly_cp=True
王敏's avatar
王敏 committed
1744
1745
1746
        )
        return cm_base

1747
    def _prepare_inputs(
1748
1749
1750
        self,
        scheduler_output: "SchedulerOutput",
        num_scheduled_tokens: np.ndarray,
1751
1752
    ) -> tuple[
        torch.Tensor,
1753
        SpecDecodeMetadata | None,
1754
    ]:
1755
1756
        """
        :return: tuple[
1757
            logits_indices, spec_decode_metadata,
1758
1759
        ]
        """
1760
1761
1762
1763
1764
1765
1766
        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.
1767
        self.input_batch.block_table.commit_block_table(num_reqs)
1768
1769
1770

        # Get request indices.
        # E.g., [2, 5, 3] -> [0, 0, 1, 1, 1, 1, 1, 2, 2, 2]
1771
        req_indices = np.repeat(self.arange_np[:num_reqs], num_scheduled_tokens)
1772

1773
1774
        # cu_num_tokens: [2, 5, 3] -> [2, 7, 10]
        # arange: [0, 1, 0, 1, 2, 3, 4, 0, 1, 2]
1775
        cu_num_tokens, arange = self._get_cumsum_and_arange(num_scheduled_tokens)
1776
1777

        # Get positions.
1778
        positions_np = self.positions.np[:total_num_scheduled_tokens]
1779
1780
1781
1782
1783
        np.add(
            self.input_batch.num_computed_tokens_cpu[req_indices],
            arange,
            out=positions_np,
        )
1784

1785
1786
        # Calculate M-RoPE positions.
        # Only relevant for models using M-RoPE (e.g, Qwen2-VL)
1787
        if self.uses_mrope:
1788
1789
            self._calc_mrope_positions(scheduler_output)

1790
1791
1792
1793
1794
        # 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)

1795
1796
1797
1798
        # 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.
1799
1800
1801
        token_indices = (
            positions_np + req_indices * self.input_batch.token_ids_cpu.shape[1]
        )
1802
        token_indices_tensor = torch.from_numpy(token_indices)
1803

1804
1805
1806
        # 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.
1807
1808
1809
1810
1811
1812
        torch.index_select(
            self.input_batch.token_ids_cpu_tensor.flatten(),
            0,
            token_indices_tensor,
            out=self.input_ids.cpu[:total_num_scheduled_tokens],
        )
1813
        if self.enable_prompt_embeds:
1814
            is_token_ids = self.input_batch.is_token_ids_tensor.flatten()
1815
1816
1817
1818
            torch.index_select(
                is_token_ids,
                0,
                token_indices_tensor,
1819
1820
                out=self.is_token_ids.cpu[:total_num_scheduled_tokens],
            )
1821
1822
1823
1824
1825
1826
1827
1828
1829
1830
1831
1832
1833
1834
1835
1836
1837
1838
1839
1840
1841
1842
1843
1844
1845
1846
1847
1848
1849
1850
1851
1852
1853

        # 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:
1854
1855
1856
                    self.inputs_embeds.cpu[
                        output_idx : output_idx + actual_num_sched
                    ].copy_(req_embeds[start_pos:actual_end])
1857
1858

                output_idx += num_sched
1859

1860
1861
        self.input_batch.block_table.compute_slot_mapping(req_indices, positions_np)
        self.input_batch.block_table.commit_slot_mapping(total_num_scheduled_tokens)
1862
1863

        # Prepare the attention metadata.
1864
        self.query_start_loc.np[0] = 0
1865
        self.query_start_loc.np[1 : num_reqs + 1] = cu_num_tokens
1866
1867
        # Note: pad query_start_loc to be non-decreasing, as kernels
        # like FlashAttention requires that
1868
        self.query_start_loc.np[num_reqs + 1 :].fill(cu_num_tokens[-1])
1869
        self.query_start_loc.copy_to_gpu()
1870
        query_start_loc = self.query_start_loc.gpu[: num_reqs + 1]
1871

1872
        self.seq_lens.np[:num_reqs] = (
1873
1874
            self.input_batch.num_computed_tokens_cpu[:num_reqs] + num_scheduled_tokens
        )
1875
        # Fill unused with 0 for full cuda graph mode.
1876
1877
        self.seq_lens.np[num_reqs:].fill(0)
        self.seq_lens.copy_to_gpu()
1878

1879
        num_tokens = [self.requests[r].num_tokens for r in self.input_batch.req_ids]
1880
1881
        num_tokens_np = np.array(num_tokens, dtype=np.int32)

1882
        # Record which requests should not be sampled,
1883
        # so that we could clear the sampled tokens before returning
1884
1885
        self.discard_request_mask.np[:num_reqs] = (
            self.seq_lens.np[:num_reqs] < num_tokens_np
1886
        )
1887
        self.discard_request_mask.copy_to_gpu(num_reqs)
1888

1889
        # Copy the tensors to the GPU.
1890
1891
1892
1893
1894
        self._prepare_input_ids(
            scheduler_output,
            total_num_scheduled_tokens,
            cu_num_tokens,
        )
1895

1896
        if self.uses_mrope:
1897
            # Only relevant for models using M-RoPE (e.g, Qwen2-VL)
guanyu1's avatar
guanyu1 committed
1898
            self._copy_mrope_positions_to_gpu(total_num_scheduled_tokens)
1899
1900
        elif self.uses_xdrope_dim > 0:
            # Only relevant for models using XD-RoPE (e.g, HunYuan-VL)
guanyu1's avatar
guanyu1 committed
1901
1902
            self._copy_xdrope_positions_to_gpu(total_num_scheduled_tokens)

1903
1904
        else:
            # Common case (1D positions)
1905
            self.positions.copy_to_gpu(total_num_scheduled_tokens)
1906

1907
        use_spec_decode = len(scheduler_output.scheduled_spec_decode_tokens) > 0
1908
1909
1910
1911
1912
1913
1914
1915
        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
1916
            num_sampled_tokens = np.ones(num_reqs, dtype=np.int32)
1917
1918
1919
1920
1921
        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)
1922
1923
1924
            # 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)
1925
1926
1927
1928
            for (
                req_id,
                draft_token_ids,
            ) in scheduler_output.scheduled_spec_decode_tokens.items():
1929
1930
                req_idx = self.input_batch.req_id_to_index[req_id]
                num_draft_tokens[req_idx] = len(draft_token_ids)
1931
1932
1933
1934
1935
                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
1936
1937
1938
1939
1940

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

1941
            spec_decode_metadata = self._calc_spec_decode_metadata(
王敏's avatar
王敏 committed
1942
                num_draft_tokens, cu_num_tokens, spec_decode_ids
1943
            )
1944
            logits_indices = spec_decode_metadata.logits_indices
1945
            num_sampled_tokens = num_draft_tokens + 1
1946
            # For DECODE only cuda graph of some attention backends (e.g., GDN).
1947
            self.num_decode_draft_tokens.np[:num_reqs] = num_decode_draft_tokens
1948
1949
            self.num_decode_draft_tokens.np[num_reqs:].fill(-1)
            self.num_decode_draft_tokens.copy_to_gpu()
1950

1951
1952
1953
1954
1955
        # Hot-Swap lora model
        if self.lora_config:
            assert (
                np.sum(num_sampled_tokens)
                <= self.vllm_config.scheduler_config.max_num_batched_tokens
1956
            )
1957
1958
1959
1960
1961
1962
1963
1964
1965
1966
1967
            self.set_active_loras(
                self.input_batch, num_scheduled_tokens, num_sampled_tokens
            )

        return (
            logits_indices,
            spec_decode_metadata,
        )

    def _build_attention_metadata(
        self,
1968
        num_tokens: int,
1969
        num_reqs: int,
1970
1971
1972
        max_query_len: int,
        num_tokens_padded: int | None = None,
        num_reqs_padded: int | None = None,
1973
1974
1975
1976
        ubatch_slices: UBatchSlices | None = None,
        logits_indices: torch.Tensor | None = None,
        use_spec_decode: bool = False,
        for_cudagraph_capture: bool = False,
1977
        num_scheduled_tokens: dict[str, int] | None = None,
1978
        cascade_attn_prefix_lens: list[list[int]] | None = None,
1979
        slot_mappings: dict[int, torch.Tensor] | None = None,
王敏's avatar
王敏 committed
1980
1981
1982
1983
1984
1985
    ) -> tuple[
        PerLayerAttnMetadata,
        CommonAttentionMetadata | None,
        torch.Tensor | None,
        torch.Tensor | None,
    ]:
1986
1987
1988
        """
        :return: tuple[attn_metadata, spec_decode_common_attn_metadata]
        """
1989
1990
        # Attention metadata is not needed for attention free models
        if len(self.kv_cache_config.kv_cache_groups) == 0:
王敏's avatar
王敏 committed
1991
1992
1993
            return {}, None, None, None

        tp_size = self.vllm_config.parallel_config.tensor_parallel_size
1994

1995
1996
        num_tokens_padded = num_tokens_padded or num_tokens
        num_reqs_padded = num_reqs_padded or num_reqs
1997
        assert num_reqs_padded is not None and num_tokens_padded is not None
1998

1999
2000
2001
        attn_metadata: PerLayerAttnMetadata = {}
        if ubatch_slices is not None:
            attn_metadata = [dict() for _ in range(len(ubatch_slices))]
2002

2003
2004
2005
2006
2007
2008
2009
2010
        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()

2011
2012
        if use_spec_decode:
            self.num_accepted_tokens.np[:num_reqs] = (
2013
2014
                self.input_batch.num_accepted_tokens_cpu[:num_reqs]
            )
2015
2016
            self.num_accepted_tokens.np[num_reqs:].fill(1)
            self.num_accepted_tokens.copy_to_gpu()
2017

2018
        kv_cache_groups = self.kv_cache_config.kv_cache_groups
2019

2020
        def _get_block_table(kv_cache_gid: int):
2021
2022
2023
            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):
2024
                blk_table_tensor = torch.zeros(
2025
                    (num_reqs_padded, 1),
2026
                    dtype=torch.int32,
2027
2028
                    device=self.device,
                )
2029
            else:
2030
                blk_table = self.input_batch.block_table[kv_cache_gid]
2031
                blk_table_tensor = blk_table.get_device_tensor(num_reqs_padded)
2032

2033
2034
2035
            # 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)
2036
            return blk_table_tensor
2037

2038
2039
2040
        assert slot_mappings is not None
        block_table_gid_0 = _get_block_table(0)
        slot_mapping_gid_0 = slot_mappings[0]
王敏's avatar
王敏 committed
2041
2042
        scatter_indexes_tensor = None
        gather_indexes_tensor = None
2043

2044
2045
        if self.model_config.enable_return_routed_experts:
            self.slot_mapping = slot_mapping_gid_0[:num_tokens].cpu().numpy()
王敏's avatar
王敏 committed
2046

2047
2048
        enable_lightly_cp = self.enable_lightly_cp and num_tokens > self.lightly_cp_threshould
        if not enable_lightly_cp:
王敏's avatar
王敏 committed
2049
2050
2051
2052
2053
2054
2055
2056
2057
2058
2059
2060
2061
2062
2063
2064
2065
2066
2067
2068
2069
            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,
                num_kv_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,
            )
        else:
            cm_base = self._prepare_cp_metadata(
                num_reqs_padded,
                max_query_len,
王敏's avatar
王敏 committed
2070
                max_seq_len,
王敏's avatar
王敏 committed
2071
2072
2073
                num_tokens,
                block_table_gid_0,
                slot_mapping_gid_0,
王敏's avatar
王敏 committed
2074
2075
2076
                self.input_batch.num_computed_tokens_cpu_tensor[
                    :num_reqs_padded
                ],
王敏's avatar
王敏 committed
2077
2078
2079
            )
            scatter_indexes_tensor = cm_base.scatter_indexes_tensor
            gather_indexes_tensor = cm_base.gather_indexes_tensor
2080
2081
2082
2083
2084
2085
2086
2087
2088
2089
2090
2091
2092
2093
2094
2095
2096
2097
2098
2099
2100
2101

        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
            )

2102
2103
2104
2105
2106
2107
2108
2109
2110
        # 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
        ] = {}

2111
2112
2113
2114
2115
2116
2117
        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]
2118
            builder = attn_group.get_metadata_builder(ubid or 0)
2119
2120
2121
2122
            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))
2123

2124
2125
2126
2127
            cascade_attn_prefix_len = (
                cascade_attn_prefix_lens[kv_cache_gid][attn_gid]
                if cascade_attn_prefix_lens
                else 0
2128
2129
            )

2130
2131
2132
2133
2134
2135
2136
2137
2138
2139
2140
2141
2142
2143
            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
                )
2144
2145
2146
2147
2148
2149
2150
2151
2152
            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,
                )
2153
2154
2155
2156
2157
2158
            else:
                attn_metadata_i = builder.build(
                    common_prefix_len=cascade_attn_prefix_len,
                    common_attn_metadata=common_attn_metadata,
                    **extra_attn_metadata_args,
                )
2159
2160
                if builder.supports_update_block_table:
                    cached_attn_metadata[cache_key] = attn_metadata_i
2161
2162
2163
2164
2165
2166
2167
2168
2169
2170
2171
2172
2173
2174
2175
2176
2177
2178
2179
2180
2181
2182
2183

            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,
2184
                for_cudagraph_capture=for_cudagraph_capture,
2185
            )
2186
            if kv_cache_gid > 0:
2187
2188
                cm.block_table_tensor = _get_block_table(kv_cache_gid)
                cm.slot_mapping = slot_mappings[kv_cache_gid]
2189

2190
                if enable_lightly_cp and cm.seq_indexes_list is not None:
王敏's avatar
王敏 committed
2191
2192
                    cm.block_table_tensor = cm.block_table_tensor[cm.seq_indexes_list]

王敏's avatar
王敏 committed
2193
            if self.speculative_config and spec_decode_common_attn_metadata is None and hasattr(self, "drafter"):
2194
                if isinstance(self.drafter, EagleProposer):
2195
                    if self.drafter.attn_layer_names[0] in kv_cache_group.layer_names:
2196
                        if enable_lightly_cp:
王敏's avatar
王敏 committed
2197
2198
2199
2200
                            spec_decode_common_attn_metadata = cm.cp_common_metadata
                        else:
                            spec_decode_common_attn_metadata = cm
                        #spec_decode_common_attn_metadata = cm
2201
                else:
2202
                    if enable_lightly_cp:
王敏's avatar
王敏 committed
2203
2204
2205
2206
                        spec_decode_common_attn_metadata = cm.cp_common_metadata
                    else:
                        spec_decode_common_attn_metadata = cm
                    #spec_decode_common_attn_metadata = cm
2207

2208
            for attn_gid in range(len(self.attn_groups[kv_cache_gid])):
2209
                if ubatch_slices is not None:
2210
2211
2212
                    for ubid, _cm in enumerate(split_attn_metadata(ubatch_slices, cm)):
                        _build_attn_group_metadata(kv_cache_gid, attn_gid, _cm, ubid)

2213
                else:
2214
                    _build_attn_group_metadata(kv_cache_gid, attn_gid, cm)
2215

2216
2217
2218
2219
2220
2221
2222
2223
2224
2225
2226
2227
2228
2229
2230
2231
2232
2233
2234
        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]
2235

王敏's avatar
王敏 committed
2236
        if (
2237
            (not self.enable_lightly_cp)
王敏's avatar
王敏 committed
2238
2239
            and spec_decode_common_attn_metadata is not None
            and (num_reqs != num_reqs_padded or num_tokens != num_tokens_padded)
2240
2241
2242
2243
2244
2245
2246
2247
        ):
            # 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)
            )

王敏's avatar
王敏 committed
2248
2249
2250
2251
2252
2253
        return (
            attn_metadata, 
            spec_decode_common_attn_metadata,
            scatter_indexes_tensor,
            gather_indexes_tensor
        )
2254

2255
2256
2257
    def _compute_cascade_attn_prefix_lens(
        self,
        num_scheduled_tokens: np.ndarray,
2258
        num_computed_tokens: np.ndarray,
2259
2260
2261
2262
2263
2264
2265
        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
        """
2266

2267
2268
2269
2270
2271
2272
2273
2274
2275
2276
2277
2278
2279
2280
        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,
2281
                        num_computed_tokens,
2282
2283
2284
2285
2286
2287
                        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
2288

2289
        return cascade_attn_prefix_lens if use_cascade_attn else None
2290

2291
2292
2293
    def _compute_cascade_attn_prefix_len(
        self,
        num_scheduled_tokens: np.ndarray,
2294
        num_computed_tokens: np.ndarray,
2295
        num_common_prefix_blocks: int,
2296
2297
        kv_cache_spec: KVCacheSpec,
        attn_metadata_builder: AttentionMetadataBuilder,
2298
2299
2300
2301
2302
2303
2304
2305
2306
2307
2308
2309
2310
2311
2312
2313
2314
2315
    ) -> 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.
        """
2316

2317
        common_prefix_len = num_common_prefix_blocks * kv_cache_spec.block_size
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
        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]
2355
        # Request 3's num_computed_tokens: 3 (i.e., [A, B, C])
2356
2357
2358
2359
2360
        # 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.
2361
        common_prefix_len = min(common_prefix_len, num_computed_tokens.min())
2362
        # common_prefix_len should be a multiple of the block size.
2363
2364
2365
2366
2367
2368
2369
2370
2371
2372
2373
        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
        )
2374
2375
        assert isinstance(kv_cache_spec, AttentionSpec)
        use_cascade = attn_metadata_builder.use_cascade_attention(
2376
2377
2378
            common_prefix_len=common_prefix_len,
            query_lens=num_scheduled_tokens,
            num_query_heads=self.num_query_heads,
2379
            num_kv_heads=kv_cache_spec.num_kv_heads,
2380
            use_alibi=self.use_alibi,
2381
            use_sliding_window=use_sliding_window,
2382
            use_local_attention=use_local_attention,
2383
            num_sms=self.num_sms,
2384
            dcp_world_size=self.dcp_world_size,
2385
2386
2387
        )
        return common_prefix_len if use_cascade else 0

guanyu1's avatar
guanyu1 committed
2388
2389
    def _calc_xdrope_positions(self, scheduler_output: "SchedulerOutput"):
        xdrope_pos_ptr = 0
2390
        for index, req_id in enumerate(self.input_batch.req_ids):
2391
            req = self.requests[req_id]
guanyu1's avatar
guanyu1 committed
2392
            assert req.xdrope_positions is not None
2393

2394
2395
            num_computed_tokens = self.input_batch.num_computed_tokens_cpu[index]
            num_scheduled_tokens = scheduler_output.num_scheduled_tokens[req_id]
2396
            num_prompt_tokens = length_from_prompt_token_ids_or_embeds(
2397
2398
                req.prompt_token_ids, req.prompt_embeds
            )
2399
2400

            if num_computed_tokens + num_scheduled_tokens > num_prompt_tokens:
2401
2402
                prompt_part_len = max(0, num_prompt_tokens - num_computed_tokens)
                completion_part_len = max(0, num_scheduled_tokens - prompt_part_len)
2403
2404
2405
2406
2407
2408
2409
            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:
guanyu1's avatar
guanyu1 committed
2410
2411
2412
                # prompt's xdrope_positions are pre-computed
                dst_start = xdrope_pos_ptr
                dst_end = xdrope_pos_ptr + prompt_part_len
2413
2414
2415
                src_start = num_computed_tokens
                src_end = num_computed_tokens + prompt_part_len

guanyu1's avatar
guanyu1 committed
2416
                self.xdrope_positions.cpu[:, dst_start:dst_end] = req.xdrope_positions[
2417
2418
                    :, src_start:src_end
                ]
guanyu1's avatar
guanyu1 committed
2419
                xdrope_pos_ptr += prompt_part_len
2420
2421

            if completion_part_len > 0:
guanyu1's avatar
guanyu1 committed
2422
2423
2424
                # compute completion's xdrope_positions on-the-fly
                dst_start = xdrope_pos_ptr
                dst_end = xdrope_pos_ptr + completion_part_len
2425

guanyu1's avatar
guanyu1 committed
2426
2427
                XDRotaryEmbedding.get_next_input_positions_tensor(
                    out=self.xdrope_positions.np,
2428
2429
2430
2431
                    out_offset=dst_start,
                    context_len=num_computed_tokens + prompt_part_len,
                    num_new_tokens=completion_part_len,
                )
2432

guanyu1's avatar
guanyu1 committed
2433
2434
                xdrope_pos_ptr += completion_part_len
                
guanyu1's avatar
guanyu1 committed
2435
2436
    def _calc_mrope_positions(self, scheduler_output: "SchedulerOutput"):
        mrope_pos_ptr = 0
guanyu1's avatar
guanyu1 committed
2437
2438
2439
2440
2441
2442
2443
        if self.use_1d_mrope:
            mrope_positions_token_major = self.mrope_positions.cpu.view(
                self.max_num_tokens + 1, 3
            )
            mrope_positions_token_major_np = self.mrope_positions.np.reshape(
                self.max_num_tokens + 1, 3
            )
2444
2445
        for index, req_id in enumerate(self.input_batch.req_ids):
            req = self.requests[req_id]
guanyu1's avatar
guanyu1 committed
2446
            assert req.mrope_positions is not None
2447

2448
2449
2450
2451
2452
            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
            )
2453

2454
2455
2456
2457
2458
2459
2460
2461
2462
2463
            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:
guanyu1's avatar
guanyu1 committed
2464
2465
2466
                # prompt's mrope_positions are pre-computed
                dst_start = mrope_pos_ptr
                dst_end = mrope_pos_ptr + prompt_part_len
2467
2468
2469
                src_start = num_computed_tokens
                src_end = num_computed_tokens + prompt_part_len

guanyu1's avatar
guanyu1 committed
2470
2471
2472
2473
2474
2475
2476
2477
                if self.use_1d_mrope:
                    mrope_positions_token_major[dst_start:dst_end, :].copy_(
                        req.mrope_positions[:, src_start:src_end].transpose(0, 1)
                    )
                else:
                    self.mrope_positions.cpu[:, dst_start:dst_end] = (
                        req.mrope_positions[:, src_start:src_end]
                    )
guanyu1's avatar
guanyu1 committed
2478
                mrope_pos_ptr += prompt_part_len
2479
2480

            if completion_part_len > 0:
guanyu1's avatar
guanyu1 committed
2481
2482
2483
                # compute completion's mrope_positions on-the-fly
                dst_start = mrope_pos_ptr
                dst_end = mrope_pos_ptr + completion_part_len
2484

guanyu1's avatar
guanyu1 committed
2485
                assert req.mrope_position_delta is not None
guanyu1's avatar
guanyu1 committed
2486
2487
2488
2489
2490
2491
2492
2493
2494
2495
2496
2497
2498
2499
2500
2501
2502
2503
2504
2505
2506
2507
                if self.use_1d_mrope:
                    values = np.arange(
                        req.mrope_position_delta
                        + num_computed_tokens
                        + prompt_part_len,
                        req.mrope_position_delta
                        + num_computed_tokens
                        + prompt_part_len
                        + completion_part_len,
                        dtype=mrope_positions_token_major_np.dtype,
                    )
                    mrope_positions_token_major_np[dst_start:dst_end, :] = values[
                        :, None
                    ]
                else:
                    MRotaryEmbedding.get_next_input_positions_tensor(
                        out=self.mrope_positions.np,
                        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,
                    )
2508

guanyu1's avatar
guanyu1 committed
2509
                mrope_pos_ptr += completion_part_len
2510

2511
2512
    def _calc_spec_decode_metadata(
        self,
2513
2514
        num_draft_tokens: np.ndarray,
        cu_num_scheduled_tokens: np.ndarray,
王敏's avatar
王敏 committed
2515
        spec_decode_ids: Optional[list[str]] = None
2516
2517
2518
2519
2520
2521
2522
2523
2524
2525
2526
2527
2528
2529
    ) -> 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
2530
2531
2532
2533

        # 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(
2534
2535
            num_sampled_tokens, cumsum_dtype=np.int32
        )
2536
        # Step 2. [0, 0, 0, 0, 103, 104, 104, 104, 206, 207, 207]
2537
        logits_indices = np.repeat(
2538
2539
            cu_num_scheduled_tokens - num_sampled_tokens, num_sampled_tokens
        )
2540
        # Step 3. [0, 1, 2, 3, 103, 104, 105, 106, 206, 207, 208]
2541
2542
2543
2544
2545
2546
        logits_indices += arange

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

        # Compute the draft logits indices.
2547
2548
2549
        # 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(
2550
2551
            num_draft_tokens, cumsum_dtype=np.int32
        )
2552
2553
        # [0, 0, 0, 5, 5, 9]
        target_logits_indices = np.repeat(
2554
2555
            cu_num_sampled_tokens - num_sampled_tokens, num_draft_tokens
        )
2556
2557
        # [0, 1, 2, 5, 6, 9]
        target_logits_indices += arange
2558
        draft_token_indices = target_logits_indices + 1
2559

2560
        # TODO: Optimize the CPU -> GPU copy.
2561
2562
2563
2564
2565
2566
2567
2568
2569
2570
2571
2572
2573
2574
2575
2576
2577
2578
2579
2580
2581
2582
2583
2584
2585
2586
2587
2588
2589
2590
2591
2592
2593
2594
2595
2596
2597
2598
2599
2600
2601
        # cu_num_draft_tokens = torch.from_numpy(cu_num_draft_tokens).to(
        #     self.device, non_blocking=True
        # )
        # cu_num_sampled_tokens = torch.from_numpy(cu_num_sampled_tokens).to(
        #     self.device, non_blocking=True
        # )
        # logits_indices = torch.from_numpy(logits_indices).to(
        #     self.device, non_blocking=True
        # )
        # target_logits_indices = torch.from_numpy(target_logits_indices).to(
        #     self.device, non_blocking=True
        # )
        # bonus_logits_indices = torch.from_numpy(bonus_logits_indices).to(
        #     self.device, non_blocking=True
        # )

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

        # Optimize the H2D in the process of creating spec decode metadata
        fused_meta_data = cu_num_draft_tokens.tolist() + cu_num_sampled_tokens.tolist()\
              + logits_indices.tolist() + target_logits_indices.tolist() + bonus_logits_indices.tolist()\
              + draft_token_indices.tolist()
        
        fused_meta_data_len = np.array([len(cu_num_draft_tokens), len(cu_num_sampled_tokens),\
                                        len(logits_indices), len(target_logits_indices),\
                                            len(bonus_logits_indices), len(draft_token_indices)], dtype=np.int32)
        cu_fused_meta_data_len = np.cumsum(fused_meta_data_len, dtype=np.int32)
        fused_meta_data = torch.tensor(
            fused_meta_data, dtype=torch.int32, pin_memory=self.pin_memory
        ).to(self.device, non_blocking=True)
        
        cu_num_draft_tokens = fused_meta_data[:cu_fused_meta_data_len[0]]
        cu_num_sampled_tokens = fused_meta_data[cu_fused_meta_data_len[0]:cu_fused_meta_data_len[1]]
        logits_indices = fused_meta_data[cu_fused_meta_data_len[1]:cu_fused_meta_data_len[2]]
        target_logits_indices = fused_meta_data[cu_fused_meta_data_len[2]:cu_fused_meta_data_len[3]]
        bonus_logits_indices = fused_meta_data[cu_fused_meta_data_len[3]:cu_fused_meta_data_len[4]]
        draft_token_indices = fused_meta_data[cu_fused_meta_data_len[4]:cu_fused_meta_data_len[5]]

2602

2603
2604
        # Compute the draft token ids.
        # draft_token_indices:      [  1,   2,   3, 105, 106, 208]
2605
        draft_token_ids = self.input_ids.gpu[logits_indices]
2606
        draft_token_ids = draft_token_ids[draft_token_indices]
2607

2608
        return SpecDecodeMetadata(
2609
2610
2611
            draft_token_ids=draft_token_ids,
            num_draft_tokens=num_draft_tokens.tolist(),
            cu_num_draft_tokens=cu_num_draft_tokens,
2612
            cu_num_sampled_tokens=cu_num_sampled_tokens,
2613
2614
2615
            target_logits_indices=target_logits_indices,
            bonus_logits_indices=bonus_logits_indices,
            logits_indices=logits_indices,
王敏's avatar
王敏 committed
2616
            spec_decode_ids=spec_decode_ids,
2617
2618
        )

2619
2620
2621
2622
2623
2624
2625
    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
2626
        self.kv_sharing_fast_prefill_logits_indices[:num_logits].copy_(logits_indices)
2627
2628
2629
2630
2631
        # 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_(
2632
2633
            logits_indices[-1].item()
        )
2634
2635
2636
2637
2638
        # 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
2639
2640
2641
        logits_indices_padded = self.kv_sharing_fast_prefill_logits_indices[
            :num_logits_padded
        ]
2642
2643
        return logits_indices_padded

2644
    def _batch_mm_inputs_from_scheduler(
2645
2646
        self,
        scheduler_output: "SchedulerOutput",
2647
2648
2649
2650
2651
    ) -> tuple[
        list[str],
        list[MultiModalKwargsItem],
        list[tuple[str, PlaceholderRange]],
    ]:
2652
        """Batch multimodal inputs from scheduled encoder inputs.
2653
2654
2655

        Args:
            scheduler_output: The scheduler output containing scheduled encoder
2656
                inputs.
2657
2658

        Returns:
2659
            A tuple of (mm_hashes, mm_kwargs, mm_lora_refs) where:
2660
2661
            - mm_hashes: List of multimodal hashes for each item
            - mm_kwargs: List of multimodal kwargs for each item
2662
            - mm_lora_refs: List of (req_id, placeholder_range) for each item
2663
        """
2664
2665
        scheduled_encoder_inputs = scheduler_output.scheduled_encoder_inputs
        if not scheduled_encoder_inputs:
2666
            return [], [], []
2667
2668

        mm_hashes = list[str]()
2669
        mm_kwargs = list[MultiModalKwargsItem]()
2670
2671
2672
        # Multimodal LoRA reference info to map each multimodal item
        # back to its request & position
        mm_lora_refs = list[tuple[str, PlaceholderRange]]()
2673
2674
        for req_id, encoder_input_ids in scheduled_encoder_inputs.items():
            req_state = self.requests[req_id]
2675
2676

            for mm_input_id in encoder_input_ids:
2677
                mm_feature = req_state.mm_features[mm_input_id]
2678
2679
                if mm_feature.data is None:
                    continue
2680
2681

                mm_hashes.append(mm_feature.identifier)
2682
                mm_kwargs.append(mm_feature.data)
2683
                mm_lora_refs.append((req_id, mm_feature.mm_position))
2684

2685
        return mm_hashes, mm_kwargs, mm_lora_refs
2686

2687
2688
2689
    def _execute_mm_encoder(
        self, scheduler_output: "SchedulerOutput"
    ) -> list[torch.Tensor]:
2690
        mm_hashes, mm_kwargs, mm_lora_refs = self._batch_mm_inputs_from_scheduler(
2691
2692
            scheduler_output
        )
2693
2694

        if not mm_kwargs:
2695
            return []
2696

2697
2698
2699
2700
2701
2702
        should_time = bool(
            self.observability_config
            and self.observability_config.enable_mm_processor_stats
            and scheduler_output.scheduled_encoder_inputs
        )

2703
2704
2705
2706
2707
2708
2709
        # 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.
2710
        model = cast(SupportsMultiModal, self.model)
2711
2712
2713
2714
2715
2716
2717
2718
2719
2720
2721
2722
2723
2724
2725
2726
2727
2728
2729
2730
2731
2732
2733
2734
2735
2736
2737
2738
2739
2740
2741
2742
2743
2744
2745
2746
2747
2748
2749
2750
2751
2752
2753
2754
2755
2756
2757
2758
2759
2760
2761
2762
2763
2764
2765
2766
2767

        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,
                )

2768
        encoder_outputs: list[torch.Tensor] = []
2769
2770
        # Track the current index in mm_kwargs/mm_lora_refs to map groups to request IDs
        current_item_idx = 0
2771
        for modality, num_items, mm_kwargs_group in group_mm_kwargs_by_modality(
2772
2773
2774
            mm_kwargs,
            device=self.device,
            pin_memory=self.pin_memory,
2775
        ):
2776
            curr_group_outputs: MultiModalEmbeddings
2777
2778

            # EVS-related change.
2779
            # (ekhvedchenia): Temporary hack to limit peak memory usage when
2780
            # processing multimodal data. This solves the issue with scheduler
2781
2782
2783
2784
            # 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)
2785
2786
2787
2788
2789
2790
2791
            # 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
            ):
2792
                curr_group_outputs_lst = list[torch.Tensor]()
2793
2794
2795
2796
2797
2798
2799
2800
2801
2802
2803
                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,
                            )
2804
                        )
2805

2806
2807
2808
                        micro_batch_outputs = model.embed_multimodal(
                            **micro_batch_mm_inputs
                        )
2809

2810
                        curr_group_outputs_lst.extend(micro_batch_outputs)
2811
2812

                curr_group_outputs = curr_group_outputs_lst
2813
2814
2815
2816
2817
2818
2819
2820
            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.
2821
2822
2823
2824
2825

                with self.timed_encoder_operation(
                    should_time, mm_lora_refs, current_item_idx, num_items
                ):
                    curr_group_outputs = model.embed_multimodal(**mm_kwargs_group)
2826

2827
2828
            sanity_check_mm_encoder_outputs(
                curr_group_outputs,
2829
                expected_num_items=num_items,
2830
            )
2831
            encoder_outputs.extend(curr_group_outputs)
2832

2833
2834
            current_item_idx += num_items

2835
        # Cache the encoder outputs by mm_hash
2836
        for mm_hash, output in zip(mm_hashes, encoder_outputs):
2837
            self.encoder_cache[mm_hash] = output
2838
2839
            logger.debug("Finish execute for mm hash %s", mm_hash)
            self.maybe_save_ec_to_connector(self.encoder_cache, mm_hash)
2840

2841
        return encoder_outputs
2842
2843

    def _gather_mm_embeddings(
2844
2845
        self,
        scheduler_output: "SchedulerOutput",
2846
        shift_computed_tokens: int = 0,
2847
2848
2849
    ) -> tuple[list[torch.Tensor], torch.Tensor]:
        total_num_scheduled_tokens = scheduler_output.total_num_scheduled_tokens

2850
2851
2852
2853
2854
        # 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]

2855
        mm_embeds = list[torch.Tensor]()
2856
        is_mm_embed = is_mm_embed_buf.cpu
2857
2858
2859
        is_mm_embed[:total_num_scheduled_tokens] = False

        req_start_idx = 0
2860
        should_sync_mrope_positions = False
2861
        should_sync_xdrope_positions = False
2862

2863
        for req_id in self.input_batch.req_ids:
2864
2865
            mm_embeds_req: list[torch.Tensor] = []

2866
            num_scheduled_tokens = scheduler_output.num_scheduled_tokens[req_id]
2867
            req_state = self.requests[req_id]
2868
            num_computed_tokens = req_state.num_computed_tokens + shift_computed_tokens
2869

2870
2871
            for mm_feature in req_state.mm_features:
                pos_info = mm_feature.mm_position
2872
2873
                start_pos = pos_info.offset
                num_encoder_tokens = pos_info.length
2874
2875
2876
2877
2878
2879
2880
2881
2882
2883
2884
2885
2886
2887
2888
2889

                # 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,
2890
2891
                    num_encoder_tokens,
                )
2892
                assert start_idx < end_idx
2893
2894
2895
2896
2897
2898
2899
                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
2900

2901
                mm_hash = mm_feature.identifier
2902
                encoder_output = self.encoder_cache.get(mm_hash, None)
2903
                assert encoder_output is not None, f"Encoder cache miss for {mm_hash}."
2904
2905
2906

                if (is_embed := pos_info.is_embed) is not None:
                    is_embed = is_embed[start_idx:end_idx]
2907
2908
2909
                    mm_embeds_item = encoder_output[curr_embeds_start:curr_embeds_end]
                else:
                    mm_embeds_item = encoder_output[start_idx:end_idx]
2910

2911
                req_start_pos = req_start_idx + start_pos - num_computed_tokens
2912
2913
2914
2915
2916
2917
2918
2919
2920
                # 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
2921
2922
2923
                mm_embeds_req.append(mm_embeds_item)

            if self.is_multimodal_pruning_enabled and self.uses_mrope:
2924
                assert req_state.mrope_positions is not None
2925
2926
2927
2928
2929
2930
2931
                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,
2932
2933
                    )
                )
2934
2935
2936
2937
                req_state.mrope_positions.copy_(new_mrope_positions)
                req_state.mrope_position_delta = new_delta

            mm_embeds.extend(mm_embeds_req)
2938
2939
            req_start_idx += num_scheduled_tokens

2940
        is_mm_embed = is_mm_embed_buf.copy_to_gpu(total_num_scheduled_tokens)
2941
2942
2943

        if should_sync_mrope_positions:
            self._calc_mrope_positions(scheduler_output)
guanyu1's avatar
guanyu1 committed
2944
            self._copy_mrope_positions_to_gpu(total_num_scheduled_tokens)
2945

2946
2947
        if should_sync_xdrope_positions:
            self._calc_xdrope_positions(scheduler_output)
guanyu1's avatar
guanyu1 committed
2948
            self._copy_xdrope_positions_to_gpu(total_num_scheduled_tokens)
2949

2950
        return mm_embeds, is_mm_embed
2951

2952
    def get_model(self) -> nn.Module:
2953
        # get raw model out of the cudagraph wrapper.
2954
        if isinstance(self.model, (CUDAGraphWrapper, UBatchWrapper)):
2955
            return self.model.unwrap()
2956
2957
        return self.model

2958
2959
2960
2961
2962
2963
2964
2965
2966
2967
2968
2969
2970
2971
2972
    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

2973
2974
2975
2976
2977
    def get_supported_pooling_tasks(self) -> list[PoolingTask]:
        model = self.get_model()
        if not is_pooling_model(model):
            return []

2978
2979
        supported_tasks = list(model.pooler.get_supported_tasks())

2980
2981
2982
2983
        if "score" in supported_tasks:
            num_labels = getattr(self.model_config.hf_config, "num_labels", 0)
            if num_labels != 1:
                supported_tasks.remove("score")
2984
                logger.debug_once("Score API is only enabled for num_labels == 1.")
2985
2986

        return supported_tasks
2987

2988
2989
2990
2991
2992
2993
2994
2995
2996
    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)
2997

2998
    def sync_and_slice_intermediate_tensors(
2999
3000
        self,
        num_tokens: int,
3001
        intermediate_tensors: IntermediateTensors | None,
3002
3003
        sync_self: bool,
    ) -> IntermediateTensors:
3004
3005
3006
        assert self.intermediate_tensors is not None

        tp = self.vllm_config.parallel_config.tensor_parallel_size
3007
        is_rs = is_residual_scattered_for_sp(self.vllm_config, num_tokens)
3008
3009
3010
3011
3012
3013

        # 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():
3014
                is_scattered = k == "residual" and is_rs
3015
                copy_len = num_tokens // tp if is_scattered else num_tokens
3016
                self.intermediate_tensors[k][:copy_len].copy_(
3017
3018
3019
3020
3021
3022
3023
3024
3025
3026
3027
3028
3029
                    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:
3030
3031
3032
3033
3034
3035
3036
        """
        Step for the EPLB (Expert Parallelism Load Balancing) state.
        """
        if not self.parallel_config.enable_eplb:
            return

        assert self.eplb_state is not None
3037
3038
        model = self.get_model()
        assert is_mixture_of_experts(model)
3039
3040
3041
        self.eplb_state.step(
            is_dummy,
            is_profile,
3042
            log_stats=self.parallel_config.eplb_config.log_balancedness,
3043
3044
        )

3045
3046
3047
3048
3049
    def _pool(
        self,
        hidden_states: torch.Tensor,
        num_scheduled_tokens: int,
        num_scheduled_tokens_np: np.ndarray,
3050
3051
3052
3053
        kv_connector_output: KVConnectorOutput | None,
    ) -> ModelRunnerOutput | AsyncModelRunnerOutput:
        num_reqs = self.input_batch.num_reqs
        assert num_reqs == len(self.input_batch.pooling_params), (
3054
3055
            "Either all or none of the requests in a batch must be pooling request"
        )
3056

3057
        hidden_states = hidden_states[:num_scheduled_tokens]
3058
        seq_lens_cpu = self.seq_lens.cpu[:num_reqs]
3059

3060
        pooling_metadata = self.input_batch.get_pooling_metadata()
3061
        pooling_metadata.build_pooling_cursor(
3062
            num_scheduled_tokens_np, seq_lens_cpu, device=hidden_states.device
3063
        )
3064

3065
3066
        model = cast(VllmModelForPooling, self.model)
        raw_pooler_output: PoolerOutput = model.pooler(
3067
            hidden_states=hidden_states, pooling_metadata=pooling_metadata
3068
        )
3069
3070
3071
3072
3073
3074
3075
3076
3077
3078
3079
3080
3081
3082
3083
3084
3085
3086
3087
3088
3089
3090
3091
3092

        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,
            )

3093
        raw_pooler_output = json_map_leaves(
3094
            lambda x: None if x is None else x.to("cpu", non_blocking=True),
3095
3096
            raw_pooler_output,
        )
3097
3098
3099
3100
        model_runner_output.pooler_output = [
            out if include else None
            for out, include in zip(raw_pooler_output, finished_mask)
        ]
3101
        self._sync_device()
3102

3103
        return model_runner_output
3104

王敏's avatar
王敏 committed
3105
3106
    def _pad_for_mla_cp(self, num_scheduled_tokens: int) -> int:
        tp_size = self.vllm_config.parallel_config.tensor_parallel_size
3107
3108
3109
3110
3111
        # if num_scheduled_tokens <= tp_size * tp_size:
        #     return num_scheduled_tokens
        # else:
        #     return round_up(num_scheduled_tokens, tp_size)
        return round_up(num_scheduled_tokens, tp_size)
王敏's avatar
王敏 committed
3112

3113
    def _pad_for_sequence_parallelism(self, num_scheduled_tokens: int) -> int:
3114
3115
        # Pad tokens to multiple of tensor_parallel_size when
        # enabled collective fusion for SP
王敏's avatar
王敏 committed
3116

3117
        if self.enable_lightly_cp and num_scheduled_tokens > self.lightly_cp_threshould:
王敏's avatar
王敏 committed
3118
            return self._pad_for_mla_cp(num_scheduled_tokens)
3119
        tp_size = self.vllm_config.parallel_config.tensor_parallel_size
3120
        if self.compilation_config.pass_config.enable_sp and tp_size > 1:
3121
3122
3123
            return round_up(num_scheduled_tokens, tp_size)
        return num_scheduled_tokens

Patrick von Platen's avatar
Patrick von Platen committed
3124
3125
3126
3127
3128
3129
3130
3131
3132
3133
3134
    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

3135
    def _preprocess(
3136
3137
        self,
        scheduler_output: "SchedulerOutput",
3138
        num_input_tokens: int,  # Padded
3139
        intermediate_tensors: IntermediateTensors | None = None,
3140
    ) -> tuple[
3141
3142
        torch.Tensor | None,
        torch.Tensor | None,
3143
        torch.Tensor,
3144
        IntermediateTensors | None,
3145
        dict[str, Any],
3146
        ECConnectorOutput | None,
3147
    ]:
3148
        num_scheduled_tokens = scheduler_output.total_num_scheduled_tokens
3149
        is_first_rank = get_pp_group().is_first_rank
3150
        is_encoder_decoder = self.model_config.is_encoder_decoder
3151

3152
3153
        # _prepare_inputs may reorder the batch, so we must gather multi
        # modal outputs after that to ensure the correct order
3154
3155
        ec_connector_output = None

3156
        if self.supports_mm_inputs and is_first_rank and not is_encoder_decoder:
3157
            # Run the multimodal encoder if any.
3158
3159
3160
3161
3162
3163
            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)
3164

3165
3166
3167
            # 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.
3168
            inputs_embeds_scheduled = self.model.embed_input_ids(
3169
3170
3171
                self.input_ids.gpu[:num_scheduled_tokens],
                multimodal_embeddings=mm_embeds,
                is_multimodal=is_mm_embed,
3172
            )
3173

3174
            # TODO(woosuk): Avoid the copy. Optimize.
3175
            self.inputs_embeds.gpu[:num_scheduled_tokens].copy_(inputs_embeds_scheduled)
3176

Patrick von Platen's avatar
Patrick von Platen committed
3177
            input_ids, inputs_embeds = self._prepare_mm_inputs(num_input_tokens)
3178
            model_kwargs = {
3179
                **self._init_model_kwargs(),
3180
3181
                **self._extract_mm_kwargs(scheduler_output),
            }
3182
        elif self.enable_prompt_embeds and is_first_rank:
3183
3184
3185
3186
3187
3188
3189
3190
3191
3192
3193
3194
            # 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).
3195
3196
3197
            token_ids_idx = (
                self.is_token_ids.gpu[:num_scheduled_tokens]
                .nonzero(as_tuple=False)
3198
                .squeeze(1)
3199
            )
3200
3201
3202
            # Some tokens ids may need to become embeds
            if token_ids_idx.numel() > 0:
                token_ids = self.input_ids.gpu[token_ids_idx]
3203
                tokens_to_embeds = self.model.embed_input_ids(input_ids=token_ids)
3204
3205
3206
                self.inputs_embeds.gpu[token_ids_idx] = tokens_to_embeds

            inputs_embeds = self.inputs_embeds.gpu[:num_input_tokens]
3207
            model_kwargs = self._init_model_kwargs()
3208
            input_ids = None
3209
        else:
3210
3211
3212
3213
            # 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.
3214
            input_ids = self.input_ids.gpu[:num_input_tokens]
3215
            inputs_embeds = None
3216
            model_kwargs = self._init_model_kwargs()
3217

guanyu1's avatar
guanyu1 committed
3218
        positions = self._get_positions(num_input_tokens)       
3219

3220
        if is_first_rank:
3221
3222
            intermediate_tensors = None
        else:
3223
            assert intermediate_tensors is not None
3224
            intermediate_tensors = self.sync_and_slice_intermediate_tensors(
3225
3226
                num_input_tokens, intermediate_tensors, True
            )
3227

3228
        if is_encoder_decoder and scheduler_output.scheduled_encoder_inputs:
3229
3230
3231
3232
3233
3234
3235
            # 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})
3236

3237
3238
3239
3240
3241
3242
        return (
            input_ids,
            inputs_embeds,
            positions,
            intermediate_tensors,
            model_kwargs,
3243
            ec_connector_output,
3244
        )
3245

3246
    def _sample(
3247
        self,
3248
3249
        logits: torch.Tensor | None,
        spec_decode_metadata: SpecDecodeMetadata | None,
3250
    ) -> SamplerOutput:
3251
        # Sample the next token and get logprobs if needed.
3252
        sampling_metadata = self.input_batch.sampling_metadata
3253
3254
3255
        # 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()
3256
        if spec_decode_metadata is None:
3257
            return self.sampler(
3258
3259
3260
                logits=logits,
                sampling_metadata=sampling_metadata,
            )
3261

3262
3263
3264
3265
3266
3267
        # 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)

3268
        sampler_output = self.rejection_sampler(
3269
            spec_decode_metadata,
王敏's avatar
王敏 committed
3270
3271
            None if self.draft_probs is None else \
                self.draft_probs.get_probs(spec_decode_metadata.spec_decode_ids),  # draft_probs
3272
            logits,
3273
3274
            sampling_metadata,
        )
3275
3276
3277
        return sampler_output

    def _bookkeeping_sync(
3278
3279
3280
        self,
        scheduler_output: "SchedulerOutput",
        sampler_output: SamplerOutput,
3281
        logits: torch.Tensor | None,
3282
3283
        hidden_states: torch.Tensor,
        num_scheduled_tokens: int,
3284
        spec_decode_metadata: SpecDecodeMetadata | None,
3285
    ) -> tuple[
3286
        dict[str, int],
3287
        LogprobsLists | None,
3288
        list[list[int]],
3289
        dict[str, LogprobsTensors | None],
3290
3291
3292
        list[str],
        dict[str, int],
        list[int],
3293
    ]:
3294
3295
3296
3297
        num_nans_in_logits = {}
        if envs.VLLM_COMPUTE_NANS_IN_LOGITS:
            num_nans_in_logits = self._get_nans_in_logits(logits)

3298
3299
3300
3301
        num_reqs = self.input_batch.num_reqs
        discard_sampled_tokens_req_indices = np.nonzero(
            self.discard_request_mask.np[:num_reqs]
        )[0]
3302
3303
3304
3305
        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)
3306

3307
3308
3309
        # 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()
3310
        req_id_to_index_output_copy = self.input_batch.req_id_to_index.copy()
3311

3312
        num_sampled_tokens = sampler_output.sampled_token_ids.shape[0]
3313
        sampled_token_ids = sampler_output.sampled_token_ids
3314
        logprobs_tensors = sampler_output.logprobs_tensors
3315
        invalid_req_indices = []
3316
        logprobs_lists = None
3317
3318
3319
3320
3321
3322
        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)
3323
3324
3325
                # 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()
3326
3327
3328

                if logprobs_tensors is not None:
                    logprobs_lists = logprobs_tensors.tolists()
3329
3330
            else:
                # Includes spec decode tokens.
3331
                valid_sampled_token_ids, logprobs_lists = RejectionSampler.parse_output(
3332
3333
                    sampled_token_ids,
                    self.input_batch.vocab_size,
3334
                    discard_sampled_tokens_req_indices,
3335
                    logprobs_tensors=logprobs_tensors,
3336
                )
3337
        else:
3338
            valid_sampled_token_ids = []
3339
            invalid_req_indices = discard_sampled_tokens_req_indices.tolist()
3340
3341
3342
3343
3344
            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.
3345
3346
3347
3348
            # 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
3349
3350
3351
3352
3353
            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
            }
3354

3355
3356
3357
3358
3359
        # 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.
3360
        req_ids = self.input_batch.req_ids
3361
3362
        for req_idx in range(num_sampled_tokens):
            if self.use_async_scheduling:
3363
                sampled_ids = [-1] if req_idx not in invalid_req_indices_set else None
3364
3365
            else:
                sampled_ids = valid_sampled_token_ids[req_idx]
3366

3367
            num_sampled_ids: int = len(sampled_ids) if sampled_ids else 0
3368

3369
3370
3371
3372
            if not sampled_ids:
                continue

            start_idx = self.input_batch.num_tokens_no_spec[req_idx]
3373
            end_idx = start_idx + num_sampled_ids
3374
3375
3376
            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: "
3377
                f"{self.max_model_len}"
3378
            )
3379

3380
            self.input_batch.token_ids_cpu[req_idx, start_idx:end_idx] = sampled_ids
3381
            self.input_batch.is_token_ids[req_idx, start_idx:end_idx] = True
3382
            self.input_batch.num_tokens_no_spec[req_idx] = end_idx
3383

3384
            req_id = req_ids[req_idx]
3385
3386
3387
            req_state = self.requests[req_id]
            req_state.output_token_ids.extend(sampled_ids)

3388
3389
3390
3391
3392
3393
        # Compute prompt logprobs if needed.
        prompt_logprobs_dict = self._get_prompt_logprobs_dict(
            hidden_states[:num_scheduled_tokens],
            scheduler_output.num_scheduled_tokens,
        )

3394
3395
3396
3397
3398
3399
3400
3401
3402
3403
        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,
        )

3404
3405
3406
3407
3408
3409
3410
3411
3412
3413
3414
3415
3416
3417
3418
    @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()

3419
3420
    def _model_forward(
        self,
3421
3422
3423
3424
        input_ids: torch.Tensor | None = None,
        positions: torch.Tensor | None = None,
        intermediate_tensors: IntermediateTensors | None = None,
        inputs_embeds: torch.Tensor | None = None,
3425
3426
3427
3428
3429
        **model_kwargs: dict[str, Any],
    ) -> Any:
        """Helper method to call the model forward pass.

        This method can be overridden by subclasses for model execution.
3430
        Motivation: We can inspect only this method versus
3431
3432
3433
3434
3435
3436
3437
3438
3439
3440
3441
3442
3443
3444
3445
3446
3447
3448
3449
3450
        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,
        )

3451
3452
3453
3454
3455
3456
3457
3458
3459
3460
3461
3462
3463
3464
3465
3466
3467
3468
3469
3470
3471
    @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
        )

3472
3473
3474
3475
3476
3477
3478
3479
3480
3481
3482
3483
3484
    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,
3485
        num_encoder_reqs: int = 0,
3486
    ) -> tuple[
3487
3488
        CUDAGraphMode,
        BatchDescriptor,
3489
        bool,
3490
3491
        torch.Tensor | None,
        CUDAGraphStat | None,
3492
    ]:
3493
3494
3495
3496
3497
3498
        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,
3499
        )
3500
3501
3502
3503
3504
        # 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
        )
3505
3506
3507
3508
3509
3510
3511

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

3512
        num_tokens_padded = self._pad_for_sequence_parallelism(num_tokens)
3513
        dispatch_cudagraph = (
3514
            lambda num_tokens, disable_full: self.cudagraph_dispatcher.dispatch(
3515
3516
3517
                num_tokens=num_tokens,
                has_lora=has_lora,
                uniform_decode=uniform_decode,
3518
                disable_full=disable_full,
3519
3520
3521
3522
3523
            )
            if not force_eager
            else (CUDAGraphMode.NONE, BatchDescriptor(num_tokens_padded))
        )

3524
        cudagraph_mode, batch_descriptor = dispatch_cudagraph(
3525
            num_tokens_padded, use_cascade_attn or has_encoder_output
3526
        )
3527
        num_tokens_padded = batch_descriptor.num_tokens
3528
3529
3530
3531
3532
3533
3534
3535
3536
        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"
            )
3537
3538
3539

        # Extra coordination when running data-parallel since we need to coordinate
        # across ranks
3540
        should_ubatch, num_tokens_across_dp = False, None
3541
3542
3543
3544
3545
3546
3547
3548
3549
        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
            )

3550
3551
3552
3553
3554
3555
3556
3557
3558
3559
3560
            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,
                )
3561
3562
            )

3563
            # Extract DP-synced values
3564
3565
3566
            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())
3567
3568
3569
3570
3571
                # 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,
                )
3572
3573
3574
3575
                # 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

3576
3577
3578
3579
3580
3581
3582
3583
3584
3585
3586
3587
        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,
3588
            should_ubatch,
3589
3590
3591
            num_tokens_across_dp,
            cudagraph_stats,
        )
3592

3593
3594
3595
3596
3597
3598
3599
3600
3601
3602
3603
3604
3605
3606
3607
3608
3609
3610
3611
3612
3613
3614
3615
3616
3617
3618
3619
3620
3621
3622
3623
3624
3625
3626
3627
    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
3628

3629
3630
3631
3632
3633
3634
3635
3636
3637
3638
3639
3640
3641
3642
3643
3644
3645
3646
3647
3648
3649
3650
3651
3652
3653
3654
3655
3656
3657
3658
3659
3660
3661
3662
3663
3664
3665
3666
3667
3668
3669
3670
3671
3672
3673
3674
3675
3676
3677
3678
3679
3680
3681
3682
3683
3684
3685
3686
3687
3688
3689
3690
3691
3692
3693
3694
3695
3696
3697
3698
3699
3700
3701
3702
    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

3703
3704
3705
3706
    @torch.inference_mode()
    def execute_model(
        self,
        scheduler_output: "SchedulerOutput",
3707
        intermediate_tensors: IntermediateTensors | None = None,
3708
    ) -> ModelRunnerOutput | AsyncModelRunnerOutput | IntermediateTensors | None:
3709
3710
3711
3712
3713
        if self.execute_model_state is not None:
            raise RuntimeError(
                "State error: sample_tokens() must be called "
                "after execute_model() returns None."
            )
3714

3715
3716
3717
3718
3719
3720
        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.")
3721

3722
3723
3724
3725
        if scheduler_output.preempted_req_ids and has_kv_transfer_group():
            get_kv_transfer_group().handle_preemptions(
                scheduler_output.preempted_req_ids
            )
3726

3727
        num_scheduled_tokens = scheduler_output.total_num_scheduled_tokens
3728
3729
3730
3731
3732
3733
        with (
            record_function_or_nullcontext("gpu_model_runner: preprocess"),
            self.synchronize_input_prep(),
        ):
            # Update persistent batch states.
            self._update_states(scheduler_output)
3734

3735
3736
            if has_ec_transfer() and get_ec_transfer().is_producer:
                with self.maybe_get_ec_connector_output(
3737
                    scheduler_output,
3738
3739
3740
3741
3742
3743
3744
3745
3746
3747
3748
3749
3750
3751
3752
3753
3754
3755
3756
3757
3758
3759
3760
3761
3762
3763
3764
3765
                    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"
3766
                )
3767

3768
3769
3770
3771
3772
3773
            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
3774

3775
3776
3777
3778
            logits_indices, spec_decode_metadata = self._prepare_inputs(
                scheduler_output,
                num_scheduled_tokens_np,
            )
3779

3780
3781
3782
3783
3784
            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(
3785
                    num_scheduled_tokens_np,
3786
3787
                    self.input_batch.num_computed_tokens_cpu[:num_reqs],
                    scheduler_output.num_common_prefix_blocks,
3788
                )
3789

3790
3791
3792
3793
3794
3795
3796
3797
3798
3799
3800
3801
3802
3803
            (
                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),
            )
3804

3805
3806
3807
3808
3809
3810
3811
3812
3813
3814
            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
3815
            if self.enable_lightly_cp and num_tokens_unpadded > self.lightly_cp_threshould:
王敏's avatar
王敏 committed
3816
                num_tokens_padded = self._pad_for_mla_cp(num_tokens_unpadded)
3817
3818
3819
3820
3821
3822
3823
3824
3825
3826
3827
3828
3829
3830
3831
3832
3833
            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,
            )

3834
3835
3836
3837
3838
3839
3840
3841
3842
3843
3844
            # 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)
            )
3845
3846
            pad_attn = cudagraph_mode == CUDAGraphMode.FULL

3847
3848
3849
3850
3851
3852
3853
3854
3855
3856
3857
3858
            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(),
                )

3859
3860
3861
            use_spec_decode = len(scheduler_output.scheduled_spec_decode_tokens) > 0
            ubatch_slices_attn = ubatch_slices_padded if pad_attn else ubatch_slices

3862
3863
3864
3865
3866
3867
3868
3869
3870
3871
3872
            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,
            )

王敏's avatar
王敏 committed
3873
3874
3875
3876
3877
3878
            (
                attn_metadata,
                spec_decode_common_attn_metadata,
                scatter_indexes_tensor,
                gather_indexes_tensor,
            ) = self._build_attention_metadata(
3879
3880
3881
3882
3883
3884
3885
3886
3887
3888
                    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,
3889
                    slot_mappings=slot_mappings_by_group,
3890
                )
3891
3892
3893
3894
3895
3896
3897

            (
                input_ids,
                inputs_embeds,
                positions,
                intermediate_tensors,
                model_kwargs,
3898
3899
3900
                ec_connector_output,
            ) = self._preprocess(
                scheduler_output, num_tokens_padded, intermediate_tensors
3901
            )
3902

3903
        # Set cudagraph mode to none if calc_kv_scales is true.
3904
3905
3906
        # KV scales calculation involves dynamic operations that are incompatible
        # with CUDA graph capture.
        if self.calculate_kv_scales:
3907
            cudagraph_mode = CUDAGraphMode.NONE
3908
3909
            # Mark KV scales as calculated after the first forward pass
            self.calculate_kv_scales = False
3910

3911
3912
3913
3914
3915
3916
3917
        # 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
        )

3918
3919
        # Run the model.
        # Use persistent buffers for CUDA graphs.
3920
        clear_kv_metadata = self.speculative_config is None
3921
3922
        with (
            set_forward_context(
3923
3924
                attn_metadata,
                self.vllm_config,
3925
                num_tokens=num_tokens_padded,
3926
                num_tokens_across_dp=num_tokens_across_dp,
3927
3928
                cudagraph_runtime_mode=cudagraph_mode,
                batch_descriptor=batch_desc,
3929
                ubatch_slices=ubatch_slices_padded,
3930
                slot_mapping=slot_mappings,
3931
                skip_compiled=has_encoder_input,
王敏's avatar
王敏 committed
3932
3933
                scatter_indexes_tensor=scatter_indexes_tensor,
                gather_indexes_tensor=gather_indexes_tensor,
3934
3935
                enable_lightly_cp=self.enable_lightly_cp and num_tokens_unpadded > self.lightly_cp_threshould,
                enable_lightly_cplb=self.enable_lightly_cplb
3936
            ),
3937
            record_function_or_nullcontext("gpu_model_runner: forward"),
3938
3939
3940
            self.maybe_get_kv_connector_output(
                scheduler_output, clear_metadata=clear_kv_metadata
            ) as kv_connector_output,
3941
        ):
3942
            model_output = self._model_forward(
3943
3944
3945
3946
3947
3948
3949
                input_ids=input_ids,
                positions=positions,
                intermediate_tensors=intermediate_tensors,
                inputs_embeds=inputs_embeds,
                **model_kwargs,
            )

3950
        with record_function_or_nullcontext("gpu_model_runner: postprocess"):
3951
            if self.use_aux_hidden_state_outputs:
3952
                # True when EAGLE 3 is used.
3953
3954
                hidden_states, aux_hidden_states = model_output
            else:
3955
                # Common case.
3956
3957
3958
                hidden_states = model_output
                aux_hidden_states = None

3959
3960
3961
3962
3963
            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)
3964
                    hidden_states.kv_connector_output = kv_connector_output
3965
                    self.kv_connector_output = kv_connector_output
3966
                    return hidden_states
3967

3968
                if self.is_pooling_model:
3969
                    # Return the pooling output.
3970
3971
3972
3973
3974
                    return self._pool(
                        hidden_states,
                        num_scheduled_tokens,
                        num_scheduled_tokens_np,
                        kv_connector_output,
3975
                    )
3976
3977

                sample_hidden_states = hidden_states[logits_indices]
3978
                logits = self.model.compute_logits(sample_hidden_states)
3979
3980
3981
3982
            else:
                # Rare case.
                assert not self.is_pooling_model

3983
                sample_hidden_states = hidden_states[logits_indices]
3984
                if not get_pp_group().is_last_rank:
3985
                    all_gather_tensors = {
3986
                        "residual": not is_residual_scattered_for_sp(
3987
                            self.vllm_config, num_tokens_padded
3988
                        )
3989
                    }
3990
                    get_pp_group().send_tensor_dict(
3991
3992
                        hidden_states.tensors,
                        all_gather_group=get_tp_group(),
3993
3994
                        all_gather_tensors=all_gather_tensors,
                    )
3995
3996
                    logits = None
                else:
3997
                    logits = self.model.compute_logits(sample_hidden_states)
3998

3999
                model_output_broadcast_data: dict[str, Any] = {}
4000
4001
4002
                if logits is not None:
                    model_output_broadcast_data["logits"] = logits.contiguous()

4003
                broadcasted = get_pp_group().broadcast_tensor_dict(
4004
4005
                    model_output_broadcast_data, src=len(get_pp_group().ranks) - 1
                )
4006
4007
                assert broadcasted is not None
                logits = broadcasted["logits"]
4008

4009
4010
4011
4012
4013
4014
4015
4016
        self.execute_model_state = ExecuteModelState(
            scheduler_output,
            logits,
            spec_decode_metadata,
            spec_decode_common_attn_metadata,
            hidden_states,
            sample_hidden_states,
            aux_hidden_states,
4017
            ec_connector_output,
4018
            cudagraph_stats,
4019
            slot_mappings,
4020
        )
4021
        self.kv_connector_output = kv_connector_output
4022
4023
4024
4025
4026
4027
        return None

    @torch.inference_mode
    def sample_tokens(
        self, grammar_output: "GrammarOutput | None"
    ) -> ModelRunnerOutput | AsyncModelRunnerOutput | IntermediateTensors:
4028
4029
4030
        kv_connector_output = self.kv_connector_output
        self.kv_connector_output = None

4031
4032
        if self.execute_model_state is None:
            # Nothing to do (PP non-final rank case), output isn't used.
4033
            if not kv_connector_output:
4034
                return None  # type: ignore[return-value]
4035
4036
4037
4038
4039
4040
4041
4042
4043

            # 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
4044

4045
4046
4047
4048
4049
4050
4051
4052
4053
        # Unpack ephemeral state.
        (
            scheduler_output,
            logits,
            spec_decode_metadata,
            spec_decode_common_attn_metadata,
            hidden_states,
            sample_hidden_states,
            aux_hidden_states,
4054
            ec_connector_output,
4055
            cudagraph_stats,
4056
            slot_mappings,
4057
4058
4059
4060
4061
4062
4063
4064
4065
        ) = 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
            )
4066

4067
        with record_function_or_nullcontext("gpu_model_runner: sample"):
4068
4069
            sampler_output = self._sample(logits, spec_decode_metadata)

4070
4071
4072
4073
        self._update_states_after_model_execute(
            sampler_output.sampled_token_ids, scheduler_output
        )

4074
4075
        self._draft_token_ids = None
        self._draft_token_req_ids = None
4076
4077
        self.input_batch.prev_sampled_token_ids = None

4078
4079
        def propose_draft_token_ids(sampled_token_ids):
            assert spec_decode_common_attn_metadata is not None
4080
            with record_function_or_nullcontext("gpu_model_runner: draft"):
4081
4082
4083
4084
4085
4086
4087
4088
4089
                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,
4090
                    slot_mappings,
4091
                )
4092
                self._copy_draft_token_ids_to_cpu(scheduler_output)
4093

4094
        spec_config = self.speculative_config
4095
4096
4097
4098
4099
        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
4100
            )
4101
4102
4103
4104
4105
            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
4106
                # as inputs, and does not need to wait for bookkeeping to finish.
4107
                assert isinstance(self.drafter, EagleProposer | DraftModelProposer)
4108
4109
4110
4111
4112
4113
4114
4115
4116
4117
4118
4119
4120
                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,
                        )
4121
                    )
4122
4123
4124
4125
4126
4127
4128
4129
4130
4131
4132
                    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
4133

4134
        with record_function_or_nullcontext("gpu_model_runner: bookkeep"):
4135
4136
4137
4138
4139
4140
4141
4142
            (
                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,
4143
4144
4145
4146
4147
            ) = self._bookkeeping_sync(
                scheduler_output,
                sampler_output,
                logits,
                hidden_states,
4148
                scheduler_output.total_num_scheduled_tokens,
4149
                spec_decode_metadata,
4150
            )
4151

4152
        if propose_drafts_after_bookkeeping:
4153
4154
4155
            # 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)
4156
4157
4158
            
        if self.speculative_config is not None:
            self.clear_kv_connector_metadata()
4159

4160
        with record_function_or_nullcontext("gpu_model_runner: eplb"):
4161
            self.eplb_step()
4162

4163
        with record_function_or_nullcontext("gpu_model_runner: ModelRunnerOutput"):
4164
4165
4166
4167
4168
4169
4170
            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.")

4171
4172
4173
4174
4175
4176
4177
            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,
4178
4179
4180
                ec_connector_output=ec_connector_output
                if self.supports_mm_inputs
                else None,
4181
                num_nans_in_logits=num_nans_in_logits,
4182
                cudagraph_stats=cudagraph_stats,
4183
            )
4184

4185
4186
        if not self.use_async_scheduling:
            return output
4187

4188
4189
4190
4191
4192
4193
4194
4195
4196
        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,
4197
                vocab_size=self.input_batch.vocab_size,
4198
4199
4200
4201
4202
            )
        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
4203
            # any requests with sampling params that require output ids.
4204
4205
4206
4207
            self.input_batch.set_async_sampled_token_ids(
                async_output.sampled_token_ids_cpu,
                async_output.async_copy_ready_event,
            )
4208

4209
        return async_output
4210

4211
    def take_draft_token_ids(self) -> DraftTokenIds | None:
4212
        if not self.num_spec_tokens or not self._draft_token_req_ids:
4213
            return None
4214
        draft_token_ids, req_ids = self._get_draft_token_ids_cpu()
4215
4216
        return DraftTokenIds(req_ids, draft_token_ids)

4217
4218
4219
    def _copy_draft_token_ids_to_cpu(
        self, scheduler_output: "SchedulerOutput", zeros_only: bool = False
    ) -> None:
4220
4221
4222
4223
4224
4225
        # 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
        ):
4226
4227
4228
            return
        # We must also set the corresponding request ids.
        self._draft_token_req_ids = self.input_batch.req_ids.copy()
4229

4230
4231
4232
4233
4234
4235
4236
4237
4238
4239
4240
4241
4242
4243
4244
4245
4246
4247
4248
4249
        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()

4250
    def _get_draft_token_ids_cpu(self) -> tuple[list[list[int]], list[str]]:
4251
        if isinstance(self._draft_token_ids, list):
4252
4253
4254
4255
            return self._draft_token_ids, self.input_batch.req_ids
        req_ids = self._draft_token_req_ids
        if req_ids is None:
            return [], []
4256
4257
4258
        assert self.draft_token_ids_event is not None
        assert self.draft_token_ids_cpu is not None
        self.draft_token_ids_event.synchronize()
4259
        return self.draft_token_ids_cpu[: len(req_ids)].tolist(), req_ids
4260

4261
4262
4263
4264
4265
4266
4267
4268
4269
4270
4271
4272
4273
    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
4274
            assert counts_cpu is not None
4275
4276
4277
4278
4279
4280
4281
4282
            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
4283
4284
        sampled_count_event = self.valid_sampled_token_count_event
        if sampled_count_event is None or prev_sampled_token_ids is None:
4285
4286
4287
            return []

        counts_cpu = self.valid_sampled_token_count_cpu
4288
4289
        assert counts_cpu is not None
        sampled_count_event.synchronize()
4290
4291
        return counts_cpu[: prev_sampled_token_ids.shape[0]].tolist()

4292
4293
4294
    def propose_draft_token_ids(
        self,
        scheduler_output: "SchedulerOutput",
4295
        sampled_token_ids: torch.Tensor | list[list[int]],
4296
4297
4298
        sampling_metadata: SamplingMetadata,
        hidden_states: torch.Tensor,
        sample_hidden_states: torch.Tensor,
4299
4300
        aux_hidden_states: list[torch.Tensor] | None,
        spec_decode_metadata: SpecDecodeMetadata | None,
4301
        common_attn_metadata: CommonAttentionMetadata,
4302
        slot_mappings: dict[str, torch.Tensor] | list[dict[str, torch.Tensor]] | None,
4303
    ) -> list[list[int]] | torch.Tensor:
4304
        num_scheduled_tokens = scheduler_output.total_num_scheduled_tokens
4305
4306
4307
        spec_config = self.speculative_config
        assert spec_config is not None
        if spec_config.method == "ngram":
4308
            assert isinstance(sampled_token_ids, list)
4309
            assert isinstance(self.drafter, NgramProposer)
4310
            draft_token_ids = self.drafter.propose(
4311
                sampled_token_ids,
4312
4313
                self.input_batch.num_tokens_no_spec,
                self.input_batch.token_ids_cpu,
4314
                slot_mappings=slot_mappings,
4315
            )
4316
        elif spec_config.method == "suffix":
4317
4318
            assert isinstance(sampled_token_ids, list)
            assert isinstance(self.drafter, SuffixDecodingProposer)
4319
4320
4321
            draft_token_ids = self.drafter.propose(
                self.input_batch, sampled_token_ids, slot_mappings=slot_mappings
            )
4322
        elif spec_config.method == "medusa":
4323
            assert isinstance(sampled_token_ids, list)
4324
            assert isinstance(self.drafter, MedusaProposer)
4325

4326
4327
            if sample_hidden_states.shape[0] == len(sampled_token_ids):
                # The input to the target model does not include draft tokens.
4328
4329
4330
4331
                hidden_states = sample_hidden_states
            else:
                indices = []
                offset = 0
4332
4333
4334
                assert spec_decode_metadata is not None, (
                    "No spec decode metadata for medusa"
                )
4335
                for num_draft, tokens in zip(
4336
4337
                    spec_decode_metadata.num_draft_tokens, sampled_token_ids
                ):
4338
4339
                    indices.append(offset + len(tokens) - 1)
                    offset += num_draft + 1
4340
                indices = torch.tensor(indices, device=self.device)
4341
4342
                hidden_states = sample_hidden_states[indices]

4343
            draft_token_ids = self.drafter.propose(
4344
4345
                target_hidden_states=hidden_states,
                sampling_metadata=sampling_metadata,
4346
                slot_mappings=slot_mappings,
4347
            )
4348
4349
        elif spec_config.use_eagle() or spec_config.uses_draft_model():
            assert isinstance(self.drafter, EagleProposer | DraftModelProposer)
4350

4351
            if spec_config.disable_padded_drafter_batch:
4352
4353
4354
                # 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.
4355
4356
                assert isinstance(sampled_token_ids, list), (
                    "sampled_token_ids should be a python list when"
4357
                    "padded-batch is disabled."
4358
                )
4359
                next_token_ids = self.drafter.prepare_next_token_ids_cpu(
4360
4361
4362
4363
4364
                    sampled_token_ids,
                    self.requests,
                    self.input_batch,
                    scheduler_output.num_scheduled_tokens,
                )
4365
4366
4367
4368
4369
            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.
4370
4371
                assert isinstance(sampled_token_ids, torch.Tensor), (
                    "sampled_token_ids should be a torch.Tensor when"
4372
                    "padded-batch is enabled."
4373
4374
                )
                next_token_ids, valid_sampled_tokens_count = (
4375
4376
4377
4378
4379
                    self.drafter.prepare_next_token_ids_padded(
                        common_attn_metadata,
                        sampled_token_ids,
                        self.requests,
                        self.input_batch,
4380
                        self.discard_request_mask.gpu,
4381
                    )
4382
                )
4383
4384
4385
                self._copy_valid_sampled_token_count(
                    next_token_ids, valid_sampled_tokens_count
                )
Jiayi Yao's avatar
Jiayi Yao committed
4386

4387
            num_rejected_tokens_gpu = None
4388
            if spec_decode_metadata is None:
4389
                token_indices_to_sample = None
4390
                # input_ids can be None for multimodal models.
4391
                target_token_ids = self.input_ids.gpu[:num_scheduled_tokens]
4392
                target_positions = self._get_positions(num_scheduled_tokens)
4393
                if self.use_aux_hidden_state_outputs:
Wentao Ye's avatar
Wentao Ye committed
4394
                    assert aux_hidden_states is not None
4395
                    target_hidden_states = torch.cat(
4396
4397
                        [h[:num_scheduled_tokens] for h in aux_hidden_states], dim=-1
                    )
4398
4399
                else:
                    target_hidden_states = hidden_states[:num_scheduled_tokens]
4400
            else:
4401
                if spec_config.disable_padded_drafter_batch:
4402
                    token_indices_to_sample = None
4403
4404
4405
4406
4407
                    common_attn_metadata, token_indices = self.drafter.prepare_inputs(
                        common_attn_metadata,
                        sampled_token_ids,
                        spec_decode_metadata.num_draft_tokens,
                    )
4408
4409
4410
4411
4412
4413
4414
4415
4416
                    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]
4417
                else:
4418
4419
4420
4421
4422
4423
4424
4425
                    (
                        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,
4426
                    )
王敏's avatar
王敏 committed
4427
4428
                    #total_num_tokens = common_attn_metadata.num_actual_tokens
                    if (
4429
                        self.enable_lightly_cp
王敏's avatar
王敏 committed
4430
4431
4432
4433
4434
4435
4436
                        and common_attn_metadata.cp_common_metadata is not None
                    ):
                        total_num_tokens = (
                            common_attn_metadata.cp_common_metadata.num_actual_tokens
                        )
                    else:
                        total_num_tokens = common_attn_metadata.num_actual_tokens
4437
4438
4439
4440
4441
4442
4443
4444
4445
4446
                    # 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]
4447

4448
            if self.supports_mm_inputs:
4449
4450
4451
4452
4453
4454
                mm_embed_inputs = self._gather_mm_embeddings(
                    scheduler_output,
                    shift_computed_tokens=1,
                )
            else:
                mm_embed_inputs = None
4455

王敏's avatar
王敏 committed
4456
            draft_result = self.drafter.propose(
4457
4458
4459
4460
                target_token_ids=target_token_ids,
                target_positions=target_positions,
                target_hidden_states=target_hidden_states,
                next_token_ids=next_token_ids,
4461
                last_token_indices=token_indices_to_sample,
4462
                sampling_metadata=sampling_metadata,
4463
                common_attn_metadata=common_attn_metadata,
4464
                mm_embed_inputs=mm_embed_inputs,
4465
                num_rejected_tokens_gpu=num_rejected_tokens_gpu,
4466
                slot_mappings=slot_mappings,
4467
            )
4468

王敏's avatar
王敏 committed
4469
4470
4471
4472
4473
4474
4475
4476
4477
4478
4479
4480
4481
            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)

4482
        return draft_token_ids
4483

4484
4485
4486
    def update_config(self, overrides: dict[str, Any]) -> None:
        allowed_config_names = {"load_config", "model_config"}
        for config_name, config_overrides in overrides.items():
4487
4488
            assert config_name in allowed_config_names, (
                f"Config `{config_name}` not supported. "
4489
                f"Allowed configs: {allowed_config_names}"
4490
            )
4491
4492
4493
4494
            config = getattr(self, config_name)
            new_config = update_config(config, config_overrides)
            setattr(self, config_name, new_config)

4495
4496
4497
4498
4499
    def load_model(self, eep_scale_up: bool = False) -> None:
        """
        Args:
            eep_scale_up: the model loading is for elastic EP scale up.
        """
4500
4501
4502
4503
4504
        logger.info_once(
            "Starting to load model %s...",
            self.model_config.model,
            scope="global",
        )
4505
4506
4507
4508
4509
        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)
        )
4510

4511
4512
4513
4514
        if self.parallel_config.enable_eplb:
            self.eplb_state = EplbState(self.parallel_config, self.device)
            eplb_models = 0

4515
4516
4517
4518
4519
4520
        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
4521
                )
4522
4523
4524
                if self.lora_config:
                    self.model = self.load_lora_model(
                        self.model, self.vllm_config, self.device
4525
                    )
4526
4527
4528
4529
4530
4531
4532
4533
4534
4535
4536
4537
4538
4539
4540
                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,
                        )
4541

4542
4543
4544
4545
4546
4547
4548
4549
4550
4551
4552
4553
4554
4555
4556
4557
4558
4559
4560
4561
4562
4563
                        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
4564

4565
4566
4567
4568
4569
4570
                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"
                        )
4571

4572
4573
4574
4575
4576
4577
4578
4579
4580
4581
                    # 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()
4582

4583
4584
4585
4586
4587
4588
4589
4590
4591
4592
4593
4594
4595
4596
                    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
4597
        logger.info_once(
4598
4599
            "Model loading took %s GiB memory and %.6f seconds",
            format_gib(self.model_memory_usage),
4600
            time_after_load - time_before_load,
4601
            scope="local",
4602
        )
4603
        prepare_communication_buffer_for_model(self.model)
4604
4605
4606
4607
        if (drafter := getattr(self, "drafter", None)) and (
            drafter_model := getattr(drafter, "model", None)
        ):
            prepare_communication_buffer_for_model(drafter_model)
4608
        mm_config = self.model_config.multimodal_config
4609
        self.is_multimodal_pruning_enabled = (
4610
            supports_multimodal_pruning(self.get_model())
4611
4612
            and mm_config is not None
            and mm_config.is_multimodal_pruning_enabled()
4613
        )
4614

4615
        if is_mixture_of_experts(self.model) and self.parallel_config.enable_eplb:
4616
4617
4618
4619
4620
4621
4622
4623
4624
4625
4626
            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(
4627
                self.model,
4628
                self.model_config,
4629
4630
4631
                global_expert_load,
                old_global_expert_indices,
                rank_mapping,
4632
            )
4633
4634
            if self.eplb_state.is_async:
                self.eplb_state.start_async_loop(rank_mapping=rank_mapping)
4635

4636
        if (
4637
4638
            self.vllm_config.compilation_config.mode
            == CompilationMode.STOCK_TORCH_COMPILE
4639
        ):
4640
            backend = self.vllm_config.compilation_config.init_backend(self.vllm_config)
4641
            compilation_counter.stock_torch_compile_count += 1
4642
            self.model.compile(fullgraph=True, backend=backend)
4643
            return
4644
        # for other compilation modes, cudagraph behavior is controlled by
4645
4646
4647
        # CudagraphWraper and CudagraphDispatcher of vllm.

        # wrap the model with full cudagraph wrapper if needed.
4648
4649
        cudagraph_mode = self.compilation_config.cudagraph_mode
        assert cudagraph_mode is not None
4650
4651
4652
4653
        if (
            cudagraph_mode.has_full_cudagraphs()
            and not self.parallel_config.use_ubatching
        ):
4654
4655
4656
            self.model = CUDAGraphWrapper(
                self.model, self.vllm_config, runtime_mode=CUDAGraphMode.FULL
            )
4657
        elif self.parallel_config.use_ubatching:
4658
            if cudagraph_mode.has_full_cudagraphs():
4659
4660
4661
                self.model = UBatchWrapper(
                    self.model, self.vllm_config, CUDAGraphMode.FULL, self.device
                )
4662
            else:
4663
4664
4665
                self.model = UBatchWrapper(
                    self.model, self.vllm_config, CUDAGraphMode.NONE, self.device
                )
4666

4667
    def _get_eagle3_aux_layers_from_config(self) -> tuple[int, ...] | None:
4668
4669
4670
4671
4672
4673
4674
4675
4676
4677
4678
4679
4680
4681
4682
4683
4684
4685
4686
4687
4688
4689
        """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
4690

4691
    def reload_weights(self) -> None:
4692
        assert getattr(self, "model", None) is not None, (
4693
            "Cannot reload weights before model is loaded."
4694
        )
4695
4696
        model_loader = get_model_loader(self.load_config)
        logger.info("Reloading weights inplace...")
4697
        model_loader.load_weights(self.get_model(), model_config=self.model_config)
4698

4699
4700
4701
4702
4703
    def save_tensorized_model(
        self,
        tensorizer_config: "TensorizerConfig",
    ) -> None:
        TensorizerLoader.save_model(
4704
            self.get_model(),
4705
            tensorizer_config=tensorizer_config,
4706
            model_config=self.model_config,
4707
4708
        )

4709
4710
4711
    def _get_prompt_logprobs_dict(
        self,
        hidden_states: torch.Tensor,
4712
        num_scheduled_tokens: dict[str, int],
4713
    ) -> dict[str, LogprobsTensors | None]:
4714
        num_prompt_logprobs_dict = self.num_prompt_logprobs
4715
4716
4717
        if not num_prompt_logprobs_dict:
            return {}

4718
        in_progress_dict = self.input_batch.in_progress_prompt_logprobs_cpu
4719
        prompt_logprobs_dict: dict[str, LogprobsTensors | None] = {}
4720
4721
4722
4723
4724

        # 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():
4725
4726
4727
4728
            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
4729
4730
4731

            # Get metadata for this request.
            request = self.requests[req_id]
4732
4733
4734
4735
            if request.prompt_token_ids is None:
                # Prompt logprobs is incompatible with prompt embeddings
                continue

4736
4737
            num_prompt_tokens = len(request.prompt_token_ids)
            prompt_token_ids = torch.tensor(request.prompt_token_ids).to(
4738
4739
                self.device, non_blocking=True
            )
4740

4741
4742
4743
4744
4745
4746
            # 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(
4747
4748
                    num_prompt_tokens - 1, num_prompt_logprobs + 1
                )
4749
4750
                in_progress_dict[req_id] = logprobs_tensors

4751
            # Determine number of logits to retrieve.
4752
4753
            start_idx = request.num_computed_tokens
            start_tok = start_idx + 1
4754
            num_remaining_tokens = num_prompt_tokens - start_tok
4755
            if num_tokens <= num_remaining_tokens:
4756
                # This is a chunk, more tokens remain.
4757
4758
4759
                # 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.
4760
4761
4762
4763
4764
                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)
4765
4766
4767
4768
4769
4770
4771
                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
4772
4773
4774
4775
4776

            # 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]
4777
            offset = self.query_start_loc.np[req_idx].item()
4778
            prompt_hidden_states = hidden_states[offset : offset + num_logits]
4779
            logits = self.model.compute_logits(prompt_hidden_states)
4780
4781
4782
4783

            # 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.
4784
            tgt_token_ids = prompt_token_ids[start_tok : start_tok + num_logits]
4785
4786

            # Compute prompt logprobs.
4787
4788
            logprobs = self.sampler.compute_logprobs(logits)
            token_ids, logprobs, ranks = self.sampler.gather_logprobs(
4789
4790
                logprobs, num_prompt_logprobs, tgt_token_ids
            )
4791
4792

            # Transfer GPU->CPU async.
4793
4794
            chunk_slice = slice(start_idx, start_idx + num_logits)
            logprobs_tensors.logprob_token_ids[chunk_slice].copy_(
4795
4796
4797
                token_ids, non_blocking=True
            )
            logprobs_tensors.logprobs[chunk_slice].copy_(logprobs, non_blocking=True)
4798
            logprobs_tensors.selected_token_ranks[chunk_slice].copy_(
4799
4800
                ranks, non_blocking=True
            )
4801
4802
4803
4804
4805

        # 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]
4806
            del in_progress_dict[req_id]
4807
4808

        # Must synchronize the non-blocking GPU->CPU transfers.
4809
        if prompt_logprobs_dict:
4810
            self._sync_device()
4811
4812
4813

        return prompt_logprobs_dict

4814
4815
    def _get_nans_in_logits(
        self,
4816
        logits: torch.Tensor | None,
4817
4818
4819
4820
4821
4822
4823
4824
4825
4826
4827
    ) -> 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])
4828
4829
4830
                    if num_nans_for_index is not None and req_index < logits.shape[0]
                    else 0
                )
4831
4832
4833
4834
            return num_nans_in_logits
        except IndexError:
            return {}

4835
    @contextmanager
4836
4837
4838
    def maybe_randomize_inputs(
        self, input_ids: torch.Tensor | None, inputs_embeds: torch.Tensor | None
    ):
4839
4840
4841
4842
        """
        Randomize input_ids if VLLM_RANDOMIZE_DP_DUMMY_INPUTS is set.
        This is to help balance expert-selection
         - during profile_run
4843
         - during DP rank dummy run
4844
        """
4845

4846
4847
4848
4849
        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
4850
        elif input_ids is not None:
4851
4852
4853
4854

            @functools.cache
            def rand_input_ids() -> torch.Tensor:
                return torch.randint_like(
4855
                    self.input_ids.gpu,
4856
4857
                    low=0,
                    high=self.model_config.get_vocab_size(),
4858
                )
4859

4860
            logger.debug_once("Randomizing dummy input_ids for DP Rank")
4861
            input_ids.copy_(rand_input_ids()[: input_ids.size(0)], non_blocking=True)
4862
4863
            yield
            input_ids.fill_(0)
4864
4865
4866
4867
4868
4869
4870
4871
4872
4873
4874
4875
4876
4877
4878
        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)
4879

4880
4881
4882
4883
4884
4885
    def _get_mm_dummy_batch(
        self,
        modality: str,
        max_items_per_batch: int,
    ) -> BatchedTensorInputs:
        """Dummy data for profiling and precompiling multimodal models."""
4886
4887
        assert self.mm_budget is not None

4888
4889
4890
        # 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,
4891
            mm_counts={modality: 1},
4892
            cache=self.mm_budget.cache,
4893
        )
4894
4895
4896
4897
4898
        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"
4899

4900
        dummy_mm_items = [dummy_mm_item] * max_items_per_batch
4901

4902
4903
4904
4905
4906
4907
4908
4909
        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,
            )
        )
4910

4911
4912
4913
4914
    @torch.inference_mode()
    def _dummy_run(
        self,
        num_tokens: int,
4915
        cudagraph_runtime_mode: CUDAGraphMode | None = None,
4916
4917
        force_attention: bool = False,
        uniform_decode: bool = False,
4918
        allow_microbatching: bool = True,
4919
4920
        skip_eplb: bool = False,
        is_profile: bool = False,
4921
        create_mixed_batch: bool = False,
4922
        remove_lora: bool = True,
4923
        activate_lora: bool = False,
Rémi Delacourt's avatar
Rémi Delacourt committed
4924
        is_graph_capturing: bool = False,
4925
    ) -> tuple[torch.Tensor, torch.Tensor]:
4926
4927
4928
4929
4930
4931
4932
        """
        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.
4933
                - if not set will determine the cudagraph mode based on using
4934
                    the self.cudagraph_dispatcher.
4935
4936
4937
4938
                - CUDAGraphMode.NONE: No cudagraph, for warm up and profile run
                - CUDAGraphMode.PIECEWISE: Piecewise cudagraph.
                - CUDAGraphMode.FULL: Full cudagraph, attention metadata is
                    needed.
4939
            force_attention: If True, always create attention metadata. Used to
4940
4941
4942
4943
                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.
4944
4945
            create_mixed_batch: If True, create a mixed batch with both decode
                (1 token) and prefill (multiple tokens) requests.
4946
            remove_lora: If False, dummy LoRAs are not destroyed after the run
4947
            activate_lora: If False, dummy_run is performed without LoRAs.
4948
        """
4949
4950
        mm_config = self.vllm_config.model_config.multimodal_config
        if mm_config and mm_config.mm_encoder_only:
4951
4952
4953
4954
            # 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([])

4955
4956
4957
4958
        assert (
            cudagraph_runtime_mode is None
            or cudagraph_runtime_mode.valid_runtime_modes()
        )
4959

4960
        # If cudagraph_mode.decode_mode() == FULL and
4961
        # cudagraph_mode.separate_routine(). This means that we are using
4962
4963
4964
4965
4966
4967
4968
4969
4970
4971
4972
        # 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.
4973
        max_query_len = self.uniform_decode_query_len if uniform_decode else num_tokens
4974

4975
4976
4977
4978
4979
        # 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
4980
4981
4982
4983
        if create_mixed_batch:
            assert not uniform_decode
            # Create mixed batch:
            # first half decode tokens, second half one prefill
4984
            num_decode_tokens = min(max_num_reqs - 1, num_tokens // 2)
4985
4986
4987
4988
            num_prefill_tokens = num_tokens - num_decode_tokens
            num_reqs = num_decode_tokens + 1

            # Create decode requests (1 token each) followed by prefill request
4989
            num_scheduled_tokens_list = [1] * num_decode_tokens + [num_prefill_tokens]
4990
4991
4992
            # Note: Overriding max_query_len to be the prefill tokens
            max_query_len = num_prefill_tokens
        elif uniform_decode:
4993
            assert not create_mixed_batch
4994
            num_reqs = min(max_num_reqs, cdiv(num_tokens, max_query_len))
4995
4996
            num_scheduled_tokens_list = [max_query_len] * num_reqs
            if num_tokens % max_query_len != 0:
4997
                num_scheduled_tokens_list[-1] = num_tokens % max_query_len
4998
4999
5000
5001
5002
5003
        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

5004
5005
        assert sum(num_scheduled_tokens_list) == num_tokens
        assert len(num_scheduled_tokens_list) == num_reqs
5006
        num_scheduled_tokens = np.array(num_scheduled_tokens_list, dtype=np.int32)
5007
5008
        num_tokens_unpadded = int(num_scheduled_tokens.sum())

5009
        num_sampled_tokens = np.ones(num_reqs, dtype=np.int32)
5010

5011
        _cudagraph_mode, batch_desc, should_ubatch, num_tokens_across_dp, _ = (
5012
5013
5014
5015
5016
5017
5018
5019
5020
5021
5022
5023
5024
5025
5026
5027
5028
5029
            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,
5030
            )
5031
        )
5032
5033
5034

        if cudagraph_runtime_mode is None:
            cudagraph_runtime_mode = _cudagraph_mode
5035
        else:
5036
5037
5038
5039
            assert cudagraph_runtime_mode == _cudagraph_mode, (
                f"Cudagraph runtime mode mismatch in dummy_run. "
                f"Expected {_cudagraph_mode}, but got {cudagraph_runtime_mode}."
            )
5040

5041
5042
5043
5044
        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
        )
5045
        ubatch_slices, ubatch_slices_padded = maybe_create_ubatch_slices(
5046
5047
5048
5049
5050
5051
5052
5053
5054
5055
            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,
5056
        )
5057

5058
        attn_metadata: PerLayerAttnMetadata | None = None
5059

5060
5061
5062
5063
5064
5065
5066
        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,
        )

5067
5068
        # If force_attention is True, we always capture attention. Otherwise,
        # it only happens for cudagraph_runtime_mode=FULL.
5069
        if force_attention or cudagraph_runtime_mode == CUDAGraphMode.FULL:
5070
5071
5072
5073
5074
5075
            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:
5076
5077
5078
5079
5080
                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
5081
            self.seq_lens.np[:num_reqs] = seq_lens
5082
5083
            self.seq_lens.np[num_reqs:] = 0
            self.seq_lens.copy_to_gpu()
5084

5085
5086
            cum_num_tokens, _ = self._get_cumsum_and_arange(num_scheduled_tokens)
            self.query_start_loc.np[1 : num_reqs + 1] = cum_num_tokens
5087
5088
            self.query_start_loc.copy_to_gpu()

5089
            pad_attn = cudagraph_runtime_mode == CUDAGraphMode.FULL
王敏's avatar
王敏 committed
5090
            attn_metadata, _, _, _ = self._build_attention_metadata(
5091
5092
5093
                num_tokens=num_tokens_unpadded,
                num_reqs=num_reqs_padded,
                max_query_len=max_query_len,
5094
                ubatch_slices=ubatch_slices_padded if pad_attn else ubatch_slices,
5095
                for_cudagraph_capture=is_graph_capturing,
5096
                slot_mappings=slot_mappings_by_group,
5097
            )
5098

5099
        with self.maybe_dummy_run_with_lora(
5100
5101
5102
5103
5104
            self.lora_config,
            num_scheduled_tokens,
            num_sampled_tokens,
            activate_lora,
            remove_lora,
5105
        ):
5106
            # Make sure padding doesn't exceed max_num_tokens
5107
            assert num_tokens_padded <= self.max_num_tokens
5108
            model_kwargs = self._init_model_kwargs()
5109
            if self.supports_mm_inputs and not self.model_config.is_encoder_decoder:
Patrick von Platen's avatar
Patrick von Platen committed
5110
5111
                input_ids, inputs_embeds = self._prepare_mm_inputs(num_tokens_padded)

5112
                model_kwargs = {
5113
                    **model_kwargs,
5114
5115
                    **self._dummy_mm_kwargs(num_reqs),
                }
5116
5117
            elif self.enable_prompt_embeds:
                input_ids = None
5118
                inputs_embeds = self.inputs_embeds.gpu[:num_tokens_padded]
5119
                model_kwargs = self._init_model_kwargs()
5120
            else:
5121
                input_ids = self.input_ids.gpu[:num_tokens_padded]
5122
                inputs_embeds = None
5123

guanyu1's avatar
guanyu1 committed
5124
            positions = self._get_positions(num_tokens_padded)
5125
5126
5127
5128
5129
5130
5131
5132
5133

            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,
5134
5135
5136
                            device=self.device,
                        )
                    )
5137
5138

                intermediate_tensors = self.sync_and_slice_intermediate_tensors(
5139
                    num_tokens_padded, None, False
5140
                )
5141

5142
            if ubatch_slices_padded is not None:
5143
5144
5145
                # Adjust values to reflect a single ubatch.
                # TODO(sage,lucas): this is cruft that should be addressed in
                #  the padding refactor.
5146
                num_tokens_padded = ubatch_slices_padded[0].num_tokens
5147
                if num_tokens_across_dp is not None:
5148
                    num_tokens_across_dp[:] = num_tokens_padded
5149

5150
            with (
5151
                self.maybe_randomize_inputs(input_ids, inputs_embeds),
5152
                set_forward_context(
5153
5154
                    attn_metadata,
                    self.vllm_config,
5155
                    num_tokens=num_tokens_padded,
5156
5157
                    num_tokens_across_dp=num_tokens_across_dp,
                    cudagraph_runtime_mode=cudagraph_runtime_mode,
5158
                    batch_descriptor=batch_desc,
5159
                    ubatch_slices=ubatch_slices_padded,
5160
                    slot_mapping=slot_mappings,
5161
5162
                    enable_lightly_cp=self.enable_lightly_cp and num_tokens_unpadded > self.lightly_cp_threshould,
                    enable_lightly_cplb=self.enable_lightly_cplb
5163
5164
                ),
            ):
5165
                outputs = self.model(
5166
5167
5168
5169
                    input_ids=input_ids,
                    positions=positions,
                    intermediate_tensors=intermediate_tensors,
                    inputs_embeds=inputs_embeds,
5170
                    **model_kwargs,
5171
                )
5172

5173
5174
5175
5176
            if self.use_aux_hidden_state_outputs:
                hidden_states, _ = outputs
            else:
                hidden_states = outputs
5177

5178
5179
5180
5181
            if self.speculative_config and (
                self.speculative_config.use_eagle()
                or self.speculative_config.uses_draft_model()
            ):
王敏's avatar
王敏 committed
5182
5183
5184
5185
5186
5187
5188
5189
5190
5191
5192
5193
5194
5195
5196
5197
5198
5199
5200
5201
5202
5203
5204
5205
5206
5207
5208
5209
5210
                #assert isinstance(self.drafter, EagleProposer | DraftModelProposer)
                if hasattr(self, "drafter") and isinstance(self.drafter, EagleProposer | DraftModelProposer):
                    assert self.speculative_config is not None
                    # 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.
                    use_cudagraphs = (
                        (
                            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

                    # 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,
                        is_graph_capturing=is_graph_capturing,
                        slot_mappings=slot_mappings,
5211
                    )
5212

5213
5214
5215
5216
5217
5218
5219
5220
5221
5222
5223
        # 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()

5224
5225
5226
5227
5228
5229
5230
5231
5232
5233
        # 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)

5234
        logit_indices = np.cumsum(num_scheduled_tokens) - 1
5235
5236
5237
5238
5239
5240
5241
5242
5243
        # logit_indices_device = torch.from_numpy(logit_indices).to(
        #     self.device, non_blocking=True
        # )
        logit_indices = logit_indices.tolist()
        logit_indices_device = async_tensor_h2d(
                    logit_indices,
                    dtype=torch.int32,
                    target_device=self.device,
                    pin_memory=True)
5244
        return hidden_states, hidden_states[logit_indices_device]
5245
5246
5247
5248
5249
5250

    @torch.inference_mode()
    def _dummy_sampler_run(
        self,
        hidden_states: torch.Tensor,
    ) -> torch.Tensor:
5251
5252
5253
        # The dummy hidden states may contain special values,
        # like `inf` or `nan`.
        # To avoid breaking the sampler, we use a random tensor here instead.
5254

5255
5256
        mm_config = self.vllm_config.model_config.multimodal_config
        if mm_config and mm_config.mm_encoder_only:
5257
5258
5259
            # MM Encoder only model no need to run sampler.
            return torch.tensor([])

5260
        hidden_states = torch.rand_like(hidden_states)
5261

5262
        logits = self.model.compute_logits(hidden_states)
5263
5264
        num_reqs = logits.size(0)

5265
        dummy_tensors = lambda v: torch.full((num_reqs,), v, device=self.device)
5266
5267
5268
5269
5270
5271
5272
5273
5274
5275
5276
5277
5278
5279
5280

        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)],
5281
            spec_token_ids=[[] for _ in range(num_reqs)],
5282
5283
            allowed_token_ids_mask=None,
            bad_words_token_ids={},
5284
            logitsprocs=LogitsProcessors(),
5285
        )
5286
        try:
5287
5288
5289
            sampler_output = self.sampler(
                logits=logits, sampling_metadata=dummy_metadata
            )
5290
        except RuntimeError as e:
5291
            if "out of memory" in str(e):
5292
5293
5294
5295
                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 "
5296
5297
                    "initializing the engine."
                ) from e
5298
5299
            else:
                raise e
5300
        if self.speculative_config:
5301
5302
            draft_token_ids = [[0] for _ in range(num_reqs)]
            dummy_spec_decode_metadata = SpecDecodeMetadata.make_dummy(
5303
5304
                draft_token_ids, self.device
            )
5305
5306

            num_tokens = sum(len(ids) for ids in draft_token_ids)
5307
5308
5309
            # draft_probs = torch.randn(
            #     num_tokens, logits.shape[-1], device=self.device,
            #     dtype=logits.dtype)
王敏's avatar
王敏 committed
5310
5311
5312
5313
5314
5315
5316
            
            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)
5317
                dummy_metadata.all_greedy = True
王敏's avatar
王敏 committed
5318

5319
5320
5321
5322
5323
            logits = torch.randn(
                num_tokens + num_reqs,
                logits.shape[-1],
                device=self.device,
                dtype=logits.dtype,
5324
            )
5325
5326
5327
            self.rejection_sampler(
                dummy_spec_decode_metadata,
                draft_probs,
5328
                logits,
5329
5330
                dummy_metadata,
            )
5331
        return sampler_output
5332

5333
    def _dummy_pooler_run_task(
5334
5335
        self,
        hidden_states: torch.Tensor,
5336
5337
        task: PoolingTask,
    ) -> PoolerOutput:
5338
5339
5340
5341
        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
5342
5343
5344
5345
        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
5346
5347
5348

        req_num_tokens = num_tokens // num_reqs

5349
        dummy_prompt_lens = torch.from_numpy(num_scheduled_tokens_np)
5350
5351
5352
        dummy_token_ids = torch.zeros(
            (num_reqs, req_num_tokens), dtype=torch.int32, device=self.device
        )
5353

5354
        model = cast(VllmModelForPooling, self.get_model())
5355
        dummy_pooling_params = PoolingParams(task=task)
5356
        dummy_pooling_params.verify(task=task, model_config=self.model_config)
5357
        to_update = model.pooler.get_pooling_updates(task)
5358
5359
        to_update.apply(dummy_pooling_params)

5360
        dummy_metadata = PoolingMetadata(
5361
5362
5363
            prompt_lens=dummy_prompt_lens,
            prompt_token_ids=dummy_token_ids,
            pooling_params=[dummy_pooling_params] * num_reqs,
5364
            pooling_states=[PoolingStates() for i in range(num_reqs)],
5365
        )
5366

5367
        dummy_metadata.build_pooling_cursor(
5368
            num_scheduled_tokens_np,
5369
5370
            seq_lens_cpu=dummy_prompt_lens,
            device=hidden_states.device,
5371
        )
5372

5373
        try:
5374
5375
5376
            return model.pooler(
                hidden_states=hidden_states, pooling_metadata=dummy_metadata
            )
5377
        except RuntimeError as e:
5378
            if "out of memory" in str(e):
5379
                raise RuntimeError(
5380
5381
5382
                    "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 "
5383
5384
                    "initializing the engine."
                ) from e
5385
5386
            else:
                raise e
5387
5388
5389
5390
5391
5392

    @torch.inference_mode()
    def _dummy_pooler_run(
        self,
        hidden_states: torch.Tensor,
    ) -> PoolerOutput:
5393
5394
        mm_config = self.vllm_config.model_config.multimodal_config
        if mm_config and mm_config.mm_encoder_only:
5395
5396
5397
            # MM Encoder only model not need to run pooler.
            return torch.tensor([])

5398
        # Find the task that has the largest output for subsequent steps
5399
5400
5401
        supported_pooling_tasks = self.get_supported_pooling_tasks()

        if not supported_pooling_tasks:
5402
5403
5404
5405
5406
5407
            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."
            )
5408

5409
        output_size = dict[PoolingTask, float]()
5410
        for task in supported_pooling_tasks:
5411
5412
            # Run a full batch with each task to ensure none of them OOMs
            output = self._dummy_pooler_run_task(hidden_states, task)
5413
            output_size[task] = sum(o.nbytes for o in output if o is not None)
5414
5415
5416
5417
            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)
5418

5419
    def profile_run(self) -> None:
5420
        # Profile with multimodal encoder & encoder cache.
5421
        if self.supports_mm_inputs:
5422
5423
            mm_config = self.model_config.multimodal_config
            if mm_config is not None and mm_config.skip_mm_profiling:
5424
                logger.info(
5425
                    "Skipping memory profiling for multimodal encoder and "
5426
5427
                    "encoder cache."
                )
5428
5429
5430
5431
5432
5433
5434
5435
            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.
5436
                    dummy_modality = mm_budget.get_modality_with_max_tokens()
5437
5438
5439
                    max_mm_items_per_batch = mm_budget.max_items_per_batch_by_modality[
                        dummy_modality
                    ]
5440
5441
5442
5443
5444
5445
5446
5447
5448

                    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,
                    )
5449

5450
5451
5452
5453
5454
                    # Create dummy batch of multimodal inputs.
                    batched_dummy_mm_inputs = self._get_mm_dummy_batch(
                        dummy_modality,
                        max_mm_items_per_batch,
                    )
5455

5456
                    # Run multimodal encoder.
5457
                    dummy_encoder_outputs = self.model.embed_multimodal(
5458
5459
                        **batched_dummy_mm_inputs
                    )
5460

5461
5462
5463
5464
                    sanity_check_mm_encoder_outputs(
                        dummy_encoder_outputs,
                        expected_num_items=max_mm_items_per_batch,
                    )
5465
5466
                    for i, output in enumerate(dummy_encoder_outputs):
                        self.encoder_cache[f"tmp_{i}"] = output
5467

5468
        # Add `is_profile` here to pre-allocate communication buffers
5469
5470
5471
        hidden_states, last_hidden_states = self._dummy_run(
            self.max_num_tokens, is_profile=True
        )
5472
        if get_pp_group().is_last_rank:
5473
5474
5475
5476
            if self.is_pooling_model:
                output = self._dummy_pooler_run(hidden_states)
            else:
                output = self._dummy_sampler_run(last_hidden_states)
5477
        else:
5478
            output = None
5479
        self._sync_device()
5480
        del hidden_states, output
5481
        self.encoder_cache.clear()
5482
        gc.collect()
5483

5484
    def capture_model(self) -> int:
5485
        if self.compilation_config.cudagraph_mode == CUDAGraphMode.NONE:
5486
            logger.warning(
5487
                "Skipping CUDA graph capture. To turn on CUDA graph capture, "
5488
5489
                "ensure `cudagraph_mode` was not manually set to `NONE`"
            )
5490
            return 0
5491

5492
5493
        compilation_counter.num_gpu_runner_capture_triggers += 1

5494
5495
        start_time = time.perf_counter()

5496
5497
5498
5499
5500
5501
5502
5503
5504
5505
5506
5507
5508
5509
        @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()
5510
                    gc.collect()
5511

5512
5513
5514
        # 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.
5515
        set_cudagraph_capturing_enabled(True)
5516
        with freeze_gc(), graph_capture(device=self.device):
5517
            start_free_gpu_memory = torch.cuda.mem_get_info()[0]
5518

5519
5520
5521
5522
            for (
                runtime_mode,
                batch_descs,
            ) in self.cudagraph_dispatcher.get_capture_descs():
5523
                self._capture_cudagraphs(
5524
5525
                    batch_descriptors=batch_descs,
                    cudagraph_runtime_mode=runtime_mode,
5526
                )
5527

5528
5529
            torch.cuda.synchronize()
            end_free_gpu_memory = torch.cuda.mem_get_info()[0]
5530
5531
5532
5533

        # 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
5534
        # we may do lazy capturing in future that still allows capturing
5535
5536
        # after here.
        set_cudagraph_capturing_enabled(False)
5537

5538
5539
5540
5541
        # Lock workspace to prevent resizing during execution.
        # Max workspace sizes should have been captured during warmup/profiling.
        lock_workspace()

5542
5543
5544
5545
        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.
5546
        logger.info_once(
5547
5548
5549
            "Graph capturing finished in %.0f secs, took %.2f GiB",
            elapsed_time,
            cuda_graph_size / (1 << 30),
5550
            scope="local",
5551
        )
5552
        return cuda_graph_size
5553

5554
5555
    def _capture_cudagraphs(
        self,
5556
        batch_descriptors: list[BatchDescriptor],
5557
5558
5559
5560
5561
5562
        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}"
5563

5564
5565
5566
5567
5568
5569
5570
5571
5572
5573
5574
5575
5576
5577
        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,
        )

5578
5579
        # Only rank 0 should print progress bar during capture
        if is_global_first_rank():
5580
5581
            batch_descriptors = tqdm(
                batch_descriptors,
5582
5583
5584
                disable=not self.load_config.use_tqdm_on_load,
                desc="Capturing CUDA graphs ({}, {})".format(
                    "decode" if uniform_decode else "mixed prefill-decode",
5585
5586
5587
                    cudagraph_runtime_mode.name,
                ),
            )
5588

5589
        # We skip EPLB here since we don't want to record dummy metrics
5590
5591
5592
5593
        for batch_desc in batch_descriptors:
            num_tokens = batch_desc.num_tokens
            activate_lora = batch_desc.has_lora

5594
            # We currently only capture ubatched graphs when its a FULL
5595
5596
5597
            # 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
5598
            allow_microbatching = (
5599
                self.parallel_config.use_ubatching
5600
5601
                and cudagraph_runtime_mode == CUDAGraphMode.FULL
                and uniform_decode
5602
5603
5604
5605
5606
                and check_ubatch_thresholds(
                    config=self.vllm_config.parallel_config,
                    num_tokens=num_tokens,
                    uniform_decode=uniform_decode,
                )
5607
            )
5608

5609
5610
            for _ in range(self.compilation_config.cudagraph_num_of_warmups):
                # Use CUDAGraphRuntimeStyle.NONE (default) for warmup.
5611
                # But be careful, warm up with `NONE` is orthogonal to
5612
5613
5614
                # 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.
5615
                dummy_run(
5616
5617
5618
                    num_tokens,
                    cudagraph_runtime_mode=CUDAGraphMode.NONE,
                    allow_microbatching=allow_microbatching,
5619
                    activate_lora=activate_lora,
5620
                )
5621
5622
5623

            # Capture run
            dummy_run(
5624
5625
5626
                num_tokens,
                cudagraph_runtime_mode=cudagraph_runtime_mode,
                allow_microbatching=allow_microbatching,
5627
                activate_lora=activate_lora,
Rémi Delacourt's avatar
Rémi Delacourt committed
5628
                is_graph_capturing=True,
5629
            )
5630
        self.maybe_remove_all_loras(self.lora_config)
5631

5632
5633
5634
5635
    def initialize_attn_backend(self, kv_cache_config: KVCacheConfig) -> None:
        """
        Initialize the attention backends and attention metadata builders.
        """
5636
        assert len(self.attn_groups) == 0, "Attention backends are already initialized"
5637

5638
5639
5640
5641
5642
5643
        class AttentionGroupKey(NamedTuple):
            attn_backend: type[AttentionBackend]
            kv_cache_spec: KVCacheSpec

        def get_attn_backends_for_group(
            kv_cache_group_spec: KVCacheGroupSpec,
5644
        ) -> tuple[dict[AttentionGroupKey, list[str]], set[type[AttentionBackend]]]:
5645
            layer_type = cast(type[Any], AttentionLayerBase)
5646
            layers = get_layers_from_vllm_config(
5647
                self.vllm_config, layer_type, kv_cache_group_spec.layer_names
5648
            )
5649
5650
            attn_backends = {}
            attn_backend_layers = defaultdict(list)
5651
            # Dedupe based on full class name; this is a bit safer than
5652
5653
5654
5655
            # 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.
5656
            for layer_name in kv_cache_group_spec.layer_names:
5657
                attn_backend = layers[layer_name].get_attn_backend()
5658
5659
5660
5661

                if layer_name in self.kv_sharing_fast_prefill_eligible_layers:
                    attn_backend = create_fast_prefill_custom_backend(
                        "FastPrefill",
5662
                        attn_backend,  # type: ignore[arg-type]
5663
5664
                    )

5665
5666
5667
                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):
5668
                    layer_kv_cache_spec = layer_kv_cache_spec.kv_cache_specs[layer_name]
5669
                key = (full_cls_name, layer_kv_cache_spec)
5670
5671
5672
                attn_backends[key] = AttentionGroupKey(
                    attn_backend, layer_kv_cache_spec
                )
5673
                attn_backend_layers[key].append(layer_name)
5674
5675
5676
5677
            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()),
            )
5678
5679

        def create_attn_groups(
5680
            attn_backends_map: dict[AttentionGroupKey, list[str]],
5681
            kv_cache_group_id: int,
5682
5683
        ) -> list[AttentionGroup]:
            attn_groups: list[AttentionGroup] = []
5684
            for (attn_backend, kv_cache_spec), layer_names in attn_backends_map.items():
5685
                attn_group = AttentionGroup(
5686
                    attn_backend,
5687
                    layer_names,
5688
                    kv_cache_spec,
5689
                    kv_cache_group_id,
5690
                )
5691

5692
5693
5694
                attn_groups.append(attn_group)
            return attn_groups

5695
        attention_backend_maps = []
5696
        attention_backend_list = []
5697
        for kv_cache_group_spec in kv_cache_config.kv_cache_groups:
5698
            attn_backends = get_attn_backends_for_group(kv_cache_group_spec)
5699
            attention_backend_maps.append(attn_backends[0])
5700
            attention_backend_list.append(attn_backends[1])
5701
5702

        # Resolve cudagraph_mode before actually initialize metadata_builders
5703
5704
5705
        self._check_and_update_cudagraph_mode(
            attention_backend_list, kv_cache_config.kv_cache_groups
        )
5706

5707
5708
5709
        # Check if attention backend supports PCP&DCP and related features.
        check_attention_cp_compatibility(self.vllm_config)

5710
5711
        for i, attn_backend_map in enumerate(attention_backend_maps):
            self.attn_groups.append(create_attn_groups(attn_backend_map, i))
5712

5713
5714
5715
5716
5717
5718
5719
5720
5721
5722
5723
5724
5725
5726
5727
    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
5728
5729
                    if not self.parallel_config.use_ubatching
                    else self.parallel_config.num_ubatches,
5730
                )
co63oc's avatar
co63oc committed
5731
        # Calculate reorder batch threshold (if needed)
5732
5733
        # Note (tdoublep): do this *after* constructing builders,
        # because some of them change the threshold at init time.
5734
5735
        self.calculate_reorder_batch_threshold()

5736
    def _check_and_update_cudagraph_mode(
5737
5738
5739
        self,
        attention_backends: list[set[type[AttentionBackend]]],
        kv_cache_groups: list[KVCacheGroupSpec],
5740
    ) -> None:
5741
        """
5742
        Resolve the cudagraph_mode when there are multiple attention
5743
        groups with potential conflicting CUDA graph support.
5744
5745
5746
        Then initialize the cudagraph_dispatcher based on the resolved
        cudagraph_mode.
        """
5747
        min_cg_support = AttentionCGSupport.ALWAYS
5748
        min_cg_backend_name = None
5749

5750
5751
5752
5753
5754
        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()
5755

5756
5757
5758
5759
5760
5761
                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__
5762
5763
        # Flexible resolve the cudagraph mode
        cudagraph_mode = self.compilation_config.cudagraph_mode
5764
        assert cudagraph_mode is not None
5765
        # check cudagraph for mixed batch is supported
5766
5767
5768
5769
5770
5771
        if (
            cudagraph_mode.mixed_mode() == CUDAGraphMode.FULL
            and min_cg_support != AttentionCGSupport.ALWAYS
        ):
            msg = (
                f"CUDAGraphMode.{cudagraph_mode.name} is not supported "
5772
                f"with {min_cg_backend_name} backend (support: "
5773
5774
                f"{min_cg_support})"
            )
5775
5776
            if min_cg_support == AttentionCGSupport.NEVER:
                # if not supported any full cudagraphs, just raise it.
5777
5778
                msg += (
                    "; please try cudagraph_mode=PIECEWISE, and "
5779
                    "make sure compilation mode is VLLM_COMPILE"
5780
                )
5781
5782
5783
5784
5785
                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"
5786
                cudagraph_mode = self.compilation_config.cudagraph_mode = (
5787
                    CUDAGraphMode.FULL_AND_PIECEWISE
5788
                )
5789
5790
            else:
                msg += "; setting cudagraph_mode=FULL_DECODE_ONLY"
5791
                cudagraph_mode = self.compilation_config.cudagraph_mode = (
5792
                    CUDAGraphMode.FULL_DECODE_ONLY
5793
                )
5794
5795
            logger.warning(msg)

5796
        # check that if we are doing decode full-cudagraphs it is supported
5797
5798
5799
5800
        if not envs.VLLM_USE_PIECEWISE:
            if (
                cudagraph_mode.decode_mode() == CUDAGraphMode.FULL
                and min_cg_support == AttentionCGSupport.NEVER
5801
            ):
5802
5803
5804
5805
                msg = (
                    f"CUDAGraphMode.{cudagraph_mode.name} is not supported "
                    f"with {min_cg_backend_name} backend (support: "
                    f"{min_cg_support})"
5806
                )
5807
5808
5809
5810
5811
5812
5813
5814
5815
5816
5817
5818
5819
5820
5821
5822
5823
5824
5825
5826
                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)
5827

5828
5829
        # check that if we are doing spec-decode + decode full-cudagraphs it is
        # supported
5830
5831
5832
5833
5834
5835
5836
5837
        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 "
5838
                f"{min_cg_backend_name} (support: {min_cg_support})"
5839
            )
5840
5841
            if self.compilation_config.splitting_ops_contain_attention():
                msg += "; setting cudagraph_mode=PIECEWISE"
5842
                cudagraph_mode = self.compilation_config.cudagraph_mode = (
5843
                    CUDAGraphMode.PIECEWISE
5844
                )
5845
5846
            else:
                msg += "; setting cudagraph_mode=NONE"
5847
                cudagraph_mode = self.compilation_config.cudagraph_mode = (
5848
                    CUDAGraphMode.NONE
5849
                )
5850
5851
5852
5853
            logger.warning(msg)

        # double check that we can support full cudagraph if they are requested
        # even after automatic downgrades
5854
5855
5856
5857
5858
5859
        if (
            cudagraph_mode.has_full_cudagraphs()
            and min_cg_support == AttentionCGSupport.NEVER
        ):
            raise ValueError(
                f"CUDAGraphMode.{cudagraph_mode.name} is not "
5860
                f"supported with {min_cg_backend_name} backend ("
5861
5862
                f"support:{min_cg_support}) "
                "; please try cudagraph_mode=PIECEWISE, "
5863
                "and make sure compilation mode is VLLM_COMPILE"
5864
            )
5865

5866
5867
5868
5869
        # 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
5870
        # Will be removed in the near future when we have separate cudagraph capture
5871
5872
5873
5874
5875
5876
5877
5878
5879
        # 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
            )
5880
5881
5882
5883
            capture_sizes = self.compilation_config.cudagraph_capture_sizes
            self.cudagraph_batch_sizes = (
                capture_sizes if capture_sizes is not None else []
            )
5884

5885
5886
        # Trigger cudagraph dispatching keys initialization after
        # resolved cudagraph mode.
5887
        self.compilation_config.cudagraph_mode = cudagraph_mode
5888
        self.cudagraph_dispatcher.initialize_cudagraph_keys(
5889
            cudagraph_mode, self.uniform_decode_query_len
5890
        )
5891

5892
        # Initialize eagle's cudagraph dispatcher if using eagle spec decode.
王敏's avatar
王敏 committed
5893
        if self.speculative_config and self.speculative_config.use_eagle() and hasattr(self, "drafter"):
5894
5895
5896
            assert isinstance(self.drafter, EagleProposer)
            self.drafter.initialize_cudagraph_keys(cudagraph_mode)

5897
5898
    def calculate_reorder_batch_threshold(self) -> None:
        """
5899
5900
5901
5902
        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.
5903
        """
5904
5905
        min_none_high = lambda a, b: a if b is None else b if a is None else min(a, b)

5906
        reorder_batch_thresholds: list[int | None] = [
5907
5908
5909
            group.get_metadata_builder().reorder_batch_threshold
            for group in self._attn_group_iterator()
        ]
5910
5911
5912
5913
5914
        # 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
5915
        self.reorder_batch_threshold = reduce(min_none_high, reorder_batch_thresholds)  # type: ignore[assignment]
5916

5917
5918
5919
    @staticmethod
    def select_common_block_size(
        kv_manager_block_size: int, attn_groups: list[AttentionGroup]
5920
5921
    ) -> int:
        """
5922
5923
5924
5925
5926
        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.
5927

5928
5929
5930
5931
5932
        Args:
            kv_manager_block_size: Block size of KV cache
            attn_groups: List of attention groups

        Returns:
5933
            The selected block size
5934
5935

        Raises:
5936
            ValueError: If no valid block size found
5937
5938
        """

王敏's avatar
王敏 committed
5939
5940
5941
5942
        #exclude indexer backend
        def _participates_in_block_size_selection(backend: type[AttentionBackend]) -> bool:
            return not getattr(backend, "exclude_from_block_size_selection", False)

5943
5944
5945
5946
5947
5948
5949
5950
        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
5951
                for supported_size in backend.get_supported_kernel_block_sizes():
5952
5953
5954
5955
5956
5957
5958
5959
5960
5961
5962
5963
                    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

zhuwenwen's avatar
zhuwenwen committed
5964
5965
5966
5967
        all_backends = [group.backend for group in attn_groups]
        backends = [
            b for b in all_backends
            if _participates_in_block_size_selection(b)
5968
            ]
zhuwenwen's avatar
zhuwenwen committed
5969

5970
5971
5972
5973
5974
5975
5976
5977
5978
5979
5980
5981
5982
5983
5984
5985
5986

        # 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
5987
            for supported_size in backend.get_supported_kernel_block_sizes()
5988
5989
            if isinstance(supported_size, int)
        )
5990

5991
5992
5993
5994
5995
5996
        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}. ")
5997

5998
5999
6000
    def may_reinitialize_input_batch(
        self, kv_cache_config: KVCacheConfig, kernel_block_sizes: list[int]
    ) -> None:
6001
6002
6003
6004
6005
6006
6007
        """
        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.
6008
            kernel_block_sizes: The kernel block sizes for each KV cache group.
6009
6010
6011
6012
        """
        block_sizes = [
            kv_cache_group.kv_cache_spec.block_size
            for kv_cache_group in kv_cache_config.kv_cache_groups
6013
            if not isinstance(kv_cache_group.kv_cache_spec, EncoderOnlyAttentionSpec)
6014
        ]
6015
6016
6017
6018
6019
6020
6021
6022
6023
6024
6025
6026
6027
6028
6029
6030
6031
6032
        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)
6033
6034
6035
6036

        if block_sizes != [self.cache_config.block_size] or kernel_block_sizes != [
            self.cache_config.block_size
        ]:
6037
6038
6039
            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
6040
6041
                "for more details."
            )
6042
6043
            self.input_batch = InputBatch(
                max_num_reqs=self.max_num_reqs,
6044
                max_model_len=max_model_len,
6045
6046
6047
6048
6049
                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,
6050
                kernel_block_sizes=kernel_block_sizes,
6051
                max_num_blocks_per_req=max_num_blocks,
6052
                is_spec_decode=bool(self.vllm_config.speculative_config),
6053
                logitsprocs=self.input_batch.logitsprocs,
6054
                logitsprocs_need_output_token_ids=self.input_batch.logitsprocs_need_output_token_ids,
6055
                is_pooling_model=self.is_pooling_model,
6056
6057
            )

6058
    def _allocate_kv_cache_tensors(
6059
6060
        self, kv_cache_config: KVCacheConfig
    ) -> dict[str, torch.Tensor]:
6061
        """
6062
6063
6064
        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.

6065
        Args:
6066
            kv_cache_config: The KV cache config
6067
        Returns:
6068
            dict[str, torch.Tensor]: A map between layer names to their
6069
            corresponding memory buffer for KV cache.
6070
        """
6071
6072
        kv_cache_raw_tensors: dict[str, torch.Tensor] = {}
        for kv_cache_tensor in kv_cache_config.kv_cache_tensors:
6073
6074
6075
            tensor = torch.zeros(
                kv_cache_tensor.size, dtype=torch.int8, device=self.device
            )
6076
6077
6078
6079
6080
            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:
6081
6082
6083
6084
            for layer_name in group.layer_names:
                if layer_name in self.runner_only_attn_layers:
                    continue
                layer_names.add(layer_name)
6085
6086
6087
        assert layer_names == set(kv_cache_raw_tensors.keys()), (
            "Some layers are not correctly initialized"
        )
6088
6089
        return kv_cache_raw_tensors

6090
6091
6092
    def _attn_group_iterator(self) -> Iterator[AttentionGroup]:
        return itertools.chain.from_iterable(self.attn_groups)

6093
    def _kv_cache_spec_attn_group_iterator(self) -> Iterator[AttentionGroup]:
6094
6095
        if not self.kv_cache_config.kv_cache_groups:
            return
6096
6097
        for attn_groups in self.attn_groups:
            yield from attn_groups
6098

6099
6100
6101
6102
6103
6104
6105
6106
6107
6108
6109
6110
6111
6112
6113
    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 = []
6114
        for kv_cache_gid, kv_cache_group in enumerate(kv_cache_config.kv_cache_groups):
6115
6116
6117
6118
6119
6120
            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):
6121
                continue
6122
            elif isinstance(kv_cache_spec, AttentionSpec):
6123
6124
6125
                # This is an attention backend that supports virtual
                # block splitting. Get the supported block sizes from
                # all backends in the group.
6126
                attn_groups = self.attn_groups[kv_cache_gid]
6127
                kv_manager_block_size = kv_cache_group.kv_cache_spec.block_size
6128
                selected_kernel_size = self.select_common_block_size(
6129
6130
6131
                    kv_manager_block_size, attn_groups
                )
                kernel_block_sizes.append(selected_kernel_size)
6132
            elif isinstance(kv_cache_spec, MambaSpec):
6133
6134
                # This is likely Mamba or other non-attention cache,
                # no splitting.
6135
                kernel_block_sizes.append(kv_cache_spec.block_size)
6136
6137
6138
6139
6140
6141
            else:
                raise NotImplementedError(
                    f"unknown kv cache spec {kv_cache_group.kv_cache_spec}"
                )
        return kernel_block_sizes

6142
6143
6144
6145
    def _reshape_kv_cache_tensors(
        self,
        kv_cache_config: KVCacheConfig,
        kv_cache_raw_tensors: dict[str, torch.Tensor],
6146
        kernel_block_sizes: list[int],
6147
    ) -> dict[str, torch.Tensor]:
6148
        """
6149
        Reshape the KV cache tensors to the desired shape and dtype.
6150

6151
        Args:
6152
6153
            kv_cache_config: The KV cache config
            kv_cache_raw_tensors: The KV cache buffer of each layer, with
6154
                correct size but uninitialized shape.
6155
            kernel_block_sizes: The kernel block sizes for each KV cache group.
6156
        Returns:
6157
            Dict[str, torch.Tensor]: A map between layer names to their
6158
6159
            corresponding memory buffer for KV cache.
        """
6160
        kv_caches: dict[str, torch.Tensor] = {}
6161
        has_attn, has_mamba = False, False
6162
6163
        for group in self._kv_cache_spec_attn_group_iterator():
            kv_cache_spec = group.kv_cache_spec
6164
            attn_backend = group.backend
6165
6166
6167
6168
            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]
6169
            for layer_name in group.layer_names:
6170
6171
                if layer_name in self.runner_only_attn_layers:
                    continue
6172
6173
                raw_tensor = kv_cache_raw_tensors[layer_name]
                assert raw_tensor.numel() % kv_cache_spec.page_size_bytes == 0
6174
                num_blocks = raw_tensor.numel() // kv_cache_spec.page_size_bytes
6175
                if isinstance(kv_cache_spec, AttentionSpec):
6176
                    has_attn = True
6177
6178
                    num_blocks_per_kv_block = (
                        kv_cache_spec.block_size // kernel_block_size
6179
6180
6181
                    )
                    kernel_num_blocks = num_blocks * num_blocks_per_kv_block

6182
                    if envs.VLLM_USE_FLASH_ATTN_PA and not self.vllm_config.model_config.use_mla:
6183
                        key_cache_shape, value_cache_shape = attn_backend.get_kv_cache_shape(
6184
6185
                            kernel_num_blocks,
                            kernel_block_size,
6186
6187
                            kv_cache_spec.num_kv_heads,
                            kv_cache_spec.head_size,
6188
6189
                            cache_dtype_str=self.cache_config.cache_dtype,
                        )
6190
6191
6192
                        dtype = kv_cache_spec.dtype
                        try:
                            key_stride_order, value_stride_order = attn_backend.get_kv_cache_stride_order()
6193
6194
                            assert len(key_stride_order) == len(key_stride_order)
                            assert len(value_stride_order) == len(value_cache_shape)
6195
                        except (AttributeError, NotImplementedError):
6196
6197
                            key_stride_order = tuple(range(len(key_cache_shape)))
                            value_stride_order = tuple(range(len(value_cache_shape)))
6198
6199
6200
6201
6202
                        # 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.
6203
6204
6205
6206
                        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)
6207
6208
6209
6210
6211
6212
6213
6214
6215
6216
6217
6218
6219
6220
6221
6222
6223
6224
6225
6226
6227
6228
6229
6230
6231
6232
6233
                        # 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(
6234
6235
                            kernel_num_blocks,
                            kernel_block_size,
6236
6237
                            kv_cache_spec.num_kv_heads,
                            kv_cache_spec.head_size,
6238
6239
                            cache_dtype_str=self.cache_config.cache_dtype,
                        )
6240
6241
                        dtype = kv_cache_spec.dtype
                        try:
6242
6243
                            kv_cache_stride_order = attn_backend.get_kv_cache_stride_order()
                            assert len(kv_cache_stride_order) == len(kv_cache_shape)
6244
                        except (AttributeError, NotImplementedError):
6245
                            kv_cache_stride_order = tuple(range(len(kv_cache_shape)))
6246
6247
6248
6249
6250
                        # 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.
6251
6252
6253
                        kv_cache_shape = tuple(
                            kv_cache_shape[i] for i in kv_cache_stride_order
                        )
6254
6255
6256
6257
6258
                        # Maintain original KV shape view.
                        inv_order = [
                            kv_cache_stride_order.index(i)
                            for i in range(len(kv_cache_stride_order))
                        ]
6259
6260
6261
6262
6263
6264
                        kv_caches[layer_name] = (
                            kv_cache_raw_tensors[layer_name]
                            .view(dtype)
                            .view(kv_cache_shape)
                            .permute(*inv_order)
                        )
6265

Chen Zhang's avatar
Chen Zhang committed
6266
                elif isinstance(kv_cache_spec, MambaSpec):
6267
                    has_mamba = True
Chen Zhang's avatar
Chen Zhang committed
6268
6269
                    raw_tensor = kv_cache_raw_tensors[layer_name]
                    state_tensors = []
6270
                    storage_offset_bytes = 0
6271
                    for shape, dtype in zip(kv_cache_spec.shapes, kv_cache_spec.dtypes):
6272
6273
                        dtype_size = get_dtype_size(dtype)
                        num_element_per_page = (
6274
6275
                            kv_cache_spec.page_size_bytes // dtype_size
                        )
Chen Zhang's avatar
Chen Zhang committed
6276
                        target_shape = (num_blocks, *shape)
6277
6278
                        stride = torch.empty(target_shape).stride()
                        target_stride = (num_element_per_page, *stride[1:])
6279
                        assert storage_offset_bytes % dtype_size == 0
6280
6281
6282
6283
                        tensor = torch.as_strided(
                            raw_tensor.view(dtype),
                            size=target_shape,
                            stride=target_stride,
6284
                            storage_offset=storage_offset_bytes // dtype_size,
6285
                        )
Chen Zhang's avatar
Chen Zhang committed
6286
                        state_tensors.append(tensor)
6287
                        storage_offset_bytes += stride[0] * dtype_size
6288
6289

                    kv_caches[layer_name] = state_tensors
6290
                else:
6291
                    raise NotImplementedError
6292
6293

        if has_attn and has_mamba:
6294
            self._update_hybrid_attention_mamba_layout(kv_caches)
6295

6296
6297
        return kv_caches

6298
    def _update_hybrid_attention_mamba_layout(
6299
        self, kv_caches: dict[str, torch.Tensor]
6300
    ) -> None:
6301
        """
6302
6303
        Update the layout of attention layers from (2, num_blocks, ...) to
        (num_blocks, 2, ...).
6304
6305

        Args:
6306
            kv_caches: The KV cache buffer of each layer.
6307
6308
        """

6309
6310
        for group in self._kv_cache_spec_attn_group_iterator():
            kv_cache_spec = group.kv_cache_spec
6311
            for layer_name in group.layer_names:
6312
                kv_cache = kv_caches[layer_name]
6313
6314
6315
6316
                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 "
6317
                        f"a tensor of shape {kv_cache.shape}"
6318
                    )
6319
                    hidden_size = kv_cache.shape[2:].numel()
6320
6321
6322
6323
                    kv_cache.as_strided_(
                        size=kv_cache.shape,
                        stride=(hidden_size, 2 * hidden_size, *kv_cache.stride()[2:]),
                    )
6324

6325
    def initialize_kv_cache_tensors(
6326
        self, kv_cache_config: KVCacheConfig, kernel_block_sizes: list[int]
6327
    ) -> dict[str, torch.Tensor]:
6328
6329
6330
6331
6332
        """
        Initialize the memory buffer for KV cache.

        Args:
            kv_cache_config: The KV cache config
6333
6334
            kernel_block_sizes: The kernel block sizes for each KV cache group.

6335
        Returns:
6336
            Dict[str, torch.Tensor]: A map between layer names to their
6337
6338
            corresponding memory buffer for KV cache.
        """
6339
6340
6341
6342
6343
6344
6345
6346
6347
6348
6349
6350
6351
6352
6353
6354
6355
6356
6357
6358
6359
6360
6361
6362

        # 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
            )
6363

6364
        # Set up cross-layer KV cache sharing
6365
6366
        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)
6367
6368
            kv_caches[layer_name] = kv_caches[target_layer_name]

6369
6370
6371
6372
6373
6374
6375
6376
6377
        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,
        )
6378
6379
6380
        return kv_caches

    def maybe_add_kv_sharing_layers_to_kv_cache_groups(
6381
6382
        self, kv_cache_config: KVCacheConfig
    ) -> None:
6383
6384
6385
6386
6387
6388
6389
6390
6391
6392
6393
6394
6395
6396
6397
6398
6399
6400
        """
        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.
6401
            attn_layers = get_layers_from_vllm_config(self.vllm_config, Attention)
6402
6403
            for layer_name in reversed(attn_layers):
                if layer_name in self.shared_kv_cache_layers:
6404
                    self.kv_sharing_fast_prefill_eligible_layers.add(layer_name)
6405
6406
                else:
                    break
6407

6408
6409
6410
6411
6412
6413
6414
    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
        """
6415
        kv_cache_config = deepcopy(kv_cache_config)
6416
        self.kv_cache_config = kv_cache_config
6417
        self.may_add_encoder_only_layers_to_kv_cache_config()
6418
        self.maybe_add_kv_sharing_layers_to_kv_cache_groups(kv_cache_config)
6419
        self.initialize_attn_backend(kv_cache_config)
6420
6421
6422
6423
6424
6425
        # 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)
6426
6427
6428
6429

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

6430
        # Reinitialize need to after initialize_attn_backend
6431
6432
6433
6434
        self.may_reinitialize_input_batch(kv_cache_config, kernel_block_sizes)
        kv_caches = self.initialize_kv_cache_tensors(
            kv_cache_config, kernel_block_sizes
        )
6435

6436
6437
6438
6439
        if self.speculative_config and (
            self.speculative_config.use_eagle()
            or self.speculative_config.uses_draft_model()
        ):
王敏's avatar
王敏 committed
6440
6441
6442
6443
6444
            #assert isinstance(self.drafter, EagleProposer | DraftModelProposer)
            if hasattr(self, "drafter") and isinstance(self.drafter, EagleProposer | DraftModelProposer):
                # validate all draft model layers belong to the same kv cache
                # group
                self.drafter.validate_same_kv_cache_group(kv_cache_config)
6445

Robert Shaw's avatar
Robert Shaw committed
6446
        if has_kv_transfer_group():
6447
            kv_transfer_group = get_kv_transfer_group()
6448
6449
6450
6451
6452
6453
6454
            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)
6455
            kv_transfer_group.set_host_xfer_buffer_ops(copy_kv_blocks)
Robert Shaw's avatar
Robert Shaw committed
6456

6457
6458
6459
6460
6461
6462
6463
6464
6465
6466
6467
6468
6469
6470
6471
6472
6473
        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,
6474
            vllm_config=self.vllm_config,
6475
6476
        )

6477
6478
6479
6480
6481
    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
6482
        encoder_only_attn_specs: dict[AttentionSpec, list[str]] = defaultdict(list)
6483
6484
6485
        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:
6486
                attn_spec: AttentionSpec = EncoderOnlyAttentionSpec(
6487
6488
6489
                    block_size=block_size,
                    num_kv_heads=attn_module.num_kv_heads,
                    head_size=attn_module.head_size,
6490
6491
                    dtype=self.kv_cache_dtype,
                )
6492
6493
6494
                encoder_only_attn_specs[attn_spec].append(layer_name)
                self.runner_only_attn_layers.add(layer_name)
        if len(encoder_only_attn_specs) > 0:
6495
6496
6497
            assert len(encoder_only_attn_specs) == 1, (
                "Only support one encoder-only attention spec now"
            )
6498
6499
            spec, layer_names = encoder_only_attn_specs.popitem()
            self.kv_cache_config.kv_cache_groups.append(
6500
6501
                KVCacheGroupSpec(layer_names=layer_names, kv_cache_spec=spec)
            )
6502

6503
    def get_kv_cache_spec(self) -> dict[str, KVCacheSpec]:
6504
        """
6505
        Generates the KVCacheSpec by parsing the kv cache format from each
6506
6507
        Attention module in the static forward context.
        Returns:
6508
            KVCacheSpec: A dictionary mapping layer names to their KV cache
6509
6510
            format. Layers that do not need KV cache are not included.
        """
6511
6512
        if has_ec_transfer() and get_ec_transfer().is_producer:
            return {}
6513
        kv_cache_spec: dict[str, KVCacheSpec] = {}
6514
6515
        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
6516
        for layer_name, attn_module in attn_layers.items():
6517
6518
6519
            if isinstance(attn_module, Attention) and (
                kv_tgt_layer := attn_module.kv_sharing_target_layer_name
            ):
6520
6521
6522
6523
6524
6525
6526
6527
6528
                # 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
6529
6530
6531
            # 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
6532

6533
        return kv_cache_spec
6534

6535
6536
6537
6538
6539
6540
6541
6542
6543
    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.
6544
        pinned = self.sampled_token_ids_pinned_cpu[: sampled_token_ids.shape[0]]
6545
6546
6547
6548
        pinned.copy_(sampled_token_ids, non_blocking=True)
        self.transfer_event.record()
        self.transfer_event.synchronize()
        return pinned.tolist()
6549
6550
6551
6552
6553
6554
6555
6556
6557
6558
6559
6560
6561
6562
6563
6564
6565
6566
6567
6568
6569
6570
6571
6572
6573
6574
6575
6576
6577
6578
6579
6580
6581
6582
6583
6584
6585
6586
6587
6588
6589
6590
6591
6592
6593
6594
6595
6596
6597
6598
6599
6600
6601
6602
6603
6604
6605
6606
6607
6608
6609
6610
6611
6612
6613
6614
6615
6616
6617
6618
6619
6620
6621
6622
6623
6624

    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,
        }