gpu_model_runner.py 243 KB
Newer Older
1
# SPDX-License-Identifier: Apache-2.0
2
# SPDX-FileCopyrightText: Copyright contributors to the vLLM project
3

4
import functools
5
import gc
6
import itertools
7
import time
8
from collections import defaultdict
9
from collections.abc import Iterator, Sequence
10
from contextlib import contextmanager
11
from copy import copy, deepcopy
12
from functools import reduce
13
from itertools import product
14
from typing import TYPE_CHECKING, Any, NamedTuple, TypeAlias, cast
15
16
17
18
19

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

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

170
171
172
173
174
175
176
from .utils import (
    AttentionGroup,
    MultiModalBudget,
    add_kv_sharing_layers_to_kv_cache_groups,
    bind_kv_cache,
    sanity_check_mm_encoder_outputs,
)
177

178
if TYPE_CHECKING:
179
    from vllm.model_executor.model_loader.tensorizer import TensorizerConfig
180
    from vllm.v1.core.sched.output import GrammarOutput, SchedulerOutput
181
182
183

logger = init_logger(__name__)

184
185
AttnMetadataDict: TypeAlias = dict[str, AttentionMetadata]
# list when ubatching is enabled
186
PerLayerAttnMetadata: TypeAlias = list[AttnMetadataDict] | AttnMetadataDict
187

188

189
190
191
192
193
194
# Wrapper for ModelRunnerOutput to support overlapped execution.
class AsyncGPUModelRunnerOutput(AsyncModelRunnerOutput):
    def __init__(
        self,
        model_runner_output: ModelRunnerOutput,
        sampled_token_ids: torch.Tensor,
195
        logprobs_tensors: LogprobsTensors | None,
196
197
        invalid_req_indices: list[int],
        async_output_copy_stream: torch.cuda.Stream,
198
        vocab_size: int,
199
200
201
202
203
    ):
        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.
204
        self.async_copy_ready_event = torch.Event()
205
206
207
208

        # 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
209
        self.vocab_size = vocab_size
210
        self._logprobs_tensors = logprobs_tensors
211
212
213
214
215

        # 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)
216
            self.sampled_token_ids_cpu = self._sampled_token_ids.to(
217
218
                "cpu", non_blocking=True
            )
219
220
221
222
223
            self._logprobs_tensors_cpu = (
                self._logprobs_tensors.to_cpu_nonblocking()
                if self._logprobs_tensors
                else None
            )
224
            self.async_copy_ready_event.record()
225
226
227

    def get_output(self) -> ModelRunnerOutput:
        """Copy the device tensors to the host and return a ModelRunnerOutput.
228

229
230
        This function blocks until the copy is finished.
        """
231
        max_gen_len = self.sampled_token_ids_cpu.shape[-1]
232
        self.async_copy_ready_event.synchronize()
233

234
235
        # Release the device tensors once the copy has completed.
        del self._logprobs_tensors
236
        del self._sampled_token_ids
237
        if max_gen_len == 1:
238
            valid_sampled_token_ids = self.sampled_token_ids_cpu.tolist()
239
240
241
            for i in self._invalid_req_indices:
                valid_sampled_token_ids[i].clear()
            cu_num_tokens = None
242
        else:
243
            valid_sampled_token_ids, cu_num_tokens = RejectionSampler.parse_output(
244
245
                self.sampled_token_ids_cpu,
                self.vocab_size,
246
247
                self._invalid_req_indices,
                return_cu_num_tokens=self._logprobs_tensors_cpu is not None,
248
            )
249
250
251

        output = self._model_runner_output
        output.sampled_token_ids = valid_sampled_token_ids
252
        if self._logprobs_tensors_cpu:
253
            output.logprobs = self._logprobs_tensors_cpu.tolists(cu_num_tokens)
254
255
256
        return output


257
258
259
260
261
262
263
264
265
266
267
268
269
270
271
272
273
274
275
276
277
278
279
280
281
282
283
284
285
286
287
288
289
290
291
292
293
294
295
296
297
298
299
300
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
        self._finished_mask = finished_mask

        # 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)
            self._raw_pooler_output_cpu = json_map_leaves(
                lambda x: None if x is None else x.to("cpu", non_blocking=True),
                self._raw_pooler_output,
            )
            self.async_copy_ready_event.record()

    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

        self._model_runner_output.pooler_output = [
            out if include else None
            for out, include in zip(self._raw_pooler_output_cpu, self._finished_mask)
        ]
        return self._model_runner_output


301
302
303
304
305
306
307
308
309
310
311
class ExecuteModelState(NamedTuple):
    """Ephemeral cached state transferred between execute_model() and
    sample_tokens(), after execute_model() returns None."""

    scheduler_output: "SchedulerOutput"
    logits: torch.Tensor
    spec_decode_metadata: SpecDecodeMetadata | None
    spec_decode_common_attn_metadata: CommonAttentionMetadata | None
    hidden_states: torch.Tensor
    sample_hidden_states: torch.Tensor
    aux_hidden_states: list[torch.Tensor] | None
312
    ec_connector_output: ECConnectorOutput | None
313
    cudagraph_stats: CUDAGraphStat | None
314
315


316
317
318
class GPUModelRunner(
    LoRAModelRunnerMixin, KVConnectorModelRunnerMixin, ECConnectorModelRunnerMixin
):
319
320
    def __init__(
        self,
321
        vllm_config: VllmConfig,
322
        device: torch.device,
323
    ):
324
325
326
        self.vllm_config = vllm_config
        self.model_config = vllm_config.model_config
        self.cache_config = vllm_config.cache_config
327
        self.compilation_config = vllm_config.compilation_config
328
329
330
331
332
333
        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
334

335
        from vllm.model_executor.models.utils import set_cpu_offload_max_bytes
336
337

        set_cpu_offload_max_bytes(int(self.cache_config.cpu_offload_gb * 1024**3))
338

339
340
341
342
        model_config = self.model_config
        cache_config = self.cache_config
        scheduler_config = self.scheduler_config
        parallel_config = self.parallel_config
343
        self.device = device
344
345
        self.pin_memory = is_pin_memory_available()
        self.dtype = self.model_config.dtype
346

347
348
349
        self.kv_cache_dtype = kv_cache_dtype_str_to_dtype(
            cache_config.cache_dtype, self.model_config
        )
350

351
        self.is_pooling_model = model_config.runner_type == "pooling"
352
        self.enable_prompt_embeds = model_config.enable_prompt_embeds
353
        self.is_multimodal_raw_input_only_model = (
354
355
            model_config.is_multimodal_raw_input_only_model
        )
356
357
        # This will be overridden in load_model()
        self.is_multimodal_pruning_enabled = False
358
        self.max_model_len = model_config.max_model_len
359
360
361

        # Always set to false after the first forward pass
        self.calculate_kv_scales = self.cache_config.calculate_kv_scales
362
        self.dcp_world_size = self.parallel_config.decode_context_parallel_size
363
        self.dcp_rank = 0 if self.dcp_world_size <= 1 else get_dcp_group().rank_in_group
364
        self.max_num_tokens = scheduler_config.max_num_batched_tokens
365
        self.max_num_reqs = scheduler_config.max_num_seqs
366

367
368
369
370
371
        # 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 = (
372
            self.parallel_config.distributed_executor_backend == "external_launcher"
373
            and len(get_pp_group().ranks) > 1
374
        )
375

376
        # Model-related.
377
        self.num_query_heads = model_config.get_num_attention_heads(parallel_config)
378
        self.inputs_embeds_size = model_config.get_inputs_embeds_size()
379
        self.attention_chunk_size = model_config.attention_chunk_size
380
        # Only relevant for models using ALiBi (e.g, MPT)
381
        self.use_alibi = model_config.uses_alibi
382

383
        self.cascade_attn_enabled = not self.model_config.disable_cascade_attn
384
        self.is_mm_prefix_lm = self.model_config.is_mm_prefix_lm
385

386
        # Multi-modal data support
387
        self.mm_registry = MULTIMODAL_REGISTRY
388
        self.uses_mrope = model_config.uses_mrope
389
        self.uses_xdrope_dim = model_config.uses_xdrope_dim
390
        self.supports_mm_inputs = self.mm_registry.supports_multimodal_inputs(
391
            model_config
392
        )
393

394
395
396
        if self.model_config.is_encoder_decoder:
            # Maximum length of the encoder input, only for encoder-decoder
            # models.
397
            self.max_encoder_len = scheduler_config.max_num_encoder_input_tokens
398
399
400
        else:
            self.max_encoder_len = 0

401
        # Sampler
402
        self.sampler = Sampler(logprobs_mode=self.model_config.logprobs_mode)
403

404
        self.eplb_state: EplbState | None = None
405
406
407
408
409
410
        """
        State of the expert parallelism load balancer.

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

411
        # Lazy initializations
412
        # self.model: nn.Module  # Set after load_model
413
        # Initialize in initialize_kv_cache
414
        self.kv_caches: list[torch.Tensor] = []
415
416
417
        # Initialize in initialize_kv_cache_tensors
        self.cross_layers_kv_cache: torch.Tensor | None = None
        self.cross_layers_attn_backend: type[AttentionBackend] | None = None
418
419
        # indexes: [kv_cache_group_id][attn_group]
        self.attn_groups: list[list[AttentionGroup]] = []
420
421
        # self.kv_cache_config: KVCacheConfig

422
423
        # mm_hash ->  encoder_output
        self.encoder_cache: dict[str, torch.Tensor] = {}
424

425
        self.use_aux_hidden_state_outputs = False
426
427
428
429
430
        # 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:
431
432
433
            self.drafter: (
                NgramProposer | SuffixDecodingProposer | EagleProposer | MedusaProposer
            )
434
435
            if self.speculative_config.method == "ngram":
                self.drafter = NgramProposer(self.vllm_config)
436
437
            elif self.speculative_config.method == "suffix":
                self.drafter = SuffixDecodingProposer(self.vllm_config)
438
            elif self.speculative_config.use_eagle():
439
                self.drafter = EagleProposer(self.vllm_config, self.device, self)
440
                if self.speculative_config.method == "eagle3":
441
442
443
                    self.use_aux_hidden_state_outputs = (
                        self.drafter.eagle3_use_aux_hidden_state
                    )
444
445
            elif self.speculative_config.method == "medusa":
                self.drafter = MedusaProposer(
446
                    vllm_config=self.vllm_config, device=self.device
447
                )
448
            else:
449
450
451
452
                raise ValueError(
                    "Unknown speculative decoding method: "
                    f"{self.speculative_config.method}"
                )
453
            self.rejection_sampler = RejectionSampler(self.sampler)
454

455
456
457
458
        self.num_spec_tokens = 0
        if self.speculative_config:
            self.num_spec_tokens = self.speculative_config.num_speculative_tokens

459
        # Request states.
460
        self.requests: dict[str, CachedRequestState] = {}
461
462
463
        # NOTE(rob): num_prompt_logprobs only includes reqs
        # that are currently in the prefill phase.
        self.num_prompt_logprobs: dict[str, int] = {}
464
        self.comm_stream = torch.cuda.Stream()
465

466
467
468
469
470
471
472
473
474
        # 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.
475
476
477
478
        logits_processors = model_config.logits_processors
        custom_logitsprocs: Sequence[str | type[LogitsProcessor]] = (
            tuple(logits_processors) if logits_processors is not None else ()
        )
479
480
        self.input_batch = InputBatch(
            max_num_reqs=self.max_num_reqs,
481
482
483
            # 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),
484
485
486
            max_num_batched_tokens=self.max_num_tokens,
            device=self.device,
            pin_memory=self.pin_memory,
487
            vocab_size=self.model_config.get_vocab_size(),
488
            block_sizes=[self.cache_config.block_size],
489
            kernel_block_sizes=[self.cache_config.block_size],
490
            is_spec_decode=bool(self.vllm_config.speculative_config),
491
            logitsprocs=build_logitsprocs(
492
493
494
                self.vllm_config,
                self.device,
                self.pin_memory,
495
                self.is_pooling_model,
496
                custom_logitsprocs,
497
            ),
498
499
500
            # 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),
501
            is_pooling_model=self.is_pooling_model,
502
            cp_kv_cache_interleave_size=self.parallel_config.cp_kv_cache_interleave_size,
503
        )
504

505
        self.use_async_scheduling = self.scheduler_config.async_scheduling
506
507
508
509
510
        # 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.
511
        self.prepare_inputs_event: torch.Event | None = None
512
513
        if self.use_async_scheduling:
            self.async_output_copy_stream = torch.cuda.Stream()
514
            self.prepare_inputs_event = torch.Event()
515

516
        # self.cudagraph_batch_sizes sorts in ascending order.
517
518
519
520
        if (
            self.compilation_config.cudagraph_capture_sizes
            and self.compilation_config.cudagraph_mode != CUDAGraphMode.NONE
        ):
521
522
            self.cudagraph_batch_sizes = sorted(
                self.compilation_config.cudagraph_capture_sizes
523
            )
524

525
        # Cache the device properties.
526
        self._init_device_properties()
527

528
        # Persistent buffers for CUDA graphs.
529
530
531
532
533
        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
        )
534
        self.seq_lens = self._make_buffer(self.max_num_reqs, dtype=torch.int32)
535
        self.encoder_seq_lens = self._make_buffer(self.max_num_reqs, dtype=torch.int32)
536
537
538
539
        if self.dcp_world_size > 1:
            self.dcp_local_seq_lens = self._make_buffer(
                self.max_num_reqs, dtype=torch.int32
            )
540
541
542
        # 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.
543
        self.inputs_embeds = self._make_buffer(
544
            self.max_num_tokens, self.inputs_embeds_size, dtype=self.dtype, numpy=False
545
546
        )
        self.is_token_ids = self._make_buffer(self.max_num_tokens, dtype=torch.bool)
547
548
        self.discard_request_mask = self._make_buffer(
            self.max_num_reqs, dtype=torch.bool
549
550
551
552
553
554
555
        )
        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
        )
556

557
558
        # Only relevant for multimodal models
        if self.supports_mm_inputs:
559
            self.is_mm_embed = self._make_buffer(self.max_num_tokens, dtype=torch.bool)
560

561
        # Only relevant for models using M-RoPE (e.g, Qwen2-VL)
562
        if self.uses_mrope:
Roger Wang's avatar
Roger Wang committed
563
564
565
566
            # 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
567
568
569
570
571
572

            # 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
573
            self.mrope_positions = self._make_buffer(
574
575
                (3, self.max_num_tokens + 1), dtype=torch.int64
            )
576

577
578
579
580
581
582
583
        # 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
            )

584
        # None in the first PP rank. The rest are set after load_model.
585
        self.intermediate_tensors: IntermediateTensors | None = None
586

587
        # OPTIMIZATION: Cache the tensors rather than creating them every step.
588
        # Keep in int64 to avoid overflow with long context
589
590
591
592
        self.arange_np = np.arange(
            max(self.max_num_reqs + 1, self.max_model_len, self.max_num_tokens),
            dtype=np.int64,
        )
593

594
595
596
597
598
        # 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] = {}
599
600
601
602
603
        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(
604
605
                self.max_num_tokens, dtype=torch.int32, device=self.device
            )
606

607
        self.uniform_decode_query_len = 1 + self.num_spec_tokens
608
609
610
611

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

612
613
        self.mm_budget = (
            MultiModalBudget(
614
                self.model_config,
615
616
617
618
619
620
                self.scheduler_config,
                self.mm_registry,
            )
            if self.supports_mm_inputs
            else None
        )
621

622
        self.reorder_batch_threshold: int | None = None
623

624
625
626
627
628
        # 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()

629
        # Cached outputs.
630
        self._draft_token_ids: list[list[int]] | torch.Tensor | None = None
631
        self.transfer_event = torch.Event()
632
        self.sampled_token_ids_pinned_cpu = torch.empty(
633
            (self.max_num_reqs, 1),
634
635
            dtype=torch.int64,
            device="cpu",
636
637
            pin_memory=self.pin_memory,
        )
638

639
640
        # Pre-allocated tensor for copying valid sampled token counts to CPU,
        # with dedicated stream for overlapping and event for coordination.
641
        self.valid_sampled_token_count_event: torch.Event | None = None
642
643
        self.valid_sampled_token_count_copy_stream: torch.cuda.Stream | None = None
        if self.use_async_scheduling and self.num_spec_tokens:
644
            self.valid_sampled_token_count_event = torch.Event()
645
646
647
648
649
650
651
652
            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,
        )

653
654
        # Ephemeral state transferred between execute_model() and sample_tokens().
        self.execute_model_state: ExecuteModelState | None = None
655
        self.kv_connector_output: KVConnectorOutput | None = None
656
        self.layerwise_nvtx_hooks_registered = False
657

658
659
660
661
    def reset_mm_cache(self) -> None:
        if self.mm_budget:
            self.mm_budget.reset_cache()

662
663
664
665
666
667
668
669
670
671
672
673
674
675
676
677
678
679
680
681
682
683
684
685
686
687
688
689
690
691
692
693
694
695
696
697
698
699
700
701
702
703
704
705
    @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)

706
707
708
709
    def _get_positions(self, num_tokens: Any):
        if isinstance(num_tokens, int):
            if self.uses_mrope:
                return self.mrope_positions.gpu[:, :num_tokens]
710
711
            if self.uses_xdrope_dim > 0:
                return self.xdrope_positions.gpu[:, :num_tokens]
712
713
714
715
            return self.positions.gpu[:num_tokens]
        else:
            if self.uses_mrope:
                return self.mrope_positions.gpu[:, num_tokens]
716
717
            if self.uses_xdrope_dim > 0:
                return self.xdrope_positions.gpu[:, num_tokens]
718
719
            return self.positions.gpu[num_tokens]

720
    def _make_buffer(
721
        self, *size: int | torch.SymInt, dtype: torch.dtype, numpy: bool = True
722
723
724
725
726
727
728
729
    ) -> CpuGpuBuffer:
        return CpuGpuBuffer(
            *size,
            dtype=dtype,
            device=self.device,
            pin_memory=self.pin_memory,
            with_numpy=numpy,
        )
730

731
    def _init_model_kwargs(self):
732
733
        model_kwargs = dict[str, Any]()

734
        if not self.is_pooling_model:
735
736
            return model_kwargs

737
738
        num_reqs = self.input_batch.num_reqs
        pooling_params = self.input_batch.get_pooling_params()
739
740
741

        token_type_id_requests = dict[int, Any]()
        for i, param in enumerate(pooling_params):
742
743
744
745
746
            if (
                param.extra_kwargs is not None
                and (token_types := param.extra_kwargs.get("compressed_token_type_ids"))
                is not None
            ):
747
748
749
750
751
                token_type_id_requests[i] = token_types

        if len(token_type_id_requests) == 0:
            return model_kwargs

752
        seq_lens = self.seq_lens.gpu[:num_reqs]
753
754
755
756
757
758
759
760
        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(
761
762
            device=self.device
        )
763
764
        return model_kwargs

765
    def _may_reorder_batch(self, scheduler_output: "SchedulerOutput") -> None:
766
767
        """
        Update the order of requests in the batch based on the attention
768
        backend's needs. For example, some attention backends (namely MLA) may
769
770
771
772
773
774
        want to separate requests based on if the attention computation will be
        compute-bound or memory-bound.

        Args:
            scheduler_output: The scheduler output.
        """
775
776
777
778
779
780
781
782
        # 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

783
784
785
786
        if self.reorder_batch_threshold is not None:
            reorder_batch_to_split_decodes_and_prefills(
                self.input_batch,
                scheduler_output,
787
788
                decode_threshold=self.reorder_batch_threshold,
            )
789

790
791
    # Note: used for model runner override.
    def _init_device_properties(self) -> None:
792
        """Initialize attributes from torch.cuda.get_device_properties"""
793
794
795
796
797
798
799
        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()

800
    def _update_states(self, scheduler_output: "SchedulerOutput") -> None:
801
802
803
804
805
806
        """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.

807
808
        The SamplingMetadata is updated and copied to the GPU if there is a
        new/resumed/paused/finished request in the batch.
809
810
        """
        # Remove finished requests from the cached states.
811
812
        for req_id in scheduler_output.finished_req_ids:
            self.requests.pop(req_id, None)
813
            self.num_prompt_logprobs.pop(req_id, None)
814
815
816
817
818
819
820
        # 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:
821
            self.input_batch.remove_request(req_id)
822
823

        # Free the cached encoder outputs.
824
825
        for mm_hash in scheduler_output.free_encoder_mm_hashes:
            self.encoder_cache.pop(mm_hash, None)
826

827
828
829
830
831
832
833
        # 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()
834
835
836
837
838
839
840
841
        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)
842
843
844
845
846
        # 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:
847
            self.input_batch.remove_request(req_id)
848

849
        reqs_to_add: list[CachedRequestState] = []
850
        # Add new requests to the cached states.
851
852
853
        for new_req_data in scheduler_output.scheduled_new_reqs:
            req_id = new_req_data.req_id
            sampling_params = new_req_data.sampling_params
854
            pooling_params = new_req_data.pooling_params
855

856
857
858
859
            if (
                sampling_params
                and sampling_params.sampling_type == SamplingType.RANDOM_SEED
            ):
860
861
862
863
864
                generator = torch.Generator(device=self.device)
                generator.manual_seed(sampling_params.seed)
            else:
                generator = None

865
866
            if self.is_pooling_model:
                assert pooling_params is not None
867
868
                task = pooling_params.task
                assert task is not None, "You did not set `task` in the API"
869

870
                model = cast(VllmModelForPooling, self.get_model())
871
                to_update = model.pooler.get_pooling_updates(task)
872
873
                to_update.apply(pooling_params)

874
            req_state = CachedRequestState(
875
                req_id=req_id,
876
                prompt_token_ids=new_req_data.prompt_token_ids,
877
                prompt_embeds=new_req_data.prompt_embeds,
878
                mm_features=new_req_data.mm_features,
879
                sampling_params=sampling_params,
880
                pooling_params=pooling_params,
881
                generator=generator,
882
883
                block_ids=new_req_data.block_ids,
                num_computed_tokens=new_req_data.num_computed_tokens,
884
                output_token_ids=[],
885
                lora_request=new_req_data.lora_request,
886
            )
887
888
            self.requests[req_id] = req_state

889
890
891
892
893
894
895
            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
                )

896
            # Only relevant for models using M-RoPE (e.g, Qwen2-VL)
897
            if self.uses_mrope:
898
                self._init_mrope_positions(req_state)
899

900
901
902
903
            # Only relevant for models using XD-RoPE (e.g, HunYuan-VL)
            if self.uses_xdrope_dim > 0:
                self._init_xdrope_positions(req_state)

904
            reqs_to_add.append(req_state)
905

906
        # Update the states of the running/resumed requests.
907
        is_last_rank = get_pp_group().is_last_rank
908
        req_data = scheduler_output.scheduled_cached_reqs
909
910
911
912
913

        # 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()

914
        for i, req_id in enumerate(req_data.req_ids):
915
            req_state = self.requests[req_id]
916
917
            num_computed_tokens = req_data.num_computed_tokens[i]
            new_block_ids = req_data.new_block_ids[i]
918
            resumed_from_preemption = req_id in req_data.resumed_req_ids
919
            num_output_tokens = req_data.num_output_tokens[i]
920
            req_index = self.input_batch.req_id_to_index.get(req_id)
921

922
923
924
925
926
927
928
929
930
931
932
933
934
935
936
937
938
939
940
941
942
943
944
            # prev_num_draft_len is used in async scheduling mode with
            # spec decode. it indicates if need to update num_computed_tokens
            # of the request. for example:
            # fist step: num_computed_tokens = 0, spec_tokens = [],
            # prev_num_draft_len = 0.
            # second step: num_computed_tokens = 100(prompt lenth),
            # spec_tokens = [a,b], prev_num_draft_len = 0.
            # third step: num_computed_tokens = 100 + 2, spec_tokens = [c,d],
            # prev_num_draft_len = 2.
            # num_computed_tokens in first step and second step does't contain
            # the spec tokens length, but in third step it contains the
            # spec tokens length. we only need to update num_computed_tokens
            # when prev_num_draft_len > 0.
            if req_state.prev_num_draft_len:
                if req_index is None:
                    req_state.prev_num_draft_len = 0
                else:
                    assert self.input_batch.prev_req_id_to_index is not None
                    prev_req_index = self.input_batch.prev_req_id_to_index[req_id]
                    num_accepted = valid_sampled_token_count[prev_req_index] - 1
                    num_rejected = req_state.prev_num_draft_len - num_accepted
                    num_computed_tokens -= num_rejected
                    req_state.output_token_ids.extend([-1] * num_accepted)
945

946
            # Update the cached states.
947
            req_state.num_computed_tokens = num_computed_tokens
948
949
950
951
952
953
954
955

            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.
956
957
958
                num_new_tokens = (
                    num_computed_tokens + len(new_token_ids) - req_state.num_tokens
                )
959
960
961
962
                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:
963
                    req_state.output_token_ids.extend(new_token_ids[-num_new_tokens:])
964
965
966
967
968
            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:
969
970
971
972
                    end_idx = (
                        self.input_batch.num_prompt_tokens[req_index]
                        + num_output_tokens
                    )
973
                    self.input_batch.num_tokens_no_spec[req_index] = end_idx
974

975
            # Update the block IDs.
976
            if not resumed_from_preemption:
977
978
                if new_block_ids is not None:
                    # Append the new blocks to the existing block IDs.
979
                    for block_ids, new_ids in zip(req_state.block_ids, new_block_ids):
980
                        block_ids.extend(new_ids)
981
            else:
982
                assert req_index is None
983
                assert new_block_ids is not None
984
985
                # The request is resumed from preemption.
                # Replace the existing block IDs with the new ones.
986
                req_state.block_ids = new_block_ids
987
988
989
990
991

            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.
992
993
994
995
996
997
998

                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:]

999
                reqs_to_add.append(req_state)
1000
1001
1002
                continue

            # Update the persistent batch.
1003
            self.input_batch.num_computed_tokens_cpu[req_index] = num_computed_tokens
1004
            if new_block_ids is not None:
1005
                self.input_batch.block_table.append_row(new_block_ids, req_index)
1006
1007
1008
1009
1010
1011
1012

            # For the last rank, we don't need to update the token_ids_cpu
            # because the sampled tokens are already cached.
            if not is_last_rank:
                # Add new_token_ids to token_ids_cpu.
                start_token_index = num_computed_tokens
                end_token_index = num_computed_tokens + len(new_token_ids)
1013
                self.input_batch.token_ids_cpu[
1014
1015
1016
                    req_index, start_token_index:end_token_index
                ] = new_token_ids
                self.input_batch.num_tokens_no_spec[req_index] = end_token_index
1017

1018
            # Add spec_token_ids to token_ids_cpu.
1019
            spec_token_ids = scheduler_output.scheduled_spec_decode_tokens.get(
1020
                req_id, []
1021
            )
1022
1023
1024
1025
1026
            num_spec_tokens = len(spec_token_ids)
            # For async scheduling, token_ids_cpu assigned from
            # spec_token_ids are placeholders and will be overwritten in
            # _prepare_input_ids.
            if num_spec_tokens:
1027
1028
1029
                start_index = self.input_batch.num_tokens_no_spec[req_index]
                end_token_index = start_index + num_spec_tokens
                self.input_batch.token_ids_cpu[
1030
1031
                    req_index, start_index:end_token_index
                ] = spec_token_ids
1032
1033
1034
1035
1036
1037

            # When speculative decoding is used with structured output,
            # the scheduler can drop draft tokens that do not
            # conform to the schema. This can result in
            # scheduler_output.scheduled_spec_decode_tokens being empty,
            # even when speculative decoding is enabled.
1038
1039
            self.input_batch.spec_token_ids[req_index].clear()
            self.input_batch.spec_token_ids[req_index].extend(spec_token_ids)
1040

1041
1042
1043
1044
1045
1046
1047
1048
1049
            # there are no draft tokens with async scheduling,
            # we clear the spec_decoding info in scheduler_output and
            # use normal sampling but rejection_sampling.
            if self.use_async_scheduling:
                req_state.prev_num_draft_len = num_spec_tokens
                if num_spec_tokens and self._draft_token_ids is None:
                    scheduler_output.total_num_scheduled_tokens -= num_spec_tokens
                    scheduler_output.num_scheduled_tokens[req_id] -= num_spec_tokens
                    scheduler_output.scheduled_spec_decode_tokens.pop(req_id, None)
1050
1051
        # Add the new or resumed requests to the persistent batch.
        # The smaller empty indices are filled first.
1052
1053
        for request in reqs_to_add:
            self.input_batch.add_request(request)
1054

1055
1056
1057
1058
1059
1060
        # Condense the batched states if there are gaps left by removed requests
        self.input_batch.condense()
        # Allow attention backend to reorder the batch, potentially
        self._may_reorder_batch(scheduler_output)
        # Refresh batch metadata with any pending updates.
        self.input_batch.refresh_metadata()
1061

1062
    def _update_states_after_model_execute(
1063
1064
        self, output_token_ids: torch.Tensor
    ) -> None:
1065
1066
1067
1068
1069
1070
1071
1072
        """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.
        """
1073
        if not self.speculative_config or not self.model_config.is_hybrid:
1074
1075
1076
            return

        # Find the number of accepted tokens for each sequence.
1077
1078
1079
1080
1081
1082
1083
1084
1085
1086
1087
1088
1089
1090
1091
1092
1093
1094
1095
1096
        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()
        )
1097
1098
1099
        for i, num_tokens in enumerate(num_accepted_tokens):
            self.input_batch.num_accepted_tokens_cpu[i] = num_tokens

1100
    def _init_mrope_positions(self, req_state: CachedRequestState):
1101
1102
        model = self.get_model()
        assert supports_mrope(model), "M-RoPE support is not implemented."
1103
1104
1105
1106
        assert req_state.prompt_token_ids is not None, (
            "M-RoPE requires prompt_token_ids to be available."
        )
        mrope_model = cast(SupportsMRoPE, model)
1107
1108

        req_state.mrope_positions, req_state.mrope_position_delta = (
1109
            mrope_model.get_mrope_input_positions(
1110
                req_state.prompt_token_ids,
1111
                req_state.mm_features,
1112
            )
1113
        )
1114

1115
1116
1117
1118
1119
1120
1121
1122
1123
1124
1125
1126
1127
    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,
        )

1128
    def _extract_mm_kwargs(
1129
        self,
1130
1131
        scheduler_output: "SchedulerOutput",
    ) -> BatchedTensorInputs:
1132
        if not scheduler_output or not self.is_multimodal_raw_input_only_model:
1133
            return {}
1134

1135
1136
        mm_kwargs = list[MultiModalKwargsItem]()
        for req in scheduler_output.scheduled_new_reqs:
1137
1138
1139
            for feature in req.mm_features:
                if feature.data is not None:
                    mm_kwargs.append(feature.data)
1140

1141
1142
1143
        # Input all modalities at once
        mm_kwargs_combined: BatchedTensorInputs = {}
        for _, _, mm_kwargs_group in group_mm_kwargs_by_modality(
1144
1145
1146
            mm_kwargs,
            device=self.device,
            pin_memory=self.pin_memory,
1147
1148
        ):
            mm_kwargs_combined.update(mm_kwargs_group)
1149

1150
        return mm_kwargs_combined
1151

1152
    def _dummy_mm_kwargs(self, num_seqs: int) -> BatchedTensorInputs:
1153
        if not self.is_multimodal_raw_input_only_model:
1154
            return {}
1155

1156
1157
1158
1159
1160
        mm_budget = self.mm_budget
        assert mm_budget is not None

        dummy_modality = mm_budget.get_modality_with_max_tokens()
        return self._get_mm_dummy_batch(dummy_modality, num_seqs)
1161

1162
1163
1164
    def _get_cumsum_and_arange(
        self,
        num_tokens: np.ndarray,
1165
        cumsum_dtype: np.dtype | None = None,
1166
1167
1168
1169
1170
1171
1172
1173
1174
1175
1176
1177
1178
1179
1180
1181
    ) -> 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

1182
    def _prepare_input_ids(
1183
1184
1185
1186
        self,
        scheduler_output: "SchedulerOutput",
        total_num_scheduled_tokens: int,
        cu_num_tokens: np.ndarray,
1187
    ) -> None:
1188
        """Prepare the input IDs for the current batch.
1189

1190
1191
1192
1193
1194
1195
1196
        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)
1197
1198
1199
            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)
1200
1201
1202
1203
1204
1205
1206
            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
1207
1208
1209
1210
        sample_flattened_indices: list[int] = []
        spec_flattened_indices: list[int] = []
        prev_common_req_indices: list[int] = []
        prev_draft_token_indices: list[int] = []
1211
1212
        indices_match = True
        max_flattened_index = -1
1213
1214
1215
        total_num_spec_tokens = 0
        scheduled_spec_tokens = scheduler_output.scheduled_spec_decode_tokens

1216
1217
1218
1219
1220
        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.
1221
1222
                draft_len = len(scheduled_spec_tokens.get(req_id, ()))
                total_num_spec_tokens += draft_len
1223
                flattened_index = cu_num_tokens[cur_index].item() - 1
1224
1225
1226
1227
1228
1229
1230
1231
1232
1233
1234
1235
1236
1237
1238
                # 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))
1239
                indices_match &= prev_index == flattened_index
1240
                max_flattened_index = max(max_flattened_index, flattened_index)
1241
1242
1243
        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:
1244
1245
1246
            # 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)
1247
1248
1249
            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)
1250
1251
        if num_commmon_tokens == 0:
            # No requests in common with the previous iteration
1252
            # So input_ids.cpu will have all the input ids.
1253
1254
1255
1256
1257
1258
1259
            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_(
1260
1261
1262
                self.input_batch.prev_sampled_token_ids[:num_commmon_tokens, 0],
                non_blocking=True,
            )
1263
1264
            if self.enable_prompt_embeds:
                self.is_token_ids.gpu[:num_commmon_tokens] = True
1265
            return
1266
        # Upload the index tensors asynchronously so the scatter can be non-blocking.
1267
1268
        sampled_tokens_index_tensor = torch.tensor(
            sample_flattened_indices, dtype=torch.int64, pin_memory=self.pin_memory
1269
        ).to(self.device, non_blocking=True)
1270
        prev_common_req_indices_tensor = torch.tensor(
1271
1272
            prev_common_req_indices, dtype=torch.int64, pin_memory=self.pin_memory
        ).to(self.device, non_blocking=True)
1273
1274
        self.input_ids.gpu.scatter_(
            dim=0,
1275
            index=sampled_tokens_index_tensor,
1276
            src=self.input_batch.prev_sampled_token_ids[
1277
1278
1279
                prev_common_req_indices_tensor, 0
            ],
        )
1280

1281
1282
1283
1284
1285
1286
1287
1288
1289
1290
1291
1292
1293
1294
1295
1296
1297
1298
1299
1300
1301
1302
1303
        # Scatter the draft tokens after the sampled tokens are scattered.
        if self._draft_token_ids is None or not spec_flattened_indices:
            return

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

        # because input_ids dtype is torch.int32,
        # so convert draft_token_ids to torch.int32 here.
        draft_token_ids = self._draft_token_ids.to(dtype=torch.int32)
        self._draft_token_ids = None

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

1304
1305
    def _get_encoder_seq_lens(
        self,
1306
        num_scheduled_tokens: dict[str, int],
1307
1308
        kv_cache_spec: KVCacheSpec,
        num_reqs: int,
1309
    ) -> tuple[torch.Tensor | None, np.ndarray | None]:
1310
        if not isinstance(kv_cache_spec, CrossAttentionSpec):
1311
            return None, None
1312

1313
1314
        # Zero out buffer for padding requests that are not actually scheduled (CGs)
        self.encoder_seq_lens.np[:num_reqs] = 0
1315
1316
        # Build encoder_seq_lens array mapping request indices to
        # encoder lengths for inputs scheduled in this batch
1317
        for req_id in num_scheduled_tokens:
1318
            req_index = self.input_batch.req_id_to_index[req_id]
1319
1320
1321
1322
1323
1324
1325
1326
1327
1328
1329
1330
1331
1332
1333
1334
            req_state = self.requests[req_id]
            if req_state.mm_features is None:
                self.encoder_seq_lens.np[req_index] = 0
                continue

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

        self.encoder_seq_lens.copy_to_gpu(num_reqs)
        encoder_seq_lens = self.encoder_seq_lens.gpu[:num_reqs]
        encoder_seq_lens_cpu = self.encoder_seq_lens.np[:num_reqs]
1335

1336
        return encoder_seq_lens, encoder_seq_lens_cpu
1337

1338
    def _prepare_inputs(
1339
1340
1341
        self,
        scheduler_output: "SchedulerOutput",
        num_scheduled_tokens: np.ndarray,
1342
1343
    ) -> tuple[
        torch.Tensor,
1344
        SpecDecodeMetadata | None,
1345
    ]:
1346
1347
        """
        :return: tuple[
1348
            logits_indices, spec_decode_metadata,
1349
1350
        ]
        """
1351
1352
1353
1354
1355
1356
1357
        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.
1358
        self.input_batch.block_table.commit_block_table(num_reqs)
1359
1360
1361

        # Get request indices.
        # E.g., [2, 5, 3] -> [0, 0, 1, 1, 1, 1, 1, 2, 2, 2]
1362
        req_indices = np.repeat(self.arange_np[:num_reqs], num_scheduled_tokens)
1363

1364
1365
        # cu_num_tokens: [2, 5, 3] -> [2, 7, 10]
        # arange: [0, 1, 0, 1, 2, 3, 4, 0, 1, 2]
1366
        cu_num_tokens, arange = self._get_cumsum_and_arange(num_scheduled_tokens)
1367
1368

        # Get positions.
1369
        positions_np = self.positions.np[:total_num_scheduled_tokens]
1370
1371
1372
1373
1374
        np.add(
            self.input_batch.num_computed_tokens_cpu[req_indices],
            arange,
            out=positions_np,
        )
1375

1376
1377
        # Calculate M-RoPE positions.
        # Only relevant for models using M-RoPE (e.g, Qwen2-VL)
1378
        if self.uses_mrope:
1379
1380
            self._calc_mrope_positions(scheduler_output)

1381
1382
1383
1384
1385
        # 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)

1386
1387
1388
1389
        # 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.
1390
1391
1392
        token_indices = (
            positions_np + req_indices * self.input_batch.token_ids_cpu.shape[1]
        )
1393
        token_indices_tensor = torch.from_numpy(token_indices)
1394

1395
1396
1397
        # 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.
1398
1399
1400
1401
1402
1403
        torch.index_select(
            self.input_batch.token_ids_cpu_tensor.flatten(),
            0,
            token_indices_tensor,
            out=self.input_ids.cpu[:total_num_scheduled_tokens],
        )
1404
        if self.enable_prompt_embeds:
1405
            is_token_ids = self.input_batch.is_token_ids_tensor.flatten()
1406
1407
1408
1409
            torch.index_select(
                is_token_ids,
                0,
                token_indices_tensor,
1410
1411
                out=self.is_token_ids.cpu[:total_num_scheduled_tokens],
            )
1412
1413
1414
1415
1416
1417
1418
1419
1420
1421
1422
1423
1424
1425
1426
1427
1428
1429
1430
1431
1432
1433
1434
1435
1436
1437
1438
1439
1440
1441
1442
1443
1444

        # 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:
1445
1446
1447
                    self.inputs_embeds.cpu[
                        output_idx : output_idx + actual_num_sched
                    ].copy_(req_embeds[start_pos:actual_end])
1448
1449

                output_idx += num_sched
1450

1451
1452
        self.input_batch.block_table.compute_slot_mapping(req_indices, positions_np)
        self.input_batch.block_table.commit_slot_mapping(total_num_scheduled_tokens)
1453
1454

        # Prepare the attention metadata.
1455
        self.query_start_loc.np[0] = 0
1456
        self.query_start_loc.np[1 : num_reqs + 1] = cu_num_tokens
1457
1458
        # Note: pad query_start_loc to be non-decreasing, as kernels
        # like FlashAttention requires that
1459
        self.query_start_loc.np[num_reqs + 1 :].fill(cu_num_tokens[-1])
1460
        self.query_start_loc.copy_to_gpu()
1461
        query_start_loc = self.query_start_loc.gpu[: num_reqs + 1]
1462

1463
        self.seq_lens.np[:num_reqs] = (
1464
1465
            self.input_batch.num_computed_tokens_cpu[:num_reqs] + num_scheduled_tokens
        )
1466
        # Fill unused with 0 for full cuda graph mode.
1467
1468
        self.seq_lens.np[num_reqs:].fill(0)
        self.seq_lens.copy_to_gpu()
1469

1470
        num_tokens = [self.requests[r].num_tokens for r in self.input_batch.req_ids]
1471
1472
        num_tokens_np = np.array(num_tokens, dtype=np.int32)

1473
        # Record which requests should not be sampled,
1474
        # so that we could clear the sampled tokens before returning
1475
1476
        self.discard_request_mask.np[:num_reqs] = (
            self.seq_lens.np[:num_reqs] < num_tokens_np
1477
        )
1478
        self.discard_request_mask.copy_to_gpu(num_reqs)
1479

1480
        # Copy the tensors to the GPU.
1481
1482
1483
1484
1485
        self._prepare_input_ids(
            scheduler_output,
            total_num_scheduled_tokens,
            cu_num_tokens,
        )
1486

1487
        if self.uses_mrope:
1488
            # Only relevant for models using M-RoPE (e.g, Qwen2-VL)
1489
1490
            self.mrope_positions.gpu[:, :total_num_scheduled_tokens].copy_(
                self.mrope_positions.cpu[:, :total_num_scheduled_tokens],
1491
1492
                non_blocking=True,
            )
1493
1494
1495
1496
1497
1498
        elif self.uses_xdrope_dim > 0:
            # Only relevant for models using XD-RoPE (e.g, HunYuan-VL)
            self.xdrope_positions.gpu[:, :total_num_scheduled_tokens].copy_(
                self.xdrope_positions.cpu[:, :total_num_scheduled_tokens],
                non_blocking=True,
            )
1499
1500
        else:
            # Common case (1D positions)
1501
            self.positions.copy_to_gpu(total_num_scheduled_tokens)
1502

1503
        use_spec_decode = len(scheduler_output.scheduled_spec_decode_tokens) > 0
1504
1505
1506
1507
1508
1509
1510
        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
1511
            num_draft_tokens = None
1512
            spec_decode_metadata = None
1513
            num_sampled_tokens = np.ones(num_reqs, dtype=np.int32)
1514
1515
1516
1517
1518
        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)
1519
1520
1521
            # 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)
1522
1523
1524
1525
            for (
                req_id,
                draft_token_ids,
            ) in scheduler_output.scheduled_spec_decode_tokens.items():
1526
1527
                req_idx = self.input_batch.req_id_to_index[req_id]
                num_draft_tokens[req_idx] = len(draft_token_ids)
1528
1529
1530
1531
1532
1533
1534
1535
                num_decode_draft_tokens[req_idx] = (
                    len(draft_token_ids)
                    if (
                        self.input_batch.num_computed_tokens_cpu[req_idx]
                        >= self.input_batch.num_prompt_tokens[req_idx]
                    )
                    else -1
                )
1536
            spec_decode_metadata = self._calc_spec_decode_metadata(
1537
1538
                num_draft_tokens, cu_num_tokens
            )
1539
            logits_indices = spec_decode_metadata.logits_indices
1540
            num_sampled_tokens = num_draft_tokens + 1
1541
            # For DECODE only cuda graph of some attention backends (e.g., GDN).
1542
            self.num_decode_draft_tokens.np[:num_reqs] = num_decode_draft_tokens
1543
1544
            self.num_decode_draft_tokens.np[num_reqs:].fill(-1)
            self.num_decode_draft_tokens.copy_to_gpu()
1545

1546
1547
1548
1549
1550
        # Hot-Swap lora model
        if self.lora_config:
            assert (
                np.sum(num_sampled_tokens)
                <= self.vllm_config.scheduler_config.max_num_batched_tokens
1551
            )
1552
1553
1554
1555
1556
1557
1558
1559
1560
1561
1562
            self.set_active_loras(
                self.input_batch, num_scheduled_tokens, num_sampled_tokens
            )

        return (
            logits_indices,
            spec_decode_metadata,
        )

    def _build_attention_metadata(
        self,
1563
        num_tokens: int,
1564
        num_reqs: int,
1565
1566
1567
        max_query_len: int,
        num_tokens_padded: int | None = None,
        num_reqs_padded: int | None = None,
1568
1569
1570
1571
        ubatch_slices: UBatchSlices | None = None,
        logits_indices: torch.Tensor | None = None,
        use_spec_decode: bool = False,
        for_cudagraph_capture: bool = False,
1572
        num_scheduled_tokens: dict[str, int] | None = None,
1573
1574
1575
1576
1577
        cascade_attn_prefix_lens: list[list[int]] | None = None,
    ) -> tuple[PerLayerAttnMetadata, CommonAttentionMetadata | None]:
        """
        :return: tuple[attn_metadata, spec_decode_common_attn_metadata]
        """
1578
1579
1580
1581
        # Attention metadata is not needed for attention free models
        if len(self.kv_cache_config.kv_cache_groups) == 0:
            return {}, None

1582
1583
        num_tokens_padded = num_tokens_padded or num_tokens
        num_reqs_padded = num_reqs_padded or num_reqs
1584
        assert num_reqs_padded is not None and num_tokens_padded is not None
1585

1586
1587
1588
        attn_metadata: PerLayerAttnMetadata = {}
        if ubatch_slices is not None:
            attn_metadata = [dict() for _ in range(len(ubatch_slices))]
1589

1590
1591
1592
1593
1594
1595
1596
1597
        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()

1598
1599
        if use_spec_decode:
            self.num_accepted_tokens.np[:num_reqs] = (
1600
1601
                self.input_batch.num_accepted_tokens_cpu[:num_reqs]
            )
1602
1603
            self.num_accepted_tokens.np[num_reqs:].fill(1)
            self.num_accepted_tokens.copy_to_gpu()
1604

1605
        kv_cache_groups = self.kv_cache_config.kv_cache_groups
1606

1607
1608
1609
1610
        def _get_block_table_and_slot_mapping(kv_cache_gid: int):
            assert num_reqs_padded is not None and num_tokens_padded is not None
            kv_cache_spec = kv_cache_groups[kv_cache_gid].kv_cache_spec
            if isinstance(kv_cache_spec, EncoderOnlyAttentionSpec):
1611
                blk_table_tensor = torch.zeros(
1612
                    (num_reqs_padded, 1),
1613
                    dtype=torch.int32,
1614
1615
1616
                    device=self.device,
                )
                slot_mapping = torch.zeros(
1617
                    (num_tokens_padded,),
1618
1619
1620
                    dtype=torch.int64,
                    device=self.device,
                )
1621
            else:
1622
                blk_table = self.input_batch.block_table[kv_cache_gid]
1623
1624
                blk_table_tensor = blk_table.get_device_tensor(num_reqs_padded)
                slot_mapping = blk_table.slot_mapping.gpu[:num_tokens_padded]
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
1655
1656
1657
1658
1659
1660
1661
1662
1663
1664
1665
1666
1667
1668
1669
1670
1671
            # Fill unused with -1. Needed for reshape_and_cache in full cuda
            # graph mode. `blk_table_tensor` -1 to match mamba PAD_SLOT_ID
            slot_mapping[num_tokens:num_tokens_padded].fill_(-1)
            blk_table_tensor[num_reqs:num_reqs_padded].fill_(-1)

            return blk_table_tensor, slot_mapping

        block_table_gid_0, slot_mapping_gid_0 = _get_block_table_and_slot_mapping(0)
        cm_base = CommonAttentionMetadata(
            query_start_loc=self.query_start_loc.gpu[: num_reqs_padded + 1],
            query_start_loc_cpu=self.query_start_loc.cpu[: num_reqs_padded + 1],
            seq_lens=self.seq_lens.gpu[:num_reqs_padded],
            _seq_lens_cpu=self.seq_lens.cpu[:num_reqs_padded],
            _num_computed_tokens_cpu=self.input_batch.num_computed_tokens_cpu_tensor[
                :num_reqs_padded
            ],
            num_reqs=num_reqs_padded,
            num_actual_tokens=num_tokens_padded,
            max_query_len=max_query_len,
            max_seq_len=max_seq_len,
            block_table_tensor=block_table_gid_0,
            slot_mapping=slot_mapping_gid_0,
            causal=True,
        )

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

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

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

1672
1673
1674
1675
1676
1677
1678
1679
1680
        # 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
        ] = {}

1681
1682
1683
1684
1685
1686
1687
        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]
1688
            builder = attn_group.get_metadata_builder(ubid or 0)
1689
1690
1691
1692
            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))
1693

1694
1695
1696
1697
1698
1699
1700
1701
1702
1703
1704
1705
1706
1707
1708
1709
1710
1711
1712
1713
            cascade_attn_prefix_len = (
                cascade_attn_prefix_lens[kv_cache_gid][attn_gid]
                if cascade_attn_prefix_lens
                else 0
            )

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

            if for_cudagraph_capture:
                attn_metadata_i = builder.build_for_cudagraph_capture(
                    common_attn_metadata
                )
1714
1715
1716
1717
1718
1719
1720
1721
1722
            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,
                )
1723
1724
1725
1726
1727
1728
            else:
                attn_metadata_i = builder.build(
                    common_prefix_len=cascade_attn_prefix_len,
                    common_attn_metadata=common_attn_metadata,
                    **extra_attn_metadata_args,
                )
1729
1730
                if builder.supports_update_block_table:
                    cached_attn_metadata[cache_key] = attn_metadata_i
1731
1732
1733
1734
1735
1736
1737
1738
1739
1740
1741
1742
1743
1744
1745
1746
1747
1748
1749
1750
1751
1752
1753

            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,
1754
            )
1755
1756
1757
1758
            if kv_cache_gid > 0:
                cm.block_table_tensor, cm.slot_mapping = (
                    _get_block_table_and_slot_mapping(kv_cache_gid)
                )
1759

1760
            if self.speculative_config and spec_decode_common_attn_metadata is None:
1761
                if isinstance(self.drafter, EagleProposer):
1762
                    if self.drafter.attn_layer_names[0] in kv_cache_group.layer_names:
1763
                        spec_decode_common_attn_metadata = cm
1764
                else:
1765
                    spec_decode_common_attn_metadata = cm
1766

1767
            for attn_gid in range(len(self.attn_groups[kv_cache_gid])):
1768
                if ubatch_slices is not None:
1769
1770
1771
                    for ubid, _cm in enumerate(split_attn_metadata(ubatch_slices, cm)):
                        _build_attn_group_metadata(kv_cache_gid, attn_gid, _cm, ubid)

1772
                else:
1773
                    _build_attn_group_metadata(kv_cache_gid, attn_gid, cm)
1774

1775
1776
1777
1778
1779
1780
1781
1782
1783
1784
1785
1786
1787
1788
1789
1790
1791
1792
1793
1794
        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]

1795
1796
1797
1798
1799
1800
1801
1802
1803
1804
        if spec_decode_common_attn_metadata is not None and (
            num_reqs != num_reqs_padded or num_tokens != num_tokens_padded
        ):
            # Currently the drafter still only uses piecewise cudagraphs (and modifies
            # the attention metadata in directly), and therefore does not want to use
            # padded attention metadata.
            spec_decode_common_attn_metadata = (
                spec_decode_common_attn_metadata.unpadded(num_tokens, num_reqs)
            )

1805
        return attn_metadata, spec_decode_common_attn_metadata
1806

1807
1808
1809
    def _compute_cascade_attn_prefix_lens(
        self,
        num_scheduled_tokens: np.ndarray,
1810
        num_computed_tokens: np.ndarray,
1811
1812
1813
1814
1815
1816
1817
        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
        """
1818

1819
1820
1821
1822
1823
1824
1825
1826
1827
1828
1829
1830
1831
1832
        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,
1833
                        num_computed_tokens,
1834
1835
1836
1837
1838
1839
1840
1841
                        num_common_prefix_blocks[kv_cache_gid],
                        attn_group.kv_cache_spec,
                        attn_group.get_metadata_builder(),
                    )
                cascade_attn_prefix_lens[kv_cache_gid].append(cascade_attn_prefix_len)
                use_cascade_attn |= cascade_attn_prefix_len > 0

        return cascade_attn_prefix_lens if use_cascade_attn else None
1842

1843
1844
1845
    def _compute_cascade_attn_prefix_len(
        self,
        num_scheduled_tokens: np.ndarray,
1846
        num_computed_tokens: np.ndarray,
1847
        num_common_prefix_blocks: int,
1848
1849
        kv_cache_spec: KVCacheSpec,
        attn_metadata_builder: AttentionMetadataBuilder,
1850
1851
1852
1853
1854
1855
1856
1857
1858
1859
1860
1861
1862
1863
1864
1865
1866
1867
    ) -> 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.
        """
1868

1869
        common_prefix_len = num_common_prefix_blocks * kv_cache_spec.block_size
1870
1871
1872
1873
1874
1875
1876
1877
1878
1879
1880
1881
1882
1883
1884
1885
1886
1887
1888
1889
1890
1891
1892
1893
1894
1895
1896
1897
1898
1899
1900
1901
1902
1903
1904
1905
1906
        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]
1907
        # Request 3's num_computed_tokens: 3 (i.e., [A, B, C])
1908
1909
1910
1911
1912
        # 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.
1913
        common_prefix_len = min(common_prefix_len, num_computed_tokens.min())
1914
        # common_prefix_len should be a multiple of the block size.
1915
1916
1917
1918
1919
1920
1921
1922
1923
1924
1925
        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
        )
1926
1927
        assert isinstance(kv_cache_spec, AttentionSpec)
        use_cascade = attn_metadata_builder.use_cascade_attention(
1928
1929
1930
            common_prefix_len=common_prefix_len,
            query_lens=num_scheduled_tokens,
            num_query_heads=self.num_query_heads,
1931
            num_kv_heads=kv_cache_spec.num_kv_heads,
1932
            use_alibi=self.use_alibi,
1933
            use_sliding_window=use_sliding_window,
1934
            use_local_attention=use_local_attention,
1935
            num_sms=self.num_sms,
1936
            dcp_world_size=self.dcp_world_size,
1937
1938
1939
        )
        return common_prefix_len if use_cascade else 0

1940
1941
    def _calc_mrope_positions(self, scheduler_output: "SchedulerOutput"):
        mrope_pos_ptr = 0
1942
        for index, req_id in enumerate(self.input_batch.req_ids):
1943
1944
1945
            req = self.requests[req_id]
            assert req.mrope_positions is not None

1946
1947
            num_computed_tokens = self.input_batch.num_computed_tokens_cpu[index]
            num_scheduled_tokens = scheduler_output.num_scheduled_tokens[req_id]
1948
            num_prompt_tokens = length_from_prompt_token_ids_or_embeds(
1949
1950
                req.prompt_token_ids, req.prompt_embeds
            )
1951
1952

            if num_computed_tokens + num_scheduled_tokens > num_prompt_tokens:
1953
1954
                prompt_part_len = max(0, num_prompt_tokens - num_computed_tokens)
                completion_part_len = max(0, num_scheduled_tokens - prompt_part_len)
1955
1956
1957
1958
1959
1960
1961
1962
1963
1964
1965
1966
1967
            else:
                prompt_part_len = num_scheduled_tokens
                completion_part_len = 0

            assert num_scheduled_tokens == prompt_part_len + completion_part_len

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

1968
1969
1970
                self.mrope_positions.cpu[:, dst_start:dst_end] = req.mrope_positions[
                    :, src_start:src_end
                ]
1971
1972
1973
1974
1975
1976
1977
                mrope_pos_ptr += prompt_part_len

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

1978
                assert req.mrope_position_delta is not None
1979
                MRotaryEmbedding.get_next_input_positions_tensor(
1980
                    out=self.mrope_positions.np,
1981
1982
1983
1984
1985
                    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,
                )
1986
1987
1988

                mrope_pos_ptr += completion_part_len

1989
1990
1991
1992
1993
1994
1995
1996
1997
1998
1999
2000
2001
2002
2003
2004
2005
2006
2007
2008
2009
2010
2011
2012
2013
2014
2015
2016
2017
2018
2019
2020
2021
2022
2023
2024
2025
2026
2027
2028
2029
2030
2031
2032
2033
2034
2035
    def _calc_xdrope_positions(self, scheduler_output: "SchedulerOutput"):
        xdrope_pos_ptr = 0
        for index, req_id in enumerate(self.input_batch.req_ids):
            req = self.requests[req_id]
            assert req.xdrope_positions is not None

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

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

            assert num_scheduled_tokens == prompt_part_len + completion_part_len

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

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

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

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

                xdrope_pos_ptr += completion_part_len

2036
2037
    def _calc_spec_decode_metadata(
        self,
2038
2039
2040
2041
2042
2043
2044
2045
2046
2047
2048
2049
2050
2051
2052
2053
        num_draft_tokens: np.ndarray,
        cu_num_scheduled_tokens: np.ndarray,
    ) -> SpecDecodeMetadata:
        # Inputs:
        # cu_num_scheduled_tokens:  [  4, 104, 107, 207, 209]
        # num_draft_tokens:         [  3,   0,   2,   0,   1]
        # Outputs:
        # cu_num_draft_tokens:      [  3,   3,   5,   5,   6]
        # logits_indices:           [  0,   1,   2,   3, 103, 104, 105, 106,
        #                            206, 207, 208]
        # target_logits_indices:    [  0,   1,   2,   5,   6,   9]
        # bonus_logits_indices:     [  3,   4,   7,   8,  10]

        # Compute the logits indices.
        # [4, 1, 3, 1, 2]
        num_sampled_tokens = num_draft_tokens + 1
2054
2055
2056
2057

        # 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(
2058
2059
            num_sampled_tokens, cumsum_dtype=np.int32
        )
2060
        # Step 2. [0, 0, 0, 0, 103, 104, 104, 104, 206, 207, 207]
2061
        logits_indices = np.repeat(
2062
2063
            cu_num_scheduled_tokens - num_sampled_tokens, num_sampled_tokens
        )
2064
        # Step 3. [0, 1, 2, 3, 103, 104, 105, 106, 206, 207, 208]
2065
2066
2067
2068
2069
2070
        logits_indices += arange

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

        # Compute the draft logits indices.
2071
2072
2073
        # 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(
2074
2075
            num_draft_tokens, cumsum_dtype=np.int32
        )
2076
2077
        # [0, 0, 0, 5, 5, 9]
        target_logits_indices = np.repeat(
2078
2079
            cu_num_sampled_tokens - num_sampled_tokens, num_draft_tokens
        )
2080
2081
2082
2083
2084
        # [0, 1, 2, 5, 6, 9]
        target_logits_indices += arange

        # TODO: Optimize the CPU -> GPU copy.
        cu_num_draft_tokens = torch.from_numpy(cu_num_draft_tokens).to(
2085
2086
            self.device, non_blocking=True
        )
2087
2088
2089
        cu_num_sampled_tokens = torch.from_numpy(cu_num_sampled_tokens).to(
            self.device, non_blocking=True
        )
2090
2091
2092
        logits_indices = torch.from_numpy(logits_indices).to(
            self.device, non_blocking=True
        )
2093
        target_logits_indices = torch.from_numpy(target_logits_indices).to(
2094
2095
            self.device, non_blocking=True
        )
2096
        bonus_logits_indices = torch.from_numpy(bonus_logits_indices).to(
2097
2098
            self.device, non_blocking=True
        )
2099

2100
2101
        # Compute the draft token ids.
        # draft_token_indices:      [  1,   2,   3, 105, 106, 208]
2102
        draft_token_ids = self.input_ids.gpu[logits_indices]
2103
2104
        draft_token_ids = draft_token_ids[target_logits_indices + 1]

2105
        return SpecDecodeMetadata(
2106
2107
2108
            draft_token_ids=draft_token_ids,
            num_draft_tokens=num_draft_tokens.tolist(),
            cu_num_draft_tokens=cu_num_draft_tokens,
2109
            cu_num_sampled_tokens=cu_num_sampled_tokens,
2110
2111
2112
2113
2114
            target_logits_indices=target_logits_indices,
            bonus_logits_indices=bonus_logits_indices,
            logits_indices=logits_indices,
        )

2115
2116
2117
2118
2119
2120
2121
    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
2122
        self.kv_sharing_fast_prefill_logits_indices[:num_logits].copy_(logits_indices)
2123
2124
2125
2126
2127
        # 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_(
2128
2129
2130
2131
2132
2133
            logits_indices[-1].item()
        )
        if (
            self.compilation_config.cudagraph_mode != CUDAGraphMode.NONE
            and num_logits <= self.cudagraph_batch_sizes[-1]
        ):
2134
2135
2136
2137
2138
            # Use piecewise CUDA graphs.
            # Add padding to the batch size.
            num_logits_padded = self.vllm_config.pad_for_cudagraph(num_logits)
        else:
            num_logits_padded = num_logits
2139
2140
2141
        logits_indices_padded = self.kv_sharing_fast_prefill_logits_indices[
            :num_logits_padded
        ]
2142
2143
        return logits_indices_padded

2144
    def _batch_mm_inputs_from_scheduler(
2145
2146
        self,
        scheduler_output: "SchedulerOutput",
2147
2148
2149
2150
2151
    ) -> tuple[
        list[str],
        list[MultiModalKwargsItem],
        list[tuple[str, PlaceholderRange]],
    ]:
2152
        """Batch multimodal inputs from scheduled encoder inputs.
2153
2154
2155

        Args:
            scheduler_output: The scheduler output containing scheduled encoder
2156
                inputs.
2157
2158

        Returns:
2159
            A tuple of (mm_hashes, mm_kwargs, mm_lora_refs) where:
2160
2161
            - mm_hashes: List of multimodal hashes for each item
            - mm_kwargs: List of multimodal kwargs for each item
2162
            - mm_lora_refs: List of (req_id, placeholder_range) for each item
2163
        """
2164
2165
        scheduled_encoder_inputs = scheduler_output.scheduled_encoder_inputs
        if not scheduled_encoder_inputs:
2166
            return [], [], []
2167
2168

        mm_hashes = list[str]()
2169
        mm_kwargs = list[MultiModalKwargsItem]()
2170
2171
2172
        # Multimodal LoRA reference info to map each multimodal item
        # back to its request & position
        mm_lora_refs = list[tuple[str, PlaceholderRange]]()
2173
2174
        for req_id, encoder_input_ids in scheduled_encoder_inputs.items():
            req_state = self.requests[req_id]
2175
2176

            for mm_input_id in encoder_input_ids:
2177
                mm_feature = req_state.mm_features[mm_input_id]
2178
2179
                if mm_feature.data is None:
                    continue
2180
2181

                mm_hashes.append(mm_feature.identifier)
2182
                mm_kwargs.append(mm_feature.data)
2183
                mm_lora_refs.append((req_id, mm_feature.mm_position))
2184

2185
        return mm_hashes, mm_kwargs, mm_lora_refs
2186

2187
2188
2189
    def _execute_mm_encoder(
        self, scheduler_output: "SchedulerOutput"
    ) -> list[torch.Tensor]:
2190
2191
2192
        mm_hashes, mm_kwargs, mm_lora_refs = self._batch_mm_inputs_from_scheduler(
            scheduler_output
        )
2193
2194

        if not mm_kwargs:
2195
            return []
2196

2197
2198
2199
2200
2201
2202
2203
        # 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.
2204
        model = cast(SupportsMultiModal, self.model)
2205
2206
2207
2208
2209
2210
2211
2212
2213
2214
2215
2216
2217
2218
2219
2220
2221
2222
2223
2224
2225
2226
2227
2228
2229
2230
2231
2232
2233
2234
2235
2236
2237
2238
2239
2240
2241
2242
2243
2244
2245
2246
2247
2248
2249
2250
2251
2252
2253
2254
2255
2256
2257
2258
2259
2260
2261

        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,
                )

2262
        encoder_outputs: list[torch.Tensor] = []
2263
        for modality, num_items, mm_kwargs_group in group_mm_kwargs_by_modality(
2264
2265
2266
            mm_kwargs,
            device=self.device,
            pin_memory=self.pin_memory,
2267
        ):
2268
            curr_group_outputs: MultiModalEmbeddings
2269
2270

            # EVS-related change.
2271
            # (ekhvedchenia): Temporary hack to limit peak memory usage when
2272
            # processing multimodal data. This solves the issue with scheduler
2273
2274
2275
2276
            # 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)
2277
2278
2279
2280
2281
2282
2283
            # 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
            ):
2284
                curr_group_outputs_lst = list[torch.Tensor]()
2285
2286
2287
2288
2289
2290
2291
2292
2293
                for video_mm_kwargs_item in filter(
                    lambda item: item.modality == "video", mm_kwargs
                ):
                    _, _, micro_batch_mm_inputs = next(
                        group_mm_kwargs_by_modality(
                            [video_mm_kwargs_item],
                            device=self.device,
                            pin_memory=self.pin_memory,
                        )
2294
                    )
2295

2296
                    micro_batch_outputs = model.embed_multimodal(
2297
2298
                        **micro_batch_mm_inputs
                    )
2299

2300
2301
2302
                    curr_group_outputs_lst.extend(micro_batch_outputs)

                curr_group_outputs = curr_group_outputs_lst
2303
2304
2305
2306
2307
2308
2309
2310
            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.
2311
                curr_group_outputs = model.embed_multimodal(**mm_kwargs_group)
2312

2313
2314
            sanity_check_mm_encoder_outputs(
                curr_group_outputs,
2315
                expected_num_items=num_items,
2316
            )
2317
            encoder_outputs.extend(curr_group_outputs)
2318

2319
        # Cache the encoder outputs by mm_hash
2320
        for mm_hash, output in zip(mm_hashes, encoder_outputs):
2321
            self.encoder_cache[mm_hash] = output
2322
2323
            logger.debug("Finish execute for mm hash %s", mm_hash)
            self.maybe_save_ec_to_connector(self.encoder_cache, mm_hash)
2324

2325
2326
        return encoder_outputs

2327
    def _gather_mm_embeddings(
2328
2329
        self,
        scheduler_output: "SchedulerOutput",
2330
        shift_computed_tokens: int = 0,
2331
2332
2333
2334
2335
2336
2337
2338
    ) -> tuple[list[torch.Tensor], torch.Tensor]:
        total_num_scheduled_tokens = scheduler_output.total_num_scheduled_tokens

        mm_embeds = list[torch.Tensor]()
        is_mm_embed = self.is_mm_embed.cpu
        is_mm_embed[:total_num_scheduled_tokens] = False

        req_start_idx = 0
2339
        should_sync_mrope_positions = False
2340
        should_sync_xdrope_positions = False
2341

2342
        for req_id in self.input_batch.req_ids:
2343
2344
            mm_embeds_req: list[torch.Tensor] = []

2345
            num_scheduled_tokens = scheduler_output.num_scheduled_tokens[req_id]
2346
            req_state = self.requests[req_id]
2347
            num_computed_tokens = req_state.num_computed_tokens + shift_computed_tokens
2348

2349
2350
            for mm_feature in req_state.mm_features:
                pos_info = mm_feature.mm_position
2351
2352
                start_pos = pos_info.offset
                num_encoder_tokens = pos_info.length
2353
2354
2355
2356
2357
2358
2359
2360
2361
2362
2363
2364
2365
2366
2367
2368

                # 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,
2369
2370
                    num_encoder_tokens,
                )
2371
                assert start_idx < end_idx
2372
2373
2374
2375
2376
2377
2378
                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
2379

2380
                mm_hash = mm_feature.identifier
2381
                encoder_output = self.encoder_cache.get(mm_hash, None)
2382
                assert encoder_output is not None, f"Encoder cache miss for {mm_hash}."
2383
2384
2385

                if (is_embed := pos_info.is_embed) is not None:
                    is_embed = is_embed[start_idx:end_idx]
2386
2387
2388
                    mm_embeds_item = encoder_output[curr_embeds_start:curr_embeds_end]
                else:
                    mm_embeds_item = encoder_output[start_idx:end_idx]
2389

2390
                req_start_pos = req_start_idx + start_pos - num_computed_tokens
2391
2392
2393
                is_mm_embed[req_start_pos + start_idx : req_start_pos + end_idx] = (
                    True if is_embed is None else is_embed
                )
2394
2395
2396
                mm_embeds_req.append(mm_embeds_item)

            if self.is_multimodal_pruning_enabled and self.uses_mrope:
2397
                assert req_state.mrope_positions is not None
2398
2399
2400
2401
2402
2403
2404
                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,
2405
2406
                    )
                )
2407
2408
2409
2410
                req_state.mrope_positions.copy_(new_mrope_positions)
                req_state.mrope_position_delta = new_delta

            mm_embeds.extend(mm_embeds_req)
2411
2412
2413
            req_start_idx += num_scheduled_tokens

        is_mm_embed = self.is_mm_embed.copy_to_gpu(total_num_scheduled_tokens)
2414
2415
2416

        if should_sync_mrope_positions:
            self._calc_mrope_positions(scheduler_output)
2417
            self.mrope_positions.copy_to_gpu(total_num_scheduled_tokens)
2418

2419
2420
2421
2422
        if should_sync_xdrope_positions:
            self._calc_xdrope_positions(scheduler_output)
            self.xdrope_positions.copy_to_gpu(total_num_scheduled_tokens)

2423
        return mm_embeds, is_mm_embed
2424

2425
    def get_model(self) -> nn.Module:
2426
        # get raw model out of the cudagraph wrapper.
2427
        if isinstance(self.model, (CUDAGraphWrapper, UBatchWrapper)):
2428
            return self.model.unwrap()
2429
2430
        return self.model

2431
2432
2433
2434
2435
2436
2437
2438
2439
2440
2441
2442
2443
2444
2445
    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

2446
2447
2448
2449
2450
    def get_supported_pooling_tasks(self) -> list[PoolingTask]:
        model = self.get_model()
        if not is_pooling_model(model):
            return []

2451
2452
        supported_tasks = list(model.pooler.get_supported_tasks())

2453
2454
2455
2456
        if "score" in supported_tasks:
            num_labels = getattr(self.model_config.hf_config, "num_labels", 0)
            if num_labels != 1:
                supported_tasks.remove("score")
2457
                logger.debug_once("Score API is only enabled for num_labels == 1.")
2458
2459

        return supported_tasks
2460

2461
2462
2463
2464
2465
2466
2467
2468
2469
2470
    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)

2471
    def sync_and_slice_intermediate_tensors(
2472
2473
        self,
        num_tokens: int,
2474
        intermediate_tensors: IntermediateTensors | None,
2475
2476
        sync_self: bool,
    ) -> IntermediateTensors:
2477
2478
2479
        assert self.intermediate_tensors is not None

        tp = self.vllm_config.parallel_config.tensor_parallel_size
2480
        is_rs = is_residual_scattered_for_sp(self.vllm_config, num_tokens)
2481
2482
2483
2484
2485
2486

        # 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():
2487
                is_scattered = k == "residual" and is_rs
2488
                copy_len = num_tokens // tp if is_scattered else num_tokens
2489
                self.intermediate_tensors[k][:copy_len].copy_(
2490
2491
2492
2493
2494
2495
2496
2497
2498
2499
2500
2501
2502
                    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:
2503
2504
2505
2506
2507
2508
2509
        """
        Step for the EPLB (Expert Parallelism Load Balancing) state.
        """
        if not self.parallel_config.enable_eplb:
            return

        assert self.eplb_state is not None
2510
2511
        model = self.get_model()
        assert is_mixture_of_experts(model)
2512
2513
2514
        self.eplb_state.step(
            is_dummy,
            is_profile,
2515
            log_stats=self.parallel_config.eplb_config.log_balancedness,
2516
2517
        )

2518
2519
2520
2521
2522
    def _pool(
        self,
        hidden_states: torch.Tensor,
        num_scheduled_tokens: int,
        num_scheduled_tokens_np: np.ndarray,
2523
2524
2525
2526
        kv_connector_output: KVConnectorOutput | None,
    ) -> ModelRunnerOutput | AsyncModelRunnerOutput:
        num_reqs = self.input_batch.num_reqs
        assert num_reqs == len(self.input_batch.pooling_params), (
2527
2528
            "Either all or none of the requests in a batch must be pooling request"
        )
2529

2530
        hidden_states = hidden_states[:num_scheduled_tokens]
2531
        seq_lens_cpu = self.seq_lens.cpu[:num_reqs]
2532

2533
        pooling_metadata = self.input_batch.get_pooling_metadata()
2534
        pooling_metadata.build_pooling_cursor(
2535
            num_scheduled_tokens_np, seq_lens_cpu, device=hidden_states.device
2536
        )
2537

2538
2539
2540
2541
2542
        model = cast(VllmModelForPooling, self.model)
        raw_pooler_output: PoolerOutput = model.pooler(
            hidden_states=hidden_states,
            pooling_metadata=pooling_metadata,
        )
2543
2544
2545
2546
2547
2548
2549
2550
2551
2552
2553
2554
2555
2556
2557
2558
2559
2560
2561
2562
2563
2564
2565
2566

        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,
            )

2567
        raw_pooler_output = json_map_leaves(
2568
            lambda x: None if x is None else x.to("cpu", non_blocking=True),
2569
2570
2571
            raw_pooler_output,
        )
        self._sync_device()
2572

2573
2574
2575
2576
2577
        model_runner_output.pooler_output = [
            out if include else None
            for out, include in zip(raw_pooler_output, finished_mask)
        ]
        return model_runner_output
2578

2579
    def _pad_for_sequence_parallelism(self, num_scheduled_tokens: int) -> int:
2580
2581
2582
        # Pad tokens to multiple of tensor_parallel_size when
        # enabled collective fusion for SP
        tp_size = self.vllm_config.parallel_config.tensor_parallel_size
2583
        if self.compilation_config.pass_config.enable_sp and tp_size > 1:
2584
2585
2586
            return round_up(num_scheduled_tokens, tp_size)
        return num_scheduled_tokens

Patrick von Platen's avatar
Patrick von Platen committed
2587
2588
2589
2590
2591
2592
2593
2594
2595
2596
2597
    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

2598
    def _preprocess(
2599
2600
        self,
        scheduler_output: "SchedulerOutput",
2601
        num_input_tokens: int,  # Padded
2602
        intermediate_tensors: IntermediateTensors | None = None,
2603
    ) -> tuple[
2604
2605
        torch.Tensor | None,
        torch.Tensor | None,
2606
        torch.Tensor,
2607
        IntermediateTensors | None,
2608
        dict[str, Any],
2609
        ECConnectorOutput | None,
2610
    ]:
2611
        num_scheduled_tokens = scheduler_output.total_num_scheduled_tokens
2612
        is_first_rank = get_pp_group().is_first_rank
2613
        is_encoder_decoder = self.model_config.is_encoder_decoder
2614

2615
2616
        # _prepare_inputs may reorder the batch, so we must gather multi
        # modal outputs after that to ensure the correct order
2617
2618
        ec_connector_output = None

2619
        if self.supports_mm_inputs and is_first_rank and not is_encoder_decoder:
2620
            # Run the multimodal encoder if any.
2621
2622
2623
2624
2625
2626
            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)
2627

2628
2629
2630
            # 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.
2631
            inputs_embeds_scheduled = self.model.embed_input_ids(
2632
2633
2634
                self.input_ids.gpu[:num_scheduled_tokens],
                multimodal_embeddings=mm_embeds,
                is_multimodal=is_mm_embed,
2635
            )
2636

2637
            # TODO(woosuk): Avoid the copy. Optimize.
2638
            self.inputs_embeds.gpu[:num_scheduled_tokens].copy_(inputs_embeds_scheduled)
2639

Patrick von Platen's avatar
Patrick von Platen committed
2640
            input_ids, inputs_embeds = self._prepare_mm_inputs(num_input_tokens)
2641
            model_kwargs = {
2642
                **self._init_model_kwargs(),
2643
2644
                **self._extract_mm_kwargs(scheduler_output),
            }
2645
        elif self.enable_prompt_embeds and is_first_rank:
2646
2647
2648
2649
2650
2651
2652
2653
2654
2655
2656
2657
            # 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).
2658
2659
2660
            token_ids_idx = (
                self.is_token_ids.gpu[:num_scheduled_tokens]
                .nonzero(as_tuple=False)
2661
                .squeeze(1)
2662
            )
2663
2664
2665
            # Some tokens ids may need to become embeds
            if token_ids_idx.numel() > 0:
                token_ids = self.input_ids.gpu[token_ids_idx]
2666
                tokens_to_embeds = self.model.embed_input_ids(input_ids=token_ids)
2667
2668
2669
                self.inputs_embeds.gpu[token_ids_idx] = tokens_to_embeds

            inputs_embeds = self.inputs_embeds.gpu[:num_input_tokens]
2670
            model_kwargs = self._init_model_kwargs()
2671
            input_ids = None
2672
        else:
2673
2674
2675
2676
            # 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.
2677
            input_ids = self.input_ids.gpu[:num_input_tokens]
2678
            inputs_embeds = None
2679
            model_kwargs = self._init_model_kwargs()
2680

2681
        if self.uses_mrope:
2682
            positions = self.mrope_positions.gpu[:, :num_input_tokens]
2683
2684
        elif self.uses_xdrope_dim > 0:
            positions = self.xdrope_positions.gpu[:, :num_input_tokens]
2685
        else:
2686
            positions = self.positions.gpu[:num_input_tokens]
2687

2688
        if is_first_rank:
2689
2690
            intermediate_tensors = None
        else:
2691
            assert intermediate_tensors is not None
2692
            intermediate_tensors = self.sync_and_slice_intermediate_tensors(
2693
2694
                num_input_tokens, intermediate_tensors, True
            )
2695

2696
        if is_encoder_decoder and scheduler_output.scheduled_encoder_inputs:
2697
2698
2699
2700
2701
2702
2703
            # 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})
2704

2705
2706
2707
2708
2709
2710
        return (
            input_ids,
            inputs_embeds,
            positions,
            intermediate_tensors,
            model_kwargs,
2711
            ec_connector_output,
2712
        )
2713

2714
    def _sample(
2715
        self,
2716
2717
        logits: torch.Tensor | None,
        spec_decode_metadata: SpecDecodeMetadata | None,
2718
    ) -> SamplerOutput:
2719
        # Sample the next token and get logprobs if needed.
2720
        sampling_metadata = self.input_batch.sampling_metadata
2721
        if spec_decode_metadata is None:
2722
2723
2724
            # 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()
2725
            return self.sampler(
2726
2727
2728
                logits=logits,
                sampling_metadata=sampling_metadata,
            )
2729

2730
        sampler_output = self.rejection_sampler(
2731
2732
            spec_decode_metadata,
            None,  # draft_probs
2733
            logits,
2734
2735
            sampling_metadata,
        )
2736
        self._update_states_after_model_execute(sampler_output.sampled_token_ids)
2737
2738
2739
        return sampler_output

    def _bookkeeping_sync(
2740
2741
2742
        self,
        scheduler_output: "SchedulerOutput",
        sampler_output: SamplerOutput,
2743
        logits: torch.Tensor | None,
2744
2745
        hidden_states: torch.Tensor,
        num_scheduled_tokens: int,
2746
        spec_decode_metadata: SpecDecodeMetadata | None,
2747
    ) -> tuple[
2748
        dict[str, int],
2749
        LogprobsLists | None,
2750
        list[list[int]],
2751
        dict[str, LogprobsTensors | None],
2752
2753
2754
        list[str],
        dict[str, int],
        list[int],
2755
    ]:
2756
2757
2758
2759
        num_nans_in_logits = {}
        if envs.VLLM_COMPUTE_NANS_IN_LOGITS:
            num_nans_in_logits = self._get_nans_in_logits(logits)

2760
2761
2762
2763
        num_reqs = self.input_batch.num_reqs
        discard_sampled_tokens_req_indices = np.nonzero(
            self.discard_request_mask.np[:num_reqs]
        )[0]
2764
2765
2766
2767
        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)
2768

2769
2770
2771
        # 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()
2772
        req_id_to_index_output_copy = self.input_batch.req_id_to_index.copy()
2773
2774

        num_sampled_tokens = sampler_output.sampled_token_ids.shape[0]
2775
        sampled_token_ids = sampler_output.sampled_token_ids
2776
        logprobs_tensors = sampler_output.logprobs_tensors
2777
        invalid_req_indices = []
2778
        cu_num_tokens: list[int] | None = None
2779
2780
2781
2782
2783
2784
        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)
2785
2786
2787
                # 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()
2788
2789
            else:
                # Includes spec decode tokens.
2790
                valid_sampled_token_ids, cu_num_tokens = RejectionSampler.parse_output(
2791
2792
                    sampled_token_ids,
                    self.input_batch.vocab_size,
2793
2794
                    discard_sampled_tokens_req_indices,
                    return_cu_num_tokens=logprobs_tensors is not None,
2795
                )
2796
        else:
2797
            valid_sampled_token_ids = []
2798
            invalid_req_indices = discard_sampled_tokens_req_indices.tolist()
2799
2800
2801
2802
2803
            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.
2804
2805
2806
2807
            # 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
2808
2809
2810
2811
2812
            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
            }
2813

2814
2815
2816
2817
2818
        # 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.
2819
        req_ids = self.input_batch.req_ids
2820
2821
        for req_idx in range(num_sampled_tokens):
            if self.use_async_scheduling:
2822
                sampled_ids = [-1] if req_idx not in invalid_req_indices_set else None
2823
2824
            else:
                sampled_ids = valid_sampled_token_ids[req_idx]
2825

2826
            num_sampled_ids: int = len(sampled_ids) if sampled_ids else 0
2827

2828
            if not sampled_ids:
2829
2830
2831
                continue

            start_idx = self.input_batch.num_tokens_no_spec[req_idx]
2832
            end_idx = start_idx + num_sampled_ids
2833
2834
2835
2836
            assert end_idx <= self.max_model_len, (
                "Sampled token IDs exceed the max model length. "
                f"Total number of tokens: {end_idx} > max_model_len: "
                f"{self.max_model_len}"
2837
            )
2838

2839
2840
            self.input_batch.token_ids_cpu[req_idx, start_idx:end_idx] = sampled_ids
            self.input_batch.is_token_ids[req_idx, start_idx:end_idx] = True
2841
            self.input_batch.num_tokens_no_spec[req_idx] = end_idx
2842

2843
            req_id = req_ids[req_idx]
2844
2845
2846
            req_state = self.requests[req_id]
            req_state.output_token_ids.extend(sampled_ids)

2847
        logprobs_lists = (
2848
            logprobs_tensors.tolists(cu_num_tokens)
2849
            if not self.use_async_scheduling and logprobs_tensors is not None
2850
2851
2852
2853
2854
2855
2856
2857
2858
            else None
        )

        # Compute prompt logprobs if needed.
        prompt_logprobs_dict = self._get_prompt_logprobs_dict(
            hidden_states[:num_scheduled_tokens],
            scheduler_output.num_scheduled_tokens,
        )

2859
2860
2861
2862
2863
2864
2865
2866
2867
2868
        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,
        )

2869
2870
2871
2872
2873
2874
2875
2876
2877
2878
2879
2880
2881
2882
2883
    @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()

2884
2885
    def _model_forward(
        self,
2886
2887
2888
2889
        input_ids: torch.Tensor | None = None,
        positions: torch.Tensor | None = None,
        intermediate_tensors: IntermediateTensors | None = None,
        inputs_embeds: torch.Tensor | None = None,
2890
2891
2892
2893
2894
        **model_kwargs: dict[str, Any],
    ) -> Any:
        """Helper method to call the model forward pass.

        This method can be overridden by subclasses for model execution.
2895
        Motivation: We can inspect only this method versus
2896
2897
2898
2899
2900
2901
2902
2903
2904
2905
2906
2907
2908
2909
2910
2911
2912
2913
2914
2915
        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,
        )

2916
2917
2918
2919
2920
2921
2922
2923
2924
2925
2926
2927
2928
2929
2930
2931
2932
2933
2934
2935
2936
    @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
        )

2937
2938
2939
2940
2941
2942
2943
2944
2945
2946
2947
2948
2949
    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,
2950
        num_encoder_reqs: int = 0,
2951
    ) -> tuple[
2952
2953
        CUDAGraphMode,
        BatchDescriptor,
2954
        bool,
2955
2956
        torch.Tensor | None,
        CUDAGraphStat | None,
2957
    ]:
2958
2959
2960
2961
2962
2963
        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,
2964
        )
2965
2966
2967
2968
2969
        # 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
        )
2970
2971
2972
2973
2974
2975
2976

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

2977
        num_tokens_padded = self._pad_for_sequence_parallelism(num_tokens)
2978
        dispatch_cudagraph = (
2979
            lambda num_tokens, disable_full: self.cudagraph_dispatcher.dispatch(
2980
2981
2982
                num_tokens=num_tokens,
                has_lora=has_lora,
                uniform_decode=uniform_decode,
2983
                disable_full=disable_full,
2984
2985
2986
2987
2988
            )
            if not force_eager
            else (CUDAGraphMode.NONE, BatchDescriptor(num_tokens_padded))
        )

2989
        cudagraph_mode, batch_descriptor = dispatch_cudagraph(
2990
            num_tokens_padded, use_cascade_attn or has_encoder_output
2991
        )
2992
        num_tokens_padded = batch_descriptor.num_tokens
2993
2994
2995
2996
2997
2998
2999
3000
3001
        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"
            )
3002
3003
3004

        # Extra coordination when running data-parallel since we need to coordinate
        # across ranks
3005
        should_ubatch, num_tokens_across_dp = False, None
3006
3007
3008
3009
3010
3011
3012
3013
3014
        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
            )

3015
3016
3017
3018
3019
3020
3021
3022
3023
3024
3025
            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,
                )
3026
3027
            )

3028
            # Extract DP-synced values
3029
3030
3031
            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())
3032
3033
3034
3035
3036
                # 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,
                )
3037
3038
3039
3040
                # 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

3041
3042
3043
3044
3045
3046
3047
3048
3049
3050
3051
3052
        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,
3053
            should_ubatch,
3054
3055
3056
            num_tokens_across_dp,
            cudagraph_stats,
        )
3057

3058
3059
3060
3061
3062
3063
3064
3065
3066
3067
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
3093
    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

3094
3095
3096
3097
    @torch.inference_mode()
    def execute_model(
        self,
        scheduler_output: "SchedulerOutput",
3098
        intermediate_tensors: IntermediateTensors | None = None,
3099
    ) -> ModelRunnerOutput | AsyncModelRunnerOutput | IntermediateTensors | None:
3100
3101
3102
3103
3104
        if self.execute_model_state is not None:
            raise RuntimeError(
                "State error: sample_tokens() must be called "
                "after execute_model() returns None."
            )
3105
3106
3107
3108
3109
3110
3111
3112
3113
3114
3115
3116
3117
3118
3119

        # self._draft_token_ids is None when `input_fits_in_drafter=False`
        # and there is no draft tokens scheduled. so it need to update the
        # spec_decoding info in scheduler_output with async_scheduling.
        # use deepcopy to avoid the modification has influence on the
        # scheduler_output in engine core process.
        # TODO(Ronald1995): deepcopy is expensive when there is a large
        # number of requests, optimize it later.
        if (
            self.use_async_scheduling
            and self.num_spec_tokens
            and self._draft_token_ids is None
        ):
            scheduler_output = deepcopy(scheduler_output)

3120
        num_scheduled_tokens = scheduler_output.total_num_scheduled_tokens
3121
3122
3123
3124
3125
3126
        with (
            record_function_or_nullcontext("gpu_model_runner: preprocess"),
            self.synchronize_input_prep(),
        ):
            # Update persistent batch states.
            self._update_states(scheduler_output)
3127

3128
3129
            if has_ec_transfer() and get_ec_transfer().is_producer:
                with self.maybe_get_ec_connector_output(
3130
                    scheduler_output,
3131
3132
3133
3134
3135
3136
3137
3138
3139
3140
3141
3142
3143
3144
3145
3146
3147
3148
3149
3150
3151
3152
3153
3154
3155
3156
3157
3158
                    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"
3159
3160
                )

3161
3162
3163
3164
3165
3166
            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
3167

3168
3169
3170
3171
            logits_indices, spec_decode_metadata = self._prepare_inputs(
                scheduler_output,
                num_scheduled_tokens_np,
            )
3172

3173
3174
3175
3176
3177
            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(
3178
                    num_scheduled_tokens_np,
3179
3180
                    self.input_batch.num_computed_tokens_cpu[:num_reqs],
                    scheduler_output.num_common_prefix_blocks,
3181
3182
                )

3183
3184
3185
3186
3187
3188
3189
3190
3191
3192
3193
3194
3195
3196
            (
                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),
            )
3197

3198
3199
3200
3201
3202
3203
3204
3205
3206
3207
3208
3209
3210
3211
3212
3213
3214
3215
3216
3217
3218
3219
3220
3221
3222
3223
3224
3225
3226
3227
3228
3229
3230
3231
3232
3233
3234
3235
3236
3237
3238
3239
3240
3241
            logger.debug(
                "Running batch with cudagraph_mode: %s, batch_descriptor: %s, "
                "should_ubatch: %s, num_tokens_across_dp: %s",
                cudagraph_mode,
                batch_desc,
                should_ubatch,
                num_tokens_across_dp,
            )

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

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

            pad_attn = cudagraph_mode == CUDAGraphMode.FULL

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

            attn_metadata, spec_decode_common_attn_metadata = (
                self._build_attention_metadata(
                    num_tokens=num_tokens_unpadded,
                    num_tokens_padded=num_tokens_padded if pad_attn else None,
                    num_reqs=num_reqs,
                    num_reqs_padded=num_reqs_padded if pad_attn else None,
                    max_query_len=max_num_scheduled_tokens,
                    ubatch_slices=ubatch_slices_attn,
                    logits_indices=logits_indices,
                    use_spec_decode=use_spec_decode,
                    num_scheduled_tokens=scheduler_output.num_scheduled_tokens,
                    cascade_attn_prefix_lens=cascade_attn_prefix_lens,
3242
                )
3243
            )
3244

3245
3246
3247
3248
3249
3250
3251
3252
3253
            (
                input_ids,
                inputs_embeds,
                positions,
                intermediate_tensors,
                model_kwargs,
                ec_connector_output,
            ) = self._preprocess(
                scheduler_output, num_tokens_padded, intermediate_tensors
3254
            )
3255

3256
        # Set cudagraph mode to none if calc_kv_scales is true.
3257
3258
3259
        # KV scales calculation involves dynamic operations that are incompatible
        # with CUDA graph capture.
        if self.calculate_kv_scales:
3260
            cudagraph_mode = CUDAGraphMode.NONE
3261
3262
            # Mark KV scales as calculated after the first forward pass
            self.calculate_kv_scales = False
3263

3264
3265
        # Run the model.
        # Use persistent buffers for CUDA graphs.
3266
3267
        with (
            set_forward_context(
3268
3269
                attn_metadata,
                self.vllm_config,
3270
                num_tokens=num_tokens_padded,
3271
                num_tokens_across_dp=num_tokens_across_dp,
3272
3273
                cudagraph_runtime_mode=cudagraph_mode,
                batch_descriptor=batch_desc,
3274
                ubatch_slices=ubatch_slices_padded,
3275
            ),
3276
            record_function_or_nullcontext("gpu_model_runner: forward"),
3277
3278
            self.maybe_get_kv_connector_output(scheduler_output) as kv_connector_output,
        ):
3279
            model_output = self._model_forward(
3280
3281
3282
3283
3284
3285
3286
                input_ids=input_ids,
                positions=positions,
                intermediate_tensors=intermediate_tensors,
                inputs_embeds=inputs_embeds,
                **model_kwargs,
            )

3287
        with record_function_or_nullcontext("gpu_model_runner: postprocess"):
3288
            if self.use_aux_hidden_state_outputs:
3289
                # True when EAGLE 3 is used.
3290
3291
                hidden_states, aux_hidden_states = model_output
            else:
3292
                # Common case.
3293
3294
3295
                hidden_states = model_output
                aux_hidden_states = None

3296
3297
3298
3299
3300
            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)
3301
                    hidden_states.kv_connector_output = kv_connector_output
3302
                    self.kv_connector_output = kv_connector_output
3303
                    return hidden_states
3304

3305
                if self.is_pooling_model:
3306
                    # Return the pooling output.
3307
3308
3309
3310
3311
                    return self._pool(
                        hidden_states,
                        num_scheduled_tokens,
                        num_scheduled_tokens_np,
                        kv_connector_output,
3312
                    )
3313
3314

                sample_hidden_states = hidden_states[logits_indices]
3315
                logits = self.model.compute_logits(sample_hidden_states)
3316
3317
3318
3319
            else:
                # Rare case.
                assert not self.is_pooling_model

3320
                sample_hidden_states = hidden_states[logits_indices]
3321
                if not get_pp_group().is_last_rank:
3322
                    all_gather_tensors = {
3323
                        "residual": not is_residual_scattered_for_sp(
3324
                            self.vllm_config, num_tokens_padded
3325
                        )
3326
                    }
3327
                    get_pp_group().send_tensor_dict(
3328
3329
                        hidden_states.tensors,
                        all_gather_group=get_tp_group(),
3330
3331
                        all_gather_tensors=all_gather_tensors,
                    )
3332
3333
                    logits = None
                else:
3334
                    logits = self.model.compute_logits(sample_hidden_states)
3335

3336
                model_output_broadcast_data: dict[str, Any] = {}
3337
3338
3339
                if logits is not None:
                    model_output_broadcast_data["logits"] = logits.contiguous()

3340
                broadcasted = get_pp_group().broadcast_tensor_dict(
3341
3342
                    model_output_broadcast_data, src=len(get_pp_group().ranks) - 1
                )
3343
3344
                assert broadcasted is not None
                logits = broadcasted["logits"]
3345

3346
3347
3348
3349
3350
3351
3352
3353
        self.execute_model_state = ExecuteModelState(
            scheduler_output,
            logits,
            spec_decode_metadata,
            spec_decode_common_attn_metadata,
            hidden_states,
            sample_hidden_states,
            aux_hidden_states,
3354
            ec_connector_output,
3355
            cudagraph_stats,
3356
        )
3357
        self.kv_connector_output = kv_connector_output
3358
3359
3360
3361
3362
3363
        return None

    @torch.inference_mode
    def sample_tokens(
        self, grammar_output: "GrammarOutput | None"
    ) -> ModelRunnerOutput | AsyncModelRunnerOutput | IntermediateTensors:
3364
3365
3366
        kv_connector_output = self.kv_connector_output
        self.kv_connector_output = None

3367
3368
        if self.execute_model_state is None:
            # Nothing to do (PP non-final rank case), output isn't used.
3369
            if not kv_connector_output:
3370
                return None  # type: ignore[return-value]
3371
3372
3373
3374
3375
3376
3377
3378
3379

            # 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
3380
3381
3382
3383
3384
3385
3386
3387
3388
3389

        # Unpack ephemeral state.
        (
            scheduler_output,
            logits,
            spec_decode_metadata,
            spec_decode_common_attn_metadata,
            hidden_states,
            sample_hidden_states,
            aux_hidden_states,
3390
            ec_connector_output,
3391
            cudagraph_stats,
3392
3393
3394
3395
3396
3397
3398
3399
3400
        ) = 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
            )
3401

3402
        with record_function_or_nullcontext("gpu_model_runner: sample"):
3403
3404
            sampler_output = self._sample(logits, spec_decode_metadata)

3405
3406
        self.input_batch.prev_sampled_token_ids = None

3407
        def propose_draft_token_ids(sampled_token_ids):
3408
            assert spec_decode_common_attn_metadata is not None
3409
            with record_function_or_nullcontext("gpu_model_runner: draft"):
3410
3411
3412
3413
3414
3415
3416
3417
3418
3419
3420
                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,
                )

3421
        spec_config = self.speculative_config
3422
        use_padded_batch_for_eagle = (
3423
3424
3425
            spec_config is not None
            and spec_config.use_eagle()
            and not spec_config.disable_padded_drafter_batch
3426
        )
3427
3428
3429
        effective_drafter_max_model_len = self.max_model_len
        if effective_drafter_max_model_len is None:
            effective_drafter_max_model_len = self.model_config.max_model_len
3430
        if (
3431
3432
3433
            spec_config is not None
            and spec_config.draft_model_config is not None
            and spec_config.draft_model_config.max_model_len is not None
3434
        ):
3435
            effective_drafter_max_model_len = (
3436
                spec_config.draft_model_config.max_model_len
3437
            )
3438
        input_fits_in_drafter = spec_decode_common_attn_metadata and (
3439
            spec_decode_common_attn_metadata.max_seq_len + self.num_spec_tokens
3440
3441
            <= effective_drafter_max_model_len
        )
3442
        if use_padded_batch_for_eagle:
3443
3444
            assert self.speculative_config is not None
            assert isinstance(self.drafter, EagleProposer)
3445
3446
3447
3448
3449
3450
            sampled_token_ids = sampler_output.sampled_token_ids
            if input_fits_in_drafter:
                # EAGLE speculative decoding can use the GPU sampled tokens
                # as inputs, and does not need to wait for bookkeeping to finish.
                propose_draft_token_ids(sampled_token_ids)
            elif self.valid_sampled_token_count_event is not None:
3451
                assert spec_decode_common_attn_metadata is not None
3452
3453
3454
3455
3456
3457
                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,
3458
                        self.discard_request_mask.gpu,
3459
3460
3461
3462
3463
                    )
                )
                self._copy_valid_sampled_token_count(
                    next_token_ids, valid_sampled_tokens_count
                )
3464

3465
        with record_function_or_nullcontext("gpu_model_runner: bookkeep"):
3466
3467
3468
3469
3470
3471
3472
3473
            (
                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,
3474
3475
3476
3477
3478
            ) = self._bookkeeping_sync(
                scheduler_output,
                sampler_output,
                logits,
                hidden_states,
3479
                scheduler_output.total_num_scheduled_tokens,
3480
                spec_decode_metadata,
3481
            )
3482

3483
3484
3485
3486
3487
        if (
            self.speculative_config
            and not use_padded_batch_for_eagle
            and input_fits_in_drafter
        ):
3488
3489
3490
            # 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)
3491

3492
        with record_function_or_nullcontext("gpu_model_runner: eplb"):
3493
            self.eplb_step()
3494
3495
3496
3497
3498
3499
3500
3501
        with record_function_or_nullcontext("gpu_model_runner: ModelRunnerOutput"):
            output = ModelRunnerOutput(
                req_ids=req_ids_output_copy,
                req_id_to_index=req_id_to_index_output_copy,
                sampled_token_ids=valid_sampled_token_ids,
                logprobs=logprobs_lists,
                prompt_logprobs_dict=prompt_logprobs_dict,
                kv_connector_output=kv_connector_output,
3502
3503
3504
                ec_connector_output=ec_connector_output
                if self.supports_mm_inputs
                else None,
3505
                num_nans_in_logits=num_nans_in_logits,
3506
                cudagraph_stats=cudagraph_stats,
3507
            )
3508

3509
3510
        if not self.use_async_scheduling:
            return output
3511
3512
3513
3514
3515
3516
3517
3518
3519
        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,
3520
                vocab_size=self.input_batch.vocab_size,
3521
3522
3523
3524
3525
            )
        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
3526
            # any requests with sampling params that require output ids.
3527
3528
3529
3530
            self.input_batch.set_async_sampled_token_ids(
                async_output.sampled_token_ids_cpu,
                async_output.async_copy_ready_event,
            )
3531
3532
3533

        return async_output

3534
    def take_draft_token_ids(self) -> DraftTokenIds | None:
3535
        if not self.num_spec_tokens:
3536
            return None
3537

3538
        req_ids = self.input_batch.req_ids
3539
3540
3541
        if self._draft_token_ids is None:
            return DraftTokenIds(req_ids, [[] for _ in req_ids])

3542
3543
3544
3545
3546
3547
3548
        if isinstance(self._draft_token_ids, torch.Tensor):
            draft_token_ids = self._draft_token_ids.tolist()
        else:
            draft_token_ids = self._draft_token_ids
        self._draft_token_ids = None
        return DraftTokenIds(req_ids, draft_token_ids)

3549
3550
3551
3552
3553
3554
3555
3556
3557
3558
3559
3560
3561
3562
3563
3564
3565
3566
3567
3568
3569
3570
3571
3572
3573
3574
3575
3576
3577
3578
3579
    def _copy_valid_sampled_token_count(
        self, next_token_ids: torch.Tensor, valid_sampled_tokens_count: torch.Tensor
    ) -> None:
        if self.valid_sampled_token_count_event is None:
            return

        default_stream = torch.cuda.current_stream()
        # Initialize a new stream to overlap the copy operation with
        # prepare_input of draft model.
        with torch.cuda.stream(self.valid_sampled_token_count_copy_stream):
            self.valid_sampled_token_count_copy_stream.wait_stream(default_stream)  # type: ignore
            counts = valid_sampled_tokens_count
            counts_cpu = self.valid_sampled_token_count_cpu
            counts_cpu[: counts.shape[0]].copy_(counts, non_blocking=True)
            self.valid_sampled_token_count_event.record()

        self.input_batch.prev_sampled_token_ids = next_token_ids.unsqueeze(1)

    def _get_valid_sampled_token_count(self) -> list[int]:
        # Wait until valid_sampled_tokens_count is copied to cpu,
        prev_sampled_token_ids = self.input_batch.prev_sampled_token_ids
        if (
            self.valid_sampled_token_count_event is None
            or prev_sampled_token_ids is None
        ):
            return []

        counts_cpu = self.valid_sampled_token_count_cpu
        self.valid_sampled_token_count_event.synchronize()
        return counts_cpu[: prev_sampled_token_ids.shape[0]].tolist()

3580
3581
3582
    def propose_draft_token_ids(
        self,
        scheduler_output: "SchedulerOutput",
3583
        sampled_token_ids: torch.Tensor | list[list[int]],
3584
3585
3586
        sampling_metadata: SamplingMetadata,
        hidden_states: torch.Tensor,
        sample_hidden_states: torch.Tensor,
3587
3588
        aux_hidden_states: list[torch.Tensor] | None,
        spec_decode_metadata: SpecDecodeMetadata | None,
3589
        common_attn_metadata: CommonAttentionMetadata,
3590
    ) -> list[list[int]] | torch.Tensor:
3591
        num_scheduled_tokens = scheduler_output.total_num_scheduled_tokens
3592
3593
3594
        spec_config = self.speculative_config
        assert spec_config is not None
        if spec_config.method == "ngram":
3595
            assert isinstance(sampled_token_ids, list)
3596
            assert isinstance(self.drafter, NgramProposer)
3597
            draft_token_ids = self.drafter.propose(
3598
3599
                sampled_token_ids,
                self.input_batch.req_ids,
3600
3601
                self.input_batch.num_tokens_no_spec,
                self.input_batch.token_ids_cpu,
3602
3603
                self.input_batch.spec_decode_unsupported_reqs,
            )
3604
        elif spec_config.method == "suffix":
3605
3606
3607
            assert isinstance(sampled_token_ids, list)
            assert isinstance(self.drafter, SuffixDecodingProposer)
            draft_token_ids = self.drafter.propose(self.input_batch, sampled_token_ids)
3608
        elif spec_config.method == "medusa":
3609
            assert isinstance(sampled_token_ids, list)
3610
            assert isinstance(self.drafter, MedusaProposer)
3611

3612
3613
            if sample_hidden_states.shape[0] == len(sampled_token_ids):
                # The input to the target model does not include draft tokens.
3614
3615
3616
3617
                hidden_states = sample_hidden_states
            else:
                indices = []
                offset = 0
3618
3619
3620
                assert spec_decode_metadata is not None, (
                    "No spec decode metadata for medusa"
                )
3621
                for num_draft, tokens in zip(
3622
3623
                    spec_decode_metadata.num_draft_tokens, sampled_token_ids
                ):
3624
                    indices.append(offset + len(tokens) - 1)
3625
                    offset += num_draft + 1
3626
                indices = torch.tensor(indices, device=self.device)
3627
3628
                hidden_states = sample_hidden_states[indices]

3629
            draft_token_ids = self.drafter.propose(
3630
3631
3632
                target_hidden_states=hidden_states,
                sampling_metadata=sampling_metadata,
            )
3633
        elif spec_config.use_eagle():
3634
            assert isinstance(self.drafter, EagleProposer)
3635

3636
            if spec_config.disable_padded_drafter_batch:
3637
3638
3639
                # 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.
3640
3641
                assert isinstance(sampled_token_ids, list), (
                    "sampled_token_ids should be a python list when"
3642
                    "padded-batch is disabled."
3643
                )
3644
                next_token_ids = self.drafter.prepare_next_token_ids_cpu(
3645
3646
3647
3648
3649
                    sampled_token_ids,
                    self.requests,
                    self.input_batch,
                    scheduler_output.num_scheduled_tokens,
                )
3650
3651
3652
3653
3654
            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.
3655
3656
                assert isinstance(sampled_token_ids, torch.Tensor), (
                    "sampled_token_ids should be a torch.Tensor when"
3657
                    "padded-batch is enabled."
3658
3659
                )
                next_token_ids, valid_sampled_tokens_count = (
3660
3661
3662
3663
3664
                    self.drafter.prepare_next_token_ids_padded(
                        common_attn_metadata,
                        sampled_token_ids,
                        self.requests,
                        self.input_batch,
3665
                        self.discard_request_mask.gpu,
3666
                    )
3667
                )
3668
3669
3670
                self._copy_valid_sampled_token_count(
                    next_token_ids, valid_sampled_tokens_count
                )
Jiayi Yao's avatar
Jiayi Yao committed
3671

3672
            num_rejected_tokens_gpu = None
3673
            if spec_decode_metadata is None:
3674
                token_indices_to_sample = None
3675
                # input_ids can be None for multimodal models.
3676
                target_token_ids = self.input_ids.gpu[:num_scheduled_tokens]
3677
                target_positions = self._get_positions(num_scheduled_tokens)
3678
                if self.use_aux_hidden_state_outputs:
Wentao Ye's avatar
Wentao Ye committed
3679
                    assert aux_hidden_states is not None
3680
                    target_hidden_states = torch.cat(
3681
3682
                        [h[:num_scheduled_tokens] for h in aux_hidden_states], dim=-1
                    )
3683
3684
                else:
                    target_hidden_states = hidden_states[:num_scheduled_tokens]
3685
            else:
3686
                if spec_config.disable_padded_drafter_batch:
3687
                    token_indices_to_sample = None
3688
3689
3690
3691
3692
                    common_attn_metadata, token_indices = self.drafter.prepare_inputs(
                        common_attn_metadata,
                        sampled_token_ids,
                        spec_decode_metadata.num_draft_tokens,
                    )
3693
3694
3695
3696
3697
3698
3699
3700
3701
                    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]
3702
                else:
3703
3704
3705
3706
3707
3708
3709
3710
                    (
                        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,
3711
                    )
3712
3713
3714
3715
3716
3717
3718
3719
3720
3721
3722
                    total_num_tokens = common_attn_metadata.num_actual_tokens
                    # When padding the batch, token_indices is just a range
                    target_token_ids = self.input_ids.gpu[:total_num_tokens]
                    target_positions = self._get_positions(total_num_tokens)
                    if self.use_aux_hidden_state_outputs:
                        assert aux_hidden_states is not None
                        target_hidden_states = torch.cat(
                            [h[:total_num_tokens] for h in aux_hidden_states], dim=-1
                        )
                    else:
                        target_hidden_states = hidden_states[:total_num_tokens]
3723

3724
            if self.supports_mm_inputs:
3725
3726
3727
3728
3729
3730
                mm_embed_inputs = self._gather_mm_embeddings(
                    scheduler_output,
                    shift_computed_tokens=1,
                )
            else:
                mm_embed_inputs = None
3731

3732
            draft_token_ids = self.drafter.propose(
3733
3734
3735
3736
                target_token_ids=target_token_ids,
                target_positions=target_positions,
                target_hidden_states=target_hidden_states,
                next_token_ids=next_token_ids,
3737
                last_token_indices=token_indices_to_sample,
3738
                sampling_metadata=sampling_metadata,
3739
                common_attn_metadata=common_attn_metadata,
3740
                mm_embed_inputs=mm_embed_inputs,
3741
                num_rejected_tokens_gpu=num_rejected_tokens_gpu,
3742
            )
3743

3744
        return draft_token_ids
3745

3746
3747
3748
    def update_config(self, overrides: dict[str, Any]) -> None:
        allowed_config_names = {"load_config", "model_config"}
        for config_name, config_overrides in overrides.items():
3749
3750
            assert config_name in allowed_config_names, (
                f"Config `{config_name}` not supported. "
3751
                f"Allowed configs: {allowed_config_names}"
3752
            )
3753
3754
3755
3756
            config = getattr(self, config_name)
            new_config = update_config(config, config_overrides)
            setattr(self, config_name, new_config)

3757
3758
3759
3760
3761
    def load_model(self, eep_scale_up: bool = False) -> None:
        """
        Args:
            eep_scale_up: the model loading is for elastic EP scale up.
        """
3762
3763
3764
3765
3766
        logger.info_once(
            "Starting to load model %s...",
            self.model_config.model,
            scope="global",
        )
3767
3768
3769
3770
3771
        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)
        )
3772

3773
3774
3775
3776
        if self.parallel_config.enable_eplb:
            self.eplb_state = EplbState(self.parallel_config, self.device)
            eplb_models = 0

3777
3778
3779
3780
3781
3782
3783
3784
3785
3786
        try:
            with DeviceMemoryProfiler() as m:
                time_before_load = time.perf_counter()
                model_loader = get_model_loader(self.load_config)
                self.model = model_loader.load_model(
                    vllm_config=self.vllm_config, model_config=self.model_config
                )
                if self.lora_config:
                    self.model = self.load_lora_model(
                        self.model, self.vllm_config, self.device
3787
                    )
3788
3789
3790
3791
3792
3793
3794
3795
3796
3797
3798
3799
3800
3801
3802
                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,
                        )
3803

3804
3805
3806
3807
3808
3809
3810
3811
3812
3813
3814
3815
3816
3817
3818
3819
3820
3821
3822
3823
3824
3825
                        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
3826

3827
3828
3829
3830
3831
3832
                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"
                        )
3833

3834
3835
3836
3837
3838
3839
3840
3841
3842
3843
3844
3845
3846
3847
3848
3849
3850
3851
3852
3853
3854
3855
3856
3857
3858
                    # Try to get auxiliary layers from speculative config,
                    # otherwise use model's default layers
                    aux_layers = self._get_eagle3_aux_layers_from_config()
                    if aux_layers:
                        logger.info(
                            "Using auxiliary layers from speculative config: %s",
                            aux_layers,
                        )
                    else:
                        aux_layers = self.model.get_eagle3_aux_hidden_state_layers()

                    self.model.set_aux_hidden_state_layers(aux_layers)
                time_after_load = time.perf_counter()
            self.model_memory_usage = m.consumed_memory
        except torch.cuda.OutOfMemoryError as e:
            msg = (
                "Failed to load model - not enough GPU memory. "
                "Try lowering --gpu-memory-utilization to free memory for weights, "
                "increasing --tensor-parallel-size, or using --quantization. "
                "See https://docs.vllm.ai/en/latest/configuration/conserving_memory/ "
                "for more tips."
            )
            combined_msg = f"{msg} (original error: {e})"
            logger.error(combined_msg)
            raise e
3859
        logger.info_once(
3860
            "Model loading took %.4f GiB memory and %.6f seconds",
3861
3862
            self.model_memory_usage / GiB_bytes,
            time_after_load - time_before_load,
3863
            scope="local",
3864
        )
3865
        prepare_communication_buffer_for_model(self.model)
3866
3867
3868
3869
        if (drafter := getattr(self, "drafter", None)) and (
            drafter_model := getattr(drafter, "model", None)
        ):
            prepare_communication_buffer_for_model(drafter_model)
3870
        mm_config = self.model_config.multimodal_config
3871
        self.is_multimodal_pruning_enabled = (
3872
            supports_multimodal_pruning(self.get_model())
3873
3874
            and mm_config is not None
            and mm_config.is_multimodal_pruning_enabled()
3875
        )
3876

3877
        if is_mixture_of_experts(self.model) and self.parallel_config.enable_eplb:
3878
3879
3880
3881
3882
3883
3884
3885
3886
3887
3888
            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(
3889
                self.model,
3890
                self.model_config,
3891
3892
3893
                global_expert_load,
                old_global_expert_indices,
                rank_mapping,
3894
            )
3895
3896
            if self.eplb_state.is_async:
                self.eplb_state.start_async_loop(rank_mapping=rank_mapping)
3897

3898
        if (
3899
3900
            self.vllm_config.compilation_config.mode
            == CompilationMode.STOCK_TORCH_COMPILE
3901
            and supports_dynamo()
3902
        ):
3903
            backend = self.vllm_config.compilation_config.init_backend(self.vllm_config)
3904
            compilation_counter.stock_torch_compile_count += 1
3905
            self.model.compile(fullgraph=True, backend=backend)
3906
            return
3907
        # for other compilation modes, cudagraph behavior is controlled by
3908
3909
3910
        # CudagraphWraper and CudagraphDispatcher of vllm.

        # wrap the model with full cudagraph wrapper if needed.
3911
3912
        cudagraph_mode = self.compilation_config.cudagraph_mode
        assert cudagraph_mode is not None
3913
3914
3915
3916
        if (
            cudagraph_mode.has_full_cudagraphs()
            and not self.parallel_config.use_ubatching
        ):
3917
3918
3919
            self.model = CUDAGraphWrapper(
                self.model, self.vllm_config, runtime_mode=CUDAGraphMode.FULL
            )
3920
        elif self.parallel_config.use_ubatching:
3921
            if cudagraph_mode.has_full_cudagraphs():
3922
3923
3924
                self.model = UBatchWrapper(
                    self.model, self.vllm_config, CUDAGraphMode.FULL, self.device
                )
3925
            else:
3926
3927
3928
                self.model = UBatchWrapper(
                    self.model, self.vllm_config, CUDAGraphMode.NONE, self.device
                )
3929

3930
    def _get_eagle3_aux_layers_from_config(self) -> tuple[int, ...] | None:
3931
3932
3933
3934
3935
3936
3937
3938
3939
3940
3941
3942
3943
3944
3945
3946
3947
3948
3949
3950
3951
3952
3953
        """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

3954
    def reload_weights(self) -> None:
3955
        assert getattr(self, "model", None) is not None, (
3956
            "Cannot reload weights before model is loaded."
3957
        )
3958
3959
        model_loader = get_model_loader(self.load_config)
        logger.info("Reloading weights inplace...")
3960
        model_loader.load_weights(self.get_model(), model_config=self.model_config)
3961

3962
3963
3964
3965
3966
    def save_tensorized_model(
        self,
        tensorizer_config: "TensorizerConfig",
    ) -> None:
        TensorizerLoader.save_model(
3967
            self.get_model(),
3968
            tensorizer_config=tensorizer_config,
3969
            model_config=self.model_config,
3970
3971
        )

3972
3973
3974
    def _get_prompt_logprobs_dict(
        self,
        hidden_states: torch.Tensor,
3975
        num_scheduled_tokens: dict[str, int],
3976
    ) -> dict[str, LogprobsTensors | None]:
3977
        num_prompt_logprobs_dict = self.num_prompt_logprobs
3978
3979
3980
        if not num_prompt_logprobs_dict:
            return {}

3981
        in_progress_dict = self.input_batch.in_progress_prompt_logprobs_cpu
3982
        prompt_logprobs_dict: dict[str, LogprobsTensors | None] = {}
3983
3984
3985
3986
3987

        # 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():
3988
3989
3990
3991
            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
3992
3993
3994

            # Get metadata for this request.
            request = self.requests[req_id]
3995
3996
3997
3998
            if request.prompt_token_ids is None:
                # Prompt logprobs is incompatible with prompt embeddings
                continue

3999
4000
            num_prompt_tokens = len(request.prompt_token_ids)
            prompt_token_ids = torch.tensor(request.prompt_token_ids).to(
4001
4002
                self.device, non_blocking=True
            )
4003

4004
4005
4006
4007
4008
4009
            # 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(
4010
4011
                    num_prompt_tokens - 1, num_prompt_logprobs + 1
                )
4012
4013
                in_progress_dict[req_id] = logprobs_tensors

4014
            # Determine number of logits to retrieve.
4015
4016
            start_idx = request.num_computed_tokens
            start_tok = start_idx + 1
4017
            num_remaining_tokens = num_prompt_tokens - start_tok
4018
            if num_tokens <= num_remaining_tokens:
4019
                # This is a chunk, more tokens remain.
4020
4021
4022
                # 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.
4023
4024
4025
4026
4027
                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)
4028
4029
4030
4031
4032
4033
4034
                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
4035
4036
4037
4038
4039

            # 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]
4040
            offset = self.query_start_loc.np[req_idx].item()
4041
            prompt_hidden_states = hidden_states[offset : offset + num_logits]
4042
            logits = self.model.compute_logits(prompt_hidden_states)
4043
4044
4045
4046

            # 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.
4047
            tgt_token_ids = prompt_token_ids[start_tok : start_tok + num_logits]
4048
4049

            # Compute prompt logprobs.
4050
4051
            logprobs = self.sampler.compute_logprobs(logits)
            token_ids, logprobs, ranks = self.sampler.gather_logprobs(
4052
4053
                logprobs, num_prompt_logprobs, tgt_token_ids
            )
4054
4055

            # Transfer GPU->CPU async.
4056
4057
            chunk_slice = slice(start_idx, start_idx + num_logits)
            logprobs_tensors.logprob_token_ids[chunk_slice].copy_(
4058
4059
4060
                token_ids, non_blocking=True
            )
            logprobs_tensors.logprobs[chunk_slice].copy_(logprobs, non_blocking=True)
4061
            logprobs_tensors.selected_token_ranks[chunk_slice].copy_(
4062
4063
                ranks, non_blocking=True
            )
4064
4065
4066
4067
4068

        # 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]
4069
            del in_progress_dict[req_id]
4070
4071

        # Must synchronize the non-blocking GPU->CPU transfers.
4072
        if prompt_logprobs_dict:
4073
            self._sync_device()
4074
4075
4076

        return prompt_logprobs_dict

4077
4078
    def _get_nans_in_logits(
        self,
4079
        logits: torch.Tensor | None,
4080
4081
4082
4083
4084
4085
4086
4087
4088
4089
4090
    ) -> 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])
4091
4092
4093
                    if num_nans_for_index is not None and req_index < logits.shape[0]
                    else 0
                )
4094
4095
4096
4097
            return num_nans_in_logits
        except IndexError:
            return {}

4098
    @contextmanager
4099
4100
4101
    def maybe_randomize_inputs(
        self, input_ids: torch.Tensor | None, inputs_embeds: torch.Tensor | None
    ):
4102
4103
4104
4105
        """
        Randomize input_ids if VLLM_RANDOMIZE_DP_DUMMY_INPUTS is set.
        This is to help balance expert-selection
         - during profile_run
4106
         - during DP rank dummy run
4107
        """
4108

4109
4110
4111
4112
        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
4113
        elif input_ids is not None:
4114
4115
4116
4117

            @functools.cache
            def rand_input_ids() -> torch.Tensor:
                return torch.randint_like(
4118
                    self.input_ids.gpu,
4119
4120
                    low=0,
                    high=self.model_config.get_vocab_size(),
4121
                )
4122

4123
            logger.debug_once("Randomizing dummy input_ids for DP Rank")
4124
            input_ids.copy_(rand_input_ids()[: input_ids.size(0)], non_blocking=True)
4125
4126
            yield
            input_ids.fill_(0)
4127
4128
4129
4130
4131
4132
4133
4134
4135
4136
4137
4138
4139
4140
4141
        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)
4142

4143
4144
4145
4146
4147
4148
    def _get_mm_dummy_batch(
        self,
        modality: str,
        max_items_per_batch: int,
    ) -> BatchedTensorInputs:
        """Dummy data for profiling and precompiling multimodal models."""
4149
4150
        assert self.mm_budget is not None

4151
        dummy_decoder_data = self.mm_registry.get_decoder_dummy_data(
4152
            model_config=self.model_config,
4153
            seq_len=self.max_model_len,
4154
            mm_counts={modality: 1},
4155
            cache=self.mm_budget.cache,
4156
4157
4158
4159
        )
        dummy_mm_data = dummy_decoder_data.multi_modal_data

        # Result in the maximum GPU consumption of the model
4160
4161
        dummy_mm_item = dummy_mm_data[modality][0]
        dummy_mm_items = [dummy_mm_item] * max_items_per_batch
4162

4163
4164
4165
4166
4167
4168
4169
4170
        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,
            )
        )
4171

4172
4173
4174
4175
    @torch.inference_mode()
    def _dummy_run(
        self,
        num_tokens: int,
4176
        cudagraph_runtime_mode: CUDAGraphMode | None = None,
4177
4178
        force_attention: bool = False,
        uniform_decode: bool = False,
4179
        allow_microbatching: bool = True,
4180
4181
        skip_eplb: bool = False,
        is_profile: bool = False,
4182
        create_mixed_batch: bool = False,
4183
        remove_lora: bool = True,
4184
        activate_lora: bool = False,
Rémi Delacourt's avatar
Rémi Delacourt committed
4185
        is_graph_capturing: bool = False,
4186
    ) -> tuple[torch.Tensor, torch.Tensor]:
4187
4188
4189
4190
4191
4192
4193
        """
        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.
4194
                - if not set will determine the cudagraph mode based on using
4195
                    the self.cudagraph_dispatcher.
4196
4197
4198
4199
                - CUDAGraphMode.NONE: No cudagraph, for warm up and profile run
                - CUDAGraphMode.PIECEWISE: Piecewise cudagraph.
                - CUDAGraphMode.FULL: Full cudagraph, attention metadata is
                    needed.
4200
            force_attention: If True, always create attention metadata. Used to
4201
4202
4203
4204
                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.
4205
4206
            create_mixed_batch: If True, create a mixed batch with both decode
                (1 token) and prefill (multiple tokens) requests.
4207
            remove_lora: If False, dummy LoRAs are not destroyed after the run
4208
            activate_lora: If False, dummy_run is performed without LoRAs.
4209
        """
4210
4211
4212
4213
4214
        if supports_mm_encoder_only(self.model):
            # The current dummy run only covers LM execution, so we can skip it.
            # mm encoder dummy run may need to add in the future.
            return torch.tensor([]), torch.tensor([])

4215
4216
4217
4218
        assert (
            cudagraph_runtime_mode is None
            or cudagraph_runtime_mode.valid_runtime_modes()
        )
4219

4220
        # If cudagraph_mode.decode_mode() == FULL and
4221
        # cudagraph_mode.separate_routine(). This means that we are using
4222
4223
4224
4225
4226
4227
4228
4229
4230
4231
4232
        # 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.
4233
        max_query_len = self.uniform_decode_query_len if uniform_decode else num_tokens
4234

4235
4236
4237
4238
4239
        # 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
4240
4241
4242
4243
        if create_mixed_batch:
            assert not uniform_decode
            # Create mixed batch:
            # first half decode tokens, second half one prefill
4244
            num_decode_tokens = min(max_num_reqs - 1, num_tokens // 2)
4245
4246
4247
4248
            num_prefill_tokens = num_tokens - num_decode_tokens
            num_reqs = num_decode_tokens + 1

            # Create decode requests (1 token each) followed by prefill request
4249
            num_scheduled_tokens_list = [1] * num_decode_tokens + [num_prefill_tokens]
4250
4251
4252
            # Note: Overriding max_query_len to be the prefill tokens
            max_query_len = num_prefill_tokens
        elif uniform_decode:
4253
            assert not create_mixed_batch
4254
            num_reqs = min(max_num_reqs, cdiv(num_tokens, max_query_len))
4255
4256
            num_scheduled_tokens_list = [max_query_len] * num_reqs
            if num_tokens % max_query_len != 0:
4257
                num_scheduled_tokens_list[-1] = num_tokens % max_query_len
4258
4259
4260
4261
4262
4263
        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

4264
4265
        assert sum(num_scheduled_tokens_list) == num_tokens
        assert len(num_scheduled_tokens_list) == num_reqs
4266
        num_scheduled_tokens = np.array(num_scheduled_tokens_list, dtype=np.int32)
4267
4268
        num_tokens_unpadded = int(num_scheduled_tokens.sum())

4269
        num_sampled_tokens = np.ones(num_reqs, dtype=np.int32)
4270

4271
        _cudagraph_mode, batch_desc, should_ubatch, num_tokens_across_dp, _ = (
4272
4273
4274
4275
4276
4277
4278
4279
4280
4281
4282
4283
4284
4285
4286
4287
4288
4289
            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,
4290
4291
            )
        )
4292
4293
4294

        if cudagraph_runtime_mode is None:
            cudagraph_runtime_mode = _cudagraph_mode
4295
        else:
4296
4297
4298
4299
4300
4301
4302
4303
4304
            assert cudagraph_runtime_mode == _cudagraph_mode, (
                f"Cudagraph runtime mode mismatch in dummy_run. "
                f"Expected {_cudagraph_mode}, but got {cudagraph_runtime_mode}."
            )

        num_tokens_padded = batch_desc.num_tokens
        num_reqs_padded = (
            batch_desc.num_reqs if batch_desc.num_reqs is not None else num_reqs
        )
4305
        ubatch_slices, ubatch_slices_padded = maybe_create_ubatch_slices(
4306
4307
4308
4309
4310
4311
4312
4313
4314
4315
            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,
4316
        )
4317

4318
        attn_metadata: PerLayerAttnMetadata | None = None
4319
4320
4321

        # If force_attention is True, we always capture attention. Otherwise,
        # it only happens for cudagraph_runtime_mode=FULL.
4322
        if force_attention or cudagraph_runtime_mode == CUDAGraphMode.FULL:
4323
4324
4325
4326
4327
4328
            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:
4329
                seq_lens = max_query_len  # type: ignore[assignment]
4330
            self.seq_lens.np[:num_reqs] = seq_lens
4331
4332
            self.seq_lens.np[num_reqs:] = 0
            self.seq_lens.copy_to_gpu()
4333

4334
4335
            cum_num_tokens, _ = self._get_cumsum_and_arange(num_scheduled_tokens)
            self.query_start_loc.np[1 : num_reqs + 1] = cum_num_tokens
4336
4337
            self.query_start_loc.copy_to_gpu()

4338
            pad_attn = cudagraph_runtime_mode == CUDAGraphMode.FULL
4339
            attn_metadata, _ = self._build_attention_metadata(
4340
4341
4342
                num_tokens=num_tokens_unpadded,
                num_reqs=num_reqs_padded,
                max_query_len=max_query_len,
4343
                ubatch_slices=ubatch_slices_padded if pad_attn else ubatch_slices,
4344
                for_cudagraph_capture=is_graph_capturing,
4345
            )
4346

4347
        with self.maybe_dummy_run_with_lora(
4348
4349
4350
4351
4352
            self.lora_config,
            num_scheduled_tokens,
            num_sampled_tokens,
            activate_lora,
            remove_lora,
4353
        ):
4354
            # Make sure padding doesn't exceed max_num_tokens
4355
            assert num_tokens_padded <= self.max_num_tokens
4356
            model_kwargs = self._init_model_kwargs()
4357
            if self.supports_mm_inputs and not self.model_config.is_encoder_decoder:
Patrick von Platen's avatar
Patrick von Platen committed
4358
4359
                input_ids, inputs_embeds = self._prepare_mm_inputs(num_tokens_padded)

4360
                model_kwargs = {
4361
                    **model_kwargs,
4362
4363
                    **self._dummy_mm_kwargs(num_reqs),
                }
4364
4365
            elif self.enable_prompt_embeds:
                input_ids = None
4366
                inputs_embeds = self.inputs_embeds.gpu[:num_tokens_padded]
4367
                model_kwargs = self._init_model_kwargs()
4368
            else:
4369
                input_ids = self.input_ids.gpu[:num_tokens_padded]
4370
                inputs_embeds = None
4371

4372
            if self.uses_mrope:
4373
                positions = self.mrope_positions.gpu[:, :num_tokens_padded]
4374
            elif self.uses_xdrope_dim > 0:
4375
                positions = self.xdrope_positions.gpu[:, :num_tokens_padded]
4376
            else:
4377
                positions = self.positions.gpu[:num_tokens_padded]
4378
4379
4380
4381
4382
4383
4384
4385
4386

            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,
4387
4388
4389
                            device=self.device,
                        )
                    )
4390
4391

                intermediate_tensors = self.sync_and_slice_intermediate_tensors(
4392
                    num_tokens_padded, None, False
4393
                )
4394

4395
            if ubatch_slices_padded is not None:
4396
4397
4398
                # Adjust values to reflect a single ubatch.
                # TODO(sage,lucas): this is cruft that should be addressed in
                #  the padding refactor.
4399
                num_tokens_padded = ubatch_slices_padded[0].num_tokens
4400
                if num_tokens_across_dp is not None:
4401
                    num_tokens_across_dp[:] = num_tokens_padded
4402

4403
            with (
4404
                self.maybe_randomize_inputs(input_ids, inputs_embeds),
4405
                set_forward_context(
4406
4407
                    attn_metadata,
                    self.vllm_config,
4408
                    num_tokens=num_tokens_padded,
4409
4410
                    num_tokens_across_dp=num_tokens_across_dp,
                    cudagraph_runtime_mode=cudagraph_runtime_mode,
4411
                    batch_descriptor=batch_desc,
4412
                    ubatch_slices=ubatch_slices_padded,
4413
4414
                ),
            ):
4415
                outputs = self.model(
4416
4417
4418
4419
                    input_ids=input_ids,
                    positions=positions,
                    intermediate_tensors=intermediate_tensors,
                    inputs_embeds=inputs_embeds,
4420
                    **model_kwargs,
4421
                )
4422

4423
4424
4425
4426
            if self.use_aux_hidden_state_outputs:
                hidden_states, _ = outputs
            else:
                hidden_states = outputs
4427

4428
            if self.speculative_config and self.speculative_config.use_eagle():
4429
                assert isinstance(self.drafter, EagleProposer)
4430
4431
4432
                # 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.
4433
                use_cudagraphs = (
4434
4435
4436
4437
4438
4439
4440
4441
4442
                    (
                        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
4443
4444
4445
4446
4447
4448
4449
4450
4451
4452
4453

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

                self.drafter.dummy_run(
                    num_tokens,
                    use_cudagraphs=use_cudagraphs,
Rémi Delacourt's avatar
Rémi Delacourt committed
4454
                    is_graph_capturing=is_graph_capturing,
4455
                )
4456

4457
4458
4459
4460
4461
4462
4463
4464
4465
4466
4467
        # 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()

4468
4469
4470
4471
4472
4473
4474
4475
4476
4477
        # 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)

4478
        logit_indices = np.cumsum(num_scheduled_tokens) - 1
4479
4480
4481
4482
        logit_indices_device = torch.from_numpy(logit_indices).to(
            self.device, non_blocking=True
        )
        return hidden_states, hidden_states[logit_indices_device]
4483
4484
4485
4486
4487
4488

    @torch.inference_mode()
    def _dummy_sampler_run(
        self,
        hidden_states: torch.Tensor,
    ) -> torch.Tensor:
4489
4490
4491
        # The dummy hidden states may contain special values,
        # like `inf` or `nan`.
        # To avoid breaking the sampler, we use a random tensor here instead.
4492
4493
4494
4495
4496

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

4497
        hidden_states = torch.rand_like(hidden_states)
4498

4499
        logits = self.model.compute_logits(hidden_states)
4500
4501
        num_reqs = logits.size(0)

4502
        dummy_tensors = lambda v: torch.full((num_reqs,), v, device=self.device)
4503
4504
4505
4506
4507
4508
4509
4510
4511
4512
4513
4514
4515
4516
4517

        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)],
4518
            spec_token_ids=[[] for _ in range(num_reqs)],
4519
4520
            allowed_token_ids_mask=None,
            bad_words_token_ids={},
4521
            logitsprocs=LogitsProcessors(),
4522
        )
4523
        try:
4524
4525
4526
            sampler_output = self.sampler(
                logits=logits, sampling_metadata=dummy_metadata
            )
4527
        except RuntimeError as e:
4528
            if "out of memory" in str(e):
4529
4530
4531
4532
                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 "
4533
4534
                    "initializing the engine."
                ) from e
4535
4536
            else:
                raise e
4537
        if self.speculative_config:
4538
4539
            draft_token_ids = [[0] for _ in range(num_reqs)]
            dummy_spec_decode_metadata = SpecDecodeMetadata.make_dummy(
4540
4541
                draft_token_ids, self.device
            )
4542
4543
4544
4545
4546
4547

            num_tokens = sum(len(ids) for ids in draft_token_ids)
            # draft_probs = torch.randn(
            #     num_tokens, logits.shape[-1], device=self.device,
            #     dtype=logits.dtype)
            draft_probs = None
4548
4549
4550
4551
4552
            logits = torch.randn(
                num_tokens + num_reqs,
                logits.shape[-1],
                device=self.device,
                dtype=logits.dtype,
4553
            )
4554
4555
4556
            self.rejection_sampler(
                dummy_spec_decode_metadata,
                draft_probs,
4557
                logits,
4558
4559
                dummy_metadata,
            )
4560
        return sampler_output
4561

4562
    def _dummy_pooler_run_task(
4563
4564
        self,
        hidden_states: torch.Tensor,
4565
4566
        task: PoolingTask,
    ) -> PoolerOutput:
4567
4568
4569
4570
        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
4571
4572
4573
4574
        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
4575
4576
4577

        req_num_tokens = num_tokens // num_reqs

4578
        dummy_prompt_lens = torch.from_numpy(num_scheduled_tokens_np)
4579
4580
4581
        dummy_token_ids = torch.zeros(
            (num_reqs, req_num_tokens), dtype=torch.int32, device=self.device
        )
4582

4583
        model = cast(VllmModelForPooling, self.get_model())
4584
        dummy_pooling_params = PoolingParams(task=task)
4585
        dummy_pooling_params.verify(task=task, model_config=self.model_config)
4586
        to_update = model.pooler.get_pooling_updates(task)
4587
4588
        to_update.apply(dummy_pooling_params)

4589
        dummy_metadata = PoolingMetadata(
4590
4591
4592
            prompt_lens=dummy_prompt_lens,
            prompt_token_ids=dummy_token_ids,
            pooling_params=[dummy_pooling_params] * num_reqs,
4593
            pooling_states=[PoolingStates() for i in range(num_reqs)],
4594
        )
4595

4596
        dummy_metadata.build_pooling_cursor(
4597
            num_scheduled_tokens_np,
4598
4599
            seq_lens_cpu=dummy_prompt_lens,
            device=hidden_states.device,
4600
        )
4601

4602
        try:
4603
4604
4605
            return model.pooler(
                hidden_states=hidden_states, pooling_metadata=dummy_metadata
            )
4606
        except RuntimeError as e:
4607
            if "out of memory" in str(e):
4608
                raise RuntimeError(
4609
4610
4611
                    "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 "
4612
4613
                    "initializing the engine."
                ) from e
4614
4615
            else:
                raise e
4616
4617
4618
4619
4620
4621

    @torch.inference_mode()
    def _dummy_pooler_run(
        self,
        hidden_states: torch.Tensor,
    ) -> PoolerOutput:
4622
4623
4624
4625
        if supports_mm_encoder_only(self.model):
            # MM Encoder only model not need to run pooler.
            return torch.tensor([])

4626
        # Find the task that has the largest output for subsequent steps
4627
4628
4629
        supported_pooling_tasks = self.get_supported_pooling_tasks()

        if not supported_pooling_tasks:
4630
4631
4632
4633
4634
4635
            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."
            )
4636

4637
        output_size = dict[PoolingTask, float]()
4638
        for task in supported_pooling_tasks:
4639
4640
            # Run a full batch with each task to ensure none of them OOMs
            output = self._dummy_pooler_run_task(hidden_states, task)
4641
            output_size[task] = sum(o.nbytes for o in output)
4642
4643
4644
4645
            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)
4646

4647
    def profile_run(self) -> None:
4648
        # Profile with multimodal encoder & encoder cache.
4649
        if self.supports_mm_inputs:
4650
4651
            mm_config = self.model_config.multimodal_config
            if mm_config is not None and mm_config.skip_mm_profiling:
4652
                logger.info(
4653
                    "Skipping memory profiling for multimodal encoder and "
4654
4655
                    "encoder cache."
                )
4656
4657
4658
4659
4660
4661
4662
4663
            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.
4664
                    dummy_modality = mm_budget.get_modality_with_max_tokens()
4665
4666
4667
                    max_mm_items_per_batch = mm_budget.max_items_per_batch_by_modality[
                        dummy_modality
                    ]
4668
4669
4670
4671
4672
4673
4674
4675
4676

                    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,
                    )
4677

4678
4679
4680
4681
4682
                    # Create dummy batch of multimodal inputs.
                    batched_dummy_mm_inputs = self._get_mm_dummy_batch(
                        dummy_modality,
                        max_mm_items_per_batch,
                    )
4683

4684
                    # Run multimodal encoder.
4685
                    dummy_encoder_outputs = self.model.embed_multimodal(
4686
4687
                        **batched_dummy_mm_inputs
                    )
4688

4689
4690
4691
4692
                    sanity_check_mm_encoder_outputs(
                        dummy_encoder_outputs,
                        expected_num_items=max_mm_items_per_batch,
                    )
4693
4694
                    for i, output in enumerate(dummy_encoder_outputs):
                        self.encoder_cache[f"tmp_{i}"] = output
4695

4696
        # Add `is_profile` here to pre-allocate communication buffers
4697
4698
4699
        hidden_states, last_hidden_states = self._dummy_run(
            self.max_num_tokens, is_profile=True
        )
4700
        if get_pp_group().is_last_rank:
4701
4702
4703
4704
            if self.is_pooling_model:
                output = self._dummy_pooler_run(hidden_states)
            else:
                output = self._dummy_sampler_run(last_hidden_states)
4705
        else:
4706
            output = None
4707
        self._sync_device()
4708
        del hidden_states, output
4709
        self.encoder_cache.clear()
4710
        gc.collect()
4711

4712
    def capture_model(self) -> int:
4713
        if self.compilation_config.cudagraph_mode == CUDAGraphMode.NONE:
4714
            logger.warning(
4715
                "Skipping CUDA graph capture. To turn on CUDA graph capture, "
4716
4717
                "ensure `cudagraph_mode` was not manually set to `NONE`"
            )
4718
            return 0
4719

4720
4721
        compilation_counter.num_gpu_runner_capture_triggers += 1

4722
4723
        start_time = time.perf_counter()

4724
4725
4726
4727
4728
4729
4730
4731
4732
4733
4734
4735
4736
4737
        @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()
4738
                    gc.collect()
4739

4740
4741
4742
        # 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.
4743
        set_cudagraph_capturing_enabled(True)
4744
        with freeze_gc(), graph_capture(device=self.device):
4745
            start_free_gpu_memory = torch.cuda.mem_get_info()[0]
4746
            cudagraph_mode = self.compilation_config.cudagraph_mode
4747
            assert cudagraph_mode is not None
4748
4749
4750
4751
4752
4753
4754
4755
4756

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

4757
4758
            if cudagraph_mode.mixed_mode() != CUDAGraphMode.NONE:
                cudagraph_runtime_mode = cudagraph_mode.mixed_mode()
4759
                # make sure we capture the largest batch size first
4760
4761
4762
                compilation_cases = list(
                    product(reversed(self.cudagraph_batch_sizes), lora_cases)
                )
4763
4764
4765
                self._capture_cudagraphs(
                    compilation_cases,
                    cudagraph_runtime_mode=cudagraph_runtime_mode,
4766
4767
                    uniform_decode=False,
                )
4768

4769
4770
            # Capture full cudagraph for uniform decode batches if we
            # don't already have full mixed prefill-decode cudagraphs.
4771
4772
4773
4774
4775
4776
4777
            if (
                cudagraph_mode.decode_mode() == CUDAGraphMode.FULL
                and cudagraph_mode.separate_routine()
            ):
                max_num_tokens = (
                    self.scheduler_config.max_num_seqs * self.uniform_decode_query_len
                )
4778
                decode_cudagraph_batch_sizes = [
4779
4780
                    x
                    for x in self.cudagraph_batch_sizes
4781
                    if max_num_tokens >= x >= self.uniform_decode_query_len
4782
                ]
4783
4784
4785
                compilation_cases_decode = list(
                    product(reversed(decode_cudagraph_batch_sizes), lora_cases)
                )
4786
4787
4788
                self._capture_cudagraphs(
                    compilation_cases=compilation_cases_decode,
                    cudagraph_runtime_mode=CUDAGraphMode.FULL,
4789
4790
                    uniform_decode=True,
                )
4791

4792
4793
4794
            torch.cuda.synchronize()
            end_free_gpu_memory = torch.cuda.mem_get_info()[0]

4795
4796
4797
        # 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
4798
        # we may do lazy capturing in future that still allows capturing
4799
4800
        # after here.
        set_cudagraph_capturing_enabled(False)
4801

4802
4803
4804
4805
        # Lock workspace to prevent resizing during execution.
        # Max workspace sizes should have been captured during warmup/profiling.
        lock_workspace()

4806
4807
4808
4809
        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.
4810
        logger.info_once(
4811
4812
4813
            "Graph capturing finished in %.0f secs, took %.2f GiB",
            elapsed_time,
            cuda_graph_size / (1 << 30),
4814
            scope="local",
4815
        )
4816
        return cuda_graph_size
4817

4818
4819
    def _capture_cudagraphs(
        self,
4820
        compilation_cases: list[tuple[int, bool]],
4821
4822
4823
4824
4825
4826
4827
        cudagraph_runtime_mode: CUDAGraphMode,
        uniform_decode: bool,
    ):
        assert (
            cudagraph_runtime_mode != CUDAGraphMode.NONE
            and cudagraph_runtime_mode.valid_runtime_modes()
        ), f"Invalid cudagraph runtime mode: {cudagraph_runtime_mode}"
4828
4829
4830
4831
4832
4833
4834
4835

        # Only rank 0 should print progress bar during capture
        if is_global_first_rank():
            compilation_cases = tqdm(
                compilation_cases,
                disable=not self.load_config.use_tqdm_on_load,
                desc="Capturing CUDA graphs ({}, {})".format(
                    "decode" if uniform_decode else "mixed prefill-decode",
4836
4837
4838
                    cudagraph_runtime_mode.name,
                ),
            )
4839

4840
        # We skip EPLB here since we don't want to record dummy metrics
4841
        for num_tokens, activate_lora in compilation_cases:
4842
            # We currently only capture ubatched graphs when its a FULL
4843
4844
4845
            # 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
4846
            allow_microbatching = (
4847
                self.parallel_config.use_ubatching
4848
4849
                and cudagraph_runtime_mode == CUDAGraphMode.FULL
                and uniform_decode
4850
4851
4852
4853
4854
                and check_ubatch_thresholds(
                    config=self.vllm_config.parallel_config,
                    num_tokens=num_tokens,
                    uniform_decode=uniform_decode,
                )
4855
            )
4856

4857
4858
4859
4860
4861
4862
            for _ in range(self.compilation_config.cudagraph_num_of_warmups):
                # Use CUDAGraphRuntimeStyle.NONE (default) for warmup.
                # But be careful, warm up with `NONE`is orthogonal to
                # if we want to warm up attention or not. This is
                # different from the case where `FULL` implies capture
                # attention while `PIECEWISE` implies no attention.
4863
4864
4865
4866
4867
4868
4869
4870
4871
                force_attention = cudagraph_runtime_mode == CUDAGraphMode.FULL
                self._dummy_run(
                    num_tokens,
                    cudagraph_runtime_mode=CUDAGraphMode.NONE,
                    force_attention=force_attention,
                    uniform_decode=uniform_decode,
                    allow_microbatching=allow_microbatching,
                    skip_eplb=True,
                    remove_lora=False,
4872
                    activate_lora=activate_lora,
4873
4874
4875
4876
4877
4878
4879
4880
                )
            self._dummy_run(
                num_tokens,
                cudagraph_runtime_mode=cudagraph_runtime_mode,
                uniform_decode=uniform_decode,
                allow_microbatching=allow_microbatching,
                skip_eplb=True,
                remove_lora=False,
4881
                activate_lora=activate_lora,
Rémi Delacourt's avatar
Rémi Delacourt committed
4882
                is_graph_capturing=True,
4883
            )
4884
        self.maybe_remove_all_loras(self.lora_config)
4885

4886
4887
4888
4889
    def initialize_attn_backend(self, kv_cache_config: KVCacheConfig) -> None:
        """
        Initialize the attention backends and attention metadata builders.
        """
4890
        assert len(self.attn_groups) == 0, "Attention backends are already initialized"
4891

4892
4893
4894
4895
4896
4897
        class AttentionGroupKey(NamedTuple):
            attn_backend: type[AttentionBackend]
            kv_cache_spec: KVCacheSpec

        def get_attn_backends_for_group(
            kv_cache_group_spec: KVCacheGroupSpec,
4898
        ) -> tuple[dict[AttentionGroupKey, list[str]], set[type[AttentionBackend]]]:
4899
            layer_type = cast(type[Any], AttentionLayerBase)
4900
            layers = get_layers_from_vllm_config(
4901
                self.vllm_config, layer_type, kv_cache_group_spec.layer_names
4902
            )
4903
4904
            attn_backends = {}
            attn_backend_layers = defaultdict(list)
4905
            # Dedupe based on full class name; this is a bit safer than
4906
4907
4908
4909
            # 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.
4910
            for layer_name in kv_cache_group_spec.layer_names:
4911
                attn_backend = layers[layer_name].get_attn_backend()
4912
4913
4914
4915

                if layer_name in self.kv_sharing_fast_prefill_eligible_layers:
                    attn_backend = create_fast_prefill_custom_backend(
                        "FastPrefill",
4916
                        attn_backend,  # type: ignore[arg-type]
4917
4918
                    )

4919
4920
4921
                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):
4922
                    layer_kv_cache_spec = layer_kv_cache_spec.kv_cache_specs[layer_name]
4923
                key = (full_cls_name, layer_kv_cache_spec)
4924
4925
4926
                attn_backends[key] = AttentionGroupKey(
                    attn_backend, layer_kv_cache_spec
                )
4927
                attn_backend_layers[key].append(layer_name)
4928
4929
4930
4931
            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()),
            )
4932
4933

        def create_attn_groups(
4934
            attn_backends_map: dict[AttentionGroupKey, list[str]],
4935
            kv_cache_group_id: int,
4936
4937
        ) -> list[AttentionGroup]:
            attn_groups: list[AttentionGroup] = []
4938
            for (attn_backend, kv_cache_spec), layer_names in attn_backends_map.items():
4939
                attn_group = AttentionGroup(
4940
                    attn_backend,
4941
                    layer_names,
4942
                    kv_cache_spec,
4943
                    kv_cache_group_id,
4944
4945
                )

4946
4947
4948
                attn_groups.append(attn_group)
            return attn_groups

4949
        attention_backend_maps = []
4950
        attention_backend_list = []
4951
        for kv_cache_group_spec in kv_cache_config.kv_cache_groups:
4952
            attn_backends = get_attn_backends_for_group(kv_cache_group_spec)
4953
            attention_backend_maps.append(attn_backends[0])
4954
            attention_backend_list.append(attn_backends[1])
4955
4956

        # Resolve cudagraph_mode before actually initialize metadata_builders
4957
4958
4959
        self._check_and_update_cudagraph_mode(
            attention_backend_list, kv_cache_config.kv_cache_groups
        )
4960

4961
4962
4963
        # Check if attention backend supports PCP&DCP and related features.
        check_attention_cp_compatibility(self.vllm_config)

4964
4965
        for i, attn_backend_map in enumerate(attention_backend_maps):
            self.attn_groups.append(create_attn_groups(attn_backend_map, i))
4966

4967
4968
4969
4970
4971
4972
4973
4974
4975
4976
4977
4978
4979
4980
4981
    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
4982
4983
                    if not self.parallel_config.use_ubatching
                    else self.parallel_config.num_ubatches,
4984
                )
co63oc's avatar
co63oc committed
4985
        # Calculate reorder batch threshold (if needed)
4986
4987
        # Note (tdoublep): do this *after* constructing builders,
        # because some of them change the threshold at init time.
4988
4989
        self.calculate_reorder_batch_threshold()

4990
    def _check_and_update_cudagraph_mode(
4991
4992
4993
        self,
        attention_backends: list[set[type[AttentionBackend]]],
        kv_cache_groups: list[KVCacheGroupSpec],
4994
    ) -> None:
4995
        """
4996
        Resolve the cudagraph_mode when there are multiple attention
4997
        groups with potential conflicting CUDA graph support.
4998
4999
5000
        Then initialize the cudagraph_dispatcher based on the resolved
        cudagraph_mode.
        """
5001
        min_cg_support = AttentionCGSupport.ALWAYS
5002
        min_cg_backend_name = None
5003

5004
5005
5006
5007
5008
5009
5010
5011
5012
5013
5014
5015
        for attn_backend_set, kv_cache_group in zip(
            attention_backends, kv_cache_groups
        ):
            for attn_backend in attn_backend_set:
                builder_cls = attn_backend.get_builder_cls()

                cg_support = builder_cls.get_cudagraph_support(
                    self.vllm_config, kv_cache_group.kv_cache_spec
                )
                if cg_support.value < min_cg_support.value:
                    min_cg_support = cg_support
                    min_cg_backend_name = attn_backend.__name__
5016
5017
        # Flexible resolve the cudagraph mode
        cudagraph_mode = self.compilation_config.cudagraph_mode
5018
        assert cudagraph_mode is not None
5019
        # check cudagraph for mixed batch is supported
5020
5021
5022
5023
5024
5025
        if (
            cudagraph_mode.mixed_mode() == CUDAGraphMode.FULL
            and min_cg_support != AttentionCGSupport.ALWAYS
        ):
            msg = (
                f"CUDAGraphMode.{cudagraph_mode.name} is not supported "
5026
                f"with {min_cg_backend_name} backend (support: "
5027
5028
                f"{min_cg_support})"
            )
5029
5030
            if min_cg_support == AttentionCGSupport.NEVER:
                # if not supported any full cudagraphs, just raise it.
5031
5032
                msg += (
                    "; please try cudagraph_mode=PIECEWISE, and "
5033
                    "make sure compilation mode is VLLM_COMPILE"
5034
                )
5035
5036
5037
5038
5039
                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"
5040
                cudagraph_mode = self.compilation_config.cudagraph_mode = (
5041
                    CUDAGraphMode.FULL_AND_PIECEWISE
5042
                )
5043
5044
            else:
                msg += "; setting cudagraph_mode=FULL_DECODE_ONLY"
5045
                cudagraph_mode = self.compilation_config.cudagraph_mode = (
5046
                    CUDAGraphMode.FULL_DECODE_ONLY
5047
                )
5048
5049
            logger.warning(msg)

5050
        # check that if we are doing decode full-cudagraphs it is supported
5051
5052
5053
5054
5055
5056
        if (
            cudagraph_mode.decode_mode() == CUDAGraphMode.FULL
            and min_cg_support == AttentionCGSupport.NEVER
        ):
            msg = (
                f"CUDAGraphMode.{cudagraph_mode.name} is not supported "
5057
                f"with {min_cg_backend_name} backend (support: "
5058
5059
                f"{min_cg_support})"
            )
5060
            if self.compilation_config.mode == CompilationMode.VLLM_COMPILE and (
5061
5062
5063
5064
5065
                self.compilation_config.splitting_ops_contain_attention()
                or self.compilation_config.use_inductor_graph_partition
            ):
                msg += (
                    "; setting cudagraph_mode=PIECEWISE because "
5066
                    "attention is compiled piecewise"
5067
5068
                )
                cudagraph_mode = self.compilation_config.cudagraph_mode = (
5069
                    CUDAGraphMode.PIECEWISE
5070
                )
5071
            else:
5072
5073
                msg += (
                    "; setting cudagraph_mode=NONE because "
5074
                    "attention is not compiled piecewise"
5075
5076
                )
                cudagraph_mode = self.compilation_config.cudagraph_mode = (
5077
                    CUDAGraphMode.NONE
5078
                )
5079
5080
            logger.warning(msg)

5081
5082
        # check that if we are doing spec-decode + decode full-cudagraphs it is
        # supported
5083
5084
5085
5086
5087
5088
5089
5090
        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 "
5091
                f"{min_cg_backend_name} (support: {min_cg_support})"
5092
            )
5093
5094
            if self.compilation_config.splitting_ops_contain_attention():
                msg += "; setting cudagraph_mode=PIECEWISE"
5095
                cudagraph_mode = self.compilation_config.cudagraph_mode = (
5096
                    CUDAGraphMode.PIECEWISE
5097
                )
5098
5099
            else:
                msg += "; setting cudagraph_mode=NONE"
5100
                cudagraph_mode = self.compilation_config.cudagraph_mode = (
5101
                    CUDAGraphMode.NONE
5102
                )
5103
5104
5105
5106
            logger.warning(msg)

        # double check that we can support full cudagraph if they are requested
        # even after automatic downgrades
5107
5108
5109
5110
5111
5112
        if (
            cudagraph_mode.has_full_cudagraphs()
            and min_cg_support == AttentionCGSupport.NEVER
        ):
            raise ValueError(
                f"CUDAGraphMode.{cudagraph_mode.name} is not "
5113
                f"supported with {min_cg_backend_name} backend ("
5114
5115
                f"support:{min_cg_support}) "
                "; please try cudagraph_mode=PIECEWISE, "
5116
                "and make sure compilation mode is VLLM_COMPILE"
5117
            )
5118

5119
5120
5121
5122
        # 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
5123
        # Will be removed in the near future when we have separate cudagraph capture
5124
5125
5126
5127
5128
5129
5130
5131
5132
        # 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
            )
5133
5134
5135
5136
            capture_sizes = self.compilation_config.cudagraph_capture_sizes
            self.cudagraph_batch_sizes = (
                capture_sizes if capture_sizes is not None else []
            )
5137

5138
5139
        # Trigger cudagraph dispatching keys initialization after
        # resolved cudagraph mode.
5140
        self.compilation_config.cudagraph_mode = cudagraph_mode
5141
        self.cudagraph_dispatcher.initialize_cudagraph_keys(
5142
            cudagraph_mode, self.uniform_decode_query_len
5143
        )
5144

5145
5146
    def calculate_reorder_batch_threshold(self) -> None:
        """
5147
5148
5149
5150
        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.
5151
        """
5152
5153
        min_none_high = lambda a, b: a if b is None else b if a is None else min(a, b)

5154
        reorder_batch_thresholds: list[int | None] = [
5155
5156
5157
            group.get_metadata_builder().reorder_batch_threshold
            for group in self._attn_group_iterator()
        ]
5158
5159
5160
5161
5162
        # 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
5163
        self.reorder_batch_threshold = reduce(min_none_high, reorder_batch_thresholds)  # type: ignore[assignment]
5164

5165
5166
5167
    @staticmethod
    def select_common_block_size(
        kv_manager_block_size: int, attn_groups: list[AttentionGroup]
5168
5169
    ) -> int:
        """
5170
5171
5172
5173
5174
        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.
5175
5176
5177
5178
5179
5180

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

        Returns:
5181
            The selected block size
5182
5183

        Raises:
5184
            ValueError: If no valid block size found
5185
5186
        """

5187
5188
5189
5190
5191
5192
5193
5194
        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
5195
                for supported_size in backend.get_supported_kernel_block_sizes():
5196
5197
5198
5199
5200
5201
5202
5203
5204
5205
5206
5207
5208
5209
5210
5211
5212
5213
5214
5215
5216
5217
5218
5219
5220
5221
5222
5223
5224
5225
                    if isinstance(supported_size, int):
                        if block_size == supported_size:
                            is_supported = True
                    elif isinstance(supported_size, MultipleOf):
                        if block_size % supported_size.base == 0:
                            is_supported = True
                    else:
                        raise ValueError(f"Unknown supported size: {supported_size}")
                if not is_supported:
                    return False
            return True

        backends = [group.backend for group in attn_groups]

        # Case 1: if the block_size of kv cache manager is supported by all backends,
        # return it directly
        if block_size_is_supported(backends, kv_manager_block_size):
            return kv_manager_block_size

        # Case 2: otherwise, the block_size must be an `int`-format supported size of
        # at least one backend. Iterate over all `int`-format supported sizes in
        # descending order and return the first one that is supported by all backends.
        # Simple proof:
        # If the supported size b is in MultipleOf(x_i) format for all attention
        # backends i, and b a factor of kv_manager_block_size, then
        # kv_manager_block_size also satisfies MultipleOf(x_i) for all i. We will
        # return kv_manager_block_size in case 1.
        all_int_supported_sizes = set(
            supported_size
            for backend in backends
5226
            for supported_size in backend.get_supported_kernel_block_sizes()
5227
5228
            if isinstance(supported_size, int)
        )
5229

5230
5231
5232
5233
5234
5235
        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}. ")
5236

5237
5238
5239
    def may_reinitialize_input_batch(
        self, kv_cache_config: KVCacheConfig, kernel_block_sizes: list[int]
    ) -> None:
5240
5241
5242
5243
5244
5245
5246
        """
        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.
5247
            kernel_block_sizes: The kernel block sizes for each KV cache group.
5248
5249
5250
5251
        """
        block_sizes = [
            kv_cache_group.kv_cache_spec.block_size
            for kv_cache_group in kv_cache_config.kv_cache_groups
5252
            if not isinstance(kv_cache_group.kv_cache_spec, EncoderOnlyAttentionSpec)
5253
        ]
5254
5255
5256
5257

        if block_sizes != [self.cache_config.block_size] or kernel_block_sizes != [
            self.cache_config.block_size
        ]:
5258
5259
5260
            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
5261
5262
                "for more details."
            )
5263
5264
            self.input_batch = InputBatch(
                max_num_reqs=self.max_num_reqs,
5265
                max_model_len=max(self.max_model_len, self.max_encoder_len),
5266
5267
5268
5269
5270
                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,
5271
                kernel_block_sizes=kernel_block_sizes,
5272
                is_spec_decode=bool(self.vllm_config.speculative_config),
5273
                logitsprocs=self.input_batch.logitsprocs,
5274
                logitsprocs_need_output_token_ids=self.input_batch.logitsprocs_need_output_token_ids,
5275
                is_pooling_model=self.is_pooling_model,
5276
                num_speculative_tokens=self.num_spec_tokens,
5277
5278
            )

5279
    def _allocate_kv_cache_tensors(
5280
5281
        self, kv_cache_config: KVCacheConfig
    ) -> dict[str, torch.Tensor]:
5282
        """
5283
5284
5285
        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.

5286
        Args:
5287
            kv_cache_config: The KV cache config
5288
        Returns:
5289
            dict[str, torch.Tensor]: A map between layer names to their
5290
            corresponding memory buffer for KV cache.
5291
        """
5292
5293
        kv_cache_raw_tensors: dict[str, torch.Tensor] = {}
        for kv_cache_tensor in kv_cache_config.kv_cache_tensors:
5294
5295
5296
            tensor = torch.zeros(
                kv_cache_tensor.size, dtype=torch.int8, device=self.device
            )
5297
5298
5299
5300
5301
            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:
5302
5303
5304
5305
            for layer_name in group.layer_names:
                if layer_name in self.runner_only_attn_layers:
                    continue
                layer_names.add(layer_name)
5306
5307
5308
        assert layer_names == set(kv_cache_raw_tensors.keys()), (
            "Some layers are not correctly initialized"
        )
5309
5310
        return kv_cache_raw_tensors

5311
5312
5313
    def _attn_group_iterator(self) -> Iterator[AttentionGroup]:
        return itertools.chain.from_iterable(self.attn_groups)

5314
    def _kv_cache_spec_attn_group_iterator(self) -> Iterator[AttentionGroup]:
5315
5316
        if not self.kv_cache_config.kv_cache_groups:
            return
5317
5318
        for attn_groups in self.attn_groups:
            yield from attn_groups
5319

5320
5321
5322
5323
5324
5325
5326
5327
5328
5329
5330
5331
5332
5333
5334
    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 = []
5335
        for kv_cache_gid, kv_cache_group in enumerate(kv_cache_config.kv_cache_groups):
5336
5337
5338
5339
5340
5341
            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):
5342
                continue
5343
            elif isinstance(kv_cache_spec, AttentionSpec):
5344
5345
5346
                # This is an attention backend that supports virtual
                # block splitting. Get the supported block sizes from
                # all backends in the group.
5347
                attn_groups = self.attn_groups[kv_cache_gid]
5348
                kv_manager_block_size = kv_cache_group.kv_cache_spec.block_size
5349
                selected_kernel_size = self.select_common_block_size(
5350
5351
5352
                    kv_manager_block_size, attn_groups
                )
                kernel_block_sizes.append(selected_kernel_size)
5353
            elif isinstance(kv_cache_spec, MambaSpec):
5354
5355
                # This is likely Mamba or other non-attention cache,
                # no splitting.
5356
                kernel_block_sizes.append(kv_cache_spec.block_size)
5357
5358
5359
5360
5361
5362
            else:
                raise NotImplementedError(
                    f"unknown kv cache spec {kv_cache_group.kv_cache_spec}"
                )
        return kernel_block_sizes

5363
5364
5365
5366
    def _reshape_kv_cache_tensors(
        self,
        kv_cache_config: KVCacheConfig,
        kv_cache_raw_tensors: dict[str, torch.Tensor],
5367
        kernel_block_sizes: list[int],
5368
    ) -> dict[str, torch.Tensor]:
5369
        """
5370
        Reshape the KV cache tensors to the desired shape and dtype.
5371

5372
        Args:
5373
5374
            kv_cache_config: The KV cache config
            kv_cache_raw_tensors: The KV cache buffer of each layer, with
5375
                correct size but uninitialized shape.
5376
            kernel_block_sizes: The kernel block sizes for each KV cache group.
5377
        Returns:
5378
            Dict[str, torch.Tensor]: A map between layer names to their
5379
5380
            corresponding memory buffer for KV cache.
        """
5381
        kv_caches: dict[str, torch.Tensor] = {}
5382
        has_attn, has_mamba = False, False
5383
5384
        for group in self._kv_cache_spec_attn_group_iterator():
            kv_cache_spec = group.kv_cache_spec
5385
            attn_backend = group.backend
5386
5387
5388
5389
            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]
5390
            for layer_name in group.layer_names:
5391
5392
                if layer_name in self.runner_only_attn_layers:
                    continue
5393
5394
                raw_tensor = kv_cache_raw_tensors[layer_name]
                assert raw_tensor.numel() % kv_cache_spec.page_size_bytes == 0
5395
                num_blocks = raw_tensor.numel() // kv_cache_spec.page_size_bytes
5396
                if isinstance(kv_cache_spec, AttentionSpec):
5397
                    has_attn = True
5398
5399
                    num_blocks_per_kv_block = (
                        kv_cache_spec.block_size // kernel_block_size
5400
5401
5402
                    )
                    kernel_num_blocks = num_blocks * num_blocks_per_kv_block

5403
                    kv_cache_shape = attn_backend.get_kv_cache_shape(
5404
                        kernel_num_blocks,
5405
                        kernel_block_size,
5406
5407
                        kv_cache_spec.num_kv_heads,
                        kv_cache_spec.head_size,
5408
5409
                        cache_dtype_str=self.cache_config.cache_dtype,
                    )
5410
                    dtype = kv_cache_spec.dtype
5411
                    try:
5412
                        kv_cache_stride_order = attn_backend.get_kv_cache_stride_order()
5413
                        assert len(kv_cache_stride_order) == len(kv_cache_shape)
5414
                    except (AttributeError, NotImplementedError):
5415
                        kv_cache_stride_order = tuple(range(len(kv_cache_shape)))
5416
5417
5418
5419
5420
                    # 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.
5421
5422
5423
                    kv_cache_shape = tuple(
                        kv_cache_shape[i] for i in kv_cache_stride_order
                    )
5424
5425
5426
5427
5428
                    # Maintain original KV shape view.
                    inv_order = [
                        kv_cache_stride_order.index(i)
                        for i in range(len(kv_cache_stride_order))
                    ]
5429
5430
5431
5432
5433
5434
                    kv_caches[layer_name] = (
                        kv_cache_raw_tensors[layer_name]
                        .view(dtype)
                        .view(kv_cache_shape)
                        .permute(*inv_order)
                    )
Chen Zhang's avatar
Chen Zhang committed
5435
                elif isinstance(kv_cache_spec, MambaSpec):
5436
                    has_mamba = True
Chen Zhang's avatar
Chen Zhang committed
5437
5438
                    raw_tensor = kv_cache_raw_tensors[layer_name]
                    state_tensors = []
5439
                    storage_offset_bytes = 0
5440
                    for shape, dtype in zip(kv_cache_spec.shapes, kv_cache_spec.dtypes):
5441
5442
                        dtype_size = get_dtype_size(dtype)
                        num_element_per_page = (
5443
5444
                            kv_cache_spec.page_size_bytes // dtype_size
                        )
Chen Zhang's avatar
Chen Zhang committed
5445
                        target_shape = (num_blocks, *shape)
5446
5447
                        stride = torch.empty(target_shape).stride()
                        target_stride = (num_element_per_page, *stride[1:])
5448
                        assert storage_offset_bytes % dtype_size == 0
5449
5450
5451
5452
                        tensor = torch.as_strided(
                            raw_tensor.view(dtype),
                            size=target_shape,
                            stride=target_stride,
5453
                            storage_offset=storage_offset_bytes // dtype_size,
5454
                        )
Chen Zhang's avatar
Chen Zhang committed
5455
                        state_tensors.append(tensor)
5456
                        storage_offset_bytes += stride[0] * dtype_size
5457
5458

                    kv_caches[layer_name] = state_tensors
5459
                else:
5460
                    raise NotImplementedError
5461
5462

        if has_attn and has_mamba:
5463
            self._update_hybrid_attention_mamba_layout(kv_caches)
5464

5465
5466
        return kv_caches

5467
    def _update_hybrid_attention_mamba_layout(
5468
5469
        self, kv_caches: dict[str, torch.Tensor]
    ) -> None:
5470
        """
5471
5472
        Update the layout of attention layers from (2, num_blocks, ...) to
        (num_blocks, 2, ...).
5473
5474

        Args:
5475
            kv_caches: The KV cache buffer of each layer.
5476
5477
        """

5478
5479
        for group in self._kv_cache_spec_attn_group_iterator():
            kv_cache_spec = group.kv_cache_spec
5480
            for layer_name in group.layer_names:
5481
                kv_cache = kv_caches[layer_name]
5482
5483
5484
5485
                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 "
5486
                        f"a tensor of shape {kv_cache.shape}"
5487
                    )
5488
                    hidden_size = kv_cache.shape[2:].numel()
5489
5490
5491
5492
                    kv_cache.as_strided_(
                        size=kv_cache.shape,
                        stride=(hidden_size, 2 * hidden_size, *kv_cache.stride()[2:]),
                    )
5493

5494
    def initialize_kv_cache_tensors(
5495
        self, kv_cache_config: KVCacheConfig, kernel_block_sizes: list[int]
5496
    ) -> dict[str, torch.Tensor]:
5497
5498
5499
5500
5501
        """
        Initialize the memory buffer for KV cache.

        Args:
            kv_cache_config: The KV cache config
5502
5503
            kernel_block_sizes: The kernel block sizes for each KV cache group.

5504
        Returns:
5505
            Dict[str, torch.Tensor]: A map between layer names to their
5506
5507
            corresponding memory buffer for KV cache.
        """
5508
5509
5510
5511
5512
5513
5514
5515
5516
5517
5518
5519
5520
5521
5522
5523
5524
5525
5526
5527
5528
5529
5530
5531

        # 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
            )
5532

5533
        # Set up cross-layer KV cache sharing
5534
5535
        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)
5536
5537
            kv_caches[layer_name] = kv_caches[target_layer_name]

5538
5539
5540
5541
5542
5543
5544
5545
5546
        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,
        )
5547
5548
5549
        return kv_caches

    def maybe_add_kv_sharing_layers_to_kv_cache_groups(
5550
5551
        self, kv_cache_config: KVCacheConfig
    ) -> None:
5552
5553
5554
5555
5556
5557
5558
5559
5560
5561
5562
5563
5564
5565
5566
5567
5568
5569
        """
        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.
5570
            attn_layers = get_layers_from_vllm_config(self.vllm_config, Attention)
5571
5572
            for layer_name in reversed(attn_layers):
                if layer_name in self.shared_kv_cache_layers:
5573
                    self.kv_sharing_fast_prefill_eligible_layers.add(layer_name)
5574
5575
                else:
                    break
5576

5577
5578
5579
5580
5581
5582
5583
    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
        """
5584
        kv_cache_config = deepcopy(kv_cache_config)
5585
        self.kv_cache_config = kv_cache_config
5586
        self.may_add_encoder_only_layers_to_kv_cache_config()
5587
        self.maybe_add_kv_sharing_layers_to_kv_cache_groups(kv_cache_config)
5588
        self.initialize_attn_backend(kv_cache_config)
5589
5590
5591
5592
5593
5594
        # 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)
5595
5596
5597
5598

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

5599
        # Reinitialize need to after initialize_attn_backend
5600
5601
5602
5603
        self.may_reinitialize_input_batch(kv_cache_config, kernel_block_sizes)
        kv_caches = self.initialize_kv_cache_tensors(
            kv_cache_config, kernel_block_sizes
        )
5604

5605
5606
5607
5608
5609
5610
        if self.speculative_config and self.speculative_config.use_eagle():
            assert isinstance(self.drafter, EagleProposer)
            # validate all draft model layers belong to the same kv cache
            # group
            self.drafter.validate_same_kv_cache_group(kv_cache_config)

Robert Shaw's avatar
Robert Shaw committed
5611
        if has_kv_transfer_group():
5612
            kv_transfer_group = get_kv_transfer_group()
5613
5614
5615
5616
5617
5618
5619
            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)
5620
            kv_transfer_group.set_host_xfer_buffer_ops(copy_kv_blocks)
Robert Shaw's avatar
Robert Shaw committed
5621

5622
5623
5624
5625
5626
    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
5627
        encoder_only_attn_specs: dict[AttentionSpec, list[str]] = defaultdict(list)
5628
5629
5630
        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:
5631
                attn_spec: AttentionSpec = EncoderOnlyAttentionSpec(
5632
5633
5634
                    block_size=block_size,
                    num_kv_heads=attn_module.num_kv_heads,
                    head_size=attn_module.head_size,
5635
5636
                    dtype=self.kv_cache_dtype,
                )
5637
5638
5639
                encoder_only_attn_specs[attn_spec].append(layer_name)
                self.runner_only_attn_layers.add(layer_name)
        if len(encoder_only_attn_specs) > 0:
5640
5641
5642
            assert len(encoder_only_attn_specs) == 1, (
                "Only support one encoder-only attention spec now"
            )
5643
5644
            spec, layer_names = encoder_only_attn_specs.popitem()
            self.kv_cache_config.kv_cache_groups.append(
5645
5646
                KVCacheGroupSpec(layer_names=layer_names, kv_cache_spec=spec)
            )
5647

5648
    def get_kv_cache_spec(self) -> dict[str, KVCacheSpec]:
5649
        """
5650
        Generates the KVCacheSpec by parsing the kv cache format from each
5651
5652
        Attention module in the static forward context.
        Returns:
5653
            KVCacheSpec: A dictionary mapping layer names to their KV cache
5654
5655
            format. Layers that do not need KV cache are not included.
        """
5656
5657
        if has_ec_transfer() and get_ec_transfer().is_producer:
            return {}
5658
        kv_cache_spec: dict[str, KVCacheSpec] = {}
5659
5660
        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
5661
        for layer_name, attn_module in attn_layers.items():
5662
5663
5664
5665
5666
5667
5668
5669
5670
5671
5672
5673
5674
5675
5676
            if isinstance(attn_module, Attention) and (
                kv_tgt_layer := attn_module.kv_sharing_target_layer_name
            ):
                # The layer doesn't need its own KV cache and will use that of
                # the target layer. We skip creating a KVCacheSpec for it, so
                # that KV cache management logic will act as this layer does
                # not exist, and doesn't allocate KV cache for the layer. This
                # enables the memory saving of cross-layer kv sharing, allowing
                # a given amount of memory to accommodate longer context lengths
                # or enable more requests to be processed simultaneously.
                self.shared_kv_cache_layers[layer_name] = kv_tgt_layer
                continue
            # Skip modules that don't need KV cache (eg encoder-only attention)
            if spec := attn_module.get_kv_cache_spec(self.vllm_config):
                kv_cache_spec[layer_name] = spec
5677

5678
        return kv_cache_spec
5679

5680
    def _to_list(self, sampled_token_ids: torch.Tensor) -> list[list[int]]:
5681
5682
5683
5684
5685
5686
5687
5688
        # 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.
5689
        pinned = self.sampled_token_ids_pinned_cpu[: sampled_token_ids.shape[0]]
5690
5691
5692
        pinned.copy_(sampled_token_ids, non_blocking=True)
        self.transfer_event.record()
        self.transfer_event.synchronize()
5693
        return pinned.tolist()