gpu_model_runner.py 245 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
from vllm.utils.mem_utils import DeviceMemoryProfiler, format_gib
97
from vllm.utils.nvtx_pytorch_hooks import PytHooks
98
from vllm.utils.platform_utils import is_pin_memory_available
99
100
101
102
103
from vllm.utils.torch_utils import (
    get_dtype_size,
    kv_cache_dtype_str_to_dtype,
    supports_dynamo,
)
104
from vllm.v1.attention.backends.gdn_attn import GDNAttentionMetadataBuilder
105
from vllm.v1.attention.backends.utils import (
106
107
108
    AttentionCGSupport,
    AttentionMetadataBuilder,
    CommonAttentionMetadata,
109
    create_fast_prefill_custom_backend,
110
    get_dcp_local_seq_lens,
111
112
113
    reorder_batch_to_split_decodes_and_prefills,
    split_attn_metadata,
)
114
from vllm.v1.cudagraph_dispatcher import CudagraphDispatcher
115
116
117
118
119
120
121
122
123
124
125
126
127
128
129
130
131
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,
132
    ECConnectorOutput,
133
    KVConnectorOutput,
134
135
136
137
138
    LogprobsLists,
    LogprobsTensors,
    ModelRunnerOutput,
    PoolerOutput,
    SamplerOutput,
139
    make_empty_encoder_model_runner_output,
140
)
141
from vllm.v1.pool.metadata import PoolingMetadata, PoolingStates
142
from vllm.v1.sample.logits_processor import LogitsProcessors, build_logitsprocs
143
from vllm.v1.sample.logits_processor.interface import LogitsProcessor
144
from vllm.v1.sample.metadata import SamplingMetadata
145
from vllm.v1.sample.rejection_sampler import RejectionSampler
146
from vllm.v1.sample.sampler import Sampler
147
from vllm.v1.spec_decode.eagle import EagleProposer
148
from vllm.v1.spec_decode.medusa import MedusaProposer
149
from vllm.v1.spec_decode.metadata import SpecDecodeMetadata
150
from vllm.v1.spec_decode.ngram_proposer import NgramProposer
151
from vllm.v1.spec_decode.suffix_decoding import SuffixDecodingProposer
152
from vllm.v1.structured_output.utils import apply_grammar_bitmask
153
from vllm.v1.utils import CpuGpuBuffer, record_function_or_nullcontext
154
from vllm.v1.worker.cp_utils import check_attention_cp_compatibility
155
from vllm.v1.worker.dp_utils import coordinate_batch_across_dp
156
from vllm.v1.worker.ec_connector_model_runner_mixin import ECConnectorModelRunnerMixin
157
from vllm.v1.worker.gpu_input_batch import CachedRequestState, InputBatch
158
from vllm.v1.worker.gpu_ubatch_wrapper import UBatchWrapper
159
from vllm.v1.worker.kv_connector_model_runner_mixin import KVConnectorModelRunnerMixin
160
from vllm.v1.worker.lora_model_runner_mixin import LoRAModelRunnerMixin
161
162
163
from vllm.v1.worker.ubatch_utils import (
    UBatchSlices,
    check_ubatch_thresholds,
164
    maybe_create_ubatch_slices,
165
)
166
from vllm.v1.worker.utils import is_residual_scattered_for_sp
167
from vllm.v1.worker.workspace import lock_workspace
168

169
170
171
172
173
174
175
from .utils import (
    AttentionGroup,
    MultiModalBudget,
    add_kv_sharing_layers_to_kv_cache_groups,
    bind_kv_cache,
    sanity_check_mm_encoder_outputs,
)
176

177
if TYPE_CHECKING:
178
    from vllm.model_executor.model_loader.tensorizer import TensorizerConfig
179
    from vllm.v1.core.sched.output import GrammarOutput, SchedulerOutput
180
181
182

logger = init_logger(__name__)

183
184
AttnMetadataDict: TypeAlias = dict[str, AttentionMetadata]
# list when ubatching is enabled
185
PerLayerAttnMetadata: TypeAlias = list[AttnMetadataDict] | AttnMetadataDict
186

187

188
189
190
191
192
193
# Wrapper for ModelRunnerOutput to support overlapped execution.
class AsyncGPUModelRunnerOutput(AsyncModelRunnerOutput):
    def __init__(
        self,
        model_runner_output: ModelRunnerOutput,
        sampled_token_ids: torch.Tensor,
194
        logprobs_tensors: LogprobsTensors | None,
195
196
        invalid_req_indices: list[int],
        async_output_copy_stream: torch.cuda.Stream,
197
        vocab_size: int,
198
199
200
201
202
    ):
        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.
203
        self.async_copy_ready_event = torch.Event()
204
205
206
207

        # 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
208
        self.vocab_size = vocab_size
209
        self._logprobs_tensors = logprobs_tensors
210
211
212
213
214

        # 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)
215
            self.sampled_token_ids_cpu = self._sampled_token_ids.to(
216
217
                "cpu", non_blocking=True
            )
218
219
220
221
222
            self._logprobs_tensors_cpu = (
                self._logprobs_tensors.to_cpu_nonblocking()
                if self._logprobs_tensors
                else None
            )
223
            self.async_copy_ready_event.record()
224
225
226

    def get_output(self) -> ModelRunnerOutput:
        """Copy the device tensors to the host and return a ModelRunnerOutput.
227

228
229
        This function blocks until the copy is finished.
        """
230
        max_gen_len = self.sampled_token_ids_cpu.shape[-1]
231
        self.async_copy_ready_event.synchronize()
232

233
234
        # Release the device tensors once the copy has completed.
        del self._logprobs_tensors
235
        del self._sampled_token_ids
236
        if max_gen_len == 1:
237
            valid_sampled_token_ids = self.sampled_token_ids_cpu.tolist()
238
239
240
            for i in self._invalid_req_indices:
                valid_sampled_token_ids[i].clear()
            cu_num_tokens = None
241
        else:
242
            valid_sampled_token_ids, cu_num_tokens = RejectionSampler.parse_output(
243
244
                self.sampled_token_ids_cpu,
                self.vocab_size,
245
246
                self._invalid_req_indices,
                return_cu_num_tokens=self._logprobs_tensors_cpu is not None,
247
            )
248
249
250

        output = self._model_runner_output
        output.sampled_token_ids = valid_sampled_token_ids
251
        if self._logprobs_tensors_cpu:
252
            output.logprobs = self._logprobs_tensors_cpu.tolists(cu_num_tokens)
253
254
255
        return output


256
257
258
259
260
261
262
263
264
265
266
267
268
269
270
271
272
273
274
275
276
class AsyncGPUPoolingModelRunnerOutput(AsyncModelRunnerOutput):
    def __init__(
        self,
        model_runner_output: ModelRunnerOutput,
        raw_pooler_output: PoolerOutput,
        finished_mask: list[bool],
        async_output_copy_stream: torch.cuda.Stream,
    ):
        self._model_runner_output = model_runner_output

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

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

        # Initiate the copy on a separate stream, but do not synchronize it.
        default_stream = torch.cuda.current_stream()
        with torch.cuda.stream(async_output_copy_stream):
            async_output_copy_stream.wait_stream(default_stream)
277
            raw_pooler_output_cpu = json_map_leaves(
278
279
280
281
                lambda x: None if x is None else x.to("cpu", non_blocking=True),
                self._raw_pooler_output,
            )
            self.async_copy_ready_event.record()
282
283
284
285
            self._model_runner_output.pooler_output = [
                out if include else None
                for out, include in zip(raw_pooler_output_cpu, finished_mask)
            ]
286
287
288
289
290
291
292
293
294
295
296
297

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

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


298
299
300
301
302
303
304
305
306
307
308
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
309
    ec_connector_output: ECConnectorOutput | None
310
    cudagraph_stats: CUDAGraphStat | None
311
312


313
314
315
class GPUModelRunner(
    LoRAModelRunnerMixin, KVConnectorModelRunnerMixin, ECConnectorModelRunnerMixin
):
316
317
    def __init__(
        self,
318
        vllm_config: VllmConfig,
319
        device: torch.device,
320
    ):
321
322
323
        self.vllm_config = vllm_config
        self.model_config = vllm_config.model_config
        self.cache_config = vllm_config.cache_config
324
        self.compilation_config = vllm_config.compilation_config
325
326
327
328
329
330
        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
331

332
        from vllm.model_executor.models.utils import set_cpu_offload_max_bytes
333
334

        set_cpu_offload_max_bytes(int(self.cache_config.cpu_offload_gb * 1024**3))
335

336
337
338
339
        model_config = self.model_config
        cache_config = self.cache_config
        scheduler_config = self.scheduler_config
        parallel_config = self.parallel_config
340
        self.device = device
341
342
        self.pin_memory = is_pin_memory_available()
        self.dtype = self.model_config.dtype
343

344
345
346
        self.kv_cache_dtype = kv_cache_dtype_str_to_dtype(
            cache_config.cache_dtype, self.model_config
        )
347

348
        self.is_pooling_model = model_config.runner_type == "pooling"
349
        self.enable_prompt_embeds = model_config.enable_prompt_embeds
350
        self.is_multimodal_raw_input_only_model = (
351
352
            model_config.is_multimodal_raw_input_only_model
        )
353
354
        # This will be overridden in load_model()
        self.is_multimodal_pruning_enabled = False
355
        self.max_model_len = model_config.max_model_len
356
357
358

        # Always set to false after the first forward pass
        self.calculate_kv_scales = self.cache_config.calculate_kv_scales
359
        self.dcp_world_size = self.parallel_config.decode_context_parallel_size
360
        self.dcp_rank = 0 if self.dcp_world_size <= 1 else get_dcp_group().rank_in_group
361
        self.max_num_tokens = scheduler_config.max_num_batched_tokens
362
        self.max_num_reqs = scheduler_config.max_num_seqs
363

364
365
366
367
368
        # 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 = (
369
            self.parallel_config.distributed_executor_backend == "external_launcher"
370
            and len(get_pp_group().ranks) > 1
371
        )
372

373
        # Model-related.
374
        self.num_query_heads = model_config.get_num_attention_heads(parallel_config)
375
        self.inputs_embeds_size = model_config.get_inputs_embeds_size()
376
        self.attention_chunk_size = model_config.attention_chunk_size
377
        # Only relevant for models using ALiBi (e.g, MPT)
378
        self.use_alibi = model_config.uses_alibi
379

380
        self.cascade_attn_enabled = not self.model_config.disable_cascade_attn
381
        self.is_mm_prefix_lm = self.model_config.is_mm_prefix_lm
382

383
        # Multi-modal data support
384
        self.mm_registry = MULTIMODAL_REGISTRY
385
        self.uses_mrope = model_config.uses_mrope
386
        self.uses_xdrope_dim = model_config.uses_xdrope_dim
387
        self.supports_mm_inputs = self.mm_registry.supports_multimodal_inputs(
388
            model_config
389
        )
390

391
392
393
        if self.model_config.is_encoder_decoder:
            # Maximum length of the encoder input, only for encoder-decoder
            # models.
394
            self.max_encoder_len = scheduler_config.max_num_encoder_input_tokens
395
396
397
        else:
            self.max_encoder_len = 0

398
        # Sampler
399
        self.sampler = Sampler(logprobs_mode=self.model_config.logprobs_mode)
400

401
        self.eplb_state: EplbState | None = None
402
403
404
405
406
407
        """
        State of the expert parallelism load balancer.

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

408
        # Lazy initializations
409
        # self.model: nn.Module  # Set after load_model
410
        # Initialize in initialize_kv_cache
411
        self.kv_caches: list[torch.Tensor] = []
412
413
414
        # Initialize in initialize_kv_cache_tensors
        self.cross_layers_kv_cache: torch.Tensor | None = None
        self.cross_layers_attn_backend: type[AttentionBackend] | None = None
415
416
        # indexes: [kv_cache_group_id][attn_group]
        self.attn_groups: list[list[AttentionGroup]] = []
417
418
        # self.kv_cache_config: KVCacheConfig

419
420
        # mm_hash ->  encoder_output
        self.encoder_cache: dict[str, torch.Tensor] = {}
421

422
        self.use_aux_hidden_state_outputs = False
423
424
425
426
427
        # 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:
428
429
430
            self.drafter: (
                NgramProposer | SuffixDecodingProposer | EagleProposer | MedusaProposer
            )
431
432
            if self.speculative_config.method == "ngram":
                self.drafter = NgramProposer(self.vllm_config)
433
434
            elif self.speculative_config.method == "suffix":
                self.drafter = SuffixDecodingProposer(self.vllm_config)
435
            elif self.speculative_config.use_eagle():
436
                self.drafter = EagleProposer(self.vllm_config, self.device, self)
437
                if self.speculative_config.method == "eagle3":
438
439
440
                    self.use_aux_hidden_state_outputs = (
                        self.drafter.eagle3_use_aux_hidden_state
                    )
441
442
            elif self.speculative_config.method == "medusa":
                self.drafter = MedusaProposer(
443
                    vllm_config=self.vllm_config, device=self.device
444
                )
445
            else:
446
447
448
449
                raise ValueError(
                    "Unknown speculative decoding method: "
                    f"{self.speculative_config.method}"
                )
450
            self.rejection_sampler = RejectionSampler(self.sampler)
451

452
453
454
455
        self.num_spec_tokens = 0
        if self.speculative_config:
            self.num_spec_tokens = self.speculative_config.num_speculative_tokens

456
        # Request states.
457
        self.requests: dict[str, CachedRequestState] = {}
458
459
460
        # NOTE(rob): num_prompt_logprobs only includes reqs
        # that are currently in the prefill phase.
        self.num_prompt_logprobs: dict[str, int] = {}
461
        self.comm_stream = torch.cuda.Stream()
462

463
464
465
466
467
468
469
470
471
        # 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.
472
473
474
475
        logits_processors = model_config.logits_processors
        custom_logitsprocs: Sequence[str | type[LogitsProcessor]] = (
            tuple(logits_processors) if logits_processors is not None else ()
        )
476
477
        self.input_batch = InputBatch(
            max_num_reqs=self.max_num_reqs,
478
479
480
            # 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),
481
482
483
            max_num_batched_tokens=self.max_num_tokens,
            device=self.device,
            pin_memory=self.pin_memory,
484
            vocab_size=self.model_config.get_vocab_size(),
485
            block_sizes=[self.cache_config.block_size],
486
            kernel_block_sizes=[self.cache_config.block_size],
487
            is_spec_decode=bool(self.vllm_config.speculative_config),
488
            logitsprocs=build_logitsprocs(
489
490
491
                self.vllm_config,
                self.device,
                self.pin_memory,
492
                self.is_pooling_model,
493
                custom_logitsprocs,
494
            ),
495
496
497
            # 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),
498
            is_pooling_model=self.is_pooling_model,
499
            cp_kv_cache_interleave_size=self.parallel_config.cp_kv_cache_interleave_size,
500
        )
501

502
        self.use_async_scheduling = self.scheduler_config.async_scheduling
503
504
505
506
507
        # 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.
508
        self.prepare_inputs_event: torch.Event | None = None
509
510
        if self.use_async_scheduling:
            self.async_output_copy_stream = torch.cuda.Stream()
511
            self.prepare_inputs_event = torch.Event()
512

513
        # self.cudagraph_batch_sizes sorts in ascending order.
514
515
516
517
        if (
            self.compilation_config.cudagraph_capture_sizes
            and self.compilation_config.cudagraph_mode != CUDAGraphMode.NONE
        ):
518
519
            self.cudagraph_batch_sizes = sorted(
                self.compilation_config.cudagraph_capture_sizes
520
            )
521

522
        # Cache the device properties.
523
        self._init_device_properties()
524

525
        # Persistent buffers for CUDA graphs.
526
527
528
529
530
        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
        )
531
        self.seq_lens = self._make_buffer(self.max_num_reqs, dtype=torch.int32)
532
        self.encoder_seq_lens = self._make_buffer(self.max_num_reqs, dtype=torch.int32)
533
534
535
536
        if self.dcp_world_size > 1:
            self.dcp_local_seq_lens = self._make_buffer(
                self.max_num_reqs, dtype=torch.int32
            )
537
538
539
        # 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.
540
        self.inputs_embeds = self._make_buffer(
541
            self.max_num_tokens, self.inputs_embeds_size, dtype=self.dtype, numpy=False
542
543
        )
        self.is_token_ids = self._make_buffer(self.max_num_tokens, dtype=torch.bool)
544
545
        self.discard_request_mask = self._make_buffer(
            self.max_num_reqs, dtype=torch.bool
546
547
548
549
550
551
552
        )
        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
        )
553

554
555
        # Only relevant for multimodal models
        if self.supports_mm_inputs:
556
557
558
559
560
561
562
            # Double buffer to avoid race condition: previous iteration's async
            # copy may still be reading from CPU while current iteration writes.
            self.is_mm_embed_buffers = [
                self._make_buffer(self.max_num_tokens, dtype=torch.bool),
                self._make_buffer(self.max_num_tokens, dtype=torch.bool),
            ]
            self.is_mm_embed_idx = 0
563

564
        # Only relevant for models using M-RoPE (e.g, Qwen2-VL)
565
        if self.uses_mrope:
Roger Wang's avatar
Roger Wang committed
566
567
568
569
            # 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
570
571
572
573
574
575

            # 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
576
            self.mrope_positions = self._make_buffer(
577
578
                (3, self.max_num_tokens + 1), dtype=torch.int64
            )
579

580
581
582
583
584
585
586
        # 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
            )

587
        # None in the first PP rank. The rest are set after load_model.
588
        self.intermediate_tensors: IntermediateTensors | None = None
589

590
        # OPTIMIZATION: Cache the tensors rather than creating them every step.
591
        # Keep in int64 to avoid overflow with long context
592
593
594
595
        self.arange_np = np.arange(
            max(self.max_num_reqs + 1, self.max_model_len, self.max_num_tokens),
            dtype=np.int64,
        )
596

597
598
599
600
601
        # 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] = {}
602
603
604
605
606
        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(
607
608
                self.max_num_tokens, dtype=torch.int32, device=self.device
            )
609

610
        self.uniform_decode_query_len = 1 + self.num_spec_tokens
611
612
613
614

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

615
616
        self.mm_budget = (
            MultiModalBudget(
617
                self.model_config,
618
619
620
621
622
623
                self.scheduler_config,
                self.mm_registry,
            )
            if self.supports_mm_inputs
            else None
        )
624

625
        self.reorder_batch_threshold: int | None = None
626

627
628
629
630
631
        # 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()

632
        # Cached outputs.
633
        self._draft_token_ids: list[list[int]] | torch.Tensor | None = None
634
        self._draft_token_req_ids: list[str] | None = None
635
        self.transfer_event = torch.Event()
636
        self.sampled_token_ids_pinned_cpu = torch.empty(
637
            (self.max_num_reqs, 1),
638
639
            dtype=torch.int64,
            device="cpu",
640
641
            pin_memory=self.pin_memory,
        )
642

643
644
        # Pre-allocated tensor for copying valid sampled token counts to CPU,
        # with dedicated stream for overlapping and event for coordination.
645
        self.valid_sampled_token_count_event: torch.Event | None = None
646
        self.valid_sampled_token_count_copy_stream: torch.cuda.Stream | None = None
647
648
649
650
651
652
653
654
655
656
657
658
659
660
661
662
663
664
665
666
667
668
669
670
        # We also copy the drafted tokens to the CPU asynchronously,
        # in case we need them for structured outputs.
        self.draft_token_ids_event: torch.Event | None = None
        self.draft_token_ids_copy_stream: torch.cuda.Stream | None = None
        self.valid_sampled_token_count_cpu: torch.Tensor | None = None
        self.draft_token_ids_cpu: torch.Tensor | None = None
        if self.num_spec_tokens:
            self.draft_token_ids_event = torch.Event()
            self.draft_token_ids_copy_stream = torch.cuda.Stream()
            self.draft_token_ids_cpu = torch.empty(
                (self.max_num_reqs, self.num_spec_tokens),
                dtype=torch.int64,
                device="cpu",
                pin_memory=self.pin_memory,
            )
            if self.use_async_scheduling:
                self.valid_sampled_token_count_event = torch.Event()
                self.valid_sampled_token_count_copy_stream = torch.cuda.Stream()
                self.valid_sampled_token_count_cpu = torch.empty(
                    self.max_num_reqs,
                    dtype=torch.int64,
                    device="cpu",
                    pin_memory=self.pin_memory,
                )
671

672
673
        # Ephemeral state transferred between execute_model() and sample_tokens().
        self.execute_model_state: ExecuteModelState | None = None
674
        self.kv_connector_output: KVConnectorOutput | None = None
675
        self.layerwise_nvtx_hooks_registered = False
676

677
678
679
680
    def reset_mm_cache(self) -> None:
        if self.mm_budget:
            self.mm_budget.reset_cache()

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
706
707
708
709
710
711
712
713
714
715
716
717
718
719
720
721
722
723
724
    @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)

725
726
727
728
    def _get_positions(self, num_tokens: Any):
        if isinstance(num_tokens, int):
            if self.uses_mrope:
                return self.mrope_positions.gpu[:, :num_tokens]
729
730
            if self.uses_xdrope_dim > 0:
                return self.xdrope_positions.gpu[:, :num_tokens]
731
732
733
734
            return self.positions.gpu[:num_tokens]
        else:
            if self.uses_mrope:
                return self.mrope_positions.gpu[:, num_tokens]
735
736
            if self.uses_xdrope_dim > 0:
                return self.xdrope_positions.gpu[:, num_tokens]
737
738
            return self.positions.gpu[num_tokens]

739
    def _make_buffer(
740
        self, *size: int | torch.SymInt, dtype: torch.dtype, numpy: bool = True
741
742
743
744
745
746
747
748
    ) -> CpuGpuBuffer:
        return CpuGpuBuffer(
            *size,
            dtype=dtype,
            device=self.device,
            pin_memory=self.pin_memory,
            with_numpy=numpy,
        )
749

750
    def _init_model_kwargs(self):
751
752
        model_kwargs = dict[str, Any]()

753
        if not self.is_pooling_model:
754
755
            return model_kwargs

756
757
        num_reqs = self.input_batch.num_reqs
        pooling_params = self.input_batch.get_pooling_params()
758
759
760

        token_type_id_requests = dict[int, Any]()
        for i, param in enumerate(pooling_params):
761
762
763
764
765
            if (
                param.extra_kwargs is not None
                and (token_types := param.extra_kwargs.get("compressed_token_type_ids"))
                is not None
            ):
766
767
768
769
770
                token_type_id_requests[i] = token_types

        if len(token_type_id_requests) == 0:
            return model_kwargs

771
        seq_lens = self.seq_lens.gpu[:num_reqs]
772
773
774
775
776
777
778
779
        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(
780
781
            device=self.device
        )
782
783
        return model_kwargs

784
    def _may_reorder_batch(self, scheduler_output: "SchedulerOutput") -> None:
785
786
        """
        Update the order of requests in the batch based on the attention
787
        backend's needs. For example, some attention backends (namely MLA) may
788
789
790
791
792
793
        want to separate requests based on if the attention computation will be
        compute-bound or memory-bound.

        Args:
            scheduler_output: The scheduler output.
        """
794
795
796
797
798
799
800
801
        # 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

802
803
804
805
        if self.reorder_batch_threshold is not None:
            reorder_batch_to_split_decodes_and_prefills(
                self.input_batch,
                scheduler_output,
806
807
                decode_threshold=self.reorder_batch_threshold,
            )
808

809
810
    # Note: used for model runner override.
    def _init_device_properties(self) -> None:
811
        """Initialize attributes from torch.cuda.get_device_properties"""
812
813
814
815
816
817
818
        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()

819
    def _update_states(self, scheduler_output: "SchedulerOutput") -> None:
820
821
822
823
824
825
        """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.

826
827
        The SamplingMetadata is updated and copied to the GPU if there is a
        new/resumed/paused/finished request in the batch.
828
829
        """
        # Remove finished requests from the cached states.
830
831
        for req_id in scheduler_output.finished_req_ids:
            self.requests.pop(req_id, None)
832
            self.num_prompt_logprobs.pop(req_id, None)
833
834
835
836
837
838
839
        # 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:
840
            self.input_batch.remove_request(req_id)
841
842

        # Free the cached encoder outputs.
843
844
        for mm_hash in scheduler_output.free_encoder_mm_hashes:
            self.encoder_cache.pop(mm_hash, None)
845

846
847
848
849
850
851
852
        # 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()
853
854
855
856
857
858
859
860
        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)
861
862
863
864
865
        # 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:
866
            self.input_batch.remove_request(req_id)
867

868
        reqs_to_add: list[CachedRequestState] = []
869
        # Add new requests to the cached states.
870
871
872
        for new_req_data in scheduler_output.scheduled_new_reqs:
            req_id = new_req_data.req_id
            sampling_params = new_req_data.sampling_params
873
            pooling_params = new_req_data.pooling_params
874

875
876
877
878
            if (
                sampling_params
                and sampling_params.sampling_type == SamplingType.RANDOM_SEED
            ):
879
880
881
882
883
                generator = torch.Generator(device=self.device)
                generator.manual_seed(sampling_params.seed)
            else:
                generator = None

884
885
            if self.is_pooling_model:
                assert pooling_params is not None
886
887
                task = pooling_params.task
                assert task is not None, "You did not set `task` in the API"
888

889
                model = cast(VllmModelForPooling, self.get_model())
890
                to_update = model.pooler.get_pooling_updates(task)
891
892
                to_update.apply(pooling_params)

893
            req_state = CachedRequestState(
894
                req_id=req_id,
895
                prompt_token_ids=new_req_data.prompt_token_ids,
896
                prompt_embeds=new_req_data.prompt_embeds,
897
                mm_features=new_req_data.mm_features,
898
                sampling_params=sampling_params,
899
                pooling_params=pooling_params,
900
                generator=generator,
901
902
                block_ids=new_req_data.block_ids,
                num_computed_tokens=new_req_data.num_computed_tokens,
903
                output_token_ids=[],
904
                lora_request=new_req_data.lora_request,
905
            )
906
907
            self.requests[req_id] = req_state

908
909
910
911
912
913
914
            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
                )

915
            # Only relevant for models using M-RoPE (e.g, Qwen2-VL)
916
            if self.uses_mrope:
917
                self._init_mrope_positions(req_state)
918

919
920
921
922
            # Only relevant for models using XD-RoPE (e.g, HunYuan-VL)
            if self.uses_xdrope_dim > 0:
                self._init_xdrope_positions(req_state)

923
            reqs_to_add.append(req_state)
924

925
        # Update the states of the running/resumed requests.
926
        is_last_rank = get_pp_group().is_last_rank
927
        req_data = scheduler_output.scheduled_cached_reqs
928
929
930
931
932

        # 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()

933
        for i, req_id in enumerate(req_data.req_ids):
934
            req_state = self.requests[req_id]
935
936
            num_computed_tokens = req_data.num_computed_tokens[i]
            new_block_ids = req_data.new_block_ids[i]
937
            resumed_from_preemption = req_id in req_data.resumed_req_ids
938
            num_output_tokens = req_data.num_output_tokens[i]
939
            req_index = self.input_batch.req_id_to_index.get(req_id)
940

941
942
943
944
945
946
947
948
949
950
951
952
953
954
955
956
957
958
959
960
961
962
963
            # 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)
964

965
            # Update the cached states.
966
            req_state.num_computed_tokens = num_computed_tokens
967
968
969
970
971
972
973
974

            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.
975
976
977
                num_new_tokens = (
                    num_computed_tokens + len(new_token_ids) - req_state.num_tokens
                )
978
979
980
981
                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:
982
                    req_state.output_token_ids.extend(new_token_ids[-num_new_tokens:])
983
984
985
986
987
            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:
988
989
990
991
                    end_idx = (
                        self.input_batch.num_prompt_tokens[req_index]
                        + num_output_tokens
                    )
992
                    self.input_batch.num_tokens_no_spec[req_index] = end_idx
993

994
            # Update the block IDs.
995
            if not resumed_from_preemption:
996
997
                if new_block_ids is not None:
                    # Append the new blocks to the existing block IDs.
998
                    for block_ids, new_ids in zip(req_state.block_ids, new_block_ids):
999
                        block_ids.extend(new_ids)
1000
            else:
1001
                assert req_index is None
1002
                assert new_block_ids is not None
1003
1004
                # The request is resumed from preemption.
                # Replace the existing block IDs with the new ones.
1005
                req_state.block_ids = new_block_ids
1006
1007
1008
1009
1010

            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.
1011
1012
1013
1014
1015
1016
1017

                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:]

1018
                reqs_to_add.append(req_state)
1019
1020
1021
                continue

            # Update the persistent batch.
1022
            self.input_batch.num_computed_tokens_cpu[req_index] = num_computed_tokens
1023
            if new_block_ids is not None:
1024
                self.input_batch.block_table.append_row(new_block_ids, req_index)
1025
1026
1027
1028
1029
1030
1031

            # 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)
1032
                self.input_batch.token_ids_cpu[
1033
1034
1035
                    req_index, start_token_index:end_token_index
                ] = new_token_ids
                self.input_batch.num_tokens_no_spec[req_index] = end_token_index
1036

1037
            # Add spec_token_ids to token_ids_cpu.
1038
            spec_token_ids = scheduler_output.scheduled_spec_decode_tokens.get(
1039
                req_id, []
1040
            )
1041
1042
1043
1044
1045
            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:
1046
1047
1048
                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[
1049
1050
                    req_index, start_index:end_token_index
                ] = spec_token_ids
1051
1052
1053
1054
1055
1056

            # 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.
1057
1058
            self.input_batch.spec_token_ids[req_index].clear()
            self.input_batch.spec_token_ids[req_index].extend(spec_token_ids)
1059

1060
1061
            if self.use_async_scheduling:
                req_state.prev_num_draft_len = num_spec_tokens
1062
1063
        # Add the new or resumed requests to the persistent batch.
        # The smaller empty indices are filled first.
1064
1065
        for request in reqs_to_add:
            self.input_batch.add_request(request)
1066

1067
1068
1069
1070
1071
1072
        # 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()
1073

1074
    def _update_states_after_model_execute(
1075
1076
        self, output_token_ids: torch.Tensor
    ) -> None:
1077
1078
1079
1080
1081
1082
1083
1084
        """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.
        """
1085
        if not self.speculative_config or not self.model_config.is_hybrid:
1086
1087
1088
            return

        # Find the number of accepted tokens for each sequence.
1089
1090
1091
1092
1093
1094
1095
1096
1097
1098
1099
1100
1101
1102
1103
1104
1105
1106
1107
1108
        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()
        )
1109
1110
1111
        for i, num_tokens in enumerate(num_accepted_tokens):
            self.input_batch.num_accepted_tokens_cpu[i] = num_tokens

1112
    def _init_mrope_positions(self, req_state: CachedRequestState):
1113
1114
        model = self.get_model()
        assert supports_mrope(model), "M-RoPE support is not implemented."
1115
1116
1117
1118
        assert req_state.prompt_token_ids is not None, (
            "M-RoPE requires prompt_token_ids to be available."
        )
        mrope_model = cast(SupportsMRoPE, model)
1119
1120

        req_state.mrope_positions, req_state.mrope_position_delta = (
1121
            mrope_model.get_mrope_input_positions(
1122
                req_state.prompt_token_ids,
1123
                req_state.mm_features,
1124
            )
1125
        )
1126

1127
1128
1129
1130
1131
1132
1133
1134
1135
1136
1137
1138
1139
    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,
        )

1140
    def _extract_mm_kwargs(
1141
        self,
1142
1143
        scheduler_output: "SchedulerOutput",
    ) -> BatchedTensorInputs:
1144
        if not scheduler_output or not self.is_multimodal_raw_input_only_model:
1145
            return {}
1146

1147
1148
        mm_kwargs = list[MultiModalKwargsItem]()
        for req in scheduler_output.scheduled_new_reqs:
1149
1150
1151
            for feature in req.mm_features:
                if feature.data is not None:
                    mm_kwargs.append(feature.data)
1152

1153
1154
1155
        # Input all modalities at once
        mm_kwargs_combined: BatchedTensorInputs = {}
        for _, _, mm_kwargs_group in group_mm_kwargs_by_modality(
1156
1157
1158
            mm_kwargs,
            device=self.device,
            pin_memory=self.pin_memory,
1159
1160
        ):
            mm_kwargs_combined.update(mm_kwargs_group)
1161

1162
        return mm_kwargs_combined
1163

1164
    def _dummy_mm_kwargs(self, num_seqs: int) -> BatchedTensorInputs:
1165
        if not self.is_multimodal_raw_input_only_model:
1166
            return {}
1167

1168
1169
1170
1171
1172
        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)
1173

1174
1175
1176
    def _get_cumsum_and_arange(
        self,
        num_tokens: np.ndarray,
1177
        cumsum_dtype: np.dtype | None = None,
1178
1179
1180
1181
1182
1183
1184
1185
1186
1187
1188
1189
1190
1191
1192
1193
    ) -> 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

1194
    def _prepare_input_ids(
1195
1196
1197
1198
        self,
        scheduler_output: "SchedulerOutput",
        total_num_scheduled_tokens: int,
        cu_num_tokens: np.ndarray,
1199
    ) -> None:
1200
        """Prepare the input IDs for the current batch.
1201

1202
1203
1204
1205
1206
1207
1208
        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)
1209
1210
1211
            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)
1212
1213
1214
1215
1216
1217
1218
            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
1219
1220
1221
1222
        sample_flattened_indices: list[int] = []
        spec_flattened_indices: list[int] = []
        prev_common_req_indices: list[int] = []
        prev_draft_token_indices: list[int] = []
1223
1224
        indices_match = True
        max_flattened_index = -1
1225
1226
1227
        total_num_spec_tokens = 0
        scheduled_spec_tokens = scheduler_output.scheduled_spec_decode_tokens

1228
1229
1230
1231
1232
        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.
1233
1234
                draft_len = len(scheduled_spec_tokens.get(req_id, ()))
                total_num_spec_tokens += draft_len
1235
                flattened_index = cu_num_tokens[cur_index].item() - 1
1236
1237
1238
1239
1240
1241
1242
1243
1244
1245
1246
1247
1248
1249
1250
                # 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))
1251
                indices_match &= prev_index == flattened_index
1252
                max_flattened_index = max(max_flattened_index, flattened_index)
1253
1254
1255
        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:
1256
1257
1258
            # 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)
1259
1260
1261
            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)
1262
1263
        if num_commmon_tokens == 0:
            # No requests in common with the previous iteration
1264
            # So input_ids.cpu will have all the input ids.
1265
1266
1267
1268
1269
1270
1271
            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_(
1272
1273
1274
                self.input_batch.prev_sampled_token_ids[:num_commmon_tokens, 0],
                non_blocking=True,
            )
1275
1276
            if self.enable_prompt_embeds:
                self.is_token_ids.gpu[:num_commmon_tokens] = True
1277
            return
1278
        # Upload the index tensors asynchronously so the scatter can be non-blocking.
1279
1280
        sampled_tokens_index_tensor = torch.tensor(
            sample_flattened_indices, dtype=torch.int64, pin_memory=self.pin_memory
1281
        ).to(self.device, non_blocking=True)
1282
        prev_common_req_indices_tensor = torch.tensor(
1283
1284
            prev_common_req_indices, dtype=torch.int64, pin_memory=self.pin_memory
        ).to(self.device, non_blocking=True)
1285
1286
        self.input_ids.gpu.scatter_(
            dim=0,
1287
            index=sampled_tokens_index_tensor,
1288
            src=self.input_batch.prev_sampled_token_ids[
1289
1290
1291
                prev_common_req_indices_tensor, 0
            ],
        )
1292

1293
1294
1295
1296
1297
1298
1299
1300
1301
1302
1303
1304
1305
1306
1307
1308
1309
1310
1311
1312
1313
1314
        # Scatter the draft tokens after the sampled tokens are scattered.
        if self._draft_token_ids is None or not spec_flattened_indices:
            return

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

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

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

1315
1316
    def _get_encoder_seq_lens(
        self,
1317
        num_scheduled_tokens: dict[str, int],
1318
1319
        kv_cache_spec: KVCacheSpec,
        num_reqs: int,
1320
    ) -> tuple[torch.Tensor | None, np.ndarray | None]:
1321
        if not isinstance(kv_cache_spec, CrossAttentionSpec):
1322
            return None, None
1323

1324
1325
        # Zero out buffer for padding requests that are not actually scheduled (CGs)
        self.encoder_seq_lens.np[:num_reqs] = 0
1326
1327
        # Build encoder_seq_lens array mapping request indices to
        # encoder lengths for inputs scheduled in this batch
1328
        for req_id in num_scheduled_tokens:
1329
            req_index = self.input_batch.req_id_to_index[req_id]
1330
1331
1332
1333
1334
1335
1336
1337
1338
1339
1340
1341
1342
1343
1344
1345
            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]
1346

1347
        return encoder_seq_lens, encoder_seq_lens_cpu
1348

1349
    def _prepare_inputs(
1350
1351
1352
        self,
        scheduler_output: "SchedulerOutput",
        num_scheduled_tokens: np.ndarray,
1353
1354
    ) -> tuple[
        torch.Tensor,
1355
        SpecDecodeMetadata | None,
1356
    ]:
1357
1358
        """
        :return: tuple[
1359
            logits_indices, spec_decode_metadata,
1360
1361
        ]
        """
1362
1363
1364
1365
1366
1367
1368
        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.
1369
        self.input_batch.block_table.commit_block_table(num_reqs)
1370
1371
1372

        # Get request indices.
        # E.g., [2, 5, 3] -> [0, 0, 1, 1, 1, 1, 1, 2, 2, 2]
1373
        req_indices = np.repeat(self.arange_np[:num_reqs], num_scheduled_tokens)
1374

1375
1376
        # cu_num_tokens: [2, 5, 3] -> [2, 7, 10]
        # arange: [0, 1, 0, 1, 2, 3, 4, 0, 1, 2]
1377
        cu_num_tokens, arange = self._get_cumsum_and_arange(num_scheduled_tokens)
1378
1379

        # Get positions.
1380
        positions_np = self.positions.np[:total_num_scheduled_tokens]
1381
1382
1383
1384
1385
        np.add(
            self.input_batch.num_computed_tokens_cpu[req_indices],
            arange,
            out=positions_np,
        )
1386

1387
1388
        # Calculate M-RoPE positions.
        # Only relevant for models using M-RoPE (e.g, Qwen2-VL)
1389
        if self.uses_mrope:
1390
1391
            self._calc_mrope_positions(scheduler_output)

1392
1393
1394
1395
1396
        # 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)

1397
1398
1399
1400
        # 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.
1401
1402
1403
        token_indices = (
            positions_np + req_indices * self.input_batch.token_ids_cpu.shape[1]
        )
1404
        token_indices_tensor = torch.from_numpy(token_indices)
1405

1406
1407
1408
        # 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.
1409
1410
1411
1412
1413
1414
        torch.index_select(
            self.input_batch.token_ids_cpu_tensor.flatten(),
            0,
            token_indices_tensor,
            out=self.input_ids.cpu[:total_num_scheduled_tokens],
        )
1415
        if self.enable_prompt_embeds:
1416
            is_token_ids = self.input_batch.is_token_ids_tensor.flatten()
1417
1418
1419
1420
            torch.index_select(
                is_token_ids,
                0,
                token_indices_tensor,
1421
1422
                out=self.is_token_ids.cpu[:total_num_scheduled_tokens],
            )
1423
1424
1425
1426
1427
1428
1429
1430
1431
1432
1433
1434
1435
1436
1437
1438
1439
1440
1441
1442
1443
1444
1445
1446
1447
1448
1449
1450
1451
1452
1453
1454
1455

        # 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:
1456
1457
1458
                    self.inputs_embeds.cpu[
                        output_idx : output_idx + actual_num_sched
                    ].copy_(req_embeds[start_pos:actual_end])
1459
1460

                output_idx += num_sched
1461

1462
1463
        self.input_batch.block_table.compute_slot_mapping(req_indices, positions_np)
        self.input_batch.block_table.commit_slot_mapping(total_num_scheduled_tokens)
1464
1465

        # Prepare the attention metadata.
1466
        self.query_start_loc.np[0] = 0
1467
        self.query_start_loc.np[1 : num_reqs + 1] = cu_num_tokens
1468
1469
        # Note: pad query_start_loc to be non-decreasing, as kernels
        # like FlashAttention requires that
1470
        self.query_start_loc.np[num_reqs + 1 :].fill(cu_num_tokens[-1])
1471
        self.query_start_loc.copy_to_gpu()
1472
        query_start_loc = self.query_start_loc.gpu[: num_reqs + 1]
1473

1474
        self.seq_lens.np[:num_reqs] = (
1475
1476
            self.input_batch.num_computed_tokens_cpu[:num_reqs] + num_scheduled_tokens
        )
1477
        # Fill unused with 0 for full cuda graph mode.
1478
1479
        self.seq_lens.np[num_reqs:].fill(0)
        self.seq_lens.copy_to_gpu()
1480

1481
        num_tokens = [self.requests[r].num_tokens for r in self.input_batch.req_ids]
1482
1483
        num_tokens_np = np.array(num_tokens, dtype=np.int32)

1484
        # Record which requests should not be sampled,
1485
        # so that we could clear the sampled tokens before returning
1486
1487
        self.discard_request_mask.np[:num_reqs] = (
            self.seq_lens.np[:num_reqs] < num_tokens_np
1488
        )
1489
        self.discard_request_mask.copy_to_gpu(num_reqs)
1490

1491
        # Copy the tensors to the GPU.
1492
1493
1494
1495
1496
        self._prepare_input_ids(
            scheduler_output,
            total_num_scheduled_tokens,
            cu_num_tokens,
        )
1497

1498
        if self.uses_mrope:
1499
            # Only relevant for models using M-RoPE (e.g, Qwen2-VL)
1500
1501
            self.mrope_positions.gpu[:, :total_num_scheduled_tokens].copy_(
                self.mrope_positions.cpu[:, :total_num_scheduled_tokens],
1502
1503
                non_blocking=True,
            )
1504
1505
1506
1507
1508
1509
        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,
            )
1510
1511
        else:
            # Common case (1D positions)
1512
            self.positions.copy_to_gpu(total_num_scheduled_tokens)
1513

1514
        use_spec_decode = len(scheduler_output.scheduled_spec_decode_tokens) > 0
1515
1516
1517
1518
1519
1520
1521
        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
1522
            num_draft_tokens = None
1523
            spec_decode_metadata = None
1524
            num_sampled_tokens = np.ones(num_reqs, dtype=np.int32)
1525
1526
1527
1528
1529
        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)
1530
1531
1532
            # 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)
1533
1534
1535
1536
            for (
                req_id,
                draft_token_ids,
            ) in scheduler_output.scheduled_spec_decode_tokens.items():
1537
1538
                req_idx = self.input_batch.req_id_to_index[req_id]
                num_draft_tokens[req_idx] = len(draft_token_ids)
1539
1540
1541
1542
1543
1544
1545
1546
                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
                )
1547
            spec_decode_metadata = self._calc_spec_decode_metadata(
1548
1549
                num_draft_tokens, cu_num_tokens
            )
1550
            logits_indices = spec_decode_metadata.logits_indices
1551
            num_sampled_tokens = num_draft_tokens + 1
1552
            # For DECODE only cuda graph of some attention backends (e.g., GDN).
1553
            self.num_decode_draft_tokens.np[:num_reqs] = num_decode_draft_tokens
1554
1555
            self.num_decode_draft_tokens.np[num_reqs:].fill(-1)
            self.num_decode_draft_tokens.copy_to_gpu()
1556

1557
1558
1559
1560
1561
        # Hot-Swap lora model
        if self.lora_config:
            assert (
                np.sum(num_sampled_tokens)
                <= self.vllm_config.scheduler_config.max_num_batched_tokens
1562
            )
1563
1564
1565
1566
1567
1568
1569
1570
1571
1572
1573
            self.set_active_loras(
                self.input_batch, num_scheduled_tokens, num_sampled_tokens
            )

        return (
            logits_indices,
            spec_decode_metadata,
        )

    def _build_attention_metadata(
        self,
1574
        num_tokens: int,
1575
        num_reqs: int,
1576
1577
1578
        max_query_len: int,
        num_tokens_padded: int | None = None,
        num_reqs_padded: int | None = None,
1579
1580
1581
1582
        ubatch_slices: UBatchSlices | None = None,
        logits_indices: torch.Tensor | None = None,
        use_spec_decode: bool = False,
        for_cudagraph_capture: bool = False,
1583
        num_scheduled_tokens: dict[str, int] | None = None,
1584
1585
1586
1587
1588
        cascade_attn_prefix_lens: list[list[int]] | None = None,
    ) -> tuple[PerLayerAttnMetadata, CommonAttentionMetadata | None]:
        """
        :return: tuple[attn_metadata, spec_decode_common_attn_metadata]
        """
1589
1590
1591
1592
        # Attention metadata is not needed for attention free models
        if len(self.kv_cache_config.kv_cache_groups) == 0:
            return {}, None

1593
1594
        num_tokens_padded = num_tokens_padded or num_tokens
        num_reqs_padded = num_reqs_padded or num_reqs
1595
        assert num_reqs_padded is not None and num_tokens_padded is not None
1596

1597
1598
1599
        attn_metadata: PerLayerAttnMetadata = {}
        if ubatch_slices is not None:
            attn_metadata = [dict() for _ in range(len(ubatch_slices))]
1600

1601
1602
1603
1604
1605
1606
1607
1608
        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()

1609
1610
        if use_spec_decode:
            self.num_accepted_tokens.np[:num_reqs] = (
1611
1612
                self.input_batch.num_accepted_tokens_cpu[:num_reqs]
            )
1613
1614
            self.num_accepted_tokens.np[num_reqs:].fill(1)
            self.num_accepted_tokens.copy_to_gpu()
1615

1616
        kv_cache_groups = self.kv_cache_config.kv_cache_groups
1617

1618
1619
1620
1621
        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):
1622
                blk_table_tensor = torch.zeros(
1623
                    (num_reqs_padded, 1),
1624
                    dtype=torch.int32,
1625
1626
1627
                    device=self.device,
                )
                slot_mapping = torch.zeros(
1628
                    (num_tokens_padded,),
1629
1630
1631
                    dtype=torch.int64,
                    device=self.device,
                )
1632
            else:
1633
                blk_table = self.input_batch.block_table[kv_cache_gid]
1634
1635
                blk_table_tensor = blk_table.get_device_tensor(num_reqs_padded)
                slot_mapping = blk_table.slot_mapping.gpu[:num_tokens_padded]
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
1672
1673
1674
1675
1676
1677
1678
1679
1680
1681
1682
            # 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
            )

1683
1684
1685
1686
1687
1688
1689
1690
1691
        # 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
        ] = {}

1692
1693
1694
1695
1696
1697
1698
        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]
1699
            builder = attn_group.get_metadata_builder(ubid or 0)
1700
1701
1702
1703
            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))
1704

1705
1706
1707
1708
1709
1710
1711
1712
1713
1714
1715
1716
1717
1718
1719
1720
1721
1722
1723
1724
            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
                )
1725
1726
1727
1728
1729
1730
1731
1732
1733
            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,
                )
1734
1735
1736
1737
1738
1739
            else:
                attn_metadata_i = builder.build(
                    common_prefix_len=cascade_attn_prefix_len,
                    common_attn_metadata=common_attn_metadata,
                    **extra_attn_metadata_args,
                )
1740
1741
                if builder.supports_update_block_table:
                    cached_attn_metadata[cache_key] = attn_metadata_i
1742
1743
1744
1745
1746
1747
1748
1749
1750
1751
1752
1753
1754
1755
1756
1757
1758
1759
1760
1761
1762
1763
1764

            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,
1765
            )
1766
1767
1768
1769
            if kv_cache_gid > 0:
                cm.block_table_tensor, cm.slot_mapping = (
                    _get_block_table_and_slot_mapping(kv_cache_gid)
                )
1770

1771
            if self.speculative_config and spec_decode_common_attn_metadata is None:
1772
                if isinstance(self.drafter, EagleProposer):
1773
                    if self.drafter.attn_layer_names[0] in kv_cache_group.layer_names:
1774
                        spec_decode_common_attn_metadata = cm
1775
                else:
1776
                    spec_decode_common_attn_metadata = cm
1777

1778
            for attn_gid in range(len(self.attn_groups[kv_cache_gid])):
1779
                if ubatch_slices is not None:
1780
1781
1782
                    for ubid, _cm in enumerate(split_attn_metadata(ubatch_slices, cm)):
                        _build_attn_group_metadata(kv_cache_gid, attn_gid, _cm, ubid)

1783
                else:
1784
                    _build_attn_group_metadata(kv_cache_gid, attn_gid, cm)
1785

1786
1787
1788
1789
1790
1791
1792
1793
1794
1795
1796
1797
1798
1799
1800
1801
1802
1803
1804
1805
        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]

1806
1807
1808
1809
1810
1811
1812
1813
1814
1815
        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)
            )

1816
        return attn_metadata, spec_decode_common_attn_metadata
1817

1818
1819
1820
    def _compute_cascade_attn_prefix_lens(
        self,
        num_scheduled_tokens: np.ndarray,
1821
        num_computed_tokens: np.ndarray,
1822
1823
1824
1825
1826
1827
1828
        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
        """
1829

1830
1831
1832
1833
1834
1835
1836
1837
1838
1839
1840
1841
1842
1843
        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,
1844
                        num_computed_tokens,
1845
1846
1847
1848
1849
1850
1851
1852
                        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
1853

1854
1855
1856
    def _compute_cascade_attn_prefix_len(
        self,
        num_scheduled_tokens: np.ndarray,
1857
        num_computed_tokens: np.ndarray,
1858
        num_common_prefix_blocks: int,
1859
1860
        kv_cache_spec: KVCacheSpec,
        attn_metadata_builder: AttentionMetadataBuilder,
1861
1862
1863
1864
1865
1866
1867
1868
1869
1870
1871
1872
1873
1874
1875
1876
1877
1878
    ) -> 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.
        """
1879

1880
        common_prefix_len = num_common_prefix_blocks * kv_cache_spec.block_size
1881
1882
1883
1884
1885
1886
1887
1888
1889
1890
1891
1892
1893
1894
1895
1896
1897
1898
1899
1900
1901
1902
1903
1904
1905
1906
1907
1908
1909
1910
1911
1912
1913
1914
1915
1916
1917
        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]
1918
        # Request 3's num_computed_tokens: 3 (i.e., [A, B, C])
1919
1920
1921
1922
1923
        # 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.
1924
        common_prefix_len = min(common_prefix_len, num_computed_tokens.min())
1925
        # common_prefix_len should be a multiple of the block size.
1926
1927
1928
1929
1930
1931
1932
1933
1934
1935
1936
        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
        )
1937
1938
        assert isinstance(kv_cache_spec, AttentionSpec)
        use_cascade = attn_metadata_builder.use_cascade_attention(
1939
1940
1941
            common_prefix_len=common_prefix_len,
            query_lens=num_scheduled_tokens,
            num_query_heads=self.num_query_heads,
1942
            num_kv_heads=kv_cache_spec.num_kv_heads,
1943
            use_alibi=self.use_alibi,
1944
            use_sliding_window=use_sliding_window,
1945
            use_local_attention=use_local_attention,
1946
            num_sms=self.num_sms,
1947
            dcp_world_size=self.dcp_world_size,
1948
1949
1950
        )
        return common_prefix_len if use_cascade else 0

1951
1952
    def _calc_mrope_positions(self, scheduler_output: "SchedulerOutput"):
        mrope_pos_ptr = 0
1953
        for index, req_id in enumerate(self.input_batch.req_ids):
1954
1955
1956
            req = self.requests[req_id]
            assert req.mrope_positions is not None

1957
1958
            num_computed_tokens = self.input_batch.num_computed_tokens_cpu[index]
            num_scheduled_tokens = scheduler_output.num_scheduled_tokens[req_id]
1959
            num_prompt_tokens = length_from_prompt_token_ids_or_embeds(
1960
1961
                req.prompt_token_ids, req.prompt_embeds
            )
1962
1963

            if num_computed_tokens + num_scheduled_tokens > num_prompt_tokens:
1964
1965
                prompt_part_len = max(0, num_prompt_tokens - num_computed_tokens)
                completion_part_len = max(0, num_scheduled_tokens - prompt_part_len)
1966
1967
1968
1969
1970
1971
1972
1973
1974
1975
1976
1977
1978
            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

1979
1980
1981
                self.mrope_positions.cpu[:, dst_start:dst_end] = req.mrope_positions[
                    :, src_start:src_end
                ]
1982
1983
1984
1985
1986
1987
1988
                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

1989
                assert req.mrope_position_delta is not None
1990
                MRotaryEmbedding.get_next_input_positions_tensor(
1991
                    out=self.mrope_positions.np,
1992
1993
1994
1995
1996
                    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,
                )
1997
1998
1999

                mrope_pos_ptr += completion_part_len

2000
2001
2002
2003
2004
2005
2006
2007
2008
2009
2010
2011
2012
2013
2014
2015
2016
2017
2018
2019
2020
2021
2022
2023
2024
2025
2026
2027
2028
2029
2030
2031
2032
2033
2034
2035
2036
2037
2038
2039
2040
2041
2042
2043
2044
2045
2046
    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

2047
2048
    def _calc_spec_decode_metadata(
        self,
2049
2050
2051
2052
2053
2054
2055
2056
2057
2058
2059
2060
2061
2062
2063
2064
        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
2065
2066
2067
2068

        # 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(
2069
2070
            num_sampled_tokens, cumsum_dtype=np.int32
        )
2071
        # Step 2. [0, 0, 0, 0, 103, 104, 104, 104, 206, 207, 207]
2072
        logits_indices = np.repeat(
2073
2074
            cu_num_scheduled_tokens - num_sampled_tokens, num_sampled_tokens
        )
2075
        # Step 3. [0, 1, 2, 3, 103, 104, 105, 106, 206, 207, 208]
2076
2077
2078
2079
2080
2081
        logits_indices += arange

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

        # Compute the draft logits indices.
2082
2083
2084
        # 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(
2085
2086
            num_draft_tokens, cumsum_dtype=np.int32
        )
2087
2088
        # [0, 0, 0, 5, 5, 9]
        target_logits_indices = np.repeat(
2089
2090
            cu_num_sampled_tokens - num_sampled_tokens, num_draft_tokens
        )
2091
2092
2093
2094
2095
        # [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(
2096
2097
            self.device, non_blocking=True
        )
2098
2099
2100
        cu_num_sampled_tokens = torch.from_numpy(cu_num_sampled_tokens).to(
            self.device, non_blocking=True
        )
2101
2102
2103
        logits_indices = torch.from_numpy(logits_indices).to(
            self.device, non_blocking=True
        )
2104
        target_logits_indices = torch.from_numpy(target_logits_indices).to(
2105
2106
            self.device, non_blocking=True
        )
2107
        bonus_logits_indices = torch.from_numpy(bonus_logits_indices).to(
2108
2109
            self.device, non_blocking=True
        )
2110

2111
2112
        # Compute the draft token ids.
        # draft_token_indices:      [  1,   2,   3, 105, 106, 208]
2113
        draft_token_ids = self.input_ids.gpu[logits_indices]
2114
2115
        draft_token_ids = draft_token_ids[target_logits_indices + 1]

2116
        return SpecDecodeMetadata(
2117
2118
2119
            draft_token_ids=draft_token_ids,
            num_draft_tokens=num_draft_tokens.tolist(),
            cu_num_draft_tokens=cu_num_draft_tokens,
2120
            cu_num_sampled_tokens=cu_num_sampled_tokens,
2121
2122
2123
2124
2125
            target_logits_indices=target_logits_indices,
            bonus_logits_indices=bonus_logits_indices,
            logits_indices=logits_indices,
        )

2126
2127
2128
2129
2130
2131
2132
    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
2133
        self.kv_sharing_fast_prefill_logits_indices[:num_logits].copy_(logits_indices)
2134
2135
2136
2137
2138
        # 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_(
2139
2140
2141
2142
2143
2144
            logits_indices[-1].item()
        )
        if (
            self.compilation_config.cudagraph_mode != CUDAGraphMode.NONE
            and num_logits <= self.cudagraph_batch_sizes[-1]
        ):
2145
2146
2147
2148
2149
            # 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
2150
2151
2152
        logits_indices_padded = self.kv_sharing_fast_prefill_logits_indices[
            :num_logits_padded
        ]
2153
2154
        return logits_indices_padded

2155
    def _batch_mm_inputs_from_scheduler(
2156
2157
        self,
        scheduler_output: "SchedulerOutput",
2158
2159
2160
2161
2162
    ) -> tuple[
        list[str],
        list[MultiModalKwargsItem],
        list[tuple[str, PlaceholderRange]],
    ]:
2163
        """Batch multimodal inputs from scheduled encoder inputs.
2164
2165
2166

        Args:
            scheduler_output: The scheduler output containing scheduled encoder
2167
                inputs.
2168
2169

        Returns:
2170
            A tuple of (mm_hashes, mm_kwargs, mm_lora_refs) where:
2171
2172
            - mm_hashes: List of multimodal hashes for each item
            - mm_kwargs: List of multimodal kwargs for each item
2173
            - mm_lora_refs: List of (req_id, placeholder_range) for each item
2174
        """
2175
2176
        scheduled_encoder_inputs = scheduler_output.scheduled_encoder_inputs
        if not scheduled_encoder_inputs:
2177
            return [], [], []
2178
2179

        mm_hashes = list[str]()
2180
        mm_kwargs = list[MultiModalKwargsItem]()
2181
2182
2183
        # Multimodal LoRA reference info to map each multimodal item
        # back to its request & position
        mm_lora_refs = list[tuple[str, PlaceholderRange]]()
2184
2185
        for req_id, encoder_input_ids in scheduled_encoder_inputs.items():
            req_state = self.requests[req_id]
2186
2187

            for mm_input_id in encoder_input_ids:
2188
                mm_feature = req_state.mm_features[mm_input_id]
2189
2190
                if mm_feature.data is None:
                    continue
2191
2192

                mm_hashes.append(mm_feature.identifier)
2193
                mm_kwargs.append(mm_feature.data)
2194
                mm_lora_refs.append((req_id, mm_feature.mm_position))
2195

2196
        return mm_hashes, mm_kwargs, mm_lora_refs
2197

2198
2199
2200
    def _execute_mm_encoder(
        self, scheduler_output: "SchedulerOutput"
    ) -> list[torch.Tensor]:
2201
2202
2203
        mm_hashes, mm_kwargs, mm_lora_refs = self._batch_mm_inputs_from_scheduler(
            scheduler_output
        )
2204
2205

        if not mm_kwargs:
2206
            return []
2207

2208
2209
2210
2211
2212
2213
2214
        # 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.
2215
        model = cast(SupportsMultiModal, self.model)
2216
2217
2218
2219
2220
2221
2222
2223
2224
2225
2226
2227
2228
2229
2230
2231
2232
2233
2234
2235
2236
2237
2238
2239
2240
2241
2242
2243
2244
2245
2246
2247
2248
2249
2250
2251
2252
2253
2254
2255
2256
2257
2258
2259
2260
2261
2262
2263
2264
2265
2266
2267
2268
2269
2270
2271
2272

        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,
                )

2273
        encoder_outputs: list[torch.Tensor] = []
2274
        for modality, num_items, mm_kwargs_group in group_mm_kwargs_by_modality(
2275
2276
2277
            mm_kwargs,
            device=self.device,
            pin_memory=self.pin_memory,
2278
        ):
2279
            curr_group_outputs: MultiModalEmbeddings
2280
2281

            # EVS-related change.
2282
            # (ekhvedchenia): Temporary hack to limit peak memory usage when
2283
            # processing multimodal data. This solves the issue with scheduler
2284
2285
2286
2287
            # 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)
2288
2289
2290
2291
2292
2293
2294
            # 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
            ):
2295
                curr_group_outputs_lst = list[torch.Tensor]()
2296
2297
2298
2299
2300
2301
2302
2303
2304
                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,
                        )
2305
                    )
2306

2307
                    micro_batch_outputs = model.embed_multimodal(
2308
2309
                        **micro_batch_mm_inputs
                    )
2310

2311
2312
2313
                    curr_group_outputs_lst.extend(micro_batch_outputs)

                curr_group_outputs = curr_group_outputs_lst
2314
2315
2316
2317
2318
2319
2320
2321
            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.
2322
                curr_group_outputs = model.embed_multimodal(**mm_kwargs_group)
2323

2324
2325
            sanity_check_mm_encoder_outputs(
                curr_group_outputs,
2326
                expected_num_items=num_items,
2327
            )
2328
            encoder_outputs.extend(curr_group_outputs)
2329

2330
        # Cache the encoder outputs by mm_hash
2331
        for mm_hash, output in zip(mm_hashes, encoder_outputs):
2332
            self.encoder_cache[mm_hash] = output
2333
2334
            logger.debug("Finish execute for mm hash %s", mm_hash)
            self.maybe_save_ec_to_connector(self.encoder_cache, mm_hash)
2335

2336
2337
        return encoder_outputs

2338
    def _gather_mm_embeddings(
2339
2340
        self,
        scheduler_output: "SchedulerOutput",
2341
        shift_computed_tokens: int = 0,
2342
2343
2344
    ) -> tuple[list[torch.Tensor], torch.Tensor]:
        total_num_scheduled_tokens = scheduler_output.total_num_scheduled_tokens

2345
2346
2347
2348
2349
        # Swap to the other buffer to avoid race condition with previous
        # iteration's async copy that may still be reading from CPU.
        self.is_mm_embed_idx = 1 - self.is_mm_embed_idx
        is_mm_embed_buf = self.is_mm_embed_buffers[self.is_mm_embed_idx]

2350
        mm_embeds = list[torch.Tensor]()
2351
        is_mm_embed = is_mm_embed_buf.cpu
2352
2353
2354
        is_mm_embed[:total_num_scheduled_tokens] = False

        req_start_idx = 0
2355
        should_sync_mrope_positions = False
2356
        should_sync_xdrope_positions = False
2357

2358
        for req_id in self.input_batch.req_ids:
2359
2360
            mm_embeds_req: list[torch.Tensor] = []

2361
            num_scheduled_tokens = scheduler_output.num_scheduled_tokens[req_id]
2362
            req_state = self.requests[req_id]
2363
            num_computed_tokens = req_state.num_computed_tokens + shift_computed_tokens
2364

2365
2366
            for mm_feature in req_state.mm_features:
                pos_info = mm_feature.mm_position
2367
2368
                start_pos = pos_info.offset
                num_encoder_tokens = pos_info.length
2369
2370
2371
2372
2373
2374
2375
2376
2377
2378
2379
2380
2381
2382
2383
2384

                # 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,
2385
2386
                    num_encoder_tokens,
                )
2387
                assert start_idx < end_idx
2388
2389
2390
2391
2392
2393
2394
                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
2395

2396
                mm_hash = mm_feature.identifier
2397
                encoder_output = self.encoder_cache.get(mm_hash, None)
2398
                assert encoder_output is not None, f"Encoder cache miss for {mm_hash}."
2399
2400
2401

                if (is_embed := pos_info.is_embed) is not None:
                    is_embed = is_embed[start_idx:end_idx]
2402
2403
2404
                    mm_embeds_item = encoder_output[curr_embeds_start:curr_embeds_end]
                else:
                    mm_embeds_item = encoder_output[start_idx:end_idx]
2405

2406
                req_start_pos = req_start_idx + start_pos - num_computed_tokens
2407
2408
2409
                is_mm_embed[req_start_pos + start_idx : req_start_pos + end_idx] = (
                    True if is_embed is None else is_embed
                )
2410
2411
2412
                mm_embeds_req.append(mm_embeds_item)

            if self.is_multimodal_pruning_enabled and self.uses_mrope:
2413
                assert req_state.mrope_positions is not None
2414
2415
2416
2417
2418
2419
2420
                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,
2421
2422
                    )
                )
2423
2424
2425
2426
                req_state.mrope_positions.copy_(new_mrope_positions)
                req_state.mrope_position_delta = new_delta

            mm_embeds.extend(mm_embeds_req)
2427
2428
            req_start_idx += num_scheduled_tokens

2429
        is_mm_embed = is_mm_embed_buf.copy_to_gpu(total_num_scheduled_tokens)
2430
2431
2432

        if should_sync_mrope_positions:
            self._calc_mrope_positions(scheduler_output)
2433
            self.mrope_positions.copy_to_gpu(total_num_scheduled_tokens)
2434

2435
2436
2437
2438
        if should_sync_xdrope_positions:
            self._calc_xdrope_positions(scheduler_output)
            self.xdrope_positions.copy_to_gpu(total_num_scheduled_tokens)

2439
        return mm_embeds, is_mm_embed
2440

2441
    def get_model(self) -> nn.Module:
2442
        # get raw model out of the cudagraph wrapper.
2443
        if isinstance(self.model, (CUDAGraphWrapper, UBatchWrapper)):
2444
            return self.model.unwrap()
2445
2446
        return self.model

2447
2448
2449
2450
2451
2452
2453
2454
2455
2456
2457
2458
2459
2460
2461
    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

2462
2463
2464
2465
2466
    def get_supported_pooling_tasks(self) -> list[PoolingTask]:
        model = self.get_model()
        if not is_pooling_model(model):
            return []

2467
2468
        supported_tasks = list(model.pooler.get_supported_tasks())

2469
2470
2471
2472
        if "score" in supported_tasks:
            num_labels = getattr(self.model_config.hf_config, "num_labels", 0)
            if num_labels != 1:
                supported_tasks.remove("score")
2473
                logger.debug_once("Score API is only enabled for num_labels == 1.")
2474
2475

        return supported_tasks
2476

2477
2478
2479
2480
2481
2482
2483
2484
2485
2486
    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)

2487
    def sync_and_slice_intermediate_tensors(
2488
2489
        self,
        num_tokens: int,
2490
        intermediate_tensors: IntermediateTensors | None,
2491
2492
        sync_self: bool,
    ) -> IntermediateTensors:
2493
2494
2495
        assert self.intermediate_tensors is not None

        tp = self.vllm_config.parallel_config.tensor_parallel_size
2496
        is_rs = is_residual_scattered_for_sp(self.vllm_config, num_tokens)
2497
2498
2499
2500
2501
2502

        # 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():
2503
                is_scattered = k == "residual" and is_rs
2504
                copy_len = num_tokens // tp if is_scattered else num_tokens
2505
                self.intermediate_tensors[k][:copy_len].copy_(
2506
2507
2508
2509
2510
2511
2512
2513
2514
2515
2516
2517
2518
                    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:
2519
2520
2521
2522
2523
2524
2525
        """
        Step for the EPLB (Expert Parallelism Load Balancing) state.
        """
        if not self.parallel_config.enable_eplb:
            return

        assert self.eplb_state is not None
2526
2527
        model = self.get_model()
        assert is_mixture_of_experts(model)
2528
2529
2530
        self.eplb_state.step(
            is_dummy,
            is_profile,
2531
            log_stats=self.parallel_config.eplb_config.log_balancedness,
2532
2533
        )

2534
2535
2536
2537
2538
    def _pool(
        self,
        hidden_states: torch.Tensor,
        num_scheduled_tokens: int,
        num_scheduled_tokens_np: np.ndarray,
2539
2540
2541
2542
        kv_connector_output: KVConnectorOutput | None,
    ) -> ModelRunnerOutput | AsyncModelRunnerOutput:
        num_reqs = self.input_batch.num_reqs
        assert num_reqs == len(self.input_batch.pooling_params), (
2543
2544
            "Either all or none of the requests in a batch must be pooling request"
        )
2545

2546
        hidden_states = hidden_states[:num_scheduled_tokens]
2547
        seq_lens_cpu = self.seq_lens.cpu[:num_reqs]
2548

2549
        pooling_metadata = self.input_batch.get_pooling_metadata()
2550
        pooling_metadata.build_pooling_cursor(
2551
            num_scheduled_tokens_np, seq_lens_cpu, device=hidden_states.device
2552
        )
2553

2554
2555
        model = cast(VllmModelForPooling, self.model)
        raw_pooler_output: PoolerOutput = model.pooler(
2556
            hidden_states=hidden_states, pooling_metadata=pooling_metadata
2557
        )
2558
2559
2560
2561
2562
2563
2564
2565
2566
2567
2568
2569
2570
2571
2572
2573
2574
2575
2576
2577
2578
2579
2580
2581

        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,
            )

2582
        raw_pooler_output = json_map_leaves(
2583
            lambda x: None if x is None else x.to("cpu", non_blocking=True),
2584
2585
            raw_pooler_output,
        )
2586
2587
2588
2589
        model_runner_output.pooler_output = [
            out if include else None
            for out, include in zip(raw_pooler_output, finished_mask)
        ]
2590
2591
        self._sync_device()

2592
        return model_runner_output
2593

2594
    def _pad_for_sequence_parallelism(self, num_scheduled_tokens: int) -> int:
2595
2596
2597
        # Pad tokens to multiple of tensor_parallel_size when
        # enabled collective fusion for SP
        tp_size = self.vllm_config.parallel_config.tensor_parallel_size
2598
        if self.compilation_config.pass_config.enable_sp and tp_size > 1:
2599
2600
2601
            return round_up(num_scheduled_tokens, tp_size)
        return num_scheduled_tokens

Patrick von Platen's avatar
Patrick von Platen committed
2602
2603
2604
2605
2606
2607
2608
2609
2610
2611
2612
    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

2613
    def _preprocess(
2614
2615
        self,
        scheduler_output: "SchedulerOutput",
2616
        num_input_tokens: int,  # Padded
2617
        intermediate_tensors: IntermediateTensors | None = None,
2618
    ) -> tuple[
2619
2620
        torch.Tensor | None,
        torch.Tensor | None,
2621
        torch.Tensor,
2622
        IntermediateTensors | None,
2623
        dict[str, Any],
2624
        ECConnectorOutput | None,
2625
    ]:
2626
        num_scheduled_tokens = scheduler_output.total_num_scheduled_tokens
2627
        is_first_rank = get_pp_group().is_first_rank
2628
        is_encoder_decoder = self.model_config.is_encoder_decoder
2629

2630
2631
        # _prepare_inputs may reorder the batch, so we must gather multi
        # modal outputs after that to ensure the correct order
2632
2633
        ec_connector_output = None

2634
        if self.supports_mm_inputs and is_first_rank and not is_encoder_decoder:
2635
            # Run the multimodal encoder if any.
2636
2637
2638
2639
2640
2641
            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)
2642

2643
2644
2645
            # 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.
2646
            inputs_embeds_scheduled = self.model.embed_input_ids(
2647
2648
2649
                self.input_ids.gpu[:num_scheduled_tokens],
                multimodal_embeddings=mm_embeds,
                is_multimodal=is_mm_embed,
2650
            )
2651

2652
            # TODO(woosuk): Avoid the copy. Optimize.
2653
            self.inputs_embeds.gpu[:num_scheduled_tokens].copy_(inputs_embeds_scheduled)
2654

Patrick von Platen's avatar
Patrick von Platen committed
2655
            input_ids, inputs_embeds = self._prepare_mm_inputs(num_input_tokens)
2656
            model_kwargs = {
2657
                **self._init_model_kwargs(),
2658
2659
                **self._extract_mm_kwargs(scheduler_output),
            }
2660
        elif self.enable_prompt_embeds and is_first_rank:
2661
2662
2663
2664
2665
2666
2667
2668
2669
2670
2671
2672
            # 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).
2673
2674
2675
            token_ids_idx = (
                self.is_token_ids.gpu[:num_scheduled_tokens]
                .nonzero(as_tuple=False)
2676
                .squeeze(1)
2677
            )
2678
2679
2680
            # Some tokens ids may need to become embeds
            if token_ids_idx.numel() > 0:
                token_ids = self.input_ids.gpu[token_ids_idx]
2681
                tokens_to_embeds = self.model.embed_input_ids(input_ids=token_ids)
2682
2683
2684
                self.inputs_embeds.gpu[token_ids_idx] = tokens_to_embeds

            inputs_embeds = self.inputs_embeds.gpu[:num_input_tokens]
2685
            model_kwargs = self._init_model_kwargs()
2686
            input_ids = None
2687
        else:
2688
2689
2690
2691
            # 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.
2692
            input_ids = self.input_ids.gpu[:num_input_tokens]
2693
            inputs_embeds = None
2694
            model_kwargs = self._init_model_kwargs()
2695

2696
        if self.uses_mrope:
2697
            positions = self.mrope_positions.gpu[:, :num_input_tokens]
2698
2699
        elif self.uses_xdrope_dim > 0:
            positions = self.xdrope_positions.gpu[:, :num_input_tokens]
2700
        else:
2701
            positions = self.positions.gpu[:num_input_tokens]
2702

2703
        if is_first_rank:
2704
2705
            intermediate_tensors = None
        else:
2706
            assert intermediate_tensors is not None
2707
            intermediate_tensors = self.sync_and_slice_intermediate_tensors(
2708
2709
                num_input_tokens, intermediate_tensors, True
            )
2710

2711
        if is_encoder_decoder and scheduler_output.scheduled_encoder_inputs:
2712
2713
2714
2715
2716
2717
2718
            # 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})
2719

2720
2721
2722
2723
2724
2725
        return (
            input_ids,
            inputs_embeds,
            positions,
            intermediate_tensors,
            model_kwargs,
2726
            ec_connector_output,
2727
        )
2728

2729
    def _sample(
2730
        self,
2731
2732
        logits: torch.Tensor | None,
        spec_decode_metadata: SpecDecodeMetadata | None,
2733
    ) -> SamplerOutput:
2734
        # Sample the next token and get logprobs if needed.
2735
        sampling_metadata = self.input_batch.sampling_metadata
2736
        if spec_decode_metadata is None:
2737
2738
2739
            # 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()
2740
            return self.sampler(
2741
2742
2743
                logits=logits,
                sampling_metadata=sampling_metadata,
            )
2744

2745
        sampler_output = self.rejection_sampler(
2746
2747
            spec_decode_metadata,
            None,  # draft_probs
2748
            logits,
2749
2750
            sampling_metadata,
        )
2751
        self._update_states_after_model_execute(sampler_output.sampled_token_ids)
2752
2753
2754
        return sampler_output

    def _bookkeeping_sync(
2755
2756
2757
        self,
        scheduler_output: "SchedulerOutput",
        sampler_output: SamplerOutput,
2758
        logits: torch.Tensor | None,
2759
2760
        hidden_states: torch.Tensor,
        num_scheduled_tokens: int,
2761
        spec_decode_metadata: SpecDecodeMetadata | None,
2762
    ) -> tuple[
2763
        dict[str, int],
2764
        LogprobsLists | None,
2765
        list[list[int]],
2766
        dict[str, LogprobsTensors | None],
2767
2768
2769
        list[str],
        dict[str, int],
        list[int],
2770
    ]:
2771
2772
2773
2774
        num_nans_in_logits = {}
        if envs.VLLM_COMPUTE_NANS_IN_LOGITS:
            num_nans_in_logits = self._get_nans_in_logits(logits)

2775
2776
2777
2778
        num_reqs = self.input_batch.num_reqs
        discard_sampled_tokens_req_indices = np.nonzero(
            self.discard_request_mask.np[:num_reqs]
        )[0]
2779
2780
2781
2782
        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)
2783

2784
2785
2786
        # 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()
2787
        req_id_to_index_output_copy = self.input_batch.req_id_to_index.copy()
2788
2789

        num_sampled_tokens = sampler_output.sampled_token_ids.shape[0]
2790
        sampled_token_ids = sampler_output.sampled_token_ids
2791
        logprobs_tensors = sampler_output.logprobs_tensors
2792
        invalid_req_indices = []
2793
        cu_num_tokens: list[int] | None = None
2794
2795
2796
2797
2798
2799
        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)
2800
2801
2802
                # 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()
2803
2804
            else:
                # Includes spec decode tokens.
2805
                valid_sampled_token_ids, cu_num_tokens = RejectionSampler.parse_output(
2806
2807
                    sampled_token_ids,
                    self.input_batch.vocab_size,
2808
2809
                    discard_sampled_tokens_req_indices,
                    return_cu_num_tokens=logprobs_tensors is not None,
2810
                )
2811
        else:
2812
            valid_sampled_token_ids = []
2813
            invalid_req_indices = discard_sampled_tokens_req_indices.tolist()
2814
2815
2816
2817
2818
            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.
2819
2820
2821
2822
            # 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
2823
2824
2825
2826
2827
            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
            }
2828

2829
2830
2831
2832
2833
        # 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.
2834
        req_ids = self.input_batch.req_ids
2835
2836
        for req_idx in range(num_sampled_tokens):
            if self.use_async_scheduling:
2837
                sampled_ids = [-1] if req_idx not in invalid_req_indices_set else None
2838
2839
            else:
                sampled_ids = valid_sampled_token_ids[req_idx]
2840

2841
            num_sampled_ids: int = len(sampled_ids) if sampled_ids else 0
2842

2843
            if not sampled_ids:
2844
2845
2846
                continue

            start_idx = self.input_batch.num_tokens_no_spec[req_idx]
2847
            end_idx = start_idx + num_sampled_ids
2848
2849
2850
2851
            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}"
2852
            )
2853

2854
2855
            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
2856
            self.input_batch.num_tokens_no_spec[req_idx] = end_idx
2857

2858
            req_id = req_ids[req_idx]
2859
2860
2861
            req_state = self.requests[req_id]
            req_state.output_token_ids.extend(sampled_ids)

2862
        logprobs_lists = (
2863
            logprobs_tensors.tolists(cu_num_tokens)
2864
            if not self.use_async_scheduling and logprobs_tensors is not None
2865
2866
2867
2868
2869
2870
2871
2872
2873
            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,
        )

2874
2875
2876
2877
2878
2879
2880
2881
2882
2883
        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,
        )

2884
2885
2886
2887
2888
2889
2890
2891
2892
2893
2894
2895
2896
2897
2898
    @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()

2899
2900
    def _model_forward(
        self,
2901
2902
2903
2904
        input_ids: torch.Tensor | None = None,
        positions: torch.Tensor | None = None,
        intermediate_tensors: IntermediateTensors | None = None,
        inputs_embeds: torch.Tensor | None = None,
2905
2906
2907
2908
2909
        **model_kwargs: dict[str, Any],
    ) -> Any:
        """Helper method to call the model forward pass.

        This method can be overridden by subclasses for model execution.
2910
        Motivation: We can inspect only this method versus
2911
2912
2913
2914
2915
2916
2917
2918
2919
2920
2921
2922
2923
2924
2925
2926
2927
2928
2929
2930
        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,
        )

2931
2932
2933
2934
2935
2936
2937
2938
2939
2940
2941
2942
2943
2944
2945
2946
2947
2948
2949
2950
2951
    @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
        )

2952
2953
2954
2955
2956
2957
2958
2959
2960
2961
2962
2963
2964
    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,
2965
        num_encoder_reqs: int = 0,
2966
    ) -> tuple[
2967
2968
        CUDAGraphMode,
        BatchDescriptor,
2969
        bool,
2970
2971
        torch.Tensor | None,
        CUDAGraphStat | None,
2972
    ]:
2973
2974
2975
2976
2977
2978
        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,
2979
        )
2980
2981
2982
2983
2984
        # 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
        )
2985
2986
2987
2988
2989
2990
2991

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

2992
        num_tokens_padded = self._pad_for_sequence_parallelism(num_tokens)
2993
        dispatch_cudagraph = (
2994
            lambda num_tokens, disable_full: self.cudagraph_dispatcher.dispatch(
2995
2996
2997
                num_tokens=num_tokens,
                has_lora=has_lora,
                uniform_decode=uniform_decode,
2998
                disable_full=disable_full,
2999
3000
3001
3002
3003
            )
            if not force_eager
            else (CUDAGraphMode.NONE, BatchDescriptor(num_tokens_padded))
        )

3004
        cudagraph_mode, batch_descriptor = dispatch_cudagraph(
3005
            num_tokens_padded, use_cascade_attn or has_encoder_output
3006
        )
3007
        num_tokens_padded = batch_descriptor.num_tokens
3008
3009
3010
3011
3012
3013
3014
3015
3016
        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"
            )
3017
3018
3019

        # Extra coordination when running data-parallel since we need to coordinate
        # across ranks
3020
        should_ubatch, num_tokens_across_dp = False, None
3021
3022
3023
3024
3025
3026
3027
3028
3029
        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
            )

3030
3031
3032
3033
3034
3035
3036
3037
3038
3039
3040
            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,
                )
3041
3042
            )

3043
            # Extract DP-synced values
3044
3045
3046
            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())
3047
3048
3049
3050
3051
                # 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,
                )
3052
3053
3054
3055
                # 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

3056
3057
3058
3059
3060
3061
3062
3063
3064
3065
3066
3067
        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,
3068
            should_ubatch,
3069
3070
3071
            num_tokens_across_dp,
            cudagraph_stats,
        )
3072

3073
3074
3075
3076
3077
3078
3079
3080
3081
3082
3083
3084
3085
3086
3087
3088
3089
3090
3091
3092
3093
3094
3095
3096
3097
3098
3099
3100
3101
3102
3103
3104
3105
3106
3107
3108
    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

3109
3110
3111
3112
    @torch.inference_mode()
    def execute_model(
        self,
        scheduler_output: "SchedulerOutput",
3113
        intermediate_tensors: IntermediateTensors | None = None,
3114
    ) -> ModelRunnerOutput | AsyncModelRunnerOutput | IntermediateTensors | None:
3115
3116
3117
3118
3119
        if self.execute_model_state is not None:
            raise RuntimeError(
                "State error: sample_tokens() must be called "
                "after execute_model() returns None."
            )
3120

3121
        num_scheduled_tokens = scheduler_output.total_num_scheduled_tokens
3122
3123
3124
3125
3126
3127
        with (
            record_function_or_nullcontext("gpu_model_runner: preprocess"),
            self.synchronize_input_prep(),
        ):
            # Update persistent batch states.
            self._update_states(scheduler_output)
3128

3129
3130
            if has_ec_transfer() and get_ec_transfer().is_producer:
                with self.maybe_get_ec_connector_output(
3131
                    scheduler_output,
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
3159
                    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"
3160
3161
                )

3162
3163
3164
3165
3166
3167
            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
3168

3169
3170
3171
3172
            logits_indices, spec_decode_metadata = self._prepare_inputs(
                scheduler_output,
                num_scheduled_tokens_np,
            )
3173

3174
3175
3176
3177
3178
            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(
3179
                    num_scheduled_tokens_np,
3180
3181
                    self.input_batch.num_computed_tokens_cpu[:num_reqs],
                    scheduler_output.num_common_prefix_blocks,
3182
3183
                )

3184
3185
3186
3187
3188
3189
3190
3191
3192
3193
3194
3195
3196
3197
            (
                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),
            )
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
3242
            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,
3243
                )
3244
            )
3245

3246
3247
3248
3249
3250
3251
3252
3253
3254
            (
                input_ids,
                inputs_embeds,
                positions,
                intermediate_tensors,
                model_kwargs,
                ec_connector_output,
            ) = self._preprocess(
                scheduler_output, num_tokens_padded, intermediate_tensors
3255
            )
3256

3257
        # Set cudagraph mode to none if calc_kv_scales is true.
3258
3259
3260
        # KV scales calculation involves dynamic operations that are incompatible
        # with CUDA graph capture.
        if self.calculate_kv_scales:
3261
            cudagraph_mode = CUDAGraphMode.NONE
3262
3263
            # Mark KV scales as calculated after the first forward pass
            self.calculate_kv_scales = False
3264

3265
3266
        # Run the model.
        # Use persistent buffers for CUDA graphs.
3267
3268
        with (
            set_forward_context(
3269
3270
                attn_metadata,
                self.vllm_config,
3271
                num_tokens=num_tokens_padded,
3272
                num_tokens_across_dp=num_tokens_across_dp,
3273
3274
                cudagraph_runtime_mode=cudagraph_mode,
                batch_descriptor=batch_desc,
3275
                ubatch_slices=ubatch_slices_padded,
3276
            ),
3277
            record_function_or_nullcontext("gpu_model_runner: forward"),
3278
3279
            self.maybe_get_kv_connector_output(scheduler_output) as kv_connector_output,
        ):
3280
            model_output = self._model_forward(
3281
3282
3283
3284
3285
3286
3287
                input_ids=input_ids,
                positions=positions,
                intermediate_tensors=intermediate_tensors,
                inputs_embeds=inputs_embeds,
                **model_kwargs,
            )

3288
        with record_function_or_nullcontext("gpu_model_runner: postprocess"):
3289
            if self.use_aux_hidden_state_outputs:
3290
                # True when EAGLE 3 is used.
3291
3292
                hidden_states, aux_hidden_states = model_output
            else:
3293
                # Common case.
3294
3295
3296
                hidden_states = model_output
                aux_hidden_states = None

3297
3298
3299
3300
3301
            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)
3302
                    hidden_states.kv_connector_output = kv_connector_output
3303
                    self.kv_connector_output = kv_connector_output
3304
                    return hidden_states
3305

3306
                if self.is_pooling_model:
3307
                    # Return the pooling output.
3308
3309
3310
3311
3312
                    return self._pool(
                        hidden_states,
                        num_scheduled_tokens,
                        num_scheduled_tokens_np,
                        kv_connector_output,
3313
                    )
3314
3315

                sample_hidden_states = hidden_states[logits_indices]
3316
                logits = self.model.compute_logits(sample_hidden_states)
3317
3318
3319
3320
            else:
                # Rare case.
                assert not self.is_pooling_model

3321
                sample_hidden_states = hidden_states[logits_indices]
3322
                if not get_pp_group().is_last_rank:
3323
                    all_gather_tensors = {
3324
                        "residual": not is_residual_scattered_for_sp(
3325
                            self.vllm_config, num_tokens_padded
3326
                        )
3327
                    }
3328
                    get_pp_group().send_tensor_dict(
3329
3330
                        hidden_states.tensors,
                        all_gather_group=get_tp_group(),
3331
3332
                        all_gather_tensors=all_gather_tensors,
                    )
3333
3334
                    logits = None
                else:
3335
                    logits = self.model.compute_logits(sample_hidden_states)
3336

3337
                model_output_broadcast_data: dict[str, Any] = {}
3338
3339
3340
                if logits is not None:
                    model_output_broadcast_data["logits"] = logits.contiguous()

3341
                broadcasted = get_pp_group().broadcast_tensor_dict(
3342
3343
                    model_output_broadcast_data, src=len(get_pp_group().ranks) - 1
                )
3344
3345
                assert broadcasted is not None
                logits = broadcasted["logits"]
3346

3347
3348
3349
3350
3351
3352
3353
3354
        self.execute_model_state = ExecuteModelState(
            scheduler_output,
            logits,
            spec_decode_metadata,
            spec_decode_common_attn_metadata,
            hidden_states,
            sample_hidden_states,
            aux_hidden_states,
3355
            ec_connector_output,
3356
            cudagraph_stats,
3357
        )
3358
        self.kv_connector_output = kv_connector_output
3359
3360
3361
3362
3363
3364
        return None

    @torch.inference_mode
    def sample_tokens(
        self, grammar_output: "GrammarOutput | None"
    ) -> ModelRunnerOutput | AsyncModelRunnerOutput | IntermediateTensors:
3365
3366
        kv_connector_output = self.kv_connector_output
        self.kv_connector_output = None
3367
3368
        self._draft_token_ids = None
        self._draft_token_req_ids = None
3369

3370
3371
        if self.execute_model_state is None:
            # Nothing to do (PP non-final rank case), output isn't used.
3372
            if not kv_connector_output:
3373
                return None  # type: ignore[return-value]
3374
3375
3376
3377
3378
3379
3380
3381
3382

            # 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
3383
3384
3385
3386
3387
3388
3389
3390
3391
3392

        # Unpack ephemeral state.
        (
            scheduler_output,
            logits,
            spec_decode_metadata,
            spec_decode_common_attn_metadata,
            hidden_states,
            sample_hidden_states,
            aux_hidden_states,
3393
            ec_connector_output,
3394
            cudagraph_stats,
3395
3396
3397
3398
3399
3400
3401
3402
3403
        ) = 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
            )
3404

3405
        with record_function_or_nullcontext("gpu_model_runner: sample"):
3406
3407
            sampler_output = self._sample(logits, spec_decode_metadata)

3408
3409
        self.input_batch.prev_sampled_token_ids = None

3410
        def propose_draft_token_ids(sampled_token_ids):
3411
            assert spec_decode_common_attn_metadata is not None
3412
            with record_function_or_nullcontext("gpu_model_runner: draft"):
3413
3414
3415
3416
3417
3418
3419
3420
3421
3422
                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,
                )
3423
                self._copy_draft_token_ids_to_cpu(scheduler_output)
3424

3425
        spec_config = self.speculative_config
3426
        use_padded_batch_for_eagle = (
3427
3428
3429
            spec_config is not None
            and spec_config.use_eagle()
            and not spec_config.disable_padded_drafter_batch
3430
        )
3431
3432
3433
        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
3434
        if (
3435
3436
3437
            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
3438
        ):
3439
            effective_drafter_max_model_len = (
3440
                spec_config.draft_model_config.max_model_len
3441
            )
3442
        input_fits_in_drafter = spec_decode_common_attn_metadata and (
3443
            spec_decode_common_attn_metadata.max_seq_len + self.num_spec_tokens
3444
3445
            <= effective_drafter_max_model_len
        )
3446
        if use_padded_batch_for_eagle:
3447
3448
            assert self.speculative_config is not None
            assert isinstance(self.drafter, EagleProposer)
3449
3450
3451
3452
3453
3454
            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:
3455
                assert spec_decode_common_attn_metadata is not None
3456
3457
3458
3459
3460
3461
                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,
3462
                        self.discard_request_mask.gpu,
3463
3464
3465
3466
3467
                    )
                )
                self._copy_valid_sampled_token_count(
                    next_token_ids, valid_sampled_tokens_count
                )
3468
3469
3470
3471
3472
3473
                # Since we couldn't run the drafter,
                # just use zeros for the draft tokens.
                self._draft_token_ids = torch.zeros(
                    1, device=self.device, dtype=torch.int32
                ).expand(len(self.input_batch.req_ids), self.num_spec_tokens)
                self._copy_draft_token_ids_to_cpu(scheduler_output, zeros_only=True)
3474

3475
        with record_function_or_nullcontext("gpu_model_runner: bookkeep"):
3476
3477
3478
3479
3480
3481
3482
3483
            (
                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,
3484
3485
3486
3487
3488
            ) = self._bookkeeping_sync(
                scheduler_output,
                sampler_output,
                logits,
                hidden_states,
3489
                scheduler_output.total_num_scheduled_tokens,
3490
                spec_decode_metadata,
3491
            )
3492

3493
3494
3495
3496
3497
        if (
            self.speculative_config
            and not use_padded_batch_for_eagle
            and input_fits_in_drafter
        ):
3498
3499
3500
            # 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)
3501

3502
        with record_function_or_nullcontext("gpu_model_runner: eplb"):
3503
            self.eplb_step()
3504
3505
3506
3507
3508
3509
3510
3511
        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,
3512
3513
3514
                ec_connector_output=ec_connector_output
                if self.supports_mm_inputs
                else None,
3515
                num_nans_in_logits=num_nans_in_logits,
3516
                cudagraph_stats=cudagraph_stats,
3517
            )
3518

3519
3520
        if not self.use_async_scheduling:
            return output
3521
3522
3523
3524
3525
3526
3527
3528
3529
        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,
3530
                vocab_size=self.input_batch.vocab_size,
3531
3532
3533
3534
3535
            )
        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
3536
            # any requests with sampling params that require output ids.
3537
3538
3539
3540
            self.input_batch.set_async_sampled_token_ids(
                async_output.sampled_token_ids_cpu,
                async_output.async_copy_ready_event,
            )
3541
3542
3543

        return async_output

3544
    def take_draft_token_ids(self) -> DraftTokenIds | None:
3545
        if not self.num_spec_tokens or not self._draft_token_req_ids:
3546
            return None
3547
3548
3549
        req_ids = self._draft_token_req_ids
        draft_token_ids = self._get_draft_token_ids_cpu(len(req_ids))
        return DraftTokenIds(req_ids, draft_token_ids)
3550

3551
3552
3553
3554
3555
3556
3557
3558
3559
3560
    def _copy_draft_token_ids_to_cpu(
        self, scheduler_output: "SchedulerOutput", zeros_only: bool = False
    ) -> None:
        struct_output = scheduler_output.has_structured_output_requests
        if self.use_async_scheduling and not struct_output:
            # Draft tokens don't need to be copied to the CPU if async
            # scheduling is in use and there are no structured output reqs.
            return
        # We must also set the corresponding request ids.
        self._draft_token_req_ids = self.input_batch.req_ids.copy()
3561

3562
3563
3564
3565
3566
3567
3568
3569
3570
3571
3572
3573
3574
3575
3576
3577
3578
3579
3580
3581
3582
3583
3584
3585
3586
3587
3588
        draft_token_ids: torch.Tensor = self._draft_token_ids
        if not torch.is_tensor(draft_token_ids):
            return
        assert self.draft_token_ids_event is not None
        assert self.draft_token_ids_copy_stream is not None
        assert self.draft_token_ids_cpu is not None
        default_stream = torch.cuda.current_stream()
        num_reqs = draft_token_ids.shape[0]
        with torch.cuda.stream(self.draft_token_ids_copy_stream):
            if not zeros_only:
                # Trigger async copy of draft token ids to cpu.
                self.draft_token_ids_copy_stream.wait_stream(default_stream)
                self.draft_token_ids_cpu[:num_reqs].copy_(
                    draft_token_ids, non_blocking=True
                )
            else:
                # No copy needed, just zero-out cpu tensor.
                self.draft_token_ids_cpu[:num_reqs] = 0
            self.draft_token_ids_event.record()

    def _get_draft_token_ids_cpu(self, num_reqs: int) -> list[list[int]]:
        if isinstance(self._draft_token_ids, list):
            return self._draft_token_ids
        assert self.draft_token_ids_event is not None
        assert self.draft_token_ids_cpu is not None
        self.draft_token_ids_event.synchronize()
        return self.draft_token_ids_cpu[:num_reqs].tolist()
3589

3590
3591
3592
3593
3594
3595
3596
3597
3598
3599
3600
3601
3602
    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
3603
            assert counts_cpu is not None
3604
3605
3606
3607
3608
3609
3610
3611
            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
3612
3613
        sampled_count_event = self.valid_sampled_token_count_event
        if sampled_count_event is None or prev_sampled_token_ids is None:
3614
3615
3616
            return []

        counts_cpu = self.valid_sampled_token_count_cpu
3617
3618
        assert counts_cpu is not None
        sampled_count_event.synchronize()
3619
3620
        return counts_cpu[: prev_sampled_token_ids.shape[0]].tolist()

3621
3622
3623
    def propose_draft_token_ids(
        self,
        scheduler_output: "SchedulerOutput",
3624
        sampled_token_ids: torch.Tensor | list[list[int]],
3625
3626
3627
        sampling_metadata: SamplingMetadata,
        hidden_states: torch.Tensor,
        sample_hidden_states: torch.Tensor,
3628
3629
        aux_hidden_states: list[torch.Tensor] | None,
        spec_decode_metadata: SpecDecodeMetadata | None,
3630
        common_attn_metadata: CommonAttentionMetadata,
3631
    ) -> list[list[int]] | torch.Tensor:
3632
        num_scheduled_tokens = scheduler_output.total_num_scheduled_tokens
3633
3634
3635
        spec_config = self.speculative_config
        assert spec_config is not None
        if spec_config.method == "ngram":
3636
            assert isinstance(sampled_token_ids, list)
3637
            assert isinstance(self.drafter, NgramProposer)
3638
            draft_token_ids = self.drafter.propose(
3639
3640
                sampled_token_ids,
                self.input_batch.req_ids,
3641
3642
                self.input_batch.num_tokens_no_spec,
                self.input_batch.token_ids_cpu,
3643
3644
                self.input_batch.spec_decode_unsupported_reqs,
            )
3645
        elif spec_config.method == "suffix":
3646
3647
3648
            assert isinstance(sampled_token_ids, list)
            assert isinstance(self.drafter, SuffixDecodingProposer)
            draft_token_ids = self.drafter.propose(self.input_batch, sampled_token_ids)
3649
        elif spec_config.method == "medusa":
3650
            assert isinstance(sampled_token_ids, list)
3651
            assert isinstance(self.drafter, MedusaProposer)
3652

3653
3654
            if sample_hidden_states.shape[0] == len(sampled_token_ids):
                # The input to the target model does not include draft tokens.
3655
3656
3657
3658
                hidden_states = sample_hidden_states
            else:
                indices = []
                offset = 0
3659
3660
3661
                assert spec_decode_metadata is not None, (
                    "No spec decode metadata for medusa"
                )
3662
                for num_draft, tokens in zip(
3663
3664
                    spec_decode_metadata.num_draft_tokens, sampled_token_ids
                ):
3665
                    indices.append(offset + len(tokens) - 1)
3666
                    offset += num_draft + 1
3667
                indices = torch.tensor(indices, device=self.device)
3668
3669
                hidden_states = sample_hidden_states[indices]

3670
            draft_token_ids = self.drafter.propose(
3671
3672
3673
                target_hidden_states=hidden_states,
                sampling_metadata=sampling_metadata,
            )
3674
        elif spec_config.use_eagle():
3675
            assert isinstance(self.drafter, EagleProposer)
3676

3677
            if spec_config.disable_padded_drafter_batch:
3678
3679
3680
                # 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.
3681
3682
                assert isinstance(sampled_token_ids, list), (
                    "sampled_token_ids should be a python list when"
3683
                    "padded-batch is disabled."
3684
                )
3685
                next_token_ids = self.drafter.prepare_next_token_ids_cpu(
3686
3687
3688
3689
3690
                    sampled_token_ids,
                    self.requests,
                    self.input_batch,
                    scheduler_output.num_scheduled_tokens,
                )
3691
3692
3693
3694
3695
            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.
3696
3697
                assert isinstance(sampled_token_ids, torch.Tensor), (
                    "sampled_token_ids should be a torch.Tensor when"
3698
                    "padded-batch is enabled."
3699
3700
                )
                next_token_ids, valid_sampled_tokens_count = (
3701
3702
3703
3704
3705
                    self.drafter.prepare_next_token_ids_padded(
                        common_attn_metadata,
                        sampled_token_ids,
                        self.requests,
                        self.input_batch,
3706
                        self.discard_request_mask.gpu,
3707
                    )
3708
                )
3709
3710
3711
                self._copy_valid_sampled_token_count(
                    next_token_ids, valid_sampled_tokens_count
                )
Jiayi Yao's avatar
Jiayi Yao committed
3712

3713
            num_rejected_tokens_gpu = None
3714
            if spec_decode_metadata is None:
3715
                token_indices_to_sample = None
3716
                # input_ids can be None for multimodal models.
3717
                target_token_ids = self.input_ids.gpu[:num_scheduled_tokens]
3718
                target_positions = self._get_positions(num_scheduled_tokens)
3719
                if self.use_aux_hidden_state_outputs:
Wentao Ye's avatar
Wentao Ye committed
3720
                    assert aux_hidden_states is not None
3721
                    target_hidden_states = torch.cat(
3722
3723
                        [h[:num_scheduled_tokens] for h in aux_hidden_states], dim=-1
                    )
3724
3725
                else:
                    target_hidden_states = hidden_states[:num_scheduled_tokens]
3726
            else:
3727
                if spec_config.disable_padded_drafter_batch:
3728
                    token_indices_to_sample = None
3729
3730
3731
3732
3733
                    common_attn_metadata, token_indices = self.drafter.prepare_inputs(
                        common_attn_metadata,
                        sampled_token_ids,
                        spec_decode_metadata.num_draft_tokens,
                    )
3734
3735
3736
3737
3738
3739
3740
3741
3742
                    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]
3743
                else:
3744
3745
3746
3747
3748
3749
3750
3751
                    (
                        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,
3752
                    )
3753
3754
3755
3756
3757
3758
3759
3760
3761
3762
3763
                    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]
3764

3765
            if self.supports_mm_inputs:
3766
3767
3768
3769
3770
3771
                mm_embed_inputs = self._gather_mm_embeddings(
                    scheduler_output,
                    shift_computed_tokens=1,
                )
            else:
                mm_embed_inputs = None
3772

3773
            draft_token_ids = self.drafter.propose(
3774
3775
3776
3777
                target_token_ids=target_token_ids,
                target_positions=target_positions,
                target_hidden_states=target_hidden_states,
                next_token_ids=next_token_ids,
3778
                last_token_indices=token_indices_to_sample,
3779
                sampling_metadata=sampling_metadata,
3780
                common_attn_metadata=common_attn_metadata,
3781
                mm_embed_inputs=mm_embed_inputs,
3782
                num_rejected_tokens_gpu=num_rejected_tokens_gpu,
3783
            )
3784

3785
        return draft_token_ids
3786

3787
3788
3789
    def update_config(self, overrides: dict[str, Any]) -> None:
        allowed_config_names = {"load_config", "model_config"}
        for config_name, config_overrides in overrides.items():
3790
3791
            assert config_name in allowed_config_names, (
                f"Config `{config_name}` not supported. "
3792
                f"Allowed configs: {allowed_config_names}"
3793
            )
3794
3795
3796
3797
            config = getattr(self, config_name)
            new_config = update_config(config, config_overrides)
            setattr(self, config_name, new_config)

3798
3799
3800
3801
3802
    def load_model(self, eep_scale_up: bool = False) -> None:
        """
        Args:
            eep_scale_up: the model loading is for elastic EP scale up.
        """
3803
3804
3805
3806
3807
        logger.info_once(
            "Starting to load model %s...",
            self.model_config.model,
            scope="global",
        )
3808
3809
3810
3811
3812
        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)
        )
3813

3814
3815
3816
3817
        if self.parallel_config.enable_eplb:
            self.eplb_state = EplbState(self.parallel_config, self.device)
            eplb_models = 0

3818
3819
3820
3821
3822
3823
3824
3825
3826
3827
        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
3828
                    )
3829
3830
3831
3832
3833
3834
3835
3836
3837
3838
3839
3840
3841
3842
3843
                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,
                        )
3844

3845
3846
3847
3848
3849
3850
3851
3852
3853
3854
3855
3856
3857
3858
3859
3860
3861
3862
3863
3864
3865
3866
                        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
3867

3868
3869
3870
3871
3872
3873
                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"
                        )
3874

3875
3876
3877
3878
3879
3880
3881
3882
3883
3884
3885
3886
3887
3888
3889
3890
3891
3892
3893
3894
3895
3896
3897
3898
3899
                    # 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
3900
        logger.info_once(
3901
3902
            "Model loading took %s GiB memory and %.6f seconds",
            format_gib(self.model_memory_usage),
3903
            time_after_load - time_before_load,
3904
            scope="local",
3905
        )
3906
        prepare_communication_buffer_for_model(self.model)
3907
3908
3909
3910
        if (drafter := getattr(self, "drafter", None)) and (
            drafter_model := getattr(drafter, "model", None)
        ):
            prepare_communication_buffer_for_model(drafter_model)
3911
        mm_config = self.model_config.multimodal_config
3912
        self.is_multimodal_pruning_enabled = (
3913
            supports_multimodal_pruning(self.get_model())
3914
3915
            and mm_config is not None
            and mm_config.is_multimodal_pruning_enabled()
3916
        )
3917

3918
        if is_mixture_of_experts(self.model) and self.parallel_config.enable_eplb:
3919
3920
3921
3922
3923
3924
3925
3926
3927
3928
3929
            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(
3930
                self.model,
3931
                self.model_config,
3932
3933
3934
                global_expert_load,
                old_global_expert_indices,
                rank_mapping,
3935
            )
3936
3937
            if self.eplb_state.is_async:
                self.eplb_state.start_async_loop(rank_mapping=rank_mapping)
3938

3939
        if (
3940
3941
            self.vllm_config.compilation_config.mode
            == CompilationMode.STOCK_TORCH_COMPILE
3942
            and supports_dynamo()
3943
        ):
3944
            backend = self.vllm_config.compilation_config.init_backend(self.vllm_config)
3945
            compilation_counter.stock_torch_compile_count += 1
3946
            self.model.compile(fullgraph=True, backend=backend)
3947
            return
3948
        # for other compilation modes, cudagraph behavior is controlled by
3949
3950
3951
        # CudagraphWraper and CudagraphDispatcher of vllm.

        # wrap the model with full cudagraph wrapper if needed.
3952
3953
        cudagraph_mode = self.compilation_config.cudagraph_mode
        assert cudagraph_mode is not None
3954
3955
3956
3957
        if (
            cudagraph_mode.has_full_cudagraphs()
            and not self.parallel_config.use_ubatching
        ):
3958
3959
3960
            self.model = CUDAGraphWrapper(
                self.model, self.vllm_config, runtime_mode=CUDAGraphMode.FULL
            )
3961
        elif self.parallel_config.use_ubatching:
3962
            if cudagraph_mode.has_full_cudagraphs():
3963
3964
3965
                self.model = UBatchWrapper(
                    self.model, self.vllm_config, CUDAGraphMode.FULL, self.device
                )
3966
            else:
3967
3968
3969
                self.model = UBatchWrapper(
                    self.model, self.vllm_config, CUDAGraphMode.NONE, self.device
                )
3970

3971
    def _get_eagle3_aux_layers_from_config(self) -> tuple[int, ...] | None:
3972
3973
3974
3975
3976
3977
3978
3979
3980
3981
3982
3983
3984
3985
3986
3987
3988
3989
3990
3991
3992
3993
3994
        """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

3995
    def reload_weights(self) -> None:
3996
        assert getattr(self, "model", None) is not None, (
3997
            "Cannot reload weights before model is loaded."
3998
        )
3999
4000
        model_loader = get_model_loader(self.load_config)
        logger.info("Reloading weights inplace...")
4001
        model_loader.load_weights(self.get_model(), model_config=self.model_config)
4002

4003
4004
4005
4006
4007
    def save_tensorized_model(
        self,
        tensorizer_config: "TensorizerConfig",
    ) -> None:
        TensorizerLoader.save_model(
4008
            self.get_model(),
4009
            tensorizer_config=tensorizer_config,
4010
            model_config=self.model_config,
4011
4012
        )

4013
4014
4015
    def _get_prompt_logprobs_dict(
        self,
        hidden_states: torch.Tensor,
4016
        num_scheduled_tokens: dict[str, int],
4017
    ) -> dict[str, LogprobsTensors | None]:
4018
        num_prompt_logprobs_dict = self.num_prompt_logprobs
4019
4020
4021
        if not num_prompt_logprobs_dict:
            return {}

4022
        in_progress_dict = self.input_batch.in_progress_prompt_logprobs_cpu
4023
        prompt_logprobs_dict: dict[str, LogprobsTensors | None] = {}
4024
4025
4026
4027
4028

        # 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():
4029
4030
4031
4032
            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
4033
4034
4035

            # Get metadata for this request.
            request = self.requests[req_id]
4036
4037
4038
4039
            if request.prompt_token_ids is None:
                # Prompt logprobs is incompatible with prompt embeddings
                continue

4040
4041
            num_prompt_tokens = len(request.prompt_token_ids)
            prompt_token_ids = torch.tensor(request.prompt_token_ids).to(
4042
4043
                self.device, non_blocking=True
            )
4044

4045
4046
4047
4048
4049
4050
            # 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(
4051
4052
                    num_prompt_tokens - 1, num_prompt_logprobs + 1
                )
4053
4054
                in_progress_dict[req_id] = logprobs_tensors

4055
            # Determine number of logits to retrieve.
4056
4057
            start_idx = request.num_computed_tokens
            start_tok = start_idx + 1
4058
            num_remaining_tokens = num_prompt_tokens - start_tok
4059
            if num_tokens <= num_remaining_tokens:
4060
                # This is a chunk, more tokens remain.
4061
4062
4063
                # 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.
4064
4065
4066
4067
4068
                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)
4069
4070
4071
4072
4073
4074
4075
                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
4076
4077
4078
4079
4080

            # 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]
4081
            offset = self.query_start_loc.np[req_idx].item()
4082
            prompt_hidden_states = hidden_states[offset : offset + num_logits]
4083
            logits = self.model.compute_logits(prompt_hidden_states)
4084
4085
4086
4087

            # 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.
4088
            tgt_token_ids = prompt_token_ids[start_tok : start_tok + num_logits]
4089
4090

            # Compute prompt logprobs.
4091
4092
            logprobs = self.sampler.compute_logprobs(logits)
            token_ids, logprobs, ranks = self.sampler.gather_logprobs(
4093
4094
                logprobs, num_prompt_logprobs, tgt_token_ids
            )
4095
4096

            # Transfer GPU->CPU async.
4097
4098
            chunk_slice = slice(start_idx, start_idx + num_logits)
            logprobs_tensors.logprob_token_ids[chunk_slice].copy_(
4099
4100
4101
                token_ids, non_blocking=True
            )
            logprobs_tensors.logprobs[chunk_slice].copy_(logprobs, non_blocking=True)
4102
            logprobs_tensors.selected_token_ranks[chunk_slice].copy_(
4103
4104
                ranks, non_blocking=True
            )
4105
4106
4107
4108
4109

        # 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]
4110
            del in_progress_dict[req_id]
4111
4112

        # Must synchronize the non-blocking GPU->CPU transfers.
4113
        if prompt_logprobs_dict:
4114
            self._sync_device()
4115
4116
4117

        return prompt_logprobs_dict

4118
4119
    def _get_nans_in_logits(
        self,
4120
        logits: torch.Tensor | None,
4121
4122
4123
4124
4125
4126
4127
4128
4129
4130
4131
    ) -> 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])
4132
4133
4134
                    if num_nans_for_index is not None and req_index < logits.shape[0]
                    else 0
                )
4135
4136
4137
4138
            return num_nans_in_logits
        except IndexError:
            return {}

4139
    @contextmanager
4140
4141
4142
    def maybe_randomize_inputs(
        self, input_ids: torch.Tensor | None, inputs_embeds: torch.Tensor | None
    ):
4143
4144
4145
4146
        """
        Randomize input_ids if VLLM_RANDOMIZE_DP_DUMMY_INPUTS is set.
        This is to help balance expert-selection
         - during profile_run
4147
         - during DP rank dummy run
4148
        """
4149

4150
4151
4152
4153
        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
4154
        elif input_ids is not None:
4155
4156
4157
4158

            @functools.cache
            def rand_input_ids() -> torch.Tensor:
                return torch.randint_like(
4159
                    self.input_ids.gpu,
4160
4161
                    low=0,
                    high=self.model_config.get_vocab_size(),
4162
                )
4163

4164
            logger.debug_once("Randomizing dummy input_ids for DP Rank")
4165
            input_ids.copy_(rand_input_ids()[: input_ids.size(0)], non_blocking=True)
4166
4167
            yield
            input_ids.fill_(0)
4168
4169
4170
4171
4172
4173
4174
4175
4176
4177
4178
4179
4180
4181
4182
        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)
4183

4184
4185
4186
4187
4188
4189
    def _get_mm_dummy_batch(
        self,
        modality: str,
        max_items_per_batch: int,
    ) -> BatchedTensorInputs:
        """Dummy data for profiling and precompiling multimodal models."""
4190
4191
        assert self.mm_budget is not None

4192
        dummy_decoder_data = self.mm_registry.get_decoder_dummy_data(
4193
            model_config=self.model_config,
4194
            seq_len=self.max_model_len,
4195
            mm_counts={modality: 1},
4196
            cache=self.mm_budget.cache,
4197
4198
4199
4200
        )
        dummy_mm_data = dummy_decoder_data.multi_modal_data

        # Result in the maximum GPU consumption of the model
4201
4202
        dummy_mm_item = dummy_mm_data[modality][0]
        dummy_mm_items = [dummy_mm_item] * max_items_per_batch
4203

4204
4205
4206
4207
4208
4209
4210
4211
        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,
            )
        )
4212

4213
4214
4215
4216
    @torch.inference_mode()
    def _dummy_run(
        self,
        num_tokens: int,
4217
        cudagraph_runtime_mode: CUDAGraphMode | None = None,
4218
4219
        force_attention: bool = False,
        uniform_decode: bool = False,
4220
        allow_microbatching: bool = True,
4221
4222
        skip_eplb: bool = False,
        is_profile: bool = False,
4223
        create_mixed_batch: bool = False,
4224
        remove_lora: bool = True,
4225
        activate_lora: bool = False,
Rémi Delacourt's avatar
Rémi Delacourt committed
4226
        is_graph_capturing: bool = False,
4227
    ) -> tuple[torch.Tensor, torch.Tensor]:
4228
4229
4230
4231
4232
4233
4234
        """
        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.
4235
                - if not set will determine the cudagraph mode based on using
4236
                    the self.cudagraph_dispatcher.
4237
4238
4239
4240
                - CUDAGraphMode.NONE: No cudagraph, for warm up and profile run
                - CUDAGraphMode.PIECEWISE: Piecewise cudagraph.
                - CUDAGraphMode.FULL: Full cudagraph, attention metadata is
                    needed.
4241
            force_attention: If True, always create attention metadata. Used to
4242
4243
4244
4245
                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.
4246
4247
            create_mixed_batch: If True, create a mixed batch with both decode
                (1 token) and prefill (multiple tokens) requests.
4248
            remove_lora: If False, dummy LoRAs are not destroyed after the run
4249
            activate_lora: If False, dummy_run is performed without LoRAs.
4250
        """
4251
4252
4253
4254
4255
        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([])

4256
4257
4258
4259
        assert (
            cudagraph_runtime_mode is None
            or cudagraph_runtime_mode.valid_runtime_modes()
        )
4260

4261
        # If cudagraph_mode.decode_mode() == FULL and
4262
        # cudagraph_mode.separate_routine(). This means that we are using
4263
4264
4265
4266
4267
4268
4269
4270
4271
4272
4273
        # 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.
4274
        max_query_len = self.uniform_decode_query_len if uniform_decode else num_tokens
4275

4276
4277
4278
4279
4280
        # 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
4281
4282
4283
4284
        if create_mixed_batch:
            assert not uniform_decode
            # Create mixed batch:
            # first half decode tokens, second half one prefill
4285
            num_decode_tokens = min(max_num_reqs - 1, num_tokens // 2)
4286
4287
4288
4289
            num_prefill_tokens = num_tokens - num_decode_tokens
            num_reqs = num_decode_tokens + 1

            # Create decode requests (1 token each) followed by prefill request
4290
            num_scheduled_tokens_list = [1] * num_decode_tokens + [num_prefill_tokens]
4291
4292
4293
            # Note: Overriding max_query_len to be the prefill tokens
            max_query_len = num_prefill_tokens
        elif uniform_decode:
4294
            assert not create_mixed_batch
4295
            num_reqs = min(max_num_reqs, cdiv(num_tokens, max_query_len))
4296
4297
            num_scheduled_tokens_list = [max_query_len] * num_reqs
            if num_tokens % max_query_len != 0:
4298
                num_scheduled_tokens_list[-1] = num_tokens % max_query_len
4299
4300
4301
4302
4303
4304
        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

4305
4306
        assert sum(num_scheduled_tokens_list) == num_tokens
        assert len(num_scheduled_tokens_list) == num_reqs
4307
        num_scheduled_tokens = np.array(num_scheduled_tokens_list, dtype=np.int32)
4308
4309
        num_tokens_unpadded = int(num_scheduled_tokens.sum())

4310
        num_sampled_tokens = np.ones(num_reqs, dtype=np.int32)
4311

4312
        _cudagraph_mode, batch_desc, should_ubatch, num_tokens_across_dp, _ = (
4313
4314
4315
4316
4317
4318
4319
4320
4321
4322
4323
4324
4325
4326
4327
4328
4329
4330
            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,
4331
4332
            )
        )
4333
4334
4335

        if cudagraph_runtime_mode is None:
            cudagraph_runtime_mode = _cudagraph_mode
4336
        else:
4337
4338
4339
4340
4341
4342
4343
4344
4345
            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
        )
4346
        ubatch_slices, ubatch_slices_padded = maybe_create_ubatch_slices(
4347
4348
4349
4350
4351
4352
4353
4354
4355
4356
            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,
4357
        )
4358

4359
        attn_metadata: PerLayerAttnMetadata | None = None
4360
4361
4362

        # If force_attention is True, we always capture attention. Otherwise,
        # it only happens for cudagraph_runtime_mode=FULL.
4363
        if force_attention or cudagraph_runtime_mode == CUDAGraphMode.FULL:
4364
4365
4366
4367
4368
4369
            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:
4370
                seq_lens = max_query_len  # type: ignore[assignment]
4371
            self.seq_lens.np[:num_reqs] = seq_lens
4372
4373
            self.seq_lens.np[num_reqs:] = 0
            self.seq_lens.copy_to_gpu()
4374

4375
4376
            cum_num_tokens, _ = self._get_cumsum_and_arange(num_scheduled_tokens)
            self.query_start_loc.np[1 : num_reqs + 1] = cum_num_tokens
4377
4378
            self.query_start_loc.copy_to_gpu()

4379
            pad_attn = cudagraph_runtime_mode == CUDAGraphMode.FULL
4380
            attn_metadata, _ = self._build_attention_metadata(
4381
4382
4383
                num_tokens=num_tokens_unpadded,
                num_reqs=num_reqs_padded,
                max_query_len=max_query_len,
4384
                ubatch_slices=ubatch_slices_padded if pad_attn else ubatch_slices,
4385
                for_cudagraph_capture=is_graph_capturing,
4386
            )
4387

4388
        with self.maybe_dummy_run_with_lora(
4389
4390
4391
4392
4393
            self.lora_config,
            num_scheduled_tokens,
            num_sampled_tokens,
            activate_lora,
            remove_lora,
4394
        ):
4395
            # Make sure padding doesn't exceed max_num_tokens
4396
            assert num_tokens_padded <= self.max_num_tokens
4397
            model_kwargs = self._init_model_kwargs()
4398
            if self.supports_mm_inputs and not self.model_config.is_encoder_decoder:
Patrick von Platen's avatar
Patrick von Platen committed
4399
4400
                input_ids, inputs_embeds = self._prepare_mm_inputs(num_tokens_padded)

4401
                model_kwargs = {
4402
                    **model_kwargs,
4403
4404
                    **self._dummy_mm_kwargs(num_reqs),
                }
4405
4406
            elif self.enable_prompt_embeds:
                input_ids = None
4407
                inputs_embeds = self.inputs_embeds.gpu[:num_tokens_padded]
4408
                model_kwargs = self._init_model_kwargs()
4409
            else:
4410
                input_ids = self.input_ids.gpu[:num_tokens_padded]
4411
                inputs_embeds = None
4412

4413
            if self.uses_mrope:
4414
                positions = self.mrope_positions.gpu[:, :num_tokens_padded]
4415
            elif self.uses_xdrope_dim > 0:
4416
                positions = self.xdrope_positions.gpu[:, :num_tokens_padded]
4417
            else:
4418
                positions = self.positions.gpu[:num_tokens_padded]
4419
4420
4421
4422
4423
4424
4425
4426
4427

            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,
4428
4429
4430
                            device=self.device,
                        )
                    )
4431
4432

                intermediate_tensors = self.sync_and_slice_intermediate_tensors(
4433
                    num_tokens_padded, None, False
4434
                )
4435

4436
            if ubatch_slices_padded is not None:
4437
4438
4439
                # Adjust values to reflect a single ubatch.
                # TODO(sage,lucas): this is cruft that should be addressed in
                #  the padding refactor.
4440
                num_tokens_padded = ubatch_slices_padded[0].num_tokens
4441
                if num_tokens_across_dp is not None:
4442
                    num_tokens_across_dp[:] = num_tokens_padded
4443

4444
            with (
4445
                self.maybe_randomize_inputs(input_ids, inputs_embeds),
4446
                set_forward_context(
4447
4448
                    attn_metadata,
                    self.vllm_config,
4449
                    num_tokens=num_tokens_padded,
4450
4451
                    num_tokens_across_dp=num_tokens_across_dp,
                    cudagraph_runtime_mode=cudagraph_runtime_mode,
4452
                    batch_descriptor=batch_desc,
4453
                    ubatch_slices=ubatch_slices_padded,
4454
4455
                ),
            ):
4456
                outputs = self.model(
4457
4458
4459
4460
                    input_ids=input_ids,
                    positions=positions,
                    intermediate_tensors=intermediate_tensors,
                    inputs_embeds=inputs_embeds,
4461
                    **model_kwargs,
4462
                )
4463

4464
4465
4466
4467
            if self.use_aux_hidden_state_outputs:
                hidden_states, _ = outputs
            else:
                hidden_states = outputs
4468

4469
            if self.speculative_config and self.speculative_config.use_eagle():
4470
                assert isinstance(self.drafter, EagleProposer)
4471
4472
4473
                # 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.
4474
                use_cudagraphs = (
4475
4476
4477
4478
4479
4480
4481
4482
4483
                    (
                        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
4484
4485
4486
4487
4488
4489
4490
4491
4492
4493
4494

                # 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
4495
                    is_graph_capturing=is_graph_capturing,
4496
                )
4497

4498
4499
4500
4501
4502
4503
4504
4505
4506
4507
4508
        # 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()

4509
4510
4511
4512
4513
4514
4515
4516
4517
4518
        # 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)

4519
        logit_indices = np.cumsum(num_scheduled_tokens) - 1
4520
4521
4522
4523
        logit_indices_device = torch.from_numpy(logit_indices).to(
            self.device, non_blocking=True
        )
        return hidden_states, hidden_states[logit_indices_device]
4524
4525
4526
4527
4528
4529

    @torch.inference_mode()
    def _dummy_sampler_run(
        self,
        hidden_states: torch.Tensor,
    ) -> torch.Tensor:
4530
4531
4532
        # The dummy hidden states may contain special values,
        # like `inf` or `nan`.
        # To avoid breaking the sampler, we use a random tensor here instead.
4533
4534
4535
4536
4537

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

4538
        hidden_states = torch.rand_like(hidden_states)
4539

4540
        logits = self.model.compute_logits(hidden_states)
4541
4542
        num_reqs = logits.size(0)

4543
        dummy_tensors = lambda v: torch.full((num_reqs,), v, device=self.device)
4544
4545
4546
4547
4548
4549
4550
4551
4552
4553
4554
4555
4556
4557
4558

        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)],
4559
            spec_token_ids=[[] for _ in range(num_reqs)],
4560
4561
            allowed_token_ids_mask=None,
            bad_words_token_ids={},
4562
            logitsprocs=LogitsProcessors(),
4563
        )
4564
        try:
4565
4566
4567
            sampler_output = self.sampler(
                logits=logits, sampling_metadata=dummy_metadata
            )
4568
        except RuntimeError as e:
4569
            if "out of memory" in str(e):
4570
4571
4572
4573
                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 "
4574
4575
                    "initializing the engine."
                ) from e
4576
4577
            else:
                raise e
4578
        if self.speculative_config:
4579
4580
            draft_token_ids = [[0] for _ in range(num_reqs)]
            dummy_spec_decode_metadata = SpecDecodeMetadata.make_dummy(
4581
4582
                draft_token_ids, self.device
            )
4583
4584
4585
4586
4587
4588

            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
4589
4590
4591
4592
4593
            logits = torch.randn(
                num_tokens + num_reqs,
                logits.shape[-1],
                device=self.device,
                dtype=logits.dtype,
4594
            )
4595
4596
4597
            self.rejection_sampler(
                dummy_spec_decode_metadata,
                draft_probs,
4598
                logits,
4599
4600
                dummy_metadata,
            )
4601
        return sampler_output
4602

4603
    def _dummy_pooler_run_task(
4604
4605
        self,
        hidden_states: torch.Tensor,
4606
4607
        task: PoolingTask,
    ) -> PoolerOutput:
4608
4609
4610
4611
        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
4612
4613
4614
4615
        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
4616
4617
4618

        req_num_tokens = num_tokens // num_reqs

4619
        dummy_prompt_lens = torch.from_numpy(num_scheduled_tokens_np)
4620
4621
4622
        dummy_token_ids = torch.zeros(
            (num_reqs, req_num_tokens), dtype=torch.int32, device=self.device
        )
4623

4624
        model = cast(VllmModelForPooling, self.get_model())
4625
        dummy_pooling_params = PoolingParams(task=task)
4626
        dummy_pooling_params.verify(task=task, model_config=self.model_config)
4627
        to_update = model.pooler.get_pooling_updates(task)
4628
4629
        to_update.apply(dummy_pooling_params)

4630
        dummy_metadata = PoolingMetadata(
4631
4632
4633
            prompt_lens=dummy_prompt_lens,
            prompt_token_ids=dummy_token_ids,
            pooling_params=[dummy_pooling_params] * num_reqs,
4634
            pooling_states=[PoolingStates() for i in range(num_reqs)],
4635
        )
4636

4637
        dummy_metadata.build_pooling_cursor(
4638
            num_scheduled_tokens_np,
4639
4640
            seq_lens_cpu=dummy_prompt_lens,
            device=hidden_states.device,
4641
        )
4642

4643
        try:
4644
4645
4646
            return model.pooler(
                hidden_states=hidden_states, pooling_metadata=dummy_metadata
            )
4647
        except RuntimeError as e:
4648
            if "out of memory" in str(e):
4649
                raise RuntimeError(
4650
4651
4652
                    "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 "
4653
4654
                    "initializing the engine."
                ) from e
4655
4656
            else:
                raise e
4657
4658
4659
4660
4661
4662

    @torch.inference_mode()
    def _dummy_pooler_run(
        self,
        hidden_states: torch.Tensor,
    ) -> PoolerOutput:
4663
4664
4665
4666
        if supports_mm_encoder_only(self.model):
            # MM Encoder only model not need to run pooler.
            return torch.tensor([])

4667
        # Find the task that has the largest output for subsequent steps
4668
4669
4670
        supported_pooling_tasks = self.get_supported_pooling_tasks()

        if not supported_pooling_tasks:
4671
4672
4673
4674
4675
4676
            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."
            )
4677

4678
        output_size = dict[PoolingTask, float]()
4679
        for task in supported_pooling_tasks:
4680
4681
            # Run a full batch with each task to ensure none of them OOMs
            output = self._dummy_pooler_run_task(hidden_states, task)
4682
            output_size[task] = sum(o.nbytes for o in output if o is not None)
4683
4684
4685
4686
            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)
4687

4688
    def profile_run(self) -> None:
4689
        # Profile with multimodal encoder & encoder cache.
4690
        if self.supports_mm_inputs:
4691
4692
            mm_config = self.model_config.multimodal_config
            if mm_config is not None and mm_config.skip_mm_profiling:
4693
                logger.info(
4694
                    "Skipping memory profiling for multimodal encoder and "
4695
4696
                    "encoder cache."
                )
4697
4698
4699
4700
4701
4702
4703
4704
            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.
4705
                    dummy_modality = mm_budget.get_modality_with_max_tokens()
4706
4707
4708
                    max_mm_items_per_batch = mm_budget.max_items_per_batch_by_modality[
                        dummy_modality
                    ]
4709
4710
4711
4712
4713
4714
4715
4716
4717

                    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,
                    )
4718

4719
4720
4721
4722
4723
                    # Create dummy batch of multimodal inputs.
                    batched_dummy_mm_inputs = self._get_mm_dummy_batch(
                        dummy_modality,
                        max_mm_items_per_batch,
                    )
4724

4725
                    # Run multimodal encoder.
4726
                    dummy_encoder_outputs = self.model.embed_multimodal(
4727
4728
                        **batched_dummy_mm_inputs
                    )
4729

4730
4731
4732
4733
                    sanity_check_mm_encoder_outputs(
                        dummy_encoder_outputs,
                        expected_num_items=max_mm_items_per_batch,
                    )
4734
4735
                    for i, output in enumerate(dummy_encoder_outputs):
                        self.encoder_cache[f"tmp_{i}"] = output
4736

4737
        # Add `is_profile` here to pre-allocate communication buffers
4738
4739
4740
        hidden_states, last_hidden_states = self._dummy_run(
            self.max_num_tokens, is_profile=True
        )
4741
        if get_pp_group().is_last_rank:
4742
4743
4744
4745
            if self.is_pooling_model:
                output = self._dummy_pooler_run(hidden_states)
            else:
                output = self._dummy_sampler_run(last_hidden_states)
4746
        else:
4747
            output = None
4748
        self._sync_device()
4749
        del hidden_states, output
4750
        self.encoder_cache.clear()
4751
        gc.collect()
4752

4753
    def capture_model(self) -> int:
4754
        if self.compilation_config.cudagraph_mode == CUDAGraphMode.NONE:
4755
            logger.warning(
4756
                "Skipping CUDA graph capture. To turn on CUDA graph capture, "
4757
4758
                "ensure `cudagraph_mode` was not manually set to `NONE`"
            )
4759
            return 0
4760

4761
4762
        compilation_counter.num_gpu_runner_capture_triggers += 1

4763
4764
        start_time = time.perf_counter()

4765
4766
4767
4768
4769
4770
4771
4772
4773
4774
4775
4776
4777
4778
        @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()
4779
                    gc.collect()
4780

4781
4782
4783
        # 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.
4784
        set_cudagraph_capturing_enabled(True)
4785
        with freeze_gc(), graph_capture(device=self.device):
4786
            start_free_gpu_memory = torch.cuda.mem_get_info()[0]
4787
            cudagraph_mode = self.compilation_config.cudagraph_mode
4788
            assert cudagraph_mode is not None
4789
4790
4791
4792
4793
4794
4795
4796
4797

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

4798
4799
            if cudagraph_mode.mixed_mode() != CUDAGraphMode.NONE:
                cudagraph_runtime_mode = cudagraph_mode.mixed_mode()
4800
                # make sure we capture the largest batch size first
4801
4802
4803
                compilation_cases = list(
                    product(reversed(self.cudagraph_batch_sizes), lora_cases)
                )
4804
4805
4806
                self._capture_cudagraphs(
                    compilation_cases,
                    cudagraph_runtime_mode=cudagraph_runtime_mode,
4807
4808
                    uniform_decode=False,
                )
4809

4810
4811
            # Capture full cudagraph for uniform decode batches if we
            # don't already have full mixed prefill-decode cudagraphs.
4812
4813
4814
4815
4816
4817
4818
            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
                )
4819
                decode_cudagraph_batch_sizes = [
4820
4821
                    x
                    for x in self.cudagraph_batch_sizes
4822
                    if max_num_tokens >= x >= self.uniform_decode_query_len
4823
                ]
4824
4825
4826
                compilation_cases_decode = list(
                    product(reversed(decode_cudagraph_batch_sizes), lora_cases)
                )
4827
4828
4829
                self._capture_cudagraphs(
                    compilation_cases=compilation_cases_decode,
                    cudagraph_runtime_mode=CUDAGraphMode.FULL,
4830
4831
                    uniform_decode=True,
                )
4832

4833
4834
4835
            torch.cuda.synchronize()
            end_free_gpu_memory = torch.cuda.mem_get_info()[0]

4836
4837
4838
        # 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
4839
        # we may do lazy capturing in future that still allows capturing
4840
4841
        # after here.
        set_cudagraph_capturing_enabled(False)
4842

4843
4844
4845
4846
        # Lock workspace to prevent resizing during execution.
        # Max workspace sizes should have been captured during warmup/profiling.
        lock_workspace()

4847
4848
4849
4850
        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.
4851
        logger.info_once(
4852
4853
4854
            "Graph capturing finished in %.0f secs, took %.2f GiB",
            elapsed_time,
            cuda_graph_size / (1 << 30),
4855
            scope="local",
4856
        )
4857
        return cuda_graph_size
4858

4859
4860
    def _capture_cudagraphs(
        self,
4861
        compilation_cases: list[tuple[int, bool]],
4862
4863
4864
4865
4866
4867
4868
        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}"
4869
4870
4871
4872
4873
4874
4875
4876

        # 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",
4877
4878
4879
                    cudagraph_runtime_mode.name,
                ),
            )
4880

4881
        # We skip EPLB here since we don't want to record dummy metrics
4882
        for num_tokens, activate_lora in compilation_cases:
4883
            # We currently only capture ubatched graphs when its a FULL
4884
4885
4886
            # 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
4887
            allow_microbatching = (
4888
                self.parallel_config.use_ubatching
4889
4890
                and cudagraph_runtime_mode == CUDAGraphMode.FULL
                and uniform_decode
4891
4892
4893
4894
4895
                and check_ubatch_thresholds(
                    config=self.vllm_config.parallel_config,
                    num_tokens=num_tokens,
                    uniform_decode=uniform_decode,
                )
4896
            )
4897

4898
4899
4900
4901
4902
4903
            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.
4904
4905
4906
4907
4908
4909
4910
4911
4912
                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,
4913
                    activate_lora=activate_lora,
4914
4915
4916
4917
4918
4919
4920
4921
                )
            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,
4922
                activate_lora=activate_lora,
Rémi Delacourt's avatar
Rémi Delacourt committed
4923
                is_graph_capturing=True,
4924
            )
4925
        self.maybe_remove_all_loras(self.lora_config)
4926

4927
4928
4929
4930
    def initialize_attn_backend(self, kv_cache_config: KVCacheConfig) -> None:
        """
        Initialize the attention backends and attention metadata builders.
        """
4931
        assert len(self.attn_groups) == 0, "Attention backends are already initialized"
4932

4933
4934
4935
4936
4937
4938
        class AttentionGroupKey(NamedTuple):
            attn_backend: type[AttentionBackend]
            kv_cache_spec: KVCacheSpec

        def get_attn_backends_for_group(
            kv_cache_group_spec: KVCacheGroupSpec,
4939
        ) -> tuple[dict[AttentionGroupKey, list[str]], set[type[AttentionBackend]]]:
4940
            layer_type = cast(type[Any], AttentionLayerBase)
4941
            layers = get_layers_from_vllm_config(
4942
                self.vllm_config, layer_type, kv_cache_group_spec.layer_names
4943
            )
4944
4945
            attn_backends = {}
            attn_backend_layers = defaultdict(list)
4946
            # Dedupe based on full class name; this is a bit safer than
4947
4948
4949
4950
            # 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.
4951
            for layer_name in kv_cache_group_spec.layer_names:
4952
                attn_backend = layers[layer_name].get_attn_backend()
4953
4954
4955
4956

                if layer_name in self.kv_sharing_fast_prefill_eligible_layers:
                    attn_backend = create_fast_prefill_custom_backend(
                        "FastPrefill",
4957
                        attn_backend,  # type: ignore[arg-type]
4958
4959
                    )

4960
4961
4962
                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):
4963
                    layer_kv_cache_spec = layer_kv_cache_spec.kv_cache_specs[layer_name]
4964
                key = (full_cls_name, layer_kv_cache_spec)
4965
4966
4967
                attn_backends[key] = AttentionGroupKey(
                    attn_backend, layer_kv_cache_spec
                )
4968
                attn_backend_layers[key].append(layer_name)
4969
4970
4971
4972
            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()),
            )
4973
4974

        def create_attn_groups(
4975
            attn_backends_map: dict[AttentionGroupKey, list[str]],
4976
            kv_cache_group_id: int,
4977
4978
        ) -> list[AttentionGroup]:
            attn_groups: list[AttentionGroup] = []
4979
            for (attn_backend, kv_cache_spec), layer_names in attn_backends_map.items():
4980
                attn_group = AttentionGroup(
4981
                    attn_backend,
4982
                    layer_names,
4983
                    kv_cache_spec,
4984
                    kv_cache_group_id,
4985
4986
                )

4987
4988
4989
                attn_groups.append(attn_group)
            return attn_groups

4990
        attention_backend_maps = []
4991
        attention_backend_list = []
4992
        for kv_cache_group_spec in kv_cache_config.kv_cache_groups:
4993
            attn_backends = get_attn_backends_for_group(kv_cache_group_spec)
4994
            attention_backend_maps.append(attn_backends[0])
4995
            attention_backend_list.append(attn_backends[1])
4996
4997

        # Resolve cudagraph_mode before actually initialize metadata_builders
4998
4999
5000
        self._check_and_update_cudagraph_mode(
            attention_backend_list, kv_cache_config.kv_cache_groups
        )
5001

5002
5003
5004
        # Check if attention backend supports PCP&DCP and related features.
        check_attention_cp_compatibility(self.vllm_config)

5005
5006
        for i, attn_backend_map in enumerate(attention_backend_maps):
            self.attn_groups.append(create_attn_groups(attn_backend_map, i))
5007

5008
5009
5010
5011
5012
5013
5014
5015
5016
5017
5018
5019
5020
5021
5022
    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
5023
5024
                    if not self.parallel_config.use_ubatching
                    else self.parallel_config.num_ubatches,
5025
                )
co63oc's avatar
co63oc committed
5026
        # Calculate reorder batch threshold (if needed)
5027
5028
        # Note (tdoublep): do this *after* constructing builders,
        # because some of them change the threshold at init time.
5029
5030
        self.calculate_reorder_batch_threshold()

5031
    def _check_and_update_cudagraph_mode(
5032
5033
5034
        self,
        attention_backends: list[set[type[AttentionBackend]]],
        kv_cache_groups: list[KVCacheGroupSpec],
5035
    ) -> None:
5036
        """
5037
        Resolve the cudagraph_mode when there are multiple attention
5038
        groups with potential conflicting CUDA graph support.
5039
5040
5041
        Then initialize the cudagraph_dispatcher based on the resolved
        cudagraph_mode.
        """
5042
        min_cg_support = AttentionCGSupport.ALWAYS
5043
        min_cg_backend_name = None
5044

5045
5046
5047
5048
5049
5050
5051
5052
5053
5054
5055
5056
        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__
5057
5058
        # Flexible resolve the cudagraph mode
        cudagraph_mode = self.compilation_config.cudagraph_mode
5059
        assert cudagraph_mode is not None
5060
        # check cudagraph for mixed batch is supported
5061
5062
5063
5064
5065
5066
        if (
            cudagraph_mode.mixed_mode() == CUDAGraphMode.FULL
            and min_cg_support != AttentionCGSupport.ALWAYS
        ):
            msg = (
                f"CUDAGraphMode.{cudagraph_mode.name} is not supported "
5067
                f"with {min_cg_backend_name} backend (support: "
5068
5069
                f"{min_cg_support})"
            )
5070
5071
            if min_cg_support == AttentionCGSupport.NEVER:
                # if not supported any full cudagraphs, just raise it.
5072
5073
                msg += (
                    "; please try cudagraph_mode=PIECEWISE, and "
5074
                    "make sure compilation mode is VLLM_COMPILE"
5075
                )
5076
5077
5078
5079
5080
                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"
5081
                cudagraph_mode = self.compilation_config.cudagraph_mode = (
5082
                    CUDAGraphMode.FULL_AND_PIECEWISE
5083
                )
5084
5085
            else:
                msg += "; setting cudagraph_mode=FULL_DECODE_ONLY"
5086
                cudagraph_mode = self.compilation_config.cudagraph_mode = (
5087
                    CUDAGraphMode.FULL_DECODE_ONLY
5088
                )
5089
5090
            logger.warning(msg)

5091
        # check that if we are doing decode full-cudagraphs it is supported
5092
5093
5094
5095
5096
5097
        if (
            cudagraph_mode.decode_mode() == CUDAGraphMode.FULL
            and min_cg_support == AttentionCGSupport.NEVER
        ):
            msg = (
                f"CUDAGraphMode.{cudagraph_mode.name} is not supported "
5098
                f"with {min_cg_backend_name} backend (support: "
5099
5100
                f"{min_cg_support})"
            )
5101
            if self.compilation_config.mode == CompilationMode.VLLM_COMPILE and (
5102
5103
5104
5105
5106
                self.compilation_config.splitting_ops_contain_attention()
                or self.compilation_config.use_inductor_graph_partition
            ):
                msg += (
                    "; setting cudagraph_mode=PIECEWISE because "
5107
                    "attention is compiled piecewise"
5108
5109
                )
                cudagraph_mode = self.compilation_config.cudagraph_mode = (
5110
                    CUDAGraphMode.PIECEWISE
5111
                )
5112
            else:
5113
5114
                msg += (
                    "; setting cudagraph_mode=NONE because "
5115
                    "attention is not compiled piecewise"
5116
5117
                )
                cudagraph_mode = self.compilation_config.cudagraph_mode = (
5118
                    CUDAGraphMode.NONE
5119
                )
5120
5121
            logger.warning(msg)

5122
5123
        # check that if we are doing spec-decode + decode full-cudagraphs it is
        # supported
5124
5125
5126
5127
5128
5129
5130
5131
        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 "
5132
                f"{min_cg_backend_name} (support: {min_cg_support})"
5133
            )
5134
5135
            if self.compilation_config.splitting_ops_contain_attention():
                msg += "; setting cudagraph_mode=PIECEWISE"
5136
                cudagraph_mode = self.compilation_config.cudagraph_mode = (
5137
                    CUDAGraphMode.PIECEWISE
5138
                )
5139
5140
            else:
                msg += "; setting cudagraph_mode=NONE"
5141
                cudagraph_mode = self.compilation_config.cudagraph_mode = (
5142
                    CUDAGraphMode.NONE
5143
                )
5144
5145
5146
5147
            logger.warning(msg)

        # double check that we can support full cudagraph if they are requested
        # even after automatic downgrades
5148
5149
5150
5151
5152
5153
        if (
            cudagraph_mode.has_full_cudagraphs()
            and min_cg_support == AttentionCGSupport.NEVER
        ):
            raise ValueError(
                f"CUDAGraphMode.{cudagraph_mode.name} is not "
5154
                f"supported with {min_cg_backend_name} backend ("
5155
5156
                f"support:{min_cg_support}) "
                "; please try cudagraph_mode=PIECEWISE, "
5157
                "and make sure compilation mode is VLLM_COMPILE"
5158
            )
5159

5160
5161
5162
5163
        # 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
5164
        # Will be removed in the near future when we have separate cudagraph capture
5165
5166
5167
5168
5169
5170
5171
5172
5173
        # 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
            )
5174
5175
5176
5177
            capture_sizes = self.compilation_config.cudagraph_capture_sizes
            self.cudagraph_batch_sizes = (
                capture_sizes if capture_sizes is not None else []
            )
5178

5179
5180
        # Trigger cudagraph dispatching keys initialization after
        # resolved cudagraph mode.
5181
        self.compilation_config.cudagraph_mode = cudagraph_mode
5182
        self.cudagraph_dispatcher.initialize_cudagraph_keys(
5183
            cudagraph_mode, self.uniform_decode_query_len
5184
        )
5185

5186
5187
    def calculate_reorder_batch_threshold(self) -> None:
        """
5188
5189
5190
5191
        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.
5192
        """
5193
5194
        min_none_high = lambda a, b: a if b is None else b if a is None else min(a, b)

5195
        reorder_batch_thresholds: list[int | None] = [
5196
5197
5198
            group.get_metadata_builder().reorder_batch_threshold
            for group in self._attn_group_iterator()
        ]
5199
5200
5201
5202
5203
        # 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
5204
        self.reorder_batch_threshold = reduce(min_none_high, reorder_batch_thresholds)  # type: ignore[assignment]
5205

5206
5207
5208
    @staticmethod
    def select_common_block_size(
        kv_manager_block_size: int, attn_groups: list[AttentionGroup]
5209
5210
    ) -> int:
        """
5211
5212
5213
5214
5215
        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.
5216
5217
5218
5219
5220
5221

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

        Returns:
5222
            The selected block size
5223
5224

        Raises:
5225
            ValueError: If no valid block size found
5226
5227
        """

5228
5229
5230
5231
5232
5233
5234
5235
        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
5236
                for supported_size in backend.get_supported_kernel_block_sizes():
5237
5238
5239
5240
5241
5242
5243
5244
5245
5246
5247
5248
5249
5250
5251
5252
5253
5254
5255
5256
5257
5258
5259
5260
5261
5262
5263
5264
5265
5266
                    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
5267
            for supported_size in backend.get_supported_kernel_block_sizes()
5268
5269
            if isinstance(supported_size, int)
        )
5270

5271
5272
5273
5274
5275
5276
        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}. ")
5277

5278
5279
5280
    def may_reinitialize_input_batch(
        self, kv_cache_config: KVCacheConfig, kernel_block_sizes: list[int]
    ) -> None:
5281
5282
5283
5284
5285
5286
5287
        """
        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.
5288
            kernel_block_sizes: The kernel block sizes for each KV cache group.
5289
5290
5291
5292
        """
        block_sizes = [
            kv_cache_group.kv_cache_spec.block_size
            for kv_cache_group in kv_cache_config.kv_cache_groups
5293
            if not isinstance(kv_cache_group.kv_cache_spec, EncoderOnlyAttentionSpec)
5294
        ]
5295
5296
5297
5298

        if block_sizes != [self.cache_config.block_size] or kernel_block_sizes != [
            self.cache_config.block_size
        ]:
5299
5300
5301
            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
5302
5303
                "for more details."
            )
5304
5305
            self.input_batch = InputBatch(
                max_num_reqs=self.max_num_reqs,
5306
                max_model_len=max(self.max_model_len, self.max_encoder_len),
5307
5308
5309
5310
5311
                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,
5312
                kernel_block_sizes=kernel_block_sizes,
5313
                is_spec_decode=bool(self.vllm_config.speculative_config),
5314
                logitsprocs=self.input_batch.logitsprocs,
5315
                logitsprocs_need_output_token_ids=self.input_batch.logitsprocs_need_output_token_ids,
5316
                is_pooling_model=self.is_pooling_model,
5317
                num_speculative_tokens=self.num_spec_tokens,
5318
5319
            )

5320
    def _allocate_kv_cache_tensors(
5321
5322
        self, kv_cache_config: KVCacheConfig
    ) -> dict[str, torch.Tensor]:
5323
        """
5324
5325
5326
        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.

5327
        Args:
5328
            kv_cache_config: The KV cache config
5329
        Returns:
5330
            dict[str, torch.Tensor]: A map between layer names to their
5331
            corresponding memory buffer for KV cache.
5332
        """
5333
5334
        kv_cache_raw_tensors: dict[str, torch.Tensor] = {}
        for kv_cache_tensor in kv_cache_config.kv_cache_tensors:
5335
5336
5337
            tensor = torch.zeros(
                kv_cache_tensor.size, dtype=torch.int8, device=self.device
            )
5338
5339
5340
5341
5342
            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:
5343
5344
5345
5346
            for layer_name in group.layer_names:
                if layer_name in self.runner_only_attn_layers:
                    continue
                layer_names.add(layer_name)
5347
5348
5349
        assert layer_names == set(kv_cache_raw_tensors.keys()), (
            "Some layers are not correctly initialized"
        )
5350
5351
        return kv_cache_raw_tensors

5352
5353
5354
    def _attn_group_iterator(self) -> Iterator[AttentionGroup]:
        return itertools.chain.from_iterable(self.attn_groups)

5355
    def _kv_cache_spec_attn_group_iterator(self) -> Iterator[AttentionGroup]:
5356
5357
        if not self.kv_cache_config.kv_cache_groups:
            return
5358
5359
        for attn_groups in self.attn_groups:
            yield from attn_groups
5360

5361
5362
5363
5364
5365
5366
5367
5368
5369
5370
5371
5372
5373
5374
5375
    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 = []
5376
        for kv_cache_gid, kv_cache_group in enumerate(kv_cache_config.kv_cache_groups):
5377
5378
5379
5380
5381
5382
            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):
5383
                continue
5384
            elif isinstance(kv_cache_spec, AttentionSpec):
5385
5386
5387
                # This is an attention backend that supports virtual
                # block splitting. Get the supported block sizes from
                # all backends in the group.
5388
                attn_groups = self.attn_groups[kv_cache_gid]
5389
                kv_manager_block_size = kv_cache_group.kv_cache_spec.block_size
5390
                selected_kernel_size = self.select_common_block_size(
5391
5392
5393
                    kv_manager_block_size, attn_groups
                )
                kernel_block_sizes.append(selected_kernel_size)
5394
            elif isinstance(kv_cache_spec, MambaSpec):
5395
5396
                # This is likely Mamba or other non-attention cache,
                # no splitting.
5397
                kernel_block_sizes.append(kv_cache_spec.block_size)
5398
5399
5400
5401
5402
5403
            else:
                raise NotImplementedError(
                    f"unknown kv cache spec {kv_cache_group.kv_cache_spec}"
                )
        return kernel_block_sizes

5404
5405
5406
5407
    def _reshape_kv_cache_tensors(
        self,
        kv_cache_config: KVCacheConfig,
        kv_cache_raw_tensors: dict[str, torch.Tensor],
5408
        kernel_block_sizes: list[int],
5409
    ) -> dict[str, torch.Tensor]:
5410
        """
5411
        Reshape the KV cache tensors to the desired shape and dtype.
5412

5413
        Args:
5414
5415
            kv_cache_config: The KV cache config
            kv_cache_raw_tensors: The KV cache buffer of each layer, with
5416
                correct size but uninitialized shape.
5417
            kernel_block_sizes: The kernel block sizes for each KV cache group.
5418
        Returns:
5419
            Dict[str, torch.Tensor]: A map between layer names to their
5420
5421
            corresponding memory buffer for KV cache.
        """
5422
        kv_caches: dict[str, torch.Tensor] = {}
5423
        has_attn, has_mamba = False, False
5424
5425
        for group in self._kv_cache_spec_attn_group_iterator():
            kv_cache_spec = group.kv_cache_spec
5426
            attn_backend = group.backend
5427
5428
5429
5430
            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]
5431
            for layer_name in group.layer_names:
5432
5433
                if layer_name in self.runner_only_attn_layers:
                    continue
5434
5435
                raw_tensor = kv_cache_raw_tensors[layer_name]
                assert raw_tensor.numel() % kv_cache_spec.page_size_bytes == 0
5436
                num_blocks = raw_tensor.numel() // kv_cache_spec.page_size_bytes
5437
                if isinstance(kv_cache_spec, AttentionSpec):
5438
                    has_attn = True
5439
5440
                    num_blocks_per_kv_block = (
                        kv_cache_spec.block_size // kernel_block_size
5441
5442
5443
                    )
                    kernel_num_blocks = num_blocks * num_blocks_per_kv_block

5444
                    kv_cache_shape = attn_backend.get_kv_cache_shape(
5445
                        kernel_num_blocks,
5446
                        kernel_block_size,
5447
5448
                        kv_cache_spec.num_kv_heads,
                        kv_cache_spec.head_size,
5449
5450
                        cache_dtype_str=self.cache_config.cache_dtype,
                    )
5451
                    dtype = kv_cache_spec.dtype
5452
                    try:
5453
                        kv_cache_stride_order = attn_backend.get_kv_cache_stride_order()
5454
                        assert len(kv_cache_stride_order) == len(kv_cache_shape)
5455
                    except (AttributeError, NotImplementedError):
5456
                        kv_cache_stride_order = tuple(range(len(kv_cache_shape)))
5457
5458
5459
5460
5461
                    # 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.
5462
5463
5464
                    kv_cache_shape = tuple(
                        kv_cache_shape[i] for i in kv_cache_stride_order
                    )
5465
5466
5467
5468
5469
                    # Maintain original KV shape view.
                    inv_order = [
                        kv_cache_stride_order.index(i)
                        for i in range(len(kv_cache_stride_order))
                    ]
5470
5471
5472
5473
5474
5475
                    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
5476
                elif isinstance(kv_cache_spec, MambaSpec):
5477
                    has_mamba = True
Chen Zhang's avatar
Chen Zhang committed
5478
5479
                    raw_tensor = kv_cache_raw_tensors[layer_name]
                    state_tensors = []
5480
                    storage_offset_bytes = 0
5481
                    for shape, dtype in zip(kv_cache_spec.shapes, kv_cache_spec.dtypes):
5482
5483
                        dtype_size = get_dtype_size(dtype)
                        num_element_per_page = (
5484
5485
                            kv_cache_spec.page_size_bytes // dtype_size
                        )
Chen Zhang's avatar
Chen Zhang committed
5486
                        target_shape = (num_blocks, *shape)
5487
5488
                        stride = torch.empty(target_shape).stride()
                        target_stride = (num_element_per_page, *stride[1:])
5489
                        assert storage_offset_bytes % dtype_size == 0
5490
5491
5492
5493
                        tensor = torch.as_strided(
                            raw_tensor.view(dtype),
                            size=target_shape,
                            stride=target_stride,
5494
                            storage_offset=storage_offset_bytes // dtype_size,
5495
                        )
Chen Zhang's avatar
Chen Zhang committed
5496
                        state_tensors.append(tensor)
5497
                        storage_offset_bytes += stride[0] * dtype_size
5498
5499

                    kv_caches[layer_name] = state_tensors
5500
                else:
5501
                    raise NotImplementedError
5502
5503

        if has_attn and has_mamba:
5504
            self._update_hybrid_attention_mamba_layout(kv_caches)
5505

5506
5507
        return kv_caches

5508
    def _update_hybrid_attention_mamba_layout(
5509
5510
        self, kv_caches: dict[str, torch.Tensor]
    ) -> None:
5511
        """
5512
5513
        Update the layout of attention layers from (2, num_blocks, ...) to
        (num_blocks, 2, ...).
5514
5515

        Args:
5516
            kv_caches: The KV cache buffer of each layer.
5517
5518
        """

5519
5520
        for group in self._kv_cache_spec_attn_group_iterator():
            kv_cache_spec = group.kv_cache_spec
5521
            for layer_name in group.layer_names:
5522
                kv_cache = kv_caches[layer_name]
5523
5524
5525
5526
                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 "
5527
                        f"a tensor of shape {kv_cache.shape}"
5528
                    )
5529
                    hidden_size = kv_cache.shape[2:].numel()
5530
5531
5532
5533
                    kv_cache.as_strided_(
                        size=kv_cache.shape,
                        stride=(hidden_size, 2 * hidden_size, *kv_cache.stride()[2:]),
                    )
5534

5535
    def initialize_kv_cache_tensors(
5536
        self, kv_cache_config: KVCacheConfig, kernel_block_sizes: list[int]
5537
    ) -> dict[str, torch.Tensor]:
5538
5539
5540
5541
5542
        """
        Initialize the memory buffer for KV cache.

        Args:
            kv_cache_config: The KV cache config
5543
5544
            kernel_block_sizes: The kernel block sizes for each KV cache group.

5545
        Returns:
5546
            Dict[str, torch.Tensor]: A map between layer names to their
5547
5548
            corresponding memory buffer for KV cache.
        """
5549
5550
5551
5552
5553
5554
5555
5556
5557
5558
5559
5560
5561
5562
5563
5564
5565
5566
5567
5568
5569
5570
5571
5572

        # 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
            )
5573

5574
        # Set up cross-layer KV cache sharing
5575
5576
        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)
5577
5578
            kv_caches[layer_name] = kv_caches[target_layer_name]

5579
5580
5581
5582
5583
5584
5585
5586
5587
        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,
        )
5588
5589
5590
        return kv_caches

    def maybe_add_kv_sharing_layers_to_kv_cache_groups(
5591
5592
        self, kv_cache_config: KVCacheConfig
    ) -> None:
5593
5594
5595
5596
5597
5598
5599
5600
5601
5602
5603
5604
5605
5606
5607
5608
5609
5610
        """
        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.
5611
            attn_layers = get_layers_from_vllm_config(self.vllm_config, Attention)
5612
5613
            for layer_name in reversed(attn_layers):
                if layer_name in self.shared_kv_cache_layers:
5614
                    self.kv_sharing_fast_prefill_eligible_layers.add(layer_name)
5615
5616
                else:
                    break
5617

5618
5619
5620
5621
5622
5623
5624
    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
        """
5625
        kv_cache_config = deepcopy(kv_cache_config)
5626
        self.kv_cache_config = kv_cache_config
5627
        self.may_add_encoder_only_layers_to_kv_cache_config()
5628
        self.maybe_add_kv_sharing_layers_to_kv_cache_groups(kv_cache_config)
5629
        self.initialize_attn_backend(kv_cache_config)
5630
5631
5632
5633
5634
5635
        # 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)
5636
5637
5638
5639

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

5640
        # Reinitialize need to after initialize_attn_backend
5641
5642
5643
5644
        self.may_reinitialize_input_batch(kv_cache_config, kernel_block_sizes)
        kv_caches = self.initialize_kv_cache_tensors(
            kv_cache_config, kernel_block_sizes
        )
5645

5646
5647
5648
5649
5650
5651
        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
5652
        if has_kv_transfer_group():
5653
            kv_transfer_group = get_kv_transfer_group()
5654
5655
5656
5657
5658
5659
5660
            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)
5661
            kv_transfer_group.set_host_xfer_buffer_ops(copy_kv_blocks)
Robert Shaw's avatar
Robert Shaw committed
5662

5663
5664
5665
5666
5667
    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
5668
        encoder_only_attn_specs: dict[AttentionSpec, list[str]] = defaultdict(list)
5669
5670
5671
        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:
5672
                attn_spec: AttentionSpec = EncoderOnlyAttentionSpec(
5673
5674
5675
                    block_size=block_size,
                    num_kv_heads=attn_module.num_kv_heads,
                    head_size=attn_module.head_size,
5676
5677
                    dtype=self.kv_cache_dtype,
                )
5678
5679
5680
                encoder_only_attn_specs[attn_spec].append(layer_name)
                self.runner_only_attn_layers.add(layer_name)
        if len(encoder_only_attn_specs) > 0:
5681
5682
5683
            assert len(encoder_only_attn_specs) == 1, (
                "Only support one encoder-only attention spec now"
            )
5684
5685
            spec, layer_names = encoder_only_attn_specs.popitem()
            self.kv_cache_config.kv_cache_groups.append(
5686
5687
                KVCacheGroupSpec(layer_names=layer_names, kv_cache_spec=spec)
            )
5688

5689
    def get_kv_cache_spec(self) -> dict[str, KVCacheSpec]:
5690
        """
5691
        Generates the KVCacheSpec by parsing the kv cache format from each
5692
5693
        Attention module in the static forward context.
        Returns:
5694
            KVCacheSpec: A dictionary mapping layer names to their KV cache
5695
5696
            format. Layers that do not need KV cache are not included.
        """
5697
5698
        if has_ec_transfer() and get_ec_transfer().is_producer:
            return {}
5699
        kv_cache_spec: dict[str, KVCacheSpec] = {}
5700
5701
        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
5702
        for layer_name, attn_module in attn_layers.items():
5703
5704
5705
5706
5707
5708
5709
5710
5711
5712
5713
5714
5715
5716
5717
            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
5718

5719
        return kv_cache_spec
5720

5721
    def _to_list(self, sampled_token_ids: torch.Tensor) -> list[list[int]]:
5722
5723
5724
5725
5726
5727
5728
5729
        # 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.
5730
        pinned = self.sampled_token_ids_pinned_cpu[: sampled_token_ids.shape[0]]
5731
5732
5733
        pinned.copy_(sampled_token_ids, non_blocking=True)
        self.transfer_event.record()
        self.transfer_event.synchronize()
5734
        return pinned.tolist()