"vllm/model_executor/models/afmoe.py" did not exist on "29283eaa7eff50c88b1efe5443921a0768d5d15e"
gpu_model_runner.py 68.6 KB
Newer Older
1
2
# SPDX-License-Identifier: Apache-2.0

3
import gc
4
import time
5
import weakref
6
from typing import TYPE_CHECKING, Optional, Union
7
8
9
10
11
12

import numpy as np
import torch
import torch.distributed
import torch.nn as nn

13
from vllm.attention import AttentionType, get_attn_backend
14
from vllm.attention.layer import Attention
15
from vllm.config import CompilationLevel, VllmConfig
16
from vllm.distributed.parallel_state import get_pp_group, graph_capture
17
from vllm.forward_context import set_forward_context
18
from vllm.inputs import INPUT_REGISTRY
19
from vllm.logger import init_logger
20
from vllm.model_executor.layers.fused_moe import FusedMoE
21
from vllm.model_executor.layers.rotary_embedding import MRotaryEmbedding
22
from vllm.model_executor.model_loader import get_model
23
from vllm.multimodal import MULTIMODAL_REGISTRY, MultiModalKwargs
24
from vllm.multimodal.utils import group_mm_inputs_by_modality
25
from vllm.sampling_params import SamplingType
26
from vllm.sequence import IntermediateTensors
27
from vllm.utils import (STR_DTYPE_TO_TORCH_DTYPE, DeviceMemoryProfiler,
28
29
                        LayerBlockType, LazyLoader, cdiv,
                        is_pin_memory_available)
30
from vllm.v1.attention.backends.flash_attn import FlashAttentionMetadata
31
from vllm.v1.core.encoder_cache_manager import compute_encoder_budget
32
33
from vllm.v1.kv_cache_interface import (FullAttentionSpec, KVCacheConfig,
                                        KVCacheSpec)
34
35
from vllm.v1.outputs import (EMPTY_MODEL_RUNNER_OUTPUT, LogprobsTensors,
                             ModelRunnerOutput)
36
from vllm.v1.sample.metadata import SamplingMetadata
37
from vllm.v1.sample.rejection_sampler import INVALID_TOKEN_ID, RejectionSampler
38
from vllm.v1.spec_decode.ngram_proposer import NgramProposer
39
from vllm.v1.spec_decode.utils import is_spec_decode_supported
40
from vllm.v1.utils import bind_kv_cache
41
from vllm.v1.worker.gpu_input_batch import CachedRequestState, InputBatch
42
from vllm.v1.worker.lora_model_runner_mixin import LoRAModelRunnerMixin
43
44

if TYPE_CHECKING:
45
46
    import xgrammar as xgr

47
    from vllm.v1.core.scheduler_output import SchedulerOutput
48
49
else:
    xgr = LazyLoader("xgr", globals(), "xgrammar")
50
51
52
53

logger = init_logger(__name__)


54
class GPUModelRunner(LoRAModelRunnerMixin):
55
56
57

    def __init__(
        self,
58
        vllm_config: VllmConfig,
59
        device: torch.device,
60
    ):
61
62
63
64
65
66
67
68
69
70
        self.vllm_config = vllm_config
        self.model_config = vllm_config.model_config
        self.cache_config = vllm_config.cache_config
        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.prompt_adapter_config = vllm_config.prompt_adapter_config
        self.observability_config = vllm_config.observability_config
71

72
73
74
75
        model_config = self.model_config
        cache_config = self.cache_config
        scheduler_config = self.scheduler_config
        parallel_config = self.parallel_config
76
        self.device = device
77
78
79
80
81
82
83
84
        self.pin_memory = is_pin_memory_available()
        self.dtype = self.model_config.dtype
        if cache_config.cache_dtype == "auto":
            self.kv_cache_dtype = self.dtype
        else:
            self.kv_cache_dtype = STR_DTYPE_TO_TORCH_DTYPE[
                cache_config.cache_dtype]

85
86
        # NOTE(woosuk): sliding_window is None for models with interleaved
        # attention. Use interleaved_sliding_window instead.
87
        self.sliding_window = model_config.get_sliding_window()
88
89
90
91
92
93
        self.interleaved_sliding_window = getattr(
            model_config.hf_text_config, "interleaved_sliding_window", None)
        self.window_size = (self.sliding_window
                            or self.interleaved_sliding_window)

        self.is_multimodal_model = model_config.is_multimodal_model
94
95
96
97
        self.block_size = cache_config.block_size
        self.max_model_len = model_config.max_model_len
        self.max_num_blocks_per_req = cdiv(self.max_model_len, self.block_size)
        self.max_num_tokens = scheduler_config.max_num_batched_tokens
98
        self.max_num_reqs = scheduler_config.max_num_seqs
99
100

        # Model-related.
101
102
        self.num_attn_layers = model_config.get_num_layers_by_block_type(
            parallel_config, LayerBlockType.attention)
103
104
        self.num_query_heads = model_config.get_num_attention_heads(
            parallel_config)
105
106
        self.num_kv_heads = model_config.get_num_kv_heads(parallel_config)
        self.head_size = model_config.get_head_size()
107
108
        self.hidden_size = model_config.get_hidden_size()

109
110
111
112
113
114
115
116
117
118
119
120
121
122
123
124
125
126
127
128
129
        self.attn_backend = get_attn_backend(
            self.head_size,
            self.dtype,
            self.kv_cache_dtype,
            self.block_size,
            self.model_config.is_attention_free,
            use_mla=self.model_config.use_mla,
        )
        if self.attn_backend is None:
            error_msg = (
                f"Error with get_att_backend: {self.head_size=}, "
                f"{self.dtype=}, {self.kv_cache_dtype=}, {self.block_size=}, "
                f"{self.model_config.is_attention_free=}, "
                f"{self.model_config.use_mla=}")
            logger.error(error_msg)
            raise NotImplementedError(
                "Non-Attention backend is not supported by V1 GPUModelRunner.")

        self.attn_metadata_builder = self.attn_backend.get_builder_cls()(
            weakref.proxy(self))

130
        # Multi-modal data support
131
132
        self.input_registry = INPUT_REGISTRY
        self.mm_registry = MULTIMODAL_REGISTRY
133
        self.uses_mrope = model_config.uses_mrope
134

135
136
137
138
139
140
        encoder_compute_budget, encoder_cache_size = compute_encoder_budget(
            model_config=model_config,
            scheduler_config=scheduler_config,
        )
        self.max_num_encoder_input_tokens = encoder_compute_budget
        self.encoder_cache_size = encoder_cache_size
141
142
143

        # Lazy initialization
        # self.model: nn.Module  # Set after load_model
144
        self.kv_caches: list[torch.Tensor] = []
145
        # req_id -> (input_id -> encoder_output)
146
        self.encoder_cache: dict[str, dict[int, torch.Tensor]] = {}
147

148
149
150
        # Set up speculative decoding.
        self.use_spec_decode = False
        if self.speculative_config:
151
            self.use_spec_decode = True
152
            self.rejection_sampler = RejectionSampler()
153
154
155
            # TODO: find a better way to check if we are using ngram.
            assert self.speculative_config.ngram_prompt_lookup_min, \
                    "Currently, only ngram spec decode is supported in V1."
156
157
158
159
160
161
162
163
164
            if get_pp_group().is_last_rank:
                self.drafter = NgramProposer()
                # Trigger Numba JIT compilation for N-gram proposer.
                # This usually takes less than 1 second.
                self.drafter.propose(
                    np.zeros(1024, dtype=np.int32),
                    self.speculative_config.ngram_prompt_lookup_min,
                    self.speculative_config.num_speculative_tokens,
                )
165

166
        # Request states.
167
        self.requests: dict[str, CachedRequestState] = {}
168
169
        # Persistent batch.
        self.input_batch = InputBatch(
170
            max_num_reqs=self.max_num_reqs,
171
172
173
174
            max_model_len=self.max_model_len,
            max_num_blocks_per_req=self.max_num_blocks_per_req,
            device=self.device,
            pin_memory=self.pin_memory,
175
            vocab_size=model_config.get_vocab_size(),
176
177
        )

178
        self.use_cuda_graph = (self.vllm_config.compilation_config.level
179
180
181
                               == CompilationLevel.PIECEWISE
                               and not self.model_config.enforce_eager)
        # TODO(woosuk): Provide an option to tune the max cudagraph batch size.
182
183
184
185
        # The convention is different.
        # self.cudagraph_batch_sizes sorts in ascending order.
        # The batch sizes in the config are in descending order.
        self.cudagraph_batch_sizes = list(
186
187
            reversed(
                self.vllm_config.compilation_config.cudagraph_capture_sizes))
188

189
190
191
192
        # Cache the device properties.
        self.device_properties = torch.cuda.get_device_properties(self.device)
        self.num_sms = self.device_properties.multi_processor_count

193
194
195
196
        # Persistent buffers for CUDA graphs.
        self.input_ids = torch.zeros(self.max_num_tokens,
                                     dtype=torch.int32,
                                     device=self.device)
197
198
199
        self.positions = torch.zeros(self.max_num_tokens,
                                     dtype=torch.int64,
                                     device=self.device)
200
201
        # None in the first PP rank. The rest are set after load_model.
        self.intermediate_tensors: Optional[IntermediateTensors] = None
202
203

        # Only relevant for models using M-RoPE (e.g, Qwen2-VL)
204
        if self.uses_mrope:
Roger Wang's avatar
Roger Wang committed
205
206
207
208
            # 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
209
210
211
212
213
214

            # 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
Roger Wang's avatar
Roger Wang committed
215
            self.mrope_positions = torch.zeros((3, self.max_num_tokens + 1),
216
217
                                               dtype=torch.int64,
                                               device=self.device)
Roger Wang's avatar
Roger Wang committed
218
219
220
221
222
            self.mrope_positions_cpu = torch.zeros(
                (3, self.max_num_tokens + 1),
                dtype=torch.int64,
                device="cpu",
                pin_memory=self.pin_memory)
223

224
225
226
227
        self.inputs_embeds = torch.zeros(
            (self.max_num_tokens, self.hidden_size),
            dtype=self.dtype,
            device=self.device)
228

229
        # OPTIMIZATION: Cache the tensors rather than creating them every step.
230
        self.arange_np = np.arange(max(self.max_num_reqs + 1,
231
232
                                       self.max_model_len,
                                       self.max_num_tokens),
233
234
235
236
                                   dtype=np.int32)
        # NOTE(woosuk): These tensors are "stateless", i.e., they are literally
        # a faster version of creating a new tensor every time. Thus, we should
        # not make any assumptions about the values in these tensors.
237
238
239
240
241
242
243
244
245
246
247
248
249
250
251
252
253
254
255
256
        self.input_ids_cpu = torch.zeros(self.max_num_tokens,
                                         dtype=torch.int32,
                                         device="cpu",
                                         pin_memory=self.pin_memory)
        self.input_ids_np = self.input_ids_cpu.numpy()
        self.positions_cpu = torch.zeros(self.max_num_tokens,
                                         dtype=torch.int64,
                                         device="cpu",
                                         pin_memory=self.pin_memory)
        self.positions_np = self.positions_cpu.numpy()
        self.slot_mapping_cpu = torch.zeros(self.max_num_tokens,
                                            dtype=torch.int32,
                                            device="cpu",
                                            pin_memory=self.pin_memory)
        self.slot_mapping_np = self.slot_mapping_cpu.numpy()
        self.query_start_loc_cpu = torch.zeros(self.max_num_reqs + 1,
                                               dtype=torch.int32,
                                               device="cpu",
                                               pin_memory=self.pin_memory)
        self.query_start_loc_np = self.query_start_loc_cpu.numpy()
257
258
259
260
261
        self.seq_lens_cpu = torch.zeros(self.max_num_reqs,
                                        dtype=torch.int32,
                                        device="cpu",
                                        pin_memory=self.pin_memory)
        self.seq_lens_np = self.seq_lens_cpu.numpy()
262

263
    def _update_states(self, scheduler_output: "SchedulerOutput") -> None:
264
265
266
267
268
269
        """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.

270
271
        The SamplingMetadata is updated and copied to the GPU if there is a
        new/resumed/paused/finished request in the batch.
272
273
        """
        # Remove finished requests from the cached states.
274
275
        for req_id in scheduler_output.finished_req_ids:
            self.requests.pop(req_id, None)
276
            self.encoder_cache.pop(req_id, None)
277
278
279
280
281
282
        # 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.
283
        removed_req_indices: list[int] = []
284
285
286
287
        for req_id in scheduler_output.finished_req_ids:
            req_index = self.input_batch.remove_request(req_id)
            if req_index is not None:
                removed_req_indices.append(req_index)
288
289
290
291
292
293
294
295

        # Free the cached encoder outputs.
        for req_id, input_id in scheduler_output.free_encoder_input_ids:
            encoder_outputs = self.encoder_cache.get(req_id)
            if encoder_outputs is not None:
                encoder_outputs.pop(input_id, None)
                if not encoder_outputs:
                    self.encoder_cache.pop(req_id, None)
296

297
298
299
300
301
302
303
304
305
306
307
308
309
        # 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()
        unscheduled_req_ids = cached_req_ids - scheduled_req_ids
        # 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:
310
            req_index = self.input_batch.remove_request(req_id)
311
312
            assert req_index is not None
            removed_req_indices.append(req_index)
313

314
        req_ids_to_add: list[str] = []
315
        # Add new requests to the cached states.
316
317
318
        for new_req_data in scheduler_output.scheduled_new_reqs:
            req_id = new_req_data.req_id
            sampling_params = new_req_data.sampling_params
319
            if sampling_params.sampling_type == SamplingType.RANDOM_SEED:
320
321
322
323
324
                generator = torch.Generator(device=self.device)
                generator.manual_seed(sampling_params.seed)
            else:
                generator = None

325
326
            self.requests[req_id] = CachedRequestState(
                req_id=req_id,
327
328
329
330
                prompt_token_ids=new_req_data.prompt_token_ids,
                prompt=new_req_data.prompt,
                mm_inputs=new_req_data.mm_inputs,
                mm_positions=new_req_data.mm_positions,
331
332
                sampling_params=sampling_params,
                generator=generator,
333
334
                block_ids=new_req_data.block_ids,
                num_computed_tokens=new_req_data.num_computed_tokens,
335
                output_token_ids=[],
336
                lora_request=new_req_data.lora_request,
337
            )
338
339

            # Only relevant for models using M-RoPE (e.g, Qwen2-VL)
340
            if self.uses_mrope:
341
342
                image_grid_thw = []
                video_grid_thw = []
Roger Wang's avatar
Roger Wang committed
343
                second_per_grid_ts = []
344
345
346
347
348
349
350
                for mm_input in self.requests[req_id].mm_inputs:
                    if mm_input.get("image_grid_thw") is not None:
                        image_grid_thw.extend(
                            mm_input["image_grid_thw"].tolist())
                    if mm_input.get("video_grid_thw") is not None:
                        video_grid_thw.extend(
                            mm_input["video_grid_thw"].tolist())
Roger Wang's avatar
Roger Wang committed
351
352
353
                    if mm_input.get("second_per_grid_ts") is not None:
                        second_per_grid_ts.extend(
                            mm_input["second_per_grid_ts"])
354
355
356
357
358
359
360

                hf_config = self.model_config.hf_config

                self.requests[req_id].mrope_positions, \
                    self.requests[req_id].mrope_position_delta = \
                    MRotaryEmbedding.get_input_positions_tensor(
                        self.requests[req_id].prompt_token_ids,
Roger Wang's avatar
Roger Wang committed
361
                        hf_config=hf_config,
362
363
                        image_grid_thw=image_grid_thw,
                        video_grid_thw=video_grid_thw,
Roger Wang's avatar
Roger Wang committed
364
                        second_per_grid_ts=second_per_grid_ts,
365
366
                    )

367
368
            req_ids_to_add.append(req_id)

369
370
371
        # Update the states of the running/resumed requests.
        for req_data in scheduler_output.scheduled_cached_reqs:
            req_id = req_data.req_id
372
373
            req_state = self.requests[req_id]

374
            # Update the cached states.
375
376
377
378
379
380
381
            num_computed_tokens = req_data.num_computed_tokens
            req_state.num_computed_tokens = num_computed_tokens
            # Add the sampled token(s) from the previous step (if any).
            # This doesn't include "unverified" tokens like spec decode tokens.
            num_new_tokens = (num_computed_tokens +
                              len(req_data.new_token_ids) -
                              req_state.num_tokens)
382
383
384
385
386
387
            if num_new_tokens == 1:
                # Avoid slicing list in most common case.
                req_state.output_token_ids.append(req_data.new_token_ids[-1])
            elif num_new_tokens > 0:
                req_state.output_token_ids.extend(
                    req_data.new_token_ids[-num_new_tokens:])
388
            # Update the block IDs.
389
390
391
392
393
394
395
396
397
398
399
400
401
402
403
404
405
406
            if not req_data.resumed_from_preemption:
                # Append the new blocks to the existing block IDs.
                req_state.block_ids.extend(req_data.new_block_ids)
            else:
                # The request is resumed from preemption.
                # Replace the existing block IDs with the new ones.
                req_state.block_ids = req_data.new_block_ids

            req_index = self.input_batch.req_id_to_index.get(req_id)
            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.
                req_ids_to_add.append(req_id)
                continue

            # Update the persistent batch.
            self.input_batch.num_computed_tokens_cpu[req_index] = (
407
                num_computed_tokens)
408
409
            self.input_batch.block_table.append_row(req_data.new_block_ids,
                                                    req_index)
410
411
412
413
414
415
            # Add new_token_ids to token_ids_cpu.
            start_token_index = num_computed_tokens
            end_token_index = num_computed_tokens + len(req_data.new_token_ids)
            self.input_batch.token_ids_cpu[
                req_index,
                start_token_index:end_token_index] = req_data.new_token_ids
416
            self.input_batch.num_tokens_no_spec[req_index] = end_token_index
417
418
            # Add spec_token_ids to token_ids_cpu.
            spec_token_ids = scheduler_output.scheduled_spec_decode_tokens.get(
419
                req_id, ())
420
421
422
423
424
425
426
            if spec_token_ids:
                start_index = end_token_index
                end_token_index += len(spec_token_ids)
                self.input_batch.token_ids_cpu[
                    req_index, start_index:end_token_index] = spec_token_ids
            # NOTE(woosuk): `num_tokens` here may include spec decode tokens.
            self.input_batch.num_tokens[req_index] = end_token_index
427

428
429
        # Check if the batch has changed. If not, we can skip copying the
        # sampling metadata from CPU to GPU.
430
431
        batch_changed = len(removed_req_indices) > 0 or len(req_ids_to_add) > 0

432
433
434
435
436
437
438
439
440
441
442
443
444
445
446
447
        # Add the new or resumed requests to the persistent batch.
        # The smaller empty indices are filled first.
        removed_req_indices = sorted(removed_req_indices, reverse=True)
        for req_id in req_ids_to_add:
            req_state = self.requests[req_id]
            if removed_req_indices:
                # Fill the empty index.
                req_index = removed_req_indices.pop()
            else:
                # Append to the end.
                req_index = None
            self.input_batch.add_request(req_state, req_index)

        # Condense the batched states if there are empty indices.
        if removed_req_indices:
            self.input_batch.condense(removed_req_indices)
448

449
450
        if batch_changed:
            self.input_batch.refresh_sampling_metadata()
451

452
    def _prepare_inputs(
453
454
        self,
        scheduler_output: "SchedulerOutput",
455
    ) -> tuple[FlashAttentionMetadata, torch.Tensor]:
456
457
458
459
460
        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

461
462
463
        # Some attention backends (namely MLA) may want to separate requests
        # based on if the attention computation will be compute-bound or
        # memory-bound. This gives them a hook to do that.
464
465
466
467
        modified_batch = self.attn_metadata_builder.reorder_batch(
            self.input_batch, scheduler_output)
        if modified_batch:
            self.input_batch.refresh_sampling_metadata()
468

469
470
        # OPTIMIZATION: Start copying the block table first.
        # This way, we can overlap the copy with the following CPU operations.
471
        self.input_batch.block_table.commit(num_reqs)
472
473
474

        # Get the number of scheduled tokens for each request.
        # TODO: The Python loop can be slow. Optimize.
475
        num_scheduled_tokens = np.empty(num_reqs, dtype=np.int32)
476
        max_num_scheduled_tokens = 0
477
        for i, req_id in enumerate(self.input_batch.req_ids):
478
            num_tokens = scheduler_output.num_scheduled_tokens[req_id]
479
            num_scheduled_tokens[i] = num_tokens
480
481
482
483
484
            max_num_scheduled_tokens = max(max_num_scheduled_tokens,
                                           num_tokens)

        # Get request indices.
        # E.g., [2, 5, 3] -> [0, 0, 1, 1, 1, 1, 1, 2, 2, 2]
485
486
        req_indices = np.repeat(self.arange_np[:num_reqs],
                                num_scheduled_tokens)
487
488
489

        # Get batched arange.
        # E.g., [2, 5, 3] -> [0, 1, 0, 1, 2, 3, 4, 0, 1, 2]
490
491
492
493
494
495
496
497
498
        # Equivalent to but faster than:
        # np.concatenate([np.arange(n) for n in num_scheduled_tokens])
        # Step 1. [2, 5, 3] -> [2, 7, 10]
        cu_num_tokens = np.cumsum(num_scheduled_tokens)
        # Step 2. [2, 7, 10] -> [0, 0, 2, 2, 2, 2, 2, 7, 7, 7]
        cumsums_offsets = np.repeat(cu_num_tokens - num_scheduled_tokens,
                                    num_scheduled_tokens)
        # Step 3. [0, 1, 0, 1, 2, 3, 4, 0, 1, 2]
        arange = self.arange_np[:total_num_scheduled_tokens] - cumsums_offsets
499
500

        # Get positions.
501
        positions_np = self.positions_np[:total_num_scheduled_tokens]
502
503
504
505
        np.add(self.input_batch.num_computed_tokens_cpu[req_indices],
               arange,
               out=positions_np)

506
507
        # Calculate M-RoPE positions.
        # Only relevant for models using M-RoPE (e.g, Qwen2-VL)
508
        if self.uses_mrope:
509
510
            self._calc_mrope_positions(scheduler_output)

511
512
513
514
        # 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.
515
516
        token_indices = (positions_np +
                         req_indices * self.input_batch.token_ids_cpu.shape[1])
517

518
519
520
521
        # 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.
        torch.index_select(self.input_batch.token_ids_cpu_tensor.flatten(),
522
                           0,
523
524
                           torch.from_numpy(token_indices),
                           out=self.input_ids_cpu[:total_num_scheduled_tokens])
525
526

        # Calculate the slot mapping.
527
528
529
530
531
        # E.g., [0, 1, 0, 1, 2, 3, 4, 0, 1, 2]
        # -> [0, 0, K, K, K + 1, K + 1, K + 2, 2 * K, 2 * K, 2 * K + 1]
        # where K is the max_num_blocks_per_req and the block size is 2.
        # NOTE(woosuk): We can't simply use `token_indices // block_size` here
        # because M (max_model_len) is not necessarily divisible by block_size.
532
533
534
535
536
        block_table_indices = (req_indices * self.max_num_blocks_per_req +
                               positions_np // self.block_size)
        # 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.
537
538
        block_table_cpu = self.input_batch.block_table.get_cpu_tensor()
        block_numbers = block_table_cpu.flatten()[block_table_indices].numpy()
539
540
541
542
        block_offsets = positions_np % self.block_size
        np.add(block_numbers * self.block_size,
               block_offsets,
               out=self.slot_mapping_np[:total_num_scheduled_tokens])
543
544

        # Prepare the attention metadata.
545
        self.query_start_loc_np[0] = 0
546
        self.query_start_loc_np[1:num_reqs + 1] = cu_num_tokens
547

548
549
550
        self.seq_lens_np[:num_reqs] = (
            self.input_batch.num_computed_tokens_cpu[:num_reqs] +
            num_scheduled_tokens)
551
552
553
554

        # Copy the tensors to the GPU.
        self.input_ids[:total_num_scheduled_tokens].copy_(
            self.input_ids_cpu[:total_num_scheduled_tokens], non_blocking=True)
555
        if self.uses_mrope:
556
557
558
559
560
561
562
563
564
            # Only relevant for models using M-RoPE (e.g, Qwen2-VL)
            self.mrope_positions[:, :total_num_scheduled_tokens].copy_(
                self.mrope_positions_cpu[:, :total_num_scheduled_tokens],
                non_blocking=True)
        else:
            # Common case (1D positions)
            self.positions[:total_num_scheduled_tokens].copy_(
                self.positions_cpu[:total_num_scheduled_tokens],
                non_blocking=True)
565
566

        # Prepare for cascade attention if needed.
567
568
569
570
        common_prefix_len = self._compute_cascade_attn_prefix_len(
            num_scheduled_tokens,
            scheduler_output.num_common_prefix_blocks,
        )
571
572
        attn_metadata = self.attn_metadata_builder.build(
            num_reqs=num_reqs,
573
            num_actual_tokens=total_num_scheduled_tokens,
574
            max_query_len=max_num_scheduled_tokens,
575
            common_prefix_len=common_prefix_len,
576
        )
577

578
579
        use_spec_decode = len(
            scheduler_output.scheduled_spec_decode_tokens) > 0
580
        if use_spec_decode:
581
582
            logits_indices = self._calc_spec_decode_metadata(
                scheduler_output, cu_num_tokens)
583
584
585
586
587
588
        else:
            # 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.
589
            logits_indices = attn_metadata.query_start_loc[1:] - 1
590

591
592
593
594
        # Hot-Swap lora model
        if self.lora_config:
            self.set_active_loras(self.input_batch, num_scheduled_tokens)

595
        return attn_metadata, logits_indices
596

597
598
599
600
601
602
603
604
605
606
607
608
609
610
611
612
613
614
615
616
617
618
619
620
621
622
623
624
625
626
627
628
629
630
631
632
633
634
635
636
637
638
639
640
641
642
643
644
645
646
647
648
649
650
651
652
653
654
655
656
657
658
659
660
661
662
663
664
665
666
667
668
669
    def _compute_cascade_attn_prefix_len(
        self,
        num_scheduled_tokens: np.ndarray,
        num_common_prefix_blocks: int,
    ) -> 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.
        """
        common_prefix_len = num_common_prefix_blocks * self.block_size
        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]
        # Request 3's num_computed_tokens: 4 (i.e., [A, B, C, D])
        # 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.
        num_reqs = len(num_scheduled_tokens)
        common_prefix_len = min(
            common_prefix_len,
            self.input_batch.num_computed_tokens_cpu[:num_reqs].min())
        # common_prefix_len should be a multiple of the block size.
        common_prefix_len = (common_prefix_len // self.block_size *
                             self.block_size)
670
        use_cascade = self.attn_backend.use_cascade_attention(
671
672
673
674
675
            common_prefix_len=common_prefix_len,
            query_lens=num_scheduled_tokens,
            num_query_heads=self.num_query_heads,
            num_kv_heads=self.num_kv_heads,
            use_alibi=False,  # FIXME
676
            use_sliding_window=self.window_size is not None,
677
678
679
680
            num_sms=self.num_sms,
        )
        return common_prefix_len if use_cascade else 0

681
682
    def _calc_mrope_positions(self, scheduler_output: "SchedulerOutput"):
        mrope_pos_ptr = 0
683
        for index, req_id in enumerate(self.input_batch.req_ids):
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
725
726
727
728
729
730
731
732
            req = self.requests[req_id]
            assert req.mrope_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 = len(req.prompt_token_ids)

            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 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

                self.mrope_positions_cpu[:, dst_start:dst_end] = \
                    req.mrope_positions[:,src_start:src_end]

                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

                self.mrope_positions_cpu[:, dst_start:dst_end] = \
                    MRotaryEmbedding.get_next_input_positions_tensor(
                        req.mrope_position_delta,
                        context_len=num_computed_tokens +
                        prompt_part_len,
                        seq_len=num_computed_tokens +
                        prompt_part_len +
                        completion_part_len,
                    )

                mrope_pos_ptr += completion_part_len

733
734
735
736
    def _calc_spec_decode_metadata(
        self,
        scheduler_output: "SchedulerOutput",
        cu_num_tokens: np.ndarray,
737
    ) -> torch.Tensor:
738
739
740
        # Get the number of spec decode tokens for each request.
        num_reqs = self.input_batch.num_reqs
        num_spec_decode_tokens = np.empty(num_reqs, dtype=np.int32)
741
        for i, req_id in enumerate(self.input_batch.req_ids):
742
743
744
745
746
747
748
749
750
751
752
753
754
755
756
757
758
759
760
761
762
763
764
765
766
767
768
769
770
771
772
773
774
775
776
777
778
            num_spec_decode_tokens[i] = len(
                scheduler_output.scheduled_spec_decode_tokens.get(req_id, ()))

        # Get spec decode logits indices.
        # E.g.,   num_scheduled_tokens: [4, 100, 3,   100, 2]
        #         cu_num_tokens:        [4, 104, 107, 207, 209]
        #         num_spec_tokens_list: [3, 0,   2,   0,   1]
        #         num_sampled_tokens:   [4, 1,   3,   1,   2]
        #         spec_decode_logits_indices:
        #                 [0, 1, 2, 3, 103, 104, 105, 106, 206, 207, 208]
        num_sampled_tokens = num_spec_decode_tokens + 1
        # logits_start_loc: [0, 103, 104, 206, 207]
        logits_start_loc = cu_num_tokens - num_sampled_tokens
        # [0, 103, 104, 206, 207] ->
        #               [0, 0, 0, 0, 103, 104, 104, 104, 206, 207, 207]
        logits_start_loc = np.repeat(logits_start_loc, num_sampled_tokens)
        # The following three lines:
        # [4, 1,   3,   1,   2] -> [0, 1, 2, 3, 0, 0, 1, 2, 0, 0, 1]
        # Step 1. [4, 1, 3, 1, 2] -> [4, 5, 8, 9, 11]
        cu_num_sampled_tokens = np.cumsum(num_sampled_tokens)
        # Step 2. [4, 5, 8, 9, 11] -> [0, 4, 5, 8, 9]
        #         -> [0, 0, 0, 0, 4, 5, 5, 5, 8, 9, 9]
        cumsums_sampled_offsets = np.repeat(
            cu_num_sampled_tokens - num_sampled_tokens, num_sampled_tokens)
        # Step 3.  [0, 1, 2, 3, 4, 5, 6, 7, 8, 9, 10]
        #       -  [0, 0, 0, 0, 4, 5, 5, 5, 8, 9, 9]
        #      -> [0, 1, 2, 3, 0, 0, 1, 2, 0, 0, 1]
        total_num_sampled_tokens = num_sampled_tokens.sum()
        sampled_arange = (self.arange_np[:total_num_sampled_tokens] -
                          cumsums_sampled_offsets)

        # [0, 0, 0, 0, 103, 104, 104, 104, 206, 207, 207] ->
        # [0, 1, 2, 3, 103, 104, 105, 106, 206, 207, 208]
        spec_decode_logits_indices = logits_start_loc + sampled_arange
        return torch.from_numpy(spec_decode_logits_indices).to(
            self.device, non_blocking=True)

779
780
781
782
783
784
    def _execute_encoder(self, scheduler_output: "SchedulerOutput"):
        scheduled_encoder_inputs = scheduler_output.scheduled_encoder_inputs
        if not scheduled_encoder_inputs:
            return

        # Batch the multi-modal inputs.
785
786
        mm_inputs: list[MultiModalKwargs] = []
        req_input_ids: list[tuple[str, int]] = []
787
788
789
790
791
        for req_id, encoder_input_ids in scheduled_encoder_inputs.items():
            req_state = self.requests[req_id]
            for input_id in encoder_input_ids:
                mm_inputs.append(req_state.mm_inputs[input_id])
                req_input_ids.append((req_id, input_id))
792
793
794
795
796
797
798
799
800
801
802
803
804
805
806
807
808
809
810
811
812
813
814
815
816
817
818
819

        # 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.
        grouped_mm_inputs_list = group_mm_inputs_by_modality(mm_inputs)

        encoder_outputs = []
        for grouped_mm_inputs in grouped_mm_inputs_list:
            batched_mm_inputs = MultiModalKwargs.batch(grouped_mm_inputs)
            batched_mm_inputs = MultiModalKwargs.as_kwargs(batched_mm_inputs,
                                                           device=self.device)

            # 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.
            curr_group_outputs = self.model.get_multimodal_embeddings(
                **batched_mm_inputs)

            for output in curr_group_outputs:
                encoder_outputs.append(output)
820
821
822
823
824
825
826
827
828
829

        # Cache the encoder outputs.
        for (req_id, input_id), output in zip(req_input_ids, encoder_outputs):
            if req_id not in self.encoder_cache:
                self.encoder_cache[req_id] = {}
            self.encoder_cache[req_id][input_id] = output

    def _gather_encoder_outputs(
        self,
        scheduler_output: "SchedulerOutput",
830
831
    ) -> list[torch.Tensor]:
        encoder_outputs: list[torch.Tensor] = []
832
        for req_id in self.input_batch.req_ids:
833
834
835
836
837
838
839
840
841
842
843
844
845
846
847
848
849
850
851
852
853
854
855
856
857
858
859
860
861
862
863
864
            num_scheduled_tokens = scheduler_output.num_scheduled_tokens[
                req_id]
            req_state = self.requests[req_id]
            num_computed_tokens = req_state.num_computed_tokens
            mm_positions = req_state.mm_positions
            for i, pos_info in enumerate(mm_positions):
                start_pos = pos_info["offset"]
                num_encoder_tokens = pos_info["length"]

                # 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,
                    num_encoder_tokens)
                assert start_idx < end_idx
                assert req_id in self.encoder_cache
                assert i in self.encoder_cache[req_id]
                encoder_output = self.encoder_cache[req_id][i]
                encoder_outputs.append(encoder_output[start_idx:end_idx])
        return encoder_outputs

865
866
867
    def get_model(self) -> nn.Module:
        return self.model

868
869
870
871
872
873
874
875
876
877
878
879
880
881
882
883
884
885
886
887
888
889
890
891
892
893
894
895
896
897
898
899
900
901
902
903
904
905
906
907
908
909
910
911
912
913
914
    def apply_grammar_bitmask(
        self,
        scheduler_output: "SchedulerOutput",
        logits: torch.Tensor,
    ):
        # Serialization of np.ndarray is much more efficient than a tensor,
        # so we receive it in that format.
        grammar_bitmask = scheduler_output.grammar_bitmask
        if grammar_bitmask is None:
            return

        # We receive the structured output bitmask from the scheduler, but the
        # indices of the requests in the batch may not match the indices of
        # the bitmask since the scheduler doesn't know how the gpu runner is
        # ordering the requests in the batch. We need to sort the bitmask to
        # match the order of the requests used here.
        struct_out_req_batch_indices: dict[str, int] = {}
        indices_match = True
        for req_id in self.input_batch.req_ids:
            mask_index = scheduler_output.structured_output_request_ids.get(
                req_id)
            if mask_index is None:
                # not a structured output request
                continue
            batch_index = self.input_batch.req_id_to_index[req_id]
            if batch_index != mask_index:
                indices_match = False
            struct_out_req_batch_indices[req_id] = batch_index

        if not indices_match:
            # Sort the bitmask to match the order of the requests
            sorted_bitmask = np.zeros_like(grammar_bitmask)
            for req_id, batch_index in struct_out_req_batch_indices.items():
                orig_index = scheduler_output.structured_output_request_ids[
                    req_id]
                sorted_bitmask[batch_index] = grammar_bitmask[orig_index]
            grammar_bitmask = sorted_bitmask

        grammar_bitmask = torch.from_numpy(grammar_bitmask)

        # TODO: compatibility with spec decode
        xgr.apply_token_bitmask_inplace(
            logits,
            grammar_bitmask.to(self.device, non_blocking=True),
            indices=list(struct_out_req_batch_indices.values()),
        )

915
916
917
918
    @torch.inference_mode()
    def execute_model(
        self,
        scheduler_output: "SchedulerOutput",
919
        intermediate_tensors: Optional[IntermediateTensors] = None,
920
    ) -> Union[ModelRunnerOutput, torch.Tensor]:
921
        self._update_states(scheduler_output)
922
923
924
        if not scheduler_output.total_num_scheduled_tokens:
            # Return empty ModelRunnerOuptut if there's no work to do.
            return EMPTY_MODEL_RUNNER_OUTPUT
925

926
927
928
929
930
931
        if self.is_multimodal_model:
            # Run the multimodal encoder if any.
            self._execute_encoder(scheduler_output)
            encoder_outputs = self._gather_encoder_outputs(scheduler_output)
        else:
            encoder_outputs = []
932
933

        # Prepare the decoder inputs.
934
        attn_metadata, logits_indices = self._prepare_inputs(scheduler_output)
935
936
937
938
939
        num_scheduled_tokens = scheduler_output.total_num_scheduled_tokens
        if (self.use_cuda_graph
                and num_scheduled_tokens <= self.cudagraph_batch_sizes[-1]):
            # Use piecewise CUDA graphs.
            # Add padding to the batch size.
940
            num_input_tokens = self.vllm_config.pad_for_cudagraph(
941
942
943
944
                num_scheduled_tokens)
        else:
            # Eager mode.
            num_input_tokens = num_scheduled_tokens
945
946
        attn_metadata.num_input_tokens = num_input_tokens

947
948
949
950
951
952
953
954
955
956
957
958
959
960
        if self.is_multimodal_model:
            # 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.
            input_ids = self.input_ids[:num_scheduled_tokens]
            if encoder_outputs:
                inputs_embeds = self.model.get_input_embeddings(
                    input_ids, encoder_outputs)
            else:
                inputs_embeds = self.model.get_input_embeddings(input_ids)
            # TODO(woosuk): Avoid the copy. Optimize.
            self.inputs_embeds[:num_scheduled_tokens].copy_(inputs_embeds)
            inputs_embeds = self.inputs_embeds[:num_input_tokens]
            input_ids = None
961
        else:
962
963
964
965
966
967
            # 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.
            input_ids = self.input_ids[:num_input_tokens]
            inputs_embeds = None
968
969
970
971
        if self.uses_mrope:
            positions = self.mrope_positions[:, :num_input_tokens]
        else:
            positions = self.positions[:num_input_tokens]
972

973
974
975
        if get_pp_group().is_first_rank:
            intermediate_tensors = None
        else:
976
977
978
979
980
            assert intermediate_tensors is not None
            assert self.intermediate_tensors is not None
            for k, v in intermediate_tensors.items():
                self.intermediate_tensors[k][:num_input_tokens].copy_(
                    v[:num_input_tokens], non_blocking=True)
981
982
983
984
985
            intermediate_tensors = IntermediateTensors({
                k: v[:num_input_tokens]
                for k, v in self.intermediate_tensors.items()
            })

986
987
        # Run the decoder.
        # Use persistent buffers for CUDA graphs.
988
        with set_forward_context(attn_metadata, self.vllm_config):
989
            hidden_states = self.model(
990
                input_ids=input_ids,
991
                positions=positions,
992
                intermediate_tensors=intermediate_tensors,
993
                inputs_embeds=inputs_embeds,
994
            )
995
        if not get_pp_group().is_last_rank:
996
            # For mid-pipeline stages, return the hidden states.
997
            return hidden_states
998

999
        hidden_states = hidden_states[:num_scheduled_tokens]
1000
1001
        sample_hidden_states = hidden_states[logits_indices]
        logits = self.model.compute_logits(sample_hidden_states, None)
1002

1003
1004
1005
1006
        # Apply structured output bitmasks if present
        if scheduler_output.grammar_bitmask is not None:
            self.apply_grammar_bitmask(scheduler_output, logits)

1007
        # Sample the next token and get logprobs if needed.
1008
1009
1010
1011
1012
1013
1014
1015
1016
        sampling_metadata = self.input_batch.sampling_metadata
        if not self.use_spec_decode:
            sampler_output = self.model.sample(
                logits=logits,
                sampling_metadata=sampling_metadata,
            )
        else:
            draft_token_ids = [
                scheduler_output.scheduled_spec_decode_tokens.get(req_id, [])
1017
                for req_id in self.input_batch.req_ids
1018
            ]
1019
1020
1021
1022
1023
1024
1025
1026
1027
1028
1029
1030
1031
1032
1033
1034
            sample_lens = [len(tokens) + 1 for tokens in draft_token_ids]
            recover_logits_idx = np.cumsum(sample_lens) - 1
            target_probs = self.rejection_sampler.compute_probs(
                logits, sampling_metadata, sample_lens)
            sampler_output = self.model.sample(
                logits=logits[recover_logits_idx, :],
                sampling_metadata=sampling_metadata,
            )
            bonus_token_ids = sampler_output.sampled_token_ids
            output_token_ids = self.rejection_sampler(
                draft_token_ids,
                None,  # draft_probs
                bonus_token_ids,
                target_probs,
                sampling_metadata)
            sampler_output.sampled_token_ids = output_token_ids
1035
1036
1037

        # TODO(woosuk): The following loop can be slow since it iterates over
        # the requests one by one. Optimize.
1038
1039
        for i, generator in self.input_batch.generators.items():
            req_id = self.input_batch.req_ids[i]
1040
1041
1042
            req_state = self.requests[req_id]
            seq_len = (req_state.num_computed_tokens +
                       scheduler_output.num_scheduled_tokens[req_id])
1043
            if seq_len < req_state.num_tokens:
1044
                # Ignore the sampled token for partial prefills.
1045
                # Rewind the generator state as if the token was not sampled.
1046
1047
                # This relies on cuda-specific torch-internal impl details
                generator.set_offset(generator.get_offset() - 4)
1048

1049
1050
        # NOTE: GPU -> CPU Sync happens here.
        # Move as many CPU operations as possible before this sync point.
1051
1052
1053
1054
1055
1056
1057
1058
1059
1060
        logprobs_tensors = sampler_output.logprobs_tensors
        logprobs_lists = logprobs_tensors.tolists() \
            if logprobs_tensors is not None else None

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

1061
        # Get the valid generated tokens.
1062
1063
1064
        sampled_token_ids = sampler_output.sampled_token_ids
        max_gen_len = sampled_token_ids.shape[-1]
        if max_gen_len == 1:
1065
            # No spec decode tokens.
1066
1067
            valid_sampled_token_ids = sampled_token_ids.tolist()
        else:
1068
            # Includes spec decode tokens.
1069
1070
            valid_mask = sampled_token_ids != INVALID_TOKEN_ID
            gen_lens = valid_mask.sum(dim=1).tolist()
1071
            # TODO(woosuk): Optimize this.
1072
1073
1074
1075
            valid_sampled_token_ids = [
                seq.tolist()
                for seq in sampled_token_ids[valid_mask].split(gen_lens)
            ]
1076

1077
1078
1079
1080
        if not self.use_spec_decode:
            spec_token_ids = None
        else:
            spec_token_ids = self.generate_draft_token_ids(
1081
                valid_sampled_token_ids, sampling_metadata)
1082

1083
        return ModelRunnerOutput(
1084
            req_ids=self.input_batch.req_ids,
1085
            req_id_to_index=self.input_batch.req_id_to_index,
1086
            sampled_token_ids=valid_sampled_token_ids,
1087
            spec_token_ids=spec_token_ids,
1088
1089
            logprobs=logprobs_lists,
            prompt_logprobs_dict=prompt_logprobs_dict,
1090
1091
        )

1092
1093
    def generate_draft_token_ids(
        self,
1094
        sampled_token_ids: list[list[int]],
1095
        sampling_metadata: SamplingMetadata,
1096
    ) -> list[list[int]]:
1097
        # TODO(woosuk): Optimize.
1098
        draft_token_ids: list[list[int]] = []
1099
1100
1101
        for i, sampled_ids in enumerate(sampled_token_ids):
            num_sampled_ids = len(sampled_ids)
            if not num_sampled_ids:
1102
1103
1104
1105
                # Skip speculative decoding.
                draft_token_ids.append([])
                continue

1106
1107
1108
1109
1110
1111
            # Skip requests that require top-p, top-k, etc.
            req_id = self.input_batch.req_ids[i]
            if not is_spec_decode_supported(req_id, self.input_batch):
                draft_token_ids.append([])
                continue

1112
1113
            # Add sampled_token_ids to token_ids_cpu.
            start_idx = self.input_batch.num_tokens_no_spec[i]
1114
1115
            end_idx = start_idx + num_sampled_ids
            self.input_batch.token_ids_cpu[i, start_idx:end_idx] = sampled_ids
1116
1117
1118
1119
1120
1121
1122
1123
1124
1125
1126
            drafter_output = self.drafter.propose(
                self.input_batch.token_ids_cpu[i, :end_idx],
                self.speculative_config.ngram_prompt_lookup_min,
                self.speculative_config.num_speculative_tokens,
            )
            if drafter_output is None or len(drafter_output) == 0:
                draft_token_ids.append([])
            else:
                draft_token_ids.append(drafter_output.tolist())
        return draft_token_ids

1127
1128
1129
    def load_model(self) -> None:
        logger.info("Starting to load model %s...", self.model_config.model)
        with DeviceMemoryProfiler() as m:  # noqa: SIM117
1130
            time_before_load = time.perf_counter()
Joe Runde's avatar
Joe Runde committed
1131
            self.model = get_model(vllm_config=self.vllm_config)
1132
1133
1134
1135
1136
1137
            if self.lora_config:
                self.model = self.load_lora_model(self.model,
                                                  self.model_config,
                                                  self.scheduler_config,
                                                  self.lora_config,
                                                  self.device)
1138
            time_after_load = time.perf_counter()
1139
        self.model_memory_usage = m.consumed_memory
1140
        logger.info("Model loading took %.4f GB and %.6f seconds",
1141
1142
                    self.model_memory_usage / float(2**30),
                    time_after_load - time_before_load)
1143

1144
1145
1146
1147
    def _get_prompt_logprobs_dict(
        self,
        hidden_states: torch.Tensor,
        scheduler_output: "SchedulerOutput",
1148
    ) -> dict[str, Optional[LogprobsTensors]]:
1149
1150
1151
1152
        num_prompt_logprobs_dict = self.input_batch.num_prompt_logprobs
        if not num_prompt_logprobs_dict:
            return {}

1153
        prompt_logprobs_dict: dict[str, Optional[LogprobsTensors]] = {}
1154
1155
1156
1157
1158
1159
1160
1161
1162
1163
1164
1165
1166
1167
1168
1169
1170
1171
1172
1173
1174
1175
1176
1177
1178
1179
1180
1181
1182
1183
1184
1185
1186
1187
1188
1189
1190
1191
1192
1193
1194
1195
1196
1197
1198
1199
1200
1201
1202
1203
1204
1205
1206
1207
1208
1209
1210
1211
1212
1213

        # 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():

            num_tokens = scheduler_output.num_scheduled_tokens[req_id]

            # Get metadata for this request.
            request = self.requests[req_id]
            num_prompt_tokens = len(request.prompt_token_ids)
            prompt_token_ids = torch.tensor(request.prompt_token_ids).to(
                self.device, non_blocking=True)

            # Determine number of logits to retrieve.
            start_tok = request.num_computed_tokens + 1
            num_remaining_tokens = num_prompt_tokens - start_tok
            if num_tokens < num_remaining_tokens:
                # This is a chunk, more tokens remain.
                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)

            # 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]
            offset = self.query_start_loc_np[req_idx].item()
            prompt_hidden_states = hidden_states[offset:offset + num_logits]
            logits = self.model.compute_logits(prompt_hidden_states, None)

            # 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.
            tgt_token_ids = prompt_token_ids[start_tok:start_tok + num_logits]

            # Compute prompt logprobs.
            logprobs = self.model.sampler.compute_logprobs(logits)
            token_ids, logprobs, ranks = self.model.sampler.gather_logprobs(
                logprobs, num_prompt_logprobs, tgt_token_ids)

            # Transfer GPU->CPU async.
            prompt_logprobs_dict[req_id] = LogprobsTensors(
                token_ids.to("cpu", non_blocking=True),
                logprobs.to("cpu", non_blocking=True),
                ranks.to("cpu", non_blocking=True),
            )

        # 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]

        # Must synchronize the non-blocking GPU->CPU transfers.
        torch.cuda.synchronize()

        return prompt_logprobs_dict

1214
1215
1216
1217
1218
    @torch.inference_mode()
    def _dummy_run(
        self,
        num_tokens: int,
    ) -> torch.Tensor:
1219

1220
1221
1222
1223
1224
1225
1226
1227
1228
1229
1230
1231
1232
        # 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
        num_reqs = max_num_reqs if num_tokens >= max_num_reqs else num_tokens
        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
        assert sum(num_scheduled_tokens_list) == num_tokens
        assert len(num_scheduled_tokens_list) == num_reqs
        num_scheduled_tokens = np.array(num_scheduled_tokens_list,
                                        dtype=np.int32)
1233

1234
1235
1236
1237
1238
1239
1240
1241
1242
1243
1244
1245
1246
1247
1248
1249
1250
1251
1252
1253
1254
1255
1256
1257
1258
1259
1260
1261
1262
1263
1264
1265
1266
1267
1268
1269
1270
1271
1272
1273
1274
1275
1276
1277
1278
1279
1280
1281
1282
1283
1284
1285
1286
1287
1288
1289
1290
1291
1292
1293
1294
1295
1296
1297
1298
1299
1300
1301
1302
1303
1304
1305
1306
        with self.maybe_dummy_run_with_lora(self.lora_config,
                                            num_scheduled_tokens):
            model = self.model
            if self.is_multimodal_model:
                input_ids = None
                inputs_embeds = self.inputs_embeds[:num_tokens]
            else:
                input_ids = self.input_ids[:num_tokens]
                inputs_embeds = None
            if self.uses_mrope:
                positions = self.mrope_positions[:, :num_tokens]
            else:
                positions = self.positions[:num_tokens]

            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,
                            device=self.device))
                intermediate_tensors = IntermediateTensors({
                    k: v[:num_tokens]
                    for k, v in self.intermediate_tensors.items()
                })

            with set_forward_context(None,
                                     self.vllm_config,
                                     num_tokens=num_tokens):
                hidden_states = model(
                    input_ids=input_ids,
                    positions=positions,
                    intermediate_tensors=intermediate_tensors,
                    inputs_embeds=inputs_embeds,
                )

        logit_indices = np.cumsum(num_scheduled_tokens) - 1
        return hidden_states[logit_indices]

    @torch.inference_mode()
    def _dummy_sampler_run(
        self,
        hidden_states: torch.Tensor,
    ) -> torch.Tensor:

        logits = self.model.compute_logits(hidden_states, None)
        num_reqs = logits.size(0)

        dummy_tensors = lambda v: torch.full(
            (num_reqs, ), v, device=self.device)

        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),
            min_p=None,
            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)],
            min_tokens={},
            logit_bias=[None for _ in range(num_reqs)],
            allowed_token_ids_mask=None,
            bad_words_token_ids={},
        )
1307
1308
1309
1310
1311
1312
1313
1314
1315
1316
1317
1318
        try:
            sampler_output = self.model.sample(
                logits=logits, sampling_metadata=dummy_metadata)
        except RuntimeError as e:
            if 'out of memory' in str(e):
                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 "
                    "initializing the engine.") from e
            else:
                raise e
1319
        return sampler_output
1320
1321

    def profile_run(self) -> None:
1322
        # Profile with multimodal encoder & encoder cache.
1323
1324
1325
        # TODO: handle encoder-decoder models once we support them.
        if (self.is_multimodal_model and self.max_num_encoder_input_tokens > 0
                and self.encoder_cache_size > 0):
1326

1327
            # NOTE: Currently model is profiled with a single non-text
1328
1329
            # modality with the max possible input tokens even when
            # it supports multiple.
1330
1331
1332
            max_tokens_by_modality_dict = (
                MULTIMODAL_REGISTRY.
                get_max_tokens_per_item_by_nonzero_modality(self.model_config))
1333
1334
1335
1336
            dummy_data_modality, max_tokens_per_mm_item = max(
                max_tokens_by_modality_dict.items(), key=lambda item: item[1])

            # Check how many items of this modality can be supported by
1337
1338
1339
1340
1341
1342
            # the encoder budget.
            encoder_budget = min(self.max_num_encoder_input_tokens,
                                 self.encoder_cache_size)

            max_num_mm_items_encoder_budget = cdiv(encoder_budget,
                                                   max_tokens_per_mm_item)
1343
1344
1345

            # Check how many items of this modality can be supported by
            # the decoder budget.
1346
1347
            max_mm_items_per_req = self.mm_registry.get_mm_limits_per_prompt(
                self.model_config)[dummy_data_modality]
1348
1349
1350
1351
1352
1353
1354
1355
1356
1357

            # NOTE: We do not consider max_num_batched_tokens on purpose
            # because the multimodal embeddings can be generated in advance
            # and chunked prefilled.
            max_num_mm_items_decoder_budget = self.max_num_reqs * \
                max_mm_items_per_req

            max_num_mm_items = min(max_num_mm_items_encoder_budget,
                                   max_num_mm_items_decoder_budget)

1358
1359
1360
1361
1362
1363
1364
1365
1366
1367
1368
1369
            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_num_mm_items, dummy_data_modality)

            # Create dummy batch of multimodal inputs.
            dummy_request_data = self.input_registry.dummy_data_for_profiling(
                model_config=self.model_config,
                seq_len=self.max_num_tokens,
                mm_registry=self.mm_registry,
            )
            dummy_mm_data = dummy_request_data.multi_modal_data
1370
1371
1372
1373
            if not isinstance(dummy_mm_data, MultiModalKwargs):
                # TODO: Delete this check once input mapper is fully removed.
                raise RuntimeError(
                    "Legacy input mapper is not supported in V1")
1374

1375
            # Dummy data definition may contain multiple multimodal items
1376
1377
1378
            # (e.g, multiple images) for a single request, therefore here we
            # always replicate first item by max_num_mm_items times since in V1
            # they are scheduled to be processed separately.
1379
1380
1381
            dummy_mm_item = dummy_mm_data.get_item(
                modality=dummy_data_modality, item_index=0)
            dummy_mm_kwargs = MultiModalKwargs.from_items([dummy_mm_item])
1382

1383
            batched_dummy_mm_inputs = MultiModalKwargs.batch(
1384
                [dummy_mm_kwargs] * max_num_mm_items)
1385
1386
1387
1388
1389
1390
1391
1392
1393
1394
1395
1396
1397
1398
1399
1400
            batched_dummy_mm_inputs = MultiModalKwargs.as_kwargs(
                batched_dummy_mm_inputs, device=self.device)

            # Run multimodal encoder.
            dummy_encoder_outputs = self.model.get_multimodal_embeddings(
                **batched_dummy_mm_inputs)
            assert len(dummy_encoder_outputs) == max_num_mm_items, (
                "Expected dimension 0 of encoder outputs to match the number "
                f"of multimodal data items: {max_num_mm_items}, got "
                f"{len(dummy_encoder_outputs)=} instead. This is most likely "
                "due to the 'get_multimodal_embeddings' method of the model "
                "not implemented correctly.")

            # Cache the dummy encoder outputs.
            self.encoder_cache["tmp"] = dict(enumerate(dummy_encoder_outputs))

1401
1402
1403
1404
1405
1406
1407
1408
        hidden_states = self._dummy_run(self.max_num_tokens)
        if get_pp_group().is_last_rank:
            sampler_output = self._dummy_sampler_run(hidden_states)
        else:
            sampler_output = None
        torch.cuda.synchronize()
        del hidden_states, sampler_output
        self.encoder_cache.clear()
1409
        gc.collect()
1410
1411

    def capture_model(self) -> None:
1412
1413
        if not self.use_cuda_graph:
            logger.warning(
1414
                "Skipping CUDA graph capture. Please add "
1415
                "-O %s to use CUDA graphs.", CompilationLevel.PIECEWISE)
1416
1417
1418
1419
1420
            return

        start_time = time.perf_counter()
        start_free_gpu_memory = torch.cuda.mem_get_info()[0]

1421
1422
1423
        # 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.
1424
        with graph_capture(device=self.device):
1425
            for num_tokens in reversed(self.cudagraph_batch_sizes):
1426
1427
                for _ in range(self.vllm_config.compilation_config.
                               cudagraph_num_of_warmups):
1428
1429
                    self._dummy_run(num_tokens)
                self._dummy_run(num_tokens)
1430
1431
1432
1433
1434
1435
1436
1437

        end_time = time.perf_counter()
        end_free_gpu_memory = torch.cuda.mem_get_info()[0]
        elapsed_time = end_time - start_time
        cuda_graph_size = start_free_gpu_memory - end_free_gpu_memory
        # This usually takes 5~20 seconds.
        logger.info("Graph capturing finished in %.0f secs, took %.2f GiB",
                    elapsed_time, cuda_graph_size / (1 << 30))
1438

1439
1440
1441
1442
    def initialize_kv_cache(self, kv_cache_config: KVCacheConfig) -> None:
        """
        Initialize KV cache based on `kv_cache_config`.
        Args:
1443
            kv_cache_config: Configuration for the KV cache, including the KV
1444
1445
1446
1447
1448
1449
1450
            cache size of each layer
        """
        if len(kv_cache_config.groups) > 1:
            raise NotImplementedError(
                "Hybrid models with more than one KV cache type are not "
                "supported yet.")

1451
        kv_caches: dict[str, torch.Tensor] = {}
1452
1453
1454
1455
1456
1457

        for layer_name, layer_spec in kv_cache_config.kv_cache_spec.items():
            tensor_config = kv_cache_config.tensors[layer_name]
            assert tensor_config.size % layer_spec.page_size_bytes == 0
            num_blocks = tensor_config.size // layer_spec.page_size_bytes
            if isinstance(layer_spec, FullAttentionSpec):
1458
                kv_cache_shape = self.attn_backend.get_kv_cache_shape(
1459
1460
1461
1462
1463
1464
1465
1466
1467
                    num_blocks, layer_spec.block_size, layer_spec.num_kv_heads,
                    layer_spec.head_size)
                dtype = layer_spec.dtype
                kv_caches[layer_name] = torch.zeros(kv_cache_shape,
                                                    dtype=dtype,
                                                    device=self.device)
            else:
                raise NotImplementedError

1468
        bind_kv_cache(
1469
            kv_caches,
1470
            self.vllm_config.compilation_config.static_forward_context,
1471
1472
1473
1474
            self.kv_caches)

    def get_kv_cache_spec(self) -> KVCacheSpec:
        """
1475
        Generates the KVCacheSpec by parsing the kv cache format from each
1476
1477
        Attention module in the static forward context.
        Returns:
1478
            KVCacheSpec: A dictionary mapping layer names to their KV cache
1479
1480
1481
1482
1483
            format. Layers that do not need KV cache are not included.
        """

        forward_ctx = self.vllm_config.compilation_config.static_forward_context
        block_size = self.vllm_config.cache_config.block_size
1484
        use_mla = self.vllm_config.model_config.use_mla
1485
1486
        kv_cache_spec: KVCacheSpec = {}
        for layer_name, attn_module in forward_ctx.items():
1487
1488
1489
            if isinstance(attn_module, FusedMoE):
                continue

1490
            # TODO: Support other attention modules, e.g., sliding window,
1491
            # cross-attention
1492
1493
1494
1495
1496
1497
1498
            assert isinstance(attn_module, Attention)
            if attn_module.attn_type == AttentionType.DECODER:
                kv_cache_spec[layer_name] = FullAttentionSpec(
                    block_size=block_size,
                    num_kv_heads=attn_module.num_kv_heads,
                    head_size=attn_module.head_size,
                    dtype=attn_module.dtype,
1499
                    use_mla=use_mla)
1500
1501
1502
1503
1504
1505
1506
1507
1508
1509
1510
            elif attn_module.attn_type in (AttentionType.ENCODER,
                                           AttentionType.ENCODER_ONLY):
                # encoder-only attention does not need KV cache.
                continue
            elif attn_module.attn_type == AttentionType.ENCODER_DECODER:
                raise NotImplementedError
            else:
                raise ValueError(
                    f"Unknown attention type: {attn_module.attn_type}")

        return kv_cache_spec