gpu_model_runner.py 51.8 KB
Newer Older
1
2
# SPDX-License-Identifier: Apache-2.0

3
import gc
4
import time
5
from typing import TYPE_CHECKING, Dict, List, Optional, Tuple, cast
6
7
8
9
10
11

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

12
13
from vllm.attention.backends.abstract import AttentionType
from vllm.attention.layer import Attention
14
from vllm.config import CompilationLevel, VllmConfig
15
from vllm.distributed.parallel_state import graph_capture
16
from vllm.forward_context import set_forward_context
17
from vllm.inputs import INPUT_REGISTRY
18
from vllm.logger import init_logger
19
from vllm.model_executor.layers.rotary_embedding import MRotaryEmbedding
20
from vllm.model_executor.model_loader import get_model
21
from vllm.multimodal import MULTIMODAL_REGISTRY, MultiModalKwargs
22
from vllm.multimodal.utils import group_mm_inputs_by_modality
23
from vllm.sampling_params import SamplingType
24
from vllm.utils import (STR_DTYPE_TO_TORCH_DTYPE, DeviceMemoryProfiler,
25
                        LayerBlockType, cdiv, is_pin_memory_available)
26
27
from vllm.v1.attention.backends.flash_attn import (FlashAttentionBackend,
                                                   FlashAttentionMetadata)
28
from vllm.v1.core.encoder_cache_manager import compute_encoder_budget
29
from vllm.v1.engine.mm_input_mapper import MMInputMapperClient
30
31
from vllm.v1.kv_cache_interface import (FullAttentionSpec, KVCacheConfig,
                                        KVCacheSpec)
32
33
from vllm.v1.outputs import ModelRunnerOutput
from vllm.v1.sample.metadata import SamplingMetadata
34
from vllm.v1.utils import bind_kv_cache
35
from vllm.v1.worker.gpu_input_batch import CachedRequestState, InputBatch
36
37
38
39
40
41
42
43
44
45
46

if TYPE_CHECKING:
    from vllm.v1.core.scheduler import SchedulerOutput

logger = init_logger(__name__)


class GPUModelRunner:

    def __init__(
        self,
47
        vllm_config: VllmConfig,
48
        device: torch.device,
49
    ):
50
51
52
53
54
55
56
57
58
59
        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
60

61
62
63
64
        model_config = self.model_config
        cache_config = self.cache_config
        scheduler_config = self.scheduler_config
        parallel_config = self.parallel_config
65
        self.device = device
66
67
68
69
70
71
72
73
        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]

74
        self.is_multimodal_model = model_config.is_multimodal_model
75
76
77
78
79
        self.sliding_window = model_config.get_sliding_window()
        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
80
        self.max_num_reqs = scheduler_config.max_num_seqs
81
82

        # Model-related.
83
84
        self.num_attn_layers = model_config.get_num_layers_by_block_type(
            parallel_config, LayerBlockType.attention)
85
86
        self.num_query_heads = model_config.get_num_attention_heads(
            parallel_config)
87
88
        self.num_kv_heads = model_config.get_num_kv_heads(parallel_config)
        self.head_size = model_config.get_head_size()
89
90
91
        self.hidden_size = model_config.get_hidden_size()

        # Multi-modal data support
92
93
        self.input_registry = INPUT_REGISTRY
        self.mm_registry = MULTIMODAL_REGISTRY
94

95
96
97
98
        # NOTE: Initialized input mapper is only used for processing dummy
        # multimodal data into multimodal kwargs for GPU memory profiling.
        self.mm_input_mapper_profiling = MMInputMapperClient(self.model_config)
        self.mm_input_mapper_profiling.use_cache = False
99

100
101
102
103
104
105
        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
106
107
108
109

        # Lazy initialization
        # self.model: nn.Module  # Set after load_model
        self.kv_caches: List[torch.Tensor] = []
110
111
        # req_id -> (input_id -> encoder_output)
        self.encoder_cache: Dict[str, Dict[int, torch.Tensor]] = {}
112
113
114
115
116

        # Request states.
        self.requests: Dict[str, CachedRequestState] = {}
        # Persistent batch.
        self.input_batch = InputBatch(
117
            max_num_reqs=self.max_num_reqs,
118
119
120
121
            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,
122
            vocab_size=model_config.get_vocab_size(),
123
124
        )

125
        self.use_cuda_graph = (self.vllm_config.compilation_config.level
126
127
128
                               == CompilationLevel.PIECEWISE
                               and not self.model_config.enforce_eager)
        # TODO(woosuk): Provide an option to tune the max cudagraph batch size.
129
130
131
132
        # 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(
133
134
            reversed(
                self.vllm_config.compilation_config.cudagraph_capture_sizes))
135

136
137
138
139
        # Cache the device properties.
        self.device_properties = torch.cuda.get_device_properties(self.device)
        self.num_sms = self.device_properties.multi_processor_count

140
141
142
143
        # Persistent buffers for CUDA graphs.
        self.input_ids = torch.zeros(self.max_num_tokens,
                                     dtype=torch.int32,
                                     device=self.device)
144
145
146
        self.positions = torch.zeros(self.max_num_tokens,
                                     dtype=torch.int64,
                                     device=self.device)
147
148
149

        # Only relevant for models using M-RoPE (e.g, Qwen2-VL)
        if self.model_config.uses_mrope:
Roger Wang's avatar
Roger Wang committed
150
151
152
153
            # 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
154
155
156
157
158
159

            # 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
160
            self.mrope_positions = torch.zeros((3, self.max_num_tokens + 1),
161
162
                                               dtype=torch.int64,
                                               device=self.device)
Roger Wang's avatar
Roger Wang committed
163
164
165
166
167
            self.mrope_positions_cpu = torch.zeros(
                (3, self.max_num_tokens + 1),
                dtype=torch.int64,
                device="cpu",
                pin_memory=self.pin_memory)
168

169
170
171
172
        self.inputs_embeds = torch.zeros(
            (self.max_num_tokens, self.hidden_size),
            dtype=self.dtype,
            device=self.device)
173

174
        # OPTIMIZATION: Cache the tensors rather than creating them every step.
175
        self.arange_np = np.arange(max(self.max_num_reqs + 1,
176
177
                                       self.max_model_len,
                                       self.max_num_tokens),
178
179
180
181
                                   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.
182
183
184
185
186
187
188
189
190
191
192
193
194
195
196
197
198
199
200
201
        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()
202
203
204
205
206
        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()
207

208
209
210
211
212
213
214
215
216
217
218
219
    def _update_states(self, scheduler_output: "SchedulerOutput") -> bool:
        """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.

        Returns:
            True if there is a new/resumed/paused/finished request in the batch.
            If False, we can skip copying SamplingMetadata to the GPU.
        """
        # Remove finished requests from the cached states.
220
221
        for req_id in scheduler_output.finished_req_ids:
            self.requests.pop(req_id, None)
222
            self.encoder_cache.pop(req_id, None)
223
224
225
226
227
228
229
230
231
232
233
        # 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.
        removed_req_indices: List[int] = []
        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)
234
235
236
237
238
239
240
241

        # 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)
242

243
244
245
246
247
248
249
250
251
252
253
254
255
        # 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:
256
            req_index = self.input_batch.remove_request(req_id)
257
258
            assert req_index is not None
            removed_req_indices.append(req_index)
259
260
261

        req_ids_to_add: List[str] = []
        # Add new requests to the cached states.
262
263
264
        for new_req_data in scheduler_output.scheduled_new_reqs:
            req_id = new_req_data.req_id
            sampling_params = new_req_data.sampling_params
265
            if sampling_params.sampling_type == SamplingType.RANDOM_SEED:
266
267
268
269
270
                generator = torch.Generator(device=self.device)
                generator.manual_seed(sampling_params.seed)
            else:
                generator = None

271
272
            self.requests[req_id] = CachedRequestState(
                req_id=req_id,
273
274
275
276
                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,
277
278
                sampling_params=sampling_params,
                generator=generator,
279
280
                block_ids=new_req_data.block_ids,
                num_computed_tokens=new_req_data.num_computed_tokens,
281
282
                output_token_ids=[],
            )
283
284
285
286
287
288
289
290
291
292
293
294
295
296
297
298
299
300
301
302
303
304
305
306
307
308
309
310
311

            # Only relevant for models using M-RoPE (e.g, Qwen2-VL)
            if self.model_config.uses_mrope:
                image_grid_thw = []
                video_grid_thw = []
                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())

                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,
                        image_grid_thw=image_grid_thw,
                        video_grid_thw=video_grid_thw,
                        image_token_id=hf_config.image_token_id,
                        video_token_id=hf_config.video_token_id,
                        vision_start_token_id=hf_config.vision_start_token_id,
                        vision_end_token_id=hf_config.vision_end_token_id,
                        spatial_merge_size=hf_config.vision_config.
                        spatial_merge_size,
                    )

312
313
            req_ids_to_add.append(req_id)

314
315
316
        # Update the states of the running/resumed requests.
        for req_data in scheduler_output.scheduled_cached_reqs:
            req_id = req_data.req_id
317
318
            req_state = self.requests[req_id]

319
320
321
322
323
324
325
326
327
328
329
330
331
332
333
334
335
336
337
338
339
340
341
342
343
            # Update the cached states.
            req_state.num_computed_tokens = req_data.num_computed_tokens
            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] = (
                req_data.num_computed_tokens)
            start_index = len(req_state.block_ids) - len(
                req_data.new_block_ids)
            self.input_batch.block_table.append_row(req_index, start_index,
                                                    req_data.new_block_ids)
344
345
346
347
348
349
350
351
352
353
354
355
356
357
358
359
360

        # 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)
361
        return len(unscheduled_req_ids) > 0 or len(req_ids_to_add) > 0
362
363
364
365
366
367
368
369
370

    def _prepare_inputs(self, scheduler_output: "SchedulerOutput"):
        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.
371
        self.input_batch.block_table.commit(num_reqs)
372
373
374
375
376
377

        # Get the number of scheduled tokens for each request.
        # TODO: The Python loop can be slow. Optimize.
        num_scheduled_tokens = []
        max_num_scheduled_tokens = 0
        for req_id in self.input_batch.req_ids[:num_reqs]:
378
            assert req_id is not None
379
380
381
382
383
384
385
386
387
            num_tokens = scheduler_output.num_scheduled_tokens[req_id]
            num_scheduled_tokens.append(num_tokens)
            max_num_scheduled_tokens = max(max_num_scheduled_tokens,
                                           num_tokens)
        num_scheduled_tokens = np.array(num_scheduled_tokens, dtype=np.int32)
        assert max_num_scheduled_tokens > 0

        # Get request indices.
        # E.g., [2, 5, 3] -> [0, 0, 1, 1, 1, 1, 1, 2, 2, 2]
388
389
        req_indices = np.repeat(self.arange_np[:num_reqs],
                                num_scheduled_tokens)
390
391
392

        # Get batched arange.
        # E.g., [2, 5, 3] -> [0, 1, 0, 1, 2, 3, 4, 0, 1, 2]
393
394
395
396
397
398
399
400
401
        # 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
402
403

        # Get positions.
404
        positions_np = self.positions_np[:total_num_scheduled_tokens]
405
406
407
408
        np.add(self.input_batch.num_computed_tokens_cpu[req_indices],
               arange,
               out=positions_np)

409
410
411
412
413
        # Calculate M-RoPE positions.
        # Only relevant for models using M-RoPE (e.g, Qwen2-VL)
        if self.model_config.uses_mrope:
            self._calc_mrope_positions(scheduler_output)

414
415
416
417
        # 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.
418
419
        token_indices = (positions_np +
                         req_indices * self.input_batch.token_ids_cpu.shape[1])
420
421
422
423
        # 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(),
424
                           0,
425
426
                           torch.from_numpy(token_indices),
                           out=self.input_ids_cpu[:total_num_scheduled_tokens])
427
428

        # Calculate the slot mapping.
429
430
431
432
433
        # 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.
434
435
436
437
438
        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.
439
440
        block_table_cpu = self.input_batch.block_table.get_cpu_tensor()
        block_numbers = block_table_cpu.flatten()[block_table_indices].numpy()
441
442
443
444
        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])
445
446

        # Prepare the attention metadata.
447
        self.query_start_loc_np[0] = 0
448
        self.query_start_loc_np[1:num_reqs + 1] = cu_num_tokens
449

450
451
452
453
        self.seq_lens_np[:num_reqs] = (
            self.input_batch.num_computed_tokens_cpu[:num_reqs] +
            num_scheduled_tokens)
        max_seq_len = self.seq_lens_np[:num_reqs].max()
454
455
456
457

        # 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)
458
459
460
461
462
463
464
465
466
467
        if self.model_config.uses_mrope:
            # 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)
468
469
        query_start_loc = self.query_start_loc_cpu[:num_reqs + 1].to(
            self.device, non_blocking=True)
470
471
        seq_lens = self.seq_lens_cpu[:num_reqs].to(self.device,
                                                   non_blocking=True)
472
473
        slot_mapping = self.slot_mapping_cpu[:total_num_scheduled_tokens].to(
            self.device, non_blocking=True).long()
474
475
476
477
478
479
480
481
482
483
484
485
486
487
488
489
490
491
492
493
494
495
496
497
498
499
500
501
502
503
504
505
506
507
508
509
510
511
512
513
514
515
516
517
518
519
520
521
522
523
524
525
526
527
528
529
530
531
532
533
534
535
536
537
538
539
540
541
542

        # Prepare for cascade attention if needed.
        common_prefix_len = (scheduler_output.num_common_prefix_blocks *
                             self.block_size)
        if common_prefix_len == 0:
            # Common case.
            use_cascade = False
        else:
            # 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.
            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)
            use_cascade = FlashAttentionBackend.use_cascade_attention(
                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
                use_sliding_window=self.sliding_window is not None,
                num_sms=self.num_sms,
            )

        if use_cascade:
            # TODO: Optimize.
            cu_prefix_query_lens = torch.tensor(
                [0, total_num_scheduled_tokens],
                dtype=torch.int32,
                device=self.device)
543
544
545
546
547
            prefix_kv_lens = torch.tensor([common_prefix_len],
                                          dtype=torch.int32,
                                          device=self.device)
            suffix_kv_lens = (self.seq_lens_np[:num_reqs] - common_prefix_len)
            suffix_kv_lens = torch.from_numpy(suffix_kv_lens).to(self.device)
548
549
        else:
            cu_prefix_query_lens = None
550
551
            prefix_kv_lens = None
            suffix_kv_lens = None
552

553
        attn_metadata = FlashAttentionMetadata(
554
            num_actual_tokens=total_num_scheduled_tokens,
555
556
557
            max_query_len=max_num_scheduled_tokens,
            query_start_loc=query_start_loc,
            max_seq_len=max_seq_len,
558
            seq_lens=seq_lens,
559
560
            block_table=(
                self.input_batch.block_table.get_device_tensor()[:num_reqs]),
561
            slot_mapping=slot_mapping,
562
563
564
            use_cascade=use_cascade,
            common_prefix_len=common_prefix_len,
            cu_prefix_query_lens=cu_prefix_query_lens,
565
566
            prefix_kv_lens=prefix_kv_lens,
            suffix_kv_lens=suffix_kv_lens,
567
        )
568
569
570
571
        # 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.
572
573
        # TODO: Support prompt logprobs.
        logits_indices = query_start_loc[1:] - 1
574
        return attn_metadata, logits_indices
575

576
577
578
579
580
581
582
583
584
585
586
587
588
589
590
591
592
593
594
595
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
    def _calc_mrope_positions(self, scheduler_output: "SchedulerOutput"):
        mrope_pos_ptr = 0
        num_reqs = self.input_batch.num_reqs
        for index, req_id in enumerate(self.input_batch.req_ids[:num_reqs]):
            assert req_id is not None

            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

631
632
    def _prepare_sampling(
        self,
633
        batch_changed: bool,
634
635
    ) -> SamplingMetadata:
        # Create the sampling metadata.
636
637
638
639
640
        req_id_output_token_ids: Dict[str, List[int]] = \
            {req_id: req.output_token_ids \
                for req_id, req in self.requests.items()}

        sampling_metadata = self.input_batch.make_sampling_metadata(
641
            req_id_output_token_ids, skip_copy=not batch_changed)
642
643
        return sampling_metadata

644
645
646
647
648
649
650
    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.
        mm_inputs: List[MultiModalKwargs] = []
651
        req_input_ids: List[Tuple[str, int]] = []
652
653
654
655
656
        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))
657
658
659
660
661
662
663
664
665
666
667
668
669
670
671
672
673
674
675
676
677
678
679
680
681
682
683
684

        # 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)
685
686
687
688
689
690
691
692
693
694
695
696
697
698

        # 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",
    ) -> List[torch.Tensor]:
        encoder_outputs: List[torch.Tensor] = []
        num_reqs = self.input_batch.num_reqs
        for req_id in self.input_batch.req_ids[:num_reqs]:
699
            assert req_id is not None
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
            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

732
733
734
    def get_model(self) -> nn.Module:
        return self.model

735
736
737
738
739
    @torch.inference_mode()
    def execute_model(
        self,
        scheduler_output: "SchedulerOutput",
    ) -> ModelRunnerOutput:
740
        batch_changed = self._update_states(scheduler_output)
741

742
743
744
745
746
747
        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 = []
748
749

        # Prepare the decoder inputs.
750
        attn_metadata, logits_indices = self._prepare_inputs(scheduler_output)
751
752
753
754
755
        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.
756
            num_input_tokens = self.vllm_config.pad_for_cudagraph(
757
758
759
760
                num_scheduled_tokens)
        else:
            # Eager mode.
            num_input_tokens = num_scheduled_tokens
761
762
        attn_metadata.num_input_tokens = num_input_tokens

763
764
765
766
767
768
769
770
771
772
773
774
775
776
        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
777
        else:
778
779
780
781
782
783
            # 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
784
785
786

        # Run the decoder.
        # Use persistent buffers for CUDA graphs.
787
        with set_forward_context(attn_metadata, self.vllm_config):
788
789
790
            positions = self.mrope_positions[:, :num_input_tokens] \
                if self.model_config.uses_mrope \
                else self.positions[:num_input_tokens]
791
            hidden_states = self.model(
792
                input_ids=input_ids,
793
                positions=positions,
794
                kv_caches=self.kv_caches,
795
                attn_metadata=None,
796
                inputs_embeds=inputs_embeds,
797
            )
798
        hidden_states = hidden_states[:num_scheduled_tokens]
799
800
801
802
        hidden_states = hidden_states[logits_indices]
        logits = self.model.compute_logits(hidden_states, None)

        # Sample the next token and get logprobs if needed.
803
        sampling_metadata = self._prepare_sampling(batch_changed)
804
805
806
807
808
809
810
811
        sampler_output = self.model.sample(
            logits=logits,
            sampling_metadata=sampling_metadata,
        )

        # TODO(woosuk): The following loop can be slow since it iterates over
        # the requests one by one. Optimize.
        num_reqs = self.input_batch.num_reqs
812
        request_seq_lens: List[Tuple[int, CachedRequestState, int]] = []
813
        for i, req_id in enumerate(self.input_batch.req_ids[:num_reqs]):
814
            assert req_id is not None
815
816
817
818
819
820
            req_state = self.requests[req_id]
            seq_len = (req_state.num_computed_tokens +
                       scheduler_output.num_scheduled_tokens[req_id])
            assert seq_len <= req_state.num_tokens
            if seq_len == req_state.num_tokens:
                # Append the sampled token to the output token ids.
821
                self.input_batch.num_tokens[i] += 1
822
823
824
                # OPTIMIZATION: Priming the state updates for later updates.
                req_state.output_token_ids.append(0)
                request_seq_lens.append((i, req_state, seq_len))
825
826
827
            else:
                # Ignore the sampled token from the partial request.
                # Rewind the generator state as if the token was not sampled.
828
                generator = self.input_batch.generators.get(i)
829
                if generator is not None:
830
831
                    # This relies on cuda-specific torch-internal impl details
                    generator.set_offset(generator.get_offset() - 4)
832

833
834
835
836
837
838
839
840
841
842
843
844
845
846
847
        # num_reqs entries should be non-None
        assert all(
            req_id is not None for req_id in
            self.input_batch.req_ids[:num_reqs]), "req_ids contains None"
        req_ids = cast(List[str], self.input_batch.req_ids[:num_reqs])

        # NOTE: GPU -> CPU Sync happens here.
        # Move as many CPU operations as possible before this sync point.
        sampled_token_ids = sampler_output.sampled_token_ids.tolist()
        # Update with the actual token ids
        for i, req_state, seq_len in request_seq_lens:
            token_id = sampled_token_ids[i]
            self.input_batch.token_ids_cpu[i, seq_len] = token_id
            req_state.output_token_ids[-1] = token_id

848
849
850
851
852
853
854
855
        if sampler_output.logprob_token_ids is None:
            logprob_token_ids = None
        else:
            logprob_token_ids = sampler_output.logprob_token_ids.cpu()
        if sampler_output.logprobs is None:
            logprobs = None
        else:
            logprobs = sampler_output.logprobs.cpu()
856

857
        model_runner_output = ModelRunnerOutput(
858
            req_ids=req_ids,
859
            req_id_to_index=self.input_batch.req_id_to_index,
860
            sampled_token_ids=sampled_token_ids,
861
862
863
864
865
866
867
868
            logprob_token_ids_cpu=logprob_token_ids,
            logprobs_cpu=logprobs,
        )
        return model_runner_output

    def load_model(self) -> None:
        logger.info("Starting to load model %s...", self.model_config.model)
        with DeviceMemoryProfiler() as m:  # noqa: SIM117
Joe Runde's avatar
Joe Runde committed
869
            self.model = get_model(vllm_config=self.vllm_config)
870
871
872
873
874

        self.model_memory_usage = m.consumed_memory
        logger.info("Loading model weights took %.4f GB",
                    self.model_memory_usage / float(2**30))

875
876
877
878
    @torch.inference_mode()
    def _dummy_run(
        self,
        num_tokens: int,
879
        kv_caches: Optional[List[torch.Tensor]] = None,
880
    ) -> torch.Tensor:
881
882
883
        model = self.model
        if kv_caches is None:
            kv_caches = self.kv_caches
884
885
886
887
888
889
        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
890
        with set_forward_context(None, self.vllm_config):
891
892
893
            positions = self.mrope_positions[:, :num_tokens] \
                if self.model_config.uses_mrope \
                else self.positions[:num_tokens]
894
            hidden_states = model(
895
                input_ids=input_ids,
896
                positions=positions,
897
898
                kv_caches=kv_caches,
                attn_metadata=None,
899
900
                inputs_embeds=inputs_embeds,
            )
901
902
903
        return hidden_states

    def profile_run(self) -> None:
904
905
906
907
908
909
910
911
912
913
914
        # use an empty tensor instead of `None`` to force Dynamo to pass
        # it by reference, rather by specializing on the value `None`.
        # the `dtype` argument does not matter, and we use `float32` as
        # a placeholder (it has wide hardware support).
        # it is important to create tensors inside the loop, rather than
        # multiplying the list, to avoid Dynamo from treating them as
        # tensor aliasing.
        dummy_kv_caches = [
            torch.tensor([], dtype=torch.float32, device=self.device)
            for _ in range(self.num_attn_layers)
        ]
915
916

        # Profile with multimodal encoder & encoder cache.
917
918
919
        # 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):
920

921
            # NOTE: Currently model is profiled with a single non-text
922
923
            # modality with the max possible input tokens even when
            # it supports multiple.
924
            max_tokens_by_modality_dict = MULTIMODAL_REGISTRY.get_max_tokens_per_item_by_nonzero_modality(  # noqa: E501
925
926
927
928
929
                self.model_config)
            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
930
931
932
933
934
935
            # 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)
936
937
938

            # Check how many items of this modality can be supported by
            # the decoder budget.
939
940
            max_mm_items_per_req = self.mm_registry.get_mm_limits_per_prompt(
                self.model_config)[dummy_data_modality]
941
942
943
944
945
946
947
948
949
950

            # 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)

951
952
953
954
955
956
957
958
959
960
961
962
963
            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

964
965
966
967
            # Dummy data definition in V0 may contain multiple multimodal items
            # (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.
968
969

            # Case when models have a merged processor, their dummy data is
970
971
            # already batched `MultiModalKwargs`, therefore we take the first
            # `MultiModalKwargsItem` from the desired modality to profile on.
972
            if isinstance(dummy_mm_data, MultiModalKwargs):
973
974
975
                dummy_mm_item = dummy_mm_data.get_item(
                    modality=dummy_data_modality, item_index=0)
                dummy_mm_kwargs = MultiModalKwargs.from_items([dummy_mm_item])
976
977
978
979

            # Case when models have dummy data explicitly defined as
            # `MultiModalDataDict`, so they need to be processed through input
            # mapper.
980
981
            # TODO (ywang96): deprecate this path once merged processor is
            # supported on all models.
982
            else:
983
                mm_kwargs_list = self.mm_input_mapper_profiling.process_inputs(
984
                    mm_data=dummy_mm_data,
985
                    mm_hashes=None,
986
987
988
989
                    mm_processor_kwargs=None,
                    precomputed_mm_inputs=None)
                dummy_mm_kwargs = mm_kwargs_list[0]

990
            batched_dummy_mm_inputs = MultiModalKwargs.batch(
991
                [dummy_mm_kwargs] * max_num_mm_items)
992
993
994
995
996
997
998
999
1000
1001
1002
1003
1004
1005
1006
1007
            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))

1008
        # Trigger compilation for general shape.
1009
        hidden_states = self._dummy_run(self.max_num_tokens, dummy_kv_caches)
1010
1011
1012
        logits = self.model.compute_logits(hidden_states, None)
        logits = logits[:self.max_num_tokens]
        # TODO(woosuk): Consider the memory usage of the sampler.
1013
        torch.cuda.synchronize()
1014
        del hidden_states, logits
1015
        self.encoder_cache.clear()
1016
        gc.collect()
1017
1018

    def capture_model(self) -> None:
1019
1020
        if not self.use_cuda_graph:
            logger.warning(
1021
                "Skipping CUDA graph capture. Please add "
1022
                "-O %s to use CUDA graphs.", CompilationLevel.PIECEWISE)
1023
1024
1025
1026
1027
            return

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

1028
1029
1030
        # 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.
1031
        with graph_capture(device=self.device):
1032
            for num_tokens in reversed(self.cudagraph_batch_sizes):
1033
1034
                for _ in range(self.vllm_config.compilation_config.
                               cudagraph_num_of_warmups):
1035
1036
                    self._dummy_run(num_tokens)
                self._dummy_run(num_tokens)
1037
1038
1039
1040
1041
1042
1043
1044

        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))
1045

1046
1047
1048
1049
1050
1051
1052
1053
1054
1055
1056
1057
1058
1059
1060
1061
1062
1063
1064
1065
1066
1067
1068
1069
1070
1071
1072
1073
1074
    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
        """
        if len(kv_cache_config.groups) > 1:
            raise NotImplementedError(
                "Hybrid models with more than one KV cache type are not "
                "supported yet.")

        kv_caches: Dict[str, torch.Tensor] = {}

        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):
                kv_cache_shape = FlashAttentionBackend.get_kv_cache_shape(
                    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

1075
        bind_kv_cache(
1076
            kv_caches,
1077
            self.vllm_config.compilation_config.static_forward_context,
1078
1079
1080
1081
1082
1083
1084
1085
1086
1087
1088
1089
1090
1091
1092
1093
1094
1095
1096
1097
1098
1099
1100
1101
1102
1103
1104
1105
1106
1107
1108
1109
1110
1111
1112
1113
            self.kv_caches)

    def get_kv_cache_spec(self) -> KVCacheSpec:
        """
        Generates the KVCacheSpec by parsing the kv cache format from each 
        Attention module in the static forward context.
        Returns:
            KVCacheSpec: A dictionary mapping layer names to their KV cache 
            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
        kv_cache_spec: KVCacheSpec = {}
        for layer_name, attn_module in forward_ctx.items():
            # TODO: Support other attention modules, e.g., sliding window,
            # cross-attention, MLA.
            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,
                )
            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