gpu_model_runner.py 49.2 KB
Newer Older
1
import gc
2
import time
3
from typing import TYPE_CHECKING, Dict, List, Tuple, cast
4
5
6
7
8
9

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

10
11
from vllm.attention.backends.abstract import AttentionType
from vllm.attention.layer import Attention
12
from vllm.config import CompilationLevel, VllmConfig
13
from vllm.distributed.parallel_state import graph_capture
14
from vllm.forward_context import set_forward_context
15
from vllm.inputs import INPUT_REGISTRY
16
from vllm.logger import init_logger
17
from vllm.model_executor.layers.rotary_embedding import MRotaryEmbedding
18
from vllm.model_executor.model_loader import get_model
19
from vllm.multimodal import MULTIMODAL_REGISTRY, MultiModalKwargs
20
from vllm.multimodal.utils import group_mm_inputs_by_modality
21
from vllm.sampling_params import SamplingType
22
from vllm.utils import (STR_DTYPE_TO_TORCH_DTYPE, DeviceMemoryProfiler,
23
                        LayerBlockType, cdiv, is_pin_memory_available)
24
25
from vllm.v1.attention.backends.flash_attn import (FlashAttentionBackend,
                                                   FlashAttentionMetadata)
26
from vllm.v1.core.encoder_cache_manager import compute_encoder_budget
27
from vllm.v1.engine.mm_input_mapper import MMInputMapperClient
28
29
from vllm.v1.kv_cache_interface import (FullAttentionSpec, KVCacheConfig,
                                        KVCacheSpec)
30
31
from vllm.v1.outputs import ModelRunnerOutput
from vllm.v1.sample.metadata import SamplingMetadata
32
from vllm.v1.utils import bind_kv_cache
33
from vllm.v1.worker.gpu_input_batch import CachedRequestState, InputBatch
34
35
36
37
38
39
40
41
42
43
44

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

logger = init_logger(__name__)


class GPUModelRunner:

    def __init__(
        self,
45
        vllm_config: VllmConfig,
46
        device: torch.device,
47
    ):
48
49
50
51
52
53
54
55
56
57
        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
58

59
60
61
62
        model_config = self.model_config
        cache_config = self.cache_config
        scheduler_config = self.scheduler_config
        parallel_config = self.parallel_config
63
        self.device = device
64
65
66
67
68
69
70
71
        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]

72
        self.is_multimodal_model = model_config.is_multimodal_model
73
74
75
76
77
        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
78
        self.max_num_reqs = scheduler_config.max_num_seqs
79
80

        # Model-related.
81
82
        self.num_attn_layers = model_config.get_num_layers_by_block_type(
            parallel_config, LayerBlockType.attention)
83
84
        self.num_query_heads = model_config.get_num_attention_heads(
            parallel_config)
85
86
        self.num_kv_heads = model_config.get_num_kv_heads(parallel_config)
        self.head_size = model_config.get_head_size()
87
88
89
        self.hidden_size = model_config.get_hidden_size()

        # Multi-modal data support
90
91
        self.input_registry = INPUT_REGISTRY
        self.mm_registry = MULTIMODAL_REGISTRY
92

93
94
95
96
        # 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
97

98
99
100
101
102
103
        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
104
105
106
107

        # Lazy initialization
        # self.model: nn.Module  # Set after load_model
        self.kv_caches: List[torch.Tensor] = []
108
109
        # req_id -> (input_id -> encoder_output)
        self.encoder_cache: Dict[str, Dict[int, torch.Tensor]] = {}
110
111
112
113
114

        # Request states.
        self.requests: Dict[str, CachedRequestState] = {}
        # Persistent batch.
        self.input_batch = InputBatch(
115
            max_num_reqs=self.max_num_reqs,
116
117
118
119
            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,
120
            vocab_size=model_config.get_vocab_size(),
121
122
        )

123
        self.use_cuda_graph = (self.vllm_config.compilation_config.level
124
125
126
                               == CompilationLevel.PIECEWISE
                               and not self.model_config.enforce_eager)
        # TODO(woosuk): Provide an option to tune the max cudagraph batch size.
127
128
129
130
131
        # 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(
            reversed(self.vllm_config.compilation_config.capture_sizes))
132

133
134
135
136
        # Cache the device properties.
        self.device_properties = torch.cuda.get_device_properties(self.device)
        self.num_sms = self.device_properties.multi_processor_count

137
138
139
140
        # Persistent buffers for CUDA graphs.
        self.input_ids = torch.zeros(self.max_num_tokens,
                                     dtype=torch.int32,
                                     device=self.device)
141
142
143
        self.positions = torch.zeros(self.max_num_tokens,
                                     dtype=torch.int64,
                                     device=self.device)
144
145
146

        # Only relevant for models using M-RoPE (e.g, Qwen2-VL)
        if self.model_config.uses_mrope:
Roger Wang's avatar
Roger Wang committed
147
148
149
150
            # 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
151
152
153
154
155
156

            # 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
157
            self.mrope_positions = torch.zeros((3, self.max_num_tokens + 1),
158
159
                                               dtype=torch.int64,
                                               device=self.device)
Roger Wang's avatar
Roger Wang committed
160
161
162
163
164
            self.mrope_positions_cpu = torch.zeros(
                (3, self.max_num_tokens + 1),
                dtype=torch.int64,
                device="cpu",
                pin_memory=self.pin_memory)
165

166
167
168
169
        self.inputs_embeds = torch.zeros(
            (self.max_num_tokens, self.hidden_size),
            dtype=self.dtype,
            device=self.device)
170

171
        # OPTIMIZATION: Cache the tensors rather than creating them every step.
172
173
        self.arange_np = np.arange(max(self.max_num_reqs + 1,
                                       self.max_model_len),
174
175
176
177
                                   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.
178
179
180
181
182
183
184
185
186
187
188
189
190
191
192
193
194
195
196
197
        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()
198
199
200
201
202
        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()
203

204
205
206
207
208
    def _update_states(self, scheduler_output: "SchedulerOutput") -> None:
        # Remove stopped requests from the cached states.
        # Keep the states of the pre-empted requests.
        for req_id in scheduler_output.finished_req_ids:
            self.requests.pop(req_id, None)
209
210
211
212
213
214
215
216
217
            self.encoder_cache.pop(req_id, None)

        # 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)
218
219
220
221
222
223
224
225
226
227
228
229
230
231
232
233
234
235
236
237
238
239
240
241
242
243
244
245
246

        # Remove the requests from the persistent batch.
        stopped_req_ids = set().union(
            scheduler_output.preempted_req_ids,
            scheduler_output.finished_req_ids,
        )
        removed_req_indices: List[int] = []
        for req_id in stopped_req_ids:
            req_index = self.input_batch.remove_request(req_id)
            if req_index is not None:
                removed_req_indices.append(req_index)

        # Update the states of the running requests.
        for req_data in scheduler_output.scheduled_running_reqs:
            req_id = req_data.req_id
            req_state = self.requests[req_id]
            req_index = self.input_batch.req_id_to_index[req_id]

            # Update the num_computed_tokens.
            req_state.num_computed_tokens = req_data.num_computed_tokens
            self.input_batch.num_computed_tokens_cpu[req_index] = (
                req_data.num_computed_tokens)

            # Update the block table.
            num_new_blocks = len(req_data.new_block_ids)
            if num_new_blocks == 0:
                continue
            start_index = len(req_state.block_ids)
            req_state.block_ids.extend(req_data.new_block_ids)
247
248
            self.input_batch.block_table.append_row(req_index, start_index,
                                                    req_data.new_block_ids)
249
250
251

        req_ids_to_add: List[str] = []
        # Add new requests to the cached states.
252
253
254
        for new_req_data in scheduler_output.scheduled_new_reqs:
            req_id = new_req_data.req_id
            sampling_params = new_req_data.sampling_params
255
            if sampling_params.sampling_type == SamplingType.RANDOM_SEED:
256
257
258
259
260
                generator = torch.Generator(device=self.device)
                generator.manual_seed(sampling_params.seed)
            else:
                generator = None

261
262
            self.requests[req_id] = CachedRequestState(
                req_id=req_id,
263
264
265
266
                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,
267
268
                sampling_params=sampling_params,
                generator=generator,
269
270
                block_ids=new_req_data.block_ids,
                num_computed_tokens=new_req_data.num_computed_tokens,
271
272
                output_token_ids=[],
            )
273
274
275
276
277
278
279
280
281
282
283
284
285
286
287
288
289
290
291
292
293
294
295
296
297
298
299
300
301

            # 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,
                    )

302
303
304
            req_ids_to_add.append(req_id)

        # Update the cached states of the resumed requests.
305
306
        for res_req_data in scheduler_output.scheduled_resumed_reqs:
            req_id = res_req_data.req_id
307
308
            req_state = self.requests[req_id]

309
310
            req_state.block_ids = res_req_data.block_ids
            req_state.num_computed_tokens = res_req_data.num_computed_tokens
311
312
313
314
315
316
317
318
319
320
321
322
323
324
325
326
327
328
329
330
331
332
333
334
335
336
337
            req_ids_to_add.append(req_id)

        # 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)

    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.
338
        self.input_batch.block_table.commit(num_reqs)
339
340
341
342
343
344

        # 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]:
345
            assert req_id is not None
346
347
348
349
350
351
352
353
354
            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]
355
356
        req_indices = np.repeat(self.arange_np[:num_reqs],
                                num_scheduled_tokens)
357
358
359

        # Get batched arange.
        # E.g., [2, 5, 3] -> [0, 1, 0, 1, 2, 3, 4, 0, 1, 2]
360
361
        arange = np.concatenate(
            [self.arange_np[:n] for n in num_scheduled_tokens])
362
363

        # Get positions.
364
        positions_np = self.positions_np[:total_num_scheduled_tokens]
365
366
367
368
        np.add(self.input_batch.num_computed_tokens_cpu[req_indices],
               arange,
               out=positions_np)

369
370
371
372
373
        # 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)

374
375
376
377
        # 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.
378
379
        token_indices = (positions_np +
                         req_indices * self.input_batch.token_ids_cpu.shape[1])
380
381
382
383
        # 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(),
384
                           0,
385
386
                           torch.from_numpy(token_indices),
                           out=self.input_ids_cpu[:total_num_scheduled_tokens])
387
388

        # Calculate the slot mapping.
389
390
391
392
393
        # 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.
394
395
396
397
398
        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.
399
400
        block_table_cpu = self.input_batch.block_table.get_cpu_tensor()
        block_numbers = block_table_cpu.flatten()[block_table_indices].numpy()
401
402
403
404
        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])
405
406

        # Prepare the attention metadata.
407
408
409
        self.query_start_loc_np[0] = 0
        np.cumsum(num_scheduled_tokens,
                  out=self.query_start_loc_np[1:num_reqs + 1])
410

411
412
413
414
        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()
415
416
417
418

        # 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)
419
420
421
422
423
424
425
426
427
428
        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)
429
430
        query_start_loc = self.query_start_loc_cpu[:num_reqs + 1].to(
            self.device, non_blocking=True)
431
432
        seq_lens = self.seq_lens_cpu[:num_reqs].to(self.device,
                                                   non_blocking=True)
433
434
        slot_mapping = self.slot_mapping_cpu[:total_num_scheduled_tokens].to(
            self.device, non_blocking=True).long()
435
436
437
438
439
440
441
442
443
444
445
446
447
448
449
450
451
452
453
454
455
456
457
458
459
460
461
462
463
464
465
466
467
468
469
470
471
472
473
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

        # 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)
504
505
506
507
508
            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)
509
510
        else:
            cu_prefix_query_lens = None
511
512
            prefix_kv_lens = None
            suffix_kv_lens = None
513

514
        attn_metadata = FlashAttentionMetadata(
515
            num_actual_tokens=total_num_scheduled_tokens,
516
517
518
            max_query_len=max_num_scheduled_tokens,
            query_start_loc=query_start_loc,
            max_seq_len=max_seq_len,
519
            seq_lens=seq_lens,
520
521
            block_table=(
                self.input_batch.block_table.get_device_tensor()[:num_reqs]),
522
            slot_mapping=slot_mapping,
523
524
525
            use_cascade=use_cascade,
            common_prefix_len=common_prefix_len,
            cu_prefix_query_lens=cu_prefix_query_lens,
526
527
            prefix_kv_lens=prefix_kv_lens,
            suffix_kv_lens=suffix_kv_lens,
528
529
530
531
532
533
534
        )
        # NOTE(woosuk): Due to chunked prefills, there can be at most 1 partial
        # request in the batch. While we should not sample any token from this
        # partial request, we do so for simplicity. We will ignore the sampled
        # token from the partial request.
        # TODO: Support prompt logprobs.
        logits_indices = query_start_loc[1:] - 1
535
        return attn_metadata, logits_indices
536

537
538
539
540
541
542
543
544
545
546
547
548
549
550
551
552
553
554
555
556
557
558
559
560
561
562
563
564
565
566
567
568
569
570
571
572
573
574
575
576
577
578
579
580
581
582
583
584
585
586
587
588
589
590
591
    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

592
593
594
595
596
597
598
599
600
601
602
603
    def _prepare_sampling(
        self,
        scheduler_output: "SchedulerOutput",
    ) -> SamplingMetadata:
        skip_copy = True
        if (scheduler_output.finished_req_ids
                or scheduler_output.preempted_req_ids):
            skip_copy = False
        if (scheduler_output.scheduled_new_reqs
                or scheduler_output.scheduled_resumed_reqs):
            skip_copy = False
        # Create the sampling metadata.
604
605
606
607
608
609
        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(
            req_id_output_token_ids, skip_copy)
610
611
        return sampling_metadata

612
613
614
615
616
617
618
    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] = []
619
        req_input_ids: List[Tuple[str, int]] = []
620
621
622
623
624
        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))
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

        # 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)
653
654
655
656
657
658
659
660
661
662
663
664
665
666

        # 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]:
667
            assert req_id is not None
668
669
670
671
672
673
674
675
676
677
678
679
680
681
682
683
684
685
686
687
688
689
690
691
692
693
694
695
696
697
698
699
            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

700
701
702
    def get_model(self) -> nn.Module:
        return self.model

703
704
705
706
707
708
    @torch.inference_mode()
    def execute_model(
        self,
        scheduler_output: "SchedulerOutput",
    ) -> ModelRunnerOutput:
        self._update_states(scheduler_output)
709

710
711
712
713
714
715
        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 = []
716
717

        # Prepare the decoder inputs.
718
        attn_metadata, logits_indices = self._prepare_inputs(scheduler_output)
719
720
721
722
723
        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.
724
            num_input_tokens = self.vllm_config.pad_for_cudagraph(
725
726
727
728
                num_scheduled_tokens)
        else:
            # Eager mode.
            num_input_tokens = num_scheduled_tokens
729
730
        attn_metadata.num_input_tokens = num_input_tokens

731
732
733
734
735
736
737
738
739
740
741
742
743
744
        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
745
        else:
746
747
748
749
750
751
            # 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
752
753
754

        # Run the decoder.
        # Use persistent buffers for CUDA graphs.
755
        with set_forward_context(attn_metadata, self.vllm_config):
756
757
758
            positions = self.mrope_positions[:, :num_input_tokens] \
                if self.model_config.uses_mrope \
                else self.positions[:num_input_tokens]
759
            hidden_states = self.model(
760
                input_ids=input_ids,
761
                positions=positions,
762
                kv_caches=self.kv_caches,
763
                attn_metadata=None,
764
                inputs_embeds=inputs_embeds,
765
            )
766
        hidden_states = hidden_states[:num_scheduled_tokens]
767
768
769
770
771
772
773
774
775
776
        hidden_states = hidden_states[logits_indices]
        logits = self.model.compute_logits(hidden_states, None)

        # Sample the next token and get logprobs if needed.
        sampling_metadata = self._prepare_sampling(scheduler_output)
        sampler_output = self.model.sample(
            logits=logits,
            sampling_metadata=sampling_metadata,
        )

777
        sampled_token_ids = sampler_output.sampled_token_ids
778
779
780
781
        # 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
        for i, req_id in enumerate(self.input_batch.req_ids[:num_reqs]):
782
            assert req_id is not None
783
784
785
786
787
788
            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.
789
                token_id = sampled_token_ids[i]
790
                self.input_batch.token_ids_cpu[i, seq_len] = token_id
791
                self.input_batch.num_tokens[i] += 1
792
793
794
795
                req_state.output_token_ids.append(token_id)
            else:
                # Ignore the sampled token from the partial request.
                # Rewind the generator state as if the token was not sampled.
796
                generator = self.input_batch.generators.get(i)
797
                if generator is not None:
798
799
                    # This relies on cuda-specific torch-internal impl details
                    generator.set_offset(generator.get_offset() - 4)
800
801
802
803
804
805
806
807
808

        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()
809
810
811
812
813
814
815

        # 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])

816
        model_runner_output = ModelRunnerOutput(
817
            req_ids=req_ids,
818
            req_id_to_index=self.input_batch.req_id_to_index,
819
            sampled_token_ids=sampled_token_ids,
820
821
822
823
824
825
826
827
            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
828
            self.model = get_model(vllm_config=self.vllm_config)
829
830
831
832
833

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

834
835
836
837
838
839
840
    @torch.inference_mode()
    def _dummy_run(
        self,
        model: nn.Module,
        num_tokens: int,
        kv_caches: List[torch.Tensor],
    ) -> torch.Tensor:
841
842
843
844
845
846
        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
847
        with set_forward_context(None, self.vllm_config):
848
849
850
            positions = self.mrope_positions[:, :num_tokens] \
                if self.model_config.uses_mrope \
                else self.positions[:num_tokens]
851
            hidden_states = model(
852
                input_ids=input_ids,
853
                positions=positions,
854
855
                kv_caches=kv_caches,
                attn_metadata=None,
856
857
                inputs_embeds=inputs_embeds,
            )
858
859
860
        return hidden_states

    def profile_run(self) -> None:
861
862
863
864
865
866
867
868
869
870
871
        # 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)
        ]
872
873

        # Profile with multimodal encoder & encoder cache.
874
875
876
        # 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):
877

878
            # NOTE: Currently model is profiled with a single non-text
879
880
            # modality with the max possible input tokens even when
            # it supports multiple.
881
            max_tokens_by_modality_dict = MULTIMODAL_REGISTRY.get_max_tokens_per_item_by_nonzero_modality(  # noqa: E501
882
883
884
885
886
                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
887
888
889
890
891
892
            # 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)
893
894
895

            # Check how many items of this modality can be supported by
            # the decoder budget.
896
897
            max_mm_items_per_req = self.mm_registry.get_mm_limits_per_prompt(
                self.model_config)[dummy_data_modality]
898
899
900
901
902
903
904
905
906
907

            # 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)

908
909
910
911
912
913
914
915
916
917
918
919
920
            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

921
922
923
924
            # 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.
925
926

            # Case when models have a merged processor, their dummy data is
927
928
            # already batched `MultiModalKwargs`, therefore we take the first
            # `MultiModalKwargsItem` from the desired modality to profile on.
929
            if isinstance(dummy_mm_data, MultiModalKwargs):
930
931
932
                dummy_mm_item = dummy_mm_data.get_item(
                    modality=dummy_data_modality, item_index=0)
                dummy_mm_kwargs = MultiModalKwargs.from_items([dummy_mm_item])
933
934
935
936

            # Case when models have dummy data explicitly defined as
            # `MultiModalDataDict`, so they need to be processed through input
            # mapper.
937
938
            # TODO (ywang96): deprecate this path once merged processor is
            # supported on all models.
939
            else:
940
                mm_kwargs_list = self.mm_input_mapper_profiling.process_inputs(
941
                    mm_data=dummy_mm_data,
942
                    mm_hashes=None,
943
944
945
946
                    mm_processor_kwargs=None,
                    precomputed_mm_inputs=None)
                dummy_mm_kwargs = mm_kwargs_list[0]

947
            batched_dummy_mm_inputs = MultiModalKwargs.batch(
948
                [dummy_mm_kwargs] * max_num_mm_items)
949
950
951
952
953
954
955
956
957
958
959
960
961
962
963
964
            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))

965
966
967
        # Trigger compilation for general shape.
        hidden_states = self._dummy_run(self.model, self.max_num_tokens,
                                        dummy_kv_caches)
968
969
970
        logits = self.model.compute_logits(hidden_states, None)
        logits = logits[:self.max_num_tokens]
        # TODO(woosuk): Consider the memory usage of the sampler.
971
        torch.cuda.synchronize()
972
        del hidden_states, logits
973
        self.encoder_cache.clear()
974
        gc.collect()
975
976

    def capture_model(self) -> None:
977
978
        if not self.use_cuda_graph:
            logger.warning(
979
                "Skipping CUDA graph capture. Please add "
980
                "-O %s to use CUDA graphs.", CompilationLevel.PIECEWISE)
981
982
983
984
985
            return

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

986
987
988
        # 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.
989
        with graph_capture(device=self.device):
990
            for num_tokens in reversed(self.cudagraph_batch_sizes):
991
992
993
                for _ in range(self.vllm_config.compilation_config.
                               cudagraph_num_of_warmups):
                    self._dummy_run(self.model, num_tokens, self.kv_caches)
994
                self._dummy_run(self.model, num_tokens, self.kv_caches)
995
996
997
998
999
1000
1001
1002

        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))
1003

1004
1005
1006
1007
1008
1009
1010
1011
1012
1013
1014
1015
1016
1017
1018
1019
1020
1021
1022
1023
1024
1025
1026
1027
1028
1029
1030
1031
1032
    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

1033
        bind_kv_cache(
1034
            kv_caches,
1035
            self.vllm_config.compilation_config.static_forward_context,
1036
1037
1038
1039
1040
1041
1042
1043
1044
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
            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