tpu_model_runner.py 92 KB
Newer Older
1
# SPDX-License-Identifier: Apache-2.0
2
# SPDX-FileCopyrightText: Copyright contributors to the vLLM project
3
import bisect
4
import gc
5
import time
6
from typing import TYPE_CHECKING, Any, cast
7
8
9
10
11
from unittest.mock import patch

import numpy as np
import torch
import torch.nn as nn
12

13
# TPU XLA related
14
import torch_xla
15
import torch_xla.core.xla_model as xm
16
import torch_xla.distributed.spmd as xs
17
18
import torch_xla.runtime as xr

19
import vllm.envs as envs
20
from vllm.attention.backends.abstract import AttentionType
21
from vllm.attention.layer import Attention, MLAAttention
22
from vllm.attention.layers.chunked_local_attention import ChunkedLocalAttention
23
from vllm.compilation.wrapper import TorchCompileWithNoGuardsWrapper
24
25
26
27
28
29
30
from vllm.config import (
    ParallelConfig,
    VllmConfig,
    get_layers_from_vllm_config,
    update_config,
)
from vllm.distributed.kv_transfer import get_kv_transfer_group, has_kv_transfer_group
31
from vllm.distributed.kv_transfer.kv_connector.utils import copy_kv_blocks
32
from vllm.forward_context import set_forward_context
33
from vllm.logger import init_logger
34
from vllm.lora.layers import BaseLayerWithLoRA
35
from vllm.model_executor.layers.attention_layer_base import AttentionLayerBase
36
from vllm.model_executor.model_loader import get_model_loader
37
from vllm.model_executor.model_loader.tpu import TPUModelLoader
38
39
40
41
from vllm.model_executor.models.interfaces import (
    SupportsMultiModal,
    supports_transcription,
)
42
from vllm.model_executor.models.interfaces_base import (
43
44
45
    is_pooling_model,
    is_text_generation_model,
)
46
from vllm.multimodal import MULTIMODAL_REGISTRY
47
48
49
50
51
from vllm.multimodal.inputs import (
    BatchedTensorInputs,
    MultiModalKwargsItem,
    PlaceholderRange,
)
52
from vllm.multimodal.utils import group_mm_kwargs_by_modality
53
from vllm.sequence import IntermediateTensors
54
from vllm.tasks import GenerationTask, PoolingTask, SupportedTask
55
from vllm.utils.math_utils import cdiv, prev_power_of_2
56
from vllm.utils.platform_utils import is_pin_memory_available
57
58
59
60
61
62
63
64
65
66
67
from vllm.v1.attention.backends.pallas import (
    TPU_STR_DTYPE_TO_TORCH_DTYPE,
    PallasAttentionBackend,
    PallasMetadata,
    get_page_size_bytes,
)
from vllm.v1.kv_cache_interface import (
    AttentionSpec,
    FullAttentionSpec,
    KVCacheConfig,
    KVCacheSpec,
68
    MLAAttentionSpec,
69
70
71
72
73
74
75
76
    SlidingWindowSpec,
)
from vllm.v1.outputs import (
    EMPTY_MODEL_RUNNER_OUTPUT,
    LogprobsLists,
    LogprobsTensors,
    ModelRunnerOutput,
)
77
78
from vllm.v1.sample.tpu.metadata import TPUSupportedSamplingMetadata
from vllm.v1.sample.tpu.sampler import Sampler as TPUSampler
79
from vllm.v1.worker.kv_connector_model_runner_mixin import (
80
81
82
    KVConnectorModelRunnerMixin,
    KVConnectorOutput,
)
83
from vllm.v1.worker.lora_model_runner_mixin import LoRAModelRunnerMixin
84
from vllm.v1.worker.tpu_input_batch import CachedRequestState, InputBatch
85

86
87
88
89
90
91
from .utils import (
    MultiModalBudget,
    add_kv_sharing_layers_to_kv_cache_groups,
    bind_kv_cache,
    sanity_check_mm_encoder_outputs,
)
92

93
if TYPE_CHECKING:
94
    from vllm.v1.core.sched.output import GrammarOutput, SchedulerOutput
95
96
97

logger = init_logger(__name__)

98
INVALID_TOKEN_ID = -1
99
100
# Smallest output size
MIN_NUM_SEQS = 8
101
102


103
104
105
106
107
108
109
110
111
112
113
114
115
116
117
118
119
120
121
122
123
124
125
126
127
128
129
130
131
132
133
134
135
136
137
#########################################################
# Ways to avoid recompilation
#########################################################
#
# The model executor has two primary components:
# 1. preparing the model and sampler inputs
# 2. executing the model and sampler.
# The core idea is to avoid any TPU computation during input preparation. For
# better compilation tracking and increased flexibility, the model execution and
# sampler are divided into several distinct components.
#
# Below are the detailed steps:
#
# Step 1
# It is recommended to avoid TPU operations when preparing the model and sampler
# inputs. CPU tensors can be prepared and transferred to the XLA device using
# cpu_tensor.to(xla_device), which only triggers CPU to TPU transfers and avoids
# compilation.
#
# Step 2
# The TPU execution should be decomposed into subgraphs (4 at the moment):
# 1. the main model
# 2. selecting hidden states for each request
# 3. sampler
# 4. encoder.
# Each subgraph should be decorated in a torch.compile. This is used to make
# sure that we have the same subgraph topology in both dummy_run and
# xecute_model. The results from these subgraphs should either be passed to
# other subgraphs, or transferred from TPU to CPU using xla_tensor.cpu() for
# subsequent processing on the CPU.
#
# Step 3
# The dummy_run should be comprehensive, ensuring all potential input shapes and
# branch predictions are included as subgraph inputs to facilitate
# pre-compilation.
138
class TPUModelRunner(LoRAModelRunnerMixin, KVConnectorModelRunnerMixin):
139
140
141
142
    def __init__(
        self,
        vllm_config: VllmConfig,
        device: torch.device,
143
        original_parallel_config: ParallelConfig | None = None,
144
145
146
147
148
149
150
    ):
        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
151
        self.original_parallel_config = original_parallel_config
152
153
154
155
156
157
158
159
160
161
        self.scheduler_config = vllm_config.scheduler_config
        self.speculative_config = vllm_config.speculative_config
        self.observability_config = vllm_config.observability_config
        self.device_config = vllm_config.device_config

        model_config = self.model_config
        cache_config = self.cache_config
        scheduler_config = self.scheduler_config
        parallel_config = self.parallel_config
        self.device = device
162
        self.check_recompilation = envs.VLLM_XLA_CHECK_RECOMPILATION
163

164
165
166
167
168
169
        # SPMD Related
        self.use_spmd = envs.VLLM_XLA_USE_SPMD
        if self.use_spmd:
            num_devices = xr.global_runtime_device_count()
            mesh_shape = (num_devices, 1)
            device_ids = np.array(range(num_devices))
170
            self.mesh = xs.Mesh(device_ids, mesh_shape, ("x", "y"))
171

172
        self.enforce_eager = model_config.enforce_eager
173
174
175
176

        self.num_xla_graphs = 0
        self._update_num_xla_graphs("init")

177
178
        self.pin_memory = is_pin_memory_available()
        self.dtype = self.model_config.dtype
179
        if cache_config.cache_dtype == "auto":
180
181
            model_dtype = self.dtype
            if isinstance(model_dtype, str):
182
                self.kv_cache_dtype = TPU_STR_DTYPE_TO_TORCH_DTYPE[model_dtype]
183
184
            else:
                self.kv_cache_dtype = model_dtype
185
        else:
186
            self.kv_cache_dtype = TPU_STR_DTYPE_TO_TORCH_DTYPE[cache_config.cache_dtype]
187
        self._hidden_states_dtype = self.dtype
188
189
190
191

        self.sliding_window = model_config.get_sliding_window()
        self.block_size = cache_config.block_size
        self.max_model_len = model_config.max_model_len
192
        self.most_model_len = envs.VLLM_TPU_MOST_MODEL_LEN
193
        self.max_num_blocks_per_req = cdiv(self.max_model_len, self.block_size)
194
195
196
197
198
        self.num_blocks_per_most_len_req = (
            cdiv(self.most_model_len, self.block_size)
            if self.most_model_len is not None
            else None
        )
199
200
        # InputBatch needs to work with sampling tensors greater than padding
        # to avoid dynamic shapes. Also, avoid suboptimal alignment.
201
        self.max_num_reqs = max(scheduler_config.max_num_seqs, MIN_NUM_SEQS)
202
203
204
        self.num_tokens_paddings = _get_token_paddings(
            min_token_size=16,
            max_token_size=scheduler_config.max_num_batched_tokens,
205
206
            padding_gap=envs.VLLM_TPU_BUCKET_PADDING_GAP,
        )
207
208
209
        # In case `max_num_tokens < max(num_tokens_paddings)` use the actual
        # padded max value to pre-allocate data structures and pre-compile.
        self.max_num_tokens = self.num_tokens_paddings[-1]
210
211
212

        # Model-related.
        self.num_attn_layers = model_config.get_num_layers_by_block_type(
213
            parallel_config, "attention"
214
215
        )
        self.num_query_heads = model_config.get_num_attention_heads(parallel_config)
216
217
        self.num_kv_heads = model_config.get_num_kv_heads(parallel_config)
        self.head_size = model_config.get_head_size()
218
        self.inputs_embeds_size = model_config.get_inputs_embeds_size()
219
        self.vocab_size = model_config.get_vocab_size()
220

221
222
223
        # Multi-modal data support
        self.mm_registry = MULTIMODAL_REGISTRY
        self.uses_mrope = model_config.uses_mrope
224
        self.supports_mm_inputs = self.mm_registry.supports_multimodal_inputs(
225
            model_config
226
        )
227
228
229
        # TODO: Support M-RoPE (e.g, Qwen2-VL)
        assert not self.uses_mrope, "TPU does not support M-RoPE yet."

230
231
232
233
234
235
236
237
238
239
        self._num_slices_per_kv_cache_update_block = (
            _get_num_slices_per_kv_cache_update_block(
                get_page_size_bytes(
                    block_size=self.block_size,
                    num_kv_heads=self.num_kv_heads,
                    head_size=self.head_size,
                    kv_cache_dtype=self.kv_cache_dtype,
                )
            )
        )
240

241
        # Lazy initialization
242
        self.model: nn.Module  # Set after load_model
243
        self.kv_caches: list[torch.Tensor] = []
244
245
        # mm_hash -> encoder_output
        self.encoder_cache: dict[str, torch.Tensor] = {}
246
247
248

        # Request states.
        self.requests: dict[str, CachedRequestState] = {}
249
250
251
        # NOTE(rob): num_prompt_logprobs only includes reqs
        # that are currently in the prefill phase.
        self.num_prompt_logprobs: dict[str, int] = {}
252

253
254
255
256
257
258
259
260
        # Initialize input batch early to avoid AttributeError in _update_states
        self.input_batch = InputBatch(
            max_num_reqs=self.max_num_reqs,
            max_model_len=self.max_model_len,
            max_num_batched_tokens=self.max_num_tokens,
            device=self.device,
            pin_memory=self.pin_memory,
            vocab_size=self.model_config.get_vocab_size(),
261
            block_sizes=[self.block_size],
262
            kernel_block_sizes=[self.cache_config.block_size],
263
264
        )

265
266
267
        # Cached torch/numpy tensor
        # The pytorch tensor and numpy array share the same buffer.
        # Sometimes the numpy op is faster so we create both.
268
269
270
        self.input_ids_cpu = torch.zeros(
            self.max_num_tokens, dtype=torch.int32, device="cpu"
        )
271

272
273
274
        self.positions_cpu = torch.zeros(
            self.max_num_tokens, dtype=torch.int32, device="cpu"
        )
275
276
        self.positions_np = self.positions_cpu.numpy()
        self.block_table_cpu = torch.zeros(
277
            (self.max_num_reqs, self.max_num_blocks_per_req),
278
            dtype=torch.int32,
279
280
            device="cpu",
        )
281
        # adjust num_reqs to avoid SMEM OOM.
282
283
284
285
286
287
288
289
290
291
        self.num_reqs_most_model_len = (
            min(
                PallasAttentionBackend.get_max_num_seqs(
                    self.most_model_len, self.block_size
                ),
                self.max_num_reqs,
            )
            if self.most_model_len is not None
            else None
        )
292
        self.num_reqs_max_model_len = min(
293
294
295
296
297
298
299
300
301
302
303
            PallasAttentionBackend.get_max_num_seqs(
                self.max_model_len, self.block_size
            ),
            self.max_num_reqs,
        )
        self.query_start_loc_cpu = torch.zeros(
            self.max_num_tokens + 1,
            dtype=torch.int32,
            device="cpu",
            pin_memory=self.pin_memory,
        )
304
305
        self.query_start_loc_np = self.query_start_loc_cpu.numpy()

306
307
308
309
310
311
        self.seq_lens_cpu = torch.zeros(
            self.max_num_tokens,
            dtype=torch.int32,
            device="cpu",
            pin_memory=self.pin_memory,
        )
312
        self.seq_lens_np = self.seq_lens_cpu.numpy()
313

314
315
        # Only relevant for multimodal models
        if self.supports_mm_inputs:
316
317
318
319
320
321
            self.is_mm_embed_cpu = torch.zeros(
                self.max_num_tokens,
                dtype=torch.bool,
                device="cpu",
                pin_memory=self.pin_memory,
            )
322

323
324
        # Range tensor with values [0 .. self.max_num_tokens - 1].
        # Used to initialize positions / context_lens / seq_lens
325
326
        # Keep in int64 to avoid overflow with long context
        self.arange_np = np.arange(self.max_num_tokens, dtype=np.int64)
327
        self.num_reqs_paddings = _get_req_paddings(
328
329
            min_req_size=MIN_NUM_SEQS, max_req_size=self.max_num_reqs
        )
330

331
332
333
334
335
336
        # Layer pairings for cross-layer KV sharing.
        # If an Attention layer `layer_name` is in the keys of this dict, it
        # means this layer will perform attention using the keys and values
        # from the KV cache of `shared_kv_cache_layers[layer_name]`.
        self.shared_kv_cache_layers: dict[str, str] = {}

337
338
339
340
341
        # tensors for structured decoding
        self.grammar_bitmask_cpu = torch.zeros(
            (self.max_num_reqs, cdiv(self.vocab_size, 32)),
            dtype=torch.int32,
            device="cpu",
342
343
            pin_memory=self.pin_memory,
        )
344
345
346
347
        self.require_structured_out_cpu = torch.zeros(
            (self.max_num_reqs, 1),
            dtype=torch.bool,
            device="cpu",
348
349
            pin_memory=self.pin_memory,
        )
350
        self.structured_decode_arange = torch.arange(
351
352
            0, 32, device="cpu", pin_memory=self.pin_memory
        )
353

354
355
        self.mm_budget = (
            MultiModalBudget(
356
                self.model_config,
357
358
359
360
361
362
                self.scheduler_config,
                self.mm_registry,
            )
            if self.supports_mm_inputs
            else None
        )
363

364
365
366
367
368
        if not self.use_spmd:
            self.sample_from_logits_func = torch.compile(
                self.sample_from_logits,
                backend="openxla",
                fullgraph=True,
369
370
                dynamic=False,
            )
371
372
373
        else:
            self.sample_from_logits_func = self.sample_from_logits

374
375
376
377
378
        # For passing scheduler_output between successive
        # execute_model() and sample_tokens() calls.
        self.scheduler_output: SchedulerOutput | None = None
        self.mm_embed_inputs: tuple[list[torch.Tensor], torch.Tensor] | None = None

379
380
381
382
    def reset_mm_cache(self) -> None:
        if self.mm_budget:
            self.mm_budget.reset_cache()

383
384
385
386
387
388
389
390
391
392
    def _update_num_xla_graphs(self, case_str):
        check_comp = self.check_recompilation and not self.enforce_eager
        if not check_comp:
            return

        total_cached_graphs = xr.get_num_cached_compilation_graph()
        new_compiled_graphs = total_cached_graphs - self.num_xla_graphs
        if new_compiled_graphs == 0:
            return

393
394
395
        logger.info(
            "Add new %d compiled XLA graphs due to %s", new_compiled_graphs, case_str
        )
396
397
398
399
400
401
402
403
404
405
406
        self.num_xla_graphs += new_compiled_graphs

    def _verify_num_xla_graphs(self, case_str):
        check_comp = self.check_recompilation and not self.enforce_eager
        if not check_comp:
            return

        curr_cached_graph = xr.get_num_cached_compilation_graph()
        assert self.num_xla_graphs == curr_cached_graph, (
            "Recompilation after warm up is detected during {}."
            " num_xla_graphs = {} curr_cached_graph = {}".format(
407
408
409
                case_str, self.num_xla_graphs, curr_cached_graph
            )
        )
410

411
412
413
414
415
416
417
418
    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:
419
            True if there is a new/resumed/paused/finished request.
420
421
422
423
424
            If False, we can skip copying SamplingMetadata to the GPU.
        """
        # Remove finished requests from the cached states.
        for req_id in scheduler_output.finished_req_ids:
            self.requests.pop(req_id, None)
425
            self.num_prompt_logprobs.pop(req_id, None)
426
427
428
429
430
431
432

        # 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.
433
        removed_req_indices: list[int] = []
434
435
436
437
438
        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)

439
        # Free the cached encoder outputs.
440
441
        for mm_hash in scheduler_output.free_encoder_mm_hashes:
            self.encoder_cache.pop(mm_hash, None)
442

443
444
445
446
447
448
449
450
451
452
453
454
455
456
457
458
459
        # 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:
            req_index = self.input_batch.remove_request(req_id)
            assert req_index is not None
            removed_req_indices.append(req_index)

460
        req_ids_to_add: list[str] = []
461
462
        # Add new requests to the cached states.
        for new_req_data in scheduler_output.scheduled_new_reqs:
463
            assert new_req_data.sampling_params is not None, (
464
                "Pooling is not supported in TPU yet"
465
            )
466
467
468
469
470
471
            req_id = new_req_data.req_id
            sampling_params = new_req_data.sampling_params

            self.requests[req_id] = CachedRequestState(
                req_id=req_id,
                prompt_token_ids=new_req_data.prompt_token_ids,
472
                prompt_embeds=new_req_data.prompt_embeds,
473
                mm_features=new_req_data.mm_features,
474
                sampling_params=sampling_params,
475
                pooling_params=None,
476
                generator=None,
477
478
479
480
481
482
                block_ids=new_req_data.block_ids,
                num_computed_tokens=new_req_data.num_computed_tokens,
                output_token_ids=[],
                lora_request=new_req_data.lora_request,
            )

483
484
485
486
487
488
489
            if sampling_params and sampling_params.prompt_logprobs is not None:
                self.num_prompt_logprobs[req_id] = (
                    self.input_batch.vocab_size
                    if sampling_params.prompt_logprobs == -1
                    else sampling_params.prompt_logprobs
                )

490
491
492
            req_ids_to_add.append(req_id)

        # Update the states of the running/resumed requests.
493
494
        req_data = scheduler_output.scheduled_cached_reqs
        for i, req_id in enumerate(req_data.req_ids):
495
            req_state = self.requests[req_id]
496
497
            num_computed_tokens = req_data.num_computed_tokens[i]
            new_block_ids = req_data.new_block_ids[i]
498
            resumed_from_preemption = req_id in req_data.resumed_req_ids
499
500

            # Update the cached states.
501
502
            req_state.num_computed_tokens = num_computed_tokens
            if not resumed_from_preemption:
503
504
                if new_block_ids is not None:
                    # Append the new blocks to the existing block IDs.
505
                    for block_ids, new_ids in zip(req_state.block_ids, new_block_ids):
506
                        block_ids.extend(new_ids)
507
            else:
508
                assert new_block_ids is not None
509
510
                # The request is resumed from preemption.
                # Replace the existing block IDs with the new ones.
511
                req_state.block_ids = new_block_ids
512
513
514
515
516
517
518
519
520
521

            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.
522
            self.input_batch.num_computed_tokens_cpu[req_index] = num_computed_tokens
523
            if new_block_ids is not None:
524
                self.input_batch.block_table.append_row(new_block_ids, req_index)
525
526
527
528
529
530

        # 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]
531
532
            # Fill the empty index or append to the end
            req_index = removed_req_indices.pop() if removed_req_indices else None
533
534
535
536
537
            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)
538

539
540
541
542
543
        return len(unscheduled_req_ids) > 0 or len(req_ids_to_add) > 0

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

544
545
546
547
548
549
550
551
552
553
554
555
556
557
558
    def get_supported_generation_tasks(self) -> list[GenerationTask]:
        model = self.get_model()
        supported_tasks = list[GenerationTask]()

        if is_text_generation_model(model):
            supported_tasks.append("generate")

        if supports_transcription(model):
            if model.supports_transcription_only:
                return ["transcription"]

            supported_tasks.append("transcription")

        return supported_tasks

559
560
561
562
563
    def get_supported_pooling_tasks(self) -> list[PoolingTask]:
        model = self.get_model()
        if not is_pooling_model(model):
            return []

564
        return list(model.pooler.get_supported_tasks())
565

566
567
568
569
570
571
572
573
574
575
    def get_supported_tasks(self) -> tuple[SupportedTask, ...]:
        tasks = list[SupportedTask]()

        if self.model_config.runner_type == "generate":
            tasks.extend(self.get_supported_generation_tasks())
        if self.model_config.runner_type == "pooling":
            tasks.extend(self.get_supported_pooling_tasks())

        return tuple(tasks)

576
    def get_kv_cache_spec(self) -> dict[str, KVCacheSpec]:
577
        """
578
        Generates the KVCacheSpec by parsing the kv cache format from each
579
580
        Attention module in the static forward context.
        Returns:
581
            KVCacheSpec: A dictionary mapping layer names to their KV cache
582
583
584
            format. Layers that do not need KV cache are not included.
        """

585
586
587
588
        layers = get_layers_from_vllm_config(
            self.vllm_config,
            AttentionLayerBase,  # type: ignore[type-abstract]
        )
589
        block_size = self.vllm_config.cache_config.block_size
590
591
        cache_dtype_str = self.vllm_config.cache_config.cache_dtype

592
        kv_cache_spec: dict[str, KVCacheSpec] = {}
593
        for layer_name, attn_module in layers.items():
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
631
632
633
634
635
636
637
            # Classic Attention path
            if isinstance(attn_module, Attention):
                if (
                    kv_tgt_layer := attn_module.kv_sharing_target_layer_name
                ) is not None:
                    # The layer doesn't need its own KV cache and will use that of
                    # the target layer. We skip creating a KVCacheSpec for it, so
                    # that KV cache management logic will act as this layer does
                    # not exist, and doesn't allocate KV cache for the layer. This
                    # enables the memory saving of cross-layer kv sharing, allowing
                    # a given amount of memory to accommodate longer context lengths
                    # or enable more requests to be processed simultaneously.
                    self.shared_kv_cache_layers[layer_name] = kv_tgt_layer
                    continue

                if attn_module.attn_type == AttentionType.DECODER:
                    if isinstance(attn_module, ChunkedLocalAttention):
                        logger.warning_once(
                            "Using irope in Pallas is not supported yet, it "
                            "will fall back to global attention for long context."
                        )
                    if attn_module.sliding_window is not None:
                        kv_cache_spec[layer_name] = SlidingWindowSpec(
                            block_size=block_size,
                            num_kv_heads=attn_module.num_kv_heads,
                            head_size=attn_module.head_size,
                            dtype=self.kv_cache_dtype,
                            sliding_window=attn_module.sliding_window,
                        )
                    else:
                        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=self.kv_cache_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
638
                else:
639
640
641
642
643
644
645
646
647
648
649
650
                    raise ValueError(f"Unknown attention type: {attn_module.attn_type}")
            # MLAAttention path
            elif isinstance(attn_module, MLAAttention):
                if layer_name in kv_cache_spec:
                    continue
                kv_cache_spec[layer_name] = MLAAttentionSpec(
                    block_size=block_size,
                    num_kv_heads=1,
                    head_size=attn_module.head_size,
                    dtype=self.kv_cache_dtype,
                    cache_dtype_str=cache_dtype_str,
                )
651
            else:
652
                continue
653
654
655

        return kv_cache_spec

656
657
658
    def _get_slot_mapping_metadata(
        self, num_reqs, num_scheduled_tokens_per_req
    ) -> np.ndarray:
659
660
661
662
663
664
665
666
667
668
669
670
        """
        Computes metadata for mapping slots to blocks in the key-value (KV)
        cache for a batch of requests.

        This function determines, for each request in the batch, how the
        scheduled tokens are distributed across memory blocks, and generates
        metadata needed to map slices of tokens to their corresponding positions
        in the KV cache.

        Args:
            num_reqs (int): Number of requests in the current batch.
            num_scheduled_tokens_per_req (int or np.ndarray): Number of tokens
671
                to be scheduled for each request.
672
673
674

        Returns:
            np.ndarray: A 2D array of shape (total_block_len, 3), where each row
675
                contains:
676
                - kv_cache_start_index (int): The starting index in the KV cache
677
                  for the corresponding slice.
678
                - new_kv_start_index (int): The starting index in the new KV
679
                  cache for the corresponding slice.
680
681
682
                - slice_len (int): The length of the slice.
        """
        slices_start = self.input_batch.num_computed_tokens_cpu[:num_reqs]
683
684
685
686
        slices_end = (
            self.input_batch.num_computed_tokens_cpu[:num_reqs]
            + num_scheduled_tokens_per_req
        )
687
688
689
690
        local_block_start_idx = slices_start // self.block_size
        local_block_end_idx = (slices_end - 1) // self.block_size
        no_repeat_req_indices = self.arange_np[:num_reqs]
        global_block_start_idx = (
691
692
            no_repeat_req_indices * self.max_num_blocks_per_req + local_block_start_idx
        )
693
694
695
696
697
698
699
        block_lens = local_block_end_idx - local_block_start_idx + 1
        global_block_start_idx = np.repeat(global_block_start_idx, block_lens)
        slice_arange = np.concatenate([self.arange_np[:n] for n in block_lens])
        global_block_indices = global_block_start_idx + slice_arange
        block_table_cpu = self.input_batch.block_table[0].get_cpu_tensor()
        block_numbers = block_table_cpu.flatten()[global_block_indices].numpy()
        total_block_len = np.sum(block_lens)
700
701
702
        slot_mapping_slices = np.repeat(
            np.array([[0, self.block_size]], dtype=np.int32), total_block_len, axis=0
        )
703
704
705
        cu_block_lens = np.zeros(len(block_lens) + 1, dtype=np.int32)
        np.cumsum(block_lens, out=cu_block_lens[1:])
        for req_idx in range(num_reqs):
706
707
708
709
710
711
            slot_mapping_slices[cu_block_lens[req_idx]][0] = (
                slices_start[req_idx] % self.block_size
            )
            slot_mapping_slices[cu_block_lens[req_idx + 1] - 1][1] = (
                slices_end[req_idx] - 1
            ) % self.block_size + 1
712
713
714
        slice_lens = slot_mapping_slices[:, 1] - slot_mapping_slices[:, 0]
        cu_slices_lens = np.zeros(len(slice_lens) + 1, dtype=np.int32)
        np.cumsum(slice_lens, out=cu_slices_lens[1:])
715
716
717
        kv_cache_start_indices = slot_mapping_slices[:, 0] + (
            block_numbers * self.block_size
        )
718
719
        new_kv_start_indices = cu_slices_lens[:-1]
        slot_mapping_metadata = np.stack(
720
721
            [kv_cache_start_indices, new_kv_start_indices, slice_lens], axis=1
        )
722
723
        return slot_mapping_metadata

724
    def _prepare_inputs(self, scheduler_output: "SchedulerOutput", start_index: int):
725
        assert scheduler_output.total_num_scheduled_tokens > 0
726
727
        num_reqs = self.input_batch.num_reqs
        assert num_reqs > 0
728
        assert start_index < num_reqs
729

730
        # Get the number of scheduled tokens for each request.
731
        use_max_model_len = self.most_model_len is None
732
733
        num_scheduled_tokens_per_req = []
        max_num_scheduled_tokens_all_reqs = 0
734
735
736
737
738
        end_index = start_index

        # Use either most_model_len or max_model_len depending on request size.
        for i in range(start_index, num_reqs):
            req_id = self.input_batch.req_ids[i]
739
            assert req_id is not None
740
            num_tokens = scheduler_output.num_scheduled_tokens[req_id]
741
742
743
744
745
            if (
                not use_max_model_len
                and self.most_model_len is not None
                and num_tokens > self.most_model_len
            ):
746
                use_max_model_len = True
747
            num_scheduled_tokens_per_req.append(num_tokens)
748
749
        if use_max_model_len:
            if len(num_scheduled_tokens_per_req) > self.num_reqs_max_model_len:
750
751
752
                num_scheduled_tokens_per_req = num_scheduled_tokens_per_req[
                    : self.num_reqs_max_model_len
                ]
753
754
755
756
                end_index = start_index + self.num_reqs_max_model_len
            else:
                end_index = num_reqs
        else:
757
            assert self.num_reqs_most_model_len is not None
758
759
760
761
            if len(num_scheduled_tokens_per_req) > self.num_reqs_most_model_len:
                num_scheduled_tokens_per_req = num_scheduled_tokens_per_req[
                    : self.num_reqs_most_model_len
                ]
762
763
764
765
                end_index = start_index + self.num_reqs_most_model_len
            else:
                end_index = num_reqs
        max_num_scheduled_tokens_all_reqs = max(num_scheduled_tokens_per_req)
766
767
768
        num_scheduled_tokens_per_req = np.array(
            num_scheduled_tokens_per_req, dtype=np.int32
        )
769
        total_num_scheduled_tokens = sum(num_scheduled_tokens_per_req)
770
771
        assert max_num_scheduled_tokens_all_reqs > 0

772
773
        num_reqs = len(num_scheduled_tokens_per_req)

774
775
776
        # Get request indices.
        # E.g., [2, 5, 3] -> [0, 0, 1, 1, 1, 1, 1, 2, 2, 2]
        # For each scheduled token, what are the corresponding req index.
777
        req_indices = np.repeat(self.arange_np[:num_reqs], num_scheduled_tokens_per_req)
778
779
780
781
782

        # Get batched arange.
        # E.g., [2, 5, 3] -> [0, 1, 0, 1, 2, 3, 4, 0, 1, 2]
        # For each scheduled token, what is its position in corresponding req.
        arange = np.concatenate(
783
784
            [self.arange_np[:n] for n in num_scheduled_tokens_per_req]
        )
785
786
787

        # Get positions.
        positions_np = self.positions_np[:total_num_scheduled_tokens]
788
789
790
791
792
        np.add(
            self.input_batch.num_computed_tokens_cpu[req_indices],
            arange,
            out=positions_np,
        )
793
794
795
796
797

        # 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.
798
799
800
        token_indices = (
            positions_np + req_indices * self.input_batch.token_ids_cpu.shape[1]
        )
801
802
803
804

        # 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.
805
806
807
808
809
810
        torch.index_select(
            self.input_batch.token_ids_cpu_tensor.flatten(),
            0,
            torch.from_numpy(token_indices),
            out=self.input_ids_cpu[:total_num_scheduled_tokens],
        )
811
812
813

        # Prepare the attention metadata.
        self.query_start_loc_np[0] = 0
814
815
816
817
        np.cumsum(
            num_scheduled_tokens_per_req, out=self.query_start_loc_np[1 : num_reqs + 1]
        )
        self.query_start_loc_np[num_reqs + 1 :] = 1
818
819

        self.seq_lens_np[:num_reqs] = (
820
821
822
            self.input_batch.num_computed_tokens_cpu[:num_reqs]
            + num_scheduled_tokens_per_req
        )
823
824

        # Do the padding and copy the tensors to the TPU.
825
        padded_total_num_scheduled_tokens = _get_padded_token_len(
826
827
            self.num_tokens_paddings, total_num_scheduled_tokens
        )
828
829
        # Zero out to avoid spurious values from prev iteration (last cp chunk)
        self.input_ids_cpu[
830
831
832
833
834
835
836
837
            total_num_scheduled_tokens:padded_total_num_scheduled_tokens
        ] = 0
        self.input_ids = self.input_ids_cpu[:padded_total_num_scheduled_tokens].to(
            self.device
        )
        self.position_ids = self.positions_cpu[:padded_total_num_scheduled_tokens].to(
            self.device
        )
838
        if use_max_model_len:
839
840
841
842
843
844
845
846
847
848
            block_tables = self.block_table_cpu[
                : self.num_reqs_max_model_len, : self.max_num_blocks_per_req
            ]
            block_tables[:num_reqs, : self.max_num_blocks_per_req] = (
                self.input_batch.block_table[0].get_cpu_tensor()[:num_reqs]
            )
            query_start_loc = self.query_start_loc_cpu[
                : self.num_reqs_max_model_len + 1
            ].to(self.device)
            seq_lens = self.seq_lens_cpu[: self.num_reqs_max_model_len].to(self.device)
849
        else:
850
            assert self.num_reqs_most_model_len is not None
851
852
853
854
855
856
857
858
859
860
861
862
            block_tables = self.block_table_cpu[
                : self.num_reqs_most_model_len, : self.num_blocks_per_most_len_req
            ]
            block_tables[:num_reqs, : self.num_blocks_per_most_len_req] = (
                self.input_batch.block_table[0].get_cpu_tensor()[
                    :num_reqs, : self.num_blocks_per_most_len_req
                ]
            )
            query_start_loc = self.query_start_loc_cpu[
                : self.num_reqs_most_model_len + 1
            ].to(self.device)
            seq_lens = self.seq_lens_cpu[: self.num_reqs_most_model_len].to(self.device)
863
        block_tables = block_tables.to(self.device)
864

865
        # Calculate the slot mapping
866
        slot_mapping_metadata = self._get_slot_mapping_metadata(
867
868
            num_reqs, num_scheduled_tokens_per_req
        )
869
        num_kv_update_slices = slot_mapping_metadata.shape[0]
870
        padded_num_slices = _get_padded_num_kv_cache_update_slices(
871
872
            padded_total_num_scheduled_tokens, self.max_num_reqs, self.block_size
        )
873
874
875
        slot_mapping_metadata = np.pad(
            slot_mapping_metadata,
            [[0, padded_num_slices - len(slot_mapping_metadata)], [0, 0]],
876
877
            constant_values=0,
        )
878
        slot_mapping_metadata = np.transpose(slot_mapping_metadata)
879
        slot_mapping_metadata = torch.tensor(slot_mapping_metadata, device=self.device)
880

881
882
883
884
885
        if self.lora_config is not None:
            # We need to respect padding when activating LoRA adapters
            padded_num_scheduled_tokens_per_req = np.copy(
                num_scheduled_tokens_per_req
            )  # Copying to avoid accidental state corruption bugs
886
            padded_num_scheduled_tokens_per_req[-1] += (
887
                padded_total_num_scheduled_tokens - total_num_scheduled_tokens
888
            )
889

890
            self.set_active_loras(self.input_batch, padded_num_scheduled_tokens_per_req)
891

892
        attn_metadata = PallasMetadata(
893
            slot_mapping=slot_mapping_metadata,
894
            block_tables=block_tables,
895
896
            context_lens=seq_lens,
            query_start_loc=query_start_loc,
897
898
899
900
901
            num_seqs=torch.tensor([num_reqs], dtype=torch.int32, device=self.device),
            num_kv_update_slices=torch.tensor(
                [num_kv_update_slices], dtype=torch.int32, device=self.device
            ),
            num_slices_per_kv_cache_update_block=self._num_slices_per_kv_cache_update_block,
902
        )
903
904
905
906
907
        # 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.
908
        padded_num_reqs = _get_padded_num_reqs_with_upper_limit(
909
910
            num_reqs, self.max_num_reqs
        )
911
912
        # Indices at which we sample (positions of last token in the sequence).
        # Padded to avoid recompiling when `num_reqs` varies.
913
        logits_indices = self.query_start_loc_cpu[1 : padded_num_reqs + 1] - 1
914
        logits_indices = logits_indices.to(self.device)
915

916
917
918
919
920
        if self.lora_config is not None:
            # We need to respect padding when activating LoRA adapters
            padded_num_scheduled_tokens_per_req = np.copy(
                num_scheduled_tokens_per_req
            )  # Copying to avoid accidental state corruption bugs
921
            padded_num_scheduled_tokens_per_req[-1] += (
922
                padded_total_num_scheduled_tokens - total_num_scheduled_tokens
923
            )
924

925
            self.set_active_loras(self.input_batch, padded_num_scheduled_tokens_per_req)
926

927
        layer_names = get_layers_from_vllm_config(self.vllm_config, Attention).keys()
928
        per_layer_attn_metadata = {
929
            layer_name: attn_metadata for layer_name in layer_names
930
        }
931
932
933
934
935
936
937
        return (
            per_layer_attn_metadata,
            logits_indices,
            padded_num_reqs,
            num_reqs,
            end_index,
        )
938

939
    def _execute_mm_encoder(self, scheduler_output: "SchedulerOutput"):
940
941
942
943
944
        scheduled_encoder_inputs = scheduler_output.scheduled_encoder_inputs
        if not scheduled_encoder_inputs:
            return

        # Batch the multi-modal inputs.
945
        mm_kwargs = list[MultiModalKwargsItem]()
946
947
        # List of tuple (mm_hash, pos_info)
        mm_hashes_pos = list[tuple[str, PlaceholderRange]]()
948
949
        for req_id, encoder_input_ids in scheduled_encoder_inputs.items():
            req_state = self.requests[req_id]
950
951

            for mm_input_id in encoder_input_ids:
952
                mm_feature = req_state.mm_features[mm_input_id]
953
954
                if mm_feature.data is None:
                    continue
955
956
957
                mm_hash = mm_feature.identifier
                mm_kwargs.append(mm_feature.data)
                mm_hashes_pos.append((mm_hash, mm_feature.mm_position))
958
959
960
961
962
963
964
965

        # 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.
966
        model = cast(SupportsMultiModal, self.model)
967
        encoder_outputs = []
968
        for _, num_items, mm_kwargs_group in group_mm_kwargs_by_modality(
969
970
971
            mm_kwargs,
            device=self.device,
            pin_memory=self.pin_memory,
972
        ):
973
974
975
976
977
978
979
            # 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.
980
            torch_xla.sync(wait=False)
981
            curr_group_outputs = model.embed_multimodal(**mm_kwargs_group)
982
            torch_xla.sync(wait=False)
983

984
985
            sanity_check_mm_encoder_outputs(
                curr_group_outputs,
986
                expected_num_items=num_items,
987
988
            )

989
990
991
992
993
994
            if isinstance(curr_group_outputs, torch.Tensor):
                encoder_outputs.append(curr_group_outputs)
            else:
                assert isinstance(curr_group_outputs, (list, tuple))
                for output in curr_group_outputs:
                    encoder_outputs.append(output)
995
996

        # Cache the encoder outputs.
997
998
999
        # NOTE (NickLucche) here we diverge from logic in other runners, as we
        # assume to only have whole mm items to process. Hence we avoid the
        # intrinsic dynamism that `scatter_mm_placeholders` introduces.
1000
        for (mm_hash, pos_info), output in zip(mm_hashes_pos, encoder_outputs):
1001
1002
1003
            assert pos_info.is_embed is None, (
                "Expected all positions to be contiguous and embeddings."
            )
1004
            self.encoder_cache[mm_hash] = output
1005
1006

    def _gather_mm_embeddings(
1007
1008
        self,
        scheduler_output: "SchedulerOutput",
1009
1010
1011
    ) -> tuple[list[torch.Tensor], torch.Tensor]:
        total_num_scheduled_tokens = scheduler_output.total_num_scheduled_tokens
        padded_total_num_scheduled_tokens = _get_padded_token_len(
1012
1013
            self.num_tokens_paddings, total_num_scheduled_tokens
        )
1014
1015
1016
1017
1018
1019

        is_mm_embed = self.is_mm_embed_cpu
        is_mm_embed[:padded_total_num_scheduled_tokens] = False
        mm_embeds = list[torch.Tensor]()
        req_start_idx = 0

1020
        for req_id in self.input_batch.req_ids:
1021
            num_scheduled_tokens = scheduler_output.num_scheduled_tokens[req_id]
1022
1023
            req_state = self.requests[req_id]
            num_computed_tokens = req_state.num_computed_tokens
1024

1025
1026
1027
1028
            # TODO unroll loop and assume/enforce --disable_chunked_mm_input
            # NOTE (NickLucche) here we diverge from logic in other runners, as
            # we assume to only have whole mm items to process. Hence we avoid
            # the intrinsic dynamism that `gather_mm_placeholders` introduces.
1029
1030
            for mm_feature in req_state.mm_features:
                pos_info = mm_feature.mm_position
1031
1032
                start_pos = pos_info.offset
                num_encoder_tokens = pos_info.length
1033
1034
1035
1036
1037
1038
1039
1040
1041
1042
1043
1044

                # 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
1045
1046
1047
1048
1049
1050
1051
1052

                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

1053
                mm_hash = mm_feature.identifier
1054
                encoder_output = self.encoder_cache.get(mm_hash, None)
1055
                assert encoder_output is not None, f"Encoder cache miss for {mm_hash}."
1056

1057
1058
1059
                assert pos_info.is_embed is None, (
                    "Expected all positions to be contiguous and embeddings."
                )
1060
1061

                req_start_pos = req_start_idx + start_pos - num_computed_tokens
1062
                is_mm_embed[req_start_pos + start_idx : req_start_pos + end_idx] = True
1063
1064

                # Only whole mm items are processed
1065
                mm_embeds.append(encoder_output)
1066

1067
1068
            req_start_idx += num_scheduled_tokens

1069
        is_mm_embed = is_mm_embed[:padded_total_num_scheduled_tokens].to(self.device)
1070
1071
1072
1073
1074
1075

        return mm_embeds, is_mm_embed

    def _get_model_inputs(
        self,
        input_ids: torch.Tensor,
1076
        mm_embed_inputs: tuple[list[torch.Tensor], torch.Tensor] | None,
1077
    ):
1078
        if self.supports_mm_inputs:
1079
1080
            mm_embeds, is_mm_embed = mm_embed_inputs or (None, None)

1081
1082
1083
            # 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.
1084
            inputs_embeds = self.model.embed_input_ids(
1085
                input_ids,
1086
                multimodal_embeddings=mm_embeds,
1087
                is_multimodal=is_mm_embed,
1088
            )
1089

1090
1091
1092
1093
1094
1095
1096
1097
            return None, inputs_embeds
        else:
            # 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.
            return input_ids, None

1098
1099
1100
1101
    @torch.no_grad()
    def execute_model(
        self,
        scheduler_output: "SchedulerOutput",
1102
        intermediate_tensors: IntermediateTensors | None = None,
1103
1104
1105
1106
1107
1108
    ) -> ModelRunnerOutput | None:
        if self.scheduler_output is not None:
            raise RuntimeError(
                "State error: sample_tokens() must be called "
                "after execute_model() returns None."
            )
1109
1110
        # Update cached state
        self._update_states(scheduler_output)
1111
        if not scheduler_output.total_num_scheduled_tokens:
1112
1113
1114
1115
            if not has_kv_transfer_group():
                # Return empty ModelRunnerOutput if there's no work to do.
                return EMPTY_MODEL_RUNNER_OUTPUT

1116
            return self.kv_connector_no_forward(scheduler_output, self.vllm_config)
1117

1118
        mm_embed_inputs = None
1119
        if self.supports_mm_inputs:
1120
            # Run the multimodal encoder if any.
1121
            self._execute_mm_encoder(scheduler_output)
1122
1123
            mm_embed_inputs = self._gather_mm_embeddings(scheduler_output)

1124
        torch_xla.sync(wait=False)
1125
1126
1127
1128
1129
1130
1131
1132
1133
1134
1135

        self.scheduler_output = scheduler_output
        self.mm_embed_inputs = mm_embed_inputs
        return None

    @torch.no_grad()
    def sample_tokens(
        self, grammar_output: "GrammarOutput | None"
    ) -> ModelRunnerOutput:
        if self.scheduler_output is None:
            # Nothing to do (PP non-final rank case), output isn't used.
1136
            return None  # type: ignore[return-value]
1137
1138
1139
1140
1141
        scheduler_output = self.scheduler_output
        mm_embed_inputs = self.mm_embed_inputs
        self.scheduler_output = None
        self.mm_embed_inputs = None

1142
        # Prepare inputs, the requests might be split into multiple
1143
1144
1145
1146
        # executions, combine the result of each execution.
        start_index = 0
        combined_selected_tokens: list[torch.Tensor] = []
        combined_logprobs: list[LogprobsLists] = []
1147
1148
1149
1150
1151
1152

        # NOTE: setup current batch's metadata for kv connector.
        # Currently, only verified with NixlConnector
        with set_forward_context(None, self.vllm_config):
            self.maybe_setup_kv_connector(scheduler_output)

1153
        while start_index < self.input_batch.num_reqs:
1154
1155
1156
            attn_metadata, logits_indices, padded_num_reqs, num_reqs, end_index = (
                self._prepare_inputs(scheduler_output, start_index)
            )
1157
            input_ids, inputs_embeds = self._get_model_inputs(
1158
1159
                self.input_ids, mm_embed_inputs
            )
1160
            torch_xla.sync(wait=False)
1161
1162
            # Run the decoder
            with set_forward_context(
1163
1164
1165
1166
                attn_metadata,
                self.vllm_config,
                num_tokens=scheduler_output.total_num_scheduled_tokens,
            ):
1167
1168
1169
1170
1171
                hidden_states = self.model(
                    input_ids=input_ids,
                    positions=self.position_ids,
                    inputs_embeds=inputs_embeds,
                )
1172
            hidden_states = self.select_hidden_states(hidden_states, logits_indices)
1173
            logits = self.compute_logits(hidden_states)
1174
1175
1176
            tpu_sampling_metadata = TPUSupportedSamplingMetadata.from_input_batch(
                self.input_batch, padded_num_reqs, self.device
            )
1177
            if grammar_output is not None:
1178
                require_struct_decoding, grammar_bitmask_padded, arange = (
1179
                    self.prepare_structured_decoding_input(logits, grammar_output)
1180
1181
1182
1183
                )
                logits = self.structured_decode(
                    require_struct_decoding, grammar_bitmask_padded, logits, arange
                )
1184
            selected_token_ids = self.sample_from_logits_func(
1185
1186
                logits, tpu_sampling_metadata
            )
1187
1188
1189
1190
            # NOTE (NickLucche) Use the original logits (before any penalties or
            # temperature scaling) for the top-k logprobs. We can't enforce it
            # due to recompilations outside torch.compiled code, so just make
            # sure `sample_from_logits` does not modify the logits in-place.
1191
1192
1193
1194
1195
            logprobs = (
                self.gather_logprobs(logits, selected_token_ids)
                if tpu_sampling_metadata.logprobs
                else None
            )
1196
1197
1198
1199
1200
1201
1202
1203
1204
1205

            # Remove padding on cpu and keep dynamic op outside of xla graph.
            selected_token_ids = selected_token_ids.cpu()[:num_reqs]

            combined_selected_tokens.append(selected_token_ids)
            if tpu_sampling_metadata.logprobs:
                combined_logprobs.append(logprobs.tolists())

            start_index = end_index

1206
1207
1208
1209
1210
        # NOTE: current kv load and save get h2d/d2h copies involved.
        # Those copies are blocking. Once they become async., kv_save
        # should be called right after each single forward pass,
        # instead of the forwards of the entire input batch.
        self.maybe_wait_for_kv_save()
1211
1212
1213
        finished_sending, finished_recving = self.get_finished_kv_transfers(
            scheduler_output
        )
1214

1215
1216
1217
1218
1219
1220
1221
1222
1223
        selected_token_ids = torch.cat(combined_selected_tokens, dim=0)
        if tpu_sampling_metadata.logprobs:

            def concat_lists(input_lists):
                result = []
                for input_list in input_lists:
                    result.extend(input_list)
                return result

1224
1225
1226
1227
1228
1229
1230
1231
1232
            logprobs_lists = LogprobsLists(
                logprob_token_ids=concat_lists(
                    [lp.logprob_token_ids for lp in combined_logprobs]
                ),
                logprobs=concat_lists([lp.logprobs for lp in combined_logprobs]),
                sampled_token_ranks=concat_lists(
                    [lp.sampled_token_ranks for lp in combined_logprobs]
                ),
            )
1233
1234
        else:
            logprobs_lists = None
1235

1236
1237
        # Update the cache state concurrently. Code above will not block until
        # we use `selected_token_ids`. Add mark_step if post-processing changes
1238
        request_seq_lens: list[tuple[int, CachedRequestState, int]] = []
1239
        discard_sampled_tokens_req_indices = []
1240
        num_reqs = self.input_batch.num_reqs
1241
1242
1243
        for i, req_id in zip(range(num_reqs), self.input_batch.req_ids):
            assert req_id is not None
            req_state = self.requests[req_id]
1244
1245
1246
1247
            seq_len = (
                req_state.num_computed_tokens
                + scheduler_output.num_scheduled_tokens[req_id]
            )
1248
1249
1250
1251
1252
1253
1254
1255
1256
1257
            if seq_len >= req_state.num_tokens:
                request_seq_lens.append((i, req_state, seq_len))
            else:
                # Ignore the sampled token from the partial request.
                # Rewind the generator state as if the token was not sampled.
                generator = self.input_batch.generators.get(i)
                if generator is not None:
                    # This relies on cuda-specific torch-internal impl details
                    generator.set_offset(generator.get_offset() - 4)

1258
1259
1260
1261
                # Record the index of the request that should not be sampled,
                # so that we could clear the sampled tokens before returning.
                discard_sampled_tokens_req_indices.append(i)

1262
        assert all(
1263
1264
            req_id is not None for req_id in self.input_batch.req_ids[:num_reqs]
        ), "req_ids contains None"
1265
        req_ids = cast(list[str], self.input_batch.req_ids[:num_reqs])
1266

1267
        prompt_logprobs_dict: dict[str, LogprobsTensors | None] = {}
1268
        for req_id in self.input_batch.req_ids[:num_reqs]:
1269
1270
            prompt_logprobs_dict[req_id] = None

1271
1272
        max_gen_len = selected_token_ids.shape[-1]
        if max_gen_len == 1:
1273
            valid_sampled_token_ids = selected_token_ids.tolist()
1274

1275
1276
1277
1278
            # Mask out the sampled tokens that should not be sampled.
            # TODO: Keep in sync with gpu_model_runner.py, in particular
            #       the "else" case here
            for i in discard_sampled_tokens_req_indices:
1279
                valid_sampled_token_ids[i].clear()
1280
1281

            # Append sampled tokens
1282
1283
1284
1285
1286
            for i, req_state, seq_len in request_seq_lens:
                token_id = valid_sampled_token_ids[i][0]
                self.input_batch.token_ids_cpu[i, seq_len] = token_id
                req_state.output_token_ids.append(token_id)
                self.input_batch.num_tokens[i] += 1
1287

1288
1289
1290
1291
        else:
            valid_mask = selected_token_ids != INVALID_TOKEN_ID
            gen_lens = valid_mask.sum(dim=1).tolist()
            valid_sampled_token_ids = [
1292
                seq.tolist() for seq in selected_token_ids[valid_mask].split(gen_lens)
1293
1294
1295
1296
            ]
            self.input_batch.num_tokens[:num_reqs] += gen_lens
            for i, req_state, seq_len in request_seq_lens:
                target_slice = slice(seq_len - gen_lens[i] + 1, seq_len + 1)
1297
1298
1299
                self.input_batch.token_ids_cpu[i, target_slice] = (
                    valid_sampled_token_ids[i]
                )
1300
1301
                req_state.output_token_ids.extend(valid_sampled_token_ids[i])

1302
1303
1304
1305
        kv_connector_output = (
            None
            if (finished_sending is None and finished_recving is None)
            else KVConnectorOutput(
1306
1307
1308
                finished_sending=finished_sending,
                finished_recving=finished_recving,
            )
1309
        )
1310

1311
        model_runner_output = ModelRunnerOutput(
1312
            req_ids=req_ids,
1313
            req_id_to_index=self.input_batch.req_id_to_index,
1314
            sampled_token_ids=valid_sampled_token_ids,
1315
            logprobs=logprobs_lists,
1316
            prompt_logprobs_dict=prompt_logprobs_dict,
1317
            pooler_output=[],
1318
1319
            kv_connector_output=kv_connector_output,
        )
1320
1321
1322
1323
1324

        # Check there are no new graphs compiled - all the graphs should be
        # captured and compiled during warm up.
        self._verify_num_xla_graphs("execute_model")

1325
1326
        return model_runner_output

1327
1328
1329
1330
1331
    def update_config(self, overrides: dict[str, Any]) -> None:
        # TODO: TPU config may need extra validation
        # https://github.com/vllm-project/vllm/pull/20095#discussion_r2201497754
        allowed_config_names = {"load_config", "model_config"}
        for config_name, config_overrides in overrides.items():
1332
1333
            assert config_name in allowed_config_names, (
                f"Config `{config_name}` not supported. "
1334
                f"Allowed configs: {allowed_config_names}"
1335
            )
1336
1337
1338
1339
            config = getattr(self, config_name)
            new_config = update_config(config, config_overrides)
            setattr(self, config_name, new_config)

1340
1341
1342
1343
1344
1345
1346
1347
1348
1349
1350
1351
1352
1353
    def load_model(self) -> None:
        self.device = self.device_config.device

        # NOTE(woosuk): While the executor assigns the TP ranks to the worker
        # process, the ranks can be different from the ranks internally assigned
        # by the xm runtime. Therefore, there is a mismatch in the rank
        # assignment between the gloo (cpu) runtime and the xm (tpu) runtime.
        # This is not a problem in linear layers because all-reduce is
        # rank-agnostic. However, it matters for all-gather as the ranks
        # determine the order of concatenating the output tensors.
        # As a workaround, we use the xm's rank assignment only when loading
        # the embedding weights.
        xm_tp_rank = xr.global_ordinal()
        with patch(
1354
1355
1356
1357
            "vllm.model_executor.layers.vocab_parallel_embedding."
            "get_tensor_model_parallel_rank",
            return_value=xm_tp_rank,
        ):
1358
1359
1360
            try:
                if self.use_spmd:
                    tpu_loader = TPUModelLoader(
1361
1362
                        load_config=self.vllm_config.load_config
                    )
1363
                    model = tpu_loader.load_model(
1364
                        vllm_config=self.vllm_config,
1365
                        model_config=self.vllm_config.model_config,
1366
1367
                        mesh=self.mesh,
                    )
1368
                else:
1369
                    model_loader = get_model_loader(self.load_config)
1370
1371
                    logger.info("Loading model from scratch...")
                    model = model_loader.load_model(
1372
1373
                        vllm_config=self.vllm_config, model_config=self.model_config
                    )
1374
1375
1376
1377
1378
1379
            except RuntimeError as e:
                raise RuntimeError(
                    f"Unable to load model, a likely reason is the model is "
                    "too large for the current device's HBM memory. "
                    "Consider switching to a smaller model "
                    "or sharding the weights on more chips. "
1380
1381
                    f"See the detailed error: {e}"
                ) from e
1382
        if self.lora_config is not None:
1383
            model = self.load_lora_model(model, self.vllm_config, self.device)
1384
            replace_set_lora(model)
1385

1386
1387
        # Sync all pending XLA execution during model initialization and weight
        # loading.
1388
        torch_xla.sync(wait=False)
1389
        xm.wait_device_ops()
1390
1391
        if not hasattr(self, "model"):
            self.model = model
1392
        self.sampler = TPUSampler()
1393

1394
    def reload_weights(self) -> None:
1395
        assert getattr(self, "model", None) is not None, (
1396
            "Cannot reload weights before model is loaded."
1397
        )
1398
1399
1400
1401
        model_loader = get_model_loader(self.load_config)
        logger.info("Reloading weights inplace...")
        model_loader.load_weights(self.model, model_config=self.model_config)

1402
    @torch.no_grad()
1403
    def _dummy_run(self, num_tokens: int, num_reqs: int, num_blocks: int) -> None:
1404
        if self.supports_mm_inputs:
1405
            input_ids = None
1406
            inputs_embeds = torch.zeros(
1407
1408
1409
                (num_tokens, self.inputs_embeds_size),
                dtype=self.dtype,
                device=self.device,
1410
            )
1411
        else:
1412
            input_ids = torch.zeros((num_tokens), dtype=torch.int32).to(self.device)
1413
            inputs_embeds = None
1414
        actual_num_reqs = min(num_tokens, num_reqs)
1415
        position_ids = torch.zeros(num_tokens, dtype=torch.int32).to(self.device)
1416
        padded_num_slices = _get_padded_num_kv_cache_update_slices(
1417
1418
1419
1420
1421
1422
1423
1424
1425
1426
1427
            num_tokens, self.max_num_reqs, self.block_size
        )
        num_kv_update_slices = torch.tensor([padded_num_slices], dtype=torch.int32).to(
            self.device
        )
        slot_mapping = torch.zeros((3, padded_num_slices), dtype=torch.int32).to(
            self.device
        )
        block_tables = torch.zeros((num_reqs, num_blocks), dtype=torch.int32).to(
            self.device
        )
1428
        query_lens = [1] * num_reqs
1429
1430
1431
1432
1433
        query_start_loc = torch.cumsum(
            torch.tensor([0] + query_lens, dtype=torch.int32), dim=0, dtype=torch.int32
        ).to(self.device)
        context_lens = torch.ones((num_reqs,), dtype=torch.int32).to(self.device)
        num_seqs = torch.tensor([actual_num_reqs], dtype=torch.int32).to(self.device)
1434
1435
1436
1437
1438
        attn_metadata = PallasMetadata(
            slot_mapping=slot_mapping,
            block_tables=block_tables,
            context_lens=context_lens,
            query_start_loc=query_start_loc,
1439
            num_seqs=num_seqs,
1440
            num_kv_update_slices=num_kv_update_slices,
1441
            num_slices_per_kv_cache_update_block=self._num_slices_per_kv_cache_update_block,
1442
        )
1443

1444
        if self.supports_mm_inputs:
1445
1446
1447
            torch._dynamo.mark_dynamic(inputs_embeds, 0)
        else:
            torch._dynamo.mark_dynamic(input_ids, 0)
1448
1449
        torch._dynamo.mark_dynamic(position_ids, 0)
        torch._dynamo.mark_dynamic(attn_metadata.slot_mapping, 0)
1450
1451
1452
        torch._dynamo.mark_dynamic(attn_metadata.block_tables, (0, 1))
        torch._dynamo.mark_dynamic(attn_metadata.context_lens, 0)
        torch._dynamo.mark_dynamic(attn_metadata.query_start_loc, 0)
1453

1454
        layer_names = get_layers_from_vllm_config(self.vllm_config, Attention).keys()
1455
        per_layer_attn_metadata = {
1456
            layer_name: attn_metadata for layer_name in layer_names
1457
1458
        }

1459
1460
1461
1462
1463
1464
1465
1466
1467
        with (
            self.maybe_select_dummy_loras(
                self.lora_config, np.array([num_tokens], dtype=np.int32)
            ),
            set_forward_context(per_layer_attn_metadata, self.vllm_config, 0),
        ):
            out = self.model(
                input_ids=input_ids, positions=position_ids, inputs_embeds=inputs_embeds
            )
1468
        self._hidden_states_dtype = out.dtype
1469

1470
1471
1472
    def _set_active_loras(
        self, prompt_lora_mapping, token_lora_mapping, lora_requests
    ) -> None:
1473
        torch_xla.sync(wait=False)  # Captures input updates
1474
1475
1476
        super()._set_active_loras(
            prompt_lora_mapping, token_lora_mapping, lora_requests
        )
1477
        torch_xla.sync(wait=False)  # Captures metadata updates
1478

1479
    def _precompile_mm_encoder(self) -> None:
1480
        if not self.supports_mm_inputs:
1481
1482
            return

1483
1484
        # Pre-compile MM encoder for all supported data modalities.
        hf_config = self.vllm_config.model_config.hf_config
1485
1486
1487
1488
1489
1490
1491

        mm_budget = self.mm_budget
        assert mm_budget is not None

        max_items_per_seq_by_modality = mm_budget.max_items_per_batch_by_modality  # noqa: E501

        for mode, max_items_per_seq in max_items_per_seq_by_modality.items():
1492
            logger.info(
1493
1494
                "Compiling Multimodal %s Encoder with different input shapes.", mode
            )
1495
1496
            start = time.perf_counter()
            # No padding for MM encoder just yet.
1497
            for num_items in range(1, max_items_per_seq + 1):
1498
1499
                logger.info("  -- mode: %s items: %d", mode, num_items)
                batched_dummy_mm_inputs = self._get_mm_dummy_batch(
1500
1501
1502
                    mode,
                    num_items,
                )
1503
                # Run multimodal encoder.
1504
                torch_xla.sync(wait=False)
1505
                mm_embeds = self.model.embed_multimodal(**batched_dummy_mm_inputs)
1506
                torch_xla.sync(wait=False)
1507
1508
1509
                num_patches = mm_embeds[0].shape[0]
                items_size = num_patches * num_items

1510
                # NOTE (NickLucche) pre-compile `embed_input_ids` when mm
1511
1512
1513
1514
1515
1516
1517
1518
                # embeddings are present. We assume `--disable-mm-chunked`,
                # hence only whole items can be scheduled. This implies we just
                # need to compile when `num_items` fit the (padded) `input_ids`
                for num_tokens in self.num_tokens_paddings:
                    if num_tokens >= items_size:
                        # XLA Workaround: if torch.zeros(..device) is used, XLA
                        # compiles a scalar+expansion op, which won't match
                        # the graph generated at runtime. CPU->TPU must be used
1519
1520
1521
                        placeholders_ids = torch.zeros(
                            num_tokens, dtype=torch.int32, device="cpu"
                        )
1522
                        # Align placeholders and actual num mm_embeddings.
1523
                        placeholders_ids[:items_size] = hf_config.image_token_index
1524
1525

                        placeholders_ids = placeholders_ids.to(self.device)
1526
1527
1528
1529

                        mm_mask = torch.tensor([False] * num_tokens)
                        mm_mask[:items_size] = True
                        mm_mask = mm_mask.to(self.device)
1530
                        # Assign outputs or the graph will be cut short.
1531
1532
1533
1534
                        a, b = self._get_model_inputs(
                            placeholders_ids,
                            mm_embed_inputs=([mm_embeds], mm_mask),
                        )
1535
                        assert a is None
1536
                        torch_xla.sync(wait=False)
1537

1538
            # Pre-compile `embed_input_ids` when mm_embeddings are not
1539
1540
            # present. Chunk is only made of text, no mm_placeholders.
            for num_tokens in self.num_tokens_paddings:
1541
1542
1543
                placeholders_ids = torch.zeros(
                    num_tokens, dtype=torch.int32, device="cpu"
                )
1544
                placeholders_ids = placeholders_ids.to(self.device)
1545
1546
1547
1548
                a, b = self._get_model_inputs(
                    placeholders_ids,
                    mm_embed_inputs=None,
                )
1549
                assert a is None
1550
                torch_xla.sync(wait=False)
1551
1552
1553
1554

            xm.wait_device_ops()
            end = time.perf_counter()
            logger.info(
1555
1556
1557
1558
                "Multimodal %s Encoder compilation finished in in %.2f [secs].",
                mode,
                end - start,
            )
1559

1560
    def _precompile_backbone(self) -> None:
1561
1562
        logger.info("Compiling the model with different input shapes.")
        start = time.perf_counter()
1563
        for num_tokens in self.num_tokens_paddings:
1564
            logger.info("  -- num_tokens: %d", num_tokens)
1565
1566
1567
            self._dummy_run(
                num_tokens, self.num_reqs_max_model_len, self.max_num_blocks_per_req
            )
1568
            if self.most_model_len is not None:
1569
1570
1571
1572
1573
                self._dummy_run(
                    num_tokens,
                    self.num_reqs_most_model_len,
                    self.num_blocks_per_most_len_req,
                )
1574
1575
        xm.wait_device_ops()
        end = time.perf_counter()
1576
        logger.info("Compilation finished in %.2f [secs].", end - start)
1577
        self._update_num_xla_graphs("model backbone")
1578

1579
1580
1581
    def _precompile_select_hidden_states(self) -> None:
        # Compile hidden state selection function for bucketed
        # n_tokens x max_num_reqs. Graph is really small so this is fine.
1582
        logger.info("Compiling select_hidden_states with different input shapes.")
1583
1584
        start = time.perf_counter()
        hsize = self.model_config.get_hidden_size()
1585
        for num_tokens in self.num_tokens_paddings:
1586
1587
1588
            dummy_hidden = torch.zeros(
                (num_tokens, hsize), device=self.device, dtype=self._hidden_states_dtype
            )
1589
1590
            torch._dynamo.mark_dynamic(dummy_hidden, 0)
            for num_reqs in self.num_reqs_paddings:
1591
                indices = torch.zeros(num_reqs, dtype=torch.int32, device=self.device)
1592
1593
                torch._dynamo.mark_dynamic(indices, 0)
                self.select_hidden_states(dummy_hidden, indices)
1594
                logger.info("  -- num_tokens: %d, num_seqs: %d", num_tokens, num_reqs)
1595
1596
1597
1598
                # Requests can't be more than tokens. But do compile for the
                # next bigger value in case num_tokens uses bucketed padding.
                if num_reqs >= min(num_tokens, self.max_num_reqs):
                    break
1599
        xm.wait_device_ops()
1600
        end = time.perf_counter()
1601
        logger.info("Compilation finished in %.2f [secs].", end - start)
1602
        self._update_num_xla_graphs("select_hidden_states")
1603

1604
1605
    def _precompile_compute_logits(self) -> None:
        logger.info("Compiling compute_logits with different input shapes.")
1606
1607
1608
        start = time.perf_counter()
        hsize = self.model_config.get_hidden_size()
        for num_reqs in self.num_reqs_paddings:
1609
1610
1611
            dummy_hidden = torch.zeros(
                (num_reqs, hsize), device=self.device, dtype=self._hidden_states_dtype
            )
1612
1613
1614
1615
1616
1617
1618
1619
1620
            torch._dynamo.mark_dynamic(dummy_hidden, 0)
            self.compute_logits(dummy_hidden)
            logger.info("  -- num_seqs: %d", num_reqs)
        xm.wait_device_ops()
        end = time.perf_counter()
        logger.info("Compilation finished in %.2f [secs].", end - start)
        self._update_num_xla_graphs("compute_logits")

    def _precompile_structured_decoding(self) -> None:
1621
        logger.info("Compiling structured_decoding with different input shapes.")
1622
1623
        start = time.perf_counter()
        for num_reqs in self.num_reqs_paddings:
1624
1625
1626
1627
1628
1629
1630
1631
1632
            dummy_logits = torch.zeros(
                (num_reqs, self.vocab_size),
                device=self.device,
                dtype=self._hidden_states_dtype,
            )
            dummy_require_struct_decoding = self.require_structured_out_cpu[
                :num_reqs
            ].to(self.device)
            dummy_grammar_bitmask = self.grammar_bitmask_cpu[:num_reqs].to(self.device)
1633
1634
1635
1636
            # The first dimension of the above 3 dummy tensors cannot be
            # mark_dynamic because some operations in structured_decode require
            # them to be static.
            arange = self.structured_decode_arange.to(self.device)
1637
1638
1639
1640
1641
1642
            self.structured_decode(
                dummy_require_struct_decoding,
                dummy_grammar_bitmask,
                dummy_logits,
                arange,
            )
1643
1644
1645
1646
1647
1648
1649
            logger.info("  -- num_seqs: %d", num_reqs)
        xm.wait_device_ops()
        end = time.perf_counter()
        logger.info("Compilation finished in %.2f [secs].", end - start)
        self._update_num_xla_graphs("structured_decoding")

    def _precompile_sample_from_logits(self) -> None:
1650
        logger.info("Compiling sample_from_logits with different input shapes.")
1651
1652
        start = time.perf_counter()
        for num_reqs in self.num_reqs_paddings:
1653
1654
1655
1656
1657
            dummy_logits = torch.zeros(
                (num_reqs, self.vocab_size),
                device=self.device,
                dtype=self._hidden_states_dtype,
            )
1658
1659
            # The first dimension of dummy_logits cannot be mark_dynamic
            # because some operations in the sampler require it to be static.
1660
1661
            for all_greedy in [False, True]:
                generate_params_if_all_greedy = not all_greedy
1662
1663
1664
1665
1666
1667
                sampling_metadata = TPUSupportedSamplingMetadata.from_input_batch(
                    self.input_batch,
                    num_reqs,
                    self.device,
                    generate_params_if_all_greedy,
                )
1668
                sampling_metadata.all_greedy = all_greedy
1669
                with self.maybe_select_dummy_loras(
1670
1671
1672
                    self.lora_config, np.array([num_reqs], dtype=np.int32)
                ):
                    self.sample_from_logits_func(dummy_logits, sampling_metadata)
1673
1674
1675
            logger.info("  -- num_seqs: %d", num_reqs)
        xm.wait_device_ops()
        end = time.perf_counter()
1676
1677
        logger.info("Compilation finished in %.2f [secs].", end - start)
        self._update_num_xla_graphs("sample_from_logits")
1678

1679
1680
1681
1682
    def _precompile_gather_logprobs(self) -> None:
        logger.info("Compiling gather_logprobs with different input shapes.")
        start = time.perf_counter()
        for num_reqs in self.num_reqs_paddings:
1683
1684
1685
1686
1687
1688
            dummy_logits = torch.zeros(
                (num_reqs, self.vocab_size),
                device=self.device,
                dtype=self._hidden_states_dtype,
            )
            dummy_tokens = torch.zeros((num_reqs, 1), dtype=torch.int64).to(self.device)
1689
            with self.maybe_select_dummy_loras(
1690
1691
                self.lora_config, np.array([num_reqs], dtype=np.int32)
            ):
1692
                self.gather_logprobs(dummy_logits, dummy_tokens)
1693
1694
1695
1696
1697
1698
            logger.info("  -- num_seqs: %d", num_reqs)
        xm.wait_device_ops()
        end = time.perf_counter()
        logger.info("Compilation finished in %.2f [secs].", end - start)
        self._update_num_xla_graphs("gather_logprobs")

1699
1700
1701
1702
    def capture_model(self) -> None:
        """
        Precompile all the subgraphs with possible input shapes.
        """
1703
1704
1705
1706
1707
1708
1709
1710
        with self.maybe_setup_dummy_loras(self.lora_config):
            self._precompile_mm_encoder()
            self._precompile_backbone()
            self._precompile_select_hidden_states()
            self._precompile_compute_logits()
            self._precompile_structured_decoding()
            self._precompile_sample_from_logits()
            self._precompile_gather_logprobs()
1711

1712
1713
1714
1715
1716
    def profile_run(
        self,
        num_tokens: int,
    ) -> None:
        # Profile with multimodal encoder & encoder cache.
1717
        if self.supports_mm_inputs:
1718
1719
            mm_config = self.model_config.multimodal_config
            if mm_config is not None and mm_config.skip_mm_profiling:
1720
                logger.info(
1721
                    "Skipping memory profiling for multimodal encoder and "
1722
1723
                    "encoder cache."
                )
1724
1725
1726
1727
1728
1729
1730
1731
1732
            else:
                mm_budget = self.mm_budget
                assert mm_budget is not None

                # TODO: handle encoder-decoder models once we support them.
                if (encoder_budget := mm_budget.get_encoder_budget()) > 0:
                    # NOTE: Currently model is profiled with a single non-text
                    # modality with the max possible input tokens even when
                    # it supports multiple.
1733
                    dummy_modality = mm_budget.get_modality_with_max_tokens()
1734
1735
1736
                    max_mm_items_per_batch = mm_budget.max_items_per_batch_by_modality[
                        dummy_modality
                    ]
1737
1738
1739
1740
1741
1742
1743
1744
1745

                    logger.info(
                        "Encoder cache will be initialized with a budget of "
                        "%s tokens, and profiled with %s %s items of the "
                        "maximum feature size.",
                        encoder_budget,
                        max_mm_items_per_batch,
                        dummy_modality,
                    )
1746

1747
1748
1749
1750
1751
                    # Create dummy batch of multimodal inputs.
                    batched_dummy_mm_inputs = self._get_mm_dummy_batch(
                        dummy_modality,
                        max_mm_items_per_batch,
                    )
1752

1753
1754
1755
1756
                    # Run multimodal encoder.
                    # Isolate encoder graph from post-processing to minimize
                    # impact of recompilation until it's fixed.
                    start = time.perf_counter()
1757
                    torch_xla.sync(wait=False)
1758
                    dummy_encoder_outputs = self.model.embed_multimodal(
1759
1760
                        **batched_dummy_mm_inputs
                    )
1761
                    torch_xla.sync(wait=False)
1762
1763
1764
1765
                    xm.wait_device_ops()
                    end = time.perf_counter()
                    logger.info(
                        "Multimodal Encoder profiling finished in %.2f [secs].",
1766
1767
                        end - start,
                    )
1768
1769
1770
1771
1772

                    sanity_check_mm_encoder_outputs(
                        dummy_encoder_outputs,
                        expected_num_items=max_mm_items_per_batch,
                    )
1773

1774
                    # Cache the dummy encoder outputs.
1775
                    self.encoder_cache["tmp"] = dict(enumerate(dummy_encoder_outputs))
1776
1777

        # Trigger compilation for general shape.
1778
1779
1780
        self._dummy_run(
            num_tokens, self.num_reqs_max_model_len, self.max_num_blocks_per_req
        )
1781
        if self.most_model_len is not None:
1782
1783
1784
1785
1786
            self._dummy_run(
                num_tokens,
                self.num_reqs_most_model_len,
                self.num_blocks_per_most_len_req,
            )
1787

1788
        torch_xla.sync(wait=False)
1789
1790
1791
1792
        xm.wait_device_ops()
        self.encoder_cache.clear()
        gc.collect()

1793
1794
1795
1796
1797
1798
1799
1800
1801
1802
1803
1804
1805
1806
1807
1808
1809
1810
    def maybe_setup_cross_layer_kv_sharing(
        self,
        kv_caches: dict[str, torch.Tensor],
        kv_cache_config: KVCacheConfig,
    ) -> None:
        """
        Add layers that re-use KV cache to KV cache group of its target layer.
        Mapping of KV cache tensors happens in `initialize_kv_cache_tensors()`
        """
        if not self.shared_kv_cache_layers:
            # No cross-layer KV sharing, return
            return

        add_kv_sharing_layers_to_kv_cache_groups(
            self.shared_kv_cache_layers,
            kv_cache_config.kv_cache_groups,
        )

1811
1812
        for layer_name, target_layer_name in self.shared_kv_cache_layers.items():
            logger.debug("%s reuses KV cache of %s", layer_name, target_layer_name)
1813
1814
            kv_caches[layer_name] = kv_caches[target_layer_name]

1815
1816
1817
1818
    def initialize_kv_cache(self, kv_cache_config: KVCacheConfig) -> None:
        """
        Initialize KV cache based on `kv_cache_config`.
        Args:
1819
            kv_cache_config: Configuration for the KV cache, including the KV
1820
1821
            cache size of each layer
        """
1822
        if len(kv_cache_config.kv_cache_groups) > 1:
1823
            raise NotImplementedError(
1824
1825
                "Hybrid models with more than one KV cache type are not supported yet."
            )
1826

1827
1828
1829
1830
        if (
            kv_cache_config.kv_cache_groups[0].kv_cache_spec.block_size
            != self.block_size
        ):
1831
1832
1833
1834
1835
1836
1837
            self.input_batch = InputBatch(
                max_num_reqs=self.max_num_reqs,
                max_model_len=self.max_model_len,
                max_num_batched_tokens=self.max_num_tokens,
                device=self.device,
                pin_memory=self.pin_memory,
                vocab_size=self.model_config.get_vocab_size(),
1838
1839
1840
                block_sizes=[
                    kv_cache_config.kv_cache_groups[0].kv_cache_spec.block_size
                ],
1841
1842
1843
                kernel_block_sizes=[
                    kv_cache_config.kv_cache_groups[0].kv_cache_spec.block_size
                ],
1844
1845
            )
        # Verify dtype compatibility between block_table_cpu and input_batch
1846
1847
1848
1849
        assert (
            self.block_table_cpu.dtype
            == self.input_batch.block_table[0].get_cpu_tensor().dtype
        )
1850

1851
1852
1853
        kv_cache_sizes = {}
        for kv_cache_tensor in kv_cache_config.kv_cache_tensors:
            assert len(kv_cache_tensor.shared_by) == 1, (
1854
1855
                "KV cache tensor shared by multiple layers is not supported in TPU."
            )
1856
            kv_cache_sizes[kv_cache_tensor.shared_by[0]] = kv_cache_tensor.size
1857

1858
        kv_caches: dict[str, torch.Tensor] = {}
1859
1860
1861
        for kv_cache_group in kv_cache_config.kv_cache_groups:
            kv_cache_spec = kv_cache_group.kv_cache_spec
            for layer_name in kv_cache_group.layer_names:
1862
1863
1864
                tensor_size = kv_cache_sizes[layer_name]
                assert tensor_size % kv_cache_spec.page_size_bytes == 0
                num_blocks = tensor_size // kv_cache_spec.page_size_bytes  # noqa
1865
                if isinstance(kv_cache_spec, AttentionSpec):
1866
1867
1868
                    if self.use_spmd:
                        num_kv_heads = kv_cache_spec.num_kv_heads
                        assert self.original_parallel_config is not None
1869
                        tp_size = self.original_parallel_config.tensor_parallel_size
1870
1871
1872
                        # TODO: Handle kv cache duplication under SPMD mode.
                        assert num_kv_heads % tp_size == 0, (
                            f"num_kv_heads {num_kv_heads} must be divisible by "
1873
1874
                            f"tp_size {tp_size} under SPMD mode"
                        )
1875
                    kv_cache_shape = PallasAttentionBackend.get_kv_cache_shape(
1876
1877
1878
1879
1880
                        num_blocks,
                        kv_cache_spec.block_size,
                        kv_cache_spec.num_kv_heads,
                        kv_cache_spec.head_size,
                    )
1881
1882
                    dtype = kv_cache_spec.dtype

1883
1884
1885
                    tpu_kv_cache = torch.zeros(kv_cache_shape, dtype=dtype).to(
                        self.device
                    )
1886

1887
                    kv_caches[layer_name] = tpu_kv_cache
1888
1889
                else:
                    raise NotImplementedError
1890

1891
1892
        # Set up cross-layer KV cache sharing if needed
        self.maybe_setup_cross_layer_kv_sharing(kv_caches, kv_cache_config)
1893

1894
1895
1896
        bind_kv_cache(
            kv_caches,
            self.vllm_config.compilation_config.static_forward_context,
1897
1898
            self.kv_caches,
        )
1899

1900
1901
1902
        if self.use_spmd:
            # Shard KV Cache
            for cache in self.kv_caches:
1903
                xs.mark_sharding(cache, self.mesh, (None, "x", None, None))
1904

1905
1906
1907
1908
        if has_kv_transfer_group():
            get_kv_transfer_group().register_kv_caches(kv_caches)
            get_kv_transfer_group().set_host_xfer_buffer_ops(copy_kv_blocks)

1909
    def reset_dynamo_cache(self):
1910
1911
1912
1913
        # NOTE: We check `is_multimodal_model` instead of `supports_mm_inputs`
        # since the compiled model object of the language backbone of a
        # multimodal model needs to be extracted via `get_language_model`.
        if self.model_config.is_multimodal_model:
1914
            compiled_model = self.model.get_language_model().model
1915
1916
        else:
            compiled_model = self.model.model
1917
        if isinstance(compiled_model, TorchCompileWithNoGuardsWrapper):
1918
1919
            logger.info("Clear dynamo cache and cached dynamo bytecode.")
            torch._dynamo.eval_frame.remove_from_cache(
1920
                compiled_model.original_code_object()
1921
            )
1922
1923
1924
            # Reset the wrapper to re-initialize.
            compiled_model.compiled = False
            TorchCompileWithNoGuardsWrapper.__init__(compiled_model)
1925

1926
1927
1928
1929
1930
    @torch.compile(backend="openxla", fullgraph=True, dynamic=False)
    def select_hidden_states(self, hidden_states, indices_do_sample):
        return hidden_states[indices_do_sample]

    @torch.compile(backend="openxla", fullgraph=True, dynamic=False)
1931
    def compute_logits(self, sample_hidden_states: torch.Tensor) -> torch.Tensor:
1932
        return self.model.compute_logits(sample_hidden_states)
1933

1934
1935
1936
    # TODO: Under SPMD mode, sample_from_logits has correctness issue.
    #       Re-enable the torch.compile once the issue is fixed in torchxla.
    # @torch.compile(backend="openxla", fullgraph=True, dynamic=False)
1937
    def sample_from_logits(
1938
1939
        self, logits: torch.Tensor, sampling_metadata: TPUSupportedSamplingMetadata
    ) -> torch.Tensor:
1940
        """
1941
        Sample with xla-friendly function. This function is to be traced
1942
1943
        separately from `forward` for lighter compilation overhead.
        """
1944
1945
1946
        if sampling_metadata.all_greedy:
            out_tokens = torch.argmax(logits, dim=-1, keepdim=True)
        else:
1947
            out_tokens = self.sampler(logits, sampling_metadata).sampled_token_ids
1948
1949
        return out_tokens

1950
    @torch.compile(backend="openxla", fullgraph=True, dynamic=False)
1951
1952
1953
    def gather_logprobs(
        self, logits: torch.Tensor, sampled_tokens: torch.Tensor
    ) -> LogprobsTensors:
1954
1955
1956
1957
1958
1959
1960
1961
1962
        """
        Gather the top_logprobs with corresponding tokens. Use a fixed number
        of logprobs as an alternative to having multiple pre-compiled graphs.
        Select the number of logprobs actually demanded by each request on CPU.
        """
        logprobs = self.sampler.compute_logprobs(logits)
        return self.sampler.gather_logprobs(
            logprobs,
            self.model_config.max_logprobs,
1963
1964
            token_ids=sampled_tokens.squeeze(-1),
        )
1965

1966
    @torch.compile(backend="openxla", fullgraph=True, dynamic=False)
1967
1968
1969
1970
1971
1972
1973
    def structured_decode(
        self,
        require_struct_decoding: torch.Tensor,
        grammar_bitmask: torch.Tensor,
        logits: torch.Tensor,
        arange: torch.Tensor,
    ) -> torch.Tensor:
1974
1975
1976
        return torch.where(
            require_struct_decoding,
            self.apply_grammar_bitmask(logits, grammar_bitmask, arange),
1977
1978
            logits,
        )
1979

1980
1981
1982
1983
    def apply_grammar_bitmask(
        self, logits: torch.Tensor, grammar_bitmask: torch.Tensor, arange: torch.Tensor
    ):
        assert logits.shape[0] == grammar_bitmask.shape[0]
1984
1985
        logits_cloned = logits.clone()
        for i in range(logits.shape[0]):
1986
1987
1988
1989
1990
            unpacked_bitmask = (
                torch.bitwise_right_shift(grammar_bitmask[i][:, None], arange[None, :])
                & 1
            ) == 0
            unpacked_bitmask = unpacked_bitmask.reshape(-1)[: self.vocab_size]
1991
            logits_cloned[i] = logits_cloned[i].masked_fill(
1992
1993
                unpacked_bitmask, -float("inf")
            )
1994
1995
        return logits_cloned

1996
1997
    def embed_multimodal(self, *args, **kwargs):
        return self.model.embed_multimodal(*args, **kwargs)
1998

1999
2000
    def embed_input_ids(self, *args, **kwargs):
        return self.model.embed_input_ids(*args, **kwargs)
2001

2002
    def prepare_structured_decoding_input(
2003
        self, logits: torch.Tensor, grammar_output: "GrammarOutput"
2004
    ) -> tuple[torch.Tensor, torch.Tensor, torch.Tensor]:
2005
        grammar_bitmask = grammar_output.grammar_bitmask
2006
2007
2008
2009
2010
2011
        num_reqs, _ = logits.shape

        # Reset pre-allocated tensors
        self.grammar_bitmask_cpu.zero_()
        self.require_structured_out_cpu.zero_()

2012
        cumulative_mask_idx = 0
2013
        for req_id in grammar_output.structured_output_request_ids:
2014
            if req_id not in self.input_batch.req_id_to_index:
2015
2016
                continue
            batch_index = self.input_batch.req_id_to_index[req_id]
2017
            self.grammar_bitmask_cpu[batch_index] = torch.from_numpy(
2018
2019
                grammar_bitmask[cumulative_mask_idx]
            )
2020
2021
2022
2023
2024
2025
            # It's not guaranteed that all requests in this batch require
            # structured output, so create a bool tensor to represent
            # the requests that need structured output.
            self.require_structured_out_cpu[batch_index] = True
            cumulative_mask_idx += 1

2026
2027
2028
2029
2030
        return (
            self.require_structured_out_cpu[:num_reqs].to(logits.device),
            self.grammar_bitmask_cpu[:num_reqs].to(logits.device),
            self.structured_decode_arange.to(logits.device),
        )
2031

2032
2033
2034
2035
2036
2037
    def _get_mm_dummy_batch(
        self,
        modality: str,
        max_items_per_batch: int,
    ) -> BatchedTensorInputs:
        """Dummy data for profiling and precompiling multimodal models."""
2038
2039
        assert self.mm_budget is not None

2040
        dummy_decoder_data = self.mm_registry.get_decoder_dummy_data(
2041
            model_config=self.model_config,
2042
            seq_len=self.max_model_len,
2043
            mm_counts={modality: 1},
2044
            cache=self.mm_budget.cache,
2045
        )
2046
2047
2048
        dummy_mm_data = dummy_decoder_data.multi_modal_data

        # Result in the maximum GPU consumption of the model
2049
2050
        dummy_mm_item = dummy_mm_data[modality][0]
        dummy_mm_items = [dummy_mm_item] * max_items_per_batch
2051

2052
2053
2054
2055
2056
2057
2058
2059
        return next(
            grouped_mm_kwargs
            for _, _, grouped_mm_kwargs in group_mm_kwargs_by_modality(
                dummy_mm_items,
                device=self.device,
                pin_memory=self.pin_memory,
            )
        )
2060

2061
2062
2063
2064
2065
2066
2067
2068
2069
2070
2071
2072

def _get_req_paddings(min_req_size: int, max_req_size: int) -> list[int]:
    logger.info("Preparing request paddings:")
    # assert min_req_size is power of 2
    assert (min_req_size & (min_req_size - 1) == 0) and min_req_size > 0
    paddings: list = []
    num = max(MIN_NUM_SEQS, min_req_size)
    while num <= max_req_size and (len(paddings) == 0 or paddings[-1] != num):
        paddings.append(num)
        logger.info("    %d", num)
        num = _get_padded_num_reqs_with_upper_limit(num + 1, max_req_size)
    return paddings
2073
2074


2075
def _get_padded_num_reqs_with_upper_limit(x: int, upper_limit: int) -> int:
2076
    res = MIN_NUM_SEQS if x <= MIN_NUM_SEQS else 1 << (x - 1).bit_length()
2077
    return min(res, upper_limit)
2078
2079


2080
2081
2082
2083
def _get_token_paddings(
    min_token_size: int, max_token_size: int, padding_gap: int
) -> list[int]:
    """Generate a list of padding size, starting from min_token_size,
2084
    ending with a number that can cover max_token_size
2085

2086
2087
2088
    If padding_gap == 0 then:
        increase 2X each time (exponential)
    else:
2089
        first increase the size to twice,
2090
        then increase the padding size by padding_gap.
2091
    """
2092
2093
    # assert min_token_size is power of 2
    assert (min_token_size & (min_token_size - 1) == 0) and min_token_size > 0
2094
2095
    paddings = []
    num = min_token_size
2096
2097

    if padding_gap == 0:
2098
        logger.info("Using exponential token paddings:")
2099
        while True:
2100
2101
            logger.info("    %d", num)
            paddings.append(num)
2102
2103
            if num >= max_token_size:
                break
2104
2105
            num *= 2
    else:
2106
        logger.info("Using incremental token paddings:")
2107
2108
2109
2110
2111
2112
2113
2114
2115
2116
        while num <= padding_gap:
            logger.info("    %d", num)
            paddings.append(num)
            num *= 2
        num //= 2
        while num < max_token_size:
            num += padding_gap
            logger.info("    %d", num)
            paddings.append(num)

2117
2118
2119
2120
    return paddings


def _get_padded_token_len(paddings: list[int], x: int) -> int:
2121
    """Return the first element in paddings list greater or equal to x."""
2122
2123
2124
    index = bisect.bisect_left(paddings, x)
    assert index < len(paddings)
    return paddings[index]
2125
2126


2127
2128
2129
def _get_padded_num_kv_cache_update_slices(
    num_tokens: int, max_num_reqs: int, page_size: int
) -> int:
2130
2131
    """Calculates the padded number of KV cache update slices to avoid
    recompilation."""
2132
2133
2134
2135
2136
    # NOTE(chengjiyao): let's say R_i is the token num for i-th request,
    # so it occupies most 2 + R_i // page_size pages. The total maximum
    # possible number of pages needed is sum(2 + R_i // page_size), which
    # is <= 2 * max_num_reqs + sum(R_i) // page_size
    # = 2 * max_num_reqs + num_tokens // page_size
2137
2138
2139
2140
2141
    padded_num_slices = 2 * max_num_reqs + num_tokens // page_size
    padded_num_slices = min(padded_num_slices, num_tokens)
    return padded_num_slices


2142
2143
2144
2145
2146
2147
2148
2149
2150
2151
2152
def _get_num_slices_per_kv_cache_update_block(page_size_bytes: int) -> int:
    """Find the optimum number of slices to copy per Pallas program instance.

    Increasing the number of slices copied in one instance of the kernel program
    will increase HBM bandwidth utilization via more in-flight DMAs.

    However, it will also use more VMEM, and experimentally, we observed
    performance regression at 128 slices on v6e, likely due to running
    out of scalar registers. Thus this function will limit the number of
    slices to 64.
    """
2153
2154
2155
    # The default vmem_limit_bytes of a pallas kernel is 32MB. Here we
    # calculate num_slices_per_block based on 16MB in case any register spills.
    vmem_limit = 16 * 1024 * 1024
2156
2157
2158
2159
2160
2161
2162
2163
    num_slices_per_block = vmem_limit // page_size_bytes
    assert num_slices_per_block > 0, "Number of slices should be positive"
    num_slices_per_block = prev_power_of_2(num_slices_per_block)
    if num_slices_per_block > 64:
        num_slices_per_block = 64
    return num_slices_per_block


2164
2165
2166
2167
2168
2169
def replace_set_lora(model):
    def _tpu_set_lora(
        self,
        index: int,
        lora_a: torch.Tensor,
        lora_b: torch.Tensor,
2170
        embeddings_tensor: torch.Tensor | None,
2171
2172
2173
2174
    ):
        # TODO: The integer index leads to a recompilation, but converting it
        # to a tensor doesn't seem to work anymore. This might be fixed with a
        # later release of torch_xla.
2175
        self._original_set_lora(index, lora_a, lora_b, embeddings_tensor)
2176
        torch_xla.sync(wait=False)
2177
2178
2179

    def _tpu_reset_lora(self, index: int):
        self._original_reset_lora(index)
2180
        torch_xla.sync(wait=False)
2181
2182
2183
2184
2185

    for _, module in model.named_modules():
        if isinstance(module, BaseLayerWithLoRA):
            module._original_set_lora = module.set_lora
            module._original_reset_lora = module.reset_lora
2186
2187
2188
2189
2190
2191
            module.set_lora = _tpu_set_lora.__get__(  # type: ignore[method-assign]
                module, module.__class__
            )
            module.reset_lora = _tpu_reset_lora.__get__(  # type: ignore[method-assign]
                module, module.__class__
            )