| 1 | # SPDX-License-Identifier: Apache-2.0 |
| 2 | # SPDX-FileCopyrightText: Copyright contributors to the vLLM project |
| 3 | from collections.abc import Callable, Iterable |
| 4 | from dataclasses import dataclass |
| 5 | from typing import Any |
| 6 | from unittest.mock import MagicMock |
| 7 | |
| 8 | import pytest |
| 9 | import torch |
| 10 | |
| 11 | from tests.v1.kv_connector.unit.utils import ( |
| 12 | EOS_TOKEN_ID, |
| 13 | create_model_runner_output, |
| 14 | create_vllm_config, |
| 15 | ) |
| 16 | from vllm import SamplingParams |
| 17 | from vllm.config import ( |
| 18 | KVEventsConfig, |
| 19 | KVTransferConfig, |
| 20 | set_current_vllm_config, |
| 21 | ) |
| 22 | from vllm.distributed.kv_transfer.kv_connector.v1 import KVConnectorRole |
| 23 | from vllm.distributed.kv_transfer.kv_connector.v1.offloading.common import ( |
| 24 | OffloadingConnectorMetadata, |
| 25 | OffloadingWorkerMetadata, |
| 26 | ) |
| 27 | from vllm.distributed.kv_transfer.kv_connector.v1.offloading_connector import ( |
| 28 | OffloadingConnector, |
| 29 | ) |
| 30 | from vllm.forward_context import ForwardContext |
| 31 | from vllm.utils.hashing import sha256 |
| 32 | from vllm.v1.core.kv_cache_utils import ( |
| 33 | get_request_block_hasher, |
| 34 | init_none_hash, |
| 35 | resolve_kv_cache_block_sizes, |
| 36 | ) |
| 37 | from vllm.v1.core.sched.async_scheduler import AsyncScheduler |
| 38 | from vllm.v1.core.sched.scheduler import Scheduler |
| 39 | from vllm.v1.kv_cache_interface import ( |
| 40 | FullAttentionSpec, |
| 41 | KVCacheConfig, |
| 42 | KVCacheGroupSpec, |
| 43 | KVCacheTensor, |
| 44 | ) |
| 45 | from vllm.v1.kv_offload.base import ( |
| 46 | CanonicalKVCaches, |
| 47 | GPULoadStoreSpec, |
| 48 | LoadStoreSpec, |
| 49 | LookupResult, |
| 50 | OffloadingManager, |
| 51 | OffloadingSpec, |
| 52 | OffloadingWorker, |
| 53 | OffloadKey, |
| 54 | PrepareStoreOutput, |
| 55 | RequestOffloadingContext, |
| 56 | TransferResult, |
| 57 | make_offload_key, |
| 58 | ) |
| 59 | from vllm.v1.kv_offload.config import OffloadingConfig |
| 60 | from vllm.v1.request import Request |
| 61 | from vllm.v1.structured_output import StructuredOutputManager |
| 62 | |
| 63 | |
| 64 | def to_key(int_hash: int) -> OffloadKey: |
| 65 | return make_offload_key(str(int_hash).encode(), 0) |
| 66 | |
| 67 | |
| 68 | def to_keys(int_hashes: Iterable[int]) -> list[OffloadKey]: |
| 69 | return [to_key(i) for i in int_hashes] |
| 70 | |
| 71 | |
| 72 | class MockLoadStoreSpec(LoadStoreSpec): |
| 73 | def __init__(self, offload_keys: Iterable[OffloadKey]): |
| 74 | self.offload_keys: list[OffloadKey] = list(offload_keys) |
| 75 | |
| 76 | def __repr__(self) -> str: |
| 77 | return repr(self.offload_keys) |
| 78 | |
| 79 | |
| 80 | class MockOffloadingWorker(OffloadingWorker): |
| 81 | def __init__(self): |
| 82 | self.transfer_specs: dict[int, tuple[LoadStoreSpec, LoadStoreSpec]] = {} |
| 83 | self.completed_transfers: list[TransferResult] = [] |
| 84 | self.waiting_jobs: set[int] = set() |
| 85 | self.completed_jobs: list[int] = [] |
| 86 | self.flushed_jobs: set[int] = set() |
| 87 | |
| 88 | def get_finished(self) -> list[TransferResult]: |
| 89 | finished = self.completed_transfers |
| 90 | self.completed_transfers = [] |
| 91 | return finished |
| 92 | |
| 93 | def submit_store( |
| 94 | self, job_id: int, src_spec: LoadStoreSpec, dst_spec: LoadStoreSpec |
| 95 | ) -> bool: # type: ignore[override] |
| 96 | self.transfer_specs[job_id] = (src_spec, dst_spec) |
| 97 | self.waiting_jobs.add(job_id) |
| 98 | return True |
| 99 | |
| 100 | def submit_load( |
| 101 | self, job_id: int, src_spec: LoadStoreSpec, dst_spec: LoadStoreSpec |
| 102 | ) -> bool: # type: ignore[override] |
| 103 | self.transfer_specs[job_id] = (src_spec, dst_spec) |
| 104 | self.waiting_jobs.add(job_id) |
| 105 | return True |
| 106 | |
| 107 | def complete_jobs(self, job_ids: set[int]) -> None: |
| 108 | for job_id in job_ids: |
| 109 | if job_id in self.waiting_jobs: |
| 110 | self.waiting_jobs.remove(job_id) |
| 111 | self.completed_jobs.append(job_id) |
| 112 | result = TransferResult( |
| 113 | job_id=job_id, |
| 114 | success=True, |
| 115 | transfer_size=None, |
| 116 | transfer_time=None, |
| 117 | ) |
| 118 | self.completed_transfers.append(result) |
| 119 | |
| 120 | def wait(self, job_ids: set[int]) -> None: |
| 121 | self.flushed_jobs |= job_ids |
| 122 | self.complete_jobs(job_ids) |
| 123 | |
| 124 | |
| 125 | class MockOffloadingSpec(OffloadingSpec): |
| 126 | def __init__(self, config: OffloadingConfig): |
| 127 | super().__init__(config) |
| 128 | |
| 129 | self.manager = MagicMock(spec=OffloadingManager) |
| 130 | self.manager.prepare_load = lambda keys, req_context: MockLoadStoreSpec(keys) |
| 131 | self.manager.lookup.return_value = LookupResult.MISS |
| 132 | self.manager.get_stats.return_value = None |
| 133 | self.manager.on_new_request.return_value = RequestOffloadingContext() |
| 134 | self.handler = MockOffloadingWorker() |
| 135 | |
| 136 | def get_manager(self) -> OffloadingManager: |
| 137 | return self.manager |
| 138 | |
| 139 | def get_worker(self, _: CanonicalKVCaches) -> OffloadingWorker: |
| 140 | return self.handler |
| 141 | |
| 142 | def complete_transfers(self): |
| 143 | self.handler.complete_jobs(self.handler.waiting_jobs.copy()) |
| 144 | |
| 145 | def get_completed_transfers(self) -> list[tuple[LoadStoreSpec, LoadStoreSpec]]: |
| 146 | specs = [ |
| 147 | self.handler.transfer_specs[job_id] |
| 148 | for job_id in self.handler.completed_jobs |
| 149 | ] |
| 150 | self.handler.completed_jobs.clear() |
| 151 | return specs |
| 152 | |
| 153 | def get_flushed_transfers(self) -> list[tuple[LoadStoreSpec, LoadStoreSpec]]: |
| 154 | specs = [ |
| 155 | self.handler.transfer_specs[job_id] for job_id in self.handler.flushed_jobs |
| 156 | ] |
| 157 | self.handler.flushed_jobs.clear() |
| 158 | return specs |
| 159 | |
| 160 | |
| 161 | @dataclass(frozen=True) |
| 162 | class GPUBlock: |
| 163 | group_idx: int |
| 164 | request_block_offset: int |
| 165 | |
| 166 | |
| 167 | @dataclass |
| 168 | class TransferSummary: |
| 169 | gpu_blocks: list[GPUBlock] |
| 170 | offload_addresses: list[Any] |
| 171 | |
| 172 | |
| 173 | class RequestRunner: |
| 174 | def __init__( |
| 175 | self, |
| 176 | block_size: int, |
| 177 | num_gpu_blocks: int, |
| 178 | blocks_per_chunk: int = 1, |
| 179 | async_scheduling: bool = True, |
| 180 | kv_cache_groups: list[KVCacheGroupSpec] | None = None, |
| 181 | extra_config_overrides: dict[str, Any] | None = None, |
| 182 | worker_count: int = 1, |
| 183 | retention_interval: int | None = None, |
| 184 | speculative_config: Any | None = None, |
| 185 | ): |
| 186 | assert blocks_per_chunk == 1 or kv_cache_groups is None, ( |
| 187 | "blocks_per_chunk > 1 requires all groups to have the same " |
| 188 | "block size, so kv_cache_groups must be None (use default group)" |
| 189 | ) |
| 190 | |
| 191 | self.blocks_per_chunk: int = blocks_per_chunk |
| 192 | self.block_size: int = block_size |
| 193 | self.num_gpu_blocks: int = num_gpu_blocks |
| 194 | self.async_scheduling: bool = async_scheduling |
| 195 | |
| 196 | self.req_id: int = -1 |
| 197 | |
| 198 | vllm_config = create_vllm_config( |
| 199 | block_size=block_size, |
| 200 | max_num_batched_tokens=1000, |
| 201 | disable_hybrid_kv_cache_manager=False, |
| 202 | ) |
| 203 | vllm_config.scheduler_config.async_scheduling = async_scheduling |
| 204 | vllm_config.parallel_config.world_size = worker_count |
| 205 | vllm_config.cache_config.prefix_cache_retention_interval = retention_interval |
| 206 | if speculative_config is not None: |
| 207 | vllm_config.speculative_config = speculative_config |
| 208 | |
| 209 | extra_config: dict[str, Any] = { |
| 210 | "spec_name": "MockOffloadingSpec", |
| 211 | "spec_module_path": "tests.v1.kv_connector.unit.offloading_connector.utils", # noqa: E501 |
| 212 | # Preserve legacy behavior for tests; new opt-in tests override. |
| 213 | "offload_prompt_only": False, |
| 214 | # Exercise the self-describing KV events path by default; |
| 215 | # opt-out tests override this to cover the legacy placeholders. |
| 216 | "self_describing_kv_events": True, |
| 217 | } |
| 218 | if blocks_per_chunk > 1: |
| 219 | extra_config["block_size"] = block_size * blocks_per_chunk |
| 220 | if extra_config_overrides: |
| 221 | extra_config.update(extra_config_overrides) |
| 222 | |
| 223 | vllm_config.kv_transfer_config = KVTransferConfig( |
| 224 | kv_connector="OffloadingConnector", |
| 225 | kv_role="kv_both", |
| 226 | kv_connector_extra_config=extra_config, |
| 227 | ) |
| 228 | vllm_config.kv_events_config = KVEventsConfig( |
| 229 | # Enable so the offloading events tracker is active, but use the |
| 230 | # null publisher: these tests drain take_events directly and a |
| 231 | # real ZMQ publisher would bind a port per test. |
| 232 | enable_kv_cache_events=True, |
| 233 | publisher="null", |
| 234 | ) |
| 235 | |
| 236 | if kv_cache_groups is None: |
| 237 | kv_cache_groups = [ |
| 238 | KVCacheGroupSpec( |
| 239 | ["layer"], |
| 240 | FullAttentionSpec( |
| 241 | block_size=block_size, |
| 242 | num_kv_heads=1, |
| 243 | head_size=1, |
| 244 | dtype=torch.float32, |
| 245 | ), |
| 246 | ) |
| 247 | ] |
| 248 | |
| 249 | # Groups overlay one allocation, sized by the largest group. |
| 250 | block_stride = max( |
| 251 | group.kv_cache_spec.page_size_bytes * len(group.layer_names) |
| 252 | for group in kv_cache_groups |
| 253 | ) |
| 254 | kv_cache_tensors = [ |
| 255 | KVCacheTensor( |
| 256 | size=block_stride * num_gpu_blocks, |
| 257 | layers=list(group.layer_names), |
| 258 | layer_stride=group.kv_cache_spec.page_size_bytes * num_gpu_blocks, |
| 259 | block_stride=group.kv_cache_spec.page_size_bytes, |
| 260 | ) |
| 261 | for group in kv_cache_groups |
| 262 | ] |
| 263 | |
| 264 | kv_cache_config = KVCacheConfig( |
| 265 | num_blocks=num_gpu_blocks, |
| 266 | kv_cache_tensors=kv_cache_tensors, |
| 267 | kv_cache_groups=kv_cache_groups, |
| 268 | ) |
| 269 | vllm_config.cache_config.num_gpu_blocks = num_gpu_blocks |
| 270 | self.num_kv_groups = len(kv_cache_config.kv_cache_groups) |
| 271 | |
| 272 | scheduler_block_size, hash_block_size = resolve_kv_cache_block_sizes( |
| 273 | kv_cache_config, vllm_config |
| 274 | ) |
| 275 | |
| 276 | scheduler_cls = AsyncScheduler if async_scheduling else Scheduler |
| 277 | self.scheduler = scheduler_cls( |
| 278 | vllm_config=vllm_config, |
| 279 | kv_cache_config=kv_cache_config, |
| 280 | log_stats=True, |
| 281 | structured_output_manager=StructuredOutputManager(vllm_config), |
| 282 | block_size=scheduler_block_size, |
| 283 | hash_block_size=hash_block_size, |
| 284 | ) |
| 285 | |
| 286 | self.worker_connector = OffloadingConnector( |
| 287 | vllm_config, KVConnectorRole.WORKER, kv_cache_config |
| 288 | ) |
| 289 | |
| 290 | # register worker kv_caches to enable OffloadingWorker creations |
| 291 | kv_caches: dict[str, torch.Tensor] = {} |
| 292 | for group in kv_cache_groups: |
| 293 | spec = group.kv_cache_spec |
| 294 | for layer_name in group.layer_names: |
| 295 | # Shape follows FlashAttention layout: |
| 296 | # Shape: (num_blocks, 2, block_size, num_kv_heads, head_size) |
| 297 | kv_caches[layer_name] = torch.empty( |
| 298 | num_gpu_blocks, |
| 299 | 2, |
| 300 | spec.block_size, |
| 301 | spec.num_kv_heads, |
| 302 | spec.head_size, |
| 303 | dtype=spec.dtype, |
| 304 | ) |
| 305 | |
| 306 | with set_current_vllm_config(vllm_config): |
| 307 | self.worker_connector.register_kv_caches(kv_caches) |
| 308 | |
| 309 | # extract connector of scheduler |
| 310 | scheduler_connector = self.scheduler.connector |
| 311 | assert scheduler_connector is not None |
| 312 | assert isinstance(scheduler_connector, OffloadingConnector) |
| 313 | self.scheduler_connector: OffloadingConnector = scheduler_connector |
| 314 | |
| 315 | # extract mocked OffloadingManager of scheduler connector |
| 316 | self.connector_scheduler = scheduler_connector.connector_scheduler |
| 317 | assert self.connector_scheduler is not None |
| 318 | manager = self.connector_scheduler.manager |
| 319 | assert isinstance(manager, MagicMock) |
| 320 | self.manager: MagicMock = manager |
| 321 | |
| 322 | num_kv_groups = len(kv_cache_config.kv_cache_groups) |
| 323 | assert len(self.connector_scheduler.config.kv_group_configs) == num_kv_groups |
| 324 | for group_config, kv_cache_group in zip( |
| 325 | self.connector_scheduler.config.kv_group_configs, |
| 326 | kv_cache_config.kv_cache_groups, |
| 327 | ): |
| 328 | tokens_per_block = kv_cache_group.kv_cache_spec.block_size |
| 329 | assert group_config.tokens_per_block == tokens_per_block |
| 330 | assert group_config.tokens_per_chunk == tokens_per_block * blocks_per_chunk |
| 331 | |
| 332 | # extract OffloadingSpec of worker_connector |
| 333 | connector_worker = self.worker_connector.connector_worker |
| 334 | assert connector_worker is not None |
| 335 | offloading_spec = connector_worker.spec |
| 336 | assert isinstance(offloading_spec, MockOffloadingSpec) |
| 337 | self.offloading_spec: MockOffloadingSpec = offloading_spec |
| 338 | |
| 339 | # mapping (offloading address) -> GPUBlock |
| 340 | self.offloaded: dict[Any, GPUBlock] = {} |
| 341 | |
| 342 | self.completed_loads: list[TransferSummary] = [] |
| 343 | self.completed_stores: list[TransferSummary] = [] |
| 344 | self.flushed_gpu_blocks: set[GPUBlock] = set() |
| 345 | self.kv_connector_stats: list[Any] = [] |
| 346 | |
| 347 | # block_id -> GPUBlock |
| 348 | self.gpu_blocks: dict[int, GPUBlock] = {} |
| 349 | |
| 350 | init_none_hash(sha256) |
| 351 | self._block_hasher = get_request_block_hasher(block_size, sha256) |
| 352 | |
| 353 | self._dummy_ctx: ForwardContext = ForwardContext( |
| 354 | no_compile_layers={}, |
| 355 | attn_metadata={}, |
| 356 | slot_mapping={}, |
| 357 | ) |
| 358 | |
| 359 | def _record_kv_connector_stats(self, engine_outputs: dict[int, Any]) -> None: |
| 360 | for output in engine_outputs.values(): |
| 361 | scheduler_stats = output.scheduler_stats |
| 362 | if scheduler_stats is not None and scheduler_stats.kv_connector_stats: |
| 363 | self.kv_connector_stats.append(scheduler_stats.kv_connector_stats) |
| 364 | |
| 365 | def new_request( |
| 366 | self, |
| 367 | token_ids: list[int], |
| 368 | kv_transfer_params: dict | None = None, |
| 369 | skip_reading_prefix_cache: bool = False, |
| 370 | ): |
| 371 | self.req_id += 1 |
| 372 | |
| 373 | sampling_params = SamplingParams( |
| 374 | max_tokens=1000, |
| 375 | skip_reading_prefix_cache=skip_reading_prefix_cache or None, |
| 376 | ) |
| 377 | sampling_params.update_from_generation_config({}, EOS_TOKEN_ID) |
| 378 | |
| 379 | req = Request( |
| 380 | request_id=str(self.req_id), |
| 381 | prompt_token_ids=token_ids, |
| 382 | sampling_params=sampling_params, |
| 383 | pooling_params=None, |
| 384 | block_hasher=self._block_hasher, |
| 385 | ) |
| 386 | if kv_transfer_params is not None: |
| 387 | req.kv_transfer_params = kv_transfer_params |
| 388 | |
| 389 | self.scheduler.add_request(req) |
| 390 | |
| 391 | def _parse_transfers(self): |
| 392 | for src_spec, dst_spec in self.offloading_spec.get_flushed_transfers(): |
| 393 | if isinstance(src_spec, GPULoadStoreSpec): |
| 394 | # store flush |
| 395 | for block_id in src_spec.block_ids: |
| 396 | self.flushed_gpu_blocks.add(self.gpu_blocks[block_id.item()]) |
| 397 | else: |
| 398 | # load flush |
| 399 | for block_id in dst_spec.block_ids: |
| 400 | self.flushed_gpu_blocks.add(self.gpu_blocks[block_id.item()]) |
| 401 | |
| 402 | blocks_per_chunk = self.blocks_per_chunk |
| 403 | |
| 404 | for src_spec, dst_spec in self.offloading_spec.get_completed_transfers(): |
| 405 | if isinstance(src_spec, GPULoadStoreSpec): |
| 406 | store = True |
| 407 | gpu_spec = src_spec |
| 408 | offload_spec = dst_spec |
| 409 | else: |
| 410 | store = False |
| 411 | gpu_spec = dst_spec |
| 412 | offload_spec = src_spec |
| 413 | |
| 414 | assert isinstance(offload_spec, MockLoadStoreSpec) |
| 415 | assert isinstance(gpu_spec, GPULoadStoreSpec) |
| 416 | assert len(gpu_spec.group_sizes) == self.num_kv_groups |
| 417 | |
| 418 | gpu_blocks: list[GPUBlock] = [] |
| 419 | for block_id in gpu_spec.block_ids: |
| 420 | gpu_blocks.append(self.gpu_blocks[block_id.item()]) |
| 421 | |
| 422 | # list of (offload_key, sub_block_offset) |
| 423 | offload_addresses: list[Any] = [] |
| 424 | for offload_key in offload_spec.offload_keys: |
| 425 | for sub_block_idx in range(blocks_per_chunk): |
| 426 | offload_addresses.append((offload_key, sub_block_idx)) |
| 427 | |
| 428 | assert gpu_spec.block_indices is not None |
| 429 | assert len(gpu_spec.block_indices) == self.num_kv_groups |
| 430 | |
| 431 | gpu_block_offset = 0 |
| 432 | offload_address_offset = 0 |
| 433 | for group_size, logical_offset in zip( |
| 434 | gpu_spec.group_sizes, gpu_spec.block_indices |
| 435 | ): |
| 436 | gpu_block_end_offset = gpu_block_offset + group_size |
| 437 | assert gpu_block_end_offset <= len(gpu_blocks) |
| 438 | |
| 439 | offload_addresses_to_skip = logical_offset % blocks_per_chunk |
| 440 | offload_addresses_end_offset = ( |
| 441 | offload_address_offset + offload_addresses_to_skip + group_size |
| 442 | ) |
| 443 | assert offload_addresses_end_offset <= len(offload_addresses) |
| 444 | |
| 445 | offload_addresses = ( |
| 446 | offload_addresses[:offload_address_offset] |
| 447 | + offload_addresses[ |
| 448 | offload_address_offset + offload_addresses_to_skip : |
| 449 | ] |
| 450 | ) |
| 451 | |
| 452 | gpu_block_offset += group_size |
| 453 | offload_address_offset += group_size |
| 454 | |
| 455 | assert gpu_block_offset == len(gpu_blocks) |
| 456 | assert offload_address_offset == len(offload_addresses) |
| 457 | |
| 458 | transfer_summary = TransferSummary(gpu_blocks, offload_addresses) |
| 459 | if store: |
| 460 | self.completed_stores.append(transfer_summary) |
| 461 | else: |
| 462 | self.completed_loads.append(transfer_summary) |
| 463 | |
| 464 | def _update_gpu_blocks(self): |
| 465 | for group_idx, manager in enumerate( |
| 466 | self.scheduler.kv_cache_manager.coordinator.single_type_managers |
| 467 | ): |
| 468 | for blocks in manager.req_to_blocks.values(): |
| 469 | for block_idx, block in enumerate(blocks): |
| 470 | self.gpu_blocks[block.block_id] = GPUBlock(group_idx, block_idx) |
| 471 | |
| 472 | def _run( |
| 473 | self, |
| 474 | decoded_tokens: list[int], |
| 475 | complete_transfers: bool, |
| 476 | post_step_fn: Callable[[], None] | None = None, |
| 477 | ): |
| 478 | """Runs multiple engine (scheduler + worker) steps. |
| 479 | Assumes a single request is running. |
| 480 | |
| 481 | Args: |
| 482 | decoded_tokens: the tokens to yield at each step. |
| 483 | complete_transfers: complete transfers immediately |
| 484 | post_step_fn: optional callback invoked after each step's |
| 485 | update_from_output(), before the next schedule(). |
| 486 | |
| 487 | """ |
| 488 | tokens_iter = iter(decoded_tokens) |
| 489 | token_id = next(tokens_iter, None) |
| 490 | prev_scheduler_output = None |
| 491 | prev_model_runner_output = None |
| 492 | while True: |
| 493 | # Strict-always-False frees the request immediately on EOS, but |
| 494 | # the worker may still have a deferred store queued. In production |
| 495 | # the next request's step drains it; in single-request tests we |
| 496 | # must keep stepping until the scheduler sees no in-flight jobs |
| 497 | # and no pending finished_req_ids awaiting build_connector_meta. |
| 498 | if ( |
| 499 | not self.scheduler.requests |
| 500 | and not self.connector_scheduler._jobs |
| 501 | and not self.scheduler.finished_req_ids |
| 502 | ): |
| 503 | break |
| 504 | |
| 505 | scheduler_output = self.scheduler.schedule() |
| 506 | self._update_gpu_blocks() |
| 507 | |
| 508 | kv_connector_metadata = scheduler_output.kv_connector_metadata |
| 509 | assert kv_connector_metadata is not None |
| 510 | assert isinstance(kv_connector_metadata, OffloadingConnectorMetadata) |
| 511 | |
| 512 | self.worker_connector.handle_preemptions(kv_connector_metadata) |
| 513 | |
| 514 | self.worker_connector.bind_connector_metadata(kv_connector_metadata) |
| 515 | self.worker_connector.start_load_kv(self._dummy_ctx) |
| 516 | |
| 517 | if complete_transfers: |
| 518 | self.offloading_spec.complete_transfers() |
| 519 | |
| 520 | self.worker_connector.wait_for_save() |
| 521 | finished_sending, finished_recving = self.worker_connector.get_finished( |
| 522 | scheduler_output.finished_req_ids |
| 523 | ) |
| 524 | worker_meta = ( |
| 525 | self.worker_connector.build_connector_worker_meta() |
| 526 | or OffloadingWorkerMetadata() |
| 527 | ) |
| 528 | |
| 529 | self.worker_connector.clear_connector_metadata() |
| 530 | |
| 531 | model_runner_output = create_model_runner_output( |
| 532 | reqs=self.scheduler.running, |
| 533 | finished_sending=finished_sending, |
| 534 | finished_recving=finished_recving, |
| 535 | token_id=token_id or 0, |
| 536 | kv_connector_worker_meta=worker_meta, |
| 537 | ) |
| 538 | |
| 539 | prev_token_id = token_id |
| 540 | if self.scheduler.running: |
| 541 | token_id = next(tokens_iter, None) |
| 542 | |
| 543 | if self.async_scheduling: |
| 544 | # in async scheduling we update the output of the previous step |
| 545 | if prev_model_runner_output is not None: |
| 546 | engine_outputs = self.scheduler.update_from_output( |
| 547 | prev_scheduler_output, prev_model_runner_output |
| 548 | ) |
| 549 | self._record_kv_connector_stats(engine_outputs) |
| 550 | prev_scheduler_output = scheduler_output |
| 551 | prev_model_runner_output = model_runner_output |
| 552 | else: |
| 553 | engine_outputs = self.scheduler.update_from_output( |
| 554 | scheduler_output, model_runner_output |
| 555 | ) |
| 556 | self._record_kv_connector_stats(engine_outputs) |
| 557 | |
| 558 | if post_step_fn is not None: |
| 559 | post_step_fn() |
| 560 | |
| 561 | if ( |
| 562 | prev_token_id == EOS_TOKEN_ID |
| 563 | and prev_token_id != token_id |
| 564 | and ( |
| 565 | self.scheduler.requests |
| 566 | or self.connector_scheduler._jobs |
| 567 | or self.scheduler.finished_req_ids |
| 568 | ) |
| 569 | ): |
| 570 | # continue for one more step to allow offloading to kick off |
| 571 | continue |
| 572 | |
| 573 | if token_id is None: |
| 574 | if self.async_scheduling: |
| 575 | # Flush the previous step's output. |
| 576 | engine_outputs = self.scheduler.update_from_output( |
| 577 | prev_scheduler_output, prev_model_runner_output |
| 578 | ) |
| 579 | self._record_kv_connector_stats(engine_outputs) |
| 580 | prev_model_runner_output = None |
| 581 | if self.scheduler.requests: |
| 582 | # Request still running, just exhausted decoded_tokens. |
| 583 | break |
| 584 | if not self.scheduler.finished_req_ids and ( |
| 585 | not complete_transfers or not self.connector_scheduler._jobs |
| 586 | ): |
| 587 | break |
| 588 | |
| 589 | self._parse_transfers() |
| 590 | |
| 591 | if EOS_TOKEN_ID in decoded_tokens: |
| 592 | assert not self.scheduler.running |
| 593 | |
| 594 | def _to_gpu_blocks( |
| 595 | self, blocks: tuple[int | tuple[int, int], ...] |
| 596 | ) -> list[GPUBlock]: |
| 597 | gpu_blocks: list[GPUBlock] = [] |
| 598 | for block in blocks: |
| 599 | if isinstance(block, int): |
| 600 | for group_idx in range(self.num_kv_groups): |
| 601 | gpu_blocks.append( |
| 602 | GPUBlock(group_idx=group_idx, request_block_offset=block) |
| 603 | ) |
| 604 | else: |
| 605 | group_idx, offset = block |
| 606 | gpu_blocks.append( |
| 607 | GPUBlock(group_idx=group_idx, request_block_offset=offset) |
| 608 | ) |
| 609 | return gpu_blocks |
| 610 | |
| 611 | def run( |
| 612 | self, |
| 613 | decoded_tokens: list[int], |
| 614 | complete_transfers: bool = True, |
| 615 | expected_stored: tuple[int | tuple[int, int], ...] = (), |
| 616 | expected_loaded: tuple[int | tuple[int, int], ...] = (), |
| 617 | expected_flushed: tuple[int | tuple[int, int], ...] = (), |
| 618 | post_step_fn: Callable[[], None] | None = None, |
| 619 | ): |
| 620 | """Runs multiple engine (scheduler + worker) steps. |
| 621 | Assumes a single request is running. |
| 622 | |
| 623 | Args: |
| 624 | decoded_tokens: the tokens to yield at each step. |
| 625 | complete_transfers: complete transfers immediately |
| 626 | expected_stored: GPU blocks |
| 627 | that are expected to be written during the run. |
| 628 | expected_loaded: GPU blocks |
| 629 | that are expected to be loaded during the run. |
| 630 | expected_flushed: GPU blocks |
| 631 | that are expected to be flushed during the run. |
| 632 | |
| 633 | A GPU block is either a (group_idx: int, request_block_offset: int) |
| 634 | or just request_block_offset: int. |
| 635 | The latter case is a convenience for representing all groups. |
| 636 | |
| 637 | """ |
| 638 | expected_stored_gpu_blocks = self._to_gpu_blocks(expected_stored) |
| 639 | expected_loaded_gpu_blocks = self._to_gpu_blocks(expected_loaded) |
| 640 | expected_flushed_gpu_blocks = self._to_gpu_blocks(expected_flushed) |
| 641 | |
| 642 | self.manager.reset_mock() |
| 643 | self._run(decoded_tokens, complete_transfers, post_step_fn=post_step_fn) |
| 644 | |
| 645 | loaded_gpu_blocks: set[GPUBlock] = set() |
| 646 | for transfer in self.completed_loads: |
| 647 | for gpu_block, offloaded_address in zip( |
| 648 | transfer.gpu_blocks, transfer.offload_addresses |
| 649 | ): |
| 650 | loaded_gpu_blocks.add(gpu_block) |
| 651 | assert gpu_block == self.offloaded[offloaded_address] |
| 652 | |
| 653 | assert set(expected_loaded_gpu_blocks) == loaded_gpu_blocks |
| 654 | self.completed_loads.clear() |
| 655 | |
| 656 | stored_gpu_blocks: set[GPUBlock] = set() |
| 657 | for transfer in self.completed_stores: |
| 658 | for gpu_block, offloaded_address in zip( |
| 659 | transfer.gpu_blocks, transfer.offload_addresses |
| 660 | ): |
| 661 | stored_gpu_blocks.add(gpu_block) |
| 662 | self.offloaded[offloaded_address] = gpu_block |
| 663 | |
| 664 | assert set(expected_stored_gpu_blocks) == stored_gpu_blocks |
| 665 | self.completed_stores.clear() |
| 666 | |
| 667 | assert set(expected_flushed_gpu_blocks) == self.flushed_gpu_blocks |
| 668 | self.flushed_gpu_blocks.clear() |
| 669 | |
| 670 | |
| 671 | @pytest.fixture |
| 672 | def request_runner(): |
| 673 | runners = [] |
| 674 | |
| 675 | def runner_factory( |
| 676 | block_size, |
| 677 | num_gpu_blocks, |
| 678 | async_scheduling, |
| 679 | blocks_per_chunk=1, |
| 680 | kv_cache_groups=None, |
| 681 | extra_config_overrides=None, |
| 682 | worker_count=1, |
| 683 | retention_interval=None, |
| 684 | speculative_config=None, |
| 685 | ): |
| 686 | runner = RequestRunner( |
| 687 | block_size=block_size, |
| 688 | num_gpu_blocks=num_gpu_blocks, |
| 689 | blocks_per_chunk=blocks_per_chunk, |
| 690 | async_scheduling=async_scheduling, |
| 691 | kv_cache_groups=kv_cache_groups, |
| 692 | extra_config_overrides=extra_config_overrides, |
| 693 | worker_count=worker_count, |
| 694 | retention_interval=retention_interval, |
| 695 | speculative_config=speculative_config, |
| 696 | ) |
| 697 | runners.append(runner) |
| 698 | return runner |
| 699 | |
| 700 | yield runner_factory # pass factory to the test |
| 701 | |
| 702 | |
| 703 | def generate_store_output(keys: Iterable[OffloadKey]): |
| 704 | keys = list(keys) |
| 705 | return PrepareStoreOutput( |
| 706 | keys_to_store=list(keys), |
| 707 | store_spec=MockLoadStoreSpec(keys), |
| 708 | evicted_keys=[], |
| 709 | ) |