| 1 | # SPDX-License-Identifier: Apache-2.0 |
| 2 | # SPDX-FileCopyrightText: Copyright contributors to the vLLM project |
| 3 | """Worker-side unit tests for SimpleCPUOffloadConnector. |
| 4 | |
| 5 | Covers the GPU->CPU store cross-stream synchronization: the store copy must be |
| 6 | ordered after the compute stream that writes the KV blocks, otherwise it can |
| 7 | read partially written / stale blocks and silently corrupt the CPU cache. |
| 8 | """ |
| 9 | |
| 10 | from __future__ import annotations |
| 11 | |
| 12 | import time |
| 13 | from contextlib import nullcontext |
| 14 | from unittest.mock import MagicMock |
| 15 | |
| 16 | import pytest |
| 17 | import torch |
| 18 | |
| 19 | from vllm.platforms import current_platform |
| 20 | |
| 21 | if not current_platform.is_cuda_alike(): |
| 22 | pytest.skip("Requires CUDA or ROCm", allow_module_level=True) |
| 23 | |
| 24 | from tests.v1.attention.utils import dense_kv_cache_tensor, dense_kv_cache_views |
| 25 | from tests.v1.kv_connector.unit.test_kv_connector_lifecycle import ( |
| 26 | _make_empty_scheduler_output, |
| 27 | ) |
| 28 | from vllm.config import CacheConfig |
| 29 | from vllm.distributed.kv_transfer.kv_connector.v1.simple_cpu_offload_connector import ( |
| 30 | SimpleCPUOffloadConnector, |
| 31 | ) |
| 32 | from vllm.v1.core.kv_cache_utils import ( |
| 33 | get_kv_cache_config_from_groups, |
| 34 | is_kv_cache_spec_uniform, |
| 35 | resolve_kv_cache_block_sizes, |
| 36 | ) |
| 37 | from vllm.v1.kv_cache_interface import ( |
| 38 | FullAttentionSpec, |
| 39 | KVCacheConfig, |
| 40 | KVCacheGroupSpec, |
| 41 | KVCacheLayout, |
| 42 | KVCacheSpec, |
| 43 | KVCacheTensor, |
| 44 | MLAAttentionSpec, |
| 45 | UniformTypeKVCacheSpecs, |
| 46 | ) |
| 47 | from vllm.v1.simple_kv_offload.copy_backend import DmaCopyBackend |
| 48 | from vllm.v1.simple_kv_offload.cuda_mem_ops import ( |
| 49 | CU_MEMCPY_SRC_ACCESS_ORDER_ANY, |
| 50 | CU_MEMCPY_SRC_ACCESS_ORDER_STREAM, |
| 51 | build_params, |
| 52 | ) |
| 53 | from vllm.v1.simple_kv_offload.disk_backend import DiskBackend |
| 54 | from vllm.v1.simple_kv_offload.metadata import ( |
| 55 | SimpleCPUOffloadMetadata, |
| 56 | SimpleCPUOffloadWorkerMetadata, |
| 57 | ) |
| 58 | from vllm.v1.simple_kv_offload.worker import SimpleCPUOffloadWorker |
| 59 | from vllm.v1.worker.gpu.kv_connector import ActiveKVConnector |
| 60 | from vllm.v1.worker.kv_connector_model_runner_mixin import KVConnectorModelRunnerMixin |
| 61 | from vllm.v1.worker.utils import allocate_kv_cache |
| 62 | |
| 63 | NUM_BLOCKS = 64 |
| 64 | BLOCK_BYTES = 4096 |
| 65 | ITERS = 30 |
| 66 | # Keep the compute stream busy so the KV write lands late; this makes the |
| 67 | # store-vs-compute race deterministic instead of timing-dependent. |
| 68 | SLEEP_CYCLES = 50_000_000 |
| 69 | |
| 70 | |
| 71 | def _make_worker( |
| 72 | kv_cache_config: KVCacheConfig | None, cpu_capacity_bytes: int |
| 73 | ) -> SimpleCPUOffloadWorker: |
| 74 | return SimpleCPUOffloadWorker(None, kv_cache_config, cpu_capacity_bytes) |
| 75 | |
| 76 | |
| 77 | def _make_backend() -> tuple[DmaCopyBackend, torch.Tensor, torch.Tensor]: |
| 78 | gpu = {"k": torch.zeros((NUM_BLOCKS, BLOCK_BYTES), dtype=torch.int8, device="cuda")} |
| 79 | cpu = { |
| 80 | "k": torch.zeros( |
| 81 | (NUM_BLOCKS, BLOCK_BYTES), dtype=torch.int8, device="cpu", pin_memory=True |
| 82 | ) |
| 83 | } |
| 84 | low_pri, _ = torch.cuda.Stream.priority_range() |
| 85 | backend = DmaCopyBackend() |
| 86 | backend.init( |
| 87 | gpu, |
| 88 | cpu, |
| 89 | gpu["k"].device, |
| 90 | torch.cuda.Stream(priority=low_pri), |
| 91 | torch.cuda.Stream(priority=low_pri), |
| 92 | ) |
| 93 | return backend, gpu["k"], cpu["k"] |
| 94 | |
| 95 | |
| 96 | @pytest.mark.parametrize("use_v2", [False, True]) |
| 97 | def test_no_forward_step_completes_cpu_store(monkeypatch, use_v2): |
| 98 | """A finished request's final store must drain without another model step.""" |
| 99 | backend, gpu, cpu = _make_backend() |
| 100 | worker = _make_worker(None, 0) |
| 101 | worker._backend = backend |
| 102 | connector = SimpleCPUOffloadConnector.__new__(SimpleCPUOffloadConnector) |
| 103 | connector.worker_handler = worker |
| 104 | output = _make_empty_scheduler_output() |
| 105 | output.kv_connector_metadata = SimpleCPUOffloadMetadata( |
| 106 | store_event=0, store_gpu_blocks=[0], store_cpu_blocks=[0] |
| 107 | ) |
| 108 | module = "vllm.v1.worker.kv_connector_model_runner_mixin" |
| 109 | monkeypatch.setattr(f"{module}.get_kv_transfer_group", lambda: connector) |
| 110 | monkeypatch.setattr(f"{module}.set_forward_context", lambda *a: nullcontext()) |
| 111 | monkeypatch.setattr(f"{module}.get_forward_context", lambda: None) |
| 112 | try: |
| 113 | gpu.fill_(91) |
| 114 | if use_v2: |
| 115 | active = ActiveKVConnector.__new__(ActiveKVConnector) |
| 116 | active.kv_connector = connector |
| 117 | active._disabled = False |
| 118 | active._pending_load_kwargs = None |
| 119 | module = "vllm.v1.worker.gpu.kv_connector" |
| 120 | monkeypatch.setattr(f"{module}.is_forward_context_available", lambda: True) |
| 121 | monkeypatch.setattr(f"{module}.get_forward_context", lambda: None) |
| 122 | result = active.no_forward(output) |
| 123 | else: |
| 124 | result = KVConnectorModelRunnerMixin.kv_connector_no_forward(output, None) |
| 125 | completion = ( |
| 126 | result.kv_connector_output.kv_connector_worker_meta |
| 127 | if result.kv_connector_output is not None |
| 128 | else None |
| 129 | ) |
| 130 | deadline = time.monotonic() + 5 |
| 131 | while completion is None and time.monotonic() < deadline: |
| 132 | connector.get_finished(set()) |
| 133 | completion = connector.build_connector_worker_meta() |
| 134 | time.sleep(0.001) |
| 135 | assert completion is not None, "CPU store never completed on the empty step" |
| 136 | assert isinstance(completion, SimpleCPUOffloadWorkerMetadata) |
| 137 | assert completion.completed_store_events == {0: 1} |
| 138 | assert torch.equal(cpu[0], gpu[0].cpu()) |
| 139 | finally: |
| 140 | backend.shutdown() |
| 141 | |
| 142 | |
| 143 | def test_deferred_store_waits_for_draft_forward(monkeypatch): |
| 144 | """Target completion must not submit the store before draft finalization.""" |
| 145 | backend, gpu, cpu = _make_backend() |
| 146 | worker = _make_worker(None, 0) |
| 147 | worker._backend = backend |
| 148 | connector = SimpleCPUOffloadConnector.__new__(SimpleCPUOffloadConnector) |
| 149 | connector.worker_handler = worker |
| 150 | output = _make_empty_scheduler_output() |
| 151 | output.kv_connector_metadata = SimpleCPUOffloadMetadata( |
| 152 | store_event=0, store_gpu_blocks=[0], store_cpu_blocks=[0] |
| 153 | ) |
| 154 | module = "vllm.v1.worker.kv_connector_model_runner_mixin" |
| 155 | monkeypatch.setattr(f"{module}.get_kv_transfer_group", lambda: connector) |
| 156 | monkeypatch.setattr(f"{module}.has_kv_transfer_group", lambda: True) |
| 157 | monkeypatch.setattr(f"{module}.get_forward_context", lambda: None) |
| 158 | launch = MagicMock(wraps=backend.launch_copy) |
| 159 | monkeypatch.setattr(backend, "launch_copy", launch) |
| 160 | try: |
| 161 | with KVConnectorModelRunnerMixin._get_kv_connector_output( |
| 162 | output, defer_finalize=True |
| 163 | ): |
| 164 | gpu.fill_(17) |
| 165 | launch.assert_not_called() |
| 166 | gpu.fill_(91) |
| 167 | KVConnectorModelRunnerMixin.finalize_kv_connector() |
| 168 | assert launch.call_count == 1 |
| 169 | completion = None |
| 170 | deadline = time.monotonic() + 5 |
| 171 | while completion is None and time.monotonic() < deadline: |
| 172 | connector.get_finished(set()) |
| 173 | completion = connector.build_connector_worker_meta() |
| 174 | time.sleep(0.001) |
| 175 | assert completion is not None |
| 176 | assert completion.completed_store_events == {0: 1} |
| 177 | assert torch.equal(cpu[0], gpu[0].cpu()) |
| 178 | assert launch.call_count == 1 |
| 179 | finally: |
| 180 | backend.shutdown() |
| 181 | |
| 182 | |
| 183 | def _drive_store( |
| 184 | backend: DmaCopyBackend, |
| 185 | gpu: torch.Tensor, |
| 186 | cpu: torch.Tensor, |
| 187 | *, |
| 188 | with_barrier: bool, |
| 189 | ) -> int: |
| 190 | """Run ITERS store cycles; return how many landed corrupted in the CPU pool. |
| 191 | |
| 192 | Each cycle writes a unique value on a compute stream (after a deliberate |
| 193 | delay) and then issues the GPU->CPU store. The store is issued *after* the |
| 194 | write in host program order, mirroring the connector's deferred-store |
| 195 | assumption. Only the compute-done event creates a real device-side |
| 196 | happens-before edge. |
| 197 | """ |
| 198 | block_ids = list(range(gpu.shape[0])) |
| 199 | compute_stream = torch.cuda.Stream() |
| 200 | corrupt = 0 |
| 201 | for it in range(ITERS): |
| 202 | val = (it % 126) + 1 # 1..126; distinct from the zero-initialized pool |
| 203 | with torch.cuda.stream(compute_stream): |
| 204 | torch.cuda._sleep(SLEEP_CYCLES) |
| 205 | gpu.fill_(val) |
| 206 | |
| 207 | wait_event = None |
| 208 | if with_barrier: |
| 209 | wait_event = torch.Event() |
| 210 | wait_event.record(compute_stream) |
| 211 | |
| 212 | store_events: list[tuple[int, torch.Event]] = [] |
| 213 | backend.launch_copy( |
| 214 | block_ids, |
| 215 | block_ids, |
| 216 | is_store=True, |
| 217 | event_idx=it, |
| 218 | events_list=store_events, |
| 219 | wait_event=wait_event, |
| 220 | ) |
| 221 | |
| 222 | deadline = time.time() + 10.0 |
| 223 | while not store_events and time.time() < deadline: |
| 224 | time.sleep(0.0005) |
| 225 | assert store_events, "background copy was never enqueued" |
| 226 | store_events[0][1].synchronize() |
| 227 | |
| 228 | if int((cpu[:, 0].to(torch.int32) != val).sum().item()): |
| 229 | corrupt += 1 |
| 230 | |
| 231 | # Drain the compute stream before returning: in the no-barrier control |
| 232 | # phase the store never waits on compute, so the host loop runs far ahead |
| 233 | # and leaves a backlog of sleep+fill kernels in flight. Without this, the |
| 234 | # leftover control-phase fills race the barrier phase's fill->copy window |
| 235 | # on the shared gpu tensor and flakily corrupt one iteration. |
| 236 | compute_stream.synchronize() |
| 237 | return corrupt |
| 238 | |
| 239 | |
| 240 | def test_store_orders_after_compute_write(): |
| 241 | """The store must wait for the compute event; without it, it races. |
| 242 | |
| 243 | Asserts both directions so the test is self-validating: the no-barrier |
| 244 | control must actually corrupt (proving the race window is exercised), and |
| 245 | the fixed path with the compute-done event must be clean. |
| 246 | """ |
| 247 | backend, gpu, cpu = _make_backend() |
| 248 | try: |
| 249 | control = _drive_store(backend, gpu, cpu, with_barrier=False) |
| 250 | fixed = _drive_store(backend, gpu, cpu, with_barrier=True) |
| 251 | finally: |
| 252 | backend.shutdown() |
| 253 | |
| 254 | assert control > 0, ( |
| 255 | "no-barrier store did not race the compute write; the test no longer " |
| 256 | "exercises the hazard it is meant to guard" |
| 257 | ) |
| 258 | assert fixed == 0, f"store raced compute even with the barrier: {fixed} corrupt" |
| 259 | |
| 260 | |
| 261 | class _RecordingBackend: |
| 262 | """Captures launch_copy calls without touching the GPU.""" |
| 263 | |
| 264 | def __init__(self) -> None: |
| 265 | self.calls: list[dict] = [] |
| 266 | |
| 267 | def launch_copy( |
| 268 | self, |
| 269 | src_blocks, |
| 270 | dst_blocks, |
| 271 | is_store, |
| 272 | event_idx, |
| 273 | events_list, |
| 274 | wait_event=None, |
| 275 | ) -> None: |
| 276 | self.calls.append({"is_store": is_store, "wait_event": wait_event}) |
| 277 | |
| 278 | |
| 279 | def test_transfer_hooks_pass_wait_event_for_store_only(): |
| 280 | """wait_for_save gates stores on a compute-done event; start_load_kv does not.""" |
| 281 | worker = _make_worker( |
| 282 | kv_cache_config=None, |
| 283 | cpu_capacity_bytes=0, |
| 284 | ) |
| 285 | recording = _RecordingBackend() |
| 286 | worker._backend = recording |
| 287 | worker._connector_metadata = SimpleCPUOffloadMetadata( |
| 288 | load_event=0, |
| 289 | load_gpu_blocks=[0], |
| 290 | load_cpu_blocks=[0], |
| 291 | store_event=1, |
| 292 | store_gpu_blocks=[1], |
| 293 | store_cpu_blocks=[1], |
| 294 | ) |
| 295 | |
| 296 | worker.start_load_kv() |
| 297 | worker.wait_for_save() |
| 298 | |
| 299 | store_calls = [c for c in recording.calls if c["is_store"]] |
| 300 | load_calls = [c for c in recording.calls if not c["is_store"]] |
| 301 | assert len(store_calls) == 1 |
| 302 | assert len(load_calls) == 1 |
| 303 | assert isinstance(store_calls[0]["wait_event"], torch.Event) |
| 304 | assert load_calls[0]["wait_event"] is None |
| 305 | |
| 306 | |
| 307 | def test_build_params_src_access_order(): |
| 308 | """build_params defaults to ANY and honors an explicit STREAM override.""" |
| 309 | gpu = {"k": torch.zeros((4, 64), dtype=torch.int8, device="cuda")} |
| 310 | cpu = {"k": torch.zeros((4, 64), dtype=torch.int8, device="cpu")} |
| 311 | stream = torch.cuda.Stream() |
| 312 | |
| 313 | default = build_params(gpu, cpu, stream) |
| 314 | assert default.attrs.srcAccessOrder == CU_MEMCPY_SRC_ACCESS_ORDER_ANY |
| 315 | |
| 316 | ordered = build_params( |
| 317 | gpu, cpu, stream, src_access_order=CU_MEMCPY_SRC_ACCESS_ORDER_STREAM |
| 318 | ) |
| 319 | assert ordered.attrs.srcAccessOrder == CU_MEMCPY_SRC_ACCESS_ORDER_STREAM |
| 320 | |
| 321 | |
| 322 | @pytest.mark.parametrize("layout", list(KVCacheLayout)) |
| 323 | def test_register_shared_kv_cache_storage(monkeypatch, layout: KVCacheLayout): |
| 324 | num_blocks = 4 |
| 325 | num_layers = 2 |
| 326 | spec = FullAttentionSpec( |
| 327 | block_size=2, |
| 328 | num_kv_heads=2, |
| 329 | head_size=2, |
| 330 | dtype=torch.float16, |
| 331 | ) |
| 332 | raw = torch.zeros( |
| 333 | num_blocks * num_layers * spec.page_size_bytes, |
| 334 | dtype=torch.int8, |
| 335 | device="cuda", |
| 336 | ) |
| 337 | caches = dense_kv_cache_views(raw, spec, num_blocks, num_layers, layout) |
| 338 | layer_names = [f"layer.{i}" for i in range(num_layers)] |
| 339 | cache_config = MagicMock( |
| 340 | num_blocks=num_blocks, |
| 341 | kv_cache_tensors=[ |
| 342 | dense_kv_cache_tensor( |
| 343 | raw, spec, num_blocks, num_layers, layout, layer_names |
| 344 | ) |
| 345 | ], |
| 346 | ) |
| 347 | worker = _make_worker( |
| 348 | kv_cache_config=cache_config, |
| 349 | cpu_capacity_bytes=raw.nbytes, |
| 350 | ) |
| 351 | worker._backend = MagicMock() |
| 352 | monkeypatch.setattr("vllm.v1.simple_kv_offload.worker.PIN_MEMORY", False) |
| 353 | |
| 354 | worker.register_kv_caches( |
| 355 | {f"layer.{layer_idx}": cache for layer_idx, cache in enumerate(caches)} |
| 356 | ) |
| 357 | |
| 358 | assert worker.gpu_kv_caches is not None |
| 359 | if layout.is_layer_compact and not layout.is_block_compact: |
| 360 | expected_regions = num_layers * spec.num_heads |
| 361 | expected_block_bytes = spec.page_size_bytes // spec.num_heads |
| 362 | elif layout.is_layer_compact: |
| 363 | expected_regions = num_layers |
| 364 | expected_block_bytes = spec.page_size_bytes |
| 365 | else: |
| 366 | expected_regions = 1 |
| 367 | expected_block_bytes = spec.page_size_bytes * num_layers |
| 368 | assert len(worker.gpu_kv_caches) == expected_regions |
| 369 | assert {cache.shape for cache in worker.gpu_kv_caches.values()} == { |
| 370 | (num_blocks, expected_block_bytes) |
| 371 | } |
| 372 | |
| 373 | |
| 374 | def test_register_kv_cache_storage_with_trailing_padding(monkeypatch): |
| 375 | num_blocks = 4 |
| 376 | block_bytes = 32 |
| 377 | cache_bytes = num_blocks * block_bytes |
| 378 | raw = torch.zeros(4096, dtype=torch.int8, device="cuda") |
| 379 | cache = raw[:cache_bytes].view(num_blocks, block_bytes) |
| 380 | worker = _make_worker( |
| 381 | kv_cache_config=MagicMock( |
| 382 | num_blocks=num_blocks, |
| 383 | kv_cache_tensors=[ |
| 384 | KVCacheTensor( |
| 385 | size=cache_bytes, |
| 386 | layers=["layer.0"], |
| 387 | layer_stride=cache_bytes, |
| 388 | block_stride=block_bytes, |
| 389 | ) |
| 390 | ], |
| 391 | ), |
| 392 | cpu_capacity_bytes=cache_bytes, |
| 393 | ) |
| 394 | worker._backend = MagicMock() |
| 395 | monkeypatch.setattr("vllm.v1.simple_kv_offload.worker.PIN_MEMORY", False) |
| 396 | |
| 397 | worker.register_kv_caches({"layer.0": cache}) |
| 398 | |
| 399 | assert worker.gpu_kv_caches is not None |
| 400 | assert list(worker.gpu_kv_caches) == ["layer.0"] |
| 401 | assert worker.gpu_kv_caches["layer.0"].shape == (num_blocks, block_bytes) |
| 402 | |
| 403 | |
| 404 | def test_register_separate_kv_head_groups(monkeypatch): |
| 405 | # LHBNC hoists the K/V head groups outside the block dim, so each layer's |
| 406 | # blocks are registered as one region per group (K, V). |
| 407 | layout = KVCacheLayout.LHBNC |
| 408 | num_blocks = 4 |
| 409 | num_layers = 2 |
| 410 | spec = FullAttentionSpec( |
| 411 | block_size=2, |
| 412 | num_kv_heads=2, |
| 413 | head_size=2, |
| 414 | dtype=torch.float16, |
| 415 | num_head_slots=2, |
| 416 | state_content_bytes=2 * 2 * 2, |
| 417 | ) |
| 418 | raw = torch.zeros( |
| 419 | num_blocks * num_layers * spec.page_size_bytes, |
| 420 | dtype=torch.int8, |
| 421 | device="cuda", |
| 422 | ) |
| 423 | caches = dense_kv_cache_views(raw, spec, num_blocks, num_layers, layout) |
| 424 | layer_names = [f"layer.{i}" for i in range(num_layers)] |
| 425 | worker = _make_worker( |
| 426 | kv_cache_config=MagicMock( |
| 427 | num_blocks=num_blocks, |
| 428 | kv_cache_tensors=[ |
| 429 | dense_kv_cache_tensor( |
| 430 | raw, spec, num_blocks, num_layers, layout, layer_names |
| 431 | ) |
| 432 | ], |
| 433 | ), |
| 434 | cpu_capacity_bytes=raw.nbytes, |
| 435 | ) |
| 436 | worker._backend = MagicMock() |
| 437 | monkeypatch.setattr("vllm.v1.simple_kv_offload.worker.PIN_MEMORY", False) |
| 438 | |
| 439 | worker.register_kv_caches( |
| 440 | {f"layer.{layer_idx}": cache for layer_idx, cache in enumerate(caches)} |
| 441 | ) |
| 442 | |
| 443 | assert worker.gpu_kv_caches is not None |
| 444 | assert len(worker.gpu_kv_caches) == num_layers * spec.num_heads |
| 445 | per_group_block_bytes = ( |
| 446 | spec.num_kv_heads * spec.block_size * spec.head_size * spec.dtype.itemsize |
| 447 | ) |
| 448 | assert {cache.shape for cache in worker.gpu_kv_caches.values()} == { |
| 449 | (num_blocks, per_group_block_bytes) |
| 450 | } |
| 451 | |
| 452 | |
| 453 | def _dsa_specs(num_layers: int, block_size: int) -> dict[str, KVCacheSpec]: |
| 454 | """A DSA model's per-layer specs (DeepSeek-V3.2, GLM-5.2). |
| 455 | |
| 456 | Each decoder layer contributes an MLA latent cache and the indexer's key |
| 457 | cache, whose pages differ in size (``MLAAttentionSpec`` off ``head_size``: |
| 458 | 512 + 64 for the latent, 128 fp8 keys + one fp32 scale per 128 elements for |
| 459 | the indexer). Both are ``MLAAttentionSpec`` with the same block size, so the |
| 460 | hybrid allocator puts them in one ``UniformTypeKVCacheSpecs`` cache group. |
| 461 | See ``DeepseekV32IndexerCache.get_kv_cache_spec`` and |
| 462 | ``MLAAttention.get_kv_cache_spec``. |
| 463 | """ |
| 464 | specs: dict[str, KVCacheSpec] = {} |
| 465 | for i in range(num_layers): |
| 466 | specs[f"model.layers.{i}.self_attn.attn"] = MLAAttentionSpec( |
| 467 | block_size=block_size, |
| 468 | num_kv_heads=1, |
| 469 | head_size=512 + 64, |
| 470 | dtype=torch.uint8, |
| 471 | cache_dtype_str="fp8", |
| 472 | ) |
| 473 | specs[f"model.layers.{i}.self_attn.indexer.k_cache"] = MLAAttentionSpec( |
| 474 | block_size=block_size, |
| 475 | num_kv_heads=1, |
| 476 | head_size=128 + 128 // 128 * 4, |
| 477 | dtype=torch.uint8, |
| 478 | ) |
| 479 | return specs |
| 480 | |
| 481 | |
| 482 | def test_register_mixed_page_sizes_in_one_cache_group(monkeypatch): |
| 483 | """Sparse-MLA models mix page sizes inside one cache group. |
| 484 | |
| 485 | The MLA latent and indexer key caches of a layer have different page sizes, |
| 486 | so the allocation has no single per-layer block size to reinterpret it with: |
| 487 | ``num_layers * (mla_page + indexer_page)`` is not a multiple of either page. |
| 488 | Registration must derive each region from the placement metadata instead. |
| 489 | """ |
| 490 | num_layers = 4 |
| 491 | block_size = 64 |
| 492 | specs = _dsa_specs(num_layers, block_size) |
| 493 | # Differing head sizes make the specs non-identical but same-type, which is |
| 494 | # what lands both caches of a layer in one group. |
| 495 | assert not is_kv_cache_spec_uniform(specs) |
| 496 | assert UniformTypeKVCacheSpecs.is_uniform_type(specs) |
| 497 | group = KVCacheGroupSpec( |
| 498 | list(specs), |
| 499 | UniformTypeKVCacheSpecs(block_size=block_size, kv_cache_specs=specs), |
| 500 | ) |
| 501 | |
| 502 | # DEEPSEEK_V32_INDEXER declares no supported layouts, so resolution lands on |
| 503 | # the default preference: layer-outermost, one region per layer. |
| 504 | layout = KVCacheLayout.LBNHC |
| 505 | vllm_config = MagicMock() |
| 506 | vllm_config.cache_config = CacheConfig() |
| 507 | vllm_config.cache_config.kv_cache_layout = layout.name |
| 508 | vllm_config.cache_config.num_gpu_blocks_override = None |
| 509 | vllm_config.attention_config.hisparse_config = None |
| 510 | vllm_config.kv_transfer_config = None |
| 511 | |
| 512 | pages = [spec.page_size_bytes for spec in specs.values()] |
| 513 | num_blocks = 4 |
| 514 | kv_cache_config = get_kv_cache_config_from_groups( |
| 515 | vllm_config, [group], sum(pages) * num_blocks |
| 516 | ) |
| 517 | assert kv_cache_config.num_blocks == num_blocks |
| 518 | assert len(set(pages)) > 1, "the mixed page sizes are what this test covers" |
| 519 | |
| 520 | kv_caches = allocate_kv_cache(kv_cache_config, torch.device("cuda"), layout) |
| 521 | worker = _make_worker( |
| 522 | kv_cache_config=kv_cache_config, |
| 523 | cpu_capacity_bytes=sum(pages) * num_blocks, |
| 524 | ) |
| 525 | worker._backend = MagicMock() |
| 526 | monkeypatch.setattr("vllm.v1.simple_kv_offload.worker.PIN_MEMORY", False) |
| 527 | |
| 528 | worker.register_kv_caches(kv_caches) |
| 529 | |
| 530 | assert worker.gpu_kv_caches is not None |
| 531 | # One region per layer, each with that layer's own page as its block stride. |
| 532 | assert sorted(worker.gpu_kv_caches) == sorted(specs) |
| 533 | assert {name: cache.shape[1] for name, cache in worker.gpu_kv_caches.items()} == { |
| 534 | name: spec.page_size_bytes for name, spec in specs.items() |
| 535 | } |
| 536 | assert all(cache.shape[0] == num_blocks for cache in worker.gpu_kv_caches.values()) |
| 537 | |
| 538 | # Every registered region must alias the bytes the model writes through, or |
| 539 | # offloaded blocks would be copied from the wrong place. |
| 540 | for name, cache in kv_caches.items(): |
| 541 | region = worker.gpu_kv_caches[name] |
| 542 | assert region.data_ptr() == cache.data_ptr() |
| 543 | assert region.stride(0) == specs[name].page_size_bytes |
| 544 | |
| 545 | |
| 546 | @pytest.mark.parametrize("rank_blocks", [1, 5]) |
| 547 | def test_register_mixed_page_sizes_odd_block_counts(monkeypatch, rank_blocks): |
| 548 | """Registration holds at block counts that are not powers of two. |
| 549 | |
| 550 | ``num_blocks`` comes out of ``available_memory // bytes_per_block``, so |
| 551 | any integer is reachable. Registration derives regions from the config's |
| 552 | own block count and strides, and the CPU pool rounds down from the same |
| 553 | per-block byte total; none of it assumes anything about how the count |
| 554 | was reached. |
| 555 | """ |
| 556 | num_layers = 4 |
| 557 | block_size = 64 |
| 558 | specs = _dsa_specs(num_layers, block_size) |
| 559 | group = KVCacheGroupSpec( |
| 560 | list(specs), |
| 561 | UniformTypeKVCacheSpecs(block_size=block_size, kv_cache_specs=specs), |
| 562 | ) |
| 563 | layout = KVCacheLayout.LBNHC |
| 564 | vllm_config = MagicMock() |
| 565 | vllm_config.cache_config = CacheConfig() |
| 566 | vllm_config.cache_config.kv_cache_layout = layout.name |
| 567 | vllm_config.cache_config.num_gpu_blocks_override = None |
| 568 | vllm_config.attention_config.hisparse_config = None |
| 569 | vllm_config.kv_transfer_config = None |
| 570 | |
| 571 | pages = [spec.page_size_bytes for spec in specs.values()] |
| 572 | kv_cache_config = get_kv_cache_config_from_groups( |
| 573 | vllm_config, [group], sum(pages) * rank_blocks |
| 574 | ) |
| 575 | assert kv_cache_config.num_blocks == rank_blocks |
| 576 | |
| 577 | kv_caches = allocate_kv_cache(kv_cache_config, torch.device("cuda"), layout) |
| 578 | worker = _make_worker( |
| 579 | kv_cache_config=kv_cache_config, |
| 580 | cpu_capacity_bytes=sum(pages) * rank_blocks, |
| 581 | ) |
| 582 | worker._backend = MagicMock() |
| 583 | monkeypatch.setattr("vllm.v1.simple_kv_offload.worker.PIN_MEMORY", False) |
| 584 | |
| 585 | worker.register_kv_caches(kv_caches) |
| 586 | |
| 587 | assert worker.gpu_kv_caches is not None |
| 588 | assert {n: c.shape for n, c in worker.gpu_kv_caches.items()} == { |
| 589 | name: (rank_blocks, spec.page_size_bytes) for name, spec in specs.items() |
| 590 | } |
| 591 | assert worker.num_cpu_blocks == rank_blocks |
| 592 | |
| 593 | |
| 594 | def test_mixed_page_byte_placement_is_dcp_invariant(): |
| 595 | """DCP scales a block's token span, not its byte placement. |
| 596 | |
| 597 | ``get_kv_cache_config_from_groups`` never reads |
| 598 | ``decode_context_parallel_size``, so every placement field registration |
| 599 | derives regions from is identical at dcp=1 and dcp=2. What DCP scales is |
| 600 | the block's token span, reported by ``resolve_kv_cache_block_sizes``; |
| 601 | checking that it doubles proves DCP was actually in effect, or the |
| 602 | identity above would be trivially true. |
| 603 | """ |
| 604 | specs = _dsa_specs(4, block_size=128) |
| 605 | |
| 606 | placements = [] |
| 607 | for dcp_size in (1, 2): |
| 608 | group = KVCacheGroupSpec( |
| 609 | list(specs), |
| 610 | UniformTypeKVCacheSpecs(block_size=128, kv_cache_specs=specs), |
| 611 | ) |
| 612 | vllm_config = MagicMock() |
| 613 | vllm_config.cache_config = CacheConfig() |
| 614 | vllm_config.cache_config.block_size = 128 |
| 615 | vllm_config.cache_config.kv_cache_layout = KVCacheLayout.LBNHC.name |
| 616 | vllm_config.cache_config.num_gpu_blocks_override = None |
| 617 | vllm_config.parallel_config.decode_context_parallel_size = dcp_size |
| 618 | vllm_config.attention_config.hisparse_config = None |
| 619 | vllm_config.kv_transfer_config = None |
| 620 | |
| 621 | page_bytes = sum(spec.page_size_bytes for spec in specs.values()) |
| 622 | config = get_kv_cache_config_from_groups(vllm_config, [group], page_bytes * 4) |
| 623 | scheduler_block_size, _ = resolve_kv_cache_block_sizes(config, vllm_config) |
| 624 | assert scheduler_block_size == 128 * dcp_size |
| 625 | placements.append( |
| 626 | [ |
| 627 | (t.size, tuple(t.layers), t.layer_stride, t.block_stride, t.offset) |
| 628 | for t in config.kv_cache_tensors |
| 629 | ] |
| 630 | ) |
| 631 | |
| 632 | assert placements[0] == placements[1] |
| 633 | |
| 634 | |
| 635 | @pytest.mark.parametrize( |
| 636 | ("block_bytes", "use_page_cache", "expect_ok"), |
| 637 | [ |
| 638 | (4096, True, True), # aligned control |
| 639 | (3 * 512, True, True), # 512-aligned only: fine for buffered I/O |
| 640 | (3 * 512, False, False), # O_DIRECT keeps enforcing 4096 alignment |
| 641 | ], |
| 642 | ) |
| 643 | def test_disk_backend_alignment_only_required_for_direct_io( |
| 644 | tmp_path, monkeypatch, block_bytes: int, use_page_cache: bool, expect_ok: bool |
| 645 | ): |
| 646 | """The 4096 stride assert belongs to O_DIRECT; page-cache I/O must not hit it. |
| 647 | |
| 648 | Hybrid models (e.g. 101 per-rank cache tensors) sum to a block stride that |
| 649 | is 512- but not 4096-aligned, which previously made disk offload unbootable |
| 650 | even with use_page_cache=True. |
| 651 | """ |
| 652 | # Host pinning is a separate lifecycle concern with its own tests; the |
| 653 | # alignment gate must not depend on registration state, so stub both ends. |
| 654 | monkeypatch.setattr( |
| 655 | "vllm.v1.simple_kv_offload.disk_backend.pin_tensor", lambda t: None |
| 656 | ) |
| 657 | monkeypatch.setattr( |
| 658 | "vllm.v1.simple_kv_offload.disk_backend.unpin_tensor", |
| 659 | lambda t: None, |
| 660 | raising=False, # harmless on branches without teardown-side unpins |
| 661 | ) |
| 662 | num_blocks = 4 |
| 663 | gpu = {"k": torch.zeros((num_blocks, block_bytes), dtype=torch.int8, device="cuda")} |
| 664 | backend = DiskBackend() |
| 665 | disk_path = tmp_path / "kv-offload.bin" |
| 666 | |
| 667 | def _init() -> None: |
| 668 | backend.init( |
| 669 | gpu, |
| 670 | gpu["k"].device, |
| 671 | torch.cuda.Stream(), |
| 672 | torch.cuda.Stream(), |
| 673 | str(disk_path), |
| 674 | num_blocks, |
| 675 | block_bytes, |
| 676 | use_page_cache=use_page_cache, |
| 677 | ) |
| 678 | |
| 679 | if not expect_ok: |
| 680 | with pytest.raises(AssertionError, match="not aligned"): |
| 681 | _init() |
| 682 | return |
| 683 | |
| 684 | _init() |
| 685 | try: |
| 686 | assert disk_path.stat().st_size == num_blocks * block_bytes |
| 687 | finally: |
| 688 | backend.shutdown() |