| 1 | # SPDX-License-Identifier: Apache-2.0 |
| 2 | # SPDX-FileCopyrightText: Copyright contributors to the vLLM project |
| 3 | """Padded-page handling in create_kv_cache_views. |
| 4 | |
| 5 | Guards that a page_size_padded spec strides the block dimension by the padded page |
| 6 | while keeping per-block content compact, so padding bytes at the end of each page are |
| 7 | never addressed by the logical view. |
| 8 | """ |
| 9 | |
| 10 | from types import SimpleNamespace |
| 11 | from typing import Any |
| 12 | |
| 13 | import numpy as np |
| 14 | import pytest |
| 15 | import torch |
| 16 | |
| 17 | import vllm.v1.hisparse.binding as attn_utils_module |
| 18 | from tests.v1.attention.utils import dense_kv_cache_views |
| 19 | from vllm.config.compilation import CUDAGraphMode |
| 20 | from vllm.v1.attention.backend import AttentionBackend, AttentionCGSupport, MultipleOf |
| 21 | from vllm.v1.core.kv_cache_utils import KVCacheBlockCopy |
| 22 | from vllm.v1.hisparse.binding import allocate_hisparse_kv_caches |
| 23 | from vllm.v1.kv_cache_interface import ( |
| 24 | FullAttentionSpec, |
| 25 | HiSparseResidentSpec, |
| 26 | KVCacheConfig, |
| 27 | KVCacheGroupSpec, |
| 28 | KVCacheLayout, |
| 29 | KVCacheTensor, |
| 30 | MLAAttentionSpec, |
| 31 | SparseCacheRole, |
| 32 | compute_layout_strides, |
| 33 | ) |
| 34 | from vllm.v1.worker.gpu import attn_utils |
| 35 | from vllm.v1.worker.gpu.attn_utils import ( |
| 36 | FastPrefillHelper, |
| 37 | get_attn_cg_support, |
| 38 | get_query_lens_mismatch_unsupported_backend, |
| 39 | ) |
| 40 | from vllm.v1.worker.gpu.cudagraph_utils import BatchExecutionDescriptor |
| 41 | from vllm.v1.worker.utils import ( |
| 42 | AttentionGroup, |
| 43 | allocate_kv_cache, |
| 44 | copy_kv_cache_blocks_inplace, |
| 45 | ) |
| 46 | |
| 47 | |
| 48 | @pytest.mark.parametrize( |
| 49 | ("enabled", "block_size", "main_sizes", "indexer_sizes", "expected"), |
| 50 | [ |
| 51 | (True, 256, [64], [64], 64), |
| 52 | (True, 64, [32, 64], [16, 32], 32), |
| 53 | (True, 64, [MultipleOf(16)], [32], 32), |
| 54 | (True, 64, [64], [32], None), |
| 55 | (False, 256, [64], [64], 256), |
| 56 | ], |
| 57 | ) |
| 58 | def test_get_kv_cache_spec_resolves_hisparse_block_size( |
| 59 | monkeypatch, enabled, block_size, main_sizes, indexer_sizes, expected |
| 60 | ): |
| 61 | """Resolve shared MLA geometry before planning; leave other specs alone.""" |
| 62 | specs = { |
| 63 | "main": MLAAttentionSpec( |
| 64 | block_size=block_size, num_kv_heads=1, head_size=576, dtype=torch.bfloat16 |
| 65 | ), |
| 66 | "indexer": MLAAttentionSpec( |
| 67 | block_size=block_size, |
| 68 | num_kv_heads=1, |
| 69 | head_size=128, |
| 70 | dtype=torch.bfloat16, |
| 71 | cache_role=SparseCacheRole.INDEXER, |
| 72 | ), |
| 73 | "dense": FullAttentionSpec( |
| 74 | block_size=block_size, num_kv_heads=1, head_size=128, dtype=torch.bfloat16 |
| 75 | ), |
| 76 | } |
| 77 | layers = {} |
| 78 | for name, sizes in zip(specs, [main_sizes, indexer_sizes, [block_size]]): |
| 79 | backend = SimpleNamespace( |
| 80 | get_name=lambda name=name: name, |
| 81 | customize_spec=AttentionBackend.customize_spec, |
| 82 | get_supported_kernel_block_sizes=lambda sizes=sizes: sizes, |
| 83 | ) |
| 84 | layers[name] = SimpleNamespace( |
| 85 | get_kv_cache_spec=lambda _, spec=specs[name]: spec, |
| 86 | get_attn_backend=lambda backend=backend: backend, |
| 87 | ) |
| 88 | monkeypatch.setattr(attn_utils, "get_layers_from_vllm_config", lambda *_: layers) |
| 89 | monkeypatch.setattr( |
| 90 | attn_utils_module, "get_hisparse_kv_cache_groups", lambda *_: [] |
| 91 | ) |
| 92 | config = SimpleNamespace( |
| 93 | attention_config=SimpleNamespace(hisparse_config=object() if enabled else None) |
| 94 | ) |
| 95 | if expected is None: |
| 96 | with pytest.raises(ValueError, match="supported by every sparse"): |
| 97 | attn_utils.get_kv_cache_spec(config) |
| 98 | return |
| 99 | |
| 100 | resolved = attn_utils.get_kv_cache_spec(config) |
| 101 | assert resolved["main"].block_size == resolved["indexer"].block_size == expected |
| 102 | assert resolved["dense"] is specs["dense"] |
| 103 | assert all(spec.block_size == block_size for spec in specs.values()) |
| 104 | |
| 105 | |
| 106 | class _FakeMetadataBuilder: |
| 107 | def __init__(self, support: AttentionCGSupport, varlen_bound: int | None = None): |
| 108 | self.support = support |
| 109 | self.varlen_bound = varlen_bound |
| 110 | |
| 111 | def get_cudagraph_support(self, *_args): |
| 112 | return self.support |
| 113 | |
| 114 | def get_varlen_cudagraph_max_query_len(self, *_args): |
| 115 | return self.varlen_bound |
| 116 | |
| 117 | |
| 118 | class _TargetBackend: |
| 119 | @classmethod |
| 120 | def supports_device_cpu_query_lens_mismatch(cls) -> bool: |
| 121 | return True |
| 122 | |
| 123 | |
| 124 | class _DraftBackend: |
| 125 | @classmethod |
| 126 | def supports_device_cpu_query_lens_mismatch(cls) -> bool: |
| 127 | return False |
| 128 | |
| 129 | |
| 130 | def test_attention_checks_preserve_global_and_target_scoped_support(): |
| 131 | spec = FullAttentionSpec( |
| 132 | block_size=16, |
| 133 | num_kv_heads=1, |
| 134 | head_size=128, |
| 135 | dtype=torch.bfloat16, |
| 136 | ) |
| 137 | target_group = AttentionGroup( |
| 138 | _TargetBackend, |
| 139 | ["target"], |
| 140 | spec, |
| 141 | 0, |
| 142 | ) |
| 143 | target_group.metadata_builders = [ |
| 144 | _FakeMetadataBuilder(AttentionCGSupport.ALWAYS) # type: ignore[list-item] |
| 145 | ] |
| 146 | draft_group = AttentionGroup( |
| 147 | _DraftBackend, |
| 148 | ["draft"], |
| 149 | spec, |
| 150 | 0, |
| 151 | ) |
| 152 | draft_group.metadata_builders = [ |
| 153 | _FakeMetadataBuilder(AttentionCGSupport.UNIFORM_BATCH) # type: ignore[list-item] |
| 154 | ] |
| 155 | groups = [[target_group, draft_group]] |
| 156 | |
| 157 | # The runner-wide execution mode must still honor the drafter's limit. |
| 158 | unfiltered = get_attn_cg_support(groups, None) |
| 159 | assert unfiltered.min_cg_support == AttentionCGSupport.UNIFORM_BATCH |
| 160 | assert unfiltered.min_cg_attn_backend == "_DraftBackend" |
| 161 | |
| 162 | # Adaptive verification validates only the target's varlen graphs. |
| 163 | target_only = get_attn_cg_support( |
| 164 | groups, |
| 165 | None, |
| 166 | checked_layer_names={"target"}, |
| 167 | ) |
| 168 | assert target_only.min_cg_support == AttentionCGSupport.ALWAYS |
| 169 | assert target_only.min_cg_attn_backend is None |
| 170 | assert ( |
| 171 | get_query_lens_mismatch_unsupported_backend( |
| 172 | groups, |
| 173 | checked_layer_names={"target"}, |
| 174 | ) |
| 175 | is None |
| 176 | ) |
| 177 | |
| 178 | # Shared target/draft groups still participate in target-scoped checks. |
| 179 | draft_group.layer_names.append("target") |
| 180 | target_with_shared_group = get_attn_cg_support( |
| 181 | groups, |
| 182 | None, |
| 183 | checked_layer_names={"target"}, |
| 184 | ) |
| 185 | assert target_with_shared_group.min_cg_support == AttentionCGSupport.UNIFORM_BATCH |
| 186 | assert ( |
| 187 | get_query_lens_mismatch_unsupported_backend( |
| 188 | groups, |
| 189 | checked_layer_names={"target"}, |
| 190 | ) |
| 191 | == "_DraftBackend" |
| 192 | ) |
| 193 | |
| 194 | |
| 195 | def test_varlen_cudagraph_unsupported_backend_checks_scoped_bounds(): |
| 196 | """ALWAYS passes without a bound, other builders need one at least as wide as |
| 197 | the requested length, and NEVER fails whatever bound a builder reports.""" |
| 198 | config: Any = SimpleNamespace() |
| 199 | spec = FullAttentionSpec( |
| 200 | block_size=16, |
| 201 | num_kv_heads=1, |
| 202 | head_size=128, |
| 203 | dtype=torch.bfloat16, |
| 204 | ) |
| 205 | |
| 206 | def group( |
| 207 | backend: Any, |
| 208 | layer_name: str, |
| 209 | support: AttentionCGSupport, |
| 210 | bound: int | None = None, |
| 211 | ): |
| 212 | builder: Any = _FakeMetadataBuilder(support, bound) |
| 213 | attn_group = AttentionGroup(backend, [layer_name], spec, 0) |
| 214 | attn_group.metadata_builders = [builder] |
| 215 | return attn_group |
| 216 | |
| 217 | target = group(_TargetBackend, "target", AttentionCGSupport.ALWAYS) |
| 218 | draft = group(_DraftBackend, "draft", AttentionCGSupport.UNIFORM_BATCH, 8) |
| 219 | never = group(_DraftBackend, "never", AttentionCGSupport.NEVER, 8) |
| 220 | |
| 221 | unsupported = attn_utils.get_varlen_cudagraph_unsupported_backend |
| 222 | assert unsupported([[target, draft]], config, 8) is None |
| 223 | assert unsupported([[target, draft]], config, 9) == ("_DraftBackend", 8) |
| 224 | assert ( |
| 225 | unsupported([[target, draft]], config, 9, checked_layer_names={"target"}) |
| 226 | is None |
| 227 | ) |
| 228 | assert unsupported([[target, never]], config, 1) == ("_DraftBackend", None) |
| 229 | |
| 230 | |
| 231 | def test_get_kv_sharing_fast_prefill_eligible_layers(monkeypatch: pytest.MonkeyPatch): |
| 232 | """Fast prefill applies to the contiguous suffix of KV-sharing layers. |
| 233 | |
| 234 | Draft-model layers register after the target model's and may share KV, so |
| 235 | they must not extend (or break) the target's eligible suffix. |
| 236 | """ |
| 237 | |
| 238 | def check( |
| 239 | layer_names: list[str], |
| 240 | shared: dict[str, str], |
| 241 | draft_layer_names: set[str] | None = None, |
| 242 | ) -> set[str]: |
| 243 | monkeypatch.setattr( |
| 244 | attn_utils, |
| 245 | "get_layers_from_vllm_config", |
| 246 | lambda *a, **k: {name: None for name in layer_names}, |
| 247 | ) |
| 248 | monkeypatch.setattr(attn_utils, "get_shared_kv_cache_layers", lambda *a: shared) |
| 249 | vllm_config = SimpleNamespace( |
| 250 | cache_config=SimpleNamespace(kv_sharing_fast_prefill=True) |
| 251 | ) |
| 252 | return attn_utils.get_kv_sharing_fast_prefill_eligible_layers( |
| 253 | vllm_config, draft_layer_names |
| 254 | ) |
| 255 | |
| 256 | # No KV sharing: nothing is eligible. |
| 257 | assert check(["t0", "t1"], {}) == set() |
| 258 | |
| 259 | # Trailing run of sharing layers (YOCO-style second half). |
| 260 | assert check(["t0", "t1", "t2", "t3"], {"t2": "t1", "t3": "t1"}) == {"t2", "t3"} |
| 261 | |
| 262 | # A non-sharing layer after a sharing one breaks the suffix. |
| 263 | assert check(["t0", "t1", "t2", "t3"], {"t1": "t0", "t3": "t0"}) == {"t3"} |
| 264 | |
| 265 | # KV-sharing draft layers at the end are collected without an exclusion... |
| 266 | assert check( |
| 267 | ["t0", "t1", "t2", "t3", "d0", "d1"], |
| 268 | {"t2": "t1", "t3": "t1", "d0": "t1", "d1": "t1"}, |
| 269 | ) == {"t2", "t3", "d0", "d1"} |
| 270 | |
| 271 | # ...so the runner excludes them: skipped, not collected, and they do not |
| 272 | # break the target's trailing run. |
| 273 | assert check( |
| 274 | ["t0", "t1", "t2", "t3", "d0", "d1"], |
| 275 | {"t2": "t1", "t3": "t1", "d0": "t1", "d1": "t1"}, |
| 276 | draft_layer_names={"d0", "d1"}, |
| 277 | ) == {"t2", "t3"} |
| 278 | |
| 279 | # Feature flag off: nothing is eligible even with sharing layers. |
| 280 | monkeypatch.setattr( |
| 281 | attn_utils, "get_layers_from_vllm_config", lambda *a, **k: {"t0": None} |
| 282 | ) |
| 283 | monkeypatch.setattr( |
| 284 | attn_utils, "get_shared_kv_cache_layers", lambda *a: {"t0": "t0"} |
| 285 | ) |
| 286 | vllm_config = SimpleNamespace( |
| 287 | cache_config=SimpleNamespace(kv_sharing_fast_prefill=False) |
| 288 | ) |
| 289 | assert attn_utils.get_kv_sharing_fast_prefill_eligible_layers(vllm_config) == set() |
| 290 | |
| 291 | |
| 292 | @pytest.mark.parametrize( |
| 293 | "num_tokens", [1, 2, 3, 4, 7, 8, 15, 16, 31, 32, 63, 64, 127, 128] |
| 294 | ) |
| 295 | @pytest.mark.parametrize("num_active_loras", [1, 2, 4]) |
| 296 | def test_fast_prefill_dispatch_preserves_active_lora_count( |
| 297 | num_tokens: int, num_active_loras: int |
| 298 | ): |
| 299 | """Fast-prefill padding must match the main dispatch's LoRA variant.""" |
| 300 | |
| 301 | class FakeCudaGraphManager: |
| 302 | device = "cpu" |
| 303 | |
| 304 | def __init__(self): |
| 305 | self.dispatch_calls = [] |
| 306 | |
| 307 | def dispatch(self, **kwargs): |
| 308 | self.dispatch_calls.append(kwargs) |
| 309 | # A captured no-LoRA graph is padded to 8 tokens, while the |
| 310 | # active-LoRA path stays eager at the unpadded token count. |
| 311 | if kwargs["num_active_loras"] == 0: |
| 312 | num_tokens = 8 |
| 313 | mode = CUDAGraphMode.PIECEWISE |
| 314 | else: |
| 315 | num_tokens = kwargs["num_tokens"] |
| 316 | mode = CUDAGraphMode.NONE |
| 317 | return BatchExecutionDescriptor( |
| 318 | cg_mode=mode, |
| 319 | num_tokens=num_tokens, |
| 320 | num_reqs=kwargs["num_reqs"], |
| 321 | num_active_loras=kwargs["num_active_loras"], |
| 322 | num_ubatches=1, |
| 323 | ) |
| 324 | |
| 325 | manager = FakeCudaGraphManager() |
| 326 | helper = FastPrefillHelper(manager, max_num_tokens=max(32, num_tokens)) |
| 327 | metadata = helper.prepare( |
| 328 | torch.arange(num_tokens, dtype=torch.int32), |
| 329 | num_reqs=1, |
| 330 | cu_num_logits_np=np.array([0, num_tokens], dtype=np.int32), |
| 331 | has_prefill=True, |
| 332 | batch_desc=BatchExecutionDescriptor( |
| 333 | cg_mode=CUDAGraphMode.NONE, |
| 334 | num_tokens=num_tokens, |
| 335 | num_reqs=1, |
| 336 | num_active_loras=num_active_loras, |
| 337 | num_ubatches=1, |
| 338 | ), |
| 339 | num_active_loras=num_active_loras, |
| 340 | ) |
| 341 | |
| 342 | assert metadata is not None |
| 343 | assert metadata.num_logits_indices == num_tokens |
| 344 | assert metadata.logits_indices_padded.shape[0] == num_tokens |
| 345 | assert metadata.max_logits_per_req == num_tokens |
| 346 | assert manager.dispatch_calls[-1]["num_active_loras"] == num_active_loras |
| 347 | |
| 348 | |
| 349 | class _FakeSharedHostRegion: |
| 350 | def __init__(self) -> None: |
| 351 | self.cleanup_calls = 0 |
| 352 | self.base_tensor = torch.empty(1, dtype=torch.int8) |
| 353 | |
| 354 | def cleanup(self) -> None: |
| 355 | self.cleanup_calls += 1 |
| 356 | |
| 357 | |
| 358 | def test_profiling_cleanup_releases_tp_shared_region_once(monkeypatch): |
| 359 | """TP-shared profiling pools must use region-aware chunk cleanup.""" |
| 360 | region = _FakeSharedHostRegion() |
| 361 | runtime = SimpleNamespace( |
| 362 | _host_cache=object(), |
| 363 | registered_host_pool=region.base_tensor, |
| 364 | hot_backing=object(), |
| 365 | shared_host_region=region, |
| 366 | ) |
| 367 | forward_context = { |
| 368 | "layer": SimpleNamespace( |
| 369 | hisparse_cache=SimpleNamespace(runtime=runtime), |
| 370 | ) |
| 371 | } |
| 372 | released = [] |
| 373 | |
| 374 | def release_pinned_state(runtimes, pinned_host_pools, shared_host_region): |
| 375 | released.append((runtimes, pinned_host_pools, shared_host_region)) |
| 376 | |
| 377 | monkeypatch.setattr( |
| 378 | attn_utils_module, |
| 379 | "release_pinned_state", |
| 380 | release_pinned_state, |
| 381 | ) |
| 382 | |
| 383 | attn_utils_module.release_hisparse_profiling_cache(forward_context) |
| 384 | |
| 385 | assert released == [([runtime], [], region)] |
| 386 | |
| 387 | |
| 388 | @pytest.mark.parametrize("failure_phase", ["allocation", "binding", "buffers"]) |
| 389 | def test_init_hisparse_rolls_back_shared_region(monkeypatch, failure_phase): |
| 390 | """A failure after mmap allocation must not leak the shared registration.""" |
| 391 | region = _FakeSharedHostRegion() |
| 392 | vllm_config = SimpleNamespace( |
| 393 | cache_config=SimpleNamespace( |
| 394 | get_resolved_kv_cache_layout=lambda: KVCacheLayout.BLHNC |
| 395 | ), |
| 396 | scheduler_config=SimpleNamespace(max_num_seqs=1, max_num_batched_tokens=1), |
| 397 | ) |
| 398 | |
| 399 | def allocate(*args): |
| 400 | args[-1].shared_region = region |
| 401 | if failure_phase == "allocation": |
| 402 | raise RuntimeError("initialization failed") |
| 403 | return {} |
| 404 | |
| 405 | def bind(**kwargs): |
| 406 | if failure_phase == "binding": |
| 407 | raise RuntimeError("initialization failed") |
| 408 | return [] |
| 409 | |
| 410 | def buffers(*args, **kwargs): |
| 411 | raise RuntimeError("initialization failed") |
| 412 | |
| 413 | monkeypatch.setattr(attn_utils_module, "allocate_hisparse_kv_caches", allocate) |
| 414 | monkeypatch.setattr(attn_utils_module, "bind_hisparse_kv_caches", bind) |
| 415 | monkeypatch.setattr( |
| 416 | attn_utils_module, "initialize_hisparse_runtime_buffers", buffers |
| 417 | ) |
| 418 | with pytest.raises(RuntimeError, match="initialization failed"): |
| 419 | attn_utils_module.init_hisparse_kv_cache( |
| 420 | SimpleNamespace(), |
| 421 | torch.device("cpu"), |
| 422 | [], |
| 423 | vllm_config, |
| 424 | {}, |
| 425 | SimpleNamespace(), |
| 426 | ) |
| 427 | assert region.cleanup_calls == 1 |
| 428 | |
| 429 | |
| 430 | def test_reshape_padded_kv_cache_strides_by_padded_page(): |
| 431 | num_blocks = 3 |
| 432 | spec = FullAttentionSpec( |
| 433 | block_size=16, |
| 434 | num_kv_heads=1, |
| 435 | head_size=2, |
| 436 | dtype=torch.float32, |
| 437 | page_size_padded=384, |
| 438 | ) |
| 439 | assert spec.real_page_size_bytes == 256 |
| 440 | |
| 441 | raw = torch.zeros(spec.page_size_bytes * num_blocks, dtype=torch.int8) |
| 442 | (kv_cache,) = dense_kv_cache_views(raw, spec, num_blocks, 1, KVCacheLayout.LBHNC) |
| 443 | |
| 444 | elem_size = 4 # float32 |
| 445 | # Content dim packs K and V: 2 * head_size. |
| 446 | assert kv_cache.shape == (num_blocks, 1, 16, 2 * spec.head_size) |
| 447 | assert kv_cache.dtype == spec.dtype |
| 448 | assert kv_cache.stride(0) == spec.page_size_padded // elem_size |
| 449 | assert kv_cache[1].storage_offset() == spec.page_size_padded // elem_size |
| 450 | # Within one block the (unpadded) content stays compact. |
| 451 | assert kv_cache[0].is_contiguous() |
| 452 | |
| 453 | |
| 454 | @pytest.mark.parametrize( |
| 455 | ( |
| 456 | "kernel_block_sizes", |
| 457 | "storage_block_size", |
| 458 | "expected_num_blocks", |
| 459 | "expected_num_states", |
| 460 | ), |
| 461 | [ |
| 462 | (None, None, 4, 64), |
| 463 | ([256], None, 4, 64), |
| 464 | ([64], None, 16, 16), |
| 465 | ([64], 256, 4, 64), |
| 466 | ], |
| 467 | ) |
| 468 | def test_allocate_compressed_mla_cache( |
| 469 | kernel_block_sizes: list[int] | None, |
| 470 | storage_block_size: int | None, |
| 471 | expected_num_blocks: int, |
| 472 | expected_num_states: int, |
| 473 | ): |
| 474 | spec = MLAAttentionSpec( |
| 475 | block_size=256, |
| 476 | num_kv_heads=1, |
| 477 | head_size=128, |
| 478 | dtype=torch.bfloat16, |
| 479 | tokens_per_state=4, |
| 480 | storage_block_size=storage_block_size, |
| 481 | ) |
| 482 | num_pages = 4 |
| 483 | config = KVCacheConfig( |
| 484 | num_blocks=num_pages, |
| 485 | kv_cache_tensors=[ |
| 486 | KVCacheTensor( |
| 487 | size=num_pages * spec.page_size_bytes, |
| 488 | layers=["layer.0"], |
| 489 | layer_stride=num_pages * spec.page_size_bytes, |
| 490 | block_stride=spec.page_size_bytes, |
| 491 | ) |
| 492 | ], |
| 493 | kv_cache_groups=[KVCacheGroupSpec(["layer.0"], spec)], |
| 494 | ) |
| 495 | |
| 496 | caches = allocate_kv_cache( |
| 497 | config, torch.device("cpu"), KVCacheLayout.LBHNC, kernel_block_sizes |
| 498 | ) |
| 499 | |
| 500 | assert caches["layer.0"].shape == (expected_num_blocks, 1, expected_num_states, 128) |
| 501 | |
| 502 | |
| 503 | @pytest.mark.parametrize("layout", list(KVCacheLayout)) |
| 504 | def test_copy_kv_cache_blocks_shared_storage(layout: KVCacheLayout): |
| 505 | num_blocks = 4 |
| 506 | num_layers = 2 |
| 507 | spec = FullAttentionSpec( |
| 508 | block_size=2, |
| 509 | num_kv_heads=2, |
| 510 | head_size=2, |
| 511 | dtype=torch.float32, |
| 512 | ) |
| 513 | raw = torch.zeros(num_blocks * num_layers * spec.page_size_bytes, dtype=torch.int8) |
| 514 | caches = dense_kv_cache_views(raw, spec, num_blocks, num_layers, layout) |
| 515 | |
| 516 | for layer_idx, cache in enumerate(caches): |
| 517 | for block_idx in range(num_blocks): |
| 518 | cache[block_idx].fill_(10 * layer_idx + block_idx) |
| 519 | |
| 520 | expected = [[cache[i].clone() for i in range(num_blocks)] for cache in caches] |
| 521 | copies = [KVCacheBlockCopy(src_block_id=0, dst_block_id=2)] |
| 522 | |
| 523 | copy_kv_cache_blocks_inplace(caches, num_blocks, copies) |
| 524 | |
| 525 | for layer_idx, cache in enumerate(caches): |
| 526 | torch.testing.assert_close(cache[2], expected[layer_idx][0]) |
| 527 | torch.testing.assert_close(cache[1], expected[layer_idx][1]) |
| 528 | |
| 529 | |
| 530 | def test_fixed_block_stride_propagates_outward_in_lhbnc(): |
| 531 | num_blocks = 3 |
| 532 | num_layers = 2 |
| 533 | spec = FullAttentionSpec( |
| 534 | block_size=2, |
| 535 | num_kv_heads=2, |
| 536 | head_size=2, |
| 537 | dtype=torch.float32, |
| 538 | ) |
| 539 | natural = compute_layout_strides(spec, num_blocks, num_layers, KVCacheLayout.LHBNC) |
| 540 | block_stride = natural[1] + 8 |
| 541 | |
| 542 | strides = compute_layout_strides( |
| 543 | spec, |
| 544 | num_blocks, |
| 545 | num_layers, |
| 546 | KVCacheLayout.LHBNC, |
| 547 | fixed_strides=(None, block_stride, None, None, None), |
| 548 | ) |
| 549 | |
| 550 | assert strides[1] == block_stride |
| 551 | assert strides[2] == block_stride * num_blocks |
| 552 | assert strides[0] == strides[2] * spec.num_heads |
| 553 | |
| 554 | |
| 555 | def test_copy_kv_cache_blocks_separate_head_groups(): |
| 556 | # LHBNC stores each head group separately, so a block's bytes are scattered |
| 557 | # across L*H regions. |
| 558 | layout = KVCacheLayout.LHBNC |
| 559 | num_blocks = 4 |
| 560 | num_layers = 2 |
| 561 | spec = FullAttentionSpec( |
| 562 | block_size=2, |
| 563 | num_kv_heads=2, |
| 564 | head_size=2, |
| 565 | dtype=torch.float32, |
| 566 | num_head_slots=2, |
| 567 | state_content_bytes=2 * 2 * 4, |
| 568 | ) |
| 569 | raw = torch.zeros(num_blocks * num_layers * spec.page_size_bytes, dtype=torch.int8) |
| 570 | caches = dense_kv_cache_views(raw, spec, num_blocks, num_layers, layout) |
| 571 | |
| 572 | for layer_idx, cache in enumerate(caches): |
| 573 | for block_idx in range(num_blocks): |
| 574 | for head_idx in range(cache.shape[1]): |
| 575 | cache[block_idx, head_idx].fill_( |
| 576 | 100 * layer_idx + 10 * head_idx + block_idx |
| 577 | ) |
| 578 | |
| 579 | expected = [[cache[i].clone() for i in range(num_blocks)] for cache in caches] |
| 580 | copy_kv_cache_blocks_inplace( |
| 581 | caches, |
| 582 | num_blocks, |
| 583 | [KVCacheBlockCopy(src_block_id=0, dst_block_id=2)], |
| 584 | ) |
| 585 | |
| 586 | for layer_idx, cache in enumerate(caches): |
| 587 | torch.testing.assert_close(cache[2], expected[layer_idx][0]) |
| 588 | torch.testing.assert_close(cache[1], expected[layer_idx][1]) |
| 589 | |
| 590 | |
| 591 | @pytest.mark.parametrize( |
| 592 | "layout,num_layers", |
| 593 | [ |
| 594 | (KVCacheLayout.LBHNC, 2), |
| 595 | # Splitting needs a manager block to be one dense page, which a |
| 596 | # block-outermost layout only gives when the block holds one layer. |
| 597 | (KVCacheLayout.BLHNC, 1), |
| 598 | ], |
| 599 | ) |
| 600 | def test_copy_kv_cache_blocks_with_virtual_block_splitting( |
| 601 | layout: KVCacheLayout, num_layers: int |
| 602 | ): |
| 603 | num_blocks = 4 |
| 604 | physical_per_logical = 2 |
| 605 | spec = FullAttentionSpec( |
| 606 | block_size=4, |
| 607 | num_kv_heads=1, |
| 608 | head_size=2, |
| 609 | dtype=torch.float32, |
| 610 | ) |
| 611 | raw = torch.zeros(num_blocks * num_layers * spec.page_size_bytes, dtype=torch.int8) |
| 612 | caches = dense_kv_cache_views( |
| 613 | raw, |
| 614 | spec, |
| 615 | num_blocks, |
| 616 | num_layers, |
| 617 | layout, |
| 618 | kernel_block_size=spec.block_size // physical_per_logical, |
| 619 | ) |
| 620 | |
| 621 | for layer_idx, cache in enumerate(caches): |
| 622 | for block_idx in range(cache.shape[0]): |
| 623 | cache[block_idx].fill_(100 * layer_idx + block_idx) |
| 624 | expected = [[cache[i].clone() for i in range(cache.shape[0])] for cache in caches] |
| 625 | |
| 626 | copy_kv_cache_blocks_inplace( |
| 627 | caches, |
| 628 | num_blocks, |
| 629 | [KVCacheBlockCopy(src_block_id=0, dst_block_id=2)], |
| 630 | ) |
| 631 | |
| 632 | dst_start = 2 * physical_per_logical |
| 633 | for layer_idx, cache in enumerate(caches): |
| 634 | for physical_idx in range(physical_per_logical): |
| 635 | torch.testing.assert_close( |
| 636 | cache[dst_start + physical_idx], expected[layer_idx][physical_idx] |
| 637 | ) |
| 638 | |
| 639 | |
| 640 | def test_allocate_hisparse_kv_caches_host_pool_and_view_less_specs(): |
| 641 | """Host tensors get their own backing; view-less specs keep the raw one.""" |
| 642 | spec = FullAttentionSpec( |
| 643 | block_size=2, num_kv_heads=1, head_size=4, dtype=torch.float32 |
| 644 | ) |
| 645 | page = spec.page_size_bytes |
| 646 | resident_spec = HiSparseResidentSpec(block_size=2, page_size=page) |
| 647 | device_size = 4 * page |
| 648 | config = KVCacheConfig( |
| 649 | num_blocks=4, |
| 650 | hisparse_host_num_blocks=3, |
| 651 | kv_cache_tensors=[ |
| 652 | KVCacheTensor( |
| 653 | size=3 * page, |
| 654 | layers=["source"], |
| 655 | layer_stride=3 * page, |
| 656 | block_stride=page, |
| 657 | host_resident=True, |
| 658 | ), |
| 659 | KVCacheTensor( |
| 660 | size=device_size, |
| 661 | layers=["indexer"], |
| 662 | layer_stride=device_size, |
| 663 | block_stride=page, |
| 664 | ), |
| 665 | KVCacheTensor( |
| 666 | size=device_size, |
| 667 | layers=["resident"], |
| 668 | layer_stride=device_size, |
| 669 | block_stride=page, |
| 670 | ), |
| 671 | ], |
| 672 | kv_cache_groups=[ |
| 673 | KVCacheGroupSpec(["source"], spec, host_resident=True), |
| 674 | KVCacheGroupSpec(["indexer"], spec), |
| 675 | KVCacheGroupSpec(["resident"], resident_spec), |
| 676 | ], |
| 677 | ) |
| 678 | host_buffers: list[torch.Tensor] = [] |
| 679 | |
| 680 | def host_allocator(size: int) -> torch.Tensor: |
| 681 | host_buffers.append(torch.zeros(size, dtype=torch.int8)) |
| 682 | return host_buffers[-1] |
| 683 | |
| 684 | caches = allocate_hisparse_kv_caches( |
| 685 | config, |
| 686 | torch.device("cpu"), |
| 687 | KVCacheLayout.LBHNC, |
| 688 | [2, 2, 2], |
| 689 | SimpleNamespace(allocate=host_allocator), |
| 690 | ) |
| 691 | assert len(config.kv_cache_tensors) == 3 |
| 692 | |
| 693 | assert [buf.numel() for buf in host_buffers] == [3 * page] |
| 694 | assert caches["source"].shape[0] == 3 |
| 695 | assert ( |
| 696 | caches["source"].untyped_storage().data_ptr() |
| 697 | == host_buffers[0].untyped_storage().data_ptr() |
| 698 | ) |
| 699 | assert caches["indexer"].shape[0] == 4 |
| 700 | backing = caches["resident"] |
| 701 | assert backing.dtype == torch.int8 and backing.numel() >= device_size |
| 702 | assert ( |
| 703 | backing.untyped_storage().data_ptr() |
| 704 | == caches["indexer"].untyped_storage().data_ptr() |
| 705 | ) |