| 1 | # SPDX-License-Identifier: Apache-2.0 |
| 2 | # SPDX-FileCopyrightText: Copyright contributors to the vLLM project |
| 3 | from dataclasses import dataclass |
| 4 | from typing import Any |
| 5 | from unittest.mock import MagicMock, patch |
| 6 | |
| 7 | import numpy as np |
| 8 | import pytest |
| 9 | import torch |
| 10 | |
| 11 | from vllm.model_executor.layers.mamba.mamba_utils import ( |
| 12 | MambaStateCopyFunc, |
| 13 | MambaStateCopyFuncsByType, |
| 14 | get_conv_copy_spec, |
| 15 | get_temporal_copy_spec, |
| 16 | ) |
| 17 | from vllm.v1.attention.backends.registry import MambaAttentionBackendEnum |
| 18 | from vllm.v1.core.sched.output import CachedRequestData, SchedulerOutput |
| 19 | from vllm.v1.kv_cache_interface import ( |
| 20 | KVCacheConfig, |
| 21 | KVCacheGroupSpec, |
| 22 | MambaSpec, |
| 23 | UniformTypeKVCacheSpecs, |
| 24 | ) |
| 25 | from vllm.v1.worker.gpu.model_states.mamba_hybrid import MambaHybridModelState |
| 26 | from vllm.v1.worker.mamba_utils import ( |
| 27 | MambaCopyBuffers, |
| 28 | MambaSpecDecodeGPUContext, |
| 29 | _reinterpret_u64_as_i64, |
| 30 | batch_memcpy, |
| 31 | collect_mamba_copy_meta, |
| 32 | do_mamba_copy_block, |
| 33 | get_mamba_groups, |
| 34 | preprocess_mamba, |
| 35 | stage_postprocess_inputs_to_gpu, |
| 36 | ) |
| 37 | |
| 38 | # Conv + temporal copy specs, in the order the tests' MambaSpec shapes expect. |
| 39 | _DEFAULT_COPY_FUNCS: tuple[MambaStateCopyFunc, ...] = ( |
| 40 | get_conv_copy_spec, |
| 41 | get_temporal_copy_spec, |
| 42 | ) |
| 43 | _COPY_FUNCS: MambaStateCopyFuncsByType = { |
| 44 | MambaAttentionBackendEnum.MAMBA2: _DEFAULT_COPY_FUNCS, |
| 45 | MambaAttentionBackendEnum.GDN_ATTN: _DEFAULT_COPY_FUNCS, |
| 46 | MambaAttentionBackendEnum.SHORT_CONV: (get_conv_copy_spec,), |
| 47 | } |
| 48 | |
| 49 | |
| 50 | def postprocess_mamba( |
| 51 | scheduler_output: "SchedulerOutput", |
| 52 | kv_cache_config: "KVCacheConfig", |
| 53 | input_batch: Any, |
| 54 | requests: dict[str, Any], |
| 55 | forward_context: dict[str, Any], |
| 56 | mamba_state_copy_funcs: MambaStateCopyFuncsByType, |
| 57 | copy_bufs: "MambaCopyBuffers", |
| 58 | ): |
| 59 | """CPU reference for the align-mode postprocess. |
| 60 | |
| 61 | Used as a golden against the GPU fused kernel (``postprocess_mamba_align_gpu``). |
| 62 | Mirrors what the production code did before the fused kernel replaced it; |
| 63 | kept here because production no longer has a CPU implementation. |
| 64 | """ |
| 65 | assert input_batch.mamba_state_idx_cpu is not None |
| 66 | num_scheduled_tokens_dict = scheduler_output.num_scheduled_tokens |
| 67 | scheduled_spec_decode_tokens_dict = scheduler_output.scheduled_spec_decode_tokens |
| 68 | num_accepted_tokens_cpu = input_batch.num_accepted_tokens_cpu |
| 69 | mamba_state_idx_cpu = input_batch.mamba_state_idx_cpu |
| 70 | mamba_group_ids = copy_bufs.mamba_group_ids |
| 71 | mamba_spec = copy_bufs.mamba_spec |
| 72 | copy_bufs.offset = 0 |
| 73 | for i, req_id in enumerate(input_batch.req_ids): |
| 74 | req_state = requests[req_id] |
| 75 | num_computed_tokens = req_state.num_computed_tokens |
| 76 | num_draft_tokens = len(scheduled_spec_decode_tokens_dict.get(req_id, [])) |
| 77 | num_scheduled_tokens = num_scheduled_tokens_dict[req_id] |
| 78 | num_accepted_tokens = num_accepted_tokens_cpu[i] |
| 79 | num_tokens_running_state = ( |
| 80 | num_computed_tokens + num_scheduled_tokens - num_draft_tokens |
| 81 | ) |
| 82 | new_num_computed_tokens = num_tokens_running_state + num_accepted_tokens - 1 |
| 83 | aligned_new_computed_tokens = ( |
| 84 | new_num_computed_tokens // mamba_spec.block_size * mamba_spec.block_size |
| 85 | ) |
| 86 | if aligned_new_computed_tokens >= num_tokens_running_state: |
| 87 | accept_token_bias = aligned_new_computed_tokens - num_tokens_running_state |
| 88 | src_block_idx = mamba_state_idx_cpu[i] |
| 89 | dest_block_idx = aligned_new_computed_tokens // mamba_spec.block_size - 1 |
| 90 | collect_mamba_copy_meta( |
| 91 | copy_bufs, |
| 92 | kv_cache_config, |
| 93 | mamba_state_copy_funcs, |
| 94 | mamba_group_ids, |
| 95 | src_block_idx, |
| 96 | dest_block_idx, |
| 97 | accept_token_bias, |
| 98 | req_state, |
| 99 | forward_context, |
| 100 | ) |
| 101 | if src_block_idx == dest_block_idx: |
| 102 | num_accepted_tokens_cpu[i] = 1 |
| 103 | do_mamba_copy_block(copy_bufs) |
| 104 | |
| 105 | |
| 106 | def _make_scheduler_output( |
| 107 | finished_req_ids: set[str], |
| 108 | preempted_req_ids: set[str] | None, |
| 109 | resumed_req_ids: set[str], |
| 110 | ) -> SchedulerOutput: |
| 111 | cached = CachedRequestData.make_empty() |
| 112 | cached.resumed_req_ids = resumed_req_ids |
| 113 | return SchedulerOutput( |
| 114 | scheduled_new_reqs=[], |
| 115 | scheduled_cached_reqs=cached, |
| 116 | num_scheduled_tokens={}, |
| 117 | total_num_scheduled_tokens=0, |
| 118 | scheduled_spec_decode_tokens={}, |
| 119 | scheduled_encoder_inputs={}, |
| 120 | num_common_prefix_blocks=[], |
| 121 | finished_req_ids=finished_req_ids, |
| 122 | free_encoder_mm_hashes=[], |
| 123 | preempted_req_ids=preempted_req_ids, |
| 124 | ) |
| 125 | |
| 126 | |
| 127 | def test_resumed_req_ids_cleared_from_mamba_state_idx(): |
| 128 | """When a request is force-preempted (e.g. reset_prefix_cache), |
| 129 | it appears in resumed_req_ids but NOT in preempted_req_ids. |
| 130 | preprocess_mamba must still clear its mamba_state_idx entry, |
| 131 | otherwise stale indices can point beyond the new block allocation. |
| 132 | """ |
| 133 | spec = MagicMock(block_size=64, num_speculative_blocks=0) |
| 134 | cache_config = MagicMock(enable_prefix_caching=True) |
| 135 | input_batch = MagicMock(req_ids=[]) |
| 136 | copy_bufs = MagicMock(mamba_group_ids=[0], mamba_spec=spec) |
| 137 | |
| 138 | mamba_state_idx: dict[str, int] = { |
| 139 | "finished": 1, |
| 140 | "preempted": 2, |
| 141 | "resumed": 3, # only in resumed_req_ids, NOT in preempted |
| 142 | "keep": 99, |
| 143 | } |
| 144 | sched = _make_scheduler_output( |
| 145 | finished_req_ids={"finished"}, |
| 146 | preempted_req_ids={"preempted"}, |
| 147 | resumed_req_ids={"resumed"}, |
| 148 | ) |
| 149 | |
| 150 | with patch( |
| 151 | "vllm.v1.worker.mamba_utils.get_mamba_groups", |
| 152 | return_value=([0], spec), |
| 153 | ): |
| 154 | preprocess_mamba( |
| 155 | sched, |
| 156 | MagicMock(), # kv_cache_config |
| 157 | cache_config, |
| 158 | mamba_state_idx, |
| 159 | input_batch, |
| 160 | {}, # requests |
| 161 | {}, # forward_context |
| 162 | {}, # mamba_state_copy_funcs |
| 163 | copy_bufs, |
| 164 | ) |
| 165 | |
| 166 | assert mamba_state_idx == {"keep": 99} |
| 167 | |
| 168 | |
| 169 | # ----------------------------------------------------------------------------- |
| 170 | # Golden tests for postprocess_mamba_fused_kernel |
| 171 | # ----------------------------------------------------------------------------- |
| 172 | |
| 173 | |
| 174 | @dataclass |
| 175 | class _TestConfig: |
| 176 | """Common test configuration for fused kernel tests.""" |
| 177 | |
| 178 | block_size: int = 16 |
| 179 | num_blocks: int = 32 |
| 180 | num_layers: int = 2 |
| 181 | num_reqs: int = 4 |
| 182 | max_num_reqs: int = 8 |
| 183 | # Conv state shape: [num_blocks, conv_width, inner_dim] |
| 184 | conv_width: int = 4 |
| 185 | conv_inner_dim: int = 64 |
| 186 | # Temporal state shape: [num_blocks, state_dim] |
| 187 | temporal_state_dim: int = 128 |
| 188 | dtype: torch.dtype = torch.float16 |
| 189 | |
| 190 | |
| 191 | class _MockCpuGpuBuffer: |
| 192 | """Mock CpuGpuBuffer for testing without pinned memory.""" |
| 193 | |
| 194 | def __init__(self, size: int, dtype: torch.dtype, device: torch.device): |
| 195 | self.cpu = torch.zeros(size, dtype=dtype, device="cpu") |
| 196 | self.gpu = torch.zeros(size, dtype=dtype, device=device) |
| 197 | self.np = self.cpu.numpy() |
| 198 | |
| 199 | def copy_to_gpu(self, n: int | None = None) -> torch.Tensor: |
| 200 | if n is None: |
| 201 | return self.gpu.copy_(self.cpu, non_blocking=True) |
| 202 | return self.gpu[:n].copy_(self.cpu[:n], non_blocking=True) |
| 203 | |
| 204 | |
| 205 | class _FakeDataPtrTensor: |
| 206 | """Tensor wrapper that exposes a controlled data_ptr for metadata tests.""" |
| 207 | |
| 208 | def __init__(self, tensor: torch.Tensor, data_ptr: int): |
| 209 | self._tensor = tensor |
| 210 | self._data_ptr = data_ptr |
| 211 | self.shape = tensor.shape |
| 212 | |
| 213 | def data_ptr(self) -> int: |
| 214 | return self._data_ptr |
| 215 | |
| 216 | def dim(self) -> int: |
| 217 | return self._tensor.dim() |
| 218 | |
| 219 | def stride(self, *args): |
| 220 | return self._tensor.stride(*args) |
| 221 | |
| 222 | def numel(self) -> int: |
| 223 | return self._tensor.numel() |
| 224 | |
| 225 | def element_size(self) -> int: |
| 226 | return self._tensor.element_size() |
| 227 | |
| 228 | def size(self, *args): |
| 229 | return self._tensor.size(*args) |
| 230 | |
| 231 | def __getitem__(self, item): |
| 232 | return self._tensor[item] |
| 233 | |
| 234 | |
| 235 | def test_reinterpret_u64_as_i64_preserves_pointer_bits(): |
| 236 | ptrs = [ |
| 237 | 0, |
| 238 | 1, |
| 239 | (1 << 63) - 1, |
| 240 | 1 << 63, |
| 241 | (1 << 63) + 1234, |
| 242 | (1 << 64) - 1, |
| 243 | ] |
| 244 | ptr_tensor = torch.zeros(len(ptrs), dtype=torch.int64) |
| 245 | |
| 246 | for idx, ptr in enumerate(ptrs): |
| 247 | ptr_tensor[idx] = _reinterpret_u64_as_i64(ptr) |
| 248 | |
| 249 | assert ptr_tensor.numpy().view(np.uint64).tolist() == ptrs |
| 250 | |
| 251 | |
| 252 | def test_gpu_context_reinterprets_high_data_ptrs_for_int64_metadata(): |
| 253 | cfg = _TestConfig(num_layers=1) |
| 254 | device = torch.device("cpu") |
| 255 | kv_cache_config = _make_kv_cache_config(cfg, ["layer_0"]) |
| 256 | gpu_ctx = _make_gpu_ctx(cfg, kv_cache_config, device) |
| 257 | conv_ptr = 1 << 63 |
| 258 | temporal_ptr = (1 << 64) - 8 |
| 259 | block_table_ptr = (1 << 63) + 42 |
| 260 | |
| 261 | conv_state = _FakeDataPtrTensor( |
| 262 | torch.empty( |
| 263 | cfg.num_blocks, |
| 264 | cfg.conv_width, |
| 265 | cfg.conv_inner_dim, |
| 266 | dtype=cfg.dtype, |
| 267 | ), |
| 268 | conv_ptr, |
| 269 | ) |
| 270 | temporal_state = _FakeDataPtrTensor( |
| 271 | torch.empty(cfg.num_blocks, cfg.temporal_state_dim, dtype=cfg.dtype), |
| 272 | temporal_ptr, |
| 273 | ) |
| 274 | block_table = _FakeDataPtrTensor( |
| 275 | torch.empty(1, 4, dtype=torch.int32), |
| 276 | block_table_ptr, |
| 277 | ) |
| 278 | forward_context = {"layer_0": _make_mock_attention(conv_state, temporal_state)} |
| 279 | |
| 280 | gpu_ctx.initialize_from_forward_context( |
| 281 | kv_cache_config, forward_context, _COPY_FUNCS, [block_table] |
| 282 | ) |
| 283 | |
| 284 | assert gpu_ctx.state_base_addrs.tolist() == [ |
| 285 | _reinterpret_u64_as_i64(conv_ptr), |
| 286 | _reinterpret_u64_as_i64(temporal_ptr), |
| 287 | ] |
| 288 | assert gpu_ctx.block_table_ptrs.tolist() == [ |
| 289 | _reinterpret_u64_as_i64(block_table_ptr) |
| 290 | ] |
| 291 | |
| 292 | |
| 293 | def _make_postprocess_scheduler_output( |
| 294 | req_ids: list[str], |
| 295 | num_scheduled_tokens: dict[str, int], |
| 296 | scheduled_spec_decode_tokens: dict[str, list] | None = None, |
| 297 | ) -> SchedulerOutput: |
| 298 | """Create a minimal SchedulerOutput for postprocess testing.""" |
| 299 | cached = CachedRequestData.make_empty() |
| 300 | return SchedulerOutput( |
| 301 | scheduled_new_reqs=[], |
| 302 | scheduled_cached_reqs=cached, |
| 303 | num_scheduled_tokens=num_scheduled_tokens, |
| 304 | total_num_scheduled_tokens=sum(num_scheduled_tokens.values()), |
| 305 | scheduled_spec_decode_tokens=scheduled_spec_decode_tokens or {}, |
| 306 | scheduled_encoder_inputs={}, |
| 307 | num_common_prefix_blocks=[], |
| 308 | finished_req_ids=set(), |
| 309 | free_encoder_mm_hashes=[], |
| 310 | preempted_req_ids=set(), |
| 311 | ) |
| 312 | |
| 313 | |
| 314 | def _make_mock_attention( |
| 315 | conv_state: torch.Tensor, temporal_state: torch.Tensor |
| 316 | ) -> MagicMock: |
| 317 | """Create a mock attention object with kv_cache.""" |
| 318 | attention = MagicMock() |
| 319 | attention.kv_cache = [conv_state, temporal_state] |
| 320 | return attention |
| 321 | |
| 322 | |
| 323 | def _make_dual_states( |
| 324 | cfg: "_TestConfig", |
| 325 | layer_names: list[str], |
| 326 | device: torch.device, |
| 327 | *, |
| 328 | num_blocks: int | None = None, |
| 329 | ) -> tuple[ |
| 330 | list[torch.Tensor], |
| 331 | list[torch.Tensor], |
| 332 | list[torch.Tensor], |
| 333 | list[torch.Tensor], |
| 334 | dict[str, MagicMock], |
| 335 | dict[str, MagicMock], |
| 336 | ]: |
| 337 | """Allocate conv+temporal state tensors for the Python path, clone them for |
| 338 | the GPU path, and build matching ``forward_context`` dicts for both. |
| 339 | |
| 340 | Returns ``(conv_py, temporal_py, conv_gpu, temporal_gpu, fwd_py, fwd_gpu)`` |
| 341 | where the four state lists are parallel to ``layer_names``. |
| 342 | """ |
| 343 | n_blocks = num_blocks if num_blocks is not None else cfg.num_blocks |
| 344 | conv_py = [ |
| 345 | torch.randn( |
| 346 | n_blocks, |
| 347 | cfg.conv_width, |
| 348 | cfg.conv_inner_dim, |
| 349 | dtype=cfg.dtype, |
| 350 | device=device, |
| 351 | ) |
| 352 | for _ in layer_names |
| 353 | ] |
| 354 | temporal_py = [ |
| 355 | torch.randn(n_blocks, cfg.temporal_state_dim, dtype=cfg.dtype, device=device) |
| 356 | for _ in layer_names |
| 357 | ] |
| 358 | conv_gpu = [s.clone() for s in conv_py] |
| 359 | temporal_gpu = [s.clone() for s in temporal_py] |
| 360 | fwd_py = { |
| 361 | name: _make_mock_attention(c, t) |
| 362 | for name, c, t in zip(layer_names, conv_py, temporal_py) |
| 363 | } |
| 364 | fwd_gpu = { |
| 365 | name: _make_mock_attention(c, t) |
| 366 | for name, c, t in zip(layer_names, conv_gpu, temporal_gpu) |
| 367 | } |
| 368 | return conv_py, temporal_py, conv_gpu, temporal_gpu, fwd_py, fwd_gpu |
| 369 | |
| 370 | |
| 371 | def _make_dual_layer_state( |
| 372 | cfg: "_TestConfig", |
| 373 | device: torch.device, |
| 374 | *, |
| 375 | num_blocks: int | None = None, |
| 376 | layer_name: str = "layer_0", |
| 377 | ) -> tuple[ |
| 378 | torch.Tensor, |
| 379 | torch.Tensor, |
| 380 | torch.Tensor, |
| 381 | torch.Tensor, |
| 382 | dict[str, MagicMock], |
| 383 | dict[str, MagicMock], |
| 384 | ]: |
| 385 | """Single-layer convenience form of ``_make_dual_states``.""" |
| 386 | conv_py, temporal_py, conv_gpu, temporal_gpu, fwd_py, fwd_gpu = _make_dual_states( |
| 387 | cfg, [layer_name], device, num_blocks=num_blocks |
| 388 | ) |
| 389 | return conv_py[0], temporal_py[0], conv_gpu[0], temporal_gpu[0], fwd_py, fwd_gpu |
| 390 | |
| 391 | |
| 392 | def _make_kv_cache_config(cfg: _TestConfig, layer_names: list[str]) -> KVCacheConfig: |
| 393 | """Create a KVCacheConfig with mamba groups.""" |
| 394 | mamba_spec = MambaSpec( |
| 395 | block_size=cfg.block_size, |
| 396 | shapes=( |
| 397 | (cfg.conv_width, cfg.conv_inner_dim), |
| 398 | (cfg.temporal_state_dim,), |
| 399 | ), |
| 400 | dtypes=(cfg.dtype, cfg.dtype), |
| 401 | mamba_cache_mode="align", |
| 402 | ) |
| 403 | group = KVCacheGroupSpec( |
| 404 | layer_names=layer_names, |
| 405 | kv_cache_spec=mamba_spec, |
| 406 | ) |
| 407 | return KVCacheConfig( |
| 408 | num_blocks=cfg.num_blocks, |
| 409 | kv_cache_tensors=[], |
| 410 | kv_cache_groups=[group], |
| 411 | ) |
| 412 | |
| 413 | |
| 414 | def _make_input_batch( |
| 415 | req_ids: list[str], |
| 416 | num_accepted_tokens: list[int], |
| 417 | mamba_state_idx: list[int], |
| 418 | ) -> MagicMock: |
| 419 | """Create a mock GPUInputBatch.""" |
| 420 | batch = MagicMock() |
| 421 | batch.req_ids = req_ids |
| 422 | batch.req_id_to_index = {rid: i for i, rid in enumerate(req_ids)} |
| 423 | # Use numpy arrays so modifications persist |
| 424 | batch.num_accepted_tokens_cpu = np.array(num_accepted_tokens, dtype=np.int32) |
| 425 | batch.mamba_state_idx_cpu = np.array(mamba_state_idx, dtype=np.int32) |
| 426 | return batch |
| 427 | |
| 428 | |
| 429 | def _make_requests( |
| 430 | req_ids: list[str], |
| 431 | num_computed_tokens: list[int], |
| 432 | block_ids_per_req: list[list[int]], |
| 433 | ) -> dict[str, MagicMock]: |
| 434 | """Create mock CachedRequestState objects.""" |
| 435 | requests = {} |
| 436 | for i, req_id in enumerate(req_ids): |
| 437 | req = MagicMock() |
| 438 | req.num_computed_tokens = num_computed_tokens[i] |
| 439 | req.block_ids = {0: block_ids_per_req[i]} # group_id=0 |
| 440 | requests[req_id] = req |
| 441 | return requests |
| 442 | |
| 443 | |
| 444 | def _make_copy_bufs( |
| 445 | cfg: _TestConfig, |
| 446 | kv_cache_config: KVCacheConfig, |
| 447 | device: torch.device, |
| 448 | copy_funcs: MambaStateCopyFuncsByType = _COPY_FUNCS, |
| 449 | ) -> MambaCopyBuffers: |
| 450 | """Create MambaCopyBuffers for the Python path.""" |
| 451 | |
| 452 | def make_buffer(n, dtype): |
| 453 | return _MockCpuGpuBuffer(n, dtype, device) |
| 454 | |
| 455 | return MambaCopyBuffers.create( |
| 456 | max_num_reqs=cfg.max_num_reqs, |
| 457 | kv_cache_config=kv_cache_config, |
| 458 | copy_funcs=copy_funcs, |
| 459 | make_buffer=make_buffer, |
| 460 | ) |
| 461 | |
| 462 | |
| 463 | def _make_gpu_ctx( |
| 464 | cfg: _TestConfig, |
| 465 | kv_cache_config: KVCacheConfig, |
| 466 | device: torch.device, |
| 467 | copy_funcs: MambaStateCopyFuncsByType = _COPY_FUNCS, |
| 468 | ) -> MambaSpecDecodeGPUContext: |
| 469 | """Create MambaSpecDecodeGPUContext for the GPU path.""" |
| 470 | |
| 471 | def make_buffer(n, dtype): |
| 472 | return _MockCpuGpuBuffer(n, dtype, device) |
| 473 | |
| 474 | return MambaSpecDecodeGPUContext.create( |
| 475 | max_num_reqs=cfg.max_num_reqs, |
| 476 | kv_cache_config=kv_cache_config, |
| 477 | copy_funcs=copy_funcs, |
| 478 | device=device, |
| 479 | make_buffer=make_buffer, |
| 480 | ) |
| 481 | |
| 482 | |
| 483 | def test_mamba_groups_support_different_state_specs(): |
| 484 | gdn_spec = MambaSpec( |
| 485 | block_size=16, |
| 486 | shapes=((4, 4), (2, 4, 4)), |
| 487 | dtypes=(torch.float16, torch.float32), |
| 488 | mamba_type=MambaAttentionBackendEnum.GDN_ATTN, |
| 489 | mamba_cache_mode="align", |
| 490 | num_speculative_blocks=3, |
| 491 | ) |
| 492 | short_conv_spec = MambaSpec( |
| 493 | block_size=16, |
| 494 | shapes=((12, 8),), |
| 495 | dtypes=(torch.float16,), |
| 496 | mamba_type=MambaAttentionBackendEnum.SHORT_CONV, |
| 497 | mamba_cache_mode="align", |
| 498 | num_speculative_blocks=3, |
| 499 | ) |
| 500 | kv_cache_config = KVCacheConfig( |
| 501 | num_blocks=8, |
| 502 | kv_cache_tensors=[], |
| 503 | kv_cache_groups=[ |
| 504 | KVCacheGroupSpec(["gdn.0", "gdn.1"], gdn_spec), |
| 505 | KVCacheGroupSpec(["ple.0"], short_conv_spec), |
| 506 | ], |
| 507 | ) |
| 508 | |
| 509 | mamba_groups = get_mamba_groups(kv_cache_config) |
| 510 | assert mamba_groups == {gdn_spec: [0], short_conv_spec: [1]} |
| 511 | |
| 512 | model_state = object.__new__(MambaHybridModelState) |
| 513 | model_state._mamba_group_ids = [] |
| 514 | model_state._mamba_spec = None |
| 515 | group_ids, representative_spec = model_state._get_mamba_group_info(kv_cache_config) |
| 516 | assert group_ids == [0, 1] |
| 517 | assert representative_spec is gdn_spec |
| 518 | |
| 519 | cfg = _TestConfig() |
| 520 | copy_bufs = _make_copy_bufs(cfg, kv_cache_config, torch.device("cpu")) |
| 521 | assert copy_bufs.src_ptrs.cpu.numel() == cfg.max_num_reqs * 5 |
| 522 | assert copy_bufs.dst_ptrs.cpu.numel() == cfg.max_num_reqs * 5 |
| 523 | assert copy_bufs.sizes.cpu.numel() == cfg.max_num_reqs * 5 |
| 524 | |
| 525 | ctx = _make_gpu_ctx(cfg, kv_cache_config, torch.device("cpu")) |
| 526 | assert ctx.num_states == 5 |
| 527 | |
| 528 | forward_context = {} |
| 529 | gdn_spec_layer_names = ["gdn.0", "gdn.1"] |
| 530 | for layer_name in gdn_spec_layer_names: |
| 531 | attention = MagicMock() |
| 532 | attention.kv_cache = ( |
| 533 | torch.empty(8, 4, 4, dtype=torch.float16), |
| 534 | torch.empty(8, 2, 4, 4, dtype=torch.float32), |
| 535 | ) |
| 536 | forward_context[layer_name] = attention |
| 537 | ple_attention = MagicMock() |
| 538 | ple_attention.kv_cache = (torch.empty(8, 12, 8, dtype=torch.float16),) |
| 539 | forward_context["ple.0"] = ple_attention |
| 540 | assert gdn_spec_layer_names == kv_cache_config.kv_cache_groups[0].layer_names |
| 541 | |
| 542 | block_tables = [ |
| 543 | torch.zeros(4, 8, dtype=torch.int32), |
| 544 | torch.zeros(4, 8, dtype=torch.int32), |
| 545 | ] |
| 546 | ctx.initialize_from_forward_context( |
| 547 | kv_cache_config, |
| 548 | forward_context, |
| 549 | _COPY_FUNCS, |
| 550 | block_tables, |
| 551 | ) |
| 552 | assert ctx.is_initialized |
| 553 | assert ctx.state_group_indices.tolist() == [0, 0, 0, 0, 1] |
| 554 | assert ctx.state_conv_widths.tolist() == [4, 0, 4, 0, 12] |
| 555 | |
| 556 | |
| 557 | def test_mamba_groups_support_mixed_specs_in_uniform_group(): |
| 558 | gdn_spec = MambaSpec( |
| 559 | block_size=16, |
| 560 | shapes=((4, 4), (2, 4, 4)), |
| 561 | dtypes=(torch.float16, torch.float32), |
| 562 | mamba_type=MambaAttentionBackendEnum.GDN_ATTN, |
| 563 | mamba_cache_mode="align", |
| 564 | num_speculative_blocks=3, |
| 565 | ) |
| 566 | short_conv_spec = MambaSpec( |
| 567 | block_size=16, |
| 568 | shapes=((12, 8),), |
| 569 | dtypes=(torch.float16,), |
| 570 | mamba_type=MambaAttentionBackendEnum.SHORT_CONV, |
| 571 | mamba_cache_mode="align", |
| 572 | num_speculative_blocks=3, |
| 573 | ) |
| 574 | layer_specs = { |
| 575 | "gdn.0": gdn_spec, |
| 576 | "gdn.1": gdn_spec, |
| 577 | "ple.0": short_conv_spec, |
| 578 | } |
| 579 | kv_cache_config = KVCacheConfig( |
| 580 | num_blocks=8, |
| 581 | kv_cache_tensors=[], |
| 582 | kv_cache_groups=[ |
| 583 | KVCacheGroupSpec( |
| 584 | list(layer_specs), |
| 585 | UniformTypeKVCacheSpecs( |
| 586 | block_size=16, |
| 587 | kv_cache_specs=layer_specs, |
| 588 | ), |
| 589 | ) |
| 590 | ], |
| 591 | ) |
| 592 | |
| 593 | assert kv_cache_config.has_mamba_layers |
| 594 | assert get_mamba_groups(kv_cache_config) == { |
| 595 | gdn_spec: [0], |
| 596 | short_conv_spec: [0], |
| 597 | } |
| 598 | |
| 599 | # The group is listed once per spec, so its id must not be repeated. |
| 600 | model_state = object.__new__(MambaHybridModelState) |
| 601 | model_state._mamba_group_ids = [] |
| 602 | model_state._mamba_spec = None |
| 603 | group_ids, representative_spec = model_state._get_mamba_group_info(kv_cache_config) |
| 604 | assert group_ids == [0] |
| 605 | assert representative_spec is gdn_spec |
| 606 | |
| 607 | cfg = _TestConfig() |
| 608 | copy_bufs = _make_copy_bufs(cfg, kv_cache_config, torch.device("cpu")) |
| 609 | assert copy_bufs.mamba_group_ids == [0] |
| 610 | assert copy_bufs.src_ptrs.cpu.numel() == cfg.max_num_reqs * 5 |
| 611 | |
| 612 | ctx = _make_gpu_ctx(cfg, kv_cache_config, torch.device("cpu")) |
| 613 | assert ctx.mamba_group_ids == [0] |
| 614 | assert ctx.num_states == 5 |
| 615 | |
| 616 | forward_context = { |
| 617 | "gdn.0": MagicMock( |
| 618 | kv_cache=( |
| 619 | torch.empty(8, 4, 4, dtype=torch.float16), |
| 620 | torch.empty(8, 2, 4, 4, dtype=torch.float32), |
| 621 | ) |
| 622 | ), |
| 623 | "gdn.1": MagicMock( |
| 624 | kv_cache=( |
| 625 | torch.empty(8, 4, 4, dtype=torch.float16), |
| 626 | torch.empty(8, 2, 4, 4, dtype=torch.float32), |
| 627 | ) |
| 628 | ), |
| 629 | "ple.0": MagicMock(kv_cache=(torch.empty(8, 12, 8, dtype=torch.float16),)), |
| 630 | } |
| 631 | ctx.initialize_from_forward_context( |
| 632 | kv_cache_config, |
| 633 | forward_context, |
| 634 | _COPY_FUNCS, |
| 635 | [torch.zeros(4, 8, dtype=torch.int32)], |
| 636 | ) |
| 637 | assert ctx.state_group_indices.tolist() == [0, 0, 0, 0, 0] |
| 638 | assert ctx.state_conv_widths.tolist() == [4, 0, 4, 0, 12] |
| 639 | |
| 640 | |
| 641 | # ----------------------------------------------------------------------------- |
| 642 | # stage_postprocess_inputs_to_gpu: single-pass staging into pinned views |
| 643 | # ----------------------------------------------------------------------------- |
| 644 | |
| 645 | |
| 646 | def _make_staging_ctx(max_num_reqs: int, device: torch.device) -> MagicMock: |
| 647 | """Build a MambaSpecDecodeGPUContext stand-in exposing only the four |
| 648 | per-request staging buffers touched by stage_postprocess_inputs_to_gpu.""" |
| 649 | ctx = MagicMock() |
| 650 | ctx.mamba_state_idx_buf = _MockCpuGpuBuffer(max_num_reqs, torch.int32, device) |
| 651 | ctx.num_scheduled_tokens_buf = _MockCpuGpuBuffer(max_num_reqs, torch.int32, device) |
| 652 | ctx.num_computed_tokens_buf = _MockCpuGpuBuffer(max_num_reqs, torch.int32, device) |
| 653 | ctx.num_draft_tokens_buf = _MockCpuGpuBuffer(max_num_reqs, torch.int32, device) |
| 654 | return ctx |
| 655 | |
| 656 | |
| 657 | def test_stage_postprocess_inputs_to_gpu_fills_pinned_views(): |
| 658 | """stage_postprocess_inputs_to_gpu writes each request's state_idx and |
| 659 | scheduled/computed/draft counts into slot `i` for `req_ids[i]`, leaves |
| 660 | slots past `num_reqs` untouched, and mirrors the values to the GPU buffers.""" |
| 661 | device = torch.device("cpu") |
| 662 | max_num_reqs = 8 |
| 663 | ctx = _make_staging_ctx(max_num_reqs, device) |
| 664 | |
| 665 | # Any negative int32 works as a sentinel: all staged values (state_idx, |
| 666 | # scheduled/computed/draft token counts) are non-negative, so a negative |
| 667 | # entry surviving in slots [num_reqs:] proves the loop didn't overrun. |
| 668 | sentinel = np.int32(-1) |
| 669 | bufs = ( |
| 670 | ctx.mamba_state_idx_buf, |
| 671 | ctx.num_scheduled_tokens_buf, |
| 672 | ctx.num_computed_tokens_buf, |
| 673 | ctx.num_draft_tokens_buf, |
| 674 | ) |
| 675 | for buf in bufs: |
| 676 | buf.np[:] = sentinel |
| 677 | |
| 678 | req_ids = ["req_a", "req_b", "req_c"] |
| 679 | num_reqs = len(req_ids) |
| 680 | scheduler_output = _make_postprocess_scheduler_output( |
| 681 | req_ids=req_ids, |
| 682 | num_scheduled_tokens={"req_a": 5, "req_b": 12, "req_c": 3}, |
| 683 | # req_b intentionally absent -> defaults to 0 drafts. |
| 684 | scheduled_spec_decode_tokens={ |
| 685 | "req_a": [1, 2], |
| 686 | "req_c": [1, 2, 3, 4], |
| 687 | }, |
| 688 | ) |
| 689 | requests = _make_requests( |
| 690 | req_ids=req_ids, |
| 691 | num_computed_tokens=[10, 20, 30], |
| 692 | block_ids_per_req=[[0], [0], [0]], |
| 693 | ) |
| 694 | mamba_state_idx = {"req_a": 100, "req_b": 200, "req_c": 300} |
| 695 | # A trailing entry past num_reqs must not be read. |
| 696 | req_ids_padded = req_ids + ["not_scheduled"] |
| 697 | |
| 698 | stage_postprocess_inputs_to_gpu( |
| 699 | ctx, |
| 700 | scheduler_output, |
| 701 | req_ids_padded, |
| 702 | num_reqs, |
| 703 | requests, |
| 704 | mamba_state_idx, |
| 705 | ) |
| 706 | |
| 707 | np.testing.assert_array_equal( |
| 708 | ctx.mamba_state_idx_buf.np[:num_reqs], [100, 200, 300] |
| 709 | ) |
| 710 | np.testing.assert_array_equal( |
| 711 | ctx.num_scheduled_tokens_buf.np[:num_reqs], [5, 12, 3] |
| 712 | ) |
| 713 | np.testing.assert_array_equal( |
| 714 | ctx.num_computed_tokens_buf.np[:num_reqs], [10, 20, 30] |
| 715 | ) |
| 716 | np.testing.assert_array_equal(ctx.num_draft_tokens_buf.np[:num_reqs], [2, 0, 4]) |
| 717 | for buf in bufs: |
| 718 | assert (buf.np[num_reqs:] == sentinel).all() |
| 719 | |
| 720 | assert torch.equal( |
| 721 | ctx.mamba_state_idx_buf.gpu[:num_reqs], |
| 722 | torch.tensor([100, 200, 300], dtype=torch.int32), |
| 723 | ) |
| 724 | assert torch.equal( |
| 725 | ctx.num_draft_tokens_buf.gpu[:num_reqs], |
| 726 | torch.tensor([2, 0, 4], dtype=torch.int32), |
| 727 | ) |
| 728 | |
| 729 | |
| 730 | def test_stage_postprocess_inputs_to_gpu_asserts_on_missing_state_idx(): |
| 731 | """If preprocess_mamba didn't populate mamba_state_idx for a req in the |
| 732 | batch, staging must fail loudly rather than silently writing a stale index.""" |
| 733 | device = torch.device("cpu") |
| 734 | ctx = _make_staging_ctx(max_num_reqs=4, device=device) |
| 735 | scheduler_output = _make_postprocess_scheduler_output( |
| 736 | req_ids=["req_a"], |
| 737 | num_scheduled_tokens={"req_a": 1}, |
| 738 | ) |
| 739 | requests = _make_requests(["req_a"], [0], [[0]]) |
| 740 | |
| 741 | # mamba_state_idx has an entry for a *different* request but not for |
| 742 | # req_a (the one being staged). This is the realistic failure mode: |
| 743 | # preprocess_mamba ran, populated some reqs, but missed this one. |
| 744 | # The safety assert in stage_postprocess_inputs_to_gpu must fire on |
| 745 | # req_a's missing entry rather than silently writing None. |
| 746 | mamba_state_idx: dict[str, int] = {"other_req": 42} |
| 747 | with pytest.raises(AssertionError, match="mamba_state_idx missing entry"): |
| 748 | stage_postprocess_inputs_to_gpu( |
| 749 | ctx, |
| 750 | scheduler_output, |
| 751 | ["req_a"], |
| 752 | 1, |
| 753 | requests, |
| 754 | mamba_state_idx, |
| 755 | ) |
| 756 | |
| 757 | |
| 758 | def test_gpu_context_ignores_auxiliary_cache_tensors() -> None: |
| 759 | device = torch.device("cpu") |
| 760 | config = _TestConfig(num_layers=1) |
| 761 | layer_names = ["layer_0"] |
| 762 | kv_cache_config = _make_kv_cache_config(config, layer_names) |
| 763 | conv_state = torch.empty( |
| 764 | config.num_blocks, |
| 765 | config.conv_width, |
| 766 | config.conv_inner_dim, |
| 767 | dtype=config.dtype, |
| 768 | ) |
| 769 | temporal_state = torch.empty( |
| 770 | config.num_blocks, config.temporal_state_dim, dtype=config.dtype |
| 771 | ) |
| 772 | attention = MagicMock() |
| 773 | attention.kv_cache = [ |
| 774 | conv_state, |
| 775 | temporal_state, |
| 776 | *(torch.empty(config.num_blocks, 1) for _ in range(4)), |
| 777 | ] |
| 778 | context = _make_gpu_ctx(config, kv_cache_config, device) |
| 779 | |
| 780 | context.initialize_from_forward_context( |
| 781 | kv_cache_config, |
| 782 | {"layer_0": attention}, |
| 783 | _COPY_FUNCS, |
| 784 | [torch.zeros(1, 1, dtype=torch.int32)], |
| 785 | ) |
| 786 | |
| 787 | assert context.is_initialized |
| 788 | assert context.state_base_addrs.tolist() == [ |
| 789 | conv_state.data_ptr(), |
| 790 | temporal_state.data_ptr(), |
| 791 | ] |
| 792 | |
| 793 | |
| 794 | def _run_gpu_postprocess( |
| 795 | gpu_ctx: MambaSpecDecodeGPUContext, |
| 796 | *, |
| 797 | kv_cache_config: KVCacheConfig, |
| 798 | forward_context: dict[str, Any], |
| 799 | copy_funcs: MambaStateCopyFuncsByType, |
| 800 | block_table: torch.Tensor, |
| 801 | req_ids: list[str], |
| 802 | num_accepted_tokens: list[int], |
| 803 | mamba_state_idx: list[int], |
| 804 | num_scheduled_tokens: dict[str, int], |
| 805 | num_computed_tokens: list[int], |
| 806 | num_draft_tokens: dict[str, int], |
| 807 | device: torch.device, |
| 808 | ) -> None: |
| 809 | """Initialize the GPU context against `block_table`, run the fused |
| 810 | postprocess kernel for `req_ids`, and synchronize.""" |
| 811 | |
| 812 | def t(values): |
| 813 | return torch.tensor(values, dtype=torch.int32, device=device) |
| 814 | |
| 815 | gpu_ctx.initialize_from_forward_context( |
| 816 | kv_cache_config, forward_context, copy_funcs, [block_table] |
| 817 | ) |
| 818 | gpu_ctx.run_fused_postprocess( |
| 819 | num_reqs=len(req_ids), |
| 820 | num_accepted_tokens_gpu=t(num_accepted_tokens), |
| 821 | mamba_state_idx_gpu=t(mamba_state_idx), |
| 822 | num_scheduled_tokens_gpu=t([num_scheduled_tokens[r] for r in req_ids]), |
| 823 | num_computed_tokens_gpu=t(num_computed_tokens), |
| 824 | num_draft_tokens_gpu=t([num_draft_tokens.get(r, 0) for r in req_ids]), |
| 825 | ) |
| 826 | torch.accelerator.synchronize() |
| 827 | |
| 828 | |
| 829 | @pytest.mark.skipif(not torch.cuda.is_available(), reason="CUDA required") |
| 830 | class TestPostprocessMambaFusedKernel: |
| 831 | """Tests for postprocess_mamba_fused_kernel comparing GPU vs CPU paths.""" |
| 832 | |
| 833 | @pytest.fixture |
| 834 | def device(self): |
| 835 | return torch.device("cuda:0") |
| 836 | |
| 837 | @pytest.fixture |
| 838 | def test_config(self): |
| 839 | return _TestConfig() |
| 840 | |
| 841 | def test_batch_memcpy_left_overlap_has_memmove_semantics(self, device): |
| 842 | batch = 128 |
| 843 | row_bytes = 32 * 1024 |
| 844 | shift = 16 |
| 845 | copy_size = row_bytes - shift |
| 846 | |
| 847 | pattern = (torch.arange(row_bytes, dtype=torch.int32, device=device) % 251).to( |
| 848 | torch.uint8 |
| 849 | ) |
| 850 | state = pattern.expand(batch, -1).clone() |
| 851 | snapshot = state.clone() |
| 852 | |
| 853 | row_stride_bytes = state.stride(0) * state.element_size() |
| 854 | row_offsets = ( |
| 855 | torch.arange(batch, dtype=torch.int64, device=device) * row_stride_bytes |
| 856 | ) |
| 857 | dst_ptrs = (row_offsets + state.data_ptr()).to(torch.uint64) |
| 858 | src_ptrs = (row_offsets + state.data_ptr() + shift).to(torch.uint64) |
| 859 | sizes = torch.full((batch,), copy_size, dtype=torch.int32, device=device) |
| 860 | |
| 861 | expected = snapshot.clone() |
| 862 | expected[:, :copy_size].copy_(snapshot[:, shift:]) |
| 863 | for _ in range(10): |
| 864 | state.copy_(snapshot) |
| 865 | batch_memcpy(src_ptrs, dst_ptrs, sizes) |
| 866 | torch.accelerator.synchronize() |
| 867 | torch.testing.assert_close(state, expected, rtol=0, atol=0) |
| 868 | |
| 869 | def test_matches_python_postprocess_mamba(self, device, test_config): |
| 870 | """Golden test: GPU kernel produces identical results to Python impl. |
| 871 | |
| 872 | This test: |
| 873 | 1. Sets up identical initial state for both paths |
| 874 | 2. Runs Python postprocess_mamba (modifies states via batch_memcpy) |
| 875 | 3. Runs GPU fused kernel (modifies states directly) |
| 876 | 4. Compares resulting state tensors and num_accepted_tokens |
| 877 | """ |
| 878 | cfg = test_config |
| 879 | torch.manual_seed(42) |
| 880 | |
| 881 | # Test scenario: 4 requests with different copy conditions |
| 882 | # Copy needed when: aligned_new_computed >= num_tokens_running_state |
| 883 | # where: num_tokens_running_state = num_computed + num_scheduled - num_draft |
| 884 | # new_num_computed = num_tokens_running_state + num_accepted - 1 |
| 885 | # aligned_new_computed = (new_num_computed // block_size) * block_size |
| 886 | req_ids = ["req_0", "req_1", "req_2", "req_3"] |
| 887 | |
| 888 | # Configure requests so some need copies, some don't |
| 889 | # block_size = 16 |
| 890 | # req_0: running=60+5-2=63, new=63+3-1=65, aligned=64 >= 63 -> COPY |
| 891 | # req_1: running=30+3-0=33, new=33+2-1=34, aligned=32 < 33 -> NO COPY |
| 892 | # req_2: running=45+8-3=50, new=50+4-1=53, aligned=48 < 50 -> NO COPY |
| 893 | # req_3: running=10+6-0=16, new=16+2-1=17, aligned=16 >= 16 -> COPY |
| 894 | num_computed_tokens = [60, 30, 45, 10] |
| 895 | num_scheduled_tokens = {"req_0": 5, "req_1": 3, "req_2": 8, "req_3": 6} |
| 896 | num_draft_tokens = {"req_0": 2, "req_1": 0, "req_2": 3, "req_3": 0} |
| 897 | num_accepted_tokens = [3, 2, 4, 2] |
| 898 | mamba_state_idx = [3, 1, 2, 0] # source block indices |
| 899 | |
| 900 | # Block IDs for each request (simulate block table) |
| 901 | block_ids_per_req = [ |
| 902 | list(range(8)), # req_0: blocks 0-7 |
| 903 | list(range(8, 16)), # req_1: blocks 8-15 |
| 904 | list(range(16, 24)), # req_2: blocks 16-23 |
| 905 | list(range(24, 32)), # req_3: blocks 24-31 |
| 906 | ] |
| 907 | |
| 908 | layer_names = [f"layer_{i}" for i in range(cfg.num_layers)] |
| 909 | kv_cache_config = _make_kv_cache_config(cfg, layer_names) |
| 910 | |
| 911 | ( |
| 912 | conv_states_py, |
| 913 | temporal_states_py, |
| 914 | conv_states_gpu, |
| 915 | temporal_states_gpu, |
| 916 | forward_context_py, |
| 917 | forward_context_gpu, |
| 918 | ) = _make_dual_states(cfg, layer_names, device) |
| 919 | |
| 920 | # --- Run Python path --- |
| 921 | scheduler_output = _make_postprocess_scheduler_output( |
| 922 | req_ids, |
| 923 | num_scheduled_tokens, |
| 924 | {k: [None] * v for k, v in num_draft_tokens.items() if v > 0}, |
| 925 | ) |
| 926 | input_batch_py = _make_input_batch( |
| 927 | req_ids, num_accepted_tokens.copy(), mamba_state_idx.copy() |
| 928 | ) |
| 929 | requests = _make_requests(req_ids, num_computed_tokens, block_ids_per_req) |
| 930 | copy_bufs = _make_copy_bufs(cfg, kv_cache_config, device) |
| 931 | |
| 932 | postprocess_mamba( |
| 933 | scheduler_output, |
| 934 | kv_cache_config, |
| 935 | input_batch_py, |
| 936 | requests, |
| 937 | forward_context_py, |
| 938 | _COPY_FUNCS, |
| 939 | copy_bufs, |
| 940 | ) |
| 941 | torch.accelerator.synchronize() |
| 942 | |
| 943 | # --- Run GPU path --- |
| 944 | gpu_ctx = _make_gpu_ctx(cfg, kv_cache_config, device) |
| 945 | |
| 946 | num_reqs = len(req_ids) |
| 947 | max_blocks = max(len(b) for b in block_ids_per_req) |
| 948 | block_table_gpu = torch.zeros( |
| 949 | num_reqs, max_blocks, dtype=torch.int32, device=device |
| 950 | ) |
| 951 | for i, block_ids in enumerate(block_ids_per_req): |
| 952 | block_table_gpu[i, : len(block_ids)] = torch.tensor( |
| 953 | block_ids, dtype=torch.int32 |
| 954 | ) |
| 955 | _run_gpu_postprocess( |
| 956 | gpu_ctx, |
| 957 | kv_cache_config=kv_cache_config, |
| 958 | forward_context=forward_context_gpu, |
| 959 | copy_funcs=_COPY_FUNCS, |
| 960 | block_table=block_table_gpu, |
| 961 | req_ids=req_ids, |
| 962 | num_accepted_tokens=num_accepted_tokens, |
| 963 | mamba_state_idx=mamba_state_idx, |
| 964 | num_scheduled_tokens=num_scheduled_tokens, |
| 965 | num_computed_tokens=num_computed_tokens, |
| 966 | num_draft_tokens=num_draft_tokens, |
| 967 | device=device, |
| 968 | ) |
| 969 | |
| 970 | # --- Compare results --- |
| 971 | # 1. Compare state tensors |
| 972 | for i in range(cfg.num_layers): |
| 973 | torch.testing.assert_close( |
| 974 | conv_states_gpu[i], |
| 975 | conv_states_py[i], |
| 976 | msg=f"Conv state mismatch at layer {i}", |
| 977 | ) |
| 978 | torch.testing.assert_close( |
| 979 | temporal_states_gpu[i], |
| 980 | temporal_states_py[i], |
| 981 | msg=f"Temporal state mismatch at layer {i}", |
| 982 | ) |
| 983 | |
| 984 | # 2. Compare num_accepted_tokens updates |
| 985 | expected_accepted = torch.tensor( |
| 986 | input_batch_py.num_accepted_tokens_cpu[:num_reqs], |
| 987 | dtype=torch.int32, |
| 988 | device=device, |
| 989 | ) |
| 990 | torch.testing.assert_close( |
| 991 | gpu_ctx.num_accepted_tokens_out[:num_reqs], |
| 992 | expected_accepted, |
| 993 | msg="num_accepted_tokens mismatch", |
| 994 | ) |
| 995 | |
| 996 | def test_no_copy_when_not_needed(self, device, test_config): |
| 997 | """Kernel should not modify state when no copy is needed.""" |
| 998 | cfg = test_config |
| 999 | torch.manual_seed(123) |
| 1000 | |
| 1001 | # Single request where no copy is needed: |
| 1002 | # running = 30 + 3 = 33, new = 33 + 1 - 1 = 33, aligned = 32 < 33 |
| 1003 | req_ids = ["req_0"] |
| 1004 | num_computed_tokens = [30] |
| 1005 | num_scheduled_tokens = {"req_0": 3} |
| 1006 | num_draft_tokens: dict[str, int] = {} |
| 1007 | num_accepted_tokens = [1] |
| 1008 | mamba_state_idx = [1] |
| 1009 | block_ids_per_req = [list(range(8))] |
| 1010 | |
| 1011 | layer_names = ["layer_0"] |
| 1012 | kv_cache_config = _make_kv_cache_config(cfg, layer_names) |
| 1013 | |
| 1014 | # Create state tensor |
| 1015 | conv_state = torch.randn( |
| 1016 | cfg.num_blocks, |
| 1017 | cfg.conv_width, |
| 1018 | cfg.conv_inner_dim, |
| 1019 | dtype=cfg.dtype, |
| 1020 | device=device, |
| 1021 | ) |
| 1022 | temporal_state = torch.randn( |
| 1023 | cfg.num_blocks, cfg.temporal_state_dim, dtype=cfg.dtype, device=device |
| 1024 | ) |
| 1025 | |
| 1026 | # Clone to verify no modification |
| 1027 | conv_state_orig = conv_state.clone() |
| 1028 | temporal_state_orig = temporal_state.clone() |
| 1029 | |
| 1030 | forward_context = {"layer_0": _make_mock_attention(conv_state, temporal_state)} |
| 1031 | |
| 1032 | gpu_ctx = _make_gpu_ctx(cfg, kv_cache_config, device) |
| 1033 | |
| 1034 | num_reqs = len(req_ids) |
| 1035 | block_table_gpu = torch.zeros(num_reqs, 8, dtype=torch.int32, device=device) |
| 1036 | block_table_gpu[0, :8] = torch.tensor(block_ids_per_req[0], dtype=torch.int32) |
| 1037 | |
| 1038 | _run_gpu_postprocess( |
| 1039 | gpu_ctx, |
| 1040 | kv_cache_config=kv_cache_config, |
| 1041 | forward_context=forward_context, |
| 1042 | copy_funcs=_COPY_FUNCS, |
| 1043 | block_table=block_table_gpu, |
| 1044 | req_ids=req_ids, |
| 1045 | num_accepted_tokens=num_accepted_tokens, |
| 1046 | mamba_state_idx=mamba_state_idx, |
| 1047 | num_scheduled_tokens=num_scheduled_tokens, |
| 1048 | num_computed_tokens=num_computed_tokens, |
| 1049 | num_draft_tokens=num_draft_tokens, |
| 1050 | device=device, |
| 1051 | ) |
| 1052 | |
| 1053 | # State should be unchanged |
| 1054 | torch.testing.assert_close(conv_state, conv_state_orig) |
| 1055 | torch.testing.assert_close(temporal_state, temporal_state_orig) |
| 1056 | |
| 1057 | @pytest.mark.parametrize("num_reqs", [1, 2, 8, 16]) |
| 1058 | def test_various_batch_sizes(self, device, test_config, num_reqs): |
| 1059 | """Verify kernel works correctly with different batch sizes.""" |
| 1060 | cfg = _TestConfig(max_num_reqs=max(16, num_reqs)) |
| 1061 | torch.manual_seed(456) |
| 1062 | |
| 1063 | req_ids = [f"req_{i}" for i in range(num_reqs)] |
| 1064 | # All requests will trigger a copy |
| 1065 | num_computed_tokens = [60] * num_reqs |
| 1066 | num_scheduled_tokens = {r: 5 for r in req_ids} |
| 1067 | num_draft_tokens = {r: 0 for r in req_ids} |
| 1068 | num_accepted_tokens = [3] * num_reqs |
| 1069 | mamba_state_idx = [3] * num_reqs |
| 1070 | # Each request gets unique blocks |
| 1071 | block_ids_per_req = [list(range(i * 8, (i + 1) * 8)) for i in range(num_reqs)] |
| 1072 | |
| 1073 | # Ensure we have enough blocks |
| 1074 | total_blocks = num_reqs * 8 |
| 1075 | cfg = _TestConfig(num_blocks=total_blocks, max_num_reqs=max(16, num_reqs)) |
| 1076 | |
| 1077 | layer_names = ["layer_0"] |
| 1078 | kv_cache_config = _make_kv_cache_config(cfg, layer_names) |
| 1079 | |
| 1080 | ( |
| 1081 | conv_state_py, |
| 1082 | temporal_state_py, |
| 1083 | conv_state_gpu, |
| 1084 | temporal_state_gpu, |
| 1085 | forward_context_py, |
| 1086 | forward_context_gpu, |
| 1087 | ) = _make_dual_layer_state(cfg, device) |
| 1088 | |
| 1089 | # Run Python path |
| 1090 | scheduler_output = _make_postprocess_scheduler_output( |
| 1091 | req_ids, |
| 1092 | num_scheduled_tokens, |
| 1093 | {k: [None] * v for k, v in num_draft_tokens.items() if v > 0}, |
| 1094 | ) |
| 1095 | input_batch_py = _make_input_batch( |
| 1096 | req_ids, num_accepted_tokens.copy(), mamba_state_idx.copy() |
| 1097 | ) |
| 1098 | requests = _make_requests(req_ids, num_computed_tokens, block_ids_per_req) |
| 1099 | copy_bufs = _make_copy_bufs(cfg, kv_cache_config, device) |
| 1100 | |
| 1101 | postprocess_mamba( |
| 1102 | scheduler_output, |
| 1103 | kv_cache_config, |
| 1104 | input_batch_py, |
| 1105 | requests, |
| 1106 | forward_context_py, |
| 1107 | _COPY_FUNCS, |
| 1108 | copy_bufs, |
| 1109 | ) |
| 1110 | torch.accelerator.synchronize() |
| 1111 | |
| 1112 | # Run GPU path |
| 1113 | gpu_ctx = _make_gpu_ctx(cfg, kv_cache_config, device) |
| 1114 | |
| 1115 | max_blocks_per_req = 8 |
| 1116 | block_table_gpu = torch.zeros( |
| 1117 | num_reqs, max_blocks_per_req, dtype=torch.int32, device=device |
| 1118 | ) |
| 1119 | for i, block_ids in enumerate(block_ids_per_req): |
| 1120 | block_table_gpu[i, : len(block_ids)] = torch.tensor( |
| 1121 | block_ids, dtype=torch.int32 |
| 1122 | ) |
| 1123 | |
| 1124 | _run_gpu_postprocess( |
| 1125 | gpu_ctx, |
| 1126 | kv_cache_config=kv_cache_config, |
| 1127 | forward_context=forward_context_gpu, |
| 1128 | copy_funcs=_COPY_FUNCS, |
| 1129 | block_table=block_table_gpu, |
| 1130 | req_ids=req_ids, |
| 1131 | num_accepted_tokens=num_accepted_tokens, |
| 1132 | mamba_state_idx=mamba_state_idx, |
| 1133 | num_scheduled_tokens=num_scheduled_tokens, |
| 1134 | num_computed_tokens=num_computed_tokens, |
| 1135 | num_draft_tokens=num_draft_tokens, |
| 1136 | device=device, |
| 1137 | ) |
| 1138 | |
| 1139 | # Compare results |
| 1140 | torch.testing.assert_close( |
| 1141 | conv_state_gpu, conv_state_py, msg="Conv state mismatch" |
| 1142 | ) |
| 1143 | torch.testing.assert_close( |
| 1144 | temporal_state_gpu, temporal_state_py, msg="Temporal state mismatch" |
| 1145 | ) |
| 1146 | |
| 1147 | def test_block_table_with_realistic_stride(self, device, test_config): |
| 1148 | """Test kernel with realistic block table strides. |
| 1149 | |
| 1150 | In real usage, the block table is pre-allocated with shape |
| 1151 | [max_num_reqs, max_num_blocks_per_req] and then sliced to |
| 1152 | [:num_reqs]. This means stride(0) = max_num_blocks_per_req, |
| 1153 | which is typically much larger than the actual blocks used. |
| 1154 | |
| 1155 | This test verifies the kernel handles non-tight strides correctly, |
| 1156 | catching bugs where stride is incorrectly treated as bytes vs elements. |
| 1157 | """ |
| 1158 | cfg = test_config |
| 1159 | torch.manual_seed(789) |
| 1160 | |
| 1161 | # Use multiple requests to exercise stride-based indexing |
| 1162 | num_reqs = 4 |
| 1163 | req_ids = [f"req_{i}" for i in range(num_reqs)] |
| 1164 | |
| 1165 | # All requests trigger copies (same setup as test_various_batch_sizes) |
| 1166 | num_computed_tokens = [60] * num_reqs |
| 1167 | num_scheduled_tokens = {r: 5 for r in req_ids} |
| 1168 | num_draft_tokens: dict[str, int] = {} |
| 1169 | num_accepted_tokens = [3] * num_reqs |
| 1170 | mamba_state_idx = [3] * num_reqs |
| 1171 | |
| 1172 | # Each request uses only 8 blocks, but we allocate much more |
| 1173 | blocks_used_per_req = 8 |
| 1174 | block_ids_per_req = [ |
| 1175 | list(range(i * blocks_used_per_req, (i + 1) * blocks_used_per_req)) |
| 1176 | for i in range(num_reqs) |
| 1177 | ] |
| 1178 | |
| 1179 | total_blocks = num_reqs * blocks_used_per_req |
| 1180 | cfg = _TestConfig(num_blocks=total_blocks, max_num_reqs=max(16, num_reqs)) |
| 1181 | |
| 1182 | layer_names = ["layer_0"] |
| 1183 | kv_cache_config = _make_kv_cache_config(cfg, layer_names) |
| 1184 | |
| 1185 | ( |
| 1186 | conv_state_py, |
| 1187 | temporal_state_py, |
| 1188 | conv_state_gpu, |
| 1189 | temporal_state_gpu, |
| 1190 | forward_context_py, |
| 1191 | forward_context_gpu, |
| 1192 | ) = _make_dual_layer_state(cfg, device) |
| 1193 | |
| 1194 | # Run Python path |
| 1195 | scheduler_output = _make_postprocess_scheduler_output( |
| 1196 | req_ids, |
| 1197 | num_scheduled_tokens, |
| 1198 | {k: [None] * v for k, v in num_draft_tokens.items() if v > 0}, |
| 1199 | ) |
| 1200 | input_batch_py = _make_input_batch( |
| 1201 | req_ids, num_accepted_tokens.copy(), mamba_state_idx.copy() |
| 1202 | ) |
| 1203 | requests = _make_requests(req_ids, num_computed_tokens, block_ids_per_req) |
| 1204 | copy_bufs = _make_copy_bufs(cfg, kv_cache_config, device) |
| 1205 | |
| 1206 | postprocess_mamba( |
| 1207 | scheduler_output, |
| 1208 | kv_cache_config, |
| 1209 | input_batch_py, |
| 1210 | requests, |
| 1211 | forward_context_py, |
| 1212 | _COPY_FUNCS, |
| 1213 | copy_bufs, |
| 1214 | ) |
| 1215 | torch.accelerator.synchronize() |
| 1216 | |
| 1217 | # Run GPU path with REALISTIC block table stride |
| 1218 | gpu_ctx = _make_gpu_ctx(cfg, kv_cache_config, device) |
| 1219 | |
| 1220 | # KEY DIFFERENCE: Create a large block table like real code does |
| 1221 | # Real system has max_num_blocks_per_req >> blocks actually used |
| 1222 | max_num_reqs_full = 16 |
| 1223 | max_blocks_per_req_full = 512 # Much larger than blocks_used_per_req=8 |
| 1224 | |
| 1225 | # Allocate full-size table (simulates pre-allocated CpuGpuBuffer) |
| 1226 | block_table_full = torch.zeros( |
| 1227 | max_num_reqs_full, max_blocks_per_req_full, dtype=torch.int32, device=device |
| 1228 | ) |
| 1229 | |
| 1230 | # Fill in actual block IDs (only first few columns used) |
| 1231 | for i, block_ids in enumerate(block_ids_per_req): |
| 1232 | block_table_full[i, : len(block_ids)] = torch.tensor( |
| 1233 | block_ids, dtype=torch.int32 |
| 1234 | ) |
| 1235 | |
| 1236 | # Slice like real code: block_table.gpu[:num_reqs] |
| 1237 | # This preserves stride(0) = 512, not 8! |
| 1238 | block_table_gpu = block_table_full[:num_reqs] |
| 1239 | |
| 1240 | # Verify stride is large (the key property we're testing) |
| 1241 | assert block_table_gpu.stride(0) == max_blocks_per_req_full, ( |
| 1242 | f"Expected stride {max_blocks_per_req_full}, " |
| 1243 | f"got {block_table_gpu.stride(0)}" |
| 1244 | ) |
| 1245 | |
| 1246 | _run_gpu_postprocess( |
| 1247 | gpu_ctx, |
| 1248 | kv_cache_config=kv_cache_config, |
| 1249 | forward_context=forward_context_gpu, |
| 1250 | copy_funcs=_COPY_FUNCS, |
| 1251 | block_table=block_table_gpu, |
| 1252 | req_ids=req_ids, |
| 1253 | num_accepted_tokens=num_accepted_tokens, |
| 1254 | mamba_state_idx=mamba_state_idx, |
| 1255 | num_scheduled_tokens=num_scheduled_tokens, |
| 1256 | num_computed_tokens=num_computed_tokens, |
| 1257 | num_draft_tokens=num_draft_tokens, |
| 1258 | device=device, |
| 1259 | ) |
| 1260 | |
| 1261 | # Compare results - this will fail if stride handling is incorrect |
| 1262 | torch.testing.assert_close( |
| 1263 | conv_state_gpu, |
| 1264 | conv_state_py, |
| 1265 | msg="Conv state mismatch - possible stride bug in kernel", |
| 1266 | ) |
| 1267 | torch.testing.assert_close( |
| 1268 | temporal_state_gpu, |
| 1269 | temporal_state_py, |
| 1270 | msg="Temporal state mismatch - possible stride bug in kernel", |
| 1271 | ) |
| 1272 | |
| 1273 | def test_src_addr_equals_dst_addr_skips_copy_and_sets_accepted_to_1( |
| 1274 | self, device, test_config |
| 1275 | ): |
| 1276 | """Test the ``src_addr == dst_addr`` early-return path in |
| 1277 | postprocess_mamba_fused_kernel matches Python behavior. |
| 1278 | |
| 1279 | When src_addr == dst_addr (source and destination memory addresses are |
| 1280 | identical), both implementations should: |
| 1281 | 1. Skip the copy (state unchanged) |
| 1282 | 2. Set num_accepted_tokens to 1 |
| 1283 | |
| 1284 | This condition occurs when: |
| 1285 | - src_block_idx == dest_block_idx (same logical block) |
| 1286 | - accept_token_bias == 0 (no offset within the block) |
| 1287 | |
| 1288 | Python reference (collect_mamba_copy_meta): |
| 1289 | if src_block_idx == dest_block_idx and accept_token_bias == 0: |
| 1290 | return # No copy added |
| 1291 | |
| 1292 | Python reference (postprocess_mamba): |
| 1293 | if src_block_idx == dest_block_idx: |
| 1294 | num_accepted_tokens_cpu[i] = 1 |
| 1295 | |
| 1296 | Test setup (block_size=16): |
| 1297 | - num_tokens_running_state = 30 + 2 - 0 = 32 |
| 1298 | - new_num_computed = 32 + 1 - 1 = 32 |
| 1299 | - aligned_new_computed = 32 |
| 1300 | - accept_token_bias = 32 - 32 = 0 |
| 1301 | - dest_block_idx = 32 // 16 - 1 = 1 |
| 1302 | - src_block_idx = 1 (set explicitly) |
| 1303 | """ |
| 1304 | cfg = test_config |
| 1305 | torch.manual_seed(1001) |
| 1306 | |
| 1307 | req_ids = ["req_0"] |
| 1308 | num_computed_tokens = [30] |
| 1309 | num_scheduled_tokens = {"req_0": 2} |
| 1310 | num_draft_tokens: dict[str, int] = {} |
| 1311 | num_accepted_tokens = [5] # Initial value, should become 1 |
| 1312 | mamba_state_idx = [1] # src_block_idx = 1 = dest_block_idx |
| 1313 | block_ids_per_req = [list(range(8))] |
| 1314 | |
| 1315 | layer_names = ["layer_0"] |
| 1316 | kv_cache_config = _make_kv_cache_config(cfg, layer_names) |
| 1317 | |
| 1318 | ( |
| 1319 | conv_state_py, |
| 1320 | temporal_state_py, |
| 1321 | conv_state_gpu, |
| 1322 | temporal_state_gpu, |
| 1323 | forward_context_py, |
| 1324 | forward_context_gpu, |
| 1325 | ) = _make_dual_layer_state(cfg, device) |
| 1326 | |
| 1327 | # Also clone to verify no modification |
| 1328 | conv_state_orig = conv_state_py.clone() |
| 1329 | temporal_state_orig = temporal_state_py.clone() |
| 1330 | |
| 1331 | # --- Run Python path --- |
| 1332 | scheduler_output = _make_postprocess_scheduler_output( |
| 1333 | req_ids, |
| 1334 | num_scheduled_tokens, |
| 1335 | {k: [None] * v for k, v in num_draft_tokens.items() if v > 0}, |
| 1336 | ) |
| 1337 | input_batch_py = _make_input_batch( |
| 1338 | req_ids, num_accepted_tokens.copy(), mamba_state_idx.copy() |
| 1339 | ) |
| 1340 | requests = _make_requests(req_ids, num_computed_tokens, block_ids_per_req) |
| 1341 | copy_bufs = _make_copy_bufs(cfg, kv_cache_config, device) |
| 1342 | |
| 1343 | postprocess_mamba( |
| 1344 | scheduler_output, |
| 1345 | kv_cache_config, |
| 1346 | input_batch_py, |
| 1347 | requests, |
| 1348 | forward_context_py, |
| 1349 | _COPY_FUNCS, |
| 1350 | copy_bufs, |
| 1351 | ) |
| 1352 | torch.accelerator.synchronize() |
| 1353 | |
| 1354 | # --- Run GPU path --- |
| 1355 | gpu_ctx = _make_gpu_ctx(cfg, kv_cache_config, device) |
| 1356 | |
| 1357 | num_reqs = len(req_ids) |
| 1358 | block_table_gpu = torch.zeros(num_reqs, 8, dtype=torch.int32, device=device) |
| 1359 | block_table_gpu[0, :8] = torch.tensor(block_ids_per_req[0], dtype=torch.int32) |
| 1360 | |
| 1361 | _run_gpu_postprocess( |
| 1362 | gpu_ctx, |
| 1363 | kv_cache_config=kv_cache_config, |
| 1364 | forward_context=forward_context_gpu, |
| 1365 | copy_funcs=_COPY_FUNCS, |
| 1366 | block_table=block_table_gpu, |
| 1367 | req_ids=req_ids, |
| 1368 | num_accepted_tokens=num_accepted_tokens, |
| 1369 | mamba_state_idx=mamba_state_idx, |
| 1370 | num_scheduled_tokens=num_scheduled_tokens, |
| 1371 | num_computed_tokens=num_computed_tokens, |
| 1372 | num_draft_tokens=num_draft_tokens, |
| 1373 | device=device, |
| 1374 | ) |
| 1375 | |
| 1376 | # --- Verify Python behavior (ground truth) --- |
| 1377 | # State should be unchanged (no copy when src_addr == dst_addr) |
| 1378 | torch.testing.assert_close( |
| 1379 | conv_state_py, |
| 1380 | conv_state_orig, |
| 1381 | msg="Python: Conv state should be unchanged when src==dst", |
| 1382 | ) |
| 1383 | torch.testing.assert_close( |
| 1384 | temporal_state_py, |
| 1385 | temporal_state_orig, |
| 1386 | msg="Python: Temporal state should be unchanged when src==dst", |
| 1387 | ) |
| 1388 | # num_accepted_tokens should be 1 |
| 1389 | assert input_batch_py.num_accepted_tokens_cpu[0] == 1, ( |
| 1390 | f"Python: num_accepted_tokens should be 1, " |
| 1391 | f"got {input_batch_py.num_accepted_tokens_cpu[0]}" |
| 1392 | ) |
| 1393 | |
| 1394 | # --- Verify GPU matches Python --- |
| 1395 | torch.testing.assert_close( |
| 1396 | conv_state_gpu, |
| 1397 | conv_state_py, |
| 1398 | msg="GPU conv state should match Python", |
| 1399 | ) |
| 1400 | torch.testing.assert_close( |
| 1401 | temporal_state_gpu, |
| 1402 | temporal_state_py, |
| 1403 | msg="GPU temporal state should match Python", |
| 1404 | ) |
| 1405 | expected_accepted = torch.tensor( |
| 1406 | input_batch_py.num_accepted_tokens_cpu[:num_reqs], |
| 1407 | dtype=torch.int32, |
| 1408 | device=device, |
| 1409 | ) |
| 1410 | torch.testing.assert_close( |
| 1411 | gpu_ctx.num_accepted_tokens_out[:num_reqs], |
| 1412 | expected_accepted, |
| 1413 | msg="GPU num_accepted_tokens should match Python", |
| 1414 | ) |
| 1415 | |
| 1416 | def test_same_block_idx_with_offset_copies_then_sets_accepted_to_1( |
| 1417 | self, device, test_config |
| 1418 | ): |
| 1419 | """Test the ``src_block_idx == dest_block_idx`` post-copy update in |
| 1420 | postprocess_mamba_fused_kernel matches Python behavior. |
| 1421 | |
| 1422 | When src_block_idx == dest_block_idx but accept_token_bias > 0, both |
| 1423 | implementations should: |
| 1424 | 1. Perform the copy (src_addr != dst_addr due to offset) |
| 1425 | 2. Set num_accepted_tokens to 1 AFTER the copy |
| 1426 | |
| 1427 | Python reference (postprocess_mamba): |
| 1428 | if src_block_idx == dest_block_idx: |
| 1429 | num_accepted_tokens_cpu[i] = 1 |
| 1430 | |
| 1431 | For conv states: copies state[block, offset:] to |
| 1432 | state[block, :] (shifted window) |
| 1433 | For temporal states: copies state[block_ids[src_idx + offset]] to |
| 1434 | state[block_ids[dest_idx]] |
| 1435 | |
| 1436 | Test setup (block_size=16): |
| 1437 | - num_tokens_running_state = 30 + 1 - 0 = 31 |
| 1438 | - new_num_computed = 31 + 2 - 1 = 32 |
| 1439 | - aligned_new_computed = 32 |
| 1440 | - accept_token_bias = 32 - 31 = 1 (> 0, so copy happens) |
| 1441 | - dest_block_idx = 32 // 16 - 1 = 1 |
| 1442 | - src_block_idx = 1 (set explicitly, == dest_block_idx) |
| 1443 | """ |
| 1444 | cfg = test_config |
| 1445 | torch.manual_seed(1002) |
| 1446 | |
| 1447 | req_ids = ["req_0"] |
| 1448 | num_computed_tokens = [30] |
| 1449 | num_scheduled_tokens = {"req_0": 1} |
| 1450 | num_draft_tokens: dict[str, int] = {} |
| 1451 | num_accepted_tokens = [2] # Results in accept_token_bias = 1 |
| 1452 | mamba_state_idx = [1] # src_block_idx = 1 = dest_block_idx |
| 1453 | block_ids_per_req = [list(range(8))] |
| 1454 | |
| 1455 | layer_names = ["layer_0"] |
| 1456 | kv_cache_config = _make_kv_cache_config(cfg, layer_names) |
| 1457 | |
| 1458 | ( |
| 1459 | conv_state_py, |
| 1460 | temporal_state_py, |
| 1461 | conv_state_gpu, |
| 1462 | temporal_state_gpu, |
| 1463 | forward_context_py, |
| 1464 | forward_context_gpu, |
| 1465 | ) = _make_dual_layer_state(cfg, device) |
| 1466 | |
| 1467 | # Clone to verify modification |
| 1468 | conv_state_orig = conv_state_py.clone() |
| 1469 | temporal_state_orig = temporal_state_py.clone() |
| 1470 | |
| 1471 | # --- Run Python path --- |
| 1472 | scheduler_output = _make_postprocess_scheduler_output( |
| 1473 | req_ids, |
| 1474 | num_scheduled_tokens, |
| 1475 | {k: [None] * v for k, v in num_draft_tokens.items() if v > 0}, |
| 1476 | ) |
| 1477 | input_batch_py = _make_input_batch( |
| 1478 | req_ids, num_accepted_tokens.copy(), mamba_state_idx.copy() |
| 1479 | ) |
| 1480 | requests = _make_requests(req_ids, num_computed_tokens, block_ids_per_req) |
| 1481 | copy_bufs = _make_copy_bufs(cfg, kv_cache_config, device) |
| 1482 | |
| 1483 | postprocess_mamba( |
| 1484 | scheduler_output, |
| 1485 | kv_cache_config, |
| 1486 | input_batch_py, |
| 1487 | requests, |
| 1488 | forward_context_py, |
| 1489 | _COPY_FUNCS, |
| 1490 | copy_bufs, |
| 1491 | ) |
| 1492 | torch.accelerator.synchronize() |
| 1493 | |
| 1494 | # --- Run GPU path --- |
| 1495 | gpu_ctx = _make_gpu_ctx(cfg, kv_cache_config, device) |
| 1496 | |
| 1497 | num_reqs = len(req_ids) |
| 1498 | block_table_gpu = torch.zeros(num_reqs, 8, dtype=torch.int32, device=device) |
| 1499 | block_table_gpu[0, :8] = torch.tensor(block_ids_per_req[0], dtype=torch.int32) |
| 1500 | |
| 1501 | _run_gpu_postprocess( |
| 1502 | gpu_ctx, |
| 1503 | kv_cache_config=kv_cache_config, |
| 1504 | forward_context=forward_context_gpu, |
| 1505 | copy_funcs=_COPY_FUNCS, |
| 1506 | block_table=block_table_gpu, |
| 1507 | req_ids=req_ids, |
| 1508 | num_accepted_tokens=num_accepted_tokens, |
| 1509 | mamba_state_idx=mamba_state_idx, |
| 1510 | num_scheduled_tokens=num_scheduled_tokens, |
| 1511 | num_computed_tokens=num_computed_tokens, |
| 1512 | num_draft_tokens=num_draft_tokens, |
| 1513 | device=device, |
| 1514 | ) |
| 1515 | |
| 1516 | # --- Verify Python behavior (ground truth) --- |
| 1517 | dest_block_id = block_ids_per_req[0][1] # dest_block_idx = 1 |
| 1518 | |
| 1519 | # This is an overlapping in-place left shift, so comparing only the |
| 1520 | # Python and fused paths can hide the same memcpy race in both. Build |
| 1521 | # the memmove result from the untouched snapshot and check each path |
| 1522 | # independently. |
| 1523 | expected_conv_state = conv_state_orig.clone() |
| 1524 | expected_conv_state[dest_block_id, :-1].copy_( |
| 1525 | conv_state_orig[dest_block_id, 1:] |
| 1526 | ) |
| 1527 | torch.testing.assert_close( |
| 1528 | conv_state_py, |
| 1529 | expected_conv_state, |
| 1530 | rtol=0, |
| 1531 | atol=0, |
| 1532 | msg="Python: overlapping conv copy should have memmove semantics", |
| 1533 | ) |
| 1534 | torch.testing.assert_close( |
| 1535 | conv_state_gpu, |
| 1536 | expected_conv_state, |
| 1537 | rtol=0, |
| 1538 | atol=0, |
| 1539 | msg="GPU: overlapping conv copy should have memmove semantics", |
| 1540 | ) |
| 1541 | |
| 1542 | # Temporal state should be modified (copy from different block) |
| 1543 | src_block_id_temporal = block_ids_per_req[0][2] # actual_src_block_idx = 2 |
| 1544 | dest_block_id_temporal = block_ids_per_req[0][1] # dest_block_idx = 1 |
| 1545 | torch.testing.assert_close( |
| 1546 | temporal_state_py[dest_block_id_temporal], |
| 1547 | temporal_state_orig[src_block_id_temporal], |
| 1548 | msg="Python: Temporal state copy should have happened", |
| 1549 | ) |
| 1550 | |
| 1551 | # num_accepted_tokens should be 1 |
| 1552 | assert input_batch_py.num_accepted_tokens_cpu[0] == 1, ( |
| 1553 | f"Python: num_accepted_tokens should be 1, " |
| 1554 | f"got {input_batch_py.num_accepted_tokens_cpu[0]}" |
| 1555 | ) |
| 1556 | |
| 1557 | # --- Verify GPU matches Python --- |
| 1558 | torch.testing.assert_close( |
| 1559 | conv_state_gpu, |
| 1560 | conv_state_py, |
| 1561 | msg="GPU conv state should match Python", |
| 1562 | ) |
| 1563 | torch.testing.assert_close( |
| 1564 | temporal_state_gpu, |
| 1565 | temporal_state_py, |
| 1566 | msg="GPU temporal state should match Python", |
| 1567 | ) |
| 1568 | expected_accepted = torch.tensor( |
| 1569 | input_batch_py.num_accepted_tokens_cpu[:num_reqs], |
| 1570 | dtype=torch.int32, |
| 1571 | device=device, |
| 1572 | ) |
| 1573 | torch.testing.assert_close( |
| 1574 | gpu_ctx.num_accepted_tokens_out[:num_reqs], |
| 1575 | expected_accepted, |
| 1576 | msg="GPU num_accepted_tokens should match Python", |
| 1577 | ) |
| 1578 | |
| 1579 | def test_different_block_idx_copies_without_setting_accepted_to_1( |
| 1580 | self, device, test_config |
| 1581 | ): |
| 1582 | """Test that neither special-case path triggers when |
| 1583 | src_block_idx != dest_block_idx, and GPU matches Python behavior. |
| 1584 | |
| 1585 | When copying between different blocks: |
| 1586 | 1. src_addr != dst_addr (different blocks = different addresses) |
| 1587 | 2. src_block_idx != dest_block_idx |
| 1588 | |
| 1589 | Therefore: |
| 1590 | - The ``src_addr == dst_addr`` early-return does NOT trigger |
| 1591 | - The ``src_block_idx == dest_block_idx`` post-copy update does NOT trigger |
| 1592 | - Copy happens normally |
| 1593 | - num_accepted_tokens remains UNCHANGED |
| 1594 | |
| 1595 | Test setup (block_size=16): |
| 1596 | - num_tokens_running_state = 60 + 3 - 0 = 63 |
| 1597 | - new_num_computed = 63 + 3 - 1 = 65 |
| 1598 | - aligned_new_computed = 64 |
| 1599 | - accept_token_bias = 64 - 63 = 1 |
| 1600 | - dest_block_idx = 64 // 16 - 1 = 3 |
| 1601 | - src_block_idx = 2 (set explicitly, != dest_block_idx) |
| 1602 | """ |
| 1603 | cfg = test_config |
| 1604 | torch.manual_seed(1003) |
| 1605 | |
| 1606 | req_ids = ["req_0"] |
| 1607 | num_computed_tokens = [60] |
| 1608 | num_scheduled_tokens = {"req_0": 3} |
| 1609 | num_draft_tokens: dict[str, int] = {} |
| 1610 | num_accepted_tokens = [3] # Should remain 3, NOT set to 1 |
| 1611 | mamba_state_idx = [2] # src_block_idx = 2, dest_block_idx will be 3 |
| 1612 | block_ids_per_req = [list(range(8))] |
| 1613 | |
| 1614 | layer_names = ["layer_0"] |
| 1615 | kv_cache_config = _make_kv_cache_config(cfg, layer_names) |
| 1616 | |
| 1617 | ( |
| 1618 | conv_state_py, |
| 1619 | temporal_state_py, |
| 1620 | conv_state_gpu, |
| 1621 | temporal_state_gpu, |
| 1622 | forward_context_py, |
| 1623 | forward_context_gpu, |
| 1624 | ) = _make_dual_layer_state(cfg, device) |
| 1625 | |
| 1626 | # Clone to verify modification |
| 1627 | conv_state_orig = conv_state_py.clone() |
| 1628 | |
| 1629 | # --- Run Python path --- |
| 1630 | scheduler_output = _make_postprocess_scheduler_output( |
| 1631 | req_ids, |
| 1632 | num_scheduled_tokens, |
| 1633 | {k: [None] * v for k, v in num_draft_tokens.items() if v > 0}, |
| 1634 | ) |
| 1635 | input_batch_py = _make_input_batch( |
| 1636 | req_ids, num_accepted_tokens.copy(), mamba_state_idx.copy() |
| 1637 | ) |
| 1638 | requests = _make_requests(req_ids, num_computed_tokens, block_ids_per_req) |
| 1639 | copy_bufs = _make_copy_bufs(cfg, kv_cache_config, device) |
| 1640 | |
| 1641 | postprocess_mamba( |
| 1642 | scheduler_output, |
| 1643 | kv_cache_config, |
| 1644 | input_batch_py, |
| 1645 | requests, |
| 1646 | forward_context_py, |
| 1647 | _COPY_FUNCS, |
| 1648 | copy_bufs, |
| 1649 | ) |
| 1650 | torch.accelerator.synchronize() |
| 1651 | |
| 1652 | # --- Run GPU path --- |
| 1653 | gpu_ctx = _make_gpu_ctx(cfg, kv_cache_config, device) |
| 1654 | |
| 1655 | num_reqs = len(req_ids) |
| 1656 | block_table_gpu = torch.zeros(num_reqs, 8, dtype=torch.int32, device=device) |
| 1657 | block_table_gpu[0, :8] = torch.tensor(block_ids_per_req[0], dtype=torch.int32) |
| 1658 | |
| 1659 | _run_gpu_postprocess( |
| 1660 | gpu_ctx, |
| 1661 | kv_cache_config=kv_cache_config, |
| 1662 | forward_context=forward_context_gpu, |
| 1663 | copy_funcs=_COPY_FUNCS, |
| 1664 | block_table=block_table_gpu, |
| 1665 | req_ids=req_ids, |
| 1666 | num_accepted_tokens=num_accepted_tokens, |
| 1667 | mamba_state_idx=mamba_state_idx, |
| 1668 | num_scheduled_tokens=num_scheduled_tokens, |
| 1669 | num_computed_tokens=num_computed_tokens, |
| 1670 | num_draft_tokens=num_draft_tokens, |
| 1671 | device=device, |
| 1672 | ) |
| 1673 | |
| 1674 | # --- Verify Python behavior (ground truth) --- |
| 1675 | dest_block_id = block_ids_per_req[0][3] # dest_block_idx = 3 |
| 1676 | |
| 1677 | # Copy DID happen (dest block should be modified) |
| 1678 | conv_changed = not torch.allclose( |
| 1679 | conv_state_py[dest_block_id], conv_state_orig[dest_block_id] |
| 1680 | ) |
| 1681 | assert conv_changed, "Python: Conv state copy should have happened" |
| 1682 | |
| 1683 | # num_accepted_tokens should NOT be changed to 1 |
| 1684 | assert input_batch_py.num_accepted_tokens_cpu[0] == num_accepted_tokens[0], ( |
| 1685 | f"Python: num_accepted_tokens should remain {num_accepted_tokens[0]}, " |
| 1686 | f"got {input_batch_py.num_accepted_tokens_cpu[0]}" |
| 1687 | ) |
| 1688 | |
| 1689 | # --- Verify GPU matches Python --- |
| 1690 | torch.testing.assert_close( |
| 1691 | conv_state_gpu, |
| 1692 | conv_state_py, |
| 1693 | msg="GPU conv state should match Python", |
| 1694 | ) |
| 1695 | torch.testing.assert_close( |
| 1696 | temporal_state_gpu, |
| 1697 | temporal_state_py, |
| 1698 | msg="GPU temporal state should match Python", |
| 1699 | ) |
| 1700 | expected_accepted = torch.tensor( |
| 1701 | input_batch_py.num_accepted_tokens_cpu[:num_reqs], |
| 1702 | dtype=torch.int32, |
| 1703 | device=device, |
| 1704 | ) |
| 1705 | torch.testing.assert_close( |
| 1706 | gpu_ctx.num_accepted_tokens_out[:num_reqs], |
| 1707 | expected_accepted, |
| 1708 | msg="GPU num_accepted_tokens should match Python", |
| 1709 | ) |
| 1710 | |
| 1711 | def test_prefix_caching_shared_block_does_not_set_accepted_to_1( |
| 1712 | self, device, test_config |
| 1713 | ): |
| 1714 | """Regression test: with prefix caching, different logical block indices |
| 1715 | can map to the same physical block. The kernel must NOT set |
| 1716 | num_accepted_tokens to 1 in that case. |
| 1717 | |
| 1718 | When src_block_idx != dest_block_idx but block_table maps both to the |
| 1719 | same physical block ID, src_addr == dst_addr. The copy is correctly |
| 1720 | skipped (self-copy is a no-op), but num_accepted_tokens must be |
| 1721 | preserved — only logical-index equality justifies setting it to 1. |
| 1722 | |
| 1723 | Test setup (block_size=16): |
| 1724 | - num_tokens_running_state = 30 + 2 - 0 = 32 |
| 1725 | - new_num_computed = 32 + 3 - 1 = 34 |
| 1726 | - aligned_new_computed = 32 |
| 1727 | - accept_token_bias = 32 - 32 = 0 |
| 1728 | - dest_block_idx = 32 // 16 - 1 = 1 |
| 1729 | - src_block_idx = 0 (set explicitly, != dest_block_idx) |
| 1730 | - block_ids = [5, 5, ...] — prefix caching: both logical indices |
| 1731 | map to the same physical block 5 |
| 1732 | """ |
| 1733 | cfg = test_config |
| 1734 | torch.manual_seed(2001) |
| 1735 | |
| 1736 | req_ids = ["req_0"] |
| 1737 | num_computed_tokens = [30] |
| 1738 | num_scheduled_tokens = {"req_0": 2} |
| 1739 | num_draft_tokens: dict[str, int] = {} |
| 1740 | num_accepted_tokens = [3] # Must stay 3, NOT become 1 |
| 1741 | mamba_state_idx = [0] # src_block_idx = 0, dest_block_idx will be 1 |
| 1742 | |
| 1743 | # Prefix caching: logical blocks 0 and 1 share physical block 5 |
| 1744 | block_ids_per_req = [[5, 5, 2, 3, 4, 6, 7, 8]] |
| 1745 | |
| 1746 | layer_names = ["layer_0"] |
| 1747 | kv_cache_config = _make_kv_cache_config(cfg, layer_names) |
| 1748 | |
| 1749 | ( |
| 1750 | conv_state_py, |
| 1751 | temporal_state_py, |
| 1752 | conv_state_gpu, |
| 1753 | temporal_state_gpu, |
| 1754 | forward_context_py, |
| 1755 | forward_context_gpu, |
| 1756 | ) = _make_dual_layer_state(cfg, device) |
| 1757 | |
| 1758 | conv_state_orig = conv_state_py.clone() |
| 1759 | temporal_state_orig = temporal_state_py.clone() |
| 1760 | |
| 1761 | # --- Run Python path --- |
| 1762 | scheduler_output = _make_postprocess_scheduler_output( |
| 1763 | req_ids, |
| 1764 | num_scheduled_tokens, |
| 1765 | {k: [None] * v for k, v in num_draft_tokens.items() if v > 0}, |
| 1766 | ) |
| 1767 | input_batch_py = _make_input_batch( |
| 1768 | req_ids, num_accepted_tokens.copy(), mamba_state_idx.copy() |
| 1769 | ) |
| 1770 | requests = _make_requests(req_ids, num_computed_tokens, block_ids_per_req) |
| 1771 | copy_bufs = _make_copy_bufs(cfg, kv_cache_config, device) |
| 1772 | |
| 1773 | postprocess_mamba( |
| 1774 | scheduler_output, |
| 1775 | kv_cache_config, |
| 1776 | input_batch_py, |
| 1777 | requests, |
| 1778 | forward_context_py, |
| 1779 | _COPY_FUNCS, |
| 1780 | copy_bufs, |
| 1781 | ) |
| 1782 | torch.accelerator.synchronize() |
| 1783 | |
| 1784 | # --- Run GPU path --- |
| 1785 | gpu_ctx = _make_gpu_ctx(cfg, kv_cache_config, device) |
| 1786 | num_reqs = len(req_ids) |
| 1787 | block_table_gpu = torch.zeros(num_reqs, 8, dtype=torch.int32, device=device) |
| 1788 | block_table_gpu[0, :8] = torch.tensor(block_ids_per_req[0], dtype=torch.int32) |
| 1789 | |
| 1790 | _run_gpu_postprocess( |
| 1791 | gpu_ctx, |
| 1792 | kv_cache_config=kv_cache_config, |
| 1793 | forward_context=forward_context_gpu, |
| 1794 | copy_funcs=_COPY_FUNCS, |
| 1795 | block_table=block_table_gpu, |
| 1796 | req_ids=req_ids, |
| 1797 | num_accepted_tokens=num_accepted_tokens, |
| 1798 | mamba_state_idx=mamba_state_idx, |
| 1799 | num_scheduled_tokens=num_scheduled_tokens, |
| 1800 | num_computed_tokens=num_computed_tokens, |
| 1801 | num_draft_tokens=num_draft_tokens, |
| 1802 | device=device, |
| 1803 | ) |
| 1804 | |
| 1805 | # --- Verify Python behavior (ground truth) --- |
| 1806 | # Copy is self-to-self (same physical block), state unchanged |
| 1807 | torch.testing.assert_close( |
| 1808 | conv_state_py, |
| 1809 | conv_state_orig, |
| 1810 | msg="Python: Conv state should be unchanged (self-copy)", |
| 1811 | ) |
| 1812 | torch.testing.assert_close( |
| 1813 | temporal_state_py, |
| 1814 | temporal_state_orig, |
| 1815 | msg="Python: Temporal state should be unchanged (self-copy)", |
| 1816 | ) |
| 1817 | # num_accepted_tokens must NOT be set to 1 (src_block_idx != dest_block_idx) |
| 1818 | assert input_batch_py.num_accepted_tokens_cpu[0] == num_accepted_tokens[0], ( |
| 1819 | f"Python: num_accepted_tokens should remain {num_accepted_tokens[0]}, " |
| 1820 | f"got {input_batch_py.num_accepted_tokens_cpu[0]}" |
| 1821 | ) |
| 1822 | |
| 1823 | # --- Verify GPU matches Python --- |
| 1824 | torch.testing.assert_close( |
| 1825 | conv_state_gpu, |
| 1826 | conv_state_py, |
| 1827 | msg="GPU conv state should match Python", |
| 1828 | ) |
| 1829 | torch.testing.assert_close( |
| 1830 | temporal_state_gpu, |
| 1831 | temporal_state_py, |
| 1832 | msg="GPU temporal state should match Python", |
| 1833 | ) |
| 1834 | expected_accepted = torch.tensor( |
| 1835 | input_batch_py.num_accepted_tokens_cpu[:num_reqs], |
| 1836 | dtype=torch.int32, |
| 1837 | device=device, |
| 1838 | ) |
| 1839 | torch.testing.assert_close( |
| 1840 | gpu_ctx.num_accepted_tokens_out[:num_reqs], |
| 1841 | expected_accepted, |
| 1842 | msg="GPU num_accepted_tokens should match Python (must NOT be 1)", |
| 1843 | ) |
| 1844 | |
| 1845 | def test_prefix_caching_nonsequential_block_ids_boundary(self, device, test_config): |
| 1846 | """Regression test: non-sequential physical block IDs under prefix caching |
| 1847 | with the needs_copy boundary at exact equality. |
| 1848 | |
| 1849 | Under PC, the block allocator assigns physical block IDs in arbitrary |
| 1850 | order (e.g., [17, 3, 42, 9] instead of [0, 1, 2, 3]). The needs_copy |
| 1851 | condition is purely token-count based and must evaluate identically |
| 1852 | regardless of the physical block IDs assigned. This test verifies that |
| 1853 | the kernel's address arithmetic (block_table lookup, stride computation) |
| 1854 | produces correct copies when physical IDs are non-sequential. |
| 1855 | |
| 1856 | Two requests exercise different boundary behaviors: |
| 1857 | - req_0: aligned_new_computed == num_tokens_running_state (exact boundary) |
| 1858 | This is the tightest edge: one fewer accepted token and no copy needed. |
| 1859 | - req_1: aligned_new_computed == num_tokens_running_state (exact boundary) |
| 1860 | Different block layout, src!=dest, real copy happens. |
| 1861 | |
| 1862 | Both use non-sequential block IDs typical of PC reuse patterns. |
| 1863 | |
| 1864 | Test setup (block_size=16): |
| 1865 | req_0: |
| 1866 | - num_tokens_running_state = 48 + 0 - 0 = 48 |
| 1867 | - new_num_computed = 48 + 1 - 1 = 48 |
| 1868 | - aligned_new_computed = 48 |
| 1869 | - needs_copy = (48 >= 48) = True (exact boundary!) |
| 1870 | - accept_token_bias = 48 - 48 = 0 |
| 1871 | - dest_block_idx = 48 // 16 - 1 = 2 |
| 1872 | - src_block_idx = 2 (same as dest -> num_accepted = 1) |
| 1873 | |
| 1874 | req_1: |
| 1875 | - num_tokens_running_state = 31 + 1 - 0 = 32 |
| 1876 | - new_num_computed = 32 + 3 - 1 = 34 |
| 1877 | - aligned_new_computed = 32 |
| 1878 | - needs_copy = (32 >= 32) = True (exact boundary!) |
| 1879 | - accept_token_bias = 32 - 32 = 0 |
| 1880 | - dest_block_idx = 32 // 16 - 1 = 1 |
| 1881 | - src_block_idx = 0 (diff from dest -> num_accepted unchanged) |
| 1882 | """ |
| 1883 | cfg = test_config |
| 1884 | torch.manual_seed(4001) |
| 1885 | |
| 1886 | req_ids = ["req_0", "req_1"] |
| 1887 | num_computed_tokens = [48, 31] |
| 1888 | num_scheduled_tokens = {"req_0": 0, "req_1": 1} |
| 1889 | num_draft_tokens: dict[str, int] = {} |
| 1890 | num_accepted_tokens = [1, 3] |
| 1891 | mamba_state_idx = [2, 0] |
| 1892 | |
| 1893 | # Non-sequential block IDs typical of prefix caching allocation |
| 1894 | block_ids_per_req = [ |
| 1895 | [17, 3, 42, 9, 25, 11, 30, 2], # req_0: scattered physical blocks |
| 1896 | [41, 7, 22, 15, 38, 19, 4, 28], # req_1: different scattered blocks |
| 1897 | ] |
| 1898 | |
| 1899 | layer_names = [f"layer_{i}" for i in range(cfg.num_layers)] |
| 1900 | # Need enough physical blocks for the scattered IDs |
| 1901 | num_blocks = 50 |
| 1902 | local_cfg = _TestConfig(num_blocks=num_blocks, max_num_reqs=cfg.max_num_reqs) |
| 1903 | kv_cache_config = _make_kv_cache_config(local_cfg, layer_names) |
| 1904 | |
| 1905 | ( |
| 1906 | conv_states_py, |
| 1907 | temporal_states_py, |
| 1908 | conv_states_gpu, |
| 1909 | temporal_states_gpu, |
| 1910 | forward_context_py, |
| 1911 | forward_context_gpu, |
| 1912 | ) = _make_dual_states(local_cfg, layer_names, device) |
| 1913 | |
| 1914 | # --- Run Python path --- |
| 1915 | scheduler_output = _make_postprocess_scheduler_output( |
| 1916 | req_ids, |
| 1917 | num_scheduled_tokens, |
| 1918 | {k: [None] * v for k, v in num_draft_tokens.items() if v > 0}, |
| 1919 | ) |
| 1920 | input_batch_py = _make_input_batch( |
| 1921 | req_ids, num_accepted_tokens.copy(), mamba_state_idx.copy() |
| 1922 | ) |
| 1923 | requests = _make_requests(req_ids, num_computed_tokens, block_ids_per_req) |
| 1924 | copy_bufs = _make_copy_bufs(local_cfg, kv_cache_config, device) |
| 1925 | |
| 1926 | postprocess_mamba( |
| 1927 | scheduler_output, |
| 1928 | kv_cache_config, |
| 1929 | input_batch_py, |
| 1930 | requests, |
| 1931 | forward_context_py, |
| 1932 | _COPY_FUNCS, |
| 1933 | copy_bufs, |
| 1934 | ) |
| 1935 | torch.accelerator.synchronize() |
| 1936 | |
| 1937 | # --- Run GPU path --- |
| 1938 | gpu_ctx = _make_gpu_ctx(local_cfg, kv_cache_config, device) |
| 1939 | num_reqs = len(req_ids) |
| 1940 | max_blocks = max(len(b) for b in block_ids_per_req) |
| 1941 | block_table_gpu = torch.zeros( |
| 1942 | num_reqs, max_blocks, dtype=torch.int32, device=device |
| 1943 | ) |
| 1944 | for i, block_ids in enumerate(block_ids_per_req): |
| 1945 | block_table_gpu[i, : len(block_ids)] = torch.tensor( |
| 1946 | block_ids, dtype=torch.int32 |
| 1947 | ) |
| 1948 | |
| 1949 | _run_gpu_postprocess( |
| 1950 | gpu_ctx, |
| 1951 | kv_cache_config=kv_cache_config, |
| 1952 | forward_context=forward_context_gpu, |
| 1953 | copy_funcs=_COPY_FUNCS, |
| 1954 | block_table=block_table_gpu, |
| 1955 | req_ids=req_ids, |
| 1956 | num_accepted_tokens=num_accepted_tokens, |
| 1957 | mamba_state_idx=mamba_state_idx, |
| 1958 | num_scheduled_tokens=num_scheduled_tokens, |
| 1959 | num_computed_tokens=num_computed_tokens, |
| 1960 | num_draft_tokens=num_draft_tokens, |
| 1961 | device=device, |
| 1962 | ) |
| 1963 | |
| 1964 | # --- Compare results --- |
| 1965 | for i in range(cfg.num_layers): |
| 1966 | torch.testing.assert_close( |
| 1967 | conv_states_gpu[i], |
| 1968 | conv_states_py[i], |
| 1969 | msg=f"Conv state mismatch at layer {i} with non-sequential block IDs", |
| 1970 | ) |
| 1971 | torch.testing.assert_close( |
| 1972 | temporal_states_gpu[i], |
| 1973 | temporal_states_py[i], |
| 1974 | msg=( |
| 1975 | f"Temporal state mismatch at layer {i} " |
| 1976 | f"with non-sequential block IDs" |
| 1977 | ), |
| 1978 | ) |
| 1979 | |
| 1980 | expected_accepted = torch.tensor( |
| 1981 | input_batch_py.num_accepted_tokens_cpu[:num_reqs], |
| 1982 | dtype=torch.int32, |
| 1983 | device=device, |
| 1984 | ) |
| 1985 | torch.testing.assert_close( |
| 1986 | gpu_ctx.num_accepted_tokens_out[:num_reqs], |
| 1987 | expected_accepted, |
| 1988 | msg="num_accepted_tokens mismatch with non-sequential block IDs", |
| 1989 | ) |
| 1990 | |
| 1991 | # Verify req_0 had num_accepted set to 1 (src==dest) and req_1 unchanged |
| 1992 | assert input_batch_py.num_accepted_tokens_cpu[0] == 1 |
| 1993 | assert input_batch_py.num_accepted_tokens_cpu[1] == num_accepted_tokens[1] |
| 1994 | |
| 1995 | def test_prefix_caching_mixed_shared_and_distinct_blocks(self, device, test_config): |
| 1996 | """Regression test: mixed batch under prefix caching where some requests |
| 1997 | have shared physical blocks (aliased) and others have distinct blocks, |
| 1998 | with the needs_copy boundary at various positions. |
| 1999 | |
| 2000 | This tests the interaction between: |
| 2001 | 1. PC block aliasing (src and dest map to same physical block) |
| 2002 | 2. The needs_copy boundary (exact equality vs well-past vs no-copy) |
| 2003 | 3. Non-sequential physical block IDs |
| 2004 | |
| 2005 | Batch of 4 requests: |
| 2006 | - req_0: needs_copy=True, src!=dest, shared physical block (PC aliased) |
| 2007 | -> copy skipped (src_addr==dst_addr), num_accepted PRESERVED |
| 2008 | - req_1: needs_copy=True, src!=dest, distinct blocks, non-sequential IDs |
| 2009 | -> real copy happens, num_accepted PRESERVED |
| 2010 | - req_2: needs_copy=False (below boundary) |
| 2011 | -> no action at all |
| 2012 | - req_3: needs_copy=True, src==dest (exact boundary, zero bias) |
| 2013 | -> copy skipped (self-copy), num_accepted SET TO 1 |
| 2014 | |
| 2015 | Test setup (block_size=16): |
| 2016 | req_0: running=30+2-0=32, new=32+3-1=34, aligned=32, 32>=32 -> COPY |
| 2017 | bias=0, dest=32//16-1=1, src=0 (!=dest) |
| 2018 | block_ids=[5,5,...] -> same physical -> skip, keep accepted=3 |
| 2019 | |
| 2020 | req_1: running=60+5-2=63, new=63+3-1=65, aligned=64, 64>=63 -> COPY |
| 2021 | bias=1, dest=64//16-1=3, src=2 (!=dest) |
| 2022 | block_ids=[41,7,22,15,...] -> distinct -> real copy, keep accepted=3 |
| 2023 | |
| 2024 | req_2: running=30+3-0=33, new=33+1-1=33, aligned=32, 32<33 -> NO COPY |
| 2025 | |
| 2026 | req_3: running=48+0-0=48, new=48+1-1=48, aligned=48, 48>=48 -> COPY |
| 2027 | bias=0, dest=48//16-1=2, src=2 (==dest) |
| 2028 | block_ids=[10,20,30,...] -> distinct IDs, same logical idx |
| 2029 | -> self-copy (src_addr==dst_addr), set accepted=1 |
| 2030 | """ |
| 2031 | cfg = test_config |
| 2032 | torch.manual_seed(5001) |
| 2033 | |
| 2034 | req_ids = ["req_0", "req_1", "req_2", "req_3"] |
| 2035 | num_computed_tokens = [30, 60, 30, 48] |
| 2036 | num_scheduled_tokens = {"req_0": 2, "req_1": 5, "req_2": 3, "req_3": 0} |
| 2037 | num_draft_tokens = {"req_1": 2} |
| 2038 | num_accepted_tokens = [3, 3, 1, 1] |
| 2039 | mamba_state_idx = [0, 2, 1, 2] |
| 2040 | |
| 2041 | # Block IDs with various PC patterns: |
| 2042 | # req_0: shared blocks (PC alias: logical 0 and 1 -> physical 5) |
| 2043 | # req_1: distinct non-sequential blocks |
| 2044 | # req_2: doesn't matter (no copy) |
| 2045 | # req_3: distinct sequential blocks (no aliasing) |
| 2046 | block_ids_per_req = [ |
| 2047 | [5, 5, 12, 18, 23, 31, 44, 2], # req_0: blocks 0,1 share phys 5 |
| 2048 | [41, 7, 22, 15, 38, 19, 4, 28], # req_1: all distinct |
| 2049 | [10, 20, 30, 40, 1, 6, 8, 14], # req_2: irrelevant |
| 2050 | [10, 20, 30, 40, 1, 6, 8, 14], # req_3: distinct, dest=src=idx 2 |
| 2051 | ] |
| 2052 | |
| 2053 | layer_names = [f"layer_{i}" for i in range(cfg.num_layers)] |
| 2054 | num_blocks = 50 |
| 2055 | local_cfg = _TestConfig(num_blocks=num_blocks, max_num_reqs=cfg.max_num_reqs) |
| 2056 | kv_cache_config = _make_kv_cache_config(local_cfg, layer_names) |
| 2057 | |
| 2058 | ( |
| 2059 | conv_states_py, |
| 2060 | temporal_states_py, |
| 2061 | conv_states_gpu, |
| 2062 | temporal_states_gpu, |
| 2063 | forward_context_py, |
| 2064 | forward_context_gpu, |
| 2065 | ) = _make_dual_states(local_cfg, layer_names, device) |
| 2066 | |
| 2067 | # --- Run Python path --- |
| 2068 | scheduler_output = _make_postprocess_scheduler_output( |
| 2069 | req_ids, |
| 2070 | num_scheduled_tokens, |
| 2071 | {k: [None] * v for k, v in num_draft_tokens.items() if v > 0}, |
| 2072 | ) |
| 2073 | input_batch_py = _make_input_batch( |
| 2074 | req_ids, num_accepted_tokens.copy(), mamba_state_idx.copy() |
| 2075 | ) |
| 2076 | requests = _make_requests(req_ids, num_computed_tokens, block_ids_per_req) |
| 2077 | copy_bufs = _make_copy_bufs(local_cfg, kv_cache_config, device) |
| 2078 | |
| 2079 | postprocess_mamba( |
| 2080 | scheduler_output, |
| 2081 | kv_cache_config, |
| 2082 | input_batch_py, |
| 2083 | requests, |
| 2084 | forward_context_py, |
| 2085 | _COPY_FUNCS, |
| 2086 | copy_bufs, |
| 2087 | ) |
| 2088 | torch.accelerator.synchronize() |
| 2089 | |
| 2090 | # --- Run GPU path --- |
| 2091 | gpu_ctx = _make_gpu_ctx(local_cfg, kv_cache_config, device) |
| 2092 | num_reqs = len(req_ids) |
| 2093 | max_blocks = max(len(b) for b in block_ids_per_req) |
| 2094 | block_table_gpu = torch.zeros( |
| 2095 | num_reqs, max_blocks, dtype=torch.int32, device=device |
| 2096 | ) |
| 2097 | for i, block_ids in enumerate(block_ids_per_req): |
| 2098 | block_table_gpu[i, : len(block_ids)] = torch.tensor( |
| 2099 | block_ids, dtype=torch.int32 |
| 2100 | ) |
| 2101 | |
| 2102 | _run_gpu_postprocess( |
| 2103 | gpu_ctx, |
| 2104 | kv_cache_config=kv_cache_config, |
| 2105 | forward_context=forward_context_gpu, |
| 2106 | copy_funcs=_COPY_FUNCS, |
| 2107 | block_table=block_table_gpu, |
| 2108 | req_ids=req_ids, |
| 2109 | num_accepted_tokens=num_accepted_tokens, |
| 2110 | mamba_state_idx=mamba_state_idx, |
| 2111 | num_scheduled_tokens=num_scheduled_tokens, |
| 2112 | num_computed_tokens=num_computed_tokens, |
| 2113 | num_draft_tokens=num_draft_tokens, |
| 2114 | device=device, |
| 2115 | ) |
| 2116 | |
| 2117 | # --- Compare all state tensors --- |
| 2118 | for i in range(cfg.num_layers): |
| 2119 | torch.testing.assert_close( |
| 2120 | conv_states_gpu[i], |
| 2121 | conv_states_py[i], |
| 2122 | msg=( |
| 2123 | f"Conv state mismatch at layer {i} — " |
| 2124 | f"mixed PC batch with shared/distinct blocks" |
| 2125 | ), |
| 2126 | ) |
| 2127 | torch.testing.assert_close( |
| 2128 | temporal_states_gpu[i], |
| 2129 | temporal_states_py[i], |
| 2130 | msg=( |
| 2131 | f"Temporal state mismatch at layer {i} — " |
| 2132 | f"mixed PC batch with shared/distinct blocks" |
| 2133 | ), |
| 2134 | ) |
| 2135 | |
| 2136 | # --- Compare num_accepted_tokens --- |
| 2137 | expected_accepted = torch.tensor( |
| 2138 | input_batch_py.num_accepted_tokens_cpu[:num_reqs], |
| 2139 | dtype=torch.int32, |
| 2140 | device=device, |
| 2141 | ) |
| 2142 | torch.testing.assert_close( |
| 2143 | gpu_ctx.num_accepted_tokens_out[:num_reqs], |
| 2144 | expected_accepted, |
| 2145 | msg="num_accepted_tokens mismatch in mixed PC batch", |
| 2146 | ) |
| 2147 | |
| 2148 | # Verify per-request expectations: |
| 2149 | # req_0: src!=dest, shared block -> preserved (3) |
| 2150 | assert input_batch_py.num_accepted_tokens_cpu[0] == 3 |
| 2151 | # req_1: src!=dest, distinct blocks -> preserved (3) |
| 2152 | assert input_batch_py.num_accepted_tokens_cpu[1] == 3 |
| 2153 | # req_2: no copy -> preserved (1) |
| 2154 | assert input_batch_py.num_accepted_tokens_cpu[2] == 1 |
| 2155 | # req_3: src==dest -> set to 1 |
| 2156 | assert input_batch_py.num_accepted_tokens_cpu[3] == 1 |
| 2157 | |
| 2158 | def test_pc_aliased_blocks_skip_must_use_logical_idx_not_addr( |
| 2159 | self, device, test_config |
| 2160 | ): |
| 2161 | """Regression test for 6466ce0d vs 959ca0fd: the kernel's early-return |
| 2162 | guard must compare logical block indices, not physical addresses. |
| 2163 | |
| 2164 | Under prefix caching, different logical blocks (src_block_idx=0, |
| 2165 | dest_block_idx=1) can map to the same physical block. When |
| 2166 | accept_token_bias=0, this makes src_addr == dst_addr for BOTH conv |
| 2167 | and temporal states. A buggy guard `if src_addr == dst_addr` would |
| 2168 | incorrectly set num_accepted_tokens=1; the correct guard is |
| 2169 | `if src_block_idx == dest_block_idx and accept_token_bias == 0`. |
| 2170 | |
| 2171 | The Python reference only sets num_accepted_tokens=1 when |
| 2172 | src_block_idx == dest_block_idx (line 79 of postprocess_mamba). |
| 2173 | With src_block_idx=0, dest_block_idx=1, num_accepted_tokens must |
| 2174 | be preserved even though the physical addresses match. |
| 2175 | |
| 2176 | Test setup (block_size=16): |
| 2177 | - num_tokens_running_state = 30 + 2 - 0 = 32 |
| 2178 | - new_num_computed = 32 + 3 - 1 = 34 |
| 2179 | - aligned_new_computed = 32 |
| 2180 | - needs_copy = (32 >= 32) = True |
| 2181 | - accept_token_bias = 32 - 32 = 0 |
| 2182 | - dest_block_idx = 32 // 16 - 1 = 1 |
| 2183 | - src_block_idx = 0 (explicitly, != dest_block_idx) |
| 2184 | - block_ids = [7, 7, ...] -> physical aliasing via prefix caching |
| 2185 | |
| 2186 | Expected: num_accepted_tokens stays 3 (not set to 1). |
| 2187 | Bug (959ca0fd): kernel saw src_addr == dst_addr, set it to 1. |
| 2188 | """ |
| 2189 | cfg = test_config |
| 2190 | torch.manual_seed(6001) |
| 2191 | |
| 2192 | req_ids = ["req_0"] |
| 2193 | num_computed_tokens = [30] |
| 2194 | num_scheduled_tokens = {"req_0": 2} |
| 2195 | num_draft_tokens: dict[str, int] = {} |
| 2196 | num_accepted_tokens = [3] |
| 2197 | mamba_state_idx = [0] # src_block_idx = 0 |
| 2198 | |
| 2199 | # Prefix caching: logical blocks 0 and 1 both map to physical block 7. |
| 2200 | block_ids_per_req = [[7, 7, 10, 11, 12, 13, 14, 15]] |
| 2201 | |
| 2202 | layer_names = ["layer_0"] |
| 2203 | kv_cache_config = _make_kv_cache_config(cfg, layer_names) |
| 2204 | |
| 2205 | ( |
| 2206 | conv_state_py, |
| 2207 | temporal_state_py, |
| 2208 | conv_state_gpu, |
| 2209 | temporal_state_gpu, |
| 2210 | forward_context_py, |
| 2211 | forward_context_gpu, |
| 2212 | ) = _make_dual_layer_state(cfg, device) |
| 2213 | |
| 2214 | # --- Run Python path --- |
| 2215 | scheduler_output = _make_postprocess_scheduler_output( |
| 2216 | req_ids, |
| 2217 | num_scheduled_tokens, |
| 2218 | {k: [None] * v for k, v in num_draft_tokens.items() if v > 0}, |
| 2219 | ) |
| 2220 | input_batch_py = _make_input_batch( |
| 2221 | req_ids, num_accepted_tokens.copy(), mamba_state_idx.copy() |
| 2222 | ) |
| 2223 | requests = _make_requests(req_ids, num_computed_tokens, block_ids_per_req) |
| 2224 | copy_bufs = _make_copy_bufs(cfg, kv_cache_config, device) |
| 2225 | |
| 2226 | postprocess_mamba( |
| 2227 | scheduler_output, |
| 2228 | kv_cache_config, |
| 2229 | input_batch_py, |
| 2230 | requests, |
| 2231 | forward_context_py, |
| 2232 | _COPY_FUNCS, |
| 2233 | copy_bufs, |
| 2234 | ) |
| 2235 | torch.accelerator.synchronize() |
| 2236 | |
| 2237 | # Python reference: src_block_idx(0) != dest_block_idx(1) -> no change |
| 2238 | assert input_batch_py.num_accepted_tokens_cpu[0] == 3, ( |
| 2239 | f"Python: num_accepted_tokens should remain 3, " |
| 2240 | f"got {input_batch_py.num_accepted_tokens_cpu[0]}" |
| 2241 | ) |
| 2242 | |
| 2243 | # --- Run GPU path --- |
| 2244 | gpu_ctx = _make_gpu_ctx(cfg, kv_cache_config, device) |
| 2245 | num_reqs = len(req_ids) |
| 2246 | block_table_gpu = torch.zeros(num_reqs, 8, dtype=torch.int32, device=device) |
| 2247 | block_table_gpu[0, :8] = torch.tensor(block_ids_per_req[0], dtype=torch.int32) |
| 2248 | |
| 2249 | _run_gpu_postprocess( |
| 2250 | gpu_ctx, |
| 2251 | kv_cache_config=kv_cache_config, |
| 2252 | forward_context=forward_context_gpu, |
| 2253 | copy_funcs=_COPY_FUNCS, |
| 2254 | block_table=block_table_gpu, |
| 2255 | req_ids=req_ids, |
| 2256 | num_accepted_tokens=num_accepted_tokens, |
| 2257 | mamba_state_idx=mamba_state_idx, |
| 2258 | num_scheduled_tokens=num_scheduled_tokens, |
| 2259 | num_computed_tokens=num_computed_tokens, |
| 2260 | num_draft_tokens=num_draft_tokens, |
| 2261 | device=device, |
| 2262 | ) |
| 2263 | |
| 2264 | # The critical assertion: kernel must NOT set num_accepted_tokens to 1 |
| 2265 | # when src_block_idx != dest_block_idx, even though src_addr == dst_addr |
| 2266 | # due to prefix caching aliasing. |
| 2267 | # |
| 2268 | # Old kernel (959ca0fd): `if src_addr == dst_addr` -> FAILS here (sets 1) |
| 2269 | # Fixed kernel (6466ce0d): `if src_block_idx == dest_block_idx and |
| 2270 | # accept_token_bias == 0` -> PASSES (preserves 3) |
| 2271 | kernel_accepted = gpu_ctx.num_accepted_tokens_out[0].item() |
| 2272 | assert kernel_accepted == 3, ( |
| 2273 | f"Kernel set num_accepted_tokens to {kernel_accepted} but expected 3. " |
| 2274 | f"The early-return guard likely compared physical addresses " |
| 2275 | f"(src_addr == dst_addr) instead of logical block indices " |
| 2276 | f"(src_block_idx == dest_block_idx). Under prefix caching, " |
| 2277 | f"different logical blocks can share the same physical block." |
| 2278 | ) |
| 2279 | |
| 2280 | # Also verify state tensors match Python |
| 2281 | torch.testing.assert_close( |
| 2282 | conv_state_gpu, |
| 2283 | conv_state_py, |
| 2284 | msg="GPU conv state should match Python", |
| 2285 | ) |
| 2286 | torch.testing.assert_close( |
| 2287 | temporal_state_gpu, |
| 2288 | temporal_state_py, |
| 2289 | msg="GPU temporal state should match Python", |
| 2290 | ) |
| 2291 | |
| 2292 | def test_as_strided_temporal_copy_size(self, device, test_config): |
| 2293 | """Regression test for 240723d46: temporal copy_size must be |
| 2294 | inner_size * elem_size, not state_block_stride. |
| 2295 | |
| 2296 | In production (gpu_model_runner.py), conv and temporal states share |
| 2297 | a raw buffer via torch.as_strided where stride(0) equals |
| 2298 | page_size_bytes / elem_size — larger than either state's natural |
| 2299 | element count. Using stride(0) as copy_size for temporal states |
| 2300 | overwrites into the next block's conv region. |
| 2301 | |
| 2302 | Layout per page (384 float16 elements = 768 bytes): |
| 2303 | [conv: 256 elems | temporal: 128 elems] |
| 2304 | |
| 2305 | The test triggers a temporal copy from block 4 to block 3. With the |
| 2306 | bug the kernel copies 768 bytes (page stride) instead of 256 bytes |
| 2307 | (128 * 2), overwriting conv_state[4] with conv_state[5]'s data. |
| 2308 | |
| 2309 | Test setup (block_size=16): |
| 2310 | - running = 60 + 5 - 2 = 63 |
| 2311 | - new = 63 + 3 - 1 = 65 |
| 2312 | - aligned = 64 >= 63 -> COPY needed |
| 2313 | - accept_token_bias = 64 - 63 = 1 |
| 2314 | - dest_block_idx = 64 // 16 - 1 = 3 |
| 2315 | - temporal: actual_src_block_idx = 3 + 1 = 4 (block_ids[4] = 4) |
| 2316 | """ |
| 2317 | cfg = test_config |
| 2318 | torch.manual_seed(7001) |
| 2319 | |
| 2320 | req_ids = ["req_0"] |
| 2321 | num_computed_tokens = [60] |
| 2322 | num_scheduled_tokens = {"req_0": 5} |
| 2323 | num_draft_tokens = {"req_0": 2} |
| 2324 | num_accepted_tokens = [3] |
| 2325 | mamba_state_idx = [3] |
| 2326 | block_ids_per_req = [list(range(8))] |
| 2327 | |
| 2328 | layer_names = ["layer_0"] |
| 2329 | kv_cache_config = _make_kv_cache_config(cfg, layer_names) |
| 2330 | |
| 2331 | # --- Production-like packed layout (mirrors gpu_model_runner.py) --- |
| 2332 | conv_shape = (cfg.conv_width, cfg.conv_inner_dim) |
| 2333 | temporal_shape = (cfg.temporal_state_dim,) |
| 2334 | dtype = cfg.dtype |
| 2335 | elem_size = torch.tensor([], dtype=dtype).element_size() |
| 2336 | |
| 2337 | conv_natural_elems = cfg.conv_width * cfg.conv_inner_dim |
| 2338 | temporal_natural_elems = cfg.temporal_state_dim |
| 2339 | page_size_bytes = (conv_natural_elems + temporal_natural_elems) * elem_size |
| 2340 | num_element_per_page = page_size_bytes // elem_size |
| 2341 | |
| 2342 | assert num_element_per_page > temporal_natural_elems, ( |
| 2343 | "Test requires padded stride; page must be larger than one state" |
| 2344 | ) |
| 2345 | |
| 2346 | raw_py = torch.randn( |
| 2347 | cfg.num_blocks * num_element_per_page, dtype=dtype, device=device |
| 2348 | ) |
| 2349 | raw_gpu = raw_py.clone() |
| 2350 | |
| 2351 | def make_views(raw): |
| 2352 | conv_tgt = (cfg.num_blocks, *conv_shape) |
| 2353 | conv_nat_stride = torch.empty(conv_tgt).stride() |
| 2354 | conv = torch.as_strided( |
| 2355 | raw, |
| 2356 | size=conv_tgt, |
| 2357 | stride=(num_element_per_page, *conv_nat_stride[1:]), |
| 2358 | storage_offset=0, |
| 2359 | ) |
| 2360 | |
| 2361 | temp_tgt = (cfg.num_blocks, *temporal_shape) |
| 2362 | temp_nat_stride = torch.empty(temp_tgt).stride() |
| 2363 | temp = torch.as_strided( |
| 2364 | raw, |
| 2365 | size=temp_tgt, |
| 2366 | stride=(num_element_per_page, *temp_nat_stride[1:]), |
| 2367 | storage_offset=conv_natural_elems, |
| 2368 | ) |
| 2369 | return conv, temp |
| 2370 | |
| 2371 | conv_py, temp_py = make_views(raw_py) |
| 2372 | conv_gpu, temp_gpu = make_views(raw_gpu) |
| 2373 | |
| 2374 | fwd_py = {"layer_0": _make_mock_attention(conv_py, temp_py)} |
| 2375 | fwd_gpu = {"layer_0": _make_mock_attention(conv_gpu, temp_gpu)} |
| 2376 | |
| 2377 | # --- Python reference --- |
| 2378 | sched = _make_postprocess_scheduler_output( |
| 2379 | req_ids, |
| 2380 | num_scheduled_tokens, |
| 2381 | {k: [None] * v for k, v in num_draft_tokens.items() if v > 0}, |
| 2382 | ) |
| 2383 | batch_py = _make_input_batch( |
| 2384 | req_ids, num_accepted_tokens.copy(), mamba_state_idx.copy() |
| 2385 | ) |
| 2386 | requests = _make_requests(req_ids, num_computed_tokens, block_ids_per_req) |
| 2387 | copy_bufs = _make_copy_bufs(cfg, kv_cache_config, device) |
| 2388 | |
| 2389 | postprocess_mamba( |
| 2390 | sched, |
| 2391 | kv_cache_config, |
| 2392 | batch_py, |
| 2393 | requests, |
| 2394 | fwd_py, |
| 2395 | _COPY_FUNCS, |
| 2396 | copy_bufs, |
| 2397 | ) |
| 2398 | torch.accelerator.synchronize() |
| 2399 | |
| 2400 | # --- GPU fused kernel --- |
| 2401 | gpu_ctx = _make_gpu_ctx(cfg, kv_cache_config, device) |
| 2402 | num_reqs = 1 |
| 2403 | block_table = torch.zeros(num_reqs, 8, dtype=torch.int32, device=device) |
| 2404 | block_table[0, :8] = torch.tensor(block_ids_per_req[0], dtype=torch.int32) |
| 2405 | |
| 2406 | _run_gpu_postprocess( |
| 2407 | gpu_ctx, |
| 2408 | kv_cache_config=kv_cache_config, |
| 2409 | forward_context=fwd_gpu, |
| 2410 | copy_funcs=_COPY_FUNCS, |
| 2411 | block_table=block_table, |
| 2412 | req_ids=req_ids, |
| 2413 | num_accepted_tokens=num_accepted_tokens, |
| 2414 | mamba_state_idx=mamba_state_idx, |
| 2415 | num_scheduled_tokens=num_scheduled_tokens, |
| 2416 | num_computed_tokens=num_computed_tokens, |
| 2417 | num_draft_tokens=num_draft_tokens, |
| 2418 | device=device, |
| 2419 | ) |
| 2420 | |
| 2421 | # --- Assertions --- |
| 2422 | # With the bug (pre-240723d46), the kernel copies page_size_bytes |
| 2423 | # (768) for temporal state instead of 256 bytes, overwriting |
| 2424 | # conv_state[4] with conv_state[5]'s data. |
| 2425 | torch.testing.assert_close( |
| 2426 | conv_gpu, |
| 2427 | conv_py, |
| 2428 | msg=( |
| 2429 | "Conv state corrupted: temporal copy_size was likely " |
| 2430 | "state_block_stride instead of inner_size * elem_size" |
| 2431 | ), |
| 2432 | ) |
| 2433 | torch.testing.assert_close( |
| 2434 | temp_gpu, |
| 2435 | temp_py, |
| 2436 | msg="Temporal state mismatch", |
| 2437 | ) |
| 2438 | |
| 2439 | expected_accepted = torch.tensor( |
| 2440 | batch_py.num_accepted_tokens_cpu[:num_reqs], |
| 2441 | dtype=torch.int32, |
| 2442 | device=device, |
| 2443 | ) |
| 2444 | torch.testing.assert_close( |
| 2445 | gpu_ctx.num_accepted_tokens_out[:num_reqs], |
| 2446 | expected_accepted, |
| 2447 | msg="num_accepted_tokens mismatch", |
| 2448 | ) |
| 2449 | |
| 2450 | def test_temporal_copy_with_bias_ge_2(self, device, test_config): |
| 2451 | """Coverage test for the temporal-state block-table stride arithmetic |
| 2452 | when ``accept_token_bias >= 2``. |
| 2453 | |
| 2454 | The kernel computes, for temporal (non-conv) states:: |
| 2455 | |
| 2456 | actual_src_block_idx = src_block_idx + accept_token_bias |
| 2457 | actual_src_block_id = block_table[req, actual_src_block_idx] |
| 2458 | |
| 2459 | A ``bias == 1`` case only reads one slot ahead of ``src_block_idx`` |
| 2460 | in the block table. This test isolates the larger-stride case, where |
| 2461 | an off-by-one would read the wrong physical block after multiple |
| 2462 | tokens are accepted across a block boundary. |
| 2463 | |
| 2464 | Setup (block_size=16): |
| 2465 | - running = 28 + 2 - 0 = 30 |
| 2466 | - new = 30 + 3 - 1 = 32 |
| 2467 | - aligned = 32 >= 30 -> COPY needed |
| 2468 | - bias = 32 - 30 = 2 (key: >= 2) |
| 2469 | - dest_idx = 32 // 16 - 1 = 1 |
| 2470 | - src_idx = 1 (same as dest -> exercises post-copy accepted=1 write) |
| 2471 | - temporal actual_src_block_idx = 1 + 2 = 3 (reads block_table[0, 3]) |
| 2472 | |
| 2473 | With identity block_ids = [0,1,2,3,...], an off-by-one that used |
| 2474 | bias=1 would copy from block_ids[2]=2 instead of block_ids[3]=3, |
| 2475 | producing a clear mismatch against the untouched snapshot. |
| 2476 | """ |
| 2477 | cfg = test_config |
| 2478 | torch.manual_seed(7002) |
| 2479 | |
| 2480 | req_ids = ["req_0"] |
| 2481 | num_computed_tokens = [28] |
| 2482 | num_scheduled_tokens = {"req_0": 2} |
| 2483 | num_draft_tokens: dict[str, int] = {} |
| 2484 | num_accepted_tokens = [3] # -> accept_token_bias = 2 |
| 2485 | mamba_state_idx = [1] # src_block_idx = 1 = dest_block_idx |
| 2486 | block_ids_per_req = [list(range(8))] |
| 2487 | |
| 2488 | layer_names = ["layer_0"] |
| 2489 | kv_cache_config = _make_kv_cache_config(cfg, layer_names) |
| 2490 | |
| 2491 | ( |
| 2492 | conv_state_py, |
| 2493 | temporal_state_py, |
| 2494 | conv_state_gpu, |
| 2495 | temporal_state_gpu, |
| 2496 | fwd_py, |
| 2497 | fwd_gpu, |
| 2498 | ) = _make_dual_layer_state(cfg, device) |
| 2499 | conv_state_orig = conv_state_py.clone() |
| 2500 | temporal_state_orig = temporal_state_py.clone() |
| 2501 | |
| 2502 | # --- Python reference --- |
| 2503 | sched = _make_postprocess_scheduler_output( |
| 2504 | req_ids, |
| 2505 | num_scheduled_tokens, |
| 2506 | {k: [None] * v for k, v in num_draft_tokens.items() if v > 0}, |
| 2507 | ) |
| 2508 | batch_py = _make_input_batch( |
| 2509 | req_ids, num_accepted_tokens.copy(), mamba_state_idx.copy() |
| 2510 | ) |
| 2511 | requests = _make_requests(req_ids, num_computed_tokens, block_ids_per_req) |
| 2512 | copy_bufs = _make_copy_bufs(cfg, kv_cache_config, device) |
| 2513 | |
| 2514 | postprocess_mamba( |
| 2515 | sched, |
| 2516 | kv_cache_config, |
| 2517 | batch_py, |
| 2518 | requests, |
| 2519 | fwd_py, |
| 2520 | _COPY_FUNCS, |
| 2521 | copy_bufs, |
| 2522 | ) |
| 2523 | torch.accelerator.synchronize() |
| 2524 | |
| 2525 | # --- GPU fused kernel --- |
| 2526 | gpu_ctx = _make_gpu_ctx(cfg, kv_cache_config, device) |
| 2527 | num_reqs = 1 |
| 2528 | block_table = torch.zeros(num_reqs, 8, dtype=torch.int32, device=device) |
| 2529 | block_table[0, :8] = torch.tensor(block_ids_per_req[0], dtype=torch.int32) |
| 2530 | |
| 2531 | _run_gpu_postprocess( |
| 2532 | gpu_ctx, |
| 2533 | kv_cache_config=kv_cache_config, |
| 2534 | forward_context=fwd_gpu, |
| 2535 | copy_funcs=_COPY_FUNCS, |
| 2536 | block_table=block_table, |
| 2537 | req_ids=req_ids, |
| 2538 | num_accepted_tokens=num_accepted_tokens, |
| 2539 | mamba_state_idx=mamba_state_idx, |
| 2540 | num_scheduled_tokens=num_scheduled_tokens, |
| 2541 | num_computed_tokens=num_computed_tokens, |
| 2542 | num_draft_tokens=num_draft_tokens, |
| 2543 | device=device, |
| 2544 | ) |
| 2545 | |
| 2546 | # --- Ground truth from untouched snapshots --- |
| 2547 | actual_src_block_id = block_ids_per_req[0][3] # == 3 |
| 2548 | dest_block_id = block_ids_per_req[0][1] # == 1 |
| 2549 | expected_conv_state = conv_state_orig.clone() |
| 2550 | expected_conv_state[dest_block_id, :-2].copy_( |
| 2551 | conv_state_orig[dest_block_id, 2:] |
| 2552 | ) |
| 2553 | torch.testing.assert_close(conv_state_py, expected_conv_state, rtol=0, atol=0) |
| 2554 | torch.testing.assert_close(conv_state_gpu, expected_conv_state, rtol=0, atol=0) |
| 2555 | |
| 2556 | # Python must have sourced temporal from block 3. |
| 2557 | torch.testing.assert_close( |
| 2558 | temporal_state_py[dest_block_id], |
| 2559 | temporal_state_orig[actual_src_block_id], |
| 2560 | rtol=0, |
| 2561 | atol=0, |
| 2562 | msg=( |
| 2563 | "Python reference did not copy from block_ids[src+bias]=3; " |
| 2564 | "test preconditions are wrong" |
| 2565 | ), |
| 2566 | ) |
| 2567 | |
| 2568 | # --- GPU kernel must match Python byte-for-byte --- |
| 2569 | torch.testing.assert_close( |
| 2570 | conv_state_gpu, |
| 2571 | conv_state_py, |
| 2572 | msg="Conv state mismatch at accept_token_bias=2", |
| 2573 | ) |
| 2574 | torch.testing.assert_close( |
| 2575 | temporal_state_gpu, |
| 2576 | temporal_state_py, |
| 2577 | msg=( |
| 2578 | "Temporal state mismatch at accept_token_bias=2: the kernel " |
| 2579 | "likely read the wrong slot of the block table " |
| 2580 | "(actual_src_block_idx stride arithmetic)" |
| 2581 | ), |
| 2582 | ) |
| 2583 | |
| 2584 | expected_accepted = torch.tensor( |
| 2585 | batch_py.num_accepted_tokens_cpu[:num_reqs], |
| 2586 | dtype=torch.int32, |
| 2587 | device=device, |
| 2588 | ) |
| 2589 | torch.testing.assert_close( |
| 2590 | gpu_ctx.num_accepted_tokens_out[:num_reqs], |
| 2591 | expected_accepted, |
| 2592 | msg="num_accepted_tokens mismatch at accept_token_bias=2", |
| 2593 | ) |
| 2594 | |
| 2595 | @pytest.mark.parametrize( |
| 2596 | "same_physical_block", [True, False], ids=["same", "distinct"] |
| 2597 | ) |
| 2598 | @pytest.mark.parametrize("accept_token_bias", [1, 2, 3]) |
| 2599 | @pytest.mark.parametrize( |
| 2600 | "dtype", |
| 2601 | [torch.float16, torch.float32, torch.float64], |
| 2602 | ids=["fp16", "fp32", "fp64"], |
| 2603 | ) |
| 2604 | def test_sd_and_ds_conv_layouts_match_snapshot( |
| 2605 | self, |
| 2606 | device, |
| 2607 | test_config, |
| 2608 | monkeypatch, |
| 2609 | accept_token_bias, |
| 2610 | same_physical_block, |
| 2611 | dtype, |
| 2612 | ): |
| 2613 | """SD and DS copies should independently match memmove semantics.""" |
| 2614 | from vllm.model_executor.layers.mamba import mamba_utils as model_mamba_utils |
| 2615 | |
| 2616 | cfg = test_config |
| 2617 | cfg.dtype = dtype |
| 2618 | torch.manual_seed(38898) |
| 2619 | |
| 2620 | req_ids = ["req_0"] |
| 2621 | # Keep new_num_computed on an aligned boundary while varying how far |
| 2622 | # below it the running state starts. This makes the copy bias exactly |
| 2623 | # ``accept_token_bias`` for each case. The 32 boundary keeps source and |
| 2624 | # destination in logical block 1; the 64 boundary copies block 2 -> 3. |
| 2625 | aligned_boundary = 32 if same_physical_block else 64 |
| 2626 | num_computed_tokens = [aligned_boundary - 2 * accept_token_bias] |
| 2627 | num_scheduled_tokens = {"req_0": accept_token_bias} |
| 2628 | num_draft_tokens: dict[str, int] = {} |
| 2629 | num_accepted_tokens = [accept_token_bias + 1] |
| 2630 | dest_block_idx = aligned_boundary // cfg.block_size - 1 |
| 2631 | src_block_idx = dest_block_idx if same_physical_block else dest_block_idx - 1 |
| 2632 | mamba_state_idx = [src_block_idx] |
| 2633 | block_ids_per_req = [list(range(8))] |
| 2634 | |
| 2635 | layer_names = ["layer_0"] |
| 2636 | kv_cache_config = _make_kv_cache_config(cfg, layer_names) |
| 2637 | |
| 2638 | num_reqs = len(req_ids) |
| 2639 | block_table_gpu = torch.zeros(num_reqs, 8, dtype=torch.int32, device=device) |
| 2640 | block_table_gpu[0, :8] = torch.tensor(block_ids_per_req[0], dtype=torch.int32) |
| 2641 | |
| 2642 | # Same logical conv state in SD and DS layouts. |
| 2643 | sd_source_conv = torch.randn( |
| 2644 | cfg.num_blocks, |
| 2645 | cfg.conv_width, |
| 2646 | cfg.conv_inner_dim, |
| 2647 | dtype=cfg.dtype, |
| 2648 | device=device, |
| 2649 | ) |
| 2650 | ds_source_conv = sd_source_conv.permute(0, 2, 1).contiguous() |
| 2651 | sd_source_temporal = torch.randn( |
| 2652 | cfg.num_blocks, cfg.temporal_state_dim, dtype=cfg.dtype, device=device |
| 2653 | ) |
| 2654 | |
| 2655 | # SD GPU path. |
| 2656 | monkeypatch.delenv("VLLM_SSM_CONV_STATE_LAYOUT", raising=False) |
| 2657 | model_mamba_utils.get_conv_state_layout.cache_clear() |
| 2658 | sd_conv = sd_source_conv.clone() |
| 2659 | sd_temporal = sd_source_temporal.clone() |
| 2660 | forward_context_sd = { |
| 2661 | "layer_0": _make_mock_attention(sd_conv, sd_temporal), |
| 2662 | } |
| 2663 | gpu_ctx_sd = _make_gpu_ctx(cfg, kv_cache_config, device) |
| 2664 | _run_gpu_postprocess( |
| 2665 | gpu_ctx_sd, |
| 2666 | kv_cache_config=kv_cache_config, |
| 2667 | forward_context=forward_context_sd, |
| 2668 | copy_funcs=_COPY_FUNCS, |
| 2669 | block_table=block_table_gpu, |
| 2670 | req_ids=req_ids, |
| 2671 | num_accepted_tokens=num_accepted_tokens, |
| 2672 | mamba_state_idx=mamba_state_idx, |
| 2673 | num_scheduled_tokens=num_scheduled_tokens, |
| 2674 | num_computed_tokens=num_computed_tokens, |
| 2675 | num_draft_tokens=num_draft_tokens, |
| 2676 | device=device, |
| 2677 | ) |
| 2678 | torch.accelerator.synchronize() |
| 2679 | |
| 2680 | src_block_id = block_ids_per_req[0][src_block_idx] |
| 2681 | dest_block_id = block_ids_per_req[0][dest_block_idx] |
| 2682 | expected_conv = sd_source_conv.clone() |
| 2683 | expected_conv[dest_block_id, :-accept_token_bias].copy_( |
| 2684 | sd_source_conv[src_block_id, accept_token_bias:] |
| 2685 | ) |
| 2686 | torch.testing.assert_close( |
| 2687 | sd_conv, |
| 2688 | expected_conv, |
| 2689 | rtol=0, |
| 2690 | atol=0, |
| 2691 | msg="SD conv copy did not match the untouched source snapshot", |
| 2692 | ) |
| 2693 | |
| 2694 | actual_temporal_src_idx = src_block_idx + accept_token_bias |
| 2695 | actual_temporal_src_id = block_ids_per_req[0][actual_temporal_src_idx] |
| 2696 | expected_temporal = sd_source_temporal.clone() |
| 2697 | expected_temporal[dest_block_id].copy_( |
| 2698 | sd_source_temporal[actual_temporal_src_id] |
| 2699 | ) |
| 2700 | torch.testing.assert_close( |
| 2701 | sd_temporal, |
| 2702 | expected_temporal, |
| 2703 | rtol=0, |
| 2704 | atol=0, |
| 2705 | msg="SD temporal copy did not match the untouched source snapshot", |
| 2706 | ) |
| 2707 | |
| 2708 | # DS GPU path on the DS twin. |
| 2709 | monkeypatch.setenv("VLLM_SSM_CONV_STATE_LAYOUT", "DS") |
| 2710 | model_mamba_utils.get_conv_state_layout.cache_clear() |
| 2711 | try: |
| 2712 | ds_conv = ds_source_conv.clone() |
| 2713 | ds_temporal = sd_source_temporal.clone() |
| 2714 | forward_context_ds = { |
| 2715 | "layer_0": _make_mock_attention(ds_conv, ds_temporal), |
| 2716 | } |
| 2717 | gpu_ctx_ds = _make_gpu_ctx(cfg, kv_cache_config, device) |
| 2718 | _run_gpu_postprocess( |
| 2719 | gpu_ctx_ds, |
| 2720 | kv_cache_config=kv_cache_config, |
| 2721 | forward_context=forward_context_ds, |
| 2722 | copy_funcs=_COPY_FUNCS, |
| 2723 | block_table=block_table_gpu, |
| 2724 | req_ids=req_ids, |
| 2725 | num_accepted_tokens=num_accepted_tokens, |
| 2726 | mamba_state_idx=mamba_state_idx, |
| 2727 | num_scheduled_tokens=num_scheduled_tokens, |
| 2728 | num_computed_tokens=num_computed_tokens, |
| 2729 | num_draft_tokens=num_draft_tokens, |
| 2730 | device=device, |
| 2731 | ) |
| 2732 | torch.accelerator.synchronize() |
| 2733 | finally: |
| 2734 | # Reset the lru cache so other tests see the default layout again. |
| 2735 | model_mamba_utils.get_conv_state_layout.cache_clear() |
| 2736 | |
| 2737 | # Validate DS independently against the snapshot; otherwise a shared |
| 2738 | # SD/DS bug would remain invisible. |
| 2739 | ds_conv_sd_layout = ds_conv.permute(0, 2, 1).contiguous() |
| 2740 | torch.testing.assert_close( |
| 2741 | ds_conv_sd_layout, |
| 2742 | expected_conv, |
| 2743 | rtol=0, |
| 2744 | atol=0, |
| 2745 | msg="DS conv copy did not match the untouched source snapshot", |
| 2746 | ) |
| 2747 | torch.testing.assert_close( |
| 2748 | ds_temporal, |
| 2749 | expected_temporal, |
| 2750 | rtol=0, |
| 2751 | atol=0, |
| 2752 | msg="DS temporal copy did not match the untouched source snapshot", |
| 2753 | ) |
| 2754 | |
| 2755 | expected_accepted = 1 if same_physical_block else accept_token_bias + 1 |
| 2756 | expected_accepted_tensor = torch.tensor( |
| 2757 | [expected_accepted], dtype=torch.int32, device=device |
| 2758 | ) |
| 2759 | torch.testing.assert_close( |
| 2760 | gpu_ctx_sd.num_accepted_tokens_out[:num_reqs], |
| 2761 | expected_accepted_tensor, |
| 2762 | rtol=0, |
| 2763 | atol=0, |
| 2764 | msg="SD num_accepted_tokens result is wrong", |
| 2765 | ) |
| 2766 | torch.testing.assert_close( |
| 2767 | gpu_ctx_ds.num_accepted_tokens_out[:num_reqs], |
| 2768 | expected_accepted_tensor, |
| 2769 | rtol=0, |
| 2770 | atol=0, |
| 2771 | msg="DS num_accepted_tokens result is wrong", |
| 2772 | ) |