| 1 | # SPDX-License-Identifier: Apache-2.0 |
| 2 | # SPDX-FileCopyrightText: Copyright contributors to the vLLM project |
| 3 | |
| 4 | from dataclasses import fields |
| 5 | from types import SimpleNamespace |
| 6 | from unittest.mock import Mock, patch |
| 7 | |
| 8 | import pytest |
| 9 | import torch |
| 10 | |
| 11 | from tests.v1.attention.utils import ( |
| 12 | BatchSpec, |
| 13 | create_common_attn_metadata, |
| 14 | create_vllm_config, |
| 15 | ) |
| 16 | from vllm.config import SpeculativeConfig |
| 17 | from vllm.config.compilation import CUDAGraphMode |
| 18 | from vllm.models.kimi_k3.nvidia.kda_metadata import ( |
| 19 | KimiK3KDAAttentionBackend, |
| 20 | KimiK3KDAMetadata, |
| 21 | KimiK3KDAMetadataBuilder, |
| 22 | _mamba_get_block_table_tensor, |
| 23 | stage_spec_decode_metadata, |
| 24 | ) |
| 25 | from vllm.models.kimi_k3.nvidia.model import KimiLinearForCausalLM |
| 26 | from vllm.v1.attention.backend import AttentionMetadataBuilder |
| 27 | from vllm.v1.attention.backends.gdn_attn import ( |
| 28 | GDNAttentionBackend, |
| 29 | GDNAttentionMetadata, |
| 30 | GDNAttentionMetadataBuilder, |
| 31 | ) |
| 32 | from vllm.v1.attention.backends.recoverssm_metadata import ( |
| 33 | RecoverSSMPostprocessMetadata, |
| 34 | ) |
| 35 | from vllm.v1.attention.backends.registry import MambaAttentionBackendEnum |
| 36 | from vllm.v1.attention.backends.utils import ( |
| 37 | NULL_BLOCK_ID, |
| 38 | mamba_get_block_table_tensor, |
| 39 | ) |
| 40 | from vllm.v1.kv_cache_interface import MambaSpec |
| 41 | from vllm.v1.worker.mamba_utils import validate_mamba_state_copy_funcs |
| 42 | |
| 43 | BLOCK_SIZE = 16 |
| 44 | DEVICE = torch.device("cpu") |
| 45 | PRUNED_METADATA_FIELDS = { |
| 46 | "chunk_indices", |
| 47 | "chunk_offsets", |
| 48 | "prefill_query_start_loc", |
| 49 | "prefill_state_indices", |
| 50 | "prefill_has_initial_state", |
| 51 | "spec_sequence_masks", |
| 52 | "spec_sequence_masks_cpu", |
| 53 | "uniform_spec_sequence_length", |
| 54 | "flashinfer_prefill_query_start_loc", |
| 55 | "flashinfer_prefill_seq_order", |
| 56 | } |
| 57 | |
| 58 | |
| 59 | def _assert_matches_shared_gdn( |
| 60 | reference: GDNAttentionMetadata, actual: KimiK3KDAMetadata |
| 61 | ): |
| 62 | assert actual.recoverssm_commit is None |
| 63 | assert actual.recoverssm_context is None |
| 64 | for field in fields(GDNAttentionMetadata): |
| 65 | actual_value = getattr(actual, field.name) |
| 66 | if field.name in PRUNED_METADATA_FIELDS: |
| 67 | assert actual_value is None |
| 68 | continue |
| 69 | expected_value = getattr(reference, field.name) |
| 70 | if ( |
| 71 | field.name in {"spec_token_indx", "non_spec_token_indx"} |
| 72 | and actual.num_spec_decodes > 0 |
| 73 | and actual.num_prefills == 0 |
| 74 | and actual.num_decodes == 0 |
| 75 | ): |
| 76 | assert actual_value is None |
| 77 | continue |
| 78 | if isinstance(actual_value, torch.Tensor): |
| 79 | torch.testing.assert_close(actual_value, expected_value) |
| 80 | elif field.name == "nums_dict": |
| 81 | assert (actual_value is None) == (expected_value is None) |
| 82 | if actual_value is not None: |
| 83 | assert actual_value[8]["tot"] == expected_value[8]["tot"] |
| 84 | torch.testing.assert_close( |
| 85 | actual_value[8]["nums"], expected_value[8]["nums"] |
| 86 | ) |
| 87 | else: |
| 88 | assert actual_value == expected_value |
| 89 | |
| 90 | |
| 91 | def _make_builder( |
| 92 | builder_cls: type[AttentionMetadataBuilder], |
| 93 | num_speculative_tokens: int, |
| 94 | full_cuda_graph: bool, |
| 95 | device: torch.device = DEVICE, |
| 96 | mamba_cache_mode: str = "none", |
| 97 | use_recoverssm: bool = False, |
| 98 | num_prefill_checkpoint_blocks: int = 0, |
| 99 | mamba_block_size: int = BLOCK_SIZE, |
| 100 | prefix_match_unit: int | None = None, |
| 101 | use_eagle: bool = False, |
| 102 | disable_eagle_block_drop: bool = False, |
| 103 | ) -> AttentionMetadataBuilder: |
| 104 | vllm_config = create_vllm_config( |
| 105 | model_name="Qwen/Qwen3.5-0.8B", |
| 106 | block_size=BLOCK_SIZE, |
| 107 | ) |
| 108 | if num_speculative_tokens: |
| 109 | vllm_config.speculative_config = SpeculativeConfig( |
| 110 | method="ngram", |
| 111 | num_speculative_tokens=num_speculative_tokens, |
| 112 | ) |
| 113 | if use_eagle: |
| 114 | vllm_config.speculative_config.method = "eagle3" |
| 115 | vllm_config.speculative_config.disable_eagle_block_drop = ( |
| 116 | disable_eagle_block_drop |
| 117 | ) |
| 118 | vllm_config.compilation_config.cudagraph_mode = ( |
| 119 | CUDAGraphMode.FULL_AND_PIECEWISE if full_cuda_graph else CUDAGraphMode.NONE |
| 120 | ) |
| 121 | vllm_config.cache_config.mamba_cache_mode = mamba_cache_mode |
| 122 | vllm_config.cache_config.use_replayssm = use_recoverssm |
| 123 | vllm_config.cache_config.use_kda_recoverssm = use_recoverssm |
| 124 | vllm_config.cache_config.prefix_match_unit = prefix_match_unit |
| 125 | builder = builder_cls( |
| 126 | kv_cache_spec=MambaSpec( |
| 127 | block_size=mamba_block_size, |
| 128 | shapes=((16, 64),), |
| 129 | dtypes=(torch.float16,), |
| 130 | mamba_cache_mode=mamba_cache_mode, |
| 131 | num_speculative_blocks=(0 if use_recoverssm else num_speculative_tokens), |
| 132 | num_prefill_checkpoint_blocks=num_prefill_checkpoint_blocks, |
| 133 | prefill_checkpoint_alignment=( |
| 134 | 16 if num_prefill_checkpoint_blocks > 0 else None |
| 135 | ), |
| 136 | ), |
| 137 | layer_names=["layer.0"], |
| 138 | vllm_config=vllm_config, |
| 139 | device=device, |
| 140 | ) |
| 141 | if use_recoverssm: |
| 142 | assert isinstance(builder, KimiK3KDAMetadataBuilder) |
| 143 | builder.recoverssm_context = Mock() |
| 144 | return builder |
| 145 | |
| 146 | |
| 147 | def test_kda_recoverssm_startup_metadata_flow_without_model(monkeypatch): |
| 148 | """Exercise KDA RecoverSSM startup without loading Kimi-K3 weights.""" |
| 149 | monkeypatch.setattr("vllm.utils.torch_utils.PIN_MEMORY", False) |
| 150 | monkeypatch.setattr("vllm.v1.attention.backends.utils.PIN_MEMORY", False) |
| 151 | layout_config = SimpleNamespace( |
| 152 | model_config=SimpleNamespace( |
| 153 | dtype=torch.bfloat16, |
| 154 | hf_config=SimpleNamespace( |
| 155 | linear_attn_config={ |
| 156 | "num_heads": 4, |
| 157 | "head_dim": 32, |
| 158 | "short_conv_kernel_size": 4, |
| 159 | } |
| 160 | ), |
| 161 | ), |
| 162 | cache_config=SimpleNamespace( |
| 163 | mamba_cache_dtype="auto", |
| 164 | mamba_ssm_cache_dtype="auto", |
| 165 | use_kda_recoverssm=True, |
| 166 | ), |
| 167 | parallel_config=SimpleNamespace(tensor_parallel_size=1), |
| 168 | speculative_config=SimpleNamespace(num_speculative_tokens=2), |
| 169 | ) |
| 170 | kv_cache_spec = MambaSpec( |
| 171 | block_size=BLOCK_SIZE, |
| 172 | shapes=KimiLinearForCausalLM.get_mamba_state_shape_from_config(layout_config), |
| 173 | dtypes=KimiLinearForCausalLM.get_mamba_state_dtype_from_config(layout_config), |
| 174 | mamba_type=MambaAttentionBackendEnum.GDN_ATTN, |
| 175 | mamba_cache_mode="align", |
| 176 | num_prefill_checkpoint_blocks=1, |
| 177 | prefill_checkpoint_alignment=16, |
| 178 | ) |
| 179 | |
| 180 | # This is the same compatibility check performed while initializing the |
| 181 | # model runner. RecoverSSM adds two workspace states that are not copied. |
| 182 | validate_mamba_state_copy_funcs( |
| 183 | {kv_cache_spec: [0]}, |
| 184 | { |
| 185 | MambaAttentionBackendEnum.GDN_ATTN: ( |
| 186 | KimiLinearForCausalLM.get_mamba_state_copy_func() |
| 187 | ) |
| 188 | }, |
| 189 | ) |
| 190 | assert len(kv_cache_spec.shapes) == 4 |
| 191 | |
| 192 | builder_config = SimpleNamespace( |
| 193 | model_config=SimpleNamespace( |
| 194 | hf_text_config=SimpleNamespace(linear_key_head_dim=32) |
| 195 | ), |
| 196 | cache_config=SimpleNamespace( |
| 197 | mamba_cache_mode="align", |
| 198 | use_kda_recoverssm=True, |
| 199 | prefix_match_unit=None, |
| 200 | ), |
| 201 | parallel_config=SimpleNamespace(decode_context_parallel_size=1), |
| 202 | speculative_config=SimpleNamespace( |
| 203 | num_speculative_tokens=2, |
| 204 | parallel_drafting=False, |
| 205 | use_eagle_block_drop=Mock(return_value=False), |
| 206 | ), |
| 207 | compilation_config=SimpleNamespace( |
| 208 | cudagraph_mode=CUDAGraphMode.NONE, |
| 209 | max_cudagraph_capture_size=None, |
| 210 | static_forward_context={}, |
| 211 | ), |
| 212 | scheduler_config=SimpleNamespace(max_num_seqs=4), |
| 213 | additional_config={}, |
| 214 | num_speculative_tokens=2, |
| 215 | use_v2_model_runner=False, |
| 216 | ) |
| 217 | builder = KimiK3KDAMetadataBuilder( |
| 218 | kv_cache_spec=kv_cache_spec, |
| 219 | layer_names=["layer.0"], |
| 220 | vllm_config=builder_config, |
| 221 | device=DEVICE, |
| 222 | ) |
| 223 | |
| 224 | # An all-prefill speculative batch used to leave an all-false spec mask |
| 225 | # alive, then access active_non_spec_mask_cpu before it was initialized. |
| 226 | prefill_common = create_common_attn_metadata( |
| 227 | BatchSpec(seq_lens=[50], query_lens=[50]), |
| 228 | BLOCK_SIZE, |
| 229 | DEVICE, |
| 230 | arange_block_indices=True, |
| 231 | ).replace(is_prefilling=torch.tensor([True])) |
| 232 | builder.mamba_aligned_state_indices = mamba_get_block_table_tensor( |
| 233 | prefill_common.block_table_tensor, |
| 234 | prefill_common.seq_lens, |
| 235 | kv_cache_spec, |
| 236 | "align", |
| 237 | ) |
| 238 | prefill_metadata = builder.build( |
| 239 | 0, |
| 240 | prefill_common, |
| 241 | num_decode_draft_tokens_cpu=torch.tensor([-1], dtype=torch.int32), |
| 242 | num_accepted_tokens=torch.ones(1, dtype=torch.int32), |
| 243 | ) |
| 244 | assert prefill_metadata.num_prefills == 1 |
| 245 | assert prefill_metadata.num_spec_decodes == 0 |
| 246 | assert prefill_metadata.checkpoint is not None |
| 247 | |
| 248 | # Creating the commit context is the first metadata path that consumes the |
| 249 | # builder's layer_names. Mock only the GPU-cache-dependent constructor. |
| 250 | builder.recoverssm_context = None |
| 251 | forward_layer = Mock() |
| 252 | builder.vllm_config.compilation_config.static_forward_context["layer.0"] = ( |
| 253 | forward_layer |
| 254 | ) |
| 255 | spec_common = create_common_attn_metadata( |
| 256 | BatchSpec(seq_lens=[20], query_lens=[3]), |
| 257 | BLOCK_SIZE, |
| 258 | DEVICE, |
| 259 | arange_block_indices=True, |
| 260 | ).replace(is_prefilling=torch.tensor([False])) |
| 261 | builder.mamba_aligned_state_indices = mamba_get_block_table_tensor( |
| 262 | spec_common.block_table_tensor, |
| 263 | spec_common.seq_lens, |
| 264 | kv_cache_spec, |
| 265 | "align", |
| 266 | ) |
| 267 | recoverssm_context = Mock() |
| 268 | with patch( |
| 269 | "vllm.models.kimi_k3.nvidia.ops.recoverssm.KDARecoverSSMCommitContext.create", |
| 270 | return_value=recoverssm_context, |
| 271 | ) as create_context: |
| 272 | spec_metadata = builder.build( |
| 273 | 0, |
| 274 | spec_common, |
| 275 | num_decode_draft_tokens_cpu=torch.tensor([2], dtype=torch.int32), |
| 276 | num_accepted_tokens=torch.ones(1, dtype=torch.int32), |
| 277 | ) |
| 278 | |
| 279 | assert spec_metadata.num_spec_decodes == 1 |
| 280 | assert spec_metadata.recoverssm_commit is not None |
| 281 | assert spec_metadata.recoverssm_context is recoverssm_context |
| 282 | create_context.assert_called_once_with( |
| 283 | [forward_layer], |
| 284 | spec_query_len=3, |
| 285 | max_num_reqs=builder.vllm_config.scheduler_config.max_num_seqs, |
| 286 | ) |
| 287 | |
| 288 | |
| 289 | @pytest.mark.skipif(not torch.cuda.is_available(), reason="requires CUDA") |
| 290 | def test_internal_checkpoint_metadata_targets_last_aligned_boundary(): |
| 291 | device = torch.device("cuda") |
| 292 | batch = BatchSpec(seq_lens=[50, 32], query_lens=[50, 16]) |
| 293 | common_attn_metadata = create_common_attn_metadata( |
| 294 | batch, BLOCK_SIZE, device, arange_block_indices=True |
| 295 | ).replace(is_prefilling=torch.tensor([True, True])) |
| 296 | builder = _make_builder( |
| 297 | KimiK3KDAMetadataBuilder, |
| 298 | num_speculative_tokens=0, |
| 299 | full_cuda_graph=False, |
| 300 | mamba_cache_mode="align", |
| 301 | num_prefill_checkpoint_blocks=1, |
| 302 | device=device, |
| 303 | ) |
| 304 | assert isinstance(builder, KimiK3KDAMetadataBuilder) |
| 305 | builder.mamba_aligned_state_indices = mamba_get_block_table_tensor( |
| 306 | common_attn_metadata.block_table_tensor, |
| 307 | common_attn_metadata.seq_lens, |
| 308 | builder.kv_cache_spec, |
| 309 | "align", |
| 310 | ) |
| 311 | actual = builder.build(0, common_attn_metadata) |
| 312 | |
| 313 | assert actual.checkpoint is not None |
| 314 | torch.testing.assert_close( |
| 315 | actual.checkpoint.state_indices, |
| 316 | torch.tensor([2, NULL_BLOCK_ID], dtype=torch.int32, device=device), |
| 317 | ) |
| 318 | torch.testing.assert_close( |
| 319 | actual.checkpoint.checkpoint_offsets, |
| 320 | torch.tensor([48, 0], dtype=torch.int32, device=device), |
| 321 | ) |
| 322 | |
| 323 | |
| 324 | @pytest.mark.parametrize( |
| 325 | ("disable_eagle_block_drop", "prefix_match_unit", "expected_offset"), |
| 326 | [(False, 16, 80), (True, 16, 96), (False, 8, None)], |
| 327 | ) |
| 328 | @pytest.mark.skipif(not torch.cuda.is_available(), reason="requires CUDA") |
| 329 | def test_spec_internal_checkpoint_metadata_targets_replay_boundary( |
| 330 | disable_eagle_block_drop: bool, |
| 331 | prefix_match_unit: int, |
| 332 | expected_offset: int | None, |
| 333 | ) -> None: |
| 334 | device = torch.device("cuda") |
| 335 | batch = BatchSpec(seq_lens=[100], query_lens=[100]) |
| 336 | common_attn_metadata = create_common_attn_metadata( |
| 337 | batch, BLOCK_SIZE, device, arange_block_indices=True |
| 338 | ) |
| 339 | common_attn_metadata = common_attn_metadata.replace( |
| 340 | is_prefilling=torch.tensor([True]), |
| 341 | block_table_tensor=common_attn_metadata.block_table_tensor + 1, |
| 342 | ) |
| 343 | builder = _make_builder( |
| 344 | KimiK3KDAMetadataBuilder, |
| 345 | num_speculative_tokens=3, |
| 346 | full_cuda_graph=False, |
| 347 | mamba_cache_mode="align", |
| 348 | num_prefill_checkpoint_blocks=1, |
| 349 | mamba_block_size=64, |
| 350 | prefix_match_unit=prefix_match_unit, |
| 351 | use_eagle=True, |
| 352 | disable_eagle_block_drop=disable_eagle_block_drop, |
| 353 | device=device, |
| 354 | ) |
| 355 | assert isinstance(builder, KimiK3KDAMetadataBuilder) |
| 356 | builder.mamba_aligned_state_indices = mamba_get_block_table_tensor( |
| 357 | common_attn_metadata.block_table_tensor, |
| 358 | common_attn_metadata.seq_lens, |
| 359 | builder.kv_cache_spec, |
| 360 | "align", |
| 361 | ) |
| 362 | actual = builder.build(0, common_attn_metadata) |
| 363 | |
| 364 | if expected_offset is None: |
| 365 | assert actual.checkpoint is None |
| 366 | return |
| 367 | assert actual.checkpoint is not None |
| 368 | torch.testing.assert_close( |
| 369 | actual.checkpoint.state_indices, |
| 370 | torch.tensor([1], dtype=torch.int32, device=device), |
| 371 | ) |
| 372 | torch.testing.assert_close( |
| 373 | actual.checkpoint.checkpoint_offsets, |
| 374 | torch.tensor([expected_offset], dtype=torch.int32, device=device), |
| 375 | ) |
| 376 | |
| 377 | |
| 378 | @pytest.mark.skipif(not torch.cuda.is_available(), reason="requires CUDA") |
| 379 | def test_internal_checkpoint_metadata_skips_unaligned_offset(): |
| 380 | device = torch.device("cuda") |
| 381 | # The checkpoint block boundary is 48, but this query starts at token 1, |
| 382 | # making the real checkpoint offset 47, which is not FlashKDA-aligned. |
| 383 | batch = BatchSpec(seq_lens=[50], query_lens=[49]) |
| 384 | common_attn_metadata = create_common_attn_metadata( |
| 385 | batch, BLOCK_SIZE, device, arange_block_indices=True |
| 386 | ).replace(is_prefilling=torch.tensor([True])) |
| 387 | builder = _make_builder( |
| 388 | KimiK3KDAMetadataBuilder, |
| 389 | num_speculative_tokens=0, |
| 390 | full_cuda_graph=False, |
| 391 | mamba_cache_mode="align", |
| 392 | num_prefill_checkpoint_blocks=1, |
| 393 | device=device, |
| 394 | ) |
| 395 | assert isinstance(builder, KimiK3KDAMetadataBuilder) |
| 396 | builder.mamba_aligned_state_indices = mamba_get_block_table_tensor( |
| 397 | common_attn_metadata.block_table_tensor, |
| 398 | common_attn_metadata.seq_lens, |
| 399 | builder.kv_cache_spec, |
| 400 | "align", |
| 401 | ) |
| 402 | actual = builder.build(0, common_attn_metadata) |
| 403 | |
| 404 | assert actual.checkpoint is None |
| 405 | |
| 406 | |
| 407 | @pytest.mark.parametrize( |
| 408 | ( |
| 409 | "batch", |
| 410 | "num_decode_draft_tokens", |
| 411 | "num_speculative_tokens", |
| 412 | "full_cuda_graph", |
| 413 | "is_prefilling", |
| 414 | "expected_uniform_spec_sequence_length", |
| 415 | ), |
| 416 | [ |
| 417 | pytest.param( |
| 418 | BatchSpec(seq_lens=[50, 30], query_lens=[3, 3]), |
| 419 | [2, 2], |
| 420 | 2, |
| 421 | False, |
| 422 | [False, False], |
| 423 | 3, |
| 424 | id="pure-spec-decode", |
| 425 | ), |
| 426 | pytest.param( |
| 427 | BatchSpec(seq_lens=[100, 65, 20], query_lens=[50, 1, 3]), |
| 428 | [-1, -1, 2], |
| 429 | 2, |
| 430 | False, |
| 431 | [True, False, False], |
| 432 | 3, |
| 433 | id="mixed-prefill-and-spec-decode", |
| 434 | ), |
| 435 | pytest.param( |
| 436 | BatchSpec(seq_lens=[40, 30], query_lens=[1, 1]), |
| 437 | None, |
| 438 | 0, |
| 439 | False, |
| 440 | [False, False], |
| 441 | None, |
| 442 | id="regular-decode", |
| 443 | ), |
| 444 | pytest.param( |
| 445 | BatchSpec(seq_lens=[40, 30], query_lens=[1, 1]), |
| 446 | [0, 0], |
| 447 | 2, |
| 448 | False, |
| 449 | [False, False], |
| 450 | None, |
| 451 | id="no-scheduled-draft-tokens", |
| 452 | ), |
| 453 | pytest.param( |
| 454 | BatchSpec(seq_lens=[50, 30], query_lens=[3, 2]), |
| 455 | [2, 1], |
| 456 | 2, |
| 457 | False, |
| 458 | [False, False], |
| 459 | None, |
| 460 | id="ragged-spec-decode", |
| 461 | ), |
| 462 | ], |
| 463 | ) |
| 464 | def test_kimi_k3_kda_metadata_matches_shared_gdn( |
| 465 | batch: BatchSpec, |
| 466 | num_decode_draft_tokens: list[int] | None, |
| 467 | num_speculative_tokens: int, |
| 468 | full_cuda_graph: bool, |
| 469 | is_prefilling: list[bool], |
| 470 | expected_uniform_spec_sequence_length: int | None, |
| 471 | ): |
| 472 | kwargs: dict[str, torch.Tensor] = {} |
| 473 | if num_decode_draft_tokens is not None: |
| 474 | kwargs = { |
| 475 | "num_decode_draft_tokens_cpu": torch.tensor( |
| 476 | num_decode_draft_tokens, dtype=torch.int32 |
| 477 | ), |
| 478 | "num_accepted_tokens": torch.ones( |
| 479 | batch.batch_size, dtype=torch.int32, device=DEVICE |
| 480 | ), |
| 481 | } |
| 482 | |
| 483 | common_attn_metadata = create_common_attn_metadata( |
| 484 | batch, BLOCK_SIZE, DEVICE |
| 485 | ).replace(is_prefilling=torch.tensor(is_prefilling, dtype=torch.bool)) |
| 486 | reference = _make_builder( |
| 487 | GDNAttentionMetadataBuilder, |
| 488 | num_speculative_tokens, |
| 489 | full_cuda_graph, |
| 490 | ).build( |
| 491 | 0, |
| 492 | common_attn_metadata, |
| 493 | **kwargs, |
| 494 | ) |
| 495 | actual = _make_builder( |
| 496 | KimiK3KDAMetadataBuilder, |
| 497 | num_speculative_tokens, |
| 498 | full_cuda_graph, |
| 499 | ).build(0, common_attn_metadata, **kwargs) |
| 500 | |
| 501 | assert isinstance(actual, KimiK3KDAMetadata) |
| 502 | _assert_matches_shared_gdn(reference, actual) |
| 503 | assert ( |
| 504 | reference.uniform_spec_sequence_length == expected_uniform_spec_sequence_length |
| 505 | ) |
| 506 | |
| 507 | |
| 508 | def test_mixed_regular_and_spec_decode_uses_packed_decode_metadata(): |
| 509 | batch = BatchSpec(seq_lens=[100, 65, 20], query_lens=[1, 1, 3]) |
| 510 | common_attn_metadata = create_common_attn_metadata( |
| 511 | batch, BLOCK_SIZE, DEVICE |
| 512 | ).replace(is_prefilling=torch.tensor([False, False, False])) |
| 513 | actual = _make_builder( |
| 514 | KimiK3KDAMetadataBuilder, |
| 515 | num_speculative_tokens=2, |
| 516 | full_cuda_graph=False, |
| 517 | ).build( |
| 518 | 0, |
| 519 | common_attn_metadata, |
| 520 | num_decode_draft_tokens_cpu=torch.tensor([-1, -1, 2], dtype=torch.int32), |
| 521 | num_accepted_tokens=torch.ones(3, dtype=torch.int32, device=DEVICE), |
| 522 | ) |
| 523 | |
| 524 | # The K3 layer dispatches the non-spec subgroup to packed decode whenever |
| 525 | # it contains no prefill request. |
| 526 | assert actual.num_decodes == 2 |
| 527 | assert actual.num_decode_tokens == 2 |
| 528 | assert actual.num_prefills == 0 |
| 529 | assert actual.num_prefill_tokens == 0 |
| 530 | assert actual.has_initial_state is None |
| 531 | assert actual.nums_dict is None |
| 532 | assert actual.non_spec_query_start_loc is None |
| 533 | torch.testing.assert_close(actual.non_spec_token_indx, torch.tensor([0, 1])) |
| 534 | torch.testing.assert_close(actual.spec_token_indx, torch.tensor([2, 3, 4])) |
| 535 | assert actual.non_spec_token_start == 0 |
| 536 | assert actual.spec_token_start == 2 |
| 537 | torch.testing.assert_close( |
| 538 | actual.spec_query_start_loc, |
| 539 | torch.tensor([0, 3], dtype=torch.int32), |
| 540 | ) |
| 541 | |
| 542 | |
| 543 | def test_mixed_regular_and_spec_decode_excludes_request_padding(): |
| 544 | batch = BatchSpec(seq_lens=[16, 65, 20], query_lens=[0, 1, 3]) |
| 545 | common_attn_metadata = create_common_attn_metadata( |
| 546 | batch, BLOCK_SIZE, DEVICE |
| 547 | ).replace(is_prefilling=torch.tensor([False, False, False])) |
| 548 | actual = _make_builder( |
| 549 | KimiK3KDAMetadataBuilder, |
| 550 | num_speculative_tokens=2, |
| 551 | full_cuda_graph=False, |
| 552 | ).build( |
| 553 | 0, |
| 554 | common_attn_metadata, |
| 555 | num_decode_draft_tokens_cpu=torch.tensor([-1, -1, 2], dtype=torch.int32), |
| 556 | num_accepted_tokens=torch.ones(3, dtype=torch.int32, device=DEVICE), |
| 557 | ) |
| 558 | |
| 559 | assert actual.num_decodes == 1 |
| 560 | assert actual.non_spec_state_indices_tensor is not None |
| 561 | assert actual.non_spec_state_indices_tensor.shape == (1,) |
| 562 | torch.testing.assert_close(actual.non_spec_token_indx, torch.tensor([0])) |
| 563 | torch.testing.assert_close(actual.spec_token_indx, torch.tensor([1, 2, 3])) |
| 564 | assert actual.non_spec_token_start == 0 |
| 565 | assert actual.spec_token_start == 1 |
| 566 | |
| 567 | |
| 568 | @pytest.mark.parametrize("mamba_cache_mode", ["none", "align"]) |
| 569 | def test_recoverssm_spec_uses_one_state_slot_and_current_window( |
| 570 | mamba_cache_mode: str, |
| 571 | ): |
| 572 | if mamba_cache_mode == "align" and not torch.cuda.is_available(): |
| 573 | pytest.skip("align metadata construction requires CUDA") |
| 574 | device = torch.device("cuda") if mamba_cache_mode == "align" else DEVICE |
| 575 | batch = BatchSpec(seq_lens=[100, 65, 20], query_lens=[1, 1, 3]) |
| 576 | common_attn_metadata = create_common_attn_metadata( |
| 577 | batch, BLOCK_SIZE, device |
| 578 | ).replace(is_prefilling=torch.tensor([True, True, False])) |
| 579 | builder = _make_builder( |
| 580 | KimiK3KDAMetadataBuilder, |
| 581 | num_speculative_tokens=2, |
| 582 | full_cuda_graph=False, |
| 583 | device=device, |
| 584 | mamba_cache_mode=mamba_cache_mode, |
| 585 | use_recoverssm=True, |
| 586 | ) |
| 587 | assert isinstance(builder, KimiK3KDAMetadataBuilder) |
| 588 | if mamba_cache_mode == "align": |
| 589 | builder.mamba_aligned_state_indices = mamba_get_block_table_tensor( |
| 590 | common_attn_metadata.block_table_tensor, |
| 591 | common_attn_metadata.seq_lens, |
| 592 | builder.kv_cache_spec, |
| 593 | mamba_cache_mode, |
| 594 | ) |
| 595 | context = builder.recoverssm_context |
| 596 | assert context is not None |
| 597 | actual = builder.build( |
| 598 | 0, |
| 599 | common_attn_metadata, |
| 600 | num_decode_draft_tokens_cpu=torch.tensor([-1, -1, 2], dtype=torch.int32), |
| 601 | num_accepted_tokens=torch.tensor([3, 2, 2], dtype=torch.int32, device=device), |
| 602 | ) |
| 603 | |
| 604 | assert actual.spec_state_indices_tensor is not None |
| 605 | assert actual.spec_state_indices_tensor.shape == (1, 1) |
| 606 | torch.testing.assert_close( |
| 607 | actual.num_accepted_tokens, |
| 608 | torch.ones(1, dtype=torch.int32, device=device), |
| 609 | ) |
| 610 | commit_metadata = actual.recoverssm_commit |
| 611 | assert commit_metadata is not None |
| 612 | torch.testing.assert_close( |
| 613 | commit_metadata.request_indices, |
| 614 | torch.tensor([2], dtype=torch.int32, device=device), |
| 615 | ) |
| 616 | assert actual.recoverssm_context is context |
| 617 | num_accepted_tokens = torch.tensor([3, 2, 1], dtype=torch.int32, device=device) |
| 618 | |
| 619 | postprocess = actual.commit_recoverssm_state(num_accepted_tokens) |
| 620 | |
| 621 | if mamba_cache_mode == "none": |
| 622 | assert commit_metadata.align is None |
| 623 | assert postprocess is None |
| 624 | else: |
| 625 | assert isinstance(postprocess, RecoverSSMPostprocessMetadata) |
| 626 | assert postprocess.num_spec_decodes == 1 |
| 627 | assert postprocess.request_indices is commit_metadata.request_indices |
| 628 | assert postprocess.block_table is common_attn_metadata.block_table_tensor |
| 629 | assert ( |
| 630 | postprocess.num_computed_tokens |
| 631 | is common_attn_metadata.compute_num_computed_tokens() |
| 632 | ) |
| 633 | assert postprocess.block_size == BLOCK_SIZE |
| 634 | args = context.commit.call_args.args |
| 635 | assert args[0] is num_accepted_tokens |
| 636 | torch.testing.assert_close(args[1], commit_metadata.state_indices[:, 0]) |
| 637 | torch.testing.assert_close(args[2], commit_metadata.query_start_loc) |
| 638 | |
| 639 | |
| 640 | def test_recoverssm_distinguishes_draftless_decode_from_one_token_prefill(): |
| 641 | batch = BatchSpec(seq_lens=[40, 30], query_lens=[1, 1]) |
| 642 | common_attn_metadata = create_common_attn_metadata( |
| 643 | batch, BLOCK_SIZE, DEVICE |
| 644 | ).replace(is_prefilling=torch.tensor([False, True])) |
| 645 | actual = _make_builder( |
| 646 | KimiK3KDAMetadataBuilder, |
| 647 | num_speculative_tokens=2, |
| 648 | full_cuda_graph=False, |
| 649 | use_recoverssm=True, |
| 650 | ).build( |
| 651 | 0, |
| 652 | common_attn_metadata, |
| 653 | num_decode_draft_tokens_cpu=torch.full((2,), -1, dtype=torch.int32), |
| 654 | num_accepted_tokens=torch.ones(2, dtype=torch.int32), |
| 655 | ) |
| 656 | |
| 657 | assert actual.num_spec_decodes == 1 |
| 658 | assert actual.num_decodes == 0 |
| 659 | assert actual.num_prefills == 1 |
| 660 | assert actual.spec_state_indices_tensor is not None |
| 661 | assert actual.spec_state_indices_tensor.shape == (1, 1) |
| 662 | torch.testing.assert_close( |
| 663 | actual.spec_query_start_loc, |
| 664 | torch.tensor([0, 1], dtype=torch.int32), |
| 665 | ) |
| 666 | |
| 667 | |
| 668 | @pytest.mark.parametrize( |
| 669 | ("seq_len", "expected_has_initial_state"), |
| 670 | [ |
| 671 | pytest.param(1, False, id="first-token-prefill"), |
| 672 | pytest.param(65, True, id="final-one-token-prefill-chunk"), |
| 673 | ], |
| 674 | ) |
| 675 | def test_mixed_one_token_prefill_and_spec_decode_uses_prefill_metadata( |
| 676 | seq_len: int, |
| 677 | expected_has_initial_state: bool, |
| 678 | ): |
| 679 | batch = BatchSpec(seq_lens=[seq_len, 20], query_lens=[1, 3]) |
| 680 | common_attn_metadata = create_common_attn_metadata( |
| 681 | batch, BLOCK_SIZE, DEVICE |
| 682 | ).replace(is_prefilling=torch.tensor([True, False])) |
| 683 | actual = _make_builder( |
| 684 | KimiK3KDAMetadataBuilder, |
| 685 | num_speculative_tokens=2, |
| 686 | full_cuda_graph=False, |
| 687 | ).build( |
| 688 | 0, |
| 689 | common_attn_metadata, |
| 690 | num_decode_draft_tokens_cpu=torch.tensor([-1, 2], dtype=torch.int32), |
| 691 | num_accepted_tokens=torch.ones(2, dtype=torch.int32, device=DEVICE), |
| 692 | ) |
| 693 | |
| 694 | assert actual.num_prefills == 1 |
| 695 | assert actual.num_prefill_tokens == 1 |
| 696 | assert actual.num_decodes == 0 |
| 697 | assert actual.num_decode_tokens == 0 |
| 698 | assert actual.has_initial_state is not None |
| 699 | assert actual.has_initial_state.tolist() == [expected_has_initial_state] |
| 700 | assert actual.non_spec_query_start_loc is not None |
| 701 | torch.testing.assert_close( |
| 702 | actual.non_spec_query_start_loc, |
| 703 | torch.tensor([0, 1], dtype=torch.int32), |
| 704 | ) |
| 705 | |
| 706 | |
| 707 | @pytest.mark.skipif(not torch.cuda.is_available(), reason="requires CUDA") |
| 708 | def test_kimi_k3_kda_cudagraph_capture_matches_shared_gdn(): |
| 709 | device = torch.device("cuda") |
| 710 | batch = BatchSpec(seq_lens=[50, 30], query_lens=[3, 3]) |
| 711 | common_attn_metadata = create_common_attn_metadata( |
| 712 | batch, BLOCK_SIZE, device |
| 713 | ).replace(is_prefilling=torch.tensor([False, False])) |
| 714 | reference = _make_builder( |
| 715 | GDNAttentionMetadataBuilder, |
| 716 | num_speculative_tokens=2, |
| 717 | full_cuda_graph=True, |
| 718 | device=device, |
| 719 | ).build_for_cudagraph_capture(common_attn_metadata) |
| 720 | actual = _make_builder( |
| 721 | KimiK3KDAMetadataBuilder, |
| 722 | num_speculative_tokens=2, |
| 723 | full_cuda_graph=True, |
| 724 | device=device, |
| 725 | ).build_for_cudagraph_capture(common_attn_metadata) |
| 726 | |
| 727 | assert isinstance(actual, KimiK3KDAMetadata) |
| 728 | _assert_matches_shared_gdn(reference, actual) |
| 729 | |
| 730 | |
| 731 | @pytest.mark.parametrize( |
| 732 | "builder_cls", |
| 733 | [ |
| 734 | GDNAttentionMetadataBuilder, |
| 735 | # The Kimi-K3 builder stages spec-decode state indices on the device. |
| 736 | pytest.param( |
| 737 | KimiK3KDAMetadataBuilder, |
| 738 | marks=pytest.mark.skipif( |
| 739 | not torch.cuda.is_available(), reason="requires CUDA" |
| 740 | ), |
| 741 | ), |
| 742 | ], |
| 743 | ) |
| 744 | def test_cudagraph_capture_metadata_avoids_device_to_host_copy( |
| 745 | builder_cls: type[AttentionMetadataBuilder], monkeypatch: pytest.MonkeyPatch |
| 746 | ): |
| 747 | """`build_for_cudagraph_capture` is not capture-only: a DP rank with nothing |
| 748 | scheduled re-stages its FULL-graph metadata through it on every dummy step, |
| 749 | so it must not synchronize the device. The host-side draft counts have to |
| 750 | come from `query_start_loc_cpu` and match the device-derived values.""" |
| 751 | num_speculative_tokens = 2 |
| 752 | batch = BatchSpec(seq_lens=[50, 30], query_lens=[3, 3]) |
| 753 | device = torch.device("cuda" if torch.cuda.is_available() else "cpu") |
| 754 | common_attn_metadata = create_common_attn_metadata( |
| 755 | batch, BLOCK_SIZE, device |
| 756 | ).replace(is_prefilling=torch.zeros(batch.batch_size, dtype=torch.bool)) |
| 757 | builder = _make_builder( |
| 758 | builder_cls, num_speculative_tokens, full_cuda_graph=True, device=device |
| 759 | ) |
| 760 | |
| 761 | device_to_host_copies: list[torch.Size] = [] |
| 762 | original_cpu = torch.Tensor.cpu |
| 763 | |
| 764 | def counting_cpu(self: torch.Tensor, *args, **kwargs): |
| 765 | device_to_host_copies.append(self.shape) |
| 766 | return original_cpu(self, *args, **kwargs) |
| 767 | |
| 768 | monkeypatch.setattr(torch.Tensor, "cpu", counting_cpu) |
| 769 | actual = builder.build_for_cudagraph_capture(common_attn_metadata) |
| 770 | monkeypatch.undo() |
| 771 | |
| 772 | assert not device_to_host_copies, device_to_host_copies |
| 773 | |
| 774 | num_accepted_tokens = torch.diff(common_attn_metadata.query_start_loc) |
| 775 | reference = builder.build( |
| 776 | 0, |
| 777 | common_attn_metadata, |
| 778 | num_accepted_tokens, |
| 779 | (num_accepted_tokens - 1).cpu(), |
| 780 | ) |
| 781 | assert actual.num_spec_decodes == batch.batch_size |
| 782 | assert actual.num_decodes == 0 |
| 783 | assert actual.num_prefills == 0 |
| 784 | for field in fields(type(actual)): |
| 785 | actual_value = getattr(actual, field.name) |
| 786 | reference_value = getattr(reference, field.name) |
| 787 | if isinstance(actual_value, torch.Tensor): |
| 788 | torch.testing.assert_close(actual_value, reference_value) |
| 789 | else: |
| 790 | assert actual_value == reference_value, field.name |
| 791 | |
| 792 | |
| 793 | @pytest.mark.skipif(not torch.cuda.is_available(), reason="requires CUDA") |
| 794 | def test_recoverssm_spec_cudagraph_stages_one_checkpoint_per_request(): |
| 795 | device = torch.device("cuda") |
| 796 | batch = BatchSpec(seq_lens=[50, 30], query_lens=[3, 3]) |
| 797 | common_attn_metadata = create_common_attn_metadata( |
| 798 | batch, BLOCK_SIZE, device |
| 799 | ).replace(is_prefilling=torch.tensor([False, False])) |
| 800 | builder = _make_builder( |
| 801 | KimiK3KDAMetadataBuilder, |
| 802 | num_speculative_tokens=2, |
| 803 | full_cuda_graph=True, |
| 804 | device=device, |
| 805 | use_recoverssm=True, |
| 806 | ) |
| 807 | assert isinstance(builder, KimiK3KDAMetadataBuilder) |
| 808 | assert builder.spec_state_indices_tensor.shape == ( |
| 809 | builder.vllm_config.scheduler_config.max_num_seqs, |
| 810 | 1, |
| 811 | ) |
| 812 | actual = builder.build_for_cudagraph_capture(common_attn_metadata) |
| 813 | |
| 814 | assert actual.spec_state_indices_tensor is not None |
| 815 | assert actual.spec_state_indices_tensor.shape == (batch.batch_size, 1) |
| 816 | assert actual.num_accepted_tokens is not None |
| 817 | torch.testing.assert_close( |
| 818 | actual.num_accepted_tokens, |
| 819 | torch.ones(batch.batch_size, dtype=torch.int32, device=device), |
| 820 | ) |
| 821 | assert actual.recoverssm_commit is not None |
| 822 | assert actual.recoverssm_commit.request_indices is None |
| 823 | |
| 824 | |
| 825 | def test_kimi_k3_kda_backend_uses_private_metadata_builder(): |
| 826 | assert KimiK3KDAAttentionBackend.get_builder_cls() is KimiK3KDAMetadataBuilder |
| 827 | assert KimiK3KDAAttentionBackend.is_ssm() |
| 828 | assert issubclass(KimiK3KDAAttentionBackend, GDNAttentionBackend) |
| 829 | assert issubclass(KimiK3KDAMetadata, GDNAttentionMetadata) |
| 830 | assert issubclass(KimiK3KDAMetadataBuilder, GDNAttentionMetadataBuilder) |
| 831 | |
| 832 | |
| 833 | def test_kimi_k3_metadata_uses_precomputed_aligned_state_indices(): |
| 834 | batch = BatchSpec(seq_lens=[40, 30], query_lens=[1, 1]) |
| 835 | common_attn_metadata = create_common_attn_metadata( |
| 836 | batch, BLOCK_SIZE, DEVICE |
| 837 | ).replace(is_prefilling=torch.tensor([False, False])) |
| 838 | builder = _make_builder( |
| 839 | KimiK3KDAMetadataBuilder, |
| 840 | num_speculative_tokens=2, |
| 841 | full_cuda_graph=False, |
| 842 | mamba_cache_mode="align", |
| 843 | ) |
| 844 | assert isinstance(builder, KimiK3KDAMetadataBuilder) |
| 845 | precomputed_indices = torch.tensor( |
| 846 | [ |
| 847 | [101, 102, 103], |
| 848 | [201, 202, 203], |
| 849 | [301, 302, 303], |
| 850 | ], |
| 851 | dtype=torch.int32, |
| 852 | ) |
| 853 | builder.mamba_aligned_state_indices = precomputed_indices |
| 854 | |
| 855 | metadata = builder.build(0, common_attn_metadata) |
| 856 | |
| 857 | torch.testing.assert_close( |
| 858 | metadata.non_spec_state_indices_tensor, |
| 859 | precomputed_indices[: batch.batch_size, 0], |
| 860 | ) |
| 861 | |
| 862 | |
| 863 | @pytest.mark.skipif(not torch.cuda.is_available(), reason="requires CUDA") |
| 864 | def test_stage_spec_decode_metadata_matches_pytorch(): |
| 865 | device = torch.device("cuda") |
| 866 | num_spec_decodes = 33 |
| 867 | batch_size = 65 |
| 868 | num_state_slots = 3 |
| 869 | state_indices = torch.arange( |
| 870 | num_spec_decodes * 32, |
| 871 | dtype=torch.int32, |
| 872 | device=device, |
| 873 | ).reshape(num_spec_decodes, 32)[:, :num_state_slots] |
| 874 | query_start_loc = ( |
| 875 | torch.arange(num_spec_decodes + 1, dtype=torch.int32, device=device) |
| 876 | * num_state_slots |
| 877 | ) |
| 878 | num_accepted_tokens = ( |
| 879 | torch.arange(num_spec_decodes, dtype=torch.int32, device=device) |
| 880 | % num_state_slots |
| 881 | + 1 |
| 882 | ) |
| 883 | |
| 884 | staged_state_indices = torch.empty( |
| 885 | (batch_size, num_state_slots), dtype=torch.int32, device=device |
| 886 | ) |
| 887 | staged_query_start_loc = torch.empty( |
| 888 | batch_size + 1, dtype=torch.int32, device=device |
| 889 | ) |
| 890 | staged_num_accepted_tokens = torch.empty( |
| 891 | batch_size, dtype=torch.int32, device=device |
| 892 | ) |
| 893 | stage_spec_decode_metadata( |
| 894 | state_indices, |
| 895 | query_start_loc, |
| 896 | num_accepted_tokens, |
| 897 | staged_state_indices, |
| 898 | staged_query_start_loc, |
| 899 | staged_num_accepted_tokens, |
| 900 | num_spec_decodes=num_spec_decodes, |
| 901 | ) |
| 902 | |
| 903 | expected_state_indices = torch.full_like(staged_state_indices, NULL_BLOCK_ID) |
| 904 | expected_state_indices[:num_spec_decodes] = state_indices |
| 905 | expected_query_start_loc = torch.full( |
| 906 | (batch_size + 1,), |
| 907 | query_start_loc[-1], |
| 908 | dtype=torch.int32, |
| 909 | device=device, |
| 910 | ) |
| 911 | expected_query_start_loc[: num_spec_decodes + 1] = query_start_loc |
| 912 | expected_num_accepted_tokens = torch.ones( |
| 913 | batch_size, dtype=torch.int32, device=device |
| 914 | ) |
| 915 | expected_num_accepted_tokens[:num_spec_decodes] = num_accepted_tokens |
| 916 | |
| 917 | torch.testing.assert_close(staged_state_indices, expected_state_indices) |
| 918 | torch.testing.assert_close(staged_query_start_loc, expected_query_start_loc) |
| 919 | torch.testing.assert_close(staged_num_accepted_tokens, expected_num_accepted_tokens) |
| 920 | |
| 921 | |
| 922 | @pytest.mark.skipif(not torch.cuda.is_available(), reason="requires CUDA") |
| 923 | def test_aligned_block_table_matches_shared_gdn(): |
| 924 | device = torch.device("cuda") |
| 925 | seq_lens = torch.tensor( |
| 926 | [0, 1, 15, 16, 17, 31, 32, 33, 511, 512, 513], |
| 927 | dtype=torch.int32, |
| 928 | device=device, |
| 929 | ).repeat(6)[:65] |
| 930 | block_table_storage = torch.arange( |
| 931 | seq_lens.numel() * 128, |
| 932 | dtype=torch.int32, |
| 933 | device=device, |
| 934 | ).reshape(seq_lens.numel(), 128) |
| 935 | block_table = block_table_storage[:, ::2] |
| 936 | kv_cache_spec = MambaSpec( |
| 937 | block_size=BLOCK_SIZE, |
| 938 | shapes=((16, 64),), |
| 939 | dtypes=(torch.float16,), |
| 940 | num_speculative_blocks=2, |
| 941 | ) |
| 942 | |
| 943 | expected = mamba_get_block_table_tensor( |
| 944 | block_table, |
| 945 | seq_lens, |
| 946 | kv_cache_spec, |
| 947 | "align", |
| 948 | ) |
| 949 | actual = _mamba_get_block_table_tensor( |
| 950 | block_table, |
| 951 | seq_lens, |
| 952 | kv_cache_spec, |
| 953 | "align", |
| 954 | ) |
| 955 | |
| 956 | torch.testing.assert_close(actual, expected) |
| 957 | |
| 958 | |
| 959 | def _build_non_spec(batch, is_prefilling, full_cuda_graph=False): |
| 960 | common_attn_metadata = create_common_attn_metadata( |
| 961 | batch, BLOCK_SIZE, DEVICE |
| 962 | ).replace( |
| 963 | is_prefilling=None |
| 964 | if is_prefilling is None |
| 965 | else torch.tensor(is_prefilling, dtype=torch.bool) |
| 966 | ) |
| 967 | builder = _make_builder( |
| 968 | KimiK3KDAMetadataBuilder, |
| 969 | num_speculative_tokens=0, |
| 970 | full_cuda_graph=full_cuda_graph, |
| 971 | ) |
| 972 | return builder, common_attn_metadata, builder.build(0, common_attn_metadata) |
| 973 | |
| 974 | |
| 975 | def test_one_token_first_chunk_excludes_padding(): |
| 976 | """Neither padding requests nor padding tokens count as prefill work.""" |
| 977 | common = create_common_attn_metadata( |
| 978 | BatchSpec(seq_lens=[100, 1, 0, 0], query_lens=[1, 1, 0, 0]), |
| 979 | BLOCK_SIZE, |
| 980 | DEVICE, |
| 981 | ).replace( |
| 982 | is_prefilling=torch.tensor([False, True, False, False], dtype=torch.bool), |
| 983 | num_actual_tokens=4, |
| 984 | ) |
| 985 | builder = _make_builder( |
| 986 | KimiK3KDAMetadataBuilder, num_speculative_tokens=0, full_cuda_graph=False |
| 987 | ) |
| 988 | actual = builder.build(0, common) |
| 989 | |
| 990 | assert actual.num_decodes == 1 |
| 991 | assert actual.num_prefills == 1 |
| 992 | assert actual.num_decode_tokens == 1 |
| 993 | assert actual.num_prefill_tokens == 1 |
| 994 | |
| 995 | |
| 996 | @pytest.mark.parametrize( |
| 997 | ("seq_len", "query_len", "is_prefilling", "num_prefills"), |
| 998 | [ |
| 999 | pytest.param(1, 1, True, 1, id="first-chunk"), |
| 1000 | pytest.param(65, 1, True, 0, id="resumed-chunk"), |
| 1001 | pytest.param(0, 0, True, 0, id="padding"), |
| 1002 | pytest.param(1, 1, None, 0, id="missing-prefill-flag"), |
| 1003 | ], |
| 1004 | ) |
| 1005 | def test_one_token_chunk_classification( |
| 1006 | seq_len, query_len, is_prefilling, num_prefills |
| 1007 | ): |
| 1008 | """Only a real first chunk with a prefill flag needs state initialization.""" |
| 1009 | _, _, actual = _build_non_spec( |
| 1010 | BatchSpec(seq_lens=[100, seq_len], query_lens=[1, query_len]), |
| 1011 | is_prefilling=None if is_prefilling is None else [False, is_prefilling], |
| 1012 | ) |
| 1013 | |
| 1014 | assert actual.num_prefills == num_prefills |
| 1015 | assert actual.num_decodes == 2 - num_prefills |
| 1016 | assert actual.num_prefill_tokens == num_prefills |
| 1017 | assert actual.num_decode_tokens == 1 + query_len - num_prefills |
| 1018 | if num_prefills: |
| 1019 | assert actual.has_initial_state is not None |
| 1020 | assert actual.has_initial_state.tolist() == [True, False] |
| 1021 | else: |
| 1022 | assert actual.has_initial_state is None |
| 1023 | |
| 1024 | |
| 1025 | def test_cudagraph_capture_batch_stays_decode_only(): |
| 1026 | """Capture rows have no history, but must still select decode kernels.""" |
| 1027 | batch = BatchSpec(seq_lens=[1] * 4, query_lens=[1] * 4) |
| 1028 | common_attn_metadata = create_common_attn_metadata( |
| 1029 | batch, BLOCK_SIZE, DEVICE |
| 1030 | ).replace(is_prefilling=torch.zeros(4, dtype=torch.bool)) |
| 1031 | builder = _make_builder( |
| 1032 | KimiK3KDAMetadataBuilder, |
| 1033 | num_speculative_tokens=0, |
| 1034 | full_cuda_graph=True, |
| 1035 | ) |
| 1036 | actual = builder.build_for_cudagraph_capture(common_attn_metadata) |
| 1037 | |
| 1038 | assert actual.num_prefills == 0 |
| 1039 | assert actual.num_decodes == 4 |
| 1040 | assert actual.has_initial_state is None |
| 1041 | staged = actual.non_spec_state_indices_tensor |
| 1042 | assert staged is not None |
| 1043 | assert staged.data_ptr() == builder.non_spec_state_indices_tensor.data_ptr() |
| 1044 | torch.testing.assert_close(staged, common_attn_metadata.block_table_tensor[:, 0]) |