| 1 | # SPDX-License-Identifier: Apache-2.0 |
| 2 | # SPDX-FileCopyrightText: Copyright contributors to the vLLM project |
| 3 | |
| 4 | from types import SimpleNamespace |
| 5 | |
| 6 | import pytest |
| 7 | import torch |
| 8 | |
| 9 | from tests.v1.attention.utils import create_vllm_config |
| 10 | from vllm.model_executor.layers.attention.sparse_mla_attention import ( |
| 11 | SparseMLACommonMetadataBuilder, |
| 12 | ) |
| 13 | from vllm.models.deepseek_v4.sparse_mla import DeepseekV4SparseMLABackend |
| 14 | from vllm.models.deepseek_v41.sparse_mla import ( |
| 15 | DeepseekV4SparseMLABackend as DeepseekV41SparseMLABackend, |
| 16 | ) |
| 17 | from vllm.v1.attention.backend import CommonAttentionMetadata |
| 18 | from vllm.v1.attention.backends.mla import indexer as indexer_module |
| 19 | from vllm.v1.attention.backends.mla.compressor_utils import ( |
| 20 | CompressedSlotMappingKernel, |
| 21 | get_compressed_slot_mapping, |
| 22 | ) |
| 23 | from vllm.v1.attention.backends.mla.indexer import ( |
| 24 | BuildPrefillChunkMetadataKernel, |
| 25 | DeepseekV4IndexerBackend, |
| 26 | DeepseekV32IndexerMetadataBuilder, |
| 27 | DeepseekV41IndexerBackend, |
| 28 | get_max_prefill_buffer_size, |
| 29 | ) |
| 30 | from vllm.v1.attention.backends.mla.sparse_utils import ( |
| 31 | ConvertReqIndexToGlobalIndexKernel, |
| 32 | ) |
| 33 | from vllm.v1.kv_cache_interface import ( |
| 34 | MLAAttentionSpec, |
| 35 | compute_layer_kv_cache_shape_bytes, |
| 36 | ) |
| 37 | from vllm.v1.worker.block_table import get_block_table_width |
| 38 | from vllm.v1.worker.utils import select_common_block_size |
| 39 | |
| 40 | |
| 41 | @pytest.mark.skipif(not torch.cuda.is_available(), reason="requires CUDA") |
| 42 | @pytest.mark.parametrize("query_lens", [[1], [6] * 16, [0, 2, 6, 0, 4], [0, 0, 0]]) |
| 43 | @pytest.mark.parametrize("padding", [0, 17]) |
| 44 | def test_fused_indexer_decode_metadata(query_lens, padding): |
| 45 | """Flatten device query boundaries and clear graph padding on every replay.""" |
| 46 | from vllm.v1.attention.ops.metadata import _indexer_decode_metadata_kernel |
| 47 | |
| 48 | device = "cuda" |
| 49 | lengths = torch.tensor(query_lens, device=device, dtype=torch.int32) |
| 50 | qsl = torch.cat( |
| 51 | [torch.zeros(1, device=device, dtype=torch.int32), lengths.cumsum(0).int()] |
| 52 | ) |
| 53 | reqs = len(query_lens) |
| 54 | n = sum(query_lens) |
| 55 | tokens = n + padding |
| 56 | capacity = max(tokens, reqs) + 31 |
| 57 | seq = lengths + 4096 |
| 58 | # Exercise block-table row padding and masked columns. |
| 59 | bt = torch.arange(reqs * 74, device=device, dtype=torch.int32).view(reqs, 74)[ |
| 60 | :, :65 |
| 61 | ] |
| 62 | out_bt = torch.full((capacity, 69), -99, device=device, dtype=torch.int32) |
| 63 | outputs = [ |
| 64 | torch.full((capacity,), -99, device=device, dtype=torch.int32) for _ in range(4) |
| 65 | ] |
| 66 | out_seq, out_lens, indices, per_req = outputs |
| 67 | grid = max(reqs, tokens + (capacity - tokens + 255) // 256) |
| 68 | _indexer_decode_metadata_kernel[(grid,)]( |
| 69 | qsl, |
| 70 | seq, |
| 71 | bt, |
| 72 | out_seq, |
| 73 | out_bt, |
| 74 | out_lens, |
| 75 | indices, |
| 76 | per_req, |
| 77 | reqs, |
| 78 | tokens, |
| 79 | capacity, |
| 80 | bt.stride(0), |
| 81 | out_bt.stride(0), |
| 82 | BLOCK_COLS=bt.shape[1], |
| 83 | num_warps=4, |
| 84 | ) |
| 85 | expected_req = torch.repeat_interleave(torch.arange(reqs, device=device), lengths) |
| 86 | expected_seq = torch.cat( |
| 87 | [torch.arange(4097, 4097 + q, device=device) for q in query_lens] |
| 88 | ) |
| 89 | torch.testing.assert_close(out_seq[:n], expected_seq.int(), rtol=0, atol=0) |
| 90 | assert torch.count_nonzero(out_seq[n:]) == 0 |
| 91 | torch.testing.assert_close(out_bt[:n, :65], bt[expected_req], rtol=0, atol=0) |
| 92 | assert torch.count_nonzero(out_bt[n:tokens, :65]) == 0 |
| 93 | torch.testing.assert_close(indices[:n], expected_req.int(), rtol=0, atol=0) |
| 94 | torch.testing.assert_close( |
| 95 | indices[n:tokens], |
| 96 | torch.arange(reqs, reqs + padding, device=device, dtype=torch.int32), |
| 97 | ) |
| 98 | assert torch.all(out_lens[:tokens] == 1) |
| 99 | torch.testing.assert_close(per_req[:reqs], lengths) |
| 100 | |
| 101 | |
| 102 | @pytest.mark.skipif(not torch.cuda.is_available(), reason="requires CUDA") |
| 103 | @pytest.mark.parametrize( |
| 104 | "query_lens", [[1], [0, 257, 1, 0, 3], [6] * 64, [2, 6, 4, 0], [0, 0, 0]] |
| 105 | ) |
| 106 | @pytest.mark.parametrize("use_sparse_mla_builder", [False, True]) |
| 107 | def test_device_token_request_mapping(query_lens, use_sparse_mla_builder): |
| 108 | """Graph replay follows device boundaries even when CPU lengths are stale.""" |
| 109 | lengths = torch.tensor(query_lens, device="cuda", dtype=torch.int32) |
| 110 | qsl = torch.cat( |
| 111 | [torch.zeros(1, device="cuda", dtype=torch.int32), lengths.cumsum(0).int()] |
| 112 | ) |
| 113 | n = sum(query_lens) |
| 114 | output = torch.full((n + 7,), -99, device="cuda", dtype=torch.int32) |
| 115 | common = CommonAttentionMetadata( |
| 116 | query_start_loc=qsl, |
| 117 | query_start_loc_cpu=qsl.cpu(), |
| 118 | seq_lens=lengths, |
| 119 | num_reqs=len(query_lens), |
| 120 | num_actual_tokens=output.numel(), |
| 121 | max_query_len=max(query_lens), |
| 122 | max_seq_len=max(query_lens), |
| 123 | block_table_tensor=torch.empty( |
| 124 | (len(query_lens), 1), device="cuda", dtype=torch.int32 |
| 125 | ), |
| 126 | slot_mapping=torch.full((n + 7,), -1, device="cuda", dtype=torch.int64), |
| 127 | ) |
| 128 | |
| 129 | def build_mapping(): |
| 130 | if use_sparse_mla_builder: |
| 131 | builder = SimpleNamespace(req_id_per_token_buffer=output) |
| 132 | return SparseMLACommonMetadataBuilder._build_req_id_per_token( |
| 133 | builder, common |
| 134 | ) |
| 135 | return common.token_to_req_indices(output) |
| 136 | |
| 137 | result = build_mapping() |
| 138 | assert result.data_ptr() == output.data_ptr() |
| 139 | expected = torch.repeat_interleave( |
| 140 | torch.arange(len(query_lens), device="cuda", dtype=torch.int32), lengths |
| 141 | ) |
| 142 | torch.testing.assert_close(output[:n], expected) |
| 143 | assert torch.count_nonzero(output[n:]) == 0 |
| 144 | graph = torch.cuda.CUDAGraph() |
| 145 | common._token_to_req_indices_cache = None |
| 146 | with torch.cuda.graph(graph): |
| 147 | build_mapping() |
| 148 | reversed_lens = lengths.flip(0) |
| 149 | qsl[1:].copy_(reversed_lens.cumsum(0)) |
| 150 | graph.replay() |
| 151 | expected = torch.repeat_interleave( |
| 152 | torch.arange(len(query_lens), device="cuda", dtype=torch.int32), reversed_lens |
| 153 | ) |
| 154 | torch.testing.assert_close(output[:n], expected) |
| 155 | assert torch.count_nonzero(output[n:]) == 0 |
| 156 | |
| 157 | |
| 158 | def test_indexer_shares_uncompressed_block_size_with_deepseek_v4_mla(): |
| 159 | """Packed MLA/indexer groups must retain 64 compressed rows per page.""" |
| 160 | kernel_block_size = select_common_block_size( |
| 161 | 256, [DeepseekV4SparseMLABackend, DeepseekV4IndexerBackend] |
| 162 | ) |
| 163 | spec = MLAAttentionSpec( |
| 164 | block_size=256, |
| 165 | num_kv_heads=1, |
| 166 | head_size=132, |
| 167 | dtype=torch.uint8, |
| 168 | tokens_per_state=4, |
| 169 | ) |
| 170 | assert compute_layer_kv_cache_shape_bytes(spec, 2, kernel_block_size) == ( |
| 171 | 2, |
| 172 | 1, |
| 173 | 64, |
| 174 | 132, |
| 175 | ) |
| 176 | |
| 177 | |
| 178 | def test_indexer_preserves_deepseek_v41_mla_block_size(): |
| 179 | """V4.1 retains its smaller pages independently of V4's C4 indexer.""" |
| 180 | block_size = DeepseekV41SparseMLABackend.get_supported_kernel_block_sizes()[0] |
| 181 | assert isinstance(block_size, int) |
| 182 | assert ( |
| 183 | select_common_block_size( |
| 184 | block_size, [DeepseekV41SparseMLABackend, DeepseekV41IndexerBackend] |
| 185 | ) |
| 186 | == block_size |
| 187 | ) |
| 188 | |
| 189 | |
| 190 | def test_indexer_warmup_normalizes_zero_compress_ratios(): |
| 191 | config = SimpleNamespace( |
| 192 | scheduler_config=SimpleNamespace(max_num_batched_tokens=8), |
| 193 | model_config=SimpleNamespace( |
| 194 | hf_text_config=SimpleNamespace( |
| 195 | compress_ratios=[0, 0, 4, 128, 0], index_kpool=32 |
| 196 | ) |
| 197 | ), |
| 198 | parallel_config=SimpleNamespace( |
| 199 | decode_context_parallel_size=1, |
| 200 | prefill_context_parallel_size=1, |
| 201 | cp_kv_cache_interleave_size=1, |
| 202 | ), |
| 203 | ) |
| 204 | |
| 205 | keys = BuildPrefillChunkMetadataKernel().get_warmup_keys(config) |
| 206 | |
| 207 | assert {key.compress_ratio for key in keys} == {1, 4, 32, 128} |
| 208 | assert {(key.query_slice_start, key.query_slice_stop) for key in keys} == { |
| 209 | (query_slice_start, query_slice_stop) |
| 210 | for query_slice_start in (1, 2, 16) |
| 211 | for query_slice_stop in (1, 2, 16) |
| 212 | } |
| 213 | |
| 214 | |
| 215 | def test_indexer_warmup_includes_pcp_normalized_dcp_key(monkeypatch): |
| 216 | monkeypatch.setattr( |
| 217 | "vllm.v1.attention.backends.mla.indexer.get_dcp_group", |
| 218 | lambda: SimpleNamespace(rank_in_group=2), |
| 219 | ) |
| 220 | config = SimpleNamespace( |
| 221 | scheduler_config=SimpleNamespace(max_num_batched_tokens=8), |
| 222 | model_config=SimpleNamespace( |
| 223 | hf_text_config=SimpleNamespace(compress_ratios=[32], index_kpool=32) |
| 224 | ), |
| 225 | parallel_config=SimpleNamespace( |
| 226 | decode_context_parallel_size=4, |
| 227 | prefill_context_parallel_size=4, |
| 228 | cp_kv_cache_interleave_size=1, |
| 229 | ), |
| 230 | ) |
| 231 | |
| 232 | keys = BuildPrefillChunkMetadataKernel().get_warmup_keys(config) |
| 233 | |
| 234 | # Triton's compile key normalizes generic i32 values to 2 and divisible |
| 235 | # i32 values (including zero) to 16. |
| 236 | assert {(key.dcp_rank, key.dcp_world) for key in keys} == {(2, 2), (16, 1)} |
| 237 | assert { |
| 238 | ( |
| 239 | key.input_variant.is_aligned("uncompressed_seq_lens"), |
| 240 | key.input_variant.is_aligned("cu_compressed_seq_lens"), |
| 241 | ) |
| 242 | for key in keys |
| 243 | } == {(False, False), (False, True), (True, False), (True, True)} |
| 244 | |
| 245 | |
| 246 | def test_compressed_slot_mapping_warmup_includes_index_kpool(): |
| 247 | config = SimpleNamespace( |
| 248 | cache_config=SimpleNamespace(block_size=256), |
| 249 | model_config=SimpleNamespace(hf_text_config=SimpleNamespace(index_kpool=32)), |
| 250 | ) |
| 251 | |
| 252 | keys = CompressedSlotMappingKernel().get_warmup_keys(config) |
| 253 | assert {(key.compress_ratio, key.block_size) for key in keys} == {(32, 2)} |
| 254 | |
| 255 | |
| 256 | @pytest.mark.skipif(not torch.cuda.is_available(), reason="requires CUDA") |
| 257 | def test_compressed_slot_mapping_inherits_padded_token_slots(): |
| 258 | """A token whose own slot is padded (SWA bounded replay) closes no |
| 259 | compressed state either.""" |
| 260 | device = torch.device("cuda") |
| 261 | query_start_loc = torch.tensor([0, 8], dtype=torch.int32, device=device) |
| 262 | seq_lens = torch.tensor([8], dtype=torch.int32, device=device) |
| 263 | block_table = torch.tensor([[3]], dtype=torch.int32, device=device) |
| 264 | slot_mapping = torch.arange(8, dtype=torch.int64, device=device) |
| 265 | slot_mapping[:4] = -1 |
| 266 | compressed = get_compressed_slot_mapping( |
| 267 | 8, |
| 268 | slot_mapping, |
| 269 | query_start_loc, |
| 270 | seq_lens, |
| 271 | block_table, |
| 272 | block_size=4, |
| 273 | compress_ratio=2, |
| 274 | ) |
| 275 | assert compressed.tolist() == [-1, -1, -1, -1, -1, 3 * 4 + 2, -1, 3 * 4 + 3] |
| 276 | |
| 277 | |
| 278 | def test_index_conversion_warmup_uses_physical_block_stride(): |
| 279 | config = SimpleNamespace( |
| 280 | cache_config=SimpleNamespace(block_size=64), |
| 281 | model_config=SimpleNamespace( |
| 282 | max_model_len=1024, |
| 283 | hf_text_config=SimpleNamespace(index_topk=2048), |
| 284 | ), |
| 285 | parallel_config=SimpleNamespace( |
| 286 | decode_context_parallel_size=1, |
| 287 | prefill_context_parallel_size=1, |
| 288 | cp_kv_cache_interleave_size=1, |
| 289 | ), |
| 290 | ) |
| 291 | |
| 292 | keys = ConvertReqIndexToGlobalIndexKernel().get_warmup_keys( |
| 293 | config, |
| 294 | block_stride_rows=4096, |
| 295 | ) |
| 296 | assert {key.block_stride_rows for key in keys} == {4096} |
| 297 | |
| 298 | |
| 299 | def test_zero_token_pcp_rank_participates_in_compressed_mapping_gather(monkeypatch): |
| 300 | builder = object.__new__(DeepseekV32IndexerMetadataBuilder) |
| 301 | builder.compress_ratio = 4 |
| 302 | builder.use_pcp = True |
| 303 | builder.pcp_world_size = 4 |
| 304 | builder.pcp_rank = 0 |
| 305 | builder.kernel_block_size = None |
| 306 | builder.kv_cache_spec = SimpleNamespace(block_size=64, num_states=64) |
| 307 | builder.compressed_slot_mapping_buffer = torch.zeros(8, dtype=torch.int64) |
| 308 | |
| 309 | def fake_get_compressed_slot_mapping( |
| 310 | num_tokens, |
| 311 | slot_mapping, |
| 312 | query_start_loc, |
| 313 | seq_lens, |
| 314 | block_table, |
| 315 | block_size, |
| 316 | compress_ratio, |
| 317 | out, |
| 318 | ): |
| 319 | assert num_tokens == 0 |
| 320 | assert block_size == 64 |
| 321 | assert compress_ratio == 4 |
| 322 | out.fill_(-1) |
| 323 | return out[:num_tokens] |
| 324 | |
| 325 | gathered_slot_mapping = torch.tensor([123, -1, -1, -1], dtype=torch.int64) |
| 326 | |
| 327 | class FakePCPGroup: |
| 328 | def __init__(self): |
| 329 | self.calls = 0 |
| 330 | |
| 331 | def all_gather(self, tensor, dim=0): |
| 332 | self.calls += 1 |
| 333 | assert dim == 0 |
| 334 | torch.testing.assert_close(tensor, torch.tensor([-1])) |
| 335 | return gathered_slot_mapping |
| 336 | |
| 337 | fake_pcp_group = FakePCPGroup() |
| 338 | monkeypatch.setattr( |
| 339 | indexer_module, |
| 340 | "get_compressed_slot_mapping", |
| 341 | fake_get_compressed_slot_mapping, |
| 342 | ) |
| 343 | monkeypatch.setattr(indexer_module, "get_pcp_group", lambda: fake_pcp_group) |
| 344 | |
| 345 | query_start_loc = torch.tensor([0, 0], dtype=torch.int32) |
| 346 | seq_lens = torch.tensor([16], dtype=torch.int32) |
| 347 | common = CommonAttentionMetadata( |
| 348 | query_start_loc=query_start_loc, |
| 349 | query_start_loc_cpu=query_start_loc, |
| 350 | seq_lens=seq_lens, |
| 351 | seq_lens_cpu_upper_bound=seq_lens, |
| 352 | num_reqs=1, |
| 353 | num_actual_tokens=0, |
| 354 | max_query_len=0, |
| 355 | max_seq_len=16, |
| 356 | block_table_tensor=torch.tensor([[0]], dtype=torch.int32), |
| 357 | slot_mapping=torch.tensor([11, 22, 33, 44], dtype=torch.int64), |
| 358 | causal=True, |
| 359 | ) |
| 360 | |
| 361 | metadata = builder.build(common_prefix_len=0, common_attn_metadata=common) |
| 362 | |
| 363 | assert fake_pcp_group.calls == 1 |
| 364 | assert metadata.num_decodes == 0 |
| 365 | assert metadata.num_decode_tokens == 0 |
| 366 | assert metadata.num_prefills == 0 |
| 367 | assert metadata.num_prefill_tokens == 0 |
| 368 | torch.testing.assert_close(metadata.slot_mapping, gathered_slot_mapping) |
| 369 | |
| 370 | |
| 371 | @pytest.mark.skipif(not torch.cuda.is_available(), reason="requires CUDA") |
| 372 | def test_indexer_builder_deepseek_v4_compressed_slot_mapping_uses_num_states(): |
| 373 | """Regression test: DeepseekV4 compression path must compute slot_mapping from |
| 374 | compressed positions, not reuse the uncompressed common metadata mapping. |
| 375 | """ |
| 376 | device = torch.device("cuda") |
| 377 | |
| 378 | # num_states = block_size // tokens_per_state = 256 // 4 = 64 |
| 379 | kv_cache_spec = MLAAttentionSpec( |
| 380 | block_size=256, |
| 381 | num_kv_heads=1, |
| 382 | head_size=128, |
| 383 | dtype=torch.bfloat16, |
| 384 | tokens_per_state=4, |
| 385 | ) |
| 386 | vllm_config = create_vllm_config(max_model_len=1024) |
| 387 | max_num_blocks = kv_cache_spec.max_num_blocks_per_req(vllm_config, 1024) |
| 388 | block_table_width = get_block_table_width(max_num_blocks, kv_cache_spec.block_size) |
| 389 | builder = DeepseekV32IndexerMetadataBuilder( |
| 390 | kv_cache_spec=kv_cache_spec, |
| 391 | layer_names=["dummy"], |
| 392 | vllm_config=vllm_config, |
| 393 | device=device, |
| 394 | block_table_width=block_table_width, |
| 395 | ) |
| 396 | |
| 397 | # Construct a single request where: |
| 398 | # - num_computed = 240 (=> compressed_pos_start = 60) |
| 399 | # - query_len = 40 (=> num_groups = 10) |
| 400 | # => compressed positions are 60..69 which cross the storage block boundary at 64. |
| 401 | query_start_loc = torch.tensor([0, 40], dtype=torch.int32, device=device) |
| 402 | query_start_loc_cpu = query_start_loc.cpu() |
| 403 | seq_lens = torch.tensor([280], dtype=torch.int32, device=device) # 240 + 40 |
| 404 | |
| 405 | # Two blocks: compressed positions 0..63 map to block 5, 64..127 map to block 7. |
| 406 | block_table_tensor = torch.tensor([[5, 7]], dtype=torch.int32, device=device) |
| 407 | |
| 408 | # Dummy uncompressed slot mapping (length == uncompressed num_actual_tokens). |
| 409 | slot_mapping = torch.full((40,), -123, dtype=torch.int64, device=device) |
| 410 | |
| 411 | common = CommonAttentionMetadata( |
| 412 | query_start_loc=query_start_loc, |
| 413 | query_start_loc_cpu=query_start_loc_cpu, |
| 414 | seq_lens=seq_lens, |
| 415 | seq_lens_cpu_upper_bound=seq_lens.cpu(), |
| 416 | num_reqs=1, |
| 417 | num_actual_tokens=40, |
| 418 | max_query_len=40, |
| 419 | max_seq_len=280, |
| 420 | block_table_tensor=block_table_tensor, |
| 421 | slot_mapping=slot_mapping, |
| 422 | causal=True, |
| 423 | ) |
| 424 | |
| 425 | md = builder.build(common_prefix_len=0, common_attn_metadata=common) |
| 426 | |
| 427 | # The compressed slot_mapping retains the original uncompressed size (40). |
| 428 | # Only every compress_ratio-th position gets a valid slot; the rest are -1. |
| 429 | assert md.slot_mapping.numel() == 40 |
| 430 | valid_slots = md.slot_mapping[md.slot_mapping >= 0] |
| 431 | assert valid_slots.numel() == 10 # 40 tokens / compress_ratio 4 |
| 432 | |
| 433 | storage_bs = kv_cache_spec.num_states # 64 |
| 434 | # Compressed positions 60..63 land in block 5, positions 64..69 in block 7. |
| 435 | expected = torch.tensor( |
| 436 | [ |
| 437 | 5 * storage_bs + 60, |
| 438 | 5 * storage_bs + 61, |
| 439 | 5 * storage_bs + 62, |
| 440 | 5 * storage_bs + 63, |
| 441 | ] |
| 442 | + [ |
| 443 | 7 * storage_bs + 0, |
| 444 | 7 * storage_bs + 1, |
| 445 | 7 * storage_bs + 2, |
| 446 | 7 * storage_bs + 3, |
| 447 | 7 * storage_bs + 4, |
| 448 | 7 * storage_bs + 5, |
| 449 | ], |
| 450 | dtype=torch.int64, |
| 451 | device=device, |
| 452 | ) |
| 453 | torch.testing.assert_close(valid_slots, expected) |
| 454 | |
| 455 | |
| 456 | @pytest.mark.parametrize("compress_ratio", [1, 4]) |
| 457 | def test_indexer_prefill_budget_matches_compressed_workspace(compress_ratio): |
| 458 | """The chunker budget is in compressed rows, like the K-gather workspace.""" |
| 459 | max_model_len = 1024 |
| 460 | kv_cache_spec = MLAAttentionSpec( |
| 461 | block_size=256, |
| 462 | num_kv_heads=1, |
| 463 | head_size=128, |
| 464 | dtype=torch.bfloat16, |
| 465 | tokens_per_state=compress_ratio, |
| 466 | ) |
| 467 | vllm_config = create_vllm_config(max_model_len=max_model_len) |
| 468 | max_num_blocks = kv_cache_spec.max_num_blocks_per_req(vllm_config, max_model_len) |
| 469 | block_table_width = get_block_table_width(max_num_blocks, kv_cache_spec.block_size) |
| 470 | builder = DeepseekV32IndexerMetadataBuilder( |
| 471 | kv_cache_spec=kv_cache_spec, |
| 472 | layer_names=["dummy"], |
| 473 | vllm_config=vllm_config, |
| 474 | device=torch.device("cpu"), |
| 475 | block_table_width=block_table_width, |
| 476 | ) |
| 477 | |
| 478 | workspace_rows = get_max_prefill_buffer_size(vllm_config) // compress_ratio |
| 479 | assert builder.max_prefill_buffer_size == workspace_rows |
| 480 | |
| 481 | # Overflows the workspace unless the chunker splits. |
| 482 | num_reqs = 3 * (workspace_rows // (max_model_len // compress_ratio)) + 1 |
| 483 | compressed_seq_lens = torch.full((num_reqs,), max_model_len // compress_ratio) |
| 484 | query_lens = torch.ones(num_reqs, dtype=torch.int64) |
| 485 | chunks = builder._split_indexer_prefill_chunks( |
| 486 | compressed_seq_lens, |
| 487 | query_lens, |
| 488 | builder.max_prefill_buffer_size, |
| 489 | max_logits_bytes=10**15, |
| 490 | ) |
| 491 | assert len(chunks) > 1 |
| 492 | for req_slice, _ in chunks: |
| 493 | assert int(compressed_seq_lens[req_slice].sum()) <= workspace_rows |