| 1 | # SPDX-License-Identifier: Apache-2.0 |
| 2 | # SPDX-FileCopyrightText: Copyright contributors to the vLLM project |
| 3 | """Tests for the FlashInfer SM90 sparse MLA backend wiring and index packing. |
| 4 | |
| 5 | The FlashInfer wrapper and top-k conversion are replaced by CPU recorders; |
| 6 | the tests pin the contract between the impl and the kernel API: page_size=1 |
| 7 | varlen rows, reserved-buffer refresh, plan parameters (dims, NoPE/rope scale, |
| 8 | causality), ckv/kpe cache splitting, and the backend's model-shape gates. |
| 9 | """ |
| 10 | |
| 11 | from types import SimpleNamespace |
| 12 | from typing import Any |
| 13 | |
| 14 | import pytest |
| 15 | import torch |
| 16 | |
| 17 | # isort: off |
| 18 | import vllm.v1.attention.backends.mla.flashinfer_mla_sparse_sm90 as sm90_mod |
| 19 | from vllm.v1.attention.backends.mla.flashinfer_mla_sparse_sm90 import ( |
| 20 | FlashInferMLASparseSM90Backend, |
| 21 | FlashInferMLASparseSM90Builder, |
| 22 | FlashInferMLASparseSM90Impl, |
| 23 | ) |
| 24 | from vllm.v1.kv_cache_interface import MLAAttentionSpec |
| 25 | # isort: on |
| 26 | |
| 27 | BLOCK_SIZE = 64 |
| 28 | HEAD = 512 |
| 29 | TOPK = 128 # triton convert requires width % 128 == 0 |
| 30 | |
| 31 | |
| 32 | def ref_convert(req_id, block_table, token_indices, BLOCK_SIZE=64, **_): |
| 33 | out = torch.full_like(token_indices, -1) |
| 34 | counts = torch.zeros(token_indices.shape[0], dtype=torch.int32) |
| 35 | for t in range(token_indices.shape[0]): |
| 36 | vals = [] |
| 37 | for j in range(token_indices.shape[1]): |
| 38 | pos = int(token_indices[t, j]) |
| 39 | if pos == -1: |
| 40 | continue |
| 41 | blk = int(block_table[int(req_id[t]), pos // BLOCK_SIZE]) |
| 42 | if blk < 0: |
| 43 | continue |
| 44 | vals.append(blk * BLOCK_SIZE + pos % BLOCK_SIZE) |
| 45 | out[t, : len(vals)] = torch.tensor(vals, dtype=out.dtype) |
| 46 | counts[t] = len(vals) |
| 47 | return out, counts |
| 48 | |
| 49 | |
| 50 | class FakeWrapper: |
| 51 | def __init__(self): |
| 52 | self.plan_args = None |
| 53 | self.run_args = None |
| 54 | self._plan_info: Any = None |
| 55 | self._planned_backend: Any = None |
| 56 | |
| 57 | def plan(self, *args, **kwargs): |
| 58 | self.plan_args = (args, kwargs) |
| 59 | |
| 60 | def run(self, *args, **kwargs): |
| 61 | q_nope, q_pe, ckv, kpe = args |
| 62 | self.run_args = (q_nope, q_pe, ckv, kpe, kwargs) |
| 63 | return torch.zeros( |
| 64 | q_nope.shape[0], q_nope.shape[1], ckv.shape[-1], dtype=torch.bfloat16 |
| 65 | ) |
| 66 | |
| 67 | |
| 68 | class FakeState: |
| 69 | def __init__(self, width, max_tokens=64): |
| 70 | self.kv_indices = torch.zeros(max_tokens * width, dtype=torch.int32) |
| 71 | self.kv_len_arr = torch.zeros(max_tokens, dtype=torch.int32) |
| 72 | self.kv_indptr = torch.zeros(max_tokens + 1, dtype=torch.int32) |
| 73 | self.wrapper = FakeWrapper() |
| 74 | self.plan_calls = [] |
| 75 | self.index_topk = 2048 |
| 76 | self.index_kpool = 4 |
| 77 | self.max_valid = 2051 |
| 78 | |
| 79 | def plan(self, num_tokens, kv_lens, cam, req_id_per_token): |
| 80 | self.plan_calls.append((num_tokens, kv_lens)) |
| 81 | |
| 82 | def pack_indices(self, slots): |
| 83 | for row in range(slots.shape[0]): |
| 84 | start, end = self.kv_indptr[row : row + 2].tolist() |
| 85 | self.kv_indices[start:end] = slots[row, : end - start].clamp(min=0) |
| 86 | |
| 87 | |
| 88 | def make_impl(qk_rope, kv_dtype="fp8_e4m3", num_heads=2, topk_width=TOPK): |
| 89 | impl = object.__new__(FlashInferMLASparseSM90Impl) |
| 90 | impl.num_heads = num_heads |
| 91 | impl.head_size = HEAD + qk_rope |
| 92 | impl.scale = (HEAD + qk_rope) ** -0.5 |
| 93 | impl.kv_lora_rank = HEAD |
| 94 | impl.qk_rope_head_dim = qk_rope |
| 95 | impl.kv_cache_dtype = kv_dtype |
| 96 | impl.use_fp8_kv_cache = kv_dtype in ("fp8", "fp8_e4m3") |
| 97 | rows = 4 |
| 98 | impl.topk_indices_buffer = torch.full((rows, topk_width), -1, dtype=torch.int32) |
| 99 | return impl, rows |
| 100 | |
| 101 | |
| 102 | def make_batch(rows, topk_rows, own_blocks): |
| 103 | req_id = torch.tensor([0] * rows, dtype=torch.int32) |
| 104 | block_table = torch.zeros(1, 16, dtype=torch.int32) |
| 105 | block_table[:, 0] = own_blocks[0] |
| 106 | topk = torch.full((rows, TOPK), -1, dtype=torch.int32) |
| 107 | for t, row in enumerate(topk_rows): |
| 108 | topk[t, : len(row)] = torch.tensor(row, dtype=torch.int32) |
| 109 | return SimpleNamespace( |
| 110 | req_id_per_token=req_id, block_table=block_table, block_size=BLOCK_SIZE |
| 111 | ) |
| 112 | |
| 113 | |
| 114 | @pytest.mark.parametrize("qk_rope,kv_dtype", [(0, "fp8_e4m3"), (64, "auto")]) |
| 115 | def test_forward_wiring(monkeypatch, qk_rope, kv_dtype): |
| 116 | impl, rows = make_impl(qk_rope, kv_dtype) |
| 117 | state = FakeState(TOPK) |
| 118 | monkeypatch.setattr( |
| 119 | sm90_mod, "triton_convert_req_index_to_global_index", ref_convert |
| 120 | ) |
| 121 | |
| 122 | # req with context 10 < topk: 8 valid + -1 padding. |
| 123 | topk_rows = [ |
| 124 | [7, 3, 1, 9, 0, 2, 5, 8] + [-1] * (TOPK - 8), |
| 125 | [4, 0, 2, 3, 1] + [-1] * (TOPK - 5), |
| 126 | [6, 5, 4] + [-1] * (TOPK - 3), |
| 127 | [2] + [-1] * (TOPK - 1), |
| 128 | ] |
| 129 | impl.topk_indices_buffer.copy_(torch.tensor(topk_rows, dtype=torch.int32)) |
| 130 | state.kv_indptr[1:5] = torch.tensor([8, 13, 16, 17], dtype=torch.int32) |
| 131 | state.kv_indptr[5:] = 17 |
| 132 | meta = make_batch(rows, topk_rows, [3]) |
| 133 | meta.state = state |
| 134 | q_nope = torch.randn(rows, impl.num_heads, HEAD) |
| 135 | q_rope = torch.randn(rows, impl.num_heads, qk_rope) |
| 136 | cache = torch.zeros( |
| 137 | 8 * BLOCK_SIZE, |
| 138 | impl.head_size, |
| 139 | dtype=torch.uint8 if impl.use_fp8_kv_cache else torch.bfloat16, |
| 140 | ) |
| 141 | |
| 142 | out, lse = impl.forward_mqa( |
| 143 | (q_nope, q_rope), cache, meta, SimpleNamespace(_k_scale_float=0.5) |
| 144 | ) |
| 145 | assert lse is None and out.shape == (rows, impl.num_heads, HEAD) |
| 146 | |
| 147 | # Reserved buffers carry this step's slots; lengths are NOT refreshed |
| 148 | # here (the builder plans them host-side before capture/replay). |
| 149 | ref_slots, ref_counts = ref_convert( |
| 150 | meta.req_id_per_token, meta.block_table, impl.topk_indices_buffer |
| 151 | ) |
| 152 | offset = 0 |
| 153 | for t in range(rows): |
| 154 | k = int(ref_counts[t]) |
| 155 | assert ( |
| 156 | state.kv_indices[offset : offset + k].tolist() == ref_slots[t, :k].tolist() |
| 157 | ) |
| 158 | offset += k |
| 159 | assert state.plan_calls == [] |
| 160 | |
| 161 | assert state.wrapper.run_args is not None |
| 162 | q_pe, ckv, kpe, kwargs = state.wrapper.run_args[1:] |
| 163 | assert q_pe.shape == (rows, impl.num_heads, qk_rope) |
| 164 | assert ckv.shape == (8 * BLOCK_SIZE, 1, HEAD) |
| 165 | assert kpe.shape[-1] == qk_rope |
| 166 | if impl.use_fp8_kv_cache: |
| 167 | assert kwargs["ckv_scale"] == 0.5 and kwargs["kpe_scale"] == 1.0 |
| 168 | else: |
| 169 | assert kwargs == {} |
| 170 | |
| 171 | |
| 172 | @pytest.mark.parametrize("use_mha", [False, True]) |
| 173 | @pytest.mark.parametrize("num_decodes", [0, 1]) |
| 174 | def test_builder_plans_only_rows_dispatched_to_mqa(monkeypatch, use_mha, num_decodes): |
| 175 | """MHA prefill rows must not make the MQA kernel read beyond its query.""" |
| 176 | builder = object.__new__(FlashInferMLASparseSM90Builder) |
| 177 | builder._adaptive_verification = False |
| 178 | builder.state = FakeState(TOPK) |
| 179 | builder._attention_layer = SimpleNamespace(_use_sparse_mha=lambda _: use_mha) |
| 180 | metadata = object.__new__(sm90_mod.FlashInferMLASparseSM90Metadata) |
| 181 | metadata.state = None |
| 182 | metadata.req_id_per_token = None |
| 183 | metadata.num_prefills = 1 |
| 184 | metadata.num_decode_tokens = num_decodes |
| 185 | monkeypatch.setattr( |
| 186 | sm90_mod.FlashInferMLASparseMetadataBuilder, |
| 187 | "build", |
| 188 | lambda *_args, **_kwargs: metadata, |
| 189 | ) |
| 190 | cam = SimpleNamespace( |
| 191 | num_reqs=num_decodes + 1, |
| 192 | query_start_loc_cpu=torch.tensor( |
| 193 | [0, 1, 6] if num_decodes else [0, 5], dtype=torch.int32 |
| 194 | ), |
| 195 | seq_lens_cpu_upper_bound=torch.tensor( |
| 196 | [1402, 5] if num_decodes else [5], dtype=torch.int32 |
| 197 | ), |
| 198 | positions=None, |
| 199 | ) |
| 200 | |
| 201 | result = builder.build(0, cam) |
| 202 | |
| 203 | assert result.state is builder.state |
| 204 | expected_lens = [1402] if num_decodes else [] |
| 205 | if not use_mha: |
| 206 | expected_lens += [1, 2, 3, 4, 5] |
| 207 | assert builder.state.plan_calls[0][0] == len(expected_lens) |
| 208 | assert builder.state.plan_calls[0][1].tolist() == expected_lens |
| 209 | |
| 210 | |
| 211 | @pytest.mark.parametrize("kv_dtype", [torch.bfloat16, torch.float8_e4m3fn]) |
| 212 | def test_plan_uses_state_params(monkeypatch, kv_dtype): |
| 213 | """The NoPE/rope dims and scale live on the builder state, not the layer. |
| 214 | |
| 215 | Without a device batch layout plan() takes exact per-row KV lengths and |
| 216 | rebuilds the schedule on every call; the indptrs are always full-size |
| 217 | with zero-query padding rows past num_tokens. An fp8 cache plans as |
| 218 | float8_e4m3fn. |
| 219 | """ |
| 220 | wrapper = FakeWrapper() |
| 221 | plan_info = [0] * sm90_mod._PLAN_INFO_LEN |
| 222 | plan_info[sm90_mod._PI_KV_INDPTR] = 64 * 4 |
| 223 | wrapper._plan_info = torch.tensor(plan_info) |
| 224 | wrapper._planned_backend = SimpleNamespace( |
| 225 | _int_workspace_buffer=torch.zeros(64, dtype=torch.int32) |
| 226 | ) |
| 227 | state = sm90_mod._SM90State.__new__(sm90_mod._SM90State) |
| 228 | state.device = torch.device("cpu") |
| 229 | state.wrapper = wrapper |
| 230 | state.num_heads = 4 |
| 231 | state.kv_dtype = kv_dtype |
| 232 | state.kv_lora_rank = HEAD |
| 233 | state.qk_rope_head_dim = 64 |
| 234 | state.sm_scale = 576**-0.5 |
| 235 | state.max_tokens = 4 |
| 236 | state.topk_width = TOPK |
| 237 | state.kv_indices = torch.zeros(4 * TOPK) |
| 238 | state._arange_cpu = torch.arange(5, dtype=torch.int32) |
| 239 | state._qo_cpu = torch.empty(5, dtype=torch.int32) |
| 240 | state._kv_cpu = torch.empty(5, dtype=torch.int32) |
| 241 | state._lens_cpu = torch.full((4,), TOPK, dtype=torch.int32) |
| 242 | state._saved_kv_end = torch.empty(0, dtype=torch.int32) |
| 243 | |
| 244 | state.plan(3, torch.tensor([2, 5, 7], dtype=torch.int32), None, None) |
| 245 | assert wrapper.plan_args is not None |
| 246 | args, kwargs = wrapper.plan_args |
| 247 | (qo, kv, indices, kv_len, heads, ckv, kpe, page, causal, scale) = args |
| 248 | assert qo.tolist() == [0, 1, 2, 3, 3] # clamp: rows past 3 have no queries |
| 249 | assert kv.tolist() == [0, 2, 7, 14, 14] |
| 250 | assert kv_len.tolist() == [2, 5, 7, 0] |
| 251 | assert (heads, ckv, kpe, page, causal) == (4, HEAD, 64, 1, False) |
| 252 | assert scale == 576**-0.5 |
| 253 | assert kwargs["q_data_type"] == torch.bfloat16 |
| 254 | assert kwargs["kv_data_type"] == kv_dtype |
| 255 | |
| 256 | # Replanning a smaller batch must clear the previous rows' lengths. |
| 257 | state.plan(1, torch.tensor([TOPK], dtype=torch.int32), None, None) |
| 258 | assert state._kv_cpu.tolist() == [0, TOPK, TOPK, TOPK, TOPK] |
| 259 | assert state._lens_cpu.tolist() == [TOPK, 0, 0, 0] |
| 260 | state.plan(0, torch.empty(0, dtype=torch.int32), None, None) |
| 261 | assert state._kv_cpu.tolist() == [0, 0, 0, 0, 0] |
| 262 | assert state._lens_cpu.tolist() == [0, 0, 0, 0] |
| 263 | |
| 264 | |
| 265 | @pytest.mark.skipif(not torch.cuda.is_available(), reason="Requires CUDA") |
| 266 | @pytest.mark.parametrize("width", [128, 2048, 2176]) |
| 267 | def test_pack_indices_replays_with_updated_offsets(width): |
| 268 | """Graph replay packs exact prefixes after row lengths and slots change.""" |
| 269 | state = sm90_mod._SM90State.__new__(sm90_mod._SM90State) |
| 270 | state.kv_indptr = torch.zeros(5, dtype=torch.int32, device="cuda") |
| 271 | state.kv_indices = torch.full((4 * width,), -99, dtype=torch.int32, device="cuda") |
| 272 | # Exercise a contiguous view with an unaligned starting address. |
| 273 | slots = torch.arange(4 * width + 1, dtype=torch.int32, device="cuda")[1:] |
| 274 | slots = slots.view(4, width) |
| 275 | state.kv_indptr.copy_(torch.tensor([0, 1, 4, 4, 4])) |
| 276 | state.pack_indices(slots) |
| 277 | torch.accelerator.synchronize() |
| 278 | graph = torch.cuda.CUDAGraph() |
| 279 | with torch.cuda.graph(graph): |
| 280 | state.pack_indices(slots) |
| 281 | |
| 282 | for lengths in ([1, 3, 0, 0], [width, 0, 17, width - 1], [0, 0, 0, 0]): |
| 283 | offsets = torch.tensor([0, *torch.tensor(lengths).cumsum(0).tolist()]) |
| 284 | state.kv_indptr.copy_(offsets) |
| 285 | state.kv_indices.fill_(-99) |
| 286 | slots.add_(1) |
| 287 | graph.replay() |
| 288 | expected = torch.cat( |
| 289 | [slots[row, :length] for row, length in enumerate(lengths)] |
| 290 | ) |
| 291 | torch.testing.assert_close(state.kv_indices[: expected.numel()], expected) |
| 292 | assert (state.kv_indices[expected.numel() :] == -99).all() |
| 293 | |
| 294 | |
| 295 | @pytest.mark.parametrize("adaptive", [False, True]) |
| 296 | def test_kv_lens_host_formula(adaptive): |
| 297 | """Per-row host lengths: context == position + 1; capped at |
| 298 | index_topk + trailing-pool remainder past the sparse threshold. |
| 299 | |
| 300 | Exact (synced) without a host upper bound; with one, a sync-free bound |
| 301 | capped at the largest valid count, which the plan clamps on device. |
| 302 | """ |
| 303 | builder = object.__new__(FlashInferMLASparseSM90Builder) |
| 304 | builder.state = FakeState(TOPK) |
| 305 | builder._adaptive_verification = adaptive |
| 306 | seq_lens = torch.tensor([100, 9, 3000], dtype=torch.int32) |
| 307 | cam = SimpleNamespace( |
| 308 | num_reqs=3, |
| 309 | query_start_loc_cpu=torch.tensor([0, 5, 7, 10], dtype=torch.int32), |
| 310 | seq_lens=seq_lens, |
| 311 | seq_lens_cpu_upper_bound=None, |
| 312 | ) |
| 313 | num_rows, lens, exact = builder._kv_lens_host(cam) |
| 314 | assert num_rows == 10 and exact |
| 315 | # req0: positions 95..99 -> ctx 96..100 (all <= 2048: full context) |
| 316 | # req1: positions 7,8 -> ctx 8,9 |
| 317 | # req2: positions 2997..2999 -> ctx 2998..3000 (> 2048: topk + ctx%4) |
| 318 | assert lens.tolist() == [96, 97, 98, 99, 100, 8, 9, 2050, 2051, 2048] |
| 319 | |
| 320 | cam.seq_lens_cpu_upper_bound = seq_lens |
| 321 | num_rows, lens, exact = builder._kv_lens_host(cam) |
| 322 | assert num_rows == 10 and not exact |
| 323 | if adaptive: |
| 324 | # Host query split may differ from the device one: batch-wide bound. |
| 325 | assert lens.tolist() == [2051] * 10 |
| 326 | else: |
| 327 | assert lens.tolist() == [96, 97, 98, 99, 100, 8, 9, 2051, 2051, 2051] |
| 328 | |
| 329 | |
| 330 | def test_kv_lens_host_empty(): |
| 331 | builder = object.__new__(FlashInferMLASparseSM90Builder) |
| 332 | cam = SimpleNamespace( |
| 333 | num_reqs=0, |
| 334 | query_start_loc_cpu=torch.tensor([0], dtype=torch.int32), |
| 335 | seq_lens=torch.zeros(0, dtype=torch.int32), |
| 336 | ) |
| 337 | num_rows, lens, _ = builder._kv_lens_host(cam) |
| 338 | assert num_rows == 0 and lens.numel() == 0 |
| 339 | |
| 340 | |
| 341 | _NO_KPOOL = object() |
| 342 | |
| 343 | |
| 344 | @pytest.mark.parametrize( |
| 345 | "spec_dtype,expected_kv_dtype", |
| 346 | [ |
| 347 | (torch.bfloat16, torch.bfloat16), |
| 348 | (torch.uint8, torch.float8_e4m3fn), |
| 349 | ], |
| 350 | ids=["bf16", "fp8"], |
| 351 | ) |
| 352 | @pytest.mark.parametrize( |
| 353 | "index_kpool,prefill_lens", |
| 354 | [ |
| 355 | (4, [2048, 2049, 2050, 2051]), |
| 356 | (None, [2048] * 4), |
| 357 | (_NO_KPOOL, [2048] * 4), |
| 358 | ], |
| 359 | ids=["kpool4", "kpool_none", "no_kpool_attr"], |
| 360 | ) |
| 361 | def test_builder_kpool_from_model_config( |
| 362 | monkeypatch, index_kpool, prefill_lens, spec_dtype, expected_kv_dtype |
| 363 | ): |
| 364 | """The builder took kpool from the KV cache spec, whose tokens_per_state is |
| 365 | 1, so the tail pool was never read. The state plans the cache's logical |
| 366 | dtype: uint8 fp8 storage becomes float8_e4m3fn.""" |
| 367 | monkeypatch.setattr( |
| 368 | sm90_mod.FlashInferMLASparseMetadataBuilder, |
| 369 | "__init__", |
| 370 | lambda *_args, **_kwargs: None, |
| 371 | ) |
| 372 | |
| 373 | class _RecordingState: |
| 374 | def __init__(self, *args, **kwargs): |
| 375 | self.kv_dtype = args[2] |
| 376 | self.index_topk = kwargs["index_topk"] |
| 377 | self.index_kpool = kwargs["index_kpool"] |
| 378 | |
| 379 | monkeypatch.setattr(sm90_mod, "_SM90State", _RecordingState) |
| 380 | impl, _ = make_impl(64) |
| 381 | spec = MLAAttentionSpec( |
| 382 | block_size=BLOCK_SIZE, num_kv_heads=1, head_size=576, dtype=spec_dtype |
| 383 | ) |
| 384 | assert spec.tokens_per_state == 1 |
| 385 | hf_config = SimpleNamespace(index_topk=2048) |
| 386 | if index_kpool is not _NO_KPOOL: |
| 387 | hf_config.index_kpool = index_kpool |
| 388 | vllm_config = SimpleNamespace( |
| 389 | compilation_config=SimpleNamespace( |
| 390 | static_forward_context={"attn": SimpleNamespace(impl=impl)} |
| 391 | ), |
| 392 | scheduler_config=SimpleNamespace( |
| 393 | max_num_batched_tokens=64, async_scheduling=False |
| 394 | ), |
| 395 | model_config=SimpleNamespace(hf_text_config=hf_config), |
| 396 | speculative_config=None, |
| 397 | ) |
| 398 | builder = FlashInferMLASparseSM90Builder( |
| 399 | spec, ["attn"], vllm_config, torch.device("cpu") |
| 400 | ) |
| 401 | assert builder.state.kv_dtype == expected_kv_dtype |
| 402 | # req0: prefill chunk ending at context 42295; req1: context <= topk. |
| 403 | cam = SimpleNamespace( |
| 404 | num_reqs=2, |
| 405 | query_start_loc_cpu=torch.tensor([0, 4, 6], dtype=torch.int32), |
| 406 | seq_lens=torch.tensor([42295, 100], dtype=torch.int32), |
| 407 | seq_lens_cpu_upper_bound=None, |
| 408 | ) |
| 409 | num_rows, lens, exact = builder._kv_lens_host(cam) |
| 410 | assert num_rows == 6 and exact |
| 411 | assert lens.tolist() == prefill_lens + [99, 100] |
| 412 | |
| 413 | |
| 414 | def _sm90_flashinfer_available(): |
| 415 | return ( |
| 416 | torch.cuda.is_available() |
| 417 | and torch.cuda.get_device_capability()[0] == 9 |
| 418 | and sm90_mod.has_flashinfer_sm90_nope_mla() |
| 419 | ) |
| 420 | |
| 421 | |
| 422 | SM90_HEADS, SM90_TOPK, SM90_KPOOL, SM90_WIDTH = 16, 2048, 4, 2176 |
| 423 | |
| 424 | |
| 425 | def _real_state(rows): |
| 426 | return sm90_mod._SM90State( |
| 427 | torch.device("cuda"), |
| 428 | SM90_HEADS, |
| 429 | torch.bfloat16, |
| 430 | rows, |
| 431 | SM90_WIDTH, |
| 432 | HEAD, |
| 433 | 0, |
| 434 | HEAD**-0.5, |
| 435 | SM90_TOPK, |
| 436 | SM90_KPOOL, |
| 437 | ) |
| 438 | |
| 439 | |
| 440 | def _clamp_batch(num_reqs, q_per_req, lo, hi, topk, kpool, width): |
| 441 | """Random batch: per-row contexts, device layout, and top-k slots.""" |
| 442 | last = torch.randint(lo + q_per_req, hi, (num_reqs,)) |
| 443 | ctx = (last[:, None] + torch.arange(1 - q_per_req, 1)).reshape(-1) |
| 444 | valid = torch.where(ctx <= topk, ctx, topk + ctx % kpool) |
| 445 | cam = SimpleNamespace( |
| 446 | seq_lens=last.to(torch.int32).cuda(), |
| 447 | query_start_loc=(torch.arange(num_reqs + 1) * q_per_req).to(torch.int32).cuda(), |
| 448 | ) |
| 449 | req_id = torch.arange(num_reqs, dtype=torch.int32).repeat_interleave(q_per_req) |
| 450 | slots = torch.randint(0, 1 << 16, (ctx.numel(), width), dtype=torch.int32) |
| 451 | slots[torch.arange(width) >= valid[:, None]] = -1 |
| 452 | return ctx, valid.to(torch.int32), cam, req_id.cuda(), slots.cuda() |
| 453 | |
| 454 | |
| 455 | @pytest.mark.skipif( |
| 456 | not _sm90_flashinfer_available(), reason="Requires SM90 FlashInfer MLA" |
| 457 | ) |
| 458 | @pytest.mark.parametrize( |
| 459 | "num_reqs,q_per_req,lo,hi", |
| 460 | [ |
| 461 | (64, 1, 1, 3000), |
| 462 | (16, 6, 2030, 2070), # MTP-style rows straddling index_topk |
| 463 | (4096, 1, 1, 3000), # clamp spans several program blocks |
| 464 | ], |
| 465 | ) |
| 466 | def test_upper_bound_plan_matches_exact(num_reqs, q_per_req, lo, hi): |
| 467 | """Sync-free planning from optimistic host bounds (async spec decode) |
| 468 | must match exact planning: the device clamp trims every work item back |
| 469 | to its row's valid count, so no -1 slot is read.""" |
| 470 | torch.manual_seed(0) |
| 471 | heads, topk, kpool, width = SM90_HEADS, SM90_TOPK, SM90_KPOOL, SM90_WIDTH |
| 472 | ctx, valid, cam, req_id, slots = _clamp_batch( |
| 473 | num_reqs, q_per_req, lo, hi, topk, kpool, width |
| 474 | ) |
| 475 | rows = ctx.numel() |
| 476 | |
| 477 | kv = torch.randn(1 << 16, 1, HEAD, device="cuda", dtype=torch.bfloat16) |
| 478 | q = torch.randn(rows, heads, HEAD, device="cuda", dtype=torch.bfloat16) |
| 479 | q_pe = q.new_empty(rows, heads, 0) |
| 480 | |
| 481 | def run(state): |
| 482 | state.pack_indices(slots) |
| 483 | return state.wrapper.run(q, q_pe, kv, kv[..., :0]) |
| 484 | |
| 485 | exact = _real_state(rows) |
| 486 | exact.plan(rows, valid, None, None) |
| 487 | ub = _real_state(rows) |
| 488 | host_bound = torch.clamp(ctx + torch.randint(0, 7, (rows,)), max=ub.max_valid) |
| 489 | ub.plan(rows, host_bound, cam, req_id) |
| 490 | out = run(ub) |
| 491 | assert not out.isnan().any() |
| 492 | torch.testing.assert_close(out, run(exact), atol=1e-2, rtol=1e-2) |
| 493 | |
| 494 | |
| 495 | @pytest.mark.skipif( |
| 496 | not _sm90_flashinfer_available(), reason="Requires SM90 FlashInfer MLA" |
| 497 | ) |
| 498 | def test_upper_bound_plan_reused_across_steps(monkeypatch): |
| 499 | """Growing contexts (draft/decode steps) reuse the padded plan and only |
| 500 | re-clamp on device, matching an exact replan every step.""" |
| 501 | torch.manual_seed(0) |
| 502 | heads, topk, kpool, width = SM90_HEADS, SM90_TOPK, SM90_KPOOL, SM90_WIDTH |
| 503 | rows = 48 |
| 504 | ctx, _, _, req_id, _ = _clamp_batch(rows, 1, 1, 3000, topk, kpool, width) |
| 505 | |
| 506 | exact, ub = _real_state(rows), _real_state(rows) |
| 507 | replans: list[int] = [] |
| 508 | plan_fn = ub._plan |
| 509 | |
| 510 | def counting_plan(num_tokens, kv_lens): |
| 511 | replans.append(num_tokens) |
| 512 | plan_fn(num_tokens, kv_lens) |
| 513 | |
| 514 | monkeypatch.setattr(ub, "_plan", counting_plan) |
| 515 | kv = torch.randn(1 << 16, 1, HEAD, device="cuda", dtype=torch.bfloat16) |
| 516 | for step in range(5): |
| 517 | step_ctx = ctx + step |
| 518 | valid = torch.where(step_ctx <= topk, step_ctx, topk + step_ctx % kpool) |
| 519 | slots = torch.randint(0, 1 << 16, (rows, width), dtype=torch.int32) |
| 520 | slots[torch.arange(width) >= valid[:, None]] = -1 |
| 521 | slots = slots.cuda() |
| 522 | cam = SimpleNamespace( |
| 523 | seq_lens=step_ctx.to(torch.int32).cuda(), |
| 524 | query_start_loc=torch.arange(rows + 1, dtype=torch.int32).cuda(), |
| 525 | ) |
| 526 | num_plans = len(replans) |
| 527 | ub.plan(rows, torch.clamp(step_ctx + 5, max=ub.max_valid), cam, req_id) |
| 528 | assert len(replans) - num_plans == (1 if step == 0 else 0) |
| 529 | exact.plan(rows, valid.to(torch.int32), None, None) |
| 530 | q = torch.randn(rows, heads, HEAD, device="cuda", dtype=torch.bfloat16) |
| 531 | outs = [] |
| 532 | for state in (ub, exact): |
| 533 | state.pack_indices(slots) |
| 534 | outs.append(state.wrapper.run(q, q[..., :0], kv, kv[..., :0])) |
| 535 | torch.testing.assert_close(outs[0], outs[1], atol=1e-2, rtol=1e-2) |
| 536 | |
| 537 | |
| 538 | def test_supports_combination_gates(monkeypatch, default_vllm_config): |
| 539 | monkeypatch.setattr(sm90_mod, "has_flashinfer_sm90_nope_mla", lambda: True) |
| 540 | call = lambda **kw: FlashInferMLASparseSM90Backend.supports_combination( |
| 541 | head_size=576, |
| 542 | dtype=torch.bfloat16, |
| 543 | kv_cache_dtype="fp8_e4m3", |
| 544 | block_size=64, |
| 545 | use_mla=True, |
| 546 | has_sink=False, |
| 547 | use_sparse=True, |
| 548 | use_mm_prefix=False, |
| 549 | device_capability=SimpleNamespace(major=9), |
| 550 | **kw, |
| 551 | ) |
| 552 | assert call() is None # no model config: only the feature gate applies |
| 553 | |
| 554 | import vllm.config as cfg |
| 555 | |
| 556 | monkeypatch.setattr( |
| 557 | cfg, |
| 558 | "get_current_vllm_config", |
| 559 | lambda: SimpleNamespace(model_config=None), |
| 560 | ) |
| 561 | assert call() is None |
| 562 | monkeypatch.setattr(sm90_mod, "has_flashinfer_sm90_nope_mla", lambda: False) |
| 563 | assert "requires FlashInfer" in (call() or "") |