| 1 | # SPDX-License-Identifier: Apache-2.0 |
| 2 | # SPDX-FileCopyrightText: Copyright contributors to the vLLM project |
| 3 | """Test: |
| 4 | |
| 5 | * Tests for MMEncoderAttention layer |
| 6 | """ |
| 7 | |
| 8 | import itertools |
| 9 | from unittest.mock import patch |
| 10 | |
| 11 | import numpy as np |
| 12 | import pytest |
| 13 | import torch |
| 14 | |
| 15 | from vllm.config import get_current_vllm_config |
| 16 | from vllm.config.multimodal import MultiModalConfig |
| 17 | from vllm.model_executor.layers.attention import MMEncoderAttention |
| 18 | from vllm.platforms import current_platform |
| 19 | from vllm.platforms.cpu import CpuPlatform |
| 20 | from vllm.platforms.cuda import CudaPlatform |
| 21 | from vllm.platforms.interface import DeviceCapability |
| 22 | from vllm.platforms.rocm import RocmPlatform, on_mi3xx |
| 23 | from vllm.utils.torch_utils import set_default_torch_dtype, set_random_seed |
| 24 | from vllm.v1.attention.backends.registry import AttentionBackendEnum |
| 25 | from vllm.v1.attention.selector import _cached_get_attn_backend |
| 26 | |
| 27 | |
| 28 | @pytest.fixture(autouse=True) |
| 29 | def reset_test_state(): |
| 30 | """Clear cached selectors and restore process-wide torch defaults.""" |
| 31 | default_device = torch.get_default_device() |
| 32 | default_dtype = torch.get_default_dtype() |
| 33 | _cached_get_attn_backend.cache_clear() |
| 34 | try: |
| 35 | yield |
| 36 | finally: |
| 37 | torch.set_default_device(default_device) |
| 38 | torch.set_default_dtype(default_dtype) |
| 39 | _cached_get_attn_backend.cache_clear() |
| 40 | |
| 41 | |
| 42 | devices = ["cpu"] |
| 43 | if current_platform.is_cuda(): |
| 44 | devices.append("cuda") |
| 45 | if current_platform.is_rocm(): |
| 46 | devices.append("hip") |
| 47 | |
| 48 | |
| 49 | @pytest.mark.parametrize("device", devices) |
| 50 | def test_mha_attn_platform(default_vllm_config, device: str): |
| 51 | """Test the attention selector between different platform and device.""" |
| 52 | torch.set_default_dtype(torch.float16) |
| 53 | |
| 54 | if device == "cpu": |
| 55 | with ( |
| 56 | patch("vllm.model_executor.models.vision.current_platform", CpuPlatform()), |
| 57 | ): |
| 58 | attn = MMEncoderAttention(16, 64, scale=1) |
| 59 | assert attn.attn_backend == AttentionBackendEnum.TORCH_SDPA |
| 60 | elif device == "hip": |
| 61 | with ( |
| 62 | patch("vllm.model_executor.models.vision.current_platform", RocmPlatform()), |
| 63 | ): |
| 64 | attn = MMEncoderAttention(16, 64, scale=1) |
| 65 | assert attn.attn_backend == AttentionBackendEnum.FLASH_ATTN |
| 66 | else: |
| 67 | # Test CUDA with head_size=64 (divisible by 32) |
| 68 | # - should use vLLM's FlashAttention |
| 69 | with ( |
| 70 | patch("vllm.model_executor.models.vision.current_platform", CudaPlatform()), |
| 71 | ): |
| 72 | attn = MMEncoderAttention(16, 64, scale=1) |
| 73 | assert attn.attn_backend == AttentionBackendEnum.FLASH_ATTN |
| 74 | |
| 75 | # Test CUDA with head_size=72 (not divisible by 32) |
| 76 | # - should use vLLM's FlashAttention |
| 77 | with ( |
| 78 | patch("vllm.model_executor.models.vision.current_platform", CudaPlatform()), |
| 79 | ): |
| 80 | attn = MMEncoderAttention(16, 72, scale=1) |
| 81 | assert attn.attn_backend == AttentionBackendEnum.FLASH_ATTN |
| 82 | |
| 83 | # Test CUDA with head_size=72 (not divisible by 32) |
| 84 | # - should use vLLM's FlashAttention |
| 85 | with ( |
| 86 | patch("vllm.model_executor.models.vision.current_platform", CudaPlatform()), |
| 87 | set_default_torch_dtype(torch.float32), |
| 88 | ): |
| 89 | attn = MMEncoderAttention(16, 72, scale=1) |
| 90 | assert attn.attn_backend == AttentionBackendEnum.TRITON_ATTN |
| 91 | |
| 92 | # Test Turing (pre-Ampere, sm_75): FlashAttention requires sm>=80, |
| 93 | # and Triton no longer supports MMA on Turing, so we expect that |
| 94 | # TORCH_SDPA is used for MMEncoderAttention. |
| 95 | with ( |
| 96 | patch("vllm.model_executor.models.vision.current_platform", CudaPlatform()), |
| 97 | patch.object( |
| 98 | CudaPlatform, |
| 99 | "get_device_capability", |
| 100 | return_value=DeviceCapability(major=7, minor=5), |
| 101 | ), |
| 102 | ): |
| 103 | attn = MMEncoderAttention(16, 64, scale=1) |
| 104 | assert attn.attn_backend == AttentionBackendEnum.TORCH_SDPA |
| 105 | |
| 106 | |
| 107 | def ref_attention( |
| 108 | query: torch.Tensor, |
| 109 | key: torch.Tensor, |
| 110 | value: torch.Tensor, |
| 111 | scale: float, |
| 112 | ) -> torch.Tensor: |
| 113 | """Native implementation of scaled dot product attention without mask: |
| 114 | - query, key, value: [batch_size, seq_len, num_heads, head_size] |
| 115 | - attn_mask: [batch_size, seq_len, seq_len] |
| 116 | """ |
| 117 | query, key, value = (x.transpose(1, 2) for x in (query, key, value)) |
| 118 | attn_weights = scale * torch.matmul(query, key.transpose(2, 3)) |
| 119 | attn_weights = torch.softmax(attn_weights, dim=-1).to(value.dtype) |
| 120 | out = torch.matmul(attn_weights, value).transpose(1, 2) |
| 121 | return out |
| 122 | |
| 123 | |
| 124 | BATCH_SIZES = [1, 16] |
| 125 | SEQ_LENS = [1] |
| 126 | VAR_SEQ_LENS = [ |
| 127 | [2, 2], |
| 128 | [2, 3, 4], |
| 129 | ] |
| 130 | NUM_HEADS = [1, 16] |
| 131 | NUM_KV_HEADS = [1] |
| 132 | HEAD_SIZES = [64, 80] |
| 133 | # flshattF and tritonflashattF supported: {torch.float16, torch.bfloat16} |
| 134 | DTYPES = ( |
| 135 | [torch.half, torch.bfloat16, torch.float] |
| 136 | if not current_platform.is_rocm() |
| 137 | else [torch.half, torch.bfloat16] |
| 138 | ) |
| 139 | CUDA_DEVICES = ["cuda"] |
| 140 | |
| 141 | |
| 142 | @pytest.mark.parametrize("batch_size", BATCH_SIZES) |
| 143 | @pytest.mark.parametrize("seq_len", SEQ_LENS) |
| 144 | @pytest.mark.parametrize("num_heads", NUM_HEADS) |
| 145 | @pytest.mark.parametrize("num_kv_heads", NUM_KV_HEADS) |
| 146 | @pytest.mark.parametrize("head_size", HEAD_SIZES) |
| 147 | @pytest.mark.parametrize("dtype", DTYPES) |
| 148 | @pytest.mark.parametrize("device", CUDA_DEVICES) |
| 149 | def test_mha_attn_forward( |
| 150 | default_vllm_config, |
| 151 | batch_size: int, |
| 152 | seq_len: int, |
| 153 | num_heads: int, |
| 154 | num_kv_heads: int, |
| 155 | head_size: int, |
| 156 | dtype: torch.dtype, |
| 157 | device: str, |
| 158 | ): |
| 159 | set_random_seed(0) |
| 160 | torch.set_default_device(device) |
| 161 | torch.set_default_dtype(dtype) |
| 162 | |
| 163 | q = torch.randn(batch_size, seq_len, num_heads * head_size) |
| 164 | k = torch.randn(batch_size, seq_len, num_kv_heads * head_size) |
| 165 | v = torch.randn(batch_size, seq_len, num_kv_heads * head_size) |
| 166 | scale = 1.0 / head_size**0.5 |
| 167 | attn = MMEncoderAttention( |
| 168 | num_heads, head_size, scale=scale, num_kv_heads=num_kv_heads |
| 169 | ) |
| 170 | output = attn(q, k, v) |
| 171 | |
| 172 | assert num_heads % num_kv_heads == 0 |
| 173 | num_queries_per_kv = num_heads // num_kv_heads |
| 174 | q = q.reshape(batch_size, seq_len, num_heads, head_size) |
| 175 | k = k.reshape(batch_size, seq_len, num_kv_heads, head_size) |
| 176 | v = v.reshape(batch_size, seq_len, num_kv_heads, head_size) |
| 177 | if num_queries_per_kv > 1: |
| 178 | k = torch.repeat_interleave(k, num_queries_per_kv, dim=2) |
| 179 | v = torch.repeat_interleave(v, num_queries_per_kv, dim=2) |
| 180 | |
| 181 | ref_output = ref_attention( |
| 182 | q, |
| 183 | k, |
| 184 | v, |
| 185 | scale=scale, |
| 186 | ).reshape(batch_size, seq_len, num_heads * head_size) |
| 187 | tol_kwargs = ( |
| 188 | dict(rtol=1e-3, atol=1e-3) |
| 189 | if attn.attn_backend == AttentionBackendEnum.TRITON_ATTN |
| 190 | else {} |
| 191 | ) |
| 192 | torch.testing.assert_close(output, ref_output, **tol_kwargs) |
| 193 | |
| 194 | |
| 195 | @pytest.mark.parametrize("var_seq_len", VAR_SEQ_LENS) |
| 196 | @pytest.mark.parametrize("num_heads", NUM_HEADS) |
| 197 | @pytest.mark.parametrize("num_kv_heads", NUM_KV_HEADS) |
| 198 | @pytest.mark.parametrize("head_size", HEAD_SIZES) |
| 199 | @pytest.mark.parametrize("dtype", DTYPES) |
| 200 | @pytest.mark.parametrize("device", CUDA_DEVICES) |
| 201 | def test_mha_attn_varlen_forward( |
| 202 | default_vllm_config, |
| 203 | var_seq_len: list[int], |
| 204 | num_heads: int, |
| 205 | num_kv_heads: int, |
| 206 | head_size: int, |
| 207 | dtype: torch.dtype, |
| 208 | device: str, |
| 209 | ): |
| 210 | set_random_seed(0) |
| 211 | torch.set_default_device(device) |
| 212 | torch.set_default_dtype(dtype) |
| 213 | |
| 214 | q = torch.randn(1, sum(var_seq_len), num_heads, head_size) |
| 215 | k = torch.randn(1, sum(var_seq_len), num_kv_heads, head_size) |
| 216 | v = torch.randn(1, sum(var_seq_len), num_kv_heads, head_size) |
| 217 | cu_seqlens = torch.tensor( |
| 218 | [0] + list(itertools.accumulate(var_seq_len)), dtype=torch.int32 |
| 219 | ) |
| 220 | scale = 1.0 / head_size**0.5 |
| 221 | attn = MMEncoderAttention( |
| 222 | num_heads, head_size, scale=scale, num_kv_heads=num_kv_heads |
| 223 | ) |
| 224 | output = attn( |
| 225 | q, k, v, cu_seqlens=cu_seqlens, max_seqlen=torch.tensor(max(var_seq_len)) |
| 226 | ) |
| 227 | |
| 228 | assert num_heads % num_kv_heads == 0 |
| 229 | num_queries_per_kv = num_heads // num_kv_heads |
| 230 | if num_queries_per_kv > 1: |
| 231 | k = torch.repeat_interleave(k, num_queries_per_kv, dim=2) |
| 232 | v = torch.repeat_interleave(v, num_queries_per_kv, dim=2) |
| 233 | |
| 234 | ref_output = [] |
| 235 | for q_i, k_i, v_i in zip( |
| 236 | torch.split(q, var_seq_len, dim=1), |
| 237 | torch.split(k, var_seq_len, dim=1), |
| 238 | torch.split(v, var_seq_len, dim=1), |
| 239 | ): |
| 240 | output_i = ref_attention( |
| 241 | q_i, |
| 242 | k_i, |
| 243 | v_i, |
| 244 | scale=scale, |
| 245 | ) |
| 246 | ref_output.append(output_i) |
| 247 | ref_output = torch.cat(ref_output, dim=1) |
| 248 | torch.testing.assert_close(output, ref_output, atol=1e-2, rtol=1e-2) |
| 249 | |
| 250 | |
| 251 | @pytest.mark.parametrize("var_seq_len", VAR_SEQ_LENS) |
| 252 | @pytest.mark.parametrize( |
| 253 | "dtype", |
| 254 | [torch.bfloat16, torch.half], |
| 255 | ) |
| 256 | @pytest.mark.parametrize("device", CUDA_DEVICES) |
| 257 | def test_mha_attn_varlen_forward_flashinfer( |
| 258 | default_vllm_config, |
| 259 | var_seq_len: list[int], |
| 260 | dtype: torch.dtype, |
| 261 | device: str, |
| 262 | ): |
| 263 | """Test MMEncoderAttention varlen forward with FLASHINFER backend (head_size=72). |
| 264 | |
| 265 | Exercises the path that uses --mm-encoder-attn-backend=FLASHINFER with |
| 266 | recomputed cu_seqlens, max_seqlen, and sequence_lengths as in qwen3_vl |
| 267 | vision encoder. |
| 268 | """ |
| 269 | pytest.importorskip("flashinfer") |
| 270 | |
| 271 | num_heads = 16 |
| 272 | head_size = 72 |
| 273 | set_random_seed(0) |
| 274 | torch.set_default_device(device) |
| 275 | torch.set_default_dtype(dtype) |
| 276 | |
| 277 | # Override vllm config so get_vit_attn_backend returns FLASHINFER (simulates |
| 278 | # --mm-encoder-attn-backend=FLASHINFER). |
| 279 | vllm_config = get_current_vllm_config() |
| 280 | old_model_config = vllm_config.model_config |
| 281 | minimal_model_config = type( |
| 282 | "MinimalModelConfig", |
| 283 | (), |
| 284 | { |
| 285 | "multimodal_config": MultiModalConfig( |
| 286 | mm_encoder_attn_backend=AttentionBackendEnum.FLASHINFER |
| 287 | ), |
| 288 | }, |
| 289 | )() |
| 290 | vllm_config.model_config = minimal_model_config |
| 291 | try: |
| 292 | total_len = sum(var_seq_len) |
| 293 | # Stride of second dim = 3 * num_heads * head_size (same as qwen2_5_vl |
| 294 | # after qkv rearrange and unbind: qkv shape (b, s, 3, head, head_dim)). |
| 295 | qkv = torch.randn(1, total_len, 3, num_heads, head_size) |
| 296 | q, k, v = qkv.unbind(dim=2) |
| 297 | |
| 298 | cu_seqlens_np = np.array( |
| 299 | [0] + list(itertools.accumulate(var_seq_len)), dtype=np.int32 |
| 300 | ) |
| 301 | hidden_size = num_heads * head_size |
| 302 | tp_size = 1 |
| 303 | |
| 304 | sequence_lengths = MMEncoderAttention.maybe_compute_seq_lens( |
| 305 | AttentionBackendEnum.FLASHINFER, |
| 306 | cu_seqlens_np, |
| 307 | device, |
| 308 | ) |
| 309 | |
| 310 | max_seqlen_val = MMEncoderAttention.compute_max_seqlen( |
| 311 | AttentionBackendEnum.FLASHINFER, cu_seqlens_np |
| 312 | ) |
| 313 | max_seqlen = torch.tensor(max_seqlen_val, device=device, dtype=torch.int32) |
| 314 | |
| 315 | cu_seqlens = MMEncoderAttention.maybe_recompute_cu_seqlens( |
| 316 | AttentionBackendEnum.FLASHINFER, |
| 317 | cu_seqlens_np, |
| 318 | hidden_size, |
| 319 | tp_size, |
| 320 | device, |
| 321 | ) |
| 322 | |
| 323 | scale = 1.0 / head_size**0.5 |
| 324 | attn = MMEncoderAttention( |
| 325 | num_heads, |
| 326 | head_size, |
| 327 | scale=scale, |
| 328 | num_kv_heads=num_heads, |
| 329 | ) |
| 330 | assert attn.attn_backend == AttentionBackendEnum.FLASHINFER |
| 331 | |
| 332 | output = attn( |
| 333 | q, |
| 334 | k, |
| 335 | v, |
| 336 | cu_seqlens=cu_seqlens, |
| 337 | max_seqlen=max_seqlen, |
| 338 | sequence_lengths=sequence_lengths, |
| 339 | ) |
| 340 | |
| 341 | ref_output = [] |
| 342 | for q_i, k_i, v_i in zip( |
| 343 | torch.split(q, var_seq_len, dim=1), |
| 344 | torch.split(k, var_seq_len, dim=1), |
| 345 | torch.split(v, var_seq_len, dim=1), |
| 346 | ): |
| 347 | output_i = ref_attention(q_i, k_i, v_i, scale=scale) |
| 348 | ref_output.append(output_i) |
| 349 | ref_output = torch.cat(ref_output, dim=1) |
| 350 | torch.testing.assert_close(output, ref_output, atol=1e-2, rtol=1e-2) |
| 351 | finally: |
| 352 | vllm_config.model_config = old_model_config |
| 353 | |
| 354 | |
| 355 | @pytest.mark.skipif( |
| 356 | not current_platform.is_rocm() or not on_mi3xx(), |
| 357 | reason="AITER FP8 attention requires MI300/MI350", |
| 358 | ) |
| 359 | @pytest.mark.parametrize("var_seq_len", [[128, 193], [256, 384, 511]]) |
| 360 | @pytest.mark.parametrize("head_size", [64, 72, 80]) |
| 361 | @pytest.mark.parametrize("dtype", [torch.bfloat16, torch.half]) |
| 362 | def test_mha_attn_varlen_forward_aiter_fp8( |
| 363 | default_vllm_config, |
| 364 | var_seq_len: list[int], |
| 365 | head_size: int, |
| 366 | dtype: torch.dtype, |
| 367 | ): |
| 368 | """Compare packed AITER FP8 ViT attention with a BF16/FP16 reference.""" |
| 369 | aiter = pytest.importorskip("aiter") |
| 370 | if not hasattr(aiter, "flash_attn_varlen_fp8_pertensor_func"): |
| 371 | pytest.skip("Installed AITER does not provide varlen FP8 attention") |
| 372 | |
| 373 | num_heads = 4 |
| 374 | set_random_seed(0) |
| 375 | torch.set_default_device("cuda") |
| 376 | torch.set_default_dtype(dtype) |
| 377 | |
| 378 | vllm_config = get_current_vllm_config() |
| 379 | old_model_config = vllm_config.model_config |
| 380 | minimal_model_config = type( |
| 381 | "MinimalModelConfig", |
| 382 | (), |
| 383 | { |
| 384 | "multimodal_config": MultiModalConfig( |
| 385 | mm_encoder_attn_backend=AttentionBackendEnum.ROCM_AITER_FA, |
| 386 | mm_encoder_attn_dtype="fp8", |
| 387 | ), |
| 388 | }, |
| 389 | )() |
| 390 | vllm_config.model_config = minimal_model_config |
| 391 | try: |
| 392 | total_len = sum(var_seq_len) |
| 393 | # Keep Q/K/V as non-contiguous views, matching interleaved ViT QKV. |
| 394 | qkv = torch.randn(1, total_len, 3, num_heads, head_size) |
| 395 | q, k, v = qkv.unbind(dim=2) |
| 396 | cu_seqlens = torch.tensor( |
| 397 | [0] + list(itertools.accumulate(var_seq_len)), dtype=torch.int32 |
| 398 | ) |
| 399 | max_seqlen = torch.tensor(max(var_seq_len), dtype=torch.int32) |
| 400 | scale = 1.0 / head_size**0.5 |
| 401 | |
| 402 | attn = MMEncoderAttention(num_heads, head_size, scale=scale) |
| 403 | assert attn.attn_backend == AttentionBackendEnum.ROCM_AITER_FA |
| 404 | assert attn.fp8_enabled |
| 405 | output = attn( |
| 406 | q, |
| 407 | k, |
| 408 | v, |
| 409 | cu_seqlens=cu_seqlens, |
| 410 | max_seqlen=max_seqlen, |
| 411 | ) |
| 412 | |
| 413 | ref_output = torch.cat( |
| 414 | [ |
| 415 | ref_attention(q_i, k_i, v_i, scale=scale) |
| 416 | for q_i, k_i, v_i in zip( |
| 417 | torch.split(q, var_seq_len, dim=1), |
| 418 | torch.split(k, var_seq_len, dim=1), |
| 419 | torch.split(v, var_seq_len, dim=1), |
| 420 | ) |
| 421 | ], |
| 422 | dim=1, |
| 423 | ) |
| 424 | diff = (output.float() - ref_output.float()).abs() |
| 425 | cosine = torch.nn.functional.cosine_similarity( |
| 426 | output.float().flatten(), |
| 427 | ref_output.float().flatten(), |
| 428 | dim=0, |
| 429 | ) |
| 430 | |
| 431 | assert output.shape == ref_output.shape |
| 432 | assert output.dtype == dtype |
| 433 | assert cosine.item() >= 0.99 |
| 434 | assert diff.mean().item() <= 0.03 |
| 435 | assert diff.max().item() <= 0.30 |
| 436 | finally: |
| 437 | vllm_config.model_config = old_model_config |
| 438 | |
| 439 | |
| 440 | @pytest.mark.skipif( |
| 441 | not current_platform.is_rocm(), reason="AITER FP8 attention requires ROCm" |
| 442 | ) |
| 443 | def test_mha_attn_aiter_fp8_rejects_unsupported_arch(default_vllm_config): |
| 444 | vllm_config = get_current_vllm_config() |
| 445 | old_model_config = vllm_config.model_config |
| 446 | vllm_config.model_config = type( |
| 447 | "MinimalModelConfig", |
| 448 | (), |
| 449 | { |
| 450 | "multimodal_config": MultiModalConfig( |
| 451 | mm_encoder_attn_backend=AttentionBackendEnum.ROCM_AITER_FA, |
| 452 | mm_encoder_attn_dtype="fp8", |
| 453 | ), |
| 454 | }, |
| 455 | )() |
| 456 | try: |
| 457 | with ( |
| 458 | patch("vllm.platforms.rocm.on_mi3xx", return_value=False), |
| 459 | pytest.raises(ValueError, match="gfx942 or gfx950"), |
| 460 | ): |
| 461 | MMEncoderAttention(4, 64) |
| 462 | finally: |
| 463 | vllm_config.model_config = old_model_config |
| 464 | |
| 465 | |
| 466 | @pytest.mark.skipif( |
| 467 | not current_platform.is_rocm(), reason="AITER FP8 attention requires ROCm" |
| 468 | ) |
| 469 | def test_mha_attn_fp8_rejects_wrong_backend(default_vllm_config): |
| 470 | vllm_config = get_current_vllm_config() |
| 471 | old_model_config = vllm_config.model_config |
| 472 | vllm_config.model_config = type( |
| 473 | "MinimalModelConfig", |
| 474 | (), |
| 475 | { |
| 476 | "multimodal_config": MultiModalConfig( |
| 477 | mm_encoder_attn_backend=AttentionBackendEnum.FLASH_ATTN, |
| 478 | mm_encoder_attn_dtype="fp8", |
| 479 | ), |
| 480 | }, |
| 481 | )() |
| 482 | try: |
| 483 | with pytest.raises(ValueError, match="requires either"): |
| 484 | MMEncoderAttention(4, 64) |
| 485 | finally: |
| 486 | vllm_config.model_config = old_model_config |