| 1 | # SPDX-License-Identifier: Apache-2.0 |
| 2 | # SPDX-FileCopyrightText: Copyright contributors to the vLLM project |
| 3 | import copy |
| 4 | |
| 5 | import pytest |
| 6 | import torch._dynamo |
| 7 | |
| 8 | from tests.compile.backend import LazyInitPass, TestBackend |
| 9 | from tests.utils import TestFP8Layer, flat_product |
| 10 | from tests.v1.attention.utils import ( |
| 11 | BatchSpec, |
| 12 | create_common_attn_metadata, |
| 13 | dense_kv_cache_views, |
| 14 | ) |
| 15 | from vllm._aiter_ops import is_aiter_found_and_supported, rocm_aiter_ops |
| 16 | from vllm._custom_ops import cutlass_scaled_fp4_mm, scaled_fp4_quant |
| 17 | from vllm.compilation.passes.fusion.attn_quant_fusion import ( |
| 18 | ATTN_OP, |
| 19 | AttnQuantFusionPass, |
| 20 | ) |
| 21 | from vllm.compilation.passes.fusion.matcher_utils import QUANT_OPS |
| 22 | from vllm.compilation.passes.fx_utils import find_op_nodes |
| 23 | from vllm.compilation.passes.utility.noop_elimination import NoOpEliminationPass |
| 24 | from vllm.compilation.passes.utility.post_cleanup import PostCleanupPass |
| 25 | from vllm.config import ( |
| 26 | AttentionConfig, |
| 27 | CacheConfig, |
| 28 | CompilationConfig, |
| 29 | CompilationMode, |
| 30 | ModelConfig, |
| 31 | PassConfig, |
| 32 | SchedulerConfig, |
| 33 | VllmConfig, |
| 34 | set_current_vllm_config, |
| 35 | ) |
| 36 | from vllm.config.cache import CacheDType |
| 37 | from vllm.forward_context import get_forward_context, set_forward_context |
| 38 | from vllm.model_executor.layers.attention import Attention |
| 39 | from vllm.model_executor.layers.quantization.utils.quant_utils import ( |
| 40 | QuantKey, |
| 41 | kFp8StaticTensorSym, |
| 42 | kNvfp4Dynamic, |
| 43 | ) |
| 44 | from vllm.platforms import current_platform |
| 45 | from vllm.utils.flashinfer import has_flashinfer |
| 46 | from vllm.v1.attention.backend import AttentionMetadata |
| 47 | from vllm.v1.attention.backends.registry import AttentionBackendEnum |
| 48 | from vllm.v1.attention.backends.utils import ( |
| 49 | get_supported_kv_cache_layouts, |
| 50 | resolve_kv_cache_layout, |
| 51 | ) |
| 52 | from vllm.v1.kv_cache_interface import ( |
| 53 | AttentionSpec, |
| 54 | get_kv_quant_mode, |
| 55 | ) |
| 56 | |
| 57 | DEVICE_TYPE = current_platform.device_type |
| 58 | FP8_DTYPE = current_platform.fp8_dtype() |
| 59 | FP4_DTYPE = torch.uint8 |
| 60 | |
| 61 | |
| 62 | class AttentionQuantPatternModel(torch.nn.Module): |
| 63 | """Base model for AttentionQuantPattern fusion.""" |
| 64 | |
| 65 | def __init__( |
| 66 | self, |
| 67 | num_qo_heads: int, |
| 68 | num_kv_heads: int, |
| 69 | head_size: int, |
| 70 | device: torch.device, |
| 71 | vllm_config: VllmConfig, |
| 72 | block_size: int, |
| 73 | **kwargs, |
| 74 | ): |
| 75 | super().__init__() |
| 76 | self.num_qo_heads = num_qo_heads |
| 77 | self.num_kv_heads = num_kv_heads |
| 78 | self.head_size = head_size |
| 79 | self.device = device |
| 80 | self.vllm_config = vllm_config |
| 81 | self.dtype = vllm_config.model_config.dtype |
| 82 | |
| 83 | self.attn = Attention( |
| 84 | num_heads=self.num_qo_heads, |
| 85 | head_size=self.head_size, |
| 86 | scale=1.0 / (self.head_size**0.5), |
| 87 | num_kv_heads=self.num_kv_heads, |
| 88 | cache_config=vllm_config.cache_config, |
| 89 | prefix="model.layers.0.self_attn.attn", |
| 90 | ) |
| 91 | self.attn._k_scale = self.attn._k_scale.to(device) |
| 92 | self.attn._v_scale = self.attn._v_scale.to(device) |
| 93 | |
| 94 | self.block_size = block_size |
| 95 | |
| 96 | # Initialize attn MetadataBuilder (match Attention.get_kv_cache_spec) |
| 97 | self.builder = self.attn.attn_backend.get_builder_cls()( |
| 98 | kv_cache_spec=AttentionSpec( |
| 99 | block_size=self.block_size, |
| 100 | num_kv_heads=self.num_kv_heads, |
| 101 | head_size=self.head_size, |
| 102 | dtype=self.attn.kv_cache_torch_dtype, |
| 103 | kv_quant_mode=get_kv_quant_mode(self.attn.kv_cache_dtype), |
| 104 | ), |
| 105 | layer_names=[self.attn.layer_name], |
| 106 | vllm_config=self.vllm_config, |
| 107 | device=self.device, |
| 108 | ) |
| 109 | |
| 110 | def build_attn_metadata(self, batch_size: int) -> AttentionMetadata: |
| 111 | """Initialize attention metadata.""" |
| 112 | # TODO (Rohan138) reuse utils from vllm/v1/worker/gpu/attn_utils.py |
| 113 | |
| 114 | # Create common attn metadata |
| 115 | batch_spec = BatchSpec(seq_lens=[1] * batch_size, query_lens=[1] * batch_size) |
| 116 | common_attn_metadata = create_common_attn_metadata( |
| 117 | batch_spec, self.block_size, self.device, arange_block_indices=True |
| 118 | ) |
| 119 | |
| 120 | max_blocks = (max(batch_spec.seq_lens) + self.block_size - 1) // self.block_size |
| 121 | num_blocks = batch_size * max_blocks |
| 122 | |
| 123 | spec = self.attn.attn_backend.customize_spec( |
| 124 | AttentionSpec( |
| 125 | block_size=self.block_size, |
| 126 | num_kv_heads=self.num_kv_heads, |
| 127 | head_size=self.head_size, |
| 128 | dtype=self.attn.kv_cache_torch_dtype, |
| 129 | kv_quant_mode=get_kv_quant_mode(self.attn.kv_cache_dtype), |
| 130 | ) |
| 131 | ) |
| 132 | supported = get_supported_kv_cache_layouts([self.attn.attn_backend]) |
| 133 | layout = resolve_kv_cache_layout( |
| 134 | self.vllm_config, [[m.name for m in supported]] |
| 135 | ) |
| 136 | raw_tensor = torch.zeros( |
| 137 | num_blocks * spec.page_size_bytes, |
| 138 | dtype=torch.int8, |
| 139 | device=self.device, |
| 140 | ) |
| 141 | kv_cache = dense_kv_cache_views( |
| 142 | raw_tensor, |
| 143 | spec, |
| 144 | num_blocks, |
| 145 | num_layers=1, |
| 146 | layout=layout, |
| 147 | )[0] |
| 148 | |
| 149 | self.attn.kv_cache = kv_cache |
| 150 | |
| 151 | # Build attn metadata |
| 152 | self.attn_metadata = self.builder.build( |
| 153 | common_prefix_len=0, common_attn_metadata=common_attn_metadata |
| 154 | ) |
| 155 | |
| 156 | return self.attn_metadata |
| 157 | |
| 158 | |
| 159 | class TestAttentionFp8StaticQuantPatternModel(AttentionQuantPatternModel): |
| 160 | """Test model for AttentionFp8StaticQuantPattern fusion.""" |
| 161 | |
| 162 | quant_key = kFp8StaticTensorSym |
| 163 | |
| 164 | def __init__(self, *args, **kwargs): |
| 165 | super().__init__(*args, **kwargs) |
| 166 | |
| 167 | hidden_size = self.num_qo_heads * self.head_size |
| 168 | self.fp8_linear = TestFP8Layer( |
| 169 | weight_shape=(hidden_size, hidden_size), |
| 170 | activation_quant_key=self.quant_key, |
| 171 | weight_quant_key=self.quant_key, |
| 172 | device=self.device, |
| 173 | input_dtype=self.dtype, |
| 174 | ) |
| 175 | |
| 176 | w = kwargs.get("w") |
| 177 | if w is not None: |
| 178 | self.fp8_linear.weight = w["weight"] |
| 179 | self.fp8_linear.weight_scale = w["wscale"] |
| 180 | self.fp8_linear.input_scale = w["scale"] |
| 181 | |
| 182 | self.w = { |
| 183 | "weight": self.fp8_linear.weight, |
| 184 | "wscale": self.fp8_linear.weight_scale, |
| 185 | "scale": self.fp8_linear.input_scale, |
| 186 | } |
| 187 | |
| 188 | def forward(self, q: torch.Tensor, k: torch.Tensor, v: torch.Tensor): |
| 189 | """Forward pass that creates the pattern to be fused.""" |
| 190 | attn_output = self.attn(q, k, v) |
| 191 | return self.fp8_linear(attn_output) |
| 192 | |
| 193 | |
| 194 | class TestAttentionAiterFp8StaticQuantPatternModel( |
| 195 | TestAttentionFp8StaticQuantPatternModel |
| 196 | ): |
| 197 | aiter_quant = True |
| 198 | |
| 199 | def __init__(self, *args, **kwargs): |
| 200 | super().__init__(*args, **kwargs) |
| 201 | self.fp8_linear.kernel.quant_fp8.use_aiter = True |
| 202 | assert self.fp8_linear.input_scale is not None |
| 203 | self.fp8_linear.input_scale.fill_(0.125) |
| 204 | |
| 205 | |
| 206 | class TestAttentionNvfp4QuantPatternModel(AttentionQuantPatternModel): |
| 207 | """Test model for AttentionNvfp4QuantPattern fusion.""" |
| 208 | |
| 209 | quant_key = kNvfp4Dynamic |
| 210 | |
| 211 | def __init__(self, *args, **kwargs): |
| 212 | super().__init__(*args, **kwargs) |
| 213 | |
| 214 | hidden_size = self.num_qo_heads * self.head_size |
| 215 | self.w = kwargs.get( |
| 216 | "w", |
| 217 | { |
| 218 | "weight": torch.randint( |
| 219 | 256, |
| 220 | (hidden_size, hidden_size // 2), |
| 221 | dtype=FP4_DTYPE, |
| 222 | device=self.device, |
| 223 | ), |
| 224 | "wscale_swizzled": torch.randn(hidden_size, hidden_size // 16).to( |
| 225 | dtype=FP8_DTYPE, device=self.device |
| 226 | ), |
| 227 | "wscale": torch.tensor([500], dtype=torch.float32, device=self.device), |
| 228 | "scale": torch.tensor([0.002], dtype=torch.float32, device=self.device), |
| 229 | }, |
| 230 | ) |
| 231 | |
| 232 | def forward(self, q: torch.Tensor, k: torch.Tensor, v: torch.Tensor): |
| 233 | """Forward pass that creates the pattern to be fused.""" |
| 234 | attn_output = self.attn(q, k, v) |
| 235 | quant_output, output_block_scale = scaled_fp4_quant( |
| 236 | attn_output, 1 / self.w["scale"] |
| 237 | ) |
| 238 | return cutlass_scaled_fp4_mm( |
| 239 | a=quant_output, |
| 240 | b=self.w["weight"], |
| 241 | block_scale_a=output_block_scale, |
| 242 | block_scale_b=self.w["wscale_swizzled"], |
| 243 | alpha=self.w["scale"] * self.w["wscale"], |
| 244 | out_dtype=attn_output.dtype, |
| 245 | ) |
| 246 | |
| 247 | |
| 248 | PATTERN_TEST_MODELS_FP8: list[tuple[str, type]] = [] |
| 249 | PATTERN_TEST_MODELS_FP4: list[tuple[str, type]] = [] |
| 250 | HEADS: list[tuple[int, int]] = [] |
| 251 | BACKENDS_FP8: list[AttentionBackendEnum] = [] |
| 252 | BACKENDS_FP4: list[AttentionBackendEnum] = [] |
| 253 | |
| 254 | if current_platform.is_cuda(): |
| 255 | HEADS = [(64, 8), (40, 8)] |
| 256 | PATTERN_TEST_MODELS_FP8 = [ |
| 257 | ( |
| 258 | "RedHatAI/Meta-Llama-3.1-8B-FP8", |
| 259 | TestAttentionFp8StaticQuantPatternModel, |
| 260 | ) |
| 261 | ] |
| 262 | PATTERN_TEST_MODELS_FP4 = [ |
| 263 | ( |
| 264 | "nvidia/Llama-3.1-8B-Instruct-NVFP4", |
| 265 | TestAttentionNvfp4QuantPatternModel, |
| 266 | ) |
| 267 | ] |
| 268 | BACKENDS_FP8 = [AttentionBackendEnum.TRITON_ATTN, AttentionBackendEnum.FLASHINFER] |
| 269 | BACKENDS_FP4 = [AttentionBackendEnum.FLASHINFER] |
| 270 | |
| 271 | elif current_platform.is_rocm(): |
| 272 | HEADS = [(32, 8), (40, 8)] |
| 273 | PATTERN_TEST_MODELS_FP8 = [ |
| 274 | ("amd/Llama-3.1-8B-Instruct-FP8-KV", TestAttentionFp8StaticQuantPatternModel) |
| 275 | ] |
| 276 | BACKENDS_FP8 = [ |
| 277 | AttentionBackendEnum.ROCM_AITER_UNIFIED_ATTN, |
| 278 | AttentionBackendEnum.ROCM_ATTN, |
| 279 | AttentionBackendEnum.TRITON_ATTN, |
| 280 | ] |
| 281 | |
| 282 | |
| 283 | @pytest.mark.parametrize("num_qo_heads, num_kv_heads", HEADS) |
| 284 | @pytest.mark.parametrize("head_size", [128]) |
| 285 | @pytest.mark.parametrize( |
| 286 | "batch_size", [7, 256, 533] if current_platform.is_cuda() else [8] |
| 287 | ) |
| 288 | @pytest.mark.parametrize("dtype", [torch.bfloat16, torch.float16]) |
| 289 | @pytest.mark.parametrize("kv_cache_dtype", ["fp8", "nvfp4"]) |
| 290 | @pytest.mark.parametrize( |
| 291 | "backend, model_name, model_class, custom_ops", |
| 292 | # Test attention+quant_fp8 fusion with custom and torch impls of QuantFP8 |
| 293 | list( |
| 294 | flat_product( |
| 295 | BACKENDS_FP8, PATTERN_TEST_MODELS_FP8, ["+quant_fp8", "-quant_fp8"] |
| 296 | ) |
| 297 | ) |
| 298 | # quant_fp4 only has the custom impl |
| 299 | + list(flat_product(BACKENDS_FP4, PATTERN_TEST_MODELS_FP4, [""])), |
| 300 | ) |
| 301 | @pytest.mark.skipif( |
| 302 | not current_platform.is_cuda_alike(), reason="Only test ROCm or CUDA" |
| 303 | ) |
| 304 | @pytest.mark.skipif(not current_platform.supports_fp8(), reason="Need FP8") |
| 305 | def test_attention_quant_pattern( |
| 306 | num_qo_heads: int, |
| 307 | num_kv_heads: int, |
| 308 | head_size: int, |
| 309 | batch_size: int, |
| 310 | dtype: torch.dtype, |
| 311 | kv_cache_dtype: CacheDType, |
| 312 | custom_ops: str, |
| 313 | model_name: str, |
| 314 | model_class: type[AttentionQuantPatternModel], |
| 315 | backend: AttentionBackendEnum, |
| 316 | dist_init, |
| 317 | disable_vllm_compile_cache, |
| 318 | ): |
| 319 | """Test AttentionStaticQuantPattern fusion pass.""" |
| 320 | if backend == AttentionBackendEnum.FLASHINFER and ( |
| 321 | not current_platform.is_device_capability((10, 0)) or not has_flashinfer() |
| 322 | ): |
| 323 | # This also captures the FP4 case |
| 324 | pytest.skip("FlashInfer attn fusion requires Blackwell and flashinfer") |
| 325 | if kv_cache_dtype == "nvfp4" and backend != AttentionBackendEnum.FLASHINFER: |
| 326 | pytest.skip("NVFP4 KV cache is only supported by FlashInfer") |
| 327 | |
| 328 | # trtllm-gen has no NVFP4-KV kernel with NVFP4 output, so with an NVFP4 KV |
| 329 | # cache the FP4 output fusion must be rejected: the standalone |
| 330 | # scaled_fp4_quant stays in the graph and the result is unchanged. FP8 output |
| 331 | # fusion with NVFP4 KV is supported. |
| 332 | expect_fusion = not ( |
| 333 | kv_cache_dtype == "nvfp4" and model_class.quant_key is kNvfp4Dynamic |
| 334 | ) |
| 335 | |
| 336 | custom_ops_list = custom_ops.split(",") if custom_ops else [] |
| 337 | |
| 338 | device = torch.device(f"{DEVICE_TYPE}:0") |
| 339 | torch.set_default_dtype(dtype) |
| 340 | torch.manual_seed(42) |
| 341 | |
| 342 | backend_cls = backend.get_class() |
| 343 | block_size = backend_cls.get_preferred_block_size(16) |
| 344 | |
| 345 | model_config = ModelConfig( |
| 346 | model=model_name, |
| 347 | max_model_len=2048, |
| 348 | dtype=dtype, |
| 349 | ) |
| 350 | vllm_config = VllmConfig( |
| 351 | model_config=model_config, |
| 352 | scheduler_config=SchedulerConfig( |
| 353 | max_num_seqs=1024, |
| 354 | max_model_len=model_config.max_model_len, |
| 355 | is_encoder_decoder=model_config.is_encoder_decoder, |
| 356 | ), |
| 357 | compilation_config=CompilationConfig( |
| 358 | mode=CompilationMode.VLLM_COMPILE, |
| 359 | custom_ops=custom_ops_list, |
| 360 | ), |
| 361 | cache_config=CacheConfig(cache_dtype=kv_cache_dtype), |
| 362 | attention_config=AttentionConfig(backend=backend), |
| 363 | ) |
| 364 | |
| 365 | # Create test inputs |
| 366 | q = torch.randn(batch_size, num_qo_heads * head_size, dtype=dtype, device=device) |
| 367 | k = torch.randn(batch_size, num_kv_heads * head_size, dtype=dtype, device=device) |
| 368 | v = torch.randn(batch_size, num_kv_heads * head_size, dtype=dtype, device=device) |
| 369 | |
| 370 | # Mark first dimension as dynamic for realistic testing |
| 371 | torch._dynamo.mark_dynamic(q, 0) |
| 372 | torch._dynamo.mark_dynamic(k, 0) |
| 373 | torch._dynamo.mark_dynamic(v, 0) |
| 374 | |
| 375 | # Run model directly without compilation and fusion |
| 376 | vllm_config_unfused = copy.deepcopy(vllm_config) |
| 377 | with ( |
| 378 | set_current_vllm_config(vllm_config_unfused), |
| 379 | set_forward_context(attn_metadata=None, vllm_config=vllm_config_unfused), |
| 380 | ): |
| 381 | model_unfused = model_class( |
| 382 | num_qo_heads=num_qo_heads, |
| 383 | num_kv_heads=num_kv_heads, |
| 384 | head_size=head_size, |
| 385 | device=device, |
| 386 | vllm_config=vllm_config_unfused, |
| 387 | block_size=block_size, |
| 388 | ) |
| 389 | model_unfused = model_unfused.to(device) |
| 390 | result_unfused_0 = model_unfused(q, k, v) # noqa: F841 HACK: See #131044 |
| 391 | |
| 392 | forward_ctx = get_forward_context() |
| 393 | forward_ctx.attn_metadata = model_unfused.build_attn_metadata(batch_size) |
| 394 | |
| 395 | # Run model directly without fusion |
| 396 | # Still compile so query QuantFP8 has closer numerics |
| 397 | compiled_unfused = torch.compile(model_unfused, fullgraph=True) |
| 398 | result_unfused = compiled_unfused(q, k, v) |
| 399 | |
| 400 | # Run model with attn fusion enabled |
| 401 | vllm_config.compilation_config.pass_config = PassConfig( |
| 402 | fuse_attn_quant=True, eliminate_noops=True |
| 403 | ) |
| 404 | with ( |
| 405 | set_current_vllm_config(vllm_config), |
| 406 | set_forward_context(attn_metadata=None, vllm_config=vllm_config), |
| 407 | ): |
| 408 | model_fused = model_class( |
| 409 | num_qo_heads=num_qo_heads, |
| 410 | num_kv_heads=num_kv_heads, |
| 411 | head_size=head_size, |
| 412 | device=device, |
| 413 | vllm_config=vllm_config, |
| 414 | w=model_unfused.w, |
| 415 | block_size=block_size, |
| 416 | ) |
| 417 | model_fused = model_fused.to(device) |
| 418 | |
| 419 | forward_ctx = get_forward_context() |
| 420 | forward_ctx.attn_metadata = model_fused.build_attn_metadata(batch_size) |
| 421 | |
| 422 | # Create test backend with fusion passes enabled |
| 423 | noop_pass = NoOpEliminationPass(vllm_config) |
| 424 | attn_pass = LazyInitPass(AttnQuantFusionPass, vllm_config) |
| 425 | cleanup_pass = PostCleanupPass(vllm_config) |
| 426 | |
| 427 | test_backend = TestBackend(noop_pass, attn_pass, cleanup_pass) |
| 428 | # HACK: See https://github.com/vllm-project/vllm/issues/31044 |
| 429 | result_fused_0 = model_fused(q, k, v) # noqa: F841 |
| 430 | |
| 431 | # Compile model with fusion enabled |
| 432 | compiled_fused = torch.compile( |
| 433 | model_fused, backend=test_backend, fullgraph=True |
| 434 | ) |
| 435 | assert compiled_fused.attn._o_scale_float is None |
| 436 | |
| 437 | result_fused = compiled_fused(q, k, v) |
| 438 | |
| 439 | if backend == AttentionBackendEnum.FLASHINFER and expect_fusion: |
| 440 | # With the Flashinfer backend after the 1st round of the forward |
| 441 | # pass, output quant scale should be loaded into the attn layer's |
| 442 | # _o_scale_float, the 2nd round should reuse the loaded |
| 443 | # _o_scale_float |
| 444 | assert compiled_fused.attn._o_scale_float is not None |
| 445 | result_fused_2 = compiled_fused(q, k, v) |
| 446 | |
| 447 | assert compiled_fused.attn._o_scale_float is not None |
| 448 | |
| 449 | torch.testing.assert_close( |
| 450 | result_unfused, result_fused_2, atol=1e-2, rtol=1e-2 |
| 451 | ) |
| 452 | elif not expect_fusion: |
| 453 | assert compiled_fused.attn._o_scale_float is None |
| 454 | |
| 455 | # Check attn fusion support |
| 456 | quant_key: QuantKey = model_class.quant_key |
| 457 | attn_fusion_supported = [ |
| 458 | layer.impl.fused_output_quant_supported(quant_key) |
| 459 | for key, layer in vllm_config.compilation_config.static_forward_context.items() |
| 460 | ] |
| 461 | if expect_fusion: |
| 462 | assert sum(attn_fusion_supported) == len(attn_fusion_supported), ( |
| 463 | "All layers should support attention fusion" |
| 464 | ) |
| 465 | else: |
| 466 | assert sum(attn_fusion_supported) == 0, ( |
| 467 | "No layer should support attention fusion" |
| 468 | ) |
| 469 | |
| 470 | # Check quantization ops in the graph before and after fusion |
| 471 | if "-quant_fp8" in custom_ops_list: |
| 472 | quant_op = torch.ops.aten.reciprocal |
| 473 | elif getattr(model_class, "aiter_quant", False): |
| 474 | quant_op = torch.ops.vllm.rocm_aiter_per_tensor_quant.default |
| 475 | else: |
| 476 | quant_op = QUANT_OPS[quant_key] |
| 477 | |
| 478 | if expect_fusion: |
| 479 | # Note: for fp8, fully_replaced=False because query quant ops remain in |
| 480 | # graph. Only output quant ops are fused into attention. |
| 481 | test_backend.check_before_ops( |
| 482 | [quant_op], fully_replaced=quant_key is kNvfp4Dynamic |
| 483 | ) |
| 484 | else: |
| 485 | # Fusion rejected: the standalone output quant must survive untouched. |
| 486 | num_pre = test_backend.op_count(quant_op, before=True) |
| 487 | assert num_pre > 0, f"Op {quant_op} not found in pre-pass graph" |
| 488 | assert test_backend.op_count(quant_op) == num_pre, ( |
| 489 | "Output quant op should remain when fusion is rejected" |
| 490 | ) |
| 491 | |
| 492 | # access the underlying `AttnQuantFusionPass` on the `LazyInitPass` |
| 493 | assert attn_pass.pass_.matched_count == sum(attn_fusion_supported) |
| 494 | |
| 495 | # Check attention ops in the graph before and after fusion |
| 496 | attn_nodes_pre = list(find_op_nodes(ATTN_OP, test_backend.graph_pre_pass)) |
| 497 | attn_nodes_post = list(find_op_nodes(ATTN_OP, test_backend.graph_post_pass)) |
| 498 | |
| 499 | assert len(attn_nodes_pre) > 0, "Should have attention nodes before fusion" |
| 500 | assert len(attn_nodes_pre) == len(attn_nodes_post), ( |
| 501 | "Should have same number of attention nodes before and after fusion" |
| 502 | ) |
| 503 | assert attn_nodes_pre[0].kwargs.get("output_scale") is None, ( |
| 504 | "Attention should not have output_scale before fusion" |
| 505 | ) |
| 506 | if expect_fusion: |
| 507 | assert attn_nodes_post[0].kwargs.get("output_scale") is not None, ( |
| 508 | "Attention should have output_scale after fusion" |
| 509 | ) |
| 510 | else: |
| 511 | assert attn_nodes_post[0].kwargs.get("output_scale") is None, ( |
| 512 | "Attention should not have output_scale when fusion is rejected" |
| 513 | ) |
| 514 | |
| 515 | assert attn_nodes_pre[0].kwargs.get("output_block_scale") is None, ( |
| 516 | "Attention should not have output_block_scale before fusion" |
| 517 | ) |
| 518 | |
| 519 | kv_cache_dummy_dep_pre_is_none = ( |
| 520 | attn_nodes_pre[0].kwargs.get("kv_cache_dummy_dep") is None |
| 521 | ) |
| 522 | kv_cache_dummy_dep_post_is_none = ( |
| 523 | attn_nodes_post[0].kwargs.get("kv_cache_dummy_dep") is None |
| 524 | ) |
| 525 | assert not (kv_cache_dummy_dep_pre_is_none ^ kv_cache_dummy_dep_post_is_none), ( |
| 526 | "The kv_cache_dummy_dep should be consistent before and after fusion" |
| 527 | ) |
| 528 | |
| 529 | if quant_key.dtype == FP8_DTYPE or not expect_fusion: |
| 530 | assert attn_nodes_post[0].kwargs.get("output_block_scale") is None, ( |
| 531 | "Attention should not have output_block_scale unless FP4 fusion happened" |
| 532 | ) |
| 533 | elif quant_key.dtype == FP4_DTYPE: |
| 534 | assert attn_nodes_post[0].kwargs.get("output_block_scale") is not None, ( |
| 535 | "Attention should have output_block_scale after FP4 fusion" |
| 536 | ) |
| 537 | |
| 538 | # Check that results are close |
| 539 | torch.testing.assert_close(result_unfused, result_fused, atol=1e-2, rtol=1e-2) |
| 540 | if getattr(model_class, "aiter_quant", False): |
| 541 | # Query quantization remains; exactly the output quant op is removed. |
| 542 | assert test_backend.op_count(quant_op, before=True) == ( |
| 543 | test_backend.op_count(quant_op) + 1 |
| 544 | ) |
| 545 | for model in (model_unfused, model_fused): |
| 546 | torch.testing.assert_close( |
| 547 | model.fp8_linear.input_scale, |
| 548 | torch.full_like(model.fp8_linear.input_scale, 0.125), |
| 549 | atol=0, |
| 550 | rtol=0, |
| 551 | ) |
| 552 | |
| 553 | |
| 554 | @pytest.mark.skipif( |
| 555 | not current_platform.is_rocm() or not is_aiter_found_and_supported(), |
| 556 | reason="Requires ROCm AITER attention and quantization", |
| 557 | ) |
| 558 | @pytest.mark.parametrize("dtype", [torch.bfloat16, torch.float16]) |
| 559 | def test_rocm_aiter_static_quant_attention( |
| 560 | dtype, dist_init, disable_vllm_compile_cache, monkeypatch |
| 561 | ): |
| 562 | """Fuse the actual AITER op without changing its static scale or outputs.""" |
| 563 | try: |
| 564 | with monkeypatch.context() as patch: |
| 565 | patch.setenv("VLLM_ROCM_USE_AITER", "1") |
| 566 | rocm_aiter_ops.refresh_env_variables() |
| 567 | test_attention_quant_pattern( |
| 568 | num_qo_heads=32, |
| 569 | num_kv_heads=8, |
| 570 | head_size=128, |
| 571 | batch_size=8, |
| 572 | dtype=dtype, |
| 573 | kv_cache_dtype="fp8", |
| 574 | custom_ops="+quant_fp8", |
| 575 | model_name="amd/Llama-3.1-8B-Instruct-FP8-KV", |
| 576 | model_class=TestAttentionAiterFp8StaticQuantPatternModel, |
| 577 | backend=AttentionBackendEnum.ROCM_AITER_UNIFIED_ATTN, |
| 578 | dist_init=dist_init, |
| 579 | disable_vllm_compile_cache=disable_vllm_compile_cache, |
| 580 | ) |
| 581 | finally: |
| 582 | rocm_aiter_ops.refresh_env_variables() |